Pith. sign in

REVIEW 4 major objections 5 minor 41 references

For QSVMs trained under local depolarising noise, the squared geometric margin — measured against the ideal kernel — is bounded above and below by explicit functions of the per-qubit noise survival probability (1−p)^{2N}, the clean margin,

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

For QSVMs under local depolarising noise, the paper claims analytic upper and lower bounds on the squared geometric margin, validated (it says) by simulation and on an IBM device, plus empirical evidence that margins predict QSVM generalisation.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Kernel bound is real, but the margin bounds rest on an invalid quadratic-form step and a p=0 contradiction; salvageable but not acceptable as is. the 4 major comments →

arxiv 2601.23084 v3 pith:ZDTII2MJ submitted 2026-01-30 quant-ph

The Cross-Kernel Margin: A Robustness Measure for Quantum Kernel Methods

Saarisha Govender , Ilya Sinayskiy This is my paper
classification quant-ph
keywords quantum kernel methodsquantum support vector machinescross-kernel margingeometric marginlocal depolarising noisegeneralisation boundsNISQkernel perturbation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces the cross-kernel margin for quantum kernel-assisted support vector machines (QSVMs): the geometric margin of a classifier trained on a noisy kernel but evaluated against the ideal kernel. It claims that under local per-qubit depolarising noise with error probability p, each noisy kernel element obeys K̃_ij ≤ (1−p)^{2LN} K_ij + 1 − (1−p)^{2LN}, and that this entrywise bound forces the squared cross-kernel margin into an explicit interval whose endpoints depend on the clean margin, the regularisation parameter C, and the noise survival probability. If correct, this gives a concrete characterisation of noise-induced margin decay for QSVMs and supports margin-based robustness analysis in the NISQ era. The paper also reports that the median geometric margin tracks test accuracy almost perfectly (Pearson r > 0.9) across four datasets, and that global depolarising noise is more optimistic than local noise at intermediate noise levels.

Core claim

On the paper's own terms: local depolarising noise applied independently to each qubit degrades quantum kernel values by a contraction-plus-additive-constant form, K̃_ij ≤ (1−p)^{2LN}K_ij + (1−(1−p)^{2LN}). Substituting this bound into the SVM dual objective and using optimality relations for the soft-margin problem, the squared geometric margin of the noisy-trained, ideal-evaluated classifier satisfies γ̃² ≤ (1−p)^{2N}γ² / [2(1−mCγ²) − (1−p)^{2N}] and γ̃² ≥ γ²[1−2(1−p)^{2N}] / [2Cmγ² − 1], subject to feasibility constraints on mCγ². These are the upper and lower cross-kernel margin bounds; at p=0 the upper bound is tight. The construction also yields a hard-margin analogue and extends to L

What carries the argument

The central object is the cross-kernel margin: the squared geometric margin 1/∥w̃∥², where w̃ uses the dual solution obtained from the noisy kernel but the norm is evaluated against the ideal kernel. The carrying mechanism is the entrywise kernel bound derived from the Kraus decomposition of the local depolarising channel via Schatten/Hölder trace inequalities, together with the KKT identity Σ ᾱ_i y_i = 0, which is used to drop the additive constant when passing from kernel elements to the quadratic form ∥w̃∥² = Σ ᾱ_i ᾱ_j y_i y_j K̃_ij. SVM duality then converts the resulting inequalities into the margin bounds.

Load-bearing premise

The load-bearing premise is that the entrywise kernel bound can be multiplied by signed SVM weights and summed to bound the squared weight norm; the exact expression (89) shows a state-dependent signed remainder, and the lower bound's validity conditions contradict the p=0 equality unless they are stronger than stated.

