acceptodds
Under review as a conference paper at ICLR 2027

Fragmentation is Efficiently Learnable by Quantum Neural Networks

Abstract

In certain classes of physical quantum systems, the exponentially large state space "fragments" into many low-dimensional, dynamically disconnected subspaces. We introduce a learning problem known as _fragment classification_, where given a quantum state input, one is interested in classifying to which subspace the state belongs. We prove that solving this learning problem is efficient on a quantum computer when the fragmentation phenomenon satisfies certain conditions. Furthermore, we give evidence supporting the classical hardness of this task by demonstrating that known dequantization techniques fail for the fragment classification problem. Consequently, this work provides a rare example of a physically motivated quantum machine learning task that is both efficient for quantum computers to perform and admits no known classical dequantization.

Primary area:
optimization
Keywords:
quantum machine learning, quantum neural networks, Hilbert space fragmentation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eBe8OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces fragment classification, a quantum learning task motivated by Hilbert-space fragmentation. Given labeled Schur-basis states, the objective is to learn a circuit that identifies sector and basis labels while generalizing across multiplicity labels not represented in training. The authors construct a symmetry-preserving QNN from local Hamiltonian generators and analyze its trainability under assumptions including polynomial effective sector dimensions and full eigenstate thermalization. The paper claims inverse-polynomial gradient scaling and the absence of spurious local minima under polynomial overparameterization. It also argues that existing algebra-based classical simulation methods cannot be directly applied without additional algebraic information. Small numerical experiments on Temperley–Lieb systems illustrate training behavior. Classical hardness remains conjectural, and no empirical quantum advantage is demonstrated.

Critical strengths

S1: The paper exploits physical symmetries through quantum operations without requiring an explicit classical description of the underlying algebra. This identifies a potentially meaningful distinction between a quantum learning procedure and existing algebra-based simulation techniques.

S2: The QNN construction also provides a concrete mechanism for generalization across multiplicity labels. Because the circuit acts identically on equivalent multiplicity sectors, this generalization follows from its structure rather than from observing exponentially many training examples.

Critical weaknesses

W1: The trainability analysis does not clearly establish the polynomial-time learning guarantee required by Definition 1. The main argument combines a gradient second-moment bound with a claimed absence of spurious local minima. These properties alone do not provide a polynomial bound on the number of optimization steps or quantum measurements needed to achieve a specified classification error. In particular, the gradient bound concerns an approximate distribution over randomized architectures, rather than a quantitative convergence guarantee along training trajectories. A complete argument connecting the landscape results to the stated poly(L,1/ϵ) learning complexity is necessary because efficient learnability is the paper’s central claim.

W2: The evidence for the intended quantum–classical distinction and its practical relevance is also limited. Section 3.3 identifies obstacles to applying particular algebra-based simulation methods, but provides no formal classical-hardness reduction. The numerical experiments contain no classical baseline comparisons and use small Temperley–Lieb instances that the authors acknowledge admit efficient classical simulation. They therefore illustrate QNN trainability without testing the setting responsible for the proposed computational advantage. Moreover, the experimental application in Section 4 is only proposed: the paper does not establish that accessible training states and the relevant physical sectors satisfy its theoretical assumptions. These limitations leave the significance of the separation insufficiently substantiated beyond the conditional theoretical construction.

Questions for authors

None.

1 comment

We thank the reviewer for their comments and suggestions regarding our manuscript. We respond to the reviewer's comments on the weaknesses of the work.

  1. It is true that we do not give an explicit bound on the number of optimization steps necessary to achieve a certain classification error. Doing so for a QNN ansatz such as the one we consider is generally not a tractable task, and the conditions on barren plateaus and local minima that we study are widely accepted proxies for efficient trainability in the quantum machine learning literature (see, e.g., arXiv:2408.11901 (ICLR 2025); Nat. Commun. 15, 7172 (2024); Nat. Commun. 15, 7171 (2024); and Nat. Commun. 16, 7907 (2025)). While these conditions do not formally guarantee that the optimization of a model will necessarily converge within a given number of steps, they eliminate the two main mechanisms that are known to prevent a QNN model from being efficiently trainable using local optimizers.

  2. We agree with the reviewer that establishing a formal classical-hardness reduction would further strengthen our results, and we hope to pursue this direction in future work. Nevertheless, we believe that identifying a learning setting that circumvents all currently known classical simulation techniques for algebraically structured quantum systems is itself a significant contribution. In particular, our results identify a concrete setting in which provable trainability does not immediately come at the expense of classical simulability, in contrast to many existing results on trainable QNNs (see, e.g., arXiv:2408.11901 (ICLR 2025) and Nat. Commun. 16, 7907 (2025)). This presents an interesting dichotomy: either our construction provides a route toward quantum computational advantage, or its classical simulation will require fundamentally new dequantization techniques that go beyond existing algebraic approaches. We believe that both possibilities are of substantial interest to the quantum machine learning community, and that identifying this setting is a valuable step toward establishing rigorous quantum–classical separations for trainable QNNs.

Reviewer_vt3tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper shows that quantum neural networks (QNNs) can provable learn fragmentation in Hilbert space. They show that for this task, QNNs do not suffer from barren plateaus and hence can be optimized easily using first order gradient methods. Moreover, they give arguments on why this problem can be classically hard for a non-quantum learning algorithm.

Critical strengths

The paper is well written and addresses a physically relevant problem that can be solved using QNNs. Their attempt to quantify classical hardness is also a step in the right direction here.

