acceptodds
Under review as a conference paper at ICLR 2027

Effective Dimension Governs Generalization in Quantum Kernel Vision Models

Abstract

Recent quantum vision models—quantum vision transformers and quantum convolutional networks—report two striking but unexplained empirical phenomena: (i) ansatze with more, or more uniformly distributed, entanglement generalize better, and (ii) injecting quantum noise can improve test accuracy rather than degrade it. These observations are currently treated as curiosities, discovered by grid search and explained, if at all, by hand. We show that both are manifestations of a single, measurable quantity: the effective dimension of the (noise-shaped) quantum feature kernel. Working primarily with quantum-kernel vision models—a quantum feature map read out by a kernel classifier—we give a spectral account in which entanglement structure and quantum noise are two knobs that move ; in an overfitting regime, contracting acts as ridge-like regularization. We analyze the mechanism: an exact decomposition of the depolarized kernel with , a contraction result (and its boundary) for amplitude damping, a kernel-machine capacity bound, and a capacity/alignment risk decomposition; the monotone contraction operative in our entangled experiments is verified empirically, not proven in general. Our informative empirical finding is that test accuracy collapses onto a single function of across distinct entangling ansatze—compressing different spectral shapes onto one curve ( over seeds). Along the one-parameter depolarizing family the collapse is instead exact by construction; we use it only to confirm the kernel decomposition to machine precision and at up to qubits, not as evidence for . Amplitude damping contracts and lifts test accuracy by up to along an inverted-U sweet spot; the effect's sign flips between the over- and under-fitting regimes; noise injection matches an explicit spectral-filtering frontier (so it is not a weak substitute for hyperparameter tuning); and the phenomenon persists in trained QViT- and QCNN-like models. Entanglement plays the complementary role of a precondition—it supplies the feature-space alignment without which the law does not hold. Our results organize two reported anecdotes into a single measurable principle for designing quantum-vision models.

Primary area:
learning theory
Keywords:
quantum machine learning, quantum kernel methods, generalization theory, effective dimension, spectral regularization

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

This paper establishes effective dimension as a unifying spectral principle governing generalization in quantum kernel vision models, showing that entanglement improves feature–label alignment while quantum noise can regularize overfitting by contracting the kernel spectrum, thereby explaining when and why both mechanisms enhance performance.

Critical strengths

The paper offers a coherent spectral perspective on two empirical observations in quantum vision models. In particular, it relates entanglement and noise to the effective dimension of the quantum feature kernel, providing a potentially useful conceptual framework for interpreting their influence on generalization. This is a meaningful strength because it attempts to move beyond purely empirical observations toward a more unified explanation.

The empirical evaluation is relatively broad and generally consistent with the proposed interpretation. The authors examine multiple entangling ansätze, noise levels, datasets, and additionally report experiments on IBM Heron hardware. Although some of the claims remain empirical rather than fully established theoretically, the consistency across these settings lends support to the practical relevance of the proposed effective-dimension viewpoint.

Critical weaknesses

The presentation raises some concerns about possible heavy automated assistance in the writing. Many paragraphs follow unusually similar lengths and rhetorical patterns, with frequent use of em dashes and highly repetitive transition structures. While these stylistic features are not sufficient to establish AI authorship, they make the manuscript read somewhat formulaically and reduce confidence that the exposition has been carefully shaped around the scientific argument itself. I would therefore encourage the authors to substantially revise the presentation and, where required by the venue, clarify the role of any generative-AI assistance.

More importantly, the central claim appears stronger than what the theoretical and empirical evidence currently establishes. The effective dimension is shown to be predictive only under additional conditions concerning spectral shape and label alignment, while several of the most important observations across different ansätze and under amplitude damping remain empirical rather than theoretically guaranteed. Moreover, part of the cleanest one-parameter collapse follows directly from the construction of the depolarizing-noise family. This gap between the broad framing of “effective dimension governs generalization” and the more conditional scope actually demonstrated is, in my view, the more consequential issue for acceptance, because it affects the paper’s main claimed contribution rather than only its presentation.

