acceptodds
Withdrawn from ICLR 2027

SHNO: Symplectic Operator Learning with Size-Independent Parameter Counts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning Hamiltonian dynamics across system sizes requires reusable parameterizations that preserve symplectic structure. We introduce the Shared Hénon Neural Operator (SHNO), which combines a tied local encoder–decoder pair shared across canonical patches, a shared HénonNet for latent evolution, and optional symplectic edge exchange between patches. The assembled operator preserves the symplectic form on the represented manifold. For fixed local architecture and propagation depth, its parameter count is independent of patch count, while computation scales linearly with the numbers of patches and edges. For compatible local target compositions, we establish a size-uniform normalized approximation bound under uniform block-approximation and Lipschitz assumptions. Experiments on Klein–Gordon, Fermi–Pasta–Ulam–Tsingou, and Toda lattices evaluate prediction accuracy, symplectic structure, and parameter reuse. With separate training at each size, the 5,574-parameter SHNO without edge exchange achieves lower Klein–Gordon rollout errors than Global Hénon ROM, FNO, and a shared residual baseline. Matched unshared models use 8–32 times as many parameters, with accuracy differences depending on the system and size. Fixed-checkpoint tests evaluate deployment from 64 to 128 and 256 canonical pairs without target-size training. On Toda, edge exchange reduces geometric mean rollout error by 62.8% relative to identity coupling with 3.6% more parameters. Code is available anonymously at https://anonymous.4open.science/r/SHNO-6547.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Symplectic operator learning, Hamiltonian dynamics, Structure-preserving learning, Hénon networks, Shared local parameterization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2FauOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Shared Hénon Neural Operator(SHNO), a compositional reduced order model (ROM) for reconstructing dynamics from given phase space observation. It construct a symplectic, one step push forward operator and learn a reusable patch block where repetition of blocks with some choice of graph connectivity of local low-dimensional patched phase space dynamics compose the predictive dynamics from high-dimensional phase state inputs.

Critical strengths
  • The idea of symplectic push forward operator in compositional structure is convincing.
Critical weaknesses
  • The numerical validation scheme is not convincing. The time stepping is fixed step size h=0.01h=0.01(L294-295), with P=r=8P=r=8, and only 3232 trajectories with 6464 time steps are used (Table 3). It is very limited and non-flexible. It would be better if author can test longer horizon (more than 128128 time steps in evaluation per trajectory) and test different patch compositions. There is no numerical algorithm block for readability. There is no visual results in phase space reconstruction quality in KG, FPUT and Toda test cases as well. These would harm the readability for general readers. making it hard to understand what is the actual advantage of SHNO.

  • A minor issue is if the system is no longer symplectic or contain observational noise, symplecticity in ROM might not be helpful. At least it is not validated in this paper.

Questions for authors
  • Have authors considered comparison with other symplectic operator learning scheme (e.g. SympGNNs , ENOs) as Shared residual nor FNO is specialized in this particular phase state space reconstruction task.
  • In Figure 5, I can clearly see energy has a slowly diverging rate up to rollout step 128128. What is the potential reason and can authors rollout 10×10\times longer to validate if energy error is going to be accumulated.
  • What is the result regarding different time stepping size hh, different choice r<Pr<P and different composition of PP? Are results robust to different configuraions of tested system (i.e. would changing (ω,γ)(\omega,\gamma) or (α,β)(\alpha,\beta) stated in L793 affect the result, what about mix-up multiple cases or create in-distribution and out-of-distribution evaluation?).
  • Would result in Table 2 considered as a negative result as transfer from N=64N=64 to N=128,256N=128,256 do lose salient accuracy?
Reviewer_Lme9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This work targets learning Hamiltonian dynamics across different system sizes while preserving symplectic structure and keeping the number of parameters fixed. SHNO shares tied symplectic encoder-decoder and Henon dynamics modules across canonical patches, adds optional symplectic edge updates between patches, and proves symplecticity on the represented manifold, size-indpendent #params, linear computational cost on sparse graphs, and conditional approximation bounds. On Klein-Gordon, FPUT, and Toda lattices, SHNO improves KG/FPUT accracy over its selected baslines using ~5600-5800 params, 8-24 x param savings against unshared controls, successful reuse of uncoupled-model checkpoints at larger sizes, and a 62.8% reduction in Toda rollout error from edge coupling with 3.6% more parameters.

