Pith. sign in

REVIEW 3 major objections 7 minor 58 references

The interplay of robustness and generalization in quantum machine learning

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Penalizing data-encoding weights makes quantum models both more robust and better at generalizing.

desk verdict The time-series experiments are the new part; the generalization claim needs a fix for the n-dependence statement and a validation split for lambda. read the letter →

arxiv 2506.08455 v1 pith:NPOTRLHE submitted 2025-06-10 quant-ph cs.LGcs.SYeess.SY

Julian Berberich , Tobias Fellner , Christian Holm This is my paper
classification quant-phcs.LGcs.SYeess.SY
keywords quantummachinelearningvariationalcircuitsadversarialrobustnessgeneralizationboundsLipschitzconstantsregularizationtrainabledataencodingtimeseriesanalysis
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

Variational quantum models are function approximators built from parameterized quantum circuits, and this paper makes the case that their robustness to input perturbations and their ability to generalize from training data are two sides of the same quantity: a Lipschitz bound that depends on the trainable weights multiplying the data. Because the bound is parameter-dependent, a regularized training loss that penalizes the norm of those encoding weights can shrink the Lipschitz constant directly, with no extra computational cost. The paper reviews this mechanism, highlights that it only works for trainable data encoding (fixed encoding cannot adjust the bound), and validates it numerically by inferring the parameter of a chaotic logistic map from short time series. The central practical claim is that a moderate regularization strength reduces test error by roughly fifty percent on that task, and makes the model markedly less sensitive to input noise.

What carries the argument

The load-bearing identity is the parameter-dependent Lipschitz bound $L_{\Theta} = 2\|M\| \sum_j \|w_j\| \|H_j\|$ for circuits made of unitaries $U_{j,\Theta_j}(x) = e^{-i(w_j^\top x + \theta_j) H_j}$. It does the work of connecting three separate ideas: robustness (output change is bounded by $L_{\Theta}\|\varepsilon\|$), generalization (the expected-risk gap is bounded by the same Lipschitz term), and trainability (only the data-encoding weights $w_j$ appear, so only trainable encoding can influence the bound during optimization). The companion mechanism is the regularized loss (9), which penalizes $\|w_j\|^2\|H_j\|^2$ and thereby turns the theory into a training algorithm.

What would settle it

Train a variational quantum model on a fixed task with increasing $\lambda$, and measure both the Lipschitz bound $L_{\Theta}$ and the worst-case test error under input noise; if shrinking $\|w_j\|$ does not lower the worst-case error or the generalization gap (with comparable training error), the claimed robustness-generalization link fails. A sharper test: construct two parameter settings with the same $L_{\Theta}$ but different generalization gaps, which would show the Lipschitz bound is not the controlling quantity.

Watch

Extended reading notes

Core claim

The central object is the bound $L_{\Theta} = 2\|M\| \sum_{j=1}^N \|w_j\| \|H_j\|$, which the paper (following its reference [21]) establishes as a Lipschitz bound for the variational quantum model $f_{\Theta}(x) = \langle 0 | U_{\Theta}(x)^\dagger M U_{\Theta}(x) | 0 \rangle$. Any perturbation of the input changes the output by at most $L_{\Theta}$ times the perturbation size, so smaller $L_{\Theta}$ means better worst-case robustness. The same Lipschitz constant appears in an informal generalization bound, $E(\Theta) \leq C_1 \|M\| \sum_j \|w_j\| \|H_j\| + C_2/\sqrt{n}$, meaning a small Lipschitz bound also shrinks the gap between training and expected test loss. The paper therefore proposes training with the regularized loss (9), which adds $\lambda \sum_j \|w_j\|^2 \|H_j\|^2$ to the empirical risk, and shows on a four-qubit circuit that larger $\lambda$ indeed reduces the Lipschitz bound, improves noise robustness, and lowers the generalization gap, with a sweet spot around $\lambda = 0.004$.

Load-bearing premise

The whole argument assumes that Theorem 1 and Theorem 2 from [21] transfer unchanged to the specific four-qubit circuit used here, in particular that its constant CNOT gates fit the unitary form (1) with $w_j = 0$ and $\theta_j = 1$, and that the informal generalization bound is dominated by its Lipschitz term at finite training set size.

