Pith. sign in

REVIEW 4 major objections 5 minor 48 references

The role of data-induced randomness in quantum machine learning classification tasks

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Data-induced randomness, measured by a new class-margin metric, limits quantum classification accuracy.

desk verdict A genuinely useful diagnostic idea, but the central theorem is false as stated and needs a real repair before the claims can stand. read the letter →

arxiv 2411.19281 v1 pith:5PUFU4HJ submitted 2024-11-28 quant-ph cs.LG

Berta Casas , Xavier Bonet-Monroig , Adrián Pérez-Salinas This is my paper · ORCID
classification quant-phcs.LG MSC 81P6868Q12 PACS 03.67.-a03.67.Lx
keywords classmarginquantummachinelearningdataembeddingdata-inducedrandomnessHaar-randomstatesshadowedt-designbinaryclassificationconcentrationofmeasure
verification ladder T0 review T1 audit T2 compute T3 formal

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 tries to establish that the randomness a data-embedding circuit induces in the quantum states it creates sets a limit on the accuracy of quantum binary classifiers. It introduces a class margin, a label-aware random variable that measures how far each point's measurement outcome sits from the decision boundary, and proves that the probability of failing to classify a random point is controlled by the margin's first two moments. The corollary is that an embedding whose states look nearly Haar-random through the classification observable cannot be classified efficiently, because the margin concentrates at the boundary and its variance vanishes exponentially. The authors check the criterion against a provably advantageous quantum learning problem, an ad-hoc task where the choice of observable decides success, and numerical feature-map and data re-uploading models.

What carries the argument

The central object is the class margin $z(x) = \langle\psi(x)|\hat{Z}_y^{(b)}|\psi(x)\rangle$, where $\hat{Z}_y^{(b)}$ equals the classification observable $\hat{O}$ for points whose true label is 0 and equals a boundary-flipped function $f(\hat{O},b)$ for label 1, chosen so that correct answers give $z(x) < b$. Its mean $\mu_1$ and variance $\sigma^2$ over the data set feed Chebyshev's inequality to yield the failure bound of Theorem 1. The variance of any projector observable over an $\hat{O}$-shadowed 2-design is $\exp(-\Omega(n))$ (Appendix A), and this concentration is what converts the moment conditions into a statement about Haar-like randomness forbidding efficient classification.

What would settle it

Exhibit a labeled data set whose embedded states form an exact 2-design (or an $\hat{O}$-shadowed 2-design for the classifier's observable) and whose labels are some known function of the state, and show a fixed-observable classifier reaches high accuracy with polynomially many copies; alternatively, compute the class-margin moments for a distribution that puts most mass just below the decision boundary and a small mass far below it, and check whether classification can succeed even though $\sigma^2/(b-\mu_1)^2$ exceeds any small constant.

Watch

Extended reading notes

Core claim

The central claim is that data-induced randomness, quantified by shadowed statistical moments of the classification observable, bounds the accuracy of QML classification from above. The argument runs through the class margin $z(x) = \langle\psi(x)|\hat{Z}_y^{(b)}|\psi(x)\rangle$, a label-dependent observable that produces values below a threshold $b$ exactly when the model classifies correctly. Theorem 1 bounds the failure probability by $\sigma^2(\hat{Z}_y^{(b)},X) \,/\, (b - \mu_1(\hat{Z}_y^{(b)},X) - \sqrt{\log(2/\delta)/(2M)})^2$, and Corollary 1 turns this into efficiency conditions: the mean margin must sit $\Omega(1/\mathrm{poly}(n))$ below $b$ and the variance must be $O(1/\mathrm{poly}(n))$. Because the variance of any projector over an $\hat{O}$-shadowed 2-design falls as $\exp(-\Omega(n))$, state sets that are essentially random through the classification observable fail these conditions, so efficient accurate classification is impossible there. The paper presents the DLP feature map as a positive example whose anti-randomness is polynomially large, and a specially built family that is a shadowed $t$-design through one observable and exponentially well classified through another.

Load-bearing premise

The conclusion that random-looking embeddings cannot be classified efficiently assumes the true labels carry no structure that is aligned with the classification observable, since the paper's own toy example shows a state family that looks Haar-random through one observable yet is easy to classify through another.

