Pith. sign in

REVIEW 5 major objections 5 minor 4 cited by

Fourier Fingerprints of Ansatzes in Quantum Machine Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Quantum models with exponentially many Fourier modes are forced to correlate them, and the correlation pattern—the Fourier fingerprint—predicts which ansatz learns best.

desk verdict New Fourier-correlation metric for ansatz selection, plausible on small benchmarks, but the random-init-to-trained transfer is unproven and App G doesn't hold up. read the letter →

arxiv 2508.20868 v1 pith:3XEADQG3 submitted 2025-08-28 quant-ph

Melvin Strobl , M. Emre Sahin , Lucas van der Horst , Eileen Kuehn , Achim Streit , Ben Jaderberg This is my paper
classification quant-ph MSC 81P68 PACS 03.67.Lx
keywords quantummachinelearningFouriermodelscoefficientcorrelationsfingerprintansatzselectionexpressibilityvariationalcircuitsjetreconstruction
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 sets out to show that an efficient quantum machine learning model is never fully flexible in the frequencies it learns. Because a typical data-encoding circuit yields exponentially many Fourier basis functions but only polynomially many trainable parameters, the Fourier coefficients must share parameters and therefore come correlated. The paper defines the Fourier fingerprint as the ansatz-specific pattern of these correlations, and the Fourier coefficient correlation (FCC) as its average. Numerically, lower FCC predicts lower mean squared error when the model is trained on random 1D and 2D Fourier series, and on a high-energy-physics jet reconstruction task, while the commonly used expressibility metric misorders the ansatzes. The authors also give an analytical argument that the expected MSE is linearly related to the weighted FCC.

What carries the argument

The central object is the Fourier fingerprint: the matrix of pairwise Pearson correlations r_Theta(omega, omega') between Fourier coefficients of the model, computed by drawing parameters uniformly from [0, 2pi), evaluating the model on a Nyquist-satisfying grid, extracting coefficients via FFT, and repeating over many random parameter samples. The scalar FCC is the average absolute correlation over all frequency pairs. It carries the argument because it quantifies how many independent 'knobs' the ansatz actually controls in frequency space: low FCC means coefficients can be tuned nearly independently, so the model can approximate arbitrary Fourier-series targets; high FCC means the optimize

What would settle it

Take a family of random 1D Fourier-series targets with independent coefficients and train each ansatz to convergence; if final MSE does not rank monotonically with FCC measured from uniform random initializations, the claimed predictive link fails. A sharper test: compute FCC from parameters sampled near the trained optima instead of uniform [0,2pi); if the ranking changes, the metric is not measuring what training respects.

Watch

Extended reading notes

Core claim

The central claim is that the constraint of efficient trainability shows up as correlations between Fourier coefficients, and that these correlations are structured, ansatz-specific, and predictive. For an n-qubit quantum Fourier model the number of Fourier terms grows as O(2^n) while the number of trainable parameters is only polynomial, so coefficients at different frequencies necessarily share parameters. Drawing many random parameter samples, extracting coefficients with a fast Fourier transform, and computing Pearson correlations between frequency pairs yields the fingerprint; FCC is the average absolute correlation over all pairs. Across eight ansatzes the paper finds that lower FCC is

Load-bearing premise

The FCC is computed from circuits whose parameters are drawn uniformly at random, and the paper assumes these random-parameter correlations match the correlations that survive after training; if they do not, the performance ranking could fail.

Editorial extensions

If this is right

  • Ansatz selection for QML can be done before training by computing FCC, which is cheaper than expressibility and, in these experiments, better ordered.
  • The value of expressibility as a performance predictor is called into question: a highly expressive ansatz can be a poor learner because its Fourier coefficients are highly correlated.
  • For datasets with uncorrelated frequency content, the design goal becomes finding efficiently implementable ansatzes with low FCC.
  • For datasets whose coefficients are themselves correlated, a deliberately high-FCC ansatz that matches the data correlations may be preferable; the paper frames FCC as an inductive bias, not an absolute quantity to minimize.
  • The FCC can be computed on quantum hardware from expectation values alone, whereas expressibility needs full state information, making FCC more practical.