Editorial extensions

If this is right

  • Training a variational quantum model with the regularized loss (9) should yield models whose prediction error degrades more slowly under input noise, exactly as seen in Figure 6 for the logistic-map task.
  • Moderate regularization improves test accuracy on unseen data: the paper reports roughly a 50% improvement relative to no regularization on the time-series inference task, and a generalization gap that drops rapidly with small $\lambda$.
  • Fixed data encoding cannot exploit this mechanism because its Lipschitz bound is set by offline design choices; trainable encoding is therefore not just an expressivity upgrade but a robustness and generalization control knob.
  • The Lipschitz-bound perspective suggests a smooth trade-off between expressivity and robustness: large $\lambda$ over-constrains the model and raises training error, so the optimal $\lambda$ is finite and problem-dependent.
  • The regularization cost is a closed-form term added to the loss, so it improves robustness without adversarial training or quantum noise injection, keeping the training procedure no more expensive than ordinary empirical-risk minimization.

Reading between the lines

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

  • The same Lipschitz-regularization recipe could be tested on other quantum learning paradigms, such as quantum reinforcement learning, where parameter-dependent bounds of the same shape are already being explored; the paper's own survey points toward that arena.
  • The mechanism mirrors classical spectral-norm and Lipschitz regularization in deep networks, suggesting that the robustness-generalization link is structural rather than quantum-specific; a direct classical-quantum comparison under identical regularization would clarify what the quantum setting adds.
  • One testable extension is to use the Lipschitz bound as a per-layer diagnostic: since $L_{\Theta}$ is a sum over encoding layers, the contribution $\|w_j\| \|H_j\|$ of each layer identifies which input features are most responsible for fragility, potentially guiding feature selection.
  • The paper's experiment infers the logistic-map parameter from 12 time steps; applying the same regularized training to higher-dimensional or out-of-distribution time series would show whether the reported improvement persists beyond the specific chaotic regime.
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

3 major / 7 minor

Summary. The manuscript is a book chapter on the interplay between adversarial robustness and generalization in variational quantum machine learning. It reviews the Lipschitz bound L_Theta = 2||M|| sum_j ||w_j|| ||H_j|| for the quantum model f_Theta in Eq. (4) (Theorem 1) and the parameter-dependent generalization bound in Eq. (13) (Theorem 2), both attributed to the authors' earlier work [21]. Based on these results, the chapter proposes, following [21], a ridge-type regularization of the encoding weights in Eq. (9) to improve robustness and generalization. The new contribution is a numerical time-series study: inferring the logistic-map parameter r from length-12 sequences using a four-qubit data-reuploading circuit. The authors report that increasing the regularization parameter lambda lowers the Lipschitz bound and improves robustness under high noise, while the test MSE exhibits a U-shape with a minimum at lambda = 0.004, corresponding to an approximately 50% improvement over lambda = 0. The code for the numerical experiments is publicly available on GitHub.

Significance. If the claims hold, the chapter usefully demonstrates that trainable data encoding provides a practical knob for robustness and generalization in QML, extending the empirical evidence for Lipschitz-based regularization to a regression and time-series setting. The exposition is clear and properly attributes the theoretical results to [21], and the numerical study is reproducible with provided code. The main weaknesses are methodological: the optimal lambda is chosen by inspecting the test-set curve without a separate validation split, and the observed generalization improvement is not causally isolated from a mere capacity-reduction effect. These issues do not undermine Theorem 1, which is on solid ground, but they weaken the paper's central empirical claim about generalization.