Editorial extensions

If this is right

  • Any QML pipeline whose embedding generates near-Haar states through the classification observable needs exponentially many measurement copies, so such embeddings cannot deliver scalable classification advantage.
  • The first two class-margin moments give a pre-training diagnostic: Monte Carlo estimates on validation data predict whether a feature map or re-uploading model will generalize before expensive optimization.
  • Provable quantum advantage in QML, as in the DLP feature map, coincides with the states being polynomially far from Haar-random through the classification hyperplanes, exactly what the class-margin criterion demands.
  • The embedding and the observable cannot be assessed separately: the same state family that is a shadowed $t$-design through one observable and useless for classification can be exponentially well classified through another observable.
  • For variational models, class margin explains why universality alone is not enough: re-uploading circuits outperform fixed feature maps in the numerics, but their margins drift toward the random regime as qubits and layers grow.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper calls the moment conditions 'necessary,' but the Chebyshev route only makes them sufficient for a small bound; a margin distribution that puts most mass just below the boundary and a small mass far below can give accurate polynomial-copy classification despite a large $\sigma^2/(b-\mu_1)^2$ ratio, so the necessity claim is stronger than the proof supports.
  • The class-margin bound treats labels as given; feeding it a generative model of concept classes would turn it into a sample-complexity statement, telling not just whether a fixed embedding fails but how many data points are needed to certify success.
  • Extending the two-moment analysis to multiclass problems, continuous-spectrum observables, and non-linearly transformed expectation values would test whether the Haar-randomness obstruction persists beyond the restricted linear-observable model considered here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces a metric called the class margin for binary quantum classification tasks, defined by combining the authors' earlier notion of observable-shadowed average randomness with the classical margin concept. For a fixed data-embedding map, a classification observable, and a threshold b, the class margin z(x) measures how far a point is from the decision boundary once the true label is taken into account. The main analytical claim (Theorem 1, Eq. 19) is a Chebyshev-type bound on the probability that a random data point is misclassified or unresolvable, expressed through the mean and variance of the class margin. Corollary 1 translates this into necessary conditions for polynomial-copy classification: the mean margin must be polynomially separated from b and the variance must be polynomially small. From these conditions the authors argue that data embeddings whose states resemble Haar-random states, as seen through the classification observable, cannot be classified efficiently. The paper then applies the framework to three examples: a discrete-logarithm-based feature map with provable quantum advantage, a toy model showing that the same state family can look random through one observable and be classifiable through another, and numerical studies of feature-map and data re-uploading variational classifiers.

Significance. If the central result were correct, the class margin would provide a useful analytical diagnostic connecting data-embedding randomness, expressibility, and classification performance, complementing barren-plateau and kernel-concentration results. The paper's strengths are its clean conceptual setup linking shadowed moments to classification, the analytically worked DLP example, and the explicit toy model in Section III.B showing that observable choice matters for apparent randomness. The numerical experiments illustrate the proposed quantities on concrete variational circuits. However, the main theorem and its corollary contain both a mathematical error in the application of Chebyshev's inequality and an algebraic error in the rearrangement, and these errors affect the exact regime (near-Haar concentration) on which the paper's headline claim rests. The qualitative conclusion is plausible and can likely be repaired with a different concentration argument, but the manuscript as written does not establish its central quantitative claim.