Reading between the lines

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

  • Beyond the paper: the FCC-performance link is only demonstrated on Fourier-series-like targets; a natural extension is to test it on non-spectral datasets, where the relevant quantity may be the overlap between the ansatz's fingerprint and the dataset's own coefficient-correlation structure rather than the average.
  • Beyond the paper: because FCC averages over all frequency pairs equally, a frequency-weighted or band-limited version focused on the target's dominant frequencies might be a stronger predictor than the unweighted scalar; the paper's own weighting experiment points in this direction.
  • Beyond the paper: the metric's predictive power hinges on sampling theta uniformly; for ansatzes with trainability barriers, the relevant correlations may be those along the optimizer's trajectory, so a version of FCC computed from the training distribution could be tested against the random-initialization version.
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

5 major / 5 minor

Summary. The paper studies quantum Fourier models (QFMs) whose number of trainable parameters is polynomial in the qubit count while the number of Fourier basis functions is exponential. It argues that this imbalance forces correlations among Fourier coefficients, that the correlation pattern is ansatz-specific ('Fourier fingerprint'), and that the average coefficient correlation (FCC) predicts ansatz performance: models with lower FCC achieve lower mean squared error on random Fourier-series tasks and on a 2D HEP jet-reconstruction task, whereas the standard expressibility metric does not. The evidence is numerical: n=6, single-layer circuits, eight ansatzes, with FCC computed from independent uniform random parameter samples and performance obtained after gradient-based training. An appendix (App. G) attempts an analytic derivation linking FCC to MSE.

Significance. If the central claim held, the FCC would be a practical, cheaper alternative to expressibility for ansatz selection, and the Fourier fingerprint would be a useful diagnostic for QML circuits. The paper has genuine strengths: the fingerprint visualization is intuitive, the complexity comparison with expressibility is concrete, the experiments are falsifiable and independently seeded, and the authors are candid about outliers and limitations (notably App. I, where FCC fails for KL divergence). The main weakness is that the analytic derivation in App. G is not valid as written, and the empirical bridge from uniformly initialized parameters to trained-parameter performance is not established. The paper is therefore a promising proof-of-concept rather than a fully supported predictive theory.

major comments (5)
  1. [Sec. 2.1 / App. E] The parameter-count argument does not prove that correlations must exist. For example, with θ uniform on [0,2π), c_ω=sin θ and c_ω'=cos θ have zero Pearson correlation despite sharing a parameter. The statement that the imbalance 'implies the existence of correlations' needs a proof or should be softened to 'can induce'. App. E's proposition is not a formal argument; it merely states that shared basis functions 'can exhibit' correlations.
  2. [Sec. 2.2 vs Sec. 3] The FCC in Eq. (4) is evaluated with θ drawn i.i.d. uniformly from [0,2π), while the MSE comparisons in Secs. 3.1–3.2 use trained θ. Nothing shows that the correlation structure over the uniform prior is rank-preserving under gradient-based training. This is the load-bearing transfer step. Please test whether FCCs computed over the distribution of trained parameters (or an ensemble of training trajectories) still predict MSE, or otherwise justify the transfer.
  3. [App. G, Eqs. (12)–(14)] This analytic link is not valid. Eq. (13) omits complex conjugation in the covariance and writes σ^2(c_ω)=E[|c_ω c_ω'|], which is dimensionally and statistically wrong. More importantly, Eq. (12) replaces E[ζ ζ*] with rΘ(c_ω',c_ω'') σ σ without any derivation; the proxy coefficients ζ are unconstrained, so this assumes the conclusion. Additionally, the algebra in Eq. (14) is incorrect: Σ_ω [FCC|Ω|^{-1}+|ĉ_ω|^2] is not FCC Σ_ω|ĉ_ω|^2. This appendix cannot support the central claim in its current form.
  4. [Fig. 3b, Fig. 5b, App. I] The evidence is weaker than the abstract and conclusion claim. Circuit 15 is an acknowledged outlier in the 2D Fourier experiment, HEA and Circuit 17 are outliers in the HEP experiment, and App. I reports that FCC and expressibility both fail for KL divergence. With only eight ansatzes at n=6, L=1, and no statistical test (e.g., rank correlation with error bars), the claim that 'FCC correctly predicts relative performance' should be restricted or supplemented with quantitative evidence and a principled treatment of outliers.
  5. [Eq. (5)] As written, FCC includes the trivial diagonal r(ω,ω)=1, which would make every FCC at least 1/|Ω| ≈ 0.14; the reported values (e.g., 4.57×10^{-3}) exclude it. Please define precisely whether the sum is over ω≠ω' and whether symmetric pairs are counted once. This is essential for reproducibility and for the numerical values reported throughout.
minor comments (5)
  1. [App. F] Tables 2 and 3 contain corrupted numerical formatting (e.g., '0.0 2 .2 ×10−5'); please fix the typesetting and verify all entries.
  2. [App. D] The weighting function w(ω,ω') is not normalized; clarify whether the weighted FCC is normalized and whether values from different ansatzes remain comparable.
  3. [App. H] The claim that FCC is 'easier to compute regardless of D' overlooks the K ≥ (2 max Ω)^D sampling requirement, which is exponential in D. Please qualify the complexity comparison.
  4. [App. C / App. G] Notation is inconsistent: σ is used for both standard deviation and variance, and Eq. 13 defines σ^2(c_ω) via a cross-frequency expression. Please clean up the statistics notation.
  5. [Fig. 2b] The label 'ci,*' for the 2D fingerprint should be defined explicitly; currently it is easy to misinterpret.