Questions for authors

Could the authors provide a more direct test that effective dimension, rather than correlated changes in the kernel spectrum, is the main factor driving the reported cross-ansatz generalization behavior?The central cross-ansatz result relies on different entangled circuits having sufficiently comparable spectral structure and label alignment, yet effective dimension is only a one-number summary of the spectrum. It would materially affect my assessment to know whether configurations with similar effective dimension but noticeably different higher-order spectral characteristics still achieve comparable accuracy, or whether the observed collapse weakens once these factors are controlled.

Could the authors clarify how robust the monotonic decrease of effective dimension under amplitude damping is across individual seeds, datasets, and ansatz choices? The theoretical analysis does not establish monotonicity for the general entangled amplitude-damping setting, while this behavior is important to the paper’s main empirical mechanism. It would therefore be helpful to report whether any non-monotonic cases were observed across the full experimental sweep, rather than only presenting averaged or selected trends.

Could the authors more carefully separate the effect of spectral contraction from ordinary kernel-scale or classifier-regularization effects? As noise changes the kernel, several quantities relevant to an SVM or kernel ridge classifier may change simultaneously. Since the paper interprets the improvement primarily through effective dimension, I would like to know whether the main conclusions remain unchanged after appropriate kernel normalization and after independently retuning the classical regularization parameter for each noise level. This control would be important for determining whether the reported benefit is genuinely specific to the proposed spectral mechanism.

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

This paper focuses on two phenomena, "more entanglement helps generalization" and "quantum noise can help". The authors insist that these phenomena are not separate by reducing the phenomena to the quantity called effective dimension. Especially, they theoretically andempirically investigate the three statements 1. The effective dimension organizes generalization, 2. Noise is a spectral regularizer, and 3. Entanglement is a precondition.

Critical strengths

This paper tackles an important problem how the entanglement and noise affect to the quantum kernel framework. The idea of using the effective dimension as a unified criterion of entanglement and noise is interesting. The topic is relevant to the community.

Critical weaknesses
  1. Theoretical results seems straightforward consequences from known results about kernel methods, and many of critical statements rely on empircal observation. Some empirical observations are included in theoretical statements (propositions and lemmas). I think the authors should separate theoretical contributions and empirical contributions, and clarify the novelty of this paper from both theoretical and empirical sides.

  2. In Lemma 2, the authors assume the constant-row-sum condition. Also, even with the constant-row-sum condition, Lemma 2 does not guarantee that dγd_{\gamma} and deffd_{eff} induce the same ordering between two different kernels. Thus, it is unclear for me why we can replace dγd_{\gamma} by deffd_{eff} and state "(P1) deffd_{eff} organizes generalization".

  3. The claim (P2) is only supported by empirical observations. Do you have any theoretical findings regarding (P2)?

  4. Prop. 4 does not seem to show the relationship between the entanglement and the alignment. Thus, it is unclear for me why Prop. 4 supports the claim (P3). Could you clarify this point?

Questions for authors

Plese see the weakness section.

Minor comments:

  • Please define AγA_{\gamma} in the main text.
Reviewer_74RWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a spectral explanation for two empirical observations reported in quantum vision models: that entanglement can improve generalization and that moderate quantum noise can sometimes improve, rather than degrade, test accuracy. The authors argue that these two phenomena are not separate and not mysterious. Entanglement structure and quantum noise are two knobs that move the effective dimension. Once alignment and spectral shape are comparable—as they are across entangled ansatze generalization is largely organized by the effective dimension; under overfitting, contracting the effective dimension acts as regularization and improves generalization up to an interior optimum.

Critical strengths
  1. The paper gives a coherent mechanistic interpretation in connecting entanglement, noise, kernel spectra, and generalization.
  2. The authors explicitly distinguish the exact depolarizing-kernel result from the amplitude-damping experiments, acknowledge the limitation.
Critical weaknesses
  1. The statement that effective dimension “governs” generalization appears too strong without sufficient evidence.
  2. The evidence on realistic quantum vision architectures and hardware remains preliminary.
  3. The related work on the impact of noise on quantum kernels or noisy quantum kernels could be more comprehensive.
  4. The presentation style is too “AI”.
Questions for authors
  1. How strongly do the authors intend the word “governs” in the title?
  2. Can the authors more directly demonstrate that entanglement itself moves the effective dimension in a systematic way?

Comments on the paper

1 comment

Thank you for the rebuttal. Some of my comments have been properly addressed. However, I still think that the theoretical contribution of this paper is limited. Thus, I decided to keep my score unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.