acceptodds
Under review as a conference paper at ICLR 2027

Weyl Algebra Equivariant Neural Networks

Abstract

Encoding physical symmetries into neural network architectures is key to improving generalization and physical consistency in domains such as molecular dynamics and particle physics. Existing equivariant networks rely primarily on Clifford algebras, which realize orthogonal-group equivariance via the sandwich action, or on partial differential operator (PDO) networks, which implicitly exploit Weyl algebra structure without an explicit unifying framework. Both approaches are limited: Clifford algebras cannot capture symplectic or differential-operator symmetries, while PDO networks lack a common algebraic language across groups. We show that the Weyl algebra—generated by position and momentum operators under the canonical commutation relation—unifies both lines of work: its subalgebra structure recovers orthogonal-group equivariance, while its differential-operator generators extend naturally to the symplectic group. Building on this, we introduce a Weyl-equivariant neural network supporting both symmetries within a single architecture, along with degree-truncation approximations for computational tractability. Empirical results on tasks involving symplectic and orthogonal symmetries demonstrate that our method matches or outperforms existing equivariant networks while offering substantially broader symmetry coverage.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Equivariant Neural Networks, Weyl Algebra, Symplectic Group, Orthogonal Group

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_wEeFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies Weyl algebra equivariant neural networks as a unificatory framework for symplectic symmetries. The idea is interesting, but the paper seems to have several mathematical flaws as explained below and so I cannot recommend acceptance of the paper.

Critical strengths

The idea of Weyl algebra equivariant neural networks is mathematically interesting and potentially valuable (although applications of symplectic geometry in machine learning already exist, I have never seen this specific type of application to equivariant neural nets).

Critical weaknesses

The paper includes several mathematical flaws. For example, the symplectic-harmonic decomposition in Section 5 is not correct: symmetric powers of the symplectic representation are irreducible. The paper also confuses symplecticity of Hamiltonian dynamics with equivariance under the full symplectic group. And the regression target is not symplectically invariant since it contains Euclidean norms in the denominator, which seems to invalidate the symmetry-based interpretation of the experiment.

In addition to these mathematical issues, practical applicability in ML is not fully convincingly established and the proposed method seems still premature to be accepted at a top conference like ICLR.

Questions for authors

Can you fix the above-mentioned mathematical issues?

Reviewer_k9h9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors use the Weyl algebra to build symplectic-group equivariant networks following the CGENN approach. They apply the method to several toy experiments with Sp(2n) symmetry and find good performance. While the idea is sound, the paper is not polished enough to be considered for acceptance, as it lacks essential information about the method and experiments, does not follow ICLR's AI use and reproducibility requirements, and the writing is poor (LLM style). I suspect the authors are aware of this, and I ask them in the future to consider if a paper is truly ready before asking for the reviewers' time.

Critical strengths
  • The authors extend the CGENN approach to the Weyl algebra, identifying the critical challenges and addressing them in their network design.
Critical weaknesses
  • The paper reads like work in progress: The main paper has only 8 pages, probably because the authors removed one experiment last-minute which is still referred to in the text. There is only one page of appendix, even though there are many things that are not properly explained in the main paper. The main paper is 50% math, yet the paper does not have a single math reference. There are no ablations. Reproducibility is poor because the authors did not submit their code, there are no details on the Hamiltonian dynamics prediction, and the existing experiment descriptions are very short. Some results do not come with uncertainties. Limitations are not discussed sufficiently. The AI use statement does not list all the required points in https://iclr.cc/Conferences/2027/AIPolicyForAuthors. The language is poor (LLM style), making it hard to read the paper.
  • The paper is not properly embedded into the literature, making it sound like the paper is a breakthrough even though they merely combine a bunch of building blocks to extend the CGENN approach to symplectic groups. They fail to discuss the relation to prior work on Sp(2n) equivariance, even though this is central to the paper.
Questions for authors
  • The main paper contains four different normalization strategies, each with specific pros and cons. The authors should settle on one approach as the default and discuss the others in an appendix.
  • L344 says that the paper includes an molecular dynamics trajectories experiment, but there is no such experiments. The paper refers to O(n) experiments several times, probably meaning this molecular dynamics example. It seems that the authors removed this experiment shortly before the deadline and failed to consistently update the paper.
  • Appendix A does not contain the full proof for Proposition 1, in contrast to the claim in L209
  • In the Hamiltonian dynamics prediction study (table 2) the WeylNN MSE is equivalent to SEGNN within uncertainties while SEGNN has slightly lower symplectic error, but the text claims WeylNN is the best
  • Table 1 and the PETRA study in Figure 1 should have uncertainties to understand if the effects are significant
  • The TM-PNN paper should be cited for the PETRA application. Currently it sounds like the authors came up with this themselves.
  • I don't see how WeylNN unifies also PDO networks
Reviewer_wKXVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a novel architecture equivariant to the symplectic group Sp(2n,R)Sp(2n, \mathcal{R}), generalizing ideas previously used in two separate branches of the equivariant deep learning literature for orthogonal groups equivariance (Clifford algebras and Clebsh-Gordan transforms).

Critical strengths

The paper introduces a novel idea combining elements from existing lines of works in an original way to generalize equivariance to the symplectic group. The proposed method is tested on different real and synthetic datasets, showing consistent gains with respect to the baselines.

Critical weaknesses

Unfortunately, I have multiple serious doubts about this manuscript.

