acceptodds
Under review as a conference paper at ICLR 2027

A General and Efficient SE(3)-Equivariant Graph Framework: Encoding Symmetries with Complete Differential Invariants and Frames

Abstract

Equivariant graph neural networks (Equiv-GNNs) have demonstrated effectiveness in modeling dynamics of multi-object systems via explicit symmetry encoding. Scalarization-based Equiv-GNNs are widely adopted in practice for their superior computational efficiency, as opposed to high-steerable models. However, most existing scalarization-based approaches rely on the empirical design of invariant scalars, which fail to fully and systematically capture system dynamics, particularly when object differential components like velocities are involved, leading to limited prediction performance. To address these issues, we propose a general, efficient, and lightweight SE(3)-equivariant graph framework with Complete Differential Invariants and Frames (CDIF). Specifically, we first systematically construct a maximal functionally independent set of basic invariant scalars derived from system states, providing a theoretically complete representational basis for arbitrary SE(3)-invariant functions. We then design local frames between objects that integrate both positional and differential components to recover directional information from the constructed invariants, which enables robust recovery of directional information while enhancing the model’s capability of dynamic modeling. Extensive experiments across diverse domains including molecular dynamics, formation control, human motion capture and particle simulation, validate that our CDIF framework outperforms state-of-the-art baselines on most benchmarks, and exhibits strong scalability for large-scale multi-object scenarios.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural network, Equivariant graph neural networks, Multi-object systems dynamics modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_WZzBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CDIF, an SE(3)-equivariant graph framework for modeling the dynamics of interacting objects. After centralizing positions to remove translational effects, the method constructs scalar features from the positions and velocities of two objects connected by an edge. A differential-invariant argument motivates a set of nine functionally independent invariants, while positional and velocity frames provide directions for equivariant state updates. The model also uses SE attention to gate the invariant features.

The experiments show that the framework achieves competitive performance. However, its claims of invariant and frame completeness are insufficiently justified. Given the central role of these claims in the paper’s contribution, I cannot recommend acceptance.

Critical strengths

S1. The empirical evaluation is broad and the reported performance is competitive. The experiments cover several distinct types of dynamics. CDIF achieves the best performance on six of ten MD17 molecules, all six formation-control configurations, two of three motion-capture prediction horizons, and four of six charged-particle configurations. These results support the practical usefulness of the architecture across different dynamics-prediction tasks.

S2. The invariant-transfer experiments provide useful evidence beyond the complete CDIF architecture. Adding the selected invariant features to EGNN and GMN improves all reported results in the corresponding comparisons, supporting their usefulness as geometric inputs. The discussion in Appendix H.3 also acknowledges the larger invariant collection’s advantage on the difficult task, providing a balanced interpretation of the feature-selection tradeoff.

Critical weaknesses

W1. The global completeness and universal-approximation claims are not valid as stated. Theorem 3.1 claims that any SO(3)-invariant scalar function of the two object states can be expressed as a function of the nine invariants in Eq. (7). However, the proof relies on local functional-dependence and differential-invariant results. The dimension count establishes that the generic quotient has nine local degrees of freedom; it does not establish that the particular nine quantities in Eq. (7) distinguish all SO(3) orbits.

A concrete counterexample is obtained by fixing qi=(1,0,0),qj=(0,1,0),q˙i=(1,1,1),q_i=(1,0,0),\quad q_j=(0,1,0),\quad \dot q_i=(1,1,1), and comparing q˙j=(1,2,0)\dot q_j=(1,2,0) and q˙j=(0,2,1)\dot q_j=(0,2,1). The nine invariants in Eq. (7) are identical in both cases in the manuscript. Nevertheless, the SO(3)-invariant polynomial qi⊤q˙jq_i^\top\dot q_j equals 11 and 00, respectively. Therefore, no single-valued function of the selected nine invariants can represent qi⊤q˙jq_i^\top\dot q_j on both configurations.