What would settle it

Take any trained QSVM with 0 < Cmγ² < 1/2 and set p=0; the claimed lower bound (56) requires γ² ≥ γ²/(1−2Cmγ²), which is strictly stronger than the exact equality γ̃² = γ². A single such instance settles whether the validity conditions are sufficient. Separately, compute the weighted sum Σ_{i,j} ᾱ_i ᾱ_j y_i y_j [K̃_ij − (1−p)^{2N}K_ij − (1−(1−p)^{2N})] for a two-qubit dataset; any positive value violates the substitution step shared by both margin bounds.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Margin decay under local depolarising noise is controlled by the survival probability (1−p)^{2N}; increasing qubit number or encoding layers compounds the noise effect.
  • The bounds give a computable range of regularisation parameters for which the noisy margin is guaranteed non-negative, allowing practitioners to certify training stability before running a noisy device.
  • The noisy-trained, ideal-evaluated margin, not the self-consistent noisy margin, is the informative robustness statistic; the paper argues the latter hides noise-induced decay.
  • In the noiseless setting, median geometric margin can serve as a proxy for QSVM generalisation, with Pearson correlations above 0.9 across the tested datasets.
  • Global depolarising noise models understate performance loss; local per-qubit noise should be preferred when modelling QSVM robustness in the NISQ era.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A tighter, state-dependent bound could replace the additive constant: the paper's exact expression (89) exhibits a signed remainder depending on the purity and Pauli coefficients of the data states, so the loose entrywise bound may be far from tight for low-purity states.
  • The label-corruption experiments establish margin–accuracy correlation in the ideal setting; applying the same corruption pipeline under kernel noise would test whether the cross-kernel margin predicts accuracy under device noise as directly as it does under label noise.
  • The bounds suggest an experimentally testable noise threshold: solving for the p at which the lower bound crosses zero should predict where QSVM test accuracy collapses, and that prediction can be checked on hardware.
  • The derivation structure should carry to other single-qubit noise channels, such as amplitude damping, by replacing (1−p)^{2N} with the channel-dependent survival factor; disagreement would show where the depolarising assumption is load-bearing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a margin-based robustness measure for quantum kernel-assisted support vector machines (QSVMs) under local depolarising noise. It derives an entrywise upper bound on noisy kernel elements, then uses SVM KKT and optimality relations to obtain upper and lower bounds on the squared geometric margin of a QSVM trained on the noisy kernel and evaluated against the ideal kernel. The main theoretical results are eqs. (42) and (56). The paper also reports empirical margin-accuracy correlations across several datasets, a comparison of local vs. global depolarising noise, and an experiment on the IBM ibm_fez device.

Significance. If the main theorem were correct, it would be a valuable analytic characterisation of noise-induced margin decay for QSVMs under a realistic, per-qubit noise model, and it would strengthen the case for margin-based generalisation analysis in quantum machine learning. The entrywise kernel bound (33) is a genuine and checkable derivation, and the empirical correlation between median geometric margin and test accuracy (r > 0.9) is interesting. However, the central step that converts the entrywise kernel bound into a quadratic-form bound is invalid, and the lower margin bound is contradicted already at p=0. The theoretical contribution is therefore not established; the useful empirical observations do not rescue the main claim.