Circularity Check

1 steps flagged · score 4.0 of 10

Empirical FCC–MSE trend is independent, but the analytic link in App. G is circular by construction.

  1. self definitional [Appendix G, Eqs. 12–14]
    "we can capture this in L as well, by assuming that each coefficient is actually not dependent on a single frequency, but on all other frequencies in Ω instead... substituting cω(θ) = Σω′ ζ(ω, ω′; θ)... (2) = Σω [ Σω′,ω′′ rΘ(cω′, cω′′)σΘ(cω′)σΘ(cω′′) + |ĉω|2 ] ... Following the weighting strategy introduced in App. D, we can set w(ω′,ω′′) = σΘ(cω′)σΘ(cω′′) and therefore forcefully restating the final loss equation ... EΘ[L(θ)] = ... FCCΘ Σω |ĉω|2 ."

    The proxy coefficient ζ is introduced ad hoc, and its cross-moments are never derived from the model. The step from E[ζ ζ*] to rΘ(cω′, cω′′)σΘ(cω′)σΘ(cω′′) simply identifies the proxy covariance with the FCC correlation matrix—the exact quantity the appendix claims to show causes the loss. Equation 14 then sets the weight to σΘσΘ, so E[L] = FCC·Σ|ĉ|² holds by definition of ζ and w, not by derivation. The word 'forcefully restating' acknowledges the construction. This is a circular analytic derivation, though it is appendix-level support; the main numerical FCC→MSE trend is measured independently and is not tautological.

full rationale

The paper's central empirical claim—that ansatzes with lower FCC trained from random initialization achieve lower MSE on random Fourier series and on the HEP task—is not circular: the FCC is computed from uniformly sampled parameters before training, while the MSE is measured after gradient-based training on independent data seeds, so the relationship is genuinely predictive rather than fitted. The expressibility comparison is an external benchmark, and the self-citations (e.g., QML-Essentials for the FFT/expressibility implementation) are implementation details, not load-bearing arguments. The one significant circular step is in Appendix G, where the claimed analytic relation between the loss and the FCC is obtained by substituting a proxy coefficient ζ and then replacing its covariance with the FCC correlations without derivation, effectively assuming the conclusion. Because the main fingerprint claim does not depend on this appendix, the overall circularity burden is moderate rather than fundamental. The weakest assumption highlighted in the reader's take—that random-initialization correlations survive training—is an empirical validity concern, not a circularity, and does not by itself raise the score.