Both configurations have nondegenerate positional and velocity frames, and the invariant Jacobian has rank nine at both states. The authors should distinguish local functional independence from global completeness, specify the intended domain, and revise the universal-approximation claims accordingly. This counterexample concerns the stated scalar factorization through Eq. (7). It does not, by itself, establish an impossibility result for the entire multilayer network.

W2. The claimed completeness of the differential frames is not established. The frames in Eq. (10) form orthonormal bases only under nondegeneracy conditions. Adding a velocity frame does not guarantee that all admissible output directions remain representable. For example, let qi=e1,qj=2e1,q˙i=e2,q˙j=2e2,q_i=e_1,\quad q_j=2e_1,\quad \dot q_i=e_2,\quad \dot q_j=2e_2,where e1,e2,e3e_1,e_2,e_3 are the standard coordinate vectors. Both qi×qjq_i\times q_j and q˙i×q˙j\dot q_i\times\dot q_j vanish, making both proposed frames undefined. Adding epsilon to the normalization denominators leaves the available frame directions within span⁡{e1,e2}\operatorname{span}\{e_1,e_2\}. However, the valid SO(3)-equivariant map g(qi,qj,q˙i,q˙j)=qi×q˙i=e3g(q_i,q_j,\dot q_i,\dot q_j)=q_i\times\dot q_i=e_3produces a direction that their linear combination cannot represent.

This example identifies a limitation of the chosen frame construction, rather than a restriction imposed by equivariance itself. The authors should specify the domain and meaning of completeness, explain how degenerate states are handled, and establish that the selected invariant features suffice to represent the required frame coefficients. Reducing the probability of degeneracy alone does not establish a completeness guarantee.

W3. The ablations do not cleanly isolate the benefit of the selected invariant set. Table 8 supports a task-dependent feature-selection tradeoff. Although the proposed nine-invariant configuration performs better on the easy and medium tasks, the larger collection improves the difficult-task error from 0.32420.3242 to 0.27810.2781. Moreover, Model 6 includes additional derived quantities and repeated triple products up to sign. Consequently, this experiment does not cleanly establish the practical advantage of reducing a standard sufficient generating set to the selected nine features. A controlled comparison on an existing task, or a more modest interpretation of the reported evidence, would strengthen the contribution.

Questions for authors

See weaknesses.

Reviewer_hd3ROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a novel SE(3) equivariant graph neural network based on scalarization, with a specific focus on differential components such as velocities. The authors provably demonstrate a complete basis for SO(3) functions, and design their method based on two edge-wise local frames.

Experiments on common datasets in the field, such as n-body systems, motion capture, and MD17 demonstrate the effectiveness of the proposed method, which outperforms or is on par with state-of-the-art baselines.

Critical strengths

The theoretical framework of scalarization-based equivariant GNNs is a very important and useful contribution in the field of geometric deep learning. The proposed method is mostly straightforward and clear.

Critical weaknesses

The experimental section largely includes datasets which are small-scale and quite saturated, namely MD17, Charged particles, and Motion Capture. It is quite hard to assess the performance of a new method in such datasets, and whether the method scales gracefully to datasets with more nodes.

The source code is not provided, which makes it hard to verify the validity and effectiveness of the proposed method.

Questions for authors
  1. What is SE attention mentioned in Section 4?
  2. There are no ablation in the main manuscript. What is the importance of each of the nine CDIs? What about the two local frames? How does the method perform with subsets of those two design choices?
  3. How do the nine invariants scale in the presence of more input variables, e.g. acceleration, or orientation?
Reviewer_rwigOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CDIF for Equivariant graph neural networks (Equiv-GNNs) and applies it to various applications including molecular dynamics, CMU frames demonstrating generalization of the approach with moderate gains in accuracy and other metrics across these compared to many of the state of the art techniques. While I agree with the central approach of how they have approached the scalarization approaches, I think there needs to be substantial justification/ thinking behind how they have currently presented their central hypothesis: namely that because the maximum number of functionally independent invariants is nine, any nine functionally independent invariants form a basis through which every SO(3)-invariant function factors, and then asserts universal approximation on that basis. Olver's work in this area that they refer to in the appendix as a Lemma, gives existence of H only locally in a neighborhood of a point, which seems to be generalized to the global context, which I am not entirely convinced (see below).