major comments (4)
  1. [Appendix A.3.1/A.3.2, eqs. (91)-(95)] The substitution of the entrywise bound K̃_ij ≤ A K_ij + B into the quadratic form is invalid because the coefficients s_i s_j = α_i α_j y_i y_j are not nonnegative. The KKT condition Σα_i y_i=0 removes only the BΣs_i s_j term; it does not control the A term. Concrete counterexample: N=2, L=1, p=0.1, training states |00⟩ and |11⟩ with labels ±1. For the local depolarising channel λ=1−4p/3=0.8667, K̃_11=K̃_22=((1+λ²)/2)²≈0.766, K̃_12=((1−λ²)/2)²≈0.0155, and A=(0.9)^4≈0.656. For the two-point SVM, the optimal noisy dual solution (for C not saturated) is v=α(1,−1) with α=2/(K̃_11+K̃_22−2K̃_12). Then v^T K̃ v ≈ 2.665 while A v^T K v ≈ 2.330, so v^T K̃ v > A v^T K v. This contradicts (93). Since (93) is used to derive (41) and (55), the margin bounds (42) and (56) are unsupported.
  2. [Section 6.3, eq. (56) and conditions (57)–(61)] At p=0 the noisy kernel coincides with the clean kernel, so the noisy margin equals the clean margin: γ̃²_α = γ²_α. Under condition (61), which is allowed when Cmγ²_α < 1/2, the right-hand side of (56) becomes γ²_α / (1 − 2Cmγ²_α) > γ²_α. Thus (56) asserts γ²_α ≥ γ²_α / (1 − 2Cmγ²_α) > γ²_α, a direct contradiction. This failure is independent of the Result 1A/1B issue and shows that the lower margin bound is false in an allowed parameter regime, not merely unproven.
  3. [Appendix A.2, eq. (89)] Equation (89) is called the exact expression for the noisy kernel, but the remainder sum is restricted to m=n≠0. The m≠n terms are not shown to vanish; nothing in the derivation up to (87) eliminates them, and the original sum in (76) contains no such restriction. The label 'exact' is therefore unjustified, and (89) cannot serve as a repair for the signed-remainder problem that invalidates the quadratic-form step in Results 1A/1B.
  4. [Section 7.1 and eqs. (47)–(49)] The feasible C′ range used in the numerical validation is not fully a priori: C′_min in (49) contains the noisy margin γ̃², the quantity being bounded. The paper estimates it on a held-out validation set and then selects datasets that have an overlapping feasible C′ range across noise levels. Although the authors acknowledge this, it means the agreement in Figs. 11–13 is a weaker test of the theory than a parameter-free bound would provide, and the dataset-selection step conditions the empirical evidence.
minor comments (5)
  1. [Section 3.3, eqs. (12)–(15)] The notation for the noise probability changes from p′_L to p_L to p without an explicit statement of the relation. In particular, p in (15) should be p_L, and the domain p_L∈[0,3/4] should be stated where the Kraus operators are introduced.
  2. [Eqs. (46)–(49) and Eq. (63)] The symbol C′ is first defined as mC before eq. (46) and later as C_0 m^{1−β} in eq. (63). These are different objects; please use distinct notation or explicitly state the reparametrization.
  3. [Appendix A.2, eq. (83)] The definition of c_{n_i} in Appendix A.2 conflicts with the c_{αβ} product constants used in Section 6.1. This makes the derivation hard to follow; distinct symbols would help.
  4. [Section 7.2, Fig. 13] The statement that the hardware noise level is 'within the particular noise range' is not quantified. The crossing points of the green horizontal line with the bounds are used to infer a noise level, but no uncertainty or error bar for the hardware margin is given.
  5. [General] The arXiv title mentions a 'cross-kernel margin', but the term is not defined in the manuscript. Either define this quantity explicitly or adjust the title to match the content.

Circularity Check

1 steps flagged

The margin-bound derivation does not assume its target, but the numerical validation is partly circular: C'_min is computed from the noisy margin it is supposed to bound.

specific steps
  1. fitted input called prediction [Section 6.2, eqs. (47)-(49); Section 7.1, Dataset Selection]
    "Notably, the expression for the minimum C ′ value contains the noisy margin, i.e. the quantity that we want to bound. ... In our numerical experiments, a closer estimate for the value of C ′ min is obtained by computing an estimate noisy margin on a separate, held-out validation set. This value is used in the C ′ min calculation and consistently provides a tighter estimate of the minimum value."