Assumptions & free parameters 2 free parameters · 9 assumptions · 0 invented entities

The central numerical claim rests on standard QFM assumptions plus an empirical transfer from random-parameter correlations to trained performance. The only genuinely ad hoc construction is the proxy-coefficient expansion in App G that converts MSE into FCC. No new physical entities are introduced.

free parameters (2)
  • weighting function w in weighted FCC = w(omega, omega') ~ (omega + omega')^-1 in App D; w(omega', omega'') = sigma_Theta(c_omega') sigma_Theta(c_omega'') in Ap
    Chosen by hand to modify the FCC. The analytical MSE-FCC relation in Eq 14 only holds after this weighting is imposed, so it is an ad hoc adjustment rather than a derived quantity.
  • HEP loss weight 0.001 for KLDiv = 0.001
    The final HEP loss is L = MSE + 0.001 * KLDiv; the weight was set empirically to improve convergence. It affects the reported physics results but not the toy Fourier series conclusions.
assumptions (9)
  • standard math The output of the QFM equals a truncated Fourier series (Eq 1 from Ref [28]).
    Foundation of the analysis; widely accepted in the QML Fourier literature.
  • domain assumption Efficiently trainable QFMs must have |theta| = O(poly(n)), with training cost linear in |theta|.
    Used in Sec 2.1 to argue that many coefficients must share parameters.
  • domain assumption Shared parameters imply Pearson correlations between Fourier coefficients over the random parameter distribution.
    The paper's central inference; it is demonstrated numerically and hedged in Sec 4, not proven for the feature maps used.
  • domain assumption Ansatz parameters are drawn i.i.d. uniformly in [0, 2 pi) and coefficient means are zero.
    Used in Eq 4 and App G. The zero mean is reported as a numerical observation in App C.
  • ad hoc to paper Target Fourier series frequencies exactly match QFM frequencies, and target coefficients are independent and uniform within the unit circle.
    Experimental design in Sec 3.1 that makes the learning task ideal; not true for general data.
  • domain assumption The HEP task can be represented by a Fourier model, and the leading pT depends mainly on ECM and E_Delta after quantile transforms.
    Feature selection in Sec 3.2; physically motivated but not proven, and it defines the simplified benchmark.
  • domain assumption The frequencies of the dataset are independent, so low FCC should be helpful.
    Explicitly stated in Sec 4 as an underlying assumption of the heuristic.
  • ad hoc to paper In App G, c_omega(theta) can be written as sum over zeta(omega, omega'; theta), with E[zeta(omega, omega'; theta) zeta*(omega, omega''; theta)] equal to r_Theta(c_omega', c_omega'') sigma_Theta(c_omega') sigma_Theta(c_omega'').
    Unjustified substitution used to derive E[L] proportional to FCC; it is an assumed equality, not a proven identity.
  • standard math Parseval and orthonormality of exponentials on X = [0, 2 pi) for the MSE expansion.
    Standard, but explicitly assumes the input interval and ignores boundary effects.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fourier Fingerprints of Ansatzes in Quantum Machine Learning." pith.science (2026). https://pith.science/paper/3XEADQG3

@misc{pith2026250820868,
  author       = {Pith},
  title        = {Pith review of: Fourier Fingerprints of Ansatzes in Quantum Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XEADQG3}},
  note         = {Machine review of arXiv:2508.20868}
}
abstract

Typical schemes to encode classical data in variational quantum machine learning (QML) lead to quantum Fourier models with $\mathcal{O}(\exp(n))$ Fourier basis functions in the number of qubits. Despite this, in order for the model to be efficiently trainable, the number of parameters must scale as $\mathcal{O}(\mathrm{poly}(n))$. This imbalance implies the existence of correlations between the Fourier modes, which depend on the structure of the circuit. In this work, we demonstrate that this phenomenon exists and show cases where these correlations can be used to predict ansatz performance. For several popular ansatzes, we numerically compute the Fourier coefficient correlations (FCCs) and construct the Fourier fingerprint, a visual representation of the correlation structure. We subsequently show how, for the problem of learning random Fourier series, the FCC correctly predicts relative performance of ansatzes whilst the widely-used expressibility metric does not. Finally, we demonstrate how our framework applies to the more challenging problem of jet reconstruction in high-energy physics. Overall, our results demonstrate how the Fourier fingerprint is a powerful new tool in the problem of optimal ansatz choice for QML.

