acceptodds
Under review as a conference paper at ICLR 2027

An Entropic Orbit-Stabiliser Theorem for Symmetry-Aware Quantum Learning

Abstract

The classical orbit-stabiliser theorem describes exact symmetry through membership in the stabiliser subgroup. We define an orbit stabilisation entropy and an average stabilisation weight , and establish the new bound , giving an entropic version of the orbit-stabiliser theorem. We further show that equals the purity of the group-twirled state. In a controlled quantum learning setting, we study whether state-level symmetry is sufficient to produce symmetry of the learned function. We first encourage state symmetry through soft regularisation using and , but do not observe a statistically robust improvement in generalisation to unseen symmetry-transformed inputs. We then derive sufficient conditions for exact function-level symmetry and implement them through hard-equivariant architectures, specifically through parameter tying. We test this across five finite groups , , , , and . The additional orbit-transfer gain from symmetry-enforcing parameter tying is significant for and , but not for the other three groups. Three controlled experiments test possible explanations for this group-dependent effect. Finally, we apply our entropic bound to prune candidate stabiliser subgroups. In subgroup-rich group families, this can substantially reduce the number of candidates that require direct verification, giving a measured end-to-end speedup at group order . We also extend our framework to mixed states and study its behaviour under controlled noise and symmetry breaking. Our work opens the door to further study of entropic symmetry measures and their applications in symmetry-aware quantum learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Quantum Machine Learning, Symmetry-Aware Learning, Equivariant Learning, Quantum Neural Networks, Group Equivariance, Quantum Symmetry, Stabilizer Theory, Group Theory, Variational Quantum Circuits, Quantum Information

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_a1PyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduced orbit stabilisation entropy and average stabilisation weight for symmetry-aware quantum learning. An entropic orbit–stabiliser bound is established, together with a connection between average stabilisation and the purity of the group-twirled state. Sufficient conditions for exact function-level symmetry are derived. Experimental results show group-dependent generalization gains from parameter tying and computational benefits from entropy-based stabiliser pruning.

Critical strengths
  1. The proposed entropy gives a continuous description of approximate stabilisation. The entropic bound and the connection between average stabilisation and twirled-state purity are clearly presented.
  2. The experiments use five finite groups, matched-untied controls, and repeated seeds. The reporting of nonsignificant soft-regularization results and group-dependent parameter-tying gains is appropriately cautious.
Critical weaknesses
  1. Soft regularization is tested at a fixed strength. Consequently, the negative empirical result does not establish whether the approach remains ineffective after suitable tuning or scheduling.
  2. The learning experiments use a small controlled classification problem, and significant parameter-tying gains occur for only two groups. The evidence for broader predictive benefit is therefore limited.
Questions for authors
  1. For stabiliser pruning, how much additional benefit does the entropy bound provide over the direct bound using total stabilisation weight, after accounting for estimation cost?
  2. How do the soft-regularization results change across regularization strengths and training schedules? Please report both orbit-transfer and IID accuracy.
  3. Can the findings be verified on a more substantive quantum learning task and larger representations, and what properties predict when parameter tying improves accuracy?
Reviewer_DhRfOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper define orbit stabilization entropy and average stabilization weight as the new conceptual contributions. They then give an equality that is an entropic version of the classical orbit-stabilizer theorem. Substantial experiments are then performed to see whether and how these new concepts aid learning under simple symmetries.

Critical strengths

The paper’s framework is simple but comprehensive. The orbit stabilization entropy is potentially useful as a soft regularizer for unstructured models, particularly when hard constraints impose an approximate or misspecified symmetry too rigidly. The use of these particular soft regularizers in this setting may be new, although this needs clearer positioning relative to prior work.

The experiments conducted are well-organized, covering many of the usual comparisons in learning under symmetries. Table 5 is particularly encouraging as a computational application, since it gives a method that can remove a large proportion of candidate subgroups as impossible.

Critical weaknesses

The paper does not position itself well in the literature. Its references to the literature are condensed into a few paragraphs. As a consequence, several ideas that the paper appears to present as its own have already appeared in the literature. For instance, closely related definitions of stabilization entropy have been studied, and the Pauli special case of Theorem 1 follows from a stronger identity in Eq. (5) of Leone et al., Learning t-doped stabilizer states (https://arxiv.org/pdf/2305.15398v1), which treats the more general family of α\alpha-Rényi entropies. The machine learning part, particularly §5.1 and §5.2, also seems insufficiently discussed in relation to prior work.

The paper’s experiments, while covering several settings, are rather weak and do not give many clear takeaways. The lack of statistically significant gains from soft symmetry enforcement is difficult to interpret in this kind of small synthetic dataset, which obeys symmetries exactly. The setting where soft constraints might be particularly useful, namely approximate or misspecified symmetry, is not tested. Hard symmetry also does not give consistently significant improvements in accuracy, although it enforces orbit consistency in every tested group. Table 4 is an interesting comparison but does not demonstrate that the distinction improves learning. Table 5 is interesting because we can rule out a large number of candidate subgroups. However, the entropic variant is unnecessary here. Indeed, the proof of Theorem 1 already gives $ |Stab_G([\psi])|\le W(\psi)\le e^{H_{\rm stab}(\psi)} $. The first inequality is a simple counting bound based on stabilization strengths. Since WW is already computed as part of computing HstabH_{\rm stab}, the authors should compare against this stronger upper bound on stabilizer size.

Questions for authors
  1. (See Weakness point 2 for the context) Is there a conceptual reason to use the bound e^H_stab instead of the stronger bound W?
  2. The soft constraints are the most interesting part of the paper. Can your framework be applied to detect approximate symmetries in experimentally prepared quantum states? (e.g. Elben et al Cross-Platform Verification of Intermediate Scale Quantum Devices)
Reviewer_TFwsOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces an entropy based on group-action overlaps and proves a lower bound involving the size of the exact stabiliser. It connects average stabilisation to the purity of the group-twirled state, extends the framework to mixed states, and studies the distinction between state symmetry and function symmetry in quantum learning. Experiments compare soft regularisation with hard-equivariant circuits and explore candidate-subgroup pruning.

Critical strengths
  1. The connection between exact stabilisers and a continuous entropy measure is interesting and clearly developed. The complementary twirling identity and mixed-state extension give the framework a coherent mathematical foundation.
  2. The experiments carefully distinguish exact function symmetry from predictive improvement. The matched-untied controls and 25-seed comparisons are valuable: parameter tying helps for Z2 and S3, while the authors explicitly report the absence of robust gains from soft regularisation (Tables 1-3).
Critical weaknesses
  1. The learning experiments use a small six-qubit synthetic task and a fixed regularisation strength. The findings are informative, but their generality beyond this controlled setting remains uncertain.
  2. The pruning application needs a stronger comparison. The proof already establishes |Stab| <= W <= exp(H_stab), where W is the unnormalised overlap sum. Thus, W gives an at least equally strong size bound; the reported speedup over exhaustive candidate verification does not establish an advantage specific to entropy.
Questions for authors
  1. Does entropy-based pruning offer any practical benefit over directly using the bound |Stab| <= W?
  2. How sensitive are the soft-regularisation results to the regularisation strength and training duration?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.