    The upper-bound test chooses C' inside [C'_min, C'_max), but C'_min is algebraically obtained by rearranging the target upper bound and explicitly contains the noisy margin γ̃² (eqs. 47-49). Using a validation-set estimate of γ̃² to compute C'_min means the chosen C' is, by construction, one for which the bound is satisfied for that estimate. The plotted 'theoretical upper bound' is therefore calibrated to the very quantity it is supposed to constrain, so the empirical confirmation is partly forced. The theorem derivation itself does not use this estimate; hence this is validation-side circularity rather than a definitional circularity of the bound.

full rationale

Walking the claimed derivation chain: the kernel-element bound (33) is derived from depolarising-channel trace inequalities and does not assume the margin result. Results 1A/1B attempt to pass from this entrywise kernel bound to a quadratic-form norm bound; whether or not that step is mathematically valid, it is not circular because it does not posit the target inequality. The final margin bounds then follow from SVM optimality and KKT relations rather than from the empirical quantities they bound. There is no load-bearing self-citation: the only author self-citation is introductory and not used to justify the central derivation. The one genuine circular element is in the validation protocol: C'_min in eqs. (47)-(49) contains the noisy margin, and Section 7.1 explicitly uses a validation-set estimate of that noisy margin to compute C'_min, so the C' used to draw the 'theoretical' bound is selected from a range defined by the quantity the bound is meant to constrain. The false quadratic-form substitution in Results 1A/1B and the p=0 contradiction in lower-bound condition (61) are serious correctness problems, but they are proof failures rather than cases where the output is equivalent to the input by construction, so they do not further raise the circularity score.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 1 invented entities

The free parameters are the CV-chosen C₀, the tunable β, and the post-hoc feasible C′ window; these set the regime in which the bounds are tested. The main unproved load-bearing step is the entrywise-to-quadratic-form substitution in Results 1A/1B, listed as an ad-hoc-to-paper axiom. No new physical entities are postulated; the 'noisy margin' is a defined observable rather than an invented degree of freedom.

free parameters (3)
  • β (regularisation scaling exponent) = β=0 (Heart, Gaussian); β=1 (Breast Cancer, Wine)
    Tunable scaling in C = C₀/m^β controlling the C′ that enters the bounds; chosen per dataset 'to allow for flexibility... to align with the constraints' (eq 63). Reported β/C₀ values appear inconsistent with Table 1 feasible ranges.
  • C₀ (initial SVM regularisation parameter) = 10 (Heart), 100 (Gaussian), 1000 (Breast Cancer, Wine)
    Chosen by 5-fold CV; determines the clean margin γ²α that enters the bound RHS and the feasibility windows.
  • feasible C′ window = [0,21.1), [32.4,112.6), [0,544.7), [0,38.9), [0.6,125)
    C′ is selected from a data-derived feasible range; datasets without an overlapping range are rejected, so the tested parameter values are chosen post hoc.
axioms (6)
  • standard math SVM dual and primal optimality: L(α,K̃) ≤ L(ᾱ,K̃) and P(wα,·) ≤ P(wᾱ,·); KKT conditions including Σαᵢyᵢ=0 and αᵢ=C for ξᵢ≠0
    Standard SVM theory (Mohri et al. [5]); used throughout §6 to relate clean and noisy dual solutions.
  • standard math Trace/Hölder/Schatten-norm inequalities and purity bound Tr(ρ²) ≤ 1 used to bound each noisy trace term by 1
    Used in §6.1 to derive the entrywise kernel bound (32).
  • domain assumption Local depolarising channel model (Kraus form, eq 15) applied independently to each qubit after each unitary layer
    The paper's chosen noise model, motivated empirically versus global noise in §5; all bounds inherit this model.
  • domain assumption Survival-probability matching p_GS=(1−p_G)^L versus p_LS=(1−p_L)^{NL} as the fair basis for comparing global and local noise
    Modelling choice in §5 equating noise strengths; the 'global noise is overly optimistic' conclusion depends on it.
  • domain assumption IQP feature map plus Hilbert-Schmidt kernel K(xᵢ,xⱼ)=Tr(ρ(xᵢ)ρ(xⱼ)) after PCA to 2–3 features
    Specific embedding used in all experiments (§4, §7); bounds are state-dependent, so results may not transfer to other encodings.
  • ad hoc to paper Unproven matrix inequality: the entrywise kernel bound K̃ ≤ AK + BJ implies the quadratic-form bound sᵀK̃s ≤ A sᵀKs for s = ᾱ∘y
    Results 1A/1B substitute an entrywise bound into a sign-indefinite quadratic form; no justification is given, and the exact expression (89) shows a state-dependent remainder. Both margin bounds rely on this step.
invented entities (1)
  • Noisy geometric margin γ̃² (margin of the dual solution trained on the noisy kernel, evaluated against the ideal kernel) independent evidence
    purpose: Quantifies how much useful information survives in the kernel during noisy training; the target of the derived bounds.
    Computable from simulations and device experiments, so falsifiable in principle; but it requires access to the ideal kernel, which real deployments do not have, limiting its practical role.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of The Cross-Kernel Margin: A Robustness Measure for Quantum Kernel Methods." pith.science (2026). https://pith.science/paper/ZDTII2MJ