Critical strengths

In my opinion, the major strength of the paper is in the approach it seeks to establish the CDIF as a framework and the theoretical underpinning it proposes. The architecture is sensible, practical and has some interesting results that seem to hold (although I did not test the code provided). The direction of asking which invariant set is sufficient is a good one and I think the approach summarizes these well in the paper, with substantial theoretical backing. Although I have some concerns in the presentation about this (see below), I think it may be fixable.

Critical weaknesses

The critical weakness is that I am not convinced that the completeness argument (Theorem 3.1) holds. While I agree with the authors with the 9 functionally independent invariants, The authors seem to conclude that because the maximum number of functionally independent invariants is nine, any nine functionally independent invariants form a basis through which every SO(3)-invariant function factors, and then assert universal approximation on that basis. However, functional independence and orbit separation are tow different things. The orbit space here is a 9-dimensional semi algebraic set with folds, not an open subset of R^9, and a set of nine independent invariants can be generically two-to-one on it.

Questions for authors

The specific set (7) presented in the paper contains eight of the ten pairwise inner products, however q_i·q̇_j and q_j·q̇_i are both absentif the one triple product. If we consider the following: take configuration A with q_i = e_1, q̇_i = e_3, q_j = e_2, q̇_j = e_1, and configuration B identical except q̇_j = −e_1. All nine invariants in (7) will agree: the four norms are 1, q_i·q̇_i = q_j·q̇_j = q_i·q_j = q̇_i·q̇_j = 0, and q̇_i·(q_i × q_j) = 1 in both cases. But q_i, q_j, q̇_i span R^3, so any rotation taking A to B is the identity, and the identity does not take e_1 to −e_1. The two configurations lie in distinct SO(3) orbits while agreeing on all nine invariants, and q_i·q̇_j, an SO(3)-invariant function, takes values +1 and −1 on them. No function H of the nine can represent it. Based on this (counter)example, I'd suggest either exhibit H explicitly for q_i·q̇_j or restate the theorem as a local result about functional independence?

The root cause is that a generating set and a maximal functionally independent set are not the same size. By the first fundamental theorem, one needs all ten inner products plus a determinant, with the Gram relations among them, eleven functions satisfying relations, not nine independent ones. This is what Villar et al showed in their prior work (2021) and I think that still holds.

Within the context of frame construction, I have a few more questions/ clarifying thoughts that would be useful.

-Within the molecular dynamics MD17 dataset, after centroid-centering, collinearity of q_i and q_j means two atoms nearly radially aligned with the molecular centroid, which is common in MD17’s elongated molecules. It would be good to cover what the implementation does exactly, namely, either clamping, epsilon, Gram-Schmidt fallback, as well as empirical frequency of near-degenerate edges.

-There seems to be a typo within Eq. 28 for the velocity updates, with each of the velocity terms appearing twice while the positional frame absent entirely? That directly contradicts the Section 3.2 claim that cross-frame coupling drives both position and velocity updates in my opinion. Can the authors clarify explicitly what was being done? As presented, either it’s a typo, in which case the paper’s second contribution is misstated in its only formal specification, or it’s in the code, in which case the ablation story should be updated accordingly.

Another question I have is the circularity in the argument about how CDIF benchmarks are tested. The FC interaction coefficients in Eqs. (69)–(71) are built from q_i·q_j, q̇_i·q̇_j, and q̇_i·(q_i × q_j), three members of CDIF’s invariant set. Based on the presentation, CDIF wins all six FC tasks, by wide margins, on a benchmark the authors designed whose ground-truth dynamics are polynomial in its own features. On the three pre-existing benchmarks the picture seems to be a more mixed bag: e.g., PONITA beats CDIF on four of ten MD17 molecules, including a large margin on Aspirin and Malonaldehyde. This should be stated plainly in the paper rather than summarized as outperforming on most benchmarks?