major comments (3)
  1. [Section 5.4, Figure 7] The optimal regularization strength lambda = 0.004 is identified by inspecting the test MSE curve without a separate validation split. This makes the reported approximately 50% test-MSE improvement an optimistic in-sample estimate rather than a fair estimate of generalization to unseen data. Please add a validation set (or nested cross-validation) and report the test performance at the validation-selected lambda, with the associated uncertainty.
  2. [Section 4, Eq. (13) and Section 5.4] The observed U-shaped test error and decreasing generalization gap are consistent with the Lipschitz-based mechanism in Theorem 2, but they are also fully consistent with a pure capacity/expressivity effect, because increasing lambda shrinks the encoding weights and thereby reduces model complexity. The statement in Section 5.4 that these findings 'support the theoretical connection between generalization and the Lipschitz bound as stated in Theorem 2' therefore overstates the evidence. Please either weaken the causal claim or design an additional experiment that separates the two mechanisms, for example by comparing models with matched representation capacity but different Lipschitz bounds.
  3. [Section 2, Eq. (1) and Section 5.2] The circuit in Figure 5 includes CNOT gates, which are described as constant gates with w_j = 0 and theta_j = 1. This is not a correct instance of the exponential form U_j = e^{-i(w_j^T x + theta_j) H_j}. A CNOT gate can be written in this form only with a suitable Hermitian generator, for example e^{-i pi/4 (I-Z) tensor (I-X)}. Please correct the description or explicitly state how constant gates are absorbed into the form (1), so that the application of Theorem 1 to the numerical circuit is unambiguous.
minor comments (7)
  1. [Section 4, after Eq. (13)] The sentence 'One can show that the bound (13) vanishes in the infinite data limit n -> infinity' is inaccurate, since the first term C1||M|| sum_j ||w_j|| ||H_j|| is independent of n; only the second term C2/sqrt(n) vanishes. Please rephrase to state that the sampling term vanishes and the bound approaches the Lipschitz term.
  2. [Figure 6 caption] The caption uses the word 'accuracy' while the y-axis label reads 'MSE (worst case)'; please use a single consistent term (for example 'mean squared error') throughout the caption and axis labels.
  3. [Section 5.3] The text says 'the worst-case mean squared error ... over 100 independently perturbed versions'; since the perturbations are random, the 'worst case' over a finite sample is not a worst case over all possible perturbations. Consider reporting the maximum over the sampled perturbations or clarifying the statistical nature of the reported quantity.
  4. [Section 5.1] The statement 'we fix the initial condition to x_1 = 0.5' leaves implicit that the same initial condition is used for all training and test sequences. Please state this explicitly, since it affects the interpretation of the learned mapping from sequences to r.
  5. [References] Reference [52] appears to be an unpublished preprint; please include the arXiv identifier or a journal reference if available.
  6. [Footnote on page 4] There is a typo: 'all involves matrix and vector norms' should read 'all involved matrix and vector norms'.
  7. [Abstract] The heading 'Key words:' should be 'Keywords:' to match standard formatting.

Circularity Check

2 steps flagged · score 4.0 of 10

Central theory is imported from a self-cited prior paper, but the new numerical experiment provides independent evidence, so the circularity is partial rather than complete.

  1. self citation load bearing [Sections 3.1, 3.2, and 4 (Eqs. 7, 9, 13)]
    "The following result from [21] provides a Lipschitz bound for the variational quantum model fΘ. Theorem 1 [21] ... LΘ = 2∥M∥∑Nj=1 ∥wj∥∥Hj∥. ... instead of the original training problem (8), it was proposed by [21] to solve the modified problem minΘ 1/n ∑nk=1 ℓ(fΘ(xk), yk) + λ∑Nj=1 ∥wj∥²∥Hj∥². ... Theorem 2 [21] (Informal). The generalization error (12) is bounded as E(Θ) ≤ C1∥M∥∑Nj=1 ∥wj∥∥Hj∥ + C2/√n."

    The chapter's theoretical backbone - that penalizing encoding weights improves robustness and, via Theorem 2, generalization - is not derived here but imported from Ref. [21]. Ref. [21] shares two authors with this chapter (Berberich and Holm), and the present text explicitly states 'it was proposed by [21]' rather than proving the results. The numerical circuit in Section 5.2 is asserted to fit the model (1) with constant CNOT gates corresponding to wj=0, but no verification or re-derivation of the theorems is given for this specific circuit. Thus the explanation of the numerical observations rests on a self-citation chain. However, the measured training/test MSE values are not computed from the bound, so the empirical results retain independent content.