@misc{pith2026260123084,
  author       = {Pith},
  title        = {Pith review of: The Cross-Kernel Margin: A Robustness Measure for Quantum Kernel Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDTII2MJ}},
  note         = {Machine review of arXiv:2601.23084}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Quantum devices in the current Noisy Intermediate-Scale Quantum (NISQ) era are inherently affected by noise, which can degrade the predictive performance of quantum machine learning models. In this work, we introduce a new margin-based robustness measure for Quantum Kernel-Assisted Support Vector Machines (QSVMs), termed the cross-kernel margin. This measure quantifies the stability of a classifier learned under a perturbed kernel relative to the ideal feature space. We derive a posteriori stability bounds for the corresponding cross-kernel inverse squared-margin under kernel perturbations using the Tikhonov-stabilised SVM dual formulation. The local depolarising noise model is then applied to this framework to induce perturbations in the kernel. The resulting bounds are numerically checked using simulations across multiple datasets and further tested using kernel matrices obtained from real quantum hardware and a noisy backend simulator. Furthermore, we empirically compare the degradation of test accuracy under local depolarising noise with the commonly used global depolarising noise model in order to motivate its use in our study. Finally, we present empirical results linking margin-based quantities with the generalisation performance of QSVMs, providing additional motivation for our margin-based robustness analysis.

Figures

Figures reproduced from arXiv: 2601.23084 by Ilya Sinayskiy, Saarisha Govender.

Figure 1
Figure 1. Figure 1: Diagram depicting the optimal linear hyper [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Diagram depicting the optimal linear hyper [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantum circuit layer implementing IQP encoding with nearest-neighbour entanglement using Hadamard [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Box-plots depicting the geometric margin distribution calculated using datasets with increasing fractions of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Plots overlaying the decreasing test accuracy and median geometric margin graphs for increasing fractions [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Plots depicting the linear correlation between test accuracy and the median geometric margin. The Pearson [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Quantum circuit used to implement local depolarising noise channels for [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Quantum circuit used to implement global depolarising noise channels for [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Equivalent quantum circuit used to implement global depolarising noise channels for [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Plots depicting the difference in the test accuracy obtained when using the global and local depolarising [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Plots showing the theoretical upper bounds (red) computed for increasing values of local depolarising [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Plots showing the theoretical lower bounds (black) computed for increasing values of local depolarising [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Plots comparing the geometric margin obtained using the ibm_fez device (depicted by the green dashed [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 2 linked inside Pith

  1. [1]

    An introduction to quantum machine learning.Contemporary Physics, 56(2):172–185, 2015

    Maria Schuld, Ilya Sinayskiy, and Francesco Petruccione. An introduction to quantum machine learning.Contemporary Physics, 56(2):172–185, 2015

    2015

  2. [2]

    Pac-bayes & margins.Advances in neural information pro- cessing systems, 15, 2002

    John Langford and John Shawe-Taylor. Pac-bayes & margins.Advances in neural information pro- cessing systems, 15, 2002

    2002

  3. [3]

    A pac-bayesian margin bound for linear classifiers: Why svms work.Advances in neural information processing systems, 13, 2000