Critical strengths
  • The tied encoder–decoder and symplectic update blocks preserve symplectic structure exactly on the represented state manifold, without relying on a training penalty.
  • Strong parameter efficiency, achieving 8-32x fewer params than the tested unshared counterpart.
  • I find no standing faults in the maths of the manuscript, it seems sound.
  • Good computational scaling
  • Transfer across sizes is demonstrated
  • Effective, inexpensive coupling.
Critical weaknesses
  • Overall, I find the paper very unconvincing on explaining WHY the audience should be interested in this work. Why is the problem impactful and why should a generic AI audience should be interested? What impact would this work bring to the AI community? If this submission is for physics specific venues, the motivation and significance of the problem might be natural; not for ICLR. In the current manuscript I am skeptical the ICLR audiences would be motivated enough to care about this work.
  • There is only one neural operator baseline, FNO, which is now a 6 years old work. Myriads of FNO variants and improvements have been pouring out ever since, authors must include advanced neural operator baselines to convince neural operator community of 2026. Here is the list of relevant baselines authors should consider: TFNO (TMLR 2024), ReFNO (TMLR2023), ENO (AISTATS 2025), Symplectic NO, FFNO (ICLR 2023), FNO with PDE Refiner (NeuRIPS 2023), CNO (NeuRIPS 2023), and also, SympGNN, which is not an FNO variant, but it combines shared parameters, graph interactions, and symplectic updates for high-dim Hamiltonian systems. Authors should include at least 2-3 of the FNO variants and probably SympGNN if they want the neural operator community to take their experiments seriously.
  • The parameter efficiency is apparent. However, the competing baselines are fairly small already. Unshared is #44k~178k, FNO is 12k, the ROM is #23k-62k, I don't see that the absolute storage saving being any issue. Yes the SHNO param is #5k but even the worst baseline occupies only 1.43MB of storage, I don't think this aspect is convincing to be impactful. Unless, authors can convince that the real impactful applications require big scaling, where the memory cost would be a serious issue with prior methods? But the size-independent parameter counts are already present in other baselines like FNO or Shared Residual.
  • Paper does not include comparative training time or inference time study against its baseline. Small #params does not guarantee small computational cost. Also, with non learning methods like ROM, the cost of neural operator method must be compared against them including the training time. Computational efficiency of SHNO is not shown in this paper.
  • The benchmarks and initial conditions cover too narrow range of dynamics. All three are 1D, nearest-neighbor lattices, with KG/FPUT initialized using small amplitude, low frequency sinusoidal states and Toda using nearly the same alternating pattern.
  • The edge ablation does not isolate the benefit of communication. Adding an edge block changes both model capacity and how states are updated, while the reported nonzero edge response only confirms that the block changes its input. It does not establish that using a neighbor’s independent state causes the improvement, particularly when Toda’s initial patches repeat the same pattern.
  • Every predictive experiment uses P=r=8 so the encoder-decoder performs an invertible change of coordinates with no dimensional compression. So the experiments do not test reconstruction quality, reduced state formulation is not empirically tested.
  • Although the maths itself looks to be intact, its scope is very limited and narrow compared to what boasted by the manuscript. It verifies SHNO's symplectic structure and computational scaling, but its error bounds assume that the local components are already uniformly accurate. It does not estabilsh that training achieves this accuracy across system sizes or that preserving symplecticity improves long-term prediction.
Questions for authors
  • Why is this problem impactful and why should the ICLR audience motivated to care about this problem? I feel the introduction and abstract needs substantial improvement on selling the significance of this work.
  • Why the neural operator baseline is only FNO? I see many neural operator baselines that appear highly relevant and far more advanced than original FNO for this problem.
  • I don't feel the #params gain is of any impact in this problem, since the conventional baselines are already small enough in memory. Could the authors convincingly suggest where this storage gain can be impactful in practice? I mean, few MB of memories is nothing in 2026.
  • Could the authors provide training / inference time and compare the end-to-end time cost with all the baselines? For neural operator methods training time should be included when comparing with numerical methods like ROM.
  • Is there theoretical guarantee or justification that preserving symplecticity improves long-term prediction? I don't find the mathematical justification in the manuscript, but it writes as if preserving symplecticity is central for good long rollout.
Reviewer_HksiOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper tries to learn Hamiltonian dynamics with one set of parameters that can be reused across different system sizes, while keeping the symplectic structure. The motivation is that many Hamiltonian systems keep the same interaction laws when the number of interacting units changes, but existing structure-preserving ROMs tie the encoder and decoder to a fixed dimension, and parameter sharing in neural operators alone gives no symplectic guarantee. The authors propose the Shared Hénon Neural Operator (SHNO), which splits the state into fixed-width canonical patches of position and momentum pairs. All patches use the same encoder, the same HénonNet for latent evolution, and a tied decoder built from one invertible symplectic map and its inverse, and optional symplectic edge maps with shared parameters couple the patches.

Critical strengths

The main strength of this paper is that it presents the encoder-decoder structure in a form that allows systematic propagation, while finding a dynamical rule that preserves the symplectic structure, modeling it, and carryng it all the way to actual experimental results. In particular, it is interesting that the modeling in (10) directly uses the so-called Henon map, which can be applied both forward and in reverse, on the local state, since this operation is known to preserve symplectic behavior. Existing operator learning methods have rarely used latent information propagation that guarantees such a strong structure. In addition, I highly value the originality of combining the rule in (10) above with the operation called “G-reflection” in (11).

Critical weaknesses

The biggest weakness of this paper is that, despite the original idea, the paper is very poorly written. It defines many functions that are pointless to introduce, and the notation for each of these functions, instead of serving as tools for a detailed understanding of the proposed method and distracts the reader and makes it impossible to consistently follow in which direction the method is going at each point.

For example, πr\pi_r and ιr\iota_r appear at line 186. What role are these functions originally supposed to play, and how are they defined “explicitly” in mathematical terms? And where is the detailed description of how they fit with (13)? I can understand the exact intention by looking at (13), but the general audience of this paper would probably not understand, in the first place, the symplecticity itself as written in the second equation of (13), and yet it is introduced without a clear definition and explanation. This kind of problem repeats up to Section 3.5, and as the discussion goes on, it leaves only unclear questions. Overall, the structure gives the impression that the implemented code already existed and the paper was written from it in a one-to-one manner.

I would also like to discuss the importance of theory. This paper builds the encoder-decoder form by composing arbitrary operators that satisfy the symplectic property, but the discussion of why this is needed in the first place is unclear. Usually, for an encoder-decoder, there should be a discussion such as how the encoded information connects to a log probability like the ELBO, or how encoding with the proposed Hénon operator strengthens the generalization of the representations. Why a Hénon-type operator in particular? Though the authors cited (Peng & Mohseni) as their foundations, one can design many kinds of encoder-decoder pairs with an inverse relation that satisfy (13).

The authors express the invertibility of this dynamical operator by misusing the encoder-decoder form that we commonly use in ML, and they should change the discussion of the entire paper.

Questions for authors

Please see the above "Critical Weaknesses" section for details!

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

Related papers

Discussion (0)

Sign in to comment.