  2. fitted input called prediction [Section 5.4, Figure 7 discussion]
    "In contrast, on the test dataset, we observe that the prediction error reaches a minimum at λ = 0.004. Relative to the case without regularization this corresponds to an improvement in performance of approximately 50%. ... In addition, we observe that the Lipschitz bound decreases monotonically with increasing λ, consistent with the original motivation for incorporating the regularization term, namely, to constrain the model's Lipschitz constant."

    The 'approximately 50% improvement' is obtained by sweeping λ and selecting the best value on the test set without a separate validation split, so it is an in-sample selection result rather than a prediction for unseen hyperparameters. The monotone decrease of the Lipschitz bound with λ is also a near-direct consequence of adding λ∑∥w∥²∥H∥² to the training objective, since that term explicitly penalizes the quantities appearing in the Lipschitz bound. Consequently, these observations partially confirm the imported theory by construction rather than by independent test, though the raw MSE values themselves are not derived from the bound.

full rationale

This is a review-and-application chapter. Theorems 1 and 2, as well as the regularized training problem, are explicitly credited to Ref. [21] instead of being re-derived. Because two of the present authors co-authored Ref. [21], the theoretical motivation of the chapter is load-bearing self-citation. That alone is not full circularity: the new numerical experiment on the logistic map is a genuine, externally meaningful test, and the measured robustness and test MSE are not computed from the Lipschitz bound. The circularity is partial because the claimed validation of the theory is weakened in two ways: (i) the central bound (13) is quoted informally and its stated 'vanishes as n→∞' behavior is internally inconsistent as written, since the Lipschitz term is n-independent; and (ii) the generalization improvement is reported after selecting λ on the test set, turning a hyperparameter sweep into an apparent confirmation. These issues are correctness and statistical-selection concerns as much as circularity. On balance, some self-citation is load-bearing, but the central empirical claim still has independent content, so a score of 4 is appropriate.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The chapter's theoretical core is imported entirely from [21], a self-cited prior paper, and the numerical section introduces one tuned hyperparameter lambda. No new axioms or entities are added, but the transfer of the bounds to the specific circuit and the finite-sample regime is assumed.

free parameters (1)
  • regularization strength lambda = 0, 0.004, 0.03 (optimal 0.004 on test set)
    Introduced as a hyperparameter in Eq. (9); the claimed benefit depends on choosing lambda=0.004, selected by inspecting the test-error curve in Figure 7, with no separate validation split described.
assumptions (3)
  • domain assumption Theorem 1 from [21]: L_Theta = 2||M|| sum_j ||w_j|| ||H_j|| is a valid Lipschitz bound for f_Theta.
    Used in Section 3.1 to justify regularization; proof is not included in this chapter.
  • domain assumption Theorem 2 from [21]: the generalization error is bounded by C1||M|| sum_j ||w_j|| ||H_j|| + C2/sqrt(n) for constants C1,C2.
    Stated informally in Section 4 and used to argue that reducing encoding weights improves generalization; proof not included.
  • domain assumption The numerical circuit (Pauli-Z and Pauli-Y rotations with CNOT entanglement layers and observable M=ZxxZxxZxxZ) belongs to the model class (1)-(4), and the constant CNOT gates can be represented in the form (1) with w_j=0, theta_j=1.
    Needed for Theorems 1 and 2 to apply to the trained models in Section 5; no explicit construction is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The interplay of robustness and generalization in quantum machine learning." pith.science (2026). https://pith.science/paper/NPOTRLHE

@misc{pith2026250608455,
  author       = {Pith},
  title        = {Pith review of: The interplay of robustness and generalization in quantum machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPOTRLHE}},
  note         = {Machine review of arXiv:2506.08455}
}
read the original abstract

While adversarial robustness and generalization have individually received substantial attention in the recent literature on quantum machine learning, their interplay is much less explored. In this chapter, we address this interplay for variational quantum models, which were recently proposed as function approximators in supervised learning. We discuss recent results quantifying both robustness and generalization via Lipschitz bounds, which explicitly depend on model parameters. Thus, they give rise to a regularization-based training approach for robust and generalizable quantum models, highlighting the importance of trainable data encoding strategies. The practical implications of the theoretical results are demonstrated with an application to time series analysis.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 50 canonical work pages