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DAGAF: A directed acyclic generative adversarial framework for joint structure learning and tabular data synthesis

    cs.LG 2026-04 conditional novelty 7.0 of 10

    A data-agnostic circuit harmonic matrix C factorises Fourier-coefficient statistics and quantum neural tangent kernels for a broad class of re-uploading parametrised quantum circuits.

  2. Quantum entanglement provides a competitive advantage in adversarial games

    quant-ph 2026-03 conditional novelty 6.0 of 10

    Entangled 8-qubit PQC feature extractors in PPO agents for Pong consistently beat separable PQCs of similar size and can match or exceed small classical MLPs in the low-parameter regime.

  3. Inherent interpretability provides inherent value in quantum machine learning

    quant-ph 2026-07 conditional novelty 4.0 of 10

    Quantum machine learning models can be valuable for their inherent mathematical interpretability, and quantum Fourier models offer a bottom-up route to Gaussian-process kernel design that random Fourier features do no...

  4. Hybrid Quantum Neural Networks: Theory, Implementations, and Applications

    quant-ph 2026-08 conditional novelty 2.0 of 10

    A balanced review of hybrid quantum neural networks, concluding that quantum layers help on structured, small-scale and quantum-native problems but do not yet beat classical models on generic benchmarks.

Reference graph

Works this paper leans on

36 extracted references · 16 canonical work pages · cited by 4 Pith papers

  1. [1]

    Parameterized quantum circuits as machine learning models

    Marcello Benedetti, Erika Lloyd, Stefan Sack, et al. “Parameterized quantum circuits as machine learning models”. In:Quantum science and technology4.4 (2019), p. 043001

    2019

  2. [2]

    The Standard Deviation of the Correlation Coefficient

    A. L. Bowley. “The Standard Deviation of the Correlation Coefficient”. In:Journal of the American Statistical Association23.161 (1928), pp. 31–34.doi: 10.1080/01621459.1928.10502991

    work page arXiv 1928

  3. [3]

    Fitting the multitemporal curve: A Fourier series approach to the missing data problem in remote sensing analysis

    Evan B Brooks, Valerie A Thomas, Randolph H Wynne, et al. “Fitting the multitemporal curve: A Fourier series approach to the missing data problem in remote sensing analysis”. In:IEEE Transactions on Geoscience and Remote Sensing50.9 (2012), pp. 3340–3353

    work page 2012

  4. [4]

    Challenges and opportunities in quantum machine learning

    Marco Cerezo, Guillaume Verdon, Hsin-Yuan Huang, et al. “Challenges and opportunities in quantum machine learning”. In:Nature computational science2.9 (2022), pp. 567–576

    work page 2022

  5. [5]

    Quantum ensemble learning with a programmable superconducting processor

    Jiachen Chen, Yaozu Wu, Zhen Yang, et al. “Quantum ensemble learning with a programmable superconducting processor”. In:npj Quantum Information11.1 (2025), p. 83

    work page 2025

  6. [6]

    An Algorithm for the Machine Calculation of Complex Fourier Series

    James W. Cooley and John W. Tukey. “An Algorithm for the Machine Calculation of Complex Fourier Series”. In:Mathematics of Computation19.90 (1965), pp. 297–301.doi: 10.1090/S0025- 5718-1965-0178586-1

    doi:10.1090/s0025- 1965

  7. [7]

    Spectral Bias in Variational Quantum Machine Learning

    Callum Duffy and Marcin Jastrzebski. “Spectral Bias in Variational Quantum Machine Learning”. In: arXiv preprint arXiv:2506.22555(2025)

    arXiv 2025

  8. [8]

    Maja Franz, Melvin Strobl, Leonid Chaichenets, et al.Out of Tune: Demystifying Noise-Effects on Quantum Fourier Models. 2025. doi: 10.48550/arxiv.2506.09527

    arXiv doi:10.48550/arxiv.2506.09527 2025

Show all 36 references
  1. [9]

    Statistics (international student edition)

    David Freedman, Robert Pisani, and Roger Purves. “Statistics (international student edition)”. In: Pisani, R. Purves, 4th edn. WW Norton & Company, New York(2007)

    2007

  2. [10]