First of all, while very familiar with the equivariant deep learning literature, I have no experience and knowledge on Weyl algebras and Symplectic groups. While I acknowledge the target audience might have more experience on the particular applications considered by the authors, I believe the presentation quality could be significantly improved and the manuscript could be made much more accessible. Throughout the manuscript, a lot of prerequisite knowledge is assumed and it is frequent that mathematical concepts or even specific notations are directly adopted without any previous explanation (see questions below). In particular, the authors seem to interchangeably use Sp(2n)Sp(2n) and Sp(2n,R)Sp(2n, \mathcal{R}), which seem to be two different groups according to Wikipedia.

Moreover, different claims through the paper appear vague (e.g. the relationship with Clifford networks and O(n)O(n) equivariance). The appendix includes some proofs and some minor experiments details, but doesn't add any additional context that could help understand the theory and the architecture.

Another serious concern is about the data. Can you provide some more context about what are the datasets used and why they represent important tasks to solve? Why is symplectic group equivariance important? In particular, the PETRA III dataset comes with no citations, can you provide some details about it? Did you gathered this data yourself? is this a contribution of this paper? Otherwise, where does this data come from exactly?

Finally, a quick search for the irreducible representations of the symplectic group Sp(n)Sp(n) led me to some claims about SympSym^p being an irreducible representation of the symplectic group (https://mathoverflow.net/a/243753). This seems partially in conflict with the decomposition in Sec 5. Considering my limited expertise with these groups I might be missing some important details, but I would appreciate it if the authors could clarify this.

Questions for authors

Def. 2: what is v1v2v_1 v2 and v2v1v2 v1? are those inner products? Or are those some other kind of products? Until there, it seems v1,v2∈Vv1, v2 \in V are simple vectors, so something is unclear. Similarly, what does ⋅1\cdot 1 mean soon after?

Line 111 what is Sm\mathfrak{S}_m? is it the set of permutations of mm elements? However, later in Sec 3.2 this symbol is defined as the quotient Sp(2n,R)/U(n)Sp(2n, \mathbb{R})/U(n), which is adding some confusion.

Sec 5: what is An\mathcal{A}_n? this is introduced as a problematic infinite dimensional construction to justify the following truncation, but An\mathcal{A}_n was not previously introduced and motivated. Why should it be considered in the first place?

Generally speaking, Definition 2 already assumes a lot of knowledge and familiarity with the notation from the reader. While explaining all background knowledge about Weyl Algebra is beyond the scope of the paper, clarifying the notation before using it would make it much easier for non-expert readers to parse the manuscript and fill their knowledge gaps. Also, the authors could consider a chapter in the appendix to provide the reader with the necessary preliminary knowledge.

You occasionally use Sp(2n,R)Sp(2n, \mathbb{R}) and Sp(2n)Sp(2n), is that the same group? Maybe the manuscript could state their equivalence once and then pick a standard notation to use.

Sec 5: is the point here analogous to typical SO(3) models with tensor product non-linearities and Clebsh-Gordan transform, where each tensor-product layer can raise the max degree of the irreps? In this sense, is the learnable projection (ii) solution analogous to the typical choice of keeping only a truncated number of degrees? Maybe a side-by-side analogy with popular equivariant models and groups could help clarifying a lot of details.

Sec 6: the symplectic capacity regression task is not very clear, could this maybe be better explained with a visualization for n=1n=1 or n=2n=2?

Sec 6: you say "SO(2n)SO(2n)-equivariance is too strong a constraint to exploit the symplectic structure fully". Why is that? How do SO(2n)SO(2n) and Sp(2n)Sp(2n) relate exactly? Can you first explain what are exactly the symmetries of the problem? Is it correct to assume Spn(2n)Spn(2n)? Also, what about the Sp(2n)Sp(2n)-MLP? In Fig. 1 it seems SO(2n)SO(2n)-MLP outperforms Sp(2n)Sp(2n)-MLP despite the task being (I assume) Sp(2n)Sp(2n) equivariant: why do you think this is the case?

Generally speaking, could you summarize the main difference between your Sp(2n)Sp(2n)-equivariant model and the Sp(2n)Sp(2n)-MLP baseline? Is it correct the difference boils down to the non-linear layers introduced in Sec 4?

Also, Fig. 1 doesn't show SO(2n)SO(2n)-MLP's performance on the PETRA III task, could you include that baseline? Same for the parity classification experiments in Table 1: why no SO(2n)SO(2n) baseline is considered?

I do not understand the point of the claim regarding the orthogonal group equivariance of Clifford-algebra networks. Afterall, the resulting architecture is Sp(2n)Sp(2n)-equivariant no O(n)O(n)-equivariant. Is this claim saying that the current architecture would be equivalent to Clifford-algebra networks if the equivariance constraints would be restricted to O(n)O(n)?

The paper concludes with a brief mention of computational limitations. Could the author further expand on that?

The appendix provides very little additional details. Could the author share some more details about the underlying implementation? Maybe a schematic overview of the architecture and of each of its layers could help clarify.

How is the integral in the Moyal trace computed for the normalization N3?

Could you provide the exact formulation of normalization N4? It's hard to understand how that works and if it is really equivariant.

Since you use some non-equivariant normalization layers, could you provide some quantitative analysis of the degree of equivariance maintained by the model after training? Is the model still approximately equivariant or does it exploit these layers to break it? Also to what extent do these layer break equivariance in the first place? Could you comment on this?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.