  1. [21]

    Training robust and generalizable quantum models,

    J. Berberich, D. Fink, D. Pranji´c, C. Tutschku, and C. Holm, “Training robust and generalizable quantum models,” Physical Review Research, vol. 6, p. 043326, 2024

    work page 2024

  2. [1]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” arXiv:1412.6572, 2014

    arXiv 2014

  3. [2]

    Intriguing properties of neural networks,

    C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus, “Intriguing properties of neural networks,” arXiv:1312.6199, 2014

    arXiv 2014

  4. [3]

    Provable defenses against adversarial examples via the convex outer adversarial polytope,

    E. Wong and Z. Kolter, “Provable defenses against adversarial examples via the convex outer adversarial polytope,” in Proc. International Conference on Machine Learning , 2018, pp. 5283–5292. [Online]. Available: https://proceedings.mlr.press/v80/wong18a/wong18a.pdf

    work page 2018

  5. [4]

    Lipschitz-margin training: Scalable certification of perturbation invariance for deep neural networks,

    Y. Tsuzuku, I. Sato, and M. Sugiyama, “Lipschitz-margin training: Scalable certification of perturbation invariance for deep neural networks,” in Proc. Advances in Neural Information Processing Systems, 2018, pp. 6541–6550. [Online]. Available: https://proceedings.neurips. cc/paper files/paper/2018/file/485843481a7edacbfce101ecb1e4d2a8-Paper.pdf

    work page arXiv 2018

  6. [5]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” arXiv:1706.06083, 2019

    arXiv 2019

  7. [6]

    A simple weight decay can improve generalization,

    A. Krogh and J. Hertz, “A simple weight decay can improve generalization,” in Advances in Neural Information Processing Systems , vol. 4, 1991. [Online]. Available: https://proceedings.neurips.cc/paper/1991/file/8eefcfdf5990e441f0fb6f3fad709e21-Paper.pdf

    work page 1991

  8. [7]

    Robustness and generalization,

    H. Xu and S. Mannor, “Robustness and generalization,” Mach. Learn. , vol. 86, no. 3, pp. 391–423, 2012

    work page 2012

Show all 58 references
  1. [8]

    Distillation as a defense to adversarial perturbations against deep neural networks,

    N. Papernot, P. McDaniel, X. Wu, S. Jha, and A. Swami, “Distillation as a defense to adversarial perturbations against deep neural networks,” inProc. IEEE Symposium on Security and Privacy (SP), 2016, pp. 582–597

    2016

  2. [9]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  3. [10]

    Challenges and opportunities in quantum machine learning,

    M. Cerezo, G. Verdon, H.-Y. Huang, L. Cincio, and P. J. Coles, “Challenges and opportunities in quantum machine learning,” Nature Computational Science, vol. 2, pp. 567–576, 2022

    2022

  4. [11]

    Quantum adversarial machine learning: status, challenges and perspectives,

    D. Edwards and D. B. Rawat, “Quantum adversarial machine learning: status, challenges and perspectives,” in Proc. 2nd IEEE Int. Conf. Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA), 2020, pp. 128–133

    2020

  5. [12]

    Vulnerability of quantum classification to adversarial perturbations,

    N. Liu and P. Wittek, “Vulnerability of quantum classification to adversarial perturbations,” Physical Review A, vol. 101, p. 062331, 2020

    2020

  6. [13]

    Quantum adversarial machine learning,

    S. Lu, L.-M. Duan, and D.-L. Deng, “Quantum adversarial machine learning,”Physical Review Research, vol. 2, p. 033212, 2020

    2020

  7. [14]

    Quantum noise protects quantum classifiers against adversaries,

    Y. Du, M.-H. Hsieh, T. Liu, D. Tao, and N. Liu, “Quantum noise protects quantum classifiers against adversaries,” Physical Review Research, vol. 3, p. 023153, 2021