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

CDIF combines position–velocity scalar features with position- and velocity-based local frames for SE(3)-equivariant dynamics prediction. It reports improvements across molecular dynamics, human motion, charged particles, and formation control, using a straightforward message-passing architecture. The results suggest practical value, but the central completeness claim is not justified as stated. The contribution also needs to be distinguished more clearly from existing scalarization and local-frame methods, and the evaluation misses several strong, directly relevant dynamics models. These issues leave me unconvinced by the paper's broader theoretical and state-of-the-art claims.

Critical strengths
  • Reported gains across several domains. Within the reported comparisons, CDIF achieves the lowest error on 6/10 MD17 tasks, 2/3 CMU Motion tasks, and 4/6 charged-particle tasks.
  • Reusable features. Adding CDIs to EGNN and GMN improves their reported accuracy, suggesting that these features can be useful beyond the full CDIF architecture.
  • Straightforward implementation. Dot products, cross products, and MLPs fit naturally into a standard message-passing pipeline, without requiring a separate higher-order tensor representation system.
Critical weaknesses
  • The global completeness claim does not hold as stated. Theorem 3.1 uses a local dimension count to justify global completeness of the specific nine invariants. Consider centered positions pi=(1,0,0)p_i=(1,0,0), pj=(0,1,0)p_j=(0,1,0) and velocity vi=(1,1,1)v_i=(1,1,1), with either vj=(1,1,0)v_j=(1,1,0) or vj=(0,1,1)v_j=(0,1,1). All nine features in Eq. (7) agree, while the rotation-invariant quantity pi⊤vjp_i^\top v_j differs. Adding a third position (−1,−1,0)(-1,-1,0) satisfies centering. This challenges the representation guarantee.

  • The novelty needs sharper justification. Previous work [1] already characterized SO(d)-invariant functions through inner products and determinants. ClofNet [2] uses the same positional frame and explicitly supports velocity scalarization. The remaining contribution appears to be the particular nine-feature selection and dual-frame parameterization. Comparisons against full velocity projections and Gram-plus-triple-product features are needed to establish their added value.

  • More importantly, the evaluation omits directly relevant models, such as ESTAG [3], EGNO [4], SEGNO [5] and GF-NODE [6]. These address overlapping dynamics-prediction tasks. Setting temporal models aside because they have a different architectural focus does not resolve this gap in the end-to-end evaluation.

  • Degenerate frames are not adequately addressed. Collinear positions, collinear velocities, or identical velocities produce zero denominators in Eq. (10). A second frame can help in some cases, but does not guarantee a valid construction. The numerical handling and robustness near these configurations remain unclear.

  • The formation-control comparison has a symmetry mismatch. Its velocity cross-product terms generally respect SO(3), but not full O(3) equivariance for ordinary vector states. Strict E(3)/O(3)-equivariant baselines therefore face a structural limitation. Some of the advantage may reflect the choice of symmetry group rather than the proposed features or frames.

I would be inclined to raise my score if the authors address the weaknesses noted above.


[1] Villar, Soledad, et al. "Scalars are universal: Equivariant machine learning, structured like classical physics." Advances in neural information processing systems 34 (2021): 28848-28863.

[2] Du, Weitao, et al. "SE (3) equivariant graph neural networks with complete local frames." International Conference on Machine Learning. PMLR, 2022.

[3] Wu, Liming, et al. "Equivariant spatio-temporal attentive graph networks to simulate physical dynamics." Advances in Neural Information Processing Systems 36 (2023): 45360-45380.

[4] Xu, Minkai, et al. "Equivariant graph neural operator for modeling 3d dynamics." arXiv preprint arXiv:2401.11037 (2024).

[5] Liu, Yang, et al. "SEGNO: Generalizing equivariant graph neural networks with physical inductive biases." International Conference on Learning Representations. Vol. 2024. 2024.

[6] Sun, Fang, et al. "Graph Fourier neural ODEs: Modeling spatial-temporal multi-scales in molecular dynamics." arXiv preprint arXiv:2411.01600 (2024).

Questions for authors

see weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.