    Quantum machine learning in high energy physics

    Wen Guan, Gabriel Perdue, Arthur Pesah, et al. “Quantum machine learning in high energy physics”.en.In: Machine Learning: Science and Technology2.1(2021).Publisher:IOPPublishing, p. 011003. doi: 10.1088/2632-2153/abc17d

    2021 doi

  3. [11]

    Robust Fourier and polynomial curve fitting

    Venkatesan Guruswami and David Zuckerman. “Robust Fourier and polynomial curve fitting”. In: 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS). IEEE. 2016, pp. 751–759

    2016

  4. [12]

    Zenodo, June 2025.doi: 10.5281/zenodo.15689696

    Yacine Haddad and Michele Grossi.Parton-to-Jet Mapping proton-proton dataset at 14 TeV. Zenodo, June 2025.doi: 10.5281/zenodo.15689696

    2025 doi

  5. [13]

    Supervised learning with quantum- enhanced feature spaces

    Vojtěch Havlíček, Antonio D Córcoles, Kristan Temme, et al. “Supervised learning with quantum- enhanced feature spaces”. In:Nature 567.7747 (2019), pp. 209–212

    2019

  6. [14]

    Potential of quantum scientific machine learning applied to weather modeling

    Ben Jaderberg, Antonio A Gentile, Atiyo Ghosh, et al. “Potential of quantum scientific machine learning applied to weather modeling”. In:Physical Review A110.5 (2024), p. 052423

    2024

  7. [15]

    Gentile, Youssef Achari Berrada, et al.Let Quantum Neural Networks Choose Their Own Frequencies

    Ben Jaderberg, Antonio A. Gentile, Youssef Achari Berrada, et al.Let Quantum Neural Networks Choose Their Own Frequencies. Tech. rep. arXiv, Sept. 2023.doi: 10.48550/arXiv.2309.03279

    2023 doi arXiv

  8. [16]

    Kingma and Jimmy Ba.Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba.Adam: A Method for Stochastic Optimization. 2017. doi: 10.48550/arXiv.1412.6980

    2017 doi arXiv

  9. [17]

    On Information and Sufficiency

    S. Kullback and R. A. Leibler. “On Information and Sufficiency”. In:The Annals of Mathematical Statistics 22.1 (Mar. 1951). Publisher: Institute of Mathematical Statistics, pp. 79–86.doi: 10.1214/aoms/1177729694

    1951

  10. [18]

    Protocols for trainable and differentiable quantum generative modeling

    Oleksandr Kyriienko, Annie E Paine, and Vincent E Elfving. “Protocols for trainable and differentiable quantum generative modeling”. In:Physical Review Research6.3 (2024), p. 033291

    2024

  11. [19]

    Quantum embeddings for machine learning

    Seth Lloyd, Maria Schuld, Aroosa Ijaz, et al. “Quantum embeddings for machine learning”. In: arXiv preprint arXiv:2001.03622(2020)

    2001 arXiv

  12. [20]

    Hela Mhiri, Leo Monbroussou, Mario Herrero-Gonzalez, et al.Constrained and Vanishing Expressivity of Quantum Fourier Models. Mar. 2024.doi: 10.48550/arXiv.2403.09417

    2024 doi arXiv

  13. [21]

    Quantum circuit learning

    Kosuke Mitarai, Makoto Negoro, Masahiro Kitagawa, et al. “Quantum circuit learning”. In: Physical Review A98.3 (2018), p. 032309

    2018

  14. [22]

    Quantum data learning for quantum simulations in high-energy physics

    Lento Nagano, Alexander Miessen, Tamiya Onodera, et al. “Quantum data learning for quantum simulations in high-energy physics”. In:Physical Review Research5.4 (2023), p. 043250. 10

    2023

  15. [23]

    Fourier expansion in variational quantum algorithms

    Nikita A. Nemkov, Evgeniy O. Kiktenko, and Aleksey K. Fedorov. “Fourier expansion in variational quantum algorithms”. In:Physical Review A108.3 (Sept. 2023), p. 032406.doi: 10.1103/PhysRevA.108.032406

    2023 doi

  16. [24]

    Experimental quantum end-to-end learning on a superconducting processor

    Xiaoxuan Pan, Xi Cao, Weiting Wang, et al. “Experimental quantum end-to-end learning on a superconducting processor”. In:npj Quantum Information9.1 (2023), p. 18