    2021

  8. [15]

    Robustness verification of quantum classifiers,

    J. Guan, W. Fang, and M. Ying, “Robustness verification of quantum classifiers,” in Proc. Int. Conf. Computer Aided Verification, 2021, pp. 151–174

    2021

  9. [16]

    Robust in practice: adversarial attacks on quantum machine learning,

    H. Liao, I. Convy, W. J. Huggins, and K. B. Whaley, “Robust in practice: adversarial attacks on quantum machine learning,” Physical Review A, vol. 103, p. 042427, 2021

    2021

  10. [17]

    Optimal provable robustness of quantum classification via quantum hypothesis testing,

    M. Weber, N. Liu, B. Li, C. Zhang, and Z. Zhao, “Optimal provable robustness of quantum classification via quantum hypothesis testing,” Quantum Information, vol. 7, p. 76, 2021

    2021

  11. [18]

    Universal adversarial examples and perturbations for quantum classifiers,

    W. Gong and D.-L.-. Deng, “Universal adversarial examples and perturbations for quantum classifiers,” National Science Review, vol. 9, no. 6, p. nwab130, 2022

    2022

  12. [19]

    Towards quantum enhanced adversarial robustness in machine learning,

    M. T. West, S.-L. Tsang, J. S. Low, C. D. Hill, C. Leckie, L. C. L. Hollenberg, S. M. Erfani, and M. Usman, “Towards quantum enhanced adversarial robustness in machine learning,” Nature Machine Intelligence, vol. 5, pp. 581–589, 2023. The interplay of robustness and generaliza...

    2023

  13. [20]

    Bench- marking adversarially robust quantum machine learning at scale,

    M. T. West, S. M. Erfani, C. Leckie, M. Sevior, L. C. L. Hollenberg, and M. Usman, “Bench- marking adversarially robust quantum machine learning at scale,” Physical Review Research, vol. 5, p. 023186, 2023

    2023

  14. [22]

    The power of quantum neural networks,

    A. Abbas, D. Sutter, C. Zoufal, A. Lucchi, A. Figalli, and S. Woerner, “The power of quantum neural networks,” Nature Computational Science, vol. 1, pp. 403–409, 2021

    2021

  15. [23]

    Generalization in quantum machine learning: a quantum information standpoint,

    L. Banchi, J. Pereira, and S. Pirandola, “Generalization in quantum machine learning: a quantum information standpoint,” PRX Quantum, vol. 2, p. 040321, 2021

    2021

  16. [24]

    Encoding-dependent general- ization bounds for parametrized quantum circuits,

    M. C. Caro, E. Gil-Fuster, J. J. Meyer, J. Eisert, and R. Sweke, “Encoding-dependent general- ization bounds for parametrized quantum circuits,” Quantum, vol. 5, p. 582, 2021

    2021

  17. [25]

    Power of data in quantum machine learning,

    H.-Y. Huang, M. Bourghton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, “Power of data in quantum machine learning,”Nature Communications, vol. 12, no. 1, pp. 1–9, 2021

    2021

  18. [26]

    Generalization in quantum machine learning from few training data,

    M. C. Caro, H.-Y. Huang, M. Cerezo, K. Sharma, A. Sornborger, L. Cincio, and P. J. Coles, “Generalization in quantum machine learning from few training data,”Nature Communications, vol. 13, p. 4919, 2022

    2022

  19. [27]

    Out-of-distribution generalization for learning quantum dynamics,

    M. C. Caro, H.-Y. Huang, N. Ezzell, J. Gibbs, A. T. Sornborger, L. Cincio, P. J. Coles, and Z. Holmes, “Out-of-distribution generalization for learning quantum dynamics,” Nat. Commun., vol. 14, no. 1, 2023

    2023

  20. [28]

    Generalization despite overfitting in quantum machine learning models,

    E. Peters and M. Schuld, “Generalization despite overfitting in quantum machine learning models,” Quantum, vol. 7, p. 1210, 2023

    2023

  21. [29]