major comments (4)
  1. [Theorem 1 / Appendix D (Eq. 19)] The application of Chebyshev's inequality in Appendix D is invalid in the regime the paper targets. The proof identifies ProbF with Prob(z ≥ b − ε), where ε = sqrt(log(2/δ)/(2M)), and then bounds this by σ²/(b − μ1 − ε)². Two-sided Chebyshev gives P(|z − μ1| ≥ k) ≤ σ²/k² only for k > 0, and the event {z ≥ b − ε} is contained in {|z − μ1| ≥ b − ε − μ1} only when b − ε ≥ μ1. Near Haar, μ1 ≈ b, so b − μ1 − ε is negative and the bound is inapplicable. As a concrete counterexample, if z takes values b ± c with probability 1/2 each, then μ1 = b, σ² = c², and choosing M so that ε = 4c gives ProbF = 1 while Eq. (19) yields c²/(4c)² = 1/16. Corollary 1 and the inference in Section II that Haar encodings violate conditions (21)–(22) depend on exactly this k ≤ 0 regime, so the main quantitative claim is currently unsupported. A one-sided concentration argument on o(x) directly, or an explicit case split with a valid one-sided inequality, is needed.
  2. [Corollary 1 (Eq. 20)] The rearrangement of Theorem 1 into Eq. (20) is algebraically incorrect. To guarantee ProbF ≤ κ one needs σ²/(b − μ1 − ε)² ≤ κ, which gives ε ≤ b − μ1 − σ/√κ and hence 2M/log(2/δ) ≥ (b − μ1 − σ/√κ)⁻². Equation (20) instead contains κ⁻¹σ and writes log2(2/δ). Consequently Eq. (22) should read σ² ≤ κ(b − μ1)² (equivalently σ ≤ √κ(b − μ1)), not σ² ≤ κ(b − μ1). These conditions define the claimed necessary conditions for efficient classification, so the error is load-bearing for the paper's main conclusion.
  3. [Lemmas 2 and 3 / Appendices E-F] The proofs of the stronger tail bounds also need repair. In Appendix E, Bernstein's inequality is stated in a form involving a deviation 2k√V, but the proof does not verify that the failure threshold k = b − ε − μ1 lies in the allowed domain, and the derivation uses the inequality (πt)^{1/t} ≥ π, which is false for t = 2. The moment condition in Eq. (23) does not match the condition obtained in the proof. Lemma 3 similarly inherits the k > 0 issue from Theorem 1, since the failure event is only a one-sided tail of z − μ1; the constant bookkeeping in Appendix F should be re-derived with explicit inequalities. These lemmas are presented as the main refined concentration results, so their proofs must be corrected or their validity restricted to the regime where b − μ1 > ε.
  4. [Section III.C, Numerical results] The conclusion that "classification is unfeasible in this scenario, as no observable can effectively discriminate the embedded data" is not supported by the numerical experiments. The simulations compute the first two moments of the class margin for a single σ(z)-based observable. Showing that this particular observable becomes random does not imply that every observable fails; Section III.B explicitly demonstrates that a state family can be random through one observable while being classifiable through another. The claim should either be restricted to the tested observable or be accompanied by an additional argument that the relevant state ensemble is Haar-like with respect to all observables.
minor comments (5)
  1. [Equation (20)] The notation log2(2/δ) appears to be a typo; the symbol should be log(2/δ) consistently, or the base should be stated explicitly.
  2. [Lemma 1 / Appendix C] The proof in Appendix C uses the specific threshold 1/2 in Eqs. (C3)–(C5), while the lemma is stated for a general threshold b. The proof should be written with a general b so that the statement follows directly.
  3. [Definition 3, Eq. (16)] The definition of f(O,b) is piecewise in terms of o(x) and b, but it is not stated what happens at the boundary o(x) = b for the two pieces; this should be clarified.
  4. [Lemmas 2 and 3] The phrase "the a quantum classifier" appears in the statements of both lemmas and should be corrected.
  5. [Appendix E] The notation is inconsistent between µt and ¯µt: the centered moments are denoted ¯µt in the main text but the proof uses µt without the bar in several places, which can confuse the reader about whether centered or raw moments are meant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central bound is derived from concentration inequalities, and the Haar-margin statements follow from a transparently defined diagnostic metric, not from a fitted parameter or self-citation chain.

full rationale

The paper's derivation chain is self-contained. Lemma 1 follows from Hoeffding's inequality; Theorem 1 applies Chebyshev to the event z >= b - epsilon; Corollary 1 rearranges the theorem; Lemmas 2 and 3 use Bernstein and sub-Gaussian bounds. The Haar-regime scalings used in the examples are computed directly in Appendices A, G and I (Dirichlet moments, DLP overlaps, Gautschi bounds), not imported as a black box. Ref [18] (which shares two authors) supplies the average-randomness framework, but the definitions are restated in Section I.B and the needed moment identities are re-derived in Appendices A and B, so the self-citation is not load-bearing. The observation that a Haar-random embedding gives mu1 = 1/2 = b for a projector at threshold b=1/2 is a straightforward consequence of the class-margin definition and the Haar mean; it is a designed diagnostic statement, not a fitted prediction with the answer baked in. Whatever concerns exist about Theorem 1's validity in the near-Haar regime (the two-sided Chebyshev step requires b - mu1 - epsilon > 0) are mathematical correctness issues, not circularity. No parameter is fitted to a subset and then renamed a prediction, and no uniqueness claim is imported from the authors' prior work.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claims rest on standard concentration inequalities, on the authors' prior shadowed-moment formulas for Haar-random states (ref [18]), on large-margin properties of the DLP feature map from ref [4], and on an unstated label-independence premise in the Haar-randomness argument. No parameters are fitted to data; the threshold b and the DLP constant c are carried from the problem setup.