Critical weaknesses

Please address the following concerns. There some crucial assumptions made in the paper, which need further clarification:

  1. Could the authors clarify how the data assumptions in Definition 1 relate to the experimental setting in Section 4? Proposition 1 establishes generalization across multiplicity labels (p), but does not appear to establish generalization across unobserved internal states (q). For the proposed variant that predicts only (\lambda), does the training data need to provide some additional coverage over each sector?

  2. I have concerns about how the ETH assumption is used here. In hypothesis 1, ETH is defined for local operators. In Lemma 5, the relevant operators seem to be rank 1 projectors. Could the authors clarify this.

  3. In the application of the Kac-Rice formula (in (96)), could the authors clarify why the first density factor in the formula p(∇l~=0)p(\nabla \tilde{l} = 0), does not grow as the authors add the epsilon perturbation to the loss function? Currently this is hand-waved in the last paragraph of the proof. I would like to see a rigorous argument here.

Questions for authors

See above.

1 comment

We thank the reviewer for their comments and suggestions regarding our manuscript. We reply to the reviewer's questions point by point.

  1. The reviewer's observation is correct that our construction generalizes across multiplicity labels pp but not internal states qq. In the context of the experimental setting where one is interested in only predicting λ\lambda, it is true that in order to train a QNN that classifies an arbitrary state in any sector, the dataset would need to include states with different qq. To obtain such a dataset, we could start from an initial state and time-evolve it under the system's Hamiltonian for randomly chosen times. If we are only interested in learning the sector λ\lambda and not qλq_\lambda, the loss function for this dataset is equivalent, up to some reweighting, to one where we start with an orthogonal set of basis states for each sector, as long as the number of states we pick is a constant factor larger than the dimension of the sector. We now sketch this argument in Section 4 of the updated and reuploaded version of the paper.

  2. The form of the full ETH in Fava et al. (2025) that we cite does not exactly correspond to the structure we are considering in Lemma 5. However, the methods they use still apply to expectations of the form we consider in our manuscript, where we are looking at correlation functions involving the local observables AA and OxO_x as well as the projector onto the initial state ∣x⟩|x\rangle. This is explicitly worked out in SciPost Phys. 18, 136 (2025) (see their Equation (18)). We now cite this reference when introducing our ETH assumption (Hypothesis 1) in the revised version of our paper.

  3. We have added a more detailed justification of this claim in the proof of Theorem 3 in the updated version of the manuscript. It follows by considering the asymptotic expression for the loss derivative distribution we compute in the proof of Theorem 1 (equation 48 in the current version), expressed as a linear combination of sums of products of i.i.d. Gaussian random variables. The probability density of each derivative (and thus the gradient) does not diverge at zero. Adding a Gaussian perturbation scaled by an arbitrarily small ϵ\epsilon clearly keeps this density bounded, which is what we require.

Reviewer_985WOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies a quantum learning problem where given a state that is a basis state promised to be in the Schur basis, one is interested in classifying to which subspace the state belongs.

Critical strengths

The problem studied is interesting and I can see its motivations. The paper is also solid with nice results if they are correct (see my comments below)

Critical weaknesses

The paper claims provable absence of barren plateaus but Theorem 1 does not support that. It bounds the second moment of an approximate that to be Ω(1/poly(L)\Omega(1/poly(L) but based the approximations have o(1)o(1) additive error per Lemma 3.

Some of the results are straightforward from existing work. The ansatz’s DLA restricted to the sectors the loss sees has dimension ≤∑λ(NλNa)2=poly(L)\le \sum_\lambda (N_\lambda N_a)^2 = poly(L) by assumption (4). Polynomial DLA dimension already gives $\Omega(1/poly) gradient variance (Ragone et al. 2024). So Prop. 2 is immediate from this.

Questions for authors

Q: How does an o(1)o(1) error justify a 1/poly(L) lower bound and hence BP absence?

Q: What is the DLA of H,A{H, A} restricted to the Λ′\Lambda' sectors, and why doesn’t Ragone et al. 2024 give Theorem 1 directly?

1 comment

We thank the reviewer for their comments and suggestions regarding our manuscript. We reply to the reviewer's questions point by point.

  1. The reviewer is correct that o(1)o(1) error does not necessarily justify a 1/poly(L)1/\textrm{poly}(L) lower bound. When we say that the loss derivatives can be approximated to an o(1)o(1) error, this should actually refer to a rescaled version of the derivatives normalized such that the variance is Θ(1)\Theta(1). Our current proofs already suffice for this, and we have modified the statements of Lemmas 2 and 3, as well as Theorem 1, in our manuscript to make this fact explicit and clear. In the course of these changes we have also simplified the proof structure and removed what was previously Lemma 4 as it was unnecessary for the current version of the proof of Theorem 1.

  2. The DLA of HH restricted to the Λ′\Lambda' sectors does have polynomial dimension. Our proof follows similar techniques to Ragone et al., and the fact that Theorem 1 holds is not very surprising. However, we cannot directly apply the results of Ragone et al. and similar prior work to immediately give these results since several technical technical details and assumptions are different in our setting. The main distinction is that, as we do not have prior knowledge of the algebraic structure of our circuits, we are unable to directly implement an ϵ\epsilon-approximate tt-design which respects the symmetries of our system. Instead, we show that assuming a common physical assumption—the eigenstate thermalization hypothesis (ETH) within each Krylov sector (our Hypothesis 1)—one can implement an ansatz which respects this algebraic structure while not requiring prior knowledge of it, and we show efficient trainability in this setting. The main contribution of this work is not just demonstrating that the loss landscape of a QNN is well-behaved under the stated conditions, but also connecting this to a learning task where a QNN model may be of practical use due to the low-dimensional algebraic structure not being easily accessible.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.