    Ralf Herbrich and Thore Graepel. A pac-bayesian margin bound for linear classifiers: Why svms work.Advances in neural information processing systems, 13, 2000

    2000

  4. [4]

    Simplified pac-bayesian margin bounds

    David McAllester. Simplified pac-bayesian margin bounds. InLearning Theory and Kernel Machines: 16th Annual Conference on Learning Theory and 7th Kernel Workshop, COLT/Kernel 2003, Wash- ington, DC, USA, August 24-27, 2003. Proceed- ings, pages 203–215. Springer, 2003

    2003

  5. [5]

    Afshin Rostamizadeh Mehryar Mohri and Ameet Talwalkar.Foundations of Machine Learning, 2nd ed.MIT Press, 2018

    2018

  6. [6]

    Near-tight margin-based generalization bounds for support vector machines

    Allan Grønlund, Lior Kamma, and Kasper Green Larsen. Near-tight margin-based generalization bounds for support vector machines. InInter- national Conference on Machine Learning, pages 3779–3788. PMLR, 2020

    2020

  7. [7]

    Generaliza- tion performance of support vector machines and other pattern classifiers.Advances in Kernel meth- ods—support vector learning, pages 43–54, 1999

    Peter Bartlett and John Shawe-Taylor. Generaliza- tion performance of support vector machines and other pattern classifiers.Advances in Kernel meth- ods—support vector learning, pages 43–54, 1999

    1999

  8. [8]

    On the uniform convergence of relative frequencies of events to their probabilities

    Vladimir Vapnik. On the uniform convergence of relative frequencies of events to their probabilities. InDoklady Akademii Nauk USSR, volume 181, pages 781–787, 1968

    1968

  9. [9]

    Bounds on error expectation for support vector machines

    Vladimir Vapnik and Olivier Chapelle. Bounds on error expectation for support vector machines. Neural computation, 12(9):2013–2036, 2000

    2013

  10. [10]

    Springer Science & Business Media, 2006

    Vladimir Vapnik.Estimation of dependences based on empirical data. Springer Science & Business Media, 2006

    2006

  11. [11]

    Support- vector networks.Machine learning, 20(3):273– 297, 1995

    Corinna Cortes and Vladimir Vapnik. Support- vector networks.Machine learning, 20(3):273– 297, 1995

    1995

  12. [12]

    Quantum computing in the nisq era and beyond.Quantum, 2:79, 2018

    John Preskill. Quantum computing in the nisq era and beyond.Quantum, 2:79, 2018

    2018

  13. [13]

    Generalization in quantum machine learning from few training data.Nature communications, 13(1):4919, 2022

    Matthias C Caro, Hsin-Yuan Huang, Marco Cerezo, Kunal Sharma, Andrew Sornborger, Lukasz Cincio, and Patrick J Coles. Generalization in quantum machine learning from few training data.Nature communications, 13(1):4919, 2022

    2022

  14. [14]

    Understanding quantum machine learn- ing also requires rethinking generalization.Nature Communications, 15(1):2277, 2024

    Elies Gil-Fuster, Jens Eisert, and Carlos Bravo- Prieto. Understanding quantum machine learn- ing also requires rethinking generalization.Nature Communications, 15(1):2277, 2024

    2024

  15. [15]

    Understanding gen- eralization in quantum machine learning with mar- gins

    Tak Hur and Daniel K Park. Understanding gen- eralization in quantum machine learning with mar- gins. InForty-second International Conference on Machine Learning, 2024

    2024

  16. [16]

    Towards understanding the power of quantum kernels in the nisq era.Quantum, 5: 531, 2021

    Xinbiao Wang, Yuxuan Du, Y ong Luo, and Dacheng Tao. Towards understanding the power of quantum kernels in the nisq era.Quantum, 5: 531, 2021

    2021

  17. [17]

    Power characterization of noisy quan- tum kernels.IEEE Transactions on Neural Net- works and Learning Systems, 2025