free parameters (3)
  • Threshold b = b = 1/2 in all examples
    The classification threshold is a free hyperparameter of the classifier; the paper's examples set b=1/2, for which the class margin simplifies to a label-flipped expectation. The analysis is not carried out for general b.
  • Constant c in DLP feature map = k = n - c log n, c unspecified constant
    Controls the size of the adversarial gap Delta = 2^{k+1}/p in Theta(1/poly(n)); the value is carried over from the DLP QML construction of Liu et al. (ref [4]), not fitted here.
  • Constant L in Lemmas 2 and 3
    Appears in the moment growth conditions of the Bernstein and sub-gaussian bounds; no value is given, and the proofs do not establish its existence for realistic encodings.
assumptions (4)
  • standard math Hoeffding, Chebyshev, Bernstein, and sub-gaussian concentration inequalities
    Used in proofs of Lemma 1, Theorem 1, and Lemmas 2-3 (Appendices C-F).
  • domain assumption Haar-random shadowed moments of projectors are given by the Dirichlet representation in ref [18]
    Appendix A computes mu1 and sigma^2 for Haar-random states using the authors' prior framework; the paper relies on these formulas without re-deriving them.
  • domain assumption The DLP feature map satisfies the large-margin hyperplane properties of ref [4]
    Lemma 4 and Theorem 2 use the four projector-overlap properties listed in Appendix G, taken from Liu et al.
  • ad hoc to paper Unstated label-independence: for a Haar-random set of embedded states to imply zero classification margin, the labels must not be correlated with the observable
    This premise underwrites the interpretation after Corollary 1 and the abstract's broad claim; it is never stated and is false in general (see Section III.B).
invented entities (1)
  • Class margin z(x)
    purpose: A label- and threshold-dependent random variable (eqs 14-16) that measures the signed distance of a data point's expected measurement from the classification boundary; used to connect data randomness to classification accuracy.
    It is a new constructed quantity, not observable in an experiment independently of the paper's own definitions; its utility is exactly the theoretical claims being evaluated. It is not a physical entity, so the graviton-problem severity does not apply.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The role of data-induced randomness in quantum machine learning classification tasks." pith.science (2026). https://pith.science/paper/5PUFU4HJ

@misc{pith2026241119281,
  author       = {Pith},
  title        = {Pith review of: The role of data-induced randomness in quantum machine learning classification tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PUFU4HJ}},
  note         = {Machine review of arXiv:2411.19281}
}
read the original abstract

Quantum machine learning (QML) has surged as a prominent area of research with the objective to go beyond the capabilities of classical machine learning models. A critical aspect of any learning task is the process of data embedding, which directly impacts model performance. Poorly designed data-embedding strategies can significantly impact the success of a learning task. Despite its importance, rigorous analyses of data-embedding effects are limited, leaving many cases without effective assessment methods. In this work, we introduce a metric for binary classification tasks, the class margin, by merging the concepts of average randomness and classification margin. This metric analytically connects data-induced randomness with classification accuracy for a given data-embedding map. We benchmark a range of data-embedding strategies through class margin, demonstrating that data-induced randomness imposes a limit on classification performance. We expect this work to provide a new approach to evaluate QML models by their data-embedding processes, addressing gaps left by existing analytical tools.

Figures

Figures reproduced from arXiv: 2411.19281 by the authors.

Figure 1
Figure 1. FIG. 1: Graphical interpretation of (a) tunable decision [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Illustration of the classification criteria and the defini [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Numerical estimation of the anti-randomness [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Numerical computations for the statistical moments [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5: Numerical computations for the statistical moments [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6: Quantum circuits for encoding the dataset [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 33 canonical work pages

  1. [1]

    P. W. Shor, SIAM Journal on Computing 26, 1484 (1997), ISSN 0097-5397

    work page 1997

  2. [2]

    Quantum Inspire - the Dutch Quantum Computer in the Cloud

    serve as an analysis of the landscape. In the feature-map case, we use a brick - and non- brick data-embedding circuits (see Appendix J for de- tails). The results of the numerical experiments are shown in Figure 4. In the training set (left column), we see that µ1(zθ(x)) concentrates around 1 /2. In contrast, σ2(zθ(x)) ap- proaches zero for both brick an...

    work page 2026

  3. [3]

    1) and 2) give information about how randomness affects model performance, while

    averaging over randomly distributed θ values with x sampled from the test set. 1) and 2) give information about how randomness affects model performance, while

    work page

  4. [4]