    Shadows of quantum machine learning,

    S. Jerbi, C. Gyurik, S. C. Marshall, R. Molteni, and V. Dunjko, “Shadows of quantum machine learning,” Nature Communications, vol. 15, p. 5676, 2024

    2024

  22. [30]

    Understanding quantum machine learning also requires rethinking generalization,

    E. Gil-Fuster, J. Eisert, and C. Bravo-Prieto, “Understanding quantum machine learning also requires rethinking generalization,” Nature Communications, vol. 15, p. 2277, 2024

    2024

  23. [31]

    Parameterized quantum circuits as machine learning models,

    M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, “Parameterized quantum circuits as machine learning models,” Quantum Sci. Technol., vol. 4, p. 043001, 2019

    2019

  24. [32]

    Supervised learning with quantum-enhanced feature spaces,

    V. Havl ´ıˇcek, A. D. C ´orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, “Supervised learning with quantum-enhanced feature spaces,”Nature, vol. 567, pp. 209–212, 2019

    2019

  25. [33]

    Quantum machine learning in feature Hilbert spaces,

    M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,” Physical Review Letters, vol. 122, p. 040504, 2019

    2019

  26. [34]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,” Nature Reviews Physics, vol. 3, pp. 625–644, 2021

    2021

  27. [35]

    Efficient and accurate estimation of Lipschitz constants for deep neural networks,

    M. Fazlyab, A. Robey, H. Hassani, M. Morari, and G. J. Pappas, “Efficient and accurate estimation of Lipschitz constants for deep neural networks,” in Advances in Neural Information Processing Systems , vol. 32, 2019. [Online]. Available: https://proceedings. neurips.cc/paper ...

    2019

  28. [36]

    A convex parameterization of robust recurrent neural networks,

    M. Revay, R. Wang, and I. R. Manchester, “A convex parameterization of robust recurrent neural networks,” IEEE Control Systems Lett., vol. 5, no. 4, pp. 1363–1368, 2021

    2021

  29. [37]

    Training robust neural networks using Lipschitz bounds,

    P. Pauli, A. Koch, J. Berberich, P. Kohler, and F. Allg ¨ower, “Training robust neural networks using Lipschitz bounds,” IEEE Control Systems Lett., vol. 6, pp. 121–126, 2022

    2022

  30. [38]

    Data re-uploading for a universal quantum classifier,

    A. P ´erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, “Data re-uploading for a universal quantum classifier,”Quantum, vol. 4, p. 226, 2020

    2020

  31. [39]

    Let quantum neural networks choose their own frequencies,

    B. Jaderberg, A. A. Gentile, Y. A. Berrada, E. Shishenina, and V. E. Elfving, “Let quantum neural networks choose their own frequencies,”Physical Review A, vol. 109, p. 042421, 2024

    2024

  32. [40]

    Stochastic gradient descent for hybrid quantum-classical optimization,

    R. Sweke, F. Wilde, J. Meyer, M. Schuld, P. K. F¨ahrmann, B. Meynard-Piganeau, and J. Eisert, “Stochastic gradient descent for hybrid quantum-classical optimization,” Quantum, vol. 4, p. 314, 2020

    2020

  33. [41]

    Robustness of quantum algorithms against coherent control errors,

    J. Berberich, D. Fink, and C. Holm, “Robustness of quantum algorithms against coherent control errors,” Physical Review A, vol. 109, p. 012417, 2024

    2024

  34. [42]

    Single-shot quantum machine learning,

    E. Recio-Armengol, J. Eisert, and J. J. Meyer, “Single-shot quantum machine learning,” Phys- ical Review A, vol. 111, p. 042420, 2025. 18 Julian Berberich, Tobias Fellner, Christian Holm

    2025

  35. [43]

    Double descent in quantum machine learning,

    M. Kempkes, A. Ijaz, E. Gil-Fuster, C. Bravo-Prieto, J. Spiegelberg, E. van Nieuwenburg, and V. Dunjko, “Double descent in quantum machine learning,”arXiv:2501.10077, 2025

    2025

  36. [44]

    Mohri, A

    M. Mohri, A. Rostamizadeh, and A. Talwalkar, Foundations of machine learning, 2nd ed., Adaptive Computation and Machine Learning . MIT Press, Cambridge, MA, 2018, iSBN: 9780262039406