    Y abo Wang, Bo Qi, Xin Wang, Tongliang Liu, and Daoyi Dong. Power characterization of noisy quan- tum kernels.IEEE Transactions on Neural Net- works and Learning Systems, 2025

    2025

  18. [18]

    Generalization error bound for quantum machine learning in nisq era—a survey.Quantum Machine Intelligence, 6(2):90, 2024

    Bikram Khanal, Pablo Rivas, Arun Sanjel, Korn Sooksatra, Ernesto Quevedo, and Alejandro Ro- driguez. Generalization error bound for quantum machine learning in nisq era—a survey.Quantum Machine Intelligence, 6(2):90, 2024

    2024

  19. [19]

    Cambridge university press, 2010

    Michael A Nielsen and Isaac L Chuang.Quantum computation and quantum information. Cambridge university press, 2010. 22

    2010

  20. [20]

    Rademacher and gaussian complexities: Risk bounds and structural results.Journal of Machine Learning Research, 3(Nov):463–482, 2002

    Peter L Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk bounds and structural results.Journal of Machine Learning Research, 3(Nov):463–482, 2002

    2002

  21. [21]

    Pseudo- dimension of quantum circuits.Quantum Machine Intelligence, 2(2):14, 2020

    Matthias C Caro and Ishaun Datta. Pseudo- dimension of quantum circuits.Quantum Machine Intelligence, 2(2):14, 2020

    2020

  22. [22]

    Generalization in quantum machine learn- ing: A quantum information standpoint.PRX Quan- tum, 2(4):040321, 2021

    Leonardo Banchi, Jason Pereira, and Stefano Pi- randola. Generalization in quantum machine learn- ing: A quantum information standpoint.PRX Quan- tum, 2(4):040321, 2021

    2021

  23. [23]

    Tobias Haug and M. S. Kim. Generalization of quantum machine learning models using quan- tum fisher information metric.Phys. Rev. Lett., 133:050603, Jul 2024. doi:10.1103/PhysRevLett. 133.050603. URLhttps://link.aps.org/doi/10. 1103/PhysRevLett.133.050603

    doi:10.1103/physrevlett 2024

  24. [24]

    Statistical complexity of quan- tum circuits.Physical Review A, 105(6):062431, 2022

    Kaifeng Bu, Dax Enshan Koh, Lu Li, Qingxian Luo, and Y aobo Zhang. Statistical complexity of quan- tum circuits.Physical Review A, 105(6):062431, 2022

    2022

  25. [25]

    Effects of quantum resources and noise on the statistical complexity of quantum circuits.Quantum Science and Technology, 8(2): 025013, 2023

    Kaifeng Bu, Dax Enshan Koh, Lu Li, Qingxian Luo, and Y aobo Zhang. Effects of quantum resources and noise on the statistical complexity of quantum circuits.Quantum Science and Technology, 8(2): 025013, 2023

    2023

  26. [26]

    Rademacher com- plexity of noisy quantum circuits.arXiv preprint arXiv:2103.03139, 2021

    Kaifeng Bu, Dax Enshan Koh, Lu Li, Qingx- ian Luo, and Y aobo Zhang. Rademacher com- plexity of noisy quantum circuits.arXiv preprint arXiv:2103.03139, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    Understanding deep learning (still) requires rethinking generaliza- tion.Communications of the ACM, 64(3):107–115, 2021

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Ben- jamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generaliza- tion.Communications of the ACM, 64(3):107–115, 2021

    2021

  28. [28]

    Spectrally-normalized margin bounds for neural networks.Advances in neural information processing systems, 30, 2017

    Peter L Bartlett, Dylan J Foster, and Matus J Tel- garsky. Spectrally-normalized margin bounds for neural networks.Advances in neural information processing systems, 30, 2017

    2017

  29. [29]

    Exponential concentration in quantum kernel methods.Nature communications, 15(1):5200, 2024