    R. P. Feynman, International Journal of Theoretical Physics 21, 467 (1982), ISSN 1572-9575

    work page 1982

  5. [5]

    Huang, M

    H.-Y. Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, Nature Commu- nications 12, 2631 (2021), ISSN 2041-1723, 2011.01938

    arXiv 2021

  6. [6]

    Y. Liu, S. Arunachalam, and K. Temme, Nature Physics 17, 1013 (2021), ISSN 1745-2481

    work page 2021

  7. [7]

    Molteni, C

    R. Molteni, C. Gyurik, and V. Dunjko, Exponential quan- tum advantages in learning quantum observables from classical data (2024), arXiv:2405.02027

    arXiv 2024

  8. [8]

    Gyurik and V

    C. Gyurik and V. Dunjko, Exponential separa- tions between classical and quantum learners (2023), arXiv:2306.16028

    arXiv 2023

Show all 48 references
  1. [9]

    Wakeham and M

    D. Wakeham and M. Schuld, Inference, interference and invariance: How the quantum fourier transform can help to learn from data (2024), arXiv:2409.00172

    2024 arXiv

  2. [10]

    Gil-Fuster, C

    E. Gil-Fuster, C. Gyurik, A. P´ erez-Salinas, and V. Dun- jko, On the relation between trainability and dequanti- zation of variational quantum learning models (2024), arXiv:2406.07072

    2024 arXiv

  3. [11]

    Schuld, Supervised quantum machine learning models are kernel methods (2021), arXiv:2101.11020

    M. Schuld, Supervised quantum machine learning models are kernel methods (2021), arXiv:2101.11020

    2021 arXiv

  4. [12]

    Havl ´ ıˇ cek, A

    V. Havl ´ ıˇ cek, A. D. C´ orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Nature 567, 209 (2019), ISSN 1476-4687

    2019

  5. [13]

    P´ erez-Salinas, A

    A. P´ erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, Quantum 4, 226 (2020)

    2020

  6. [14]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications 9, 4812 (2018), ISSN 2041-1723

    2018

  7. [15]

    Holmes, K

    Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, PRX Quantum 3, 010313 (2022)

    2022

  8. [16]

    S. Sim, P. D. Johnson, and A. Aspuru-Guzik, Advanced Quantum Technologies 2, 1900070 (2019), ISSN 2511- 9044

    2019

  9. [17]

    Cerezo, A

    M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Nature Communications 12, 1791 (2021), ISSN 2041-1723

    2021

  10. [18]

    Larocca, N

    M. Larocca, N. Ju, D. Garc ´ ıa-Mart ´ ın, P. J. Coles, and M. Cerezo, Nature Computational Science 3, 542 (2023)

    2023

  11. [19]

    Larocca, P

    M. Larocca, P. Czarnik, K. Sharma, G. Muraleedharan, P. J. Coles, and M. Cerezo, Quantum6, 824 (2022), ISSN 2521-327X, 2105.14377

    2022 arXiv

  12. [20]

    Bonet-Monroig, H

    X. Bonet-Monroig, H. Wang, and A. P´ erez-Salinas,Ver- ifying randomness in sets of quantum states via observ- ables (2024), arXiv:2404.16211

    2024 arXiv

  13. [21]

    P. L. Bartlett, P. M. Long, and R. C. Williamson, 52, 434 (1996), ISSN 0022-0000

    1996

  14. [22]

    V. N. Vapnik, The Nature of Statistical Learning Theory (Springer, 2000), ISBN 978-1-4419-3160-3 978-1-4757- 3264-1

    2000

  15. [23]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, et al., Nature Reviews Physics 3, 625 (2021)

    2021

  16. [24]

    Bharti, A

    K. Bharti, A. Cervera-Lierta, T. H. Kyaw, T. Haug, S. Alperin-Lea, A. Anand, M. Degroote, H. Heimonen, J. S. Kottmann, T. Menke, et al., Reviews of Modern Physics 94, 015004 (2022)

    2022

  17. [25]

    Schuld, R

    M. Schuld, R. Sweke, and J. J. Meyer, Physical Re- view A 103, 032430 (2021), ISSN 2469-9926, 2469-9934, 2008.08605

    2021 arXiv

  18. [26]

    Cortes and V

    C. Cortes and V. Vapnik, Machine Learning 20, 273 (1995), ISSN 1573-0565

    1995

  19. [27]

    Sch¨ olkopf, R

    B. Sch¨ olkopf, R. Herbrich, and A. J. Smola, in In- ternational conference on computational learning theory (Springer, 2001), pp. 416–426. 12