    2018

  37. [45]

    Quantum circuit learning,

    K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, “Quantum circuit learning,” Physical Review A, vol. 98, p. 032309, 2018

    2018

  38. [46]

    Circuit-centric quantum classifiers,

    M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, “Circuit-centric quantum classifiers,” Physical Review A, vol. 101, p. 032308, 2020

    2020

  39. [47]

    RobustnessGeneralizationQML,

    T. Fellner, “RobustnessGeneralizationQML,” May 2025. [Online]. Available: https: //github.com/tobias-fllnr/RobustnessGeneralizationQML

    2025

  40. [48]

    Simple mathematical models with very complicated dynamics,

    R. M. May, “Simple mathematical models with very complicated dynamics,”Nature, vol. 261, no. 5560, pp. 459–467, Jun. 1976

    1976

  41. [49]

    Devaney, An Introduction To Chaotic Dynamical Systems, 2nd ed

    R. Devaney, An Introduction To Chaotic Dynamical Systems, 2nd ed. Boca Raton: CRC Press, Mar. 2018

    2018

  42. [50]

    Climate Predictions: The Chaos and Complexity in Climate Models,

    D. T. Mihailovi ´c, G. Mimi´c, and I. Arseni´c, “Climate Predictions: The Chaos and Complexity in Climate Models,” Advances in Meteorology, vol. 2014, no. 1, p. 878249, 2014

    2014

  43. [51]

    Chaos and nonlinear forecastability in economics and finance,

    B. LeBaron, B. T. Grenfell, R. M. May, and H. Tong, “Chaos and nonlinear forecastability in economics and finance,” Philosophical Transactions of the Royal Society of London. Series A: Physical and Engineering Sciences, vol. 348, no. 1688, pp. 397–404, Sep. 1994

    1994

  44. [52]

    Quantum vs. classical: A comprehensive benchmark study for predicting time series with variational quantum machine learning,

    T. Fellner, D. Kreplin, S. Tovey, and C. Holm, “Quantum vs. classical: A comprehensive benchmark study for predicting time series with variational quantum machine learning,” Apr. 2025

    2025

  45. [53]

    Pennylane: Automatic differentiation of hybrid quantum-classical computations,

    V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso- Linaje, B. AkashNarayanan, A. Asadi, J. M. Arrazola, U. Azad, S. Banning, C. Blank, T. R. Bromley, B. A. Cordier, J. Ceroni, A. Delgado, O. D. Matteo, A. Dusko, T. Garg, D. Guala, A. Hayes...

    2018 arXiv

  46. [54]

    Adam: a method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: a method for stochastic optimization,”arXiv:1412.6980, 2014

    2014 arXiv

  47. [55]

    A comparative analysis of adversarial robustness for quantum and classical machine learning,

    M. Wendlinger, K. Tscharke, and P. Debus, “A comparative analysis of adversarial robustness for quantum and classical machine learning,” in Proc. IEEE Int. Conf. Quantum Computing and Engineering (QCE), 2024, pp. 1447–1457

    2024

  48. [56]

    Robustness and generalization in quantum reinforcement learning via Lipschitz regularization,

    N. Meyer, J. Berberich, C. Mutschler, and D. D. Scherer, “Robustness and generalization in quantum reinforcement learning via Lipschitz regularization,”Quantum Machine Intelligence, 2024, submitted, available at https://arxiv.org/abs/2410.21117

    2024 arXiv

  49. [57]

    Adversarial quantum machine learning: an information-theoretic generalization analysis,

    P. Georgiou, S. T. Jose, and O. Simeone, “Adversarial quantum machine learning: an information-theoretic generalization analysis,” in Proc. IEEE Int. Symp. Information Theory (ISIT), 2024, pp. 789–794

    2024

  50. [58]

    On the generalization of adversarially trained quantum classifiers,

    P. Georgiou, A. M. Thomas, and S. T. Jose, “On the generalization of adversarially trained quantum classifiers,” arXiv:2504.17690, 2025

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.