    Supanut Thanasilp, Samson Wang, Marco Cerezo, and Zoë Holmes. Exponential concentration in quantum kernel methods.Nature communications, 15(1):5200, 2024

    2024

  30. [30]

    A training algorithm for optimal margin classifiers

    Bernhard E Boser, Isabelle M Guyon, and Vladimir N Vapnik. A training algorithm for optimal margin classifiers. InProceedings of the fifth an- nual workshop on Computational learning theory, pages 144–152, 1992

    1992

  31. [31]

    Quantum ma- chine learning in feature hilbert spaces.Physical review letters, 122(4):040504, 2019

    Maria Schuld and Nathan Killoran. Quantum ma- chine learning in feature hilbert spaces.Physical review letters, 122(4):040504, 2019

    2019

  32. [32]

    Robert J Lyon, BW Stappers, Sally Cooper, John Martin Brooke, and Joshua D Knowles. Fifty years of pulsar candidate selection: from simple fil- ters to a new principled real-time classification ap- proach.Monthly Notices of the Royal Astronomical Society, 459(1):1104–1123, 2016

    2016

  33. [33]

    Supervised learning with quantum-enhanced feature spaces

    Vojt ˇech Havlí ˇcek, Antonio D Córcoles, Kristan Temme, Aram W Harrow, Abhinav Kandala, Jerry M Chow, and Jay M Gambetta. Supervised learning with quantum-enhanced feature spaces. Nature, 567(7747):209–212, 2019

    2019

  34. [34]

    Cross-validation., 2019

    Daniel Berrar et al. Cross-validation., 2019

    2019

  35. [35]

    Principal component analysis.Chemometrics and intelligent laboratory systems, 2(1-3):37–52, 1987

    Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis.Chemometrics and intelligent laboratory systems, 2(1-3):37–52, 1987

    1987

  36. [36]

    Learnability of quan- tum neural networks.PRX quantum, 2(4):040337, 2021

    Yuxuan Du, Min-Hsiu Hsieh, Tongliang Liu, Shan Y ou, and Dacheng Tao. Learnability of quan- tum neural networks.PRX quantum, 2(4):040337, 2021

    2021

  37. [37]

    K-fold cross validation for error rate estimate in support vector machines

    Davide Anguita, Alessandro Ghio, Sandro Ridella, and Dario Sterpi. K-fold cross validation for error rate estimate in support vector machines. InDMIN, pages 291–297, 2009

    2009

  38. [38]

    Pennylane: Automatic differentiation of hybrid quantum-classical compu- tations.arXiv preprint arXiv:1811.04968, 2018

    Ville Bergholm, Josh Izaac, Maria Schuld, Chris- tian Gogolin, Shahnawaz Ahmed, Vishnu Ajith, M Sohaib Alam, Guillermo Alonso-Linaje, B Akash- Narayanan, Ali Asadi, et al. Pennylane: Automatic differentiation of hybrid quantum-classical compu- tations.arXiv preprint arXiv:1811.04968, 2018. 23

    Pith/arXiv arXiv 2018

  39. [39]

    Wood, Jake Lishman, Julien Gacon, Simon Martiel, Paul D

    Ali Javadi-Abhari, Matthew Treinish, Kevin Krsulich, Christopher J. Wood, Jake Lishman, Julien Gacon, Simon Martiel, Paul D. Nation, Lev S. Bishop, An- drew W. Cross, Blake R. Johnson, and Jay M. Gambetta. Quantum computing with Qiskit, 2024

    2024

  40. [40]

    com/, 2025

    IBM Quantum.https://quantum.cloud.ibm. com/, 2025

    2025

  41. [41]

    M" or "B

    Kolby Nottingham Markelle Kelly, Rachel Longjohn. The uci machine learning repository.https:// archive.ics.uci.edu, 1987. 24 A Appendix A.1 Kernel Bound (LLayers) In this section, we present bounds for the case of ker- nel elements affected byLapplications of a local depo- larising noise channel for anN-qubit system. Before we rework our bound to includeL...

    1987

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.