    2001

  20. [28]

    Delsarte, Journal of Combinatorial Theory, Series A 20, 230 (1976), ISSN 0097-3165

    P. Delsarte, Journal of Combinatorial Theory, Series A 20, 230 (1976), ISSN 0097-3165

    1976

  21. [29]

    Ambainis and J

    A. Ambainis and J. Emerson, Quantum t-designs: T- wise independence in the quantum world (2007), quant- ph/0701126

    2007

  22. [30]

    Peruzzo, J

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, Nature Communications 5, 4213 (2014), ISSN 2041-1723

    2014

  23. [31]

    Preskill, Quantum 2, 79 (2018)

    J. Preskill, Quantum 2, 79 (2018)

    2018

  24. [32]

    Arrasmith, M

    A. Arrasmith, M. Cerezo, P. Czarnik, L. Cincio, and P. J. Coles, Quantum 5 (2021)

    2021

  25. [33]

    Jerbi, L

    S. Jerbi, L. J. Fiderer, H. Poulsen Nautrup, J. M. K¨ ubler, H. J. Briegel, and V. Dunjko, Nature Communications 14, 517 (2023), ISSN 2041-1723

    2023

  26. [34]

    J. M. K¨ ubler, S. Buchholz, and B. Sch¨ olkopf,The Induc- tive Bias of Quantum Kernels (2021), arXiv:2106.03747

    2021 arXiv

  27. [35]

    Rebentrost, M

    P. Rebentrost, M. Mohseni, and S. Lloyd, Physical Re- view Letters 113, 130503 (2014)

    2014

  28. [36]

    J. G. Vidal and D. O. Theis, Input Redundancy for Pa- rameterized Quantum Circuits (2020), arXiv:1901.11434

    2020 arXiv

  29. [37]

    M. C. Caro, E. Gil-Fuster, J. J. Meyer, J. Eisert, and R. Sweke, Quantum 5, 582 (2021)

    2021

  30. [38]

    M. C. Caro, H.-Y. Huang, N. Ezzell, J. Gibbs, A. T. Sornborger, L. Cincio, P. J. Coles, and Z. Holmes, Nature Communications 14, 3751 (2023), ISSN 2041-1723

    2023

  31. [39]

    Hur and D

    T. Hur and D. K. Park, Understanding generalization in quantum machine learning with margins (2024), arXiv: 2411.06919

    2024

  32. [40]

    Olkin and H

    I. Olkin and H. Rubin, The Annals of Mathematical Statistics 35, 261 (1964)

    1964

  33. [41]

    C. W. Helstrom, Quantum Detection and Estimation Theory, no. v. 123 in Mathematics in Science and En- gineering (Academic Press, New York, 1976), ISBN 978- 0-12-340050-5

    1976

  34. [42]

    P´ erez-Salinas, D

    A. P´ erez-Salinas, D. L´ opez-N´ u˜ nez, A. Garc ´ ıa-S´ aez, P. Forn-D ´ ıaz, and J. I. Latorre, Physical Review A104, 012405 (2021), ISSN 2469-9926, 2469-9934

    2021

  35. [43]

    Barthe and A

    A. Barthe and A. P´ erez-Salinas, Gradients and fre- quency profiles of quantum re-uploading models (2023), arXiv:2311.10822

    2023 arXiv

  36. [44]

    Thanasilp, S

    S. Thanasilp, S. Wang, M. Cerezo, and Z. Holmes, Nature Communications 15, 5200 (2024)

    2024

  37. [45]

    S.N.Bernstein, Ann. Sci. Inst. Sav. Ukraine 4 (1924)

    1924

  38. [46]

    Bercu, B

    B. Bercu, B. Delyon, and E. Rio, Concentration Inequal- ities for Sums and Martingales , SpringerBriefs in Math- ematics (Springer International Publishing, 2015), ISBN 978-3-319-22099-4

    2015

  39. [47]

    J. G. Wendel, The American Mathematical Monthly 55, 563 (1948), ISSN 00029890, 19300972

    1948

  40. [48]

    1 2 − √x⌊n/2⌋x⌈n/2⌉ t# = E

    W. Gautschi, Journal of Mathematics and Physics 38, 77 (1959). 13 Appendix A: Analytic expression of the variance for Haar random-states In this section, we derive an analytic expression for the variance of a given observable ˆO when the family of states S = {|ψ(x)⟩} forms, at...

    1959

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.