    2023

  17. [25]

    Evan Peters and Maria Schuld.Generalization despite overfitting in quantum machine learning models. Tech. rep. arXiv, Sept. 2022.doi: 10.48550/arXiv.2209.05523

    2022 doi arXiv

  18. [26]

    A Lie Algebraic Theory of Barren Plateaus for Deep Parameterized Quantum Circuits

    Michael Ragone, Bojko N. Bakalov, Frédéric Sauvage, et al. “A Lie Algebraic Theory of Barren Plateaus for Deep Parameterized Quantum Circuits”. In:Nature Communications15.1 (Aug. 2024), p. 7172.doi: 10.1038/s41467-024-49909-3

    2024 doi

  19. [27]

    Circuit-centric quantum classifiers

    Maria Schuld, Alex Bocharov, Krysta Svore, et al. “Circuit-centric quantum classifiers”. In: Physical Review A101.3 (Mar. 2020), p. 032308.doi: 10.1103/PhysRevA.101.032308

    2020 doi

  20. [28]

    The effect of data encoding on the expressive power of variational quantum machine learning models

    Maria Schuld, Ryan Sweke, and Johannes Jakob Meyer. “The effect of data encoding on the expressive power of variational quantum machine learning models”. In:Physical Review A103.3 (Mar. 2021), p. 032430.doi: 10.1103/PhysRevA.103.032430

    2021 doi

  21. [29]

    Particle track reconstruction with noisy intermediate-scale quantum computers

    Tim Schwägerl, Cigdem Issever, Karl Jansen, et al. “Particle track reconstruction with noisy intermediate-scale quantum computers”. In: (2023).doi: 10.48550/arXiv.2303.13249

    2023 doi arXiv

  22. [30]

    Communication in the Presence of Noise

    C.E. Shannon. “Communication in the Presence of Noise”. In:Proceedings of the IRE37.1 (1949), pp. 10–21.doi: 10.1109/JRPROC.1949.232969

    1949

  23. [31]

    Exponential data encoding for quantum supervised learning

    Seongwook Shin, Yong-Siah Teo, and Hyunseok Jeong. “Exponential data encoding for quantum supervised learning”. In:Physical Review A107.1 (2023), p. 012422

    2023

  24. [32]

    Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum-Classical Algorithms

    Sukin Sim, Peter D. Johnson, and Alán Aspuru-Guzik. “Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum-Classical Algorithms”. en. In:Advanced Quantum Technologies2.12 (2019), p. 1900070.doi: 10.1002/qute.201900070

    2019 doi

  25. [33]

    Melvin Strobl, Maja Franz, Eileen Kuehn, et al.QML Essentials – A framework for working with Quantum Fourier Models. 2025. doi: 10.48550/arXiv.2506.06695

    2025 doi

  26. [34]

    Potential and limitations of random Fourier features for dequantizing quantum machine learning

    Ryan Sweke, Erik Recio-Armengol, Sofiene Jerbi, et al. “Potential and limitations of random Fourier features for dequantizing quantum machine learning”. In:Quantum 9 (Feb. 2025), p. 1640. doi: 10.22331/q-2025-02-20-1640

    2025 doi

  27. [35]

    Particle Track Reconstruction with Quantum Algorithms

    Cenk Tüysüz, Federico Carminati, Bilge Demirköz, et al. “Particle Track Reconstruction with Quantum Algorithms”. In:EPJ Web of Conferences245 (2020). Ed. by C. Doglioni, D. Kim, G.A. Stewart, et al., p. 09013.doi: 10.1051/epjconf/202024509013

    2020

  28. [36]

    Scherer.Fourier Analysis of Vari- ational Quantum Circuits for Supervised Learning

    Marco Wiedmann, Maniraman Periyasamy, and Daniel D. Scherer.Fourier Analysis of Vari- ational Quantum Circuits for Supervised Learning. Nov. 2024. doi: 10.48550/arXiv.2411. 03450. 11 A Ansatzes Fig. 6 shows the ansatzes used in this work. Specifically, we depict only a single ...

    2024 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.