Pith. sign in

REVIEW 100 references

Q-FLAIR learns quantum feature-maps by reconstructing the model output as a sine curve from three measurements, making gate, feature, and weight selection classical.

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

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

By classically reconstructing quantum model outputs, Q-FLAIR selects gates, features, and weights with O(M) quantum evaluations per iteration, decoupling quantum cost from feature dimension and enabling >90% MNIST accuracy on IBM hardware.

T0 review reviewed 2026-08-04 challenge →

arxiv 2510.03389 v2 pith:U7IP7ASZ submitted 2025-10-03 quant-ph cs.LG

Quantum feature-map learning with reduced resource overhead

Jonas J\"ager , Philipp Els\"asser , Elham Torabian This is my paper
classification quant-ph cs.LG MSC 81P6868Q1268T05
keywords quantum feature mapsQ-FLAIRanalytic reconstructionadaptive ansatzquantum machine learningfeature selectionvariational quantum classifierquantum kernels
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper introduces Q-FLAIR, an algorithm that grows a quantum feature-map circuit gate by gate while offloading the expensive choices to a classical computer. For any candidate gate that is a rotation with generator A²=I, the expectation value of the model as a function of the rotation angle is exactly a sinusoid f(α)=a cos(α−b)+c. Three expectation-value measurements determine that curve, after which feature selection and weight optimization for that gate are performed entirely classically. This reduces the quantum cost per iteration from a product of gate-pool size, feature dimension, and loss evaluations to a constant number of evaluations per gate candidate. The claimed consequence is that high-dimensional data becomes trainable on current hardware: the paper reports learning a feature map for full-resolution 784-pixel MNIST (digits 3 vs 5) on an IBM device in four hours, surpassing 90% accuracy.

Core claim

The central discovery is that iterative quantum feature-map construction does not need to query the quantum computer for every feature–weight combination. Because appending a rotation gate with generator A²=I to a fixed circuit turns the expectation value of any observable into a sinusoid in the rotation angle, three circuit evaluations recover the entire model-output curve. The algorithm then scans the gate pool once per iteration, reconstructs each candidate's curve, and classically solves the inner optimization over the weight θ and the outer selection over the feature index k. For weight-data-dependent gates the substitution α=θ x_k makes feature selection classical, so the quantum resou

What carries the argument

The load-bearing object is the analytic reconstruction of the expectation value as f(α)=a cos(α−b)+c, obtained from three estimates at α₀, α₀+π/2, and α₀−π/2. Because rotation gates with A²=I are 2π-periodic and produce exactly this sine form, the measured curve can be extrapolated to any rotation angle. In Q-FLAIR the reconstruction is evaluated once per gate candidate per data point, and then the loss under that candidate is evaluated classically; the same curve is reused for all features by substituting α=θ x_k. This turns the combined search over gates, features, and weights into a one-dimensional classical optimization per gate, which is the mechanism that reduces the quantum evaluation

Load-bearing premise

The method works only if every added gate is a rotation with generator satisfying A²=I, so the output is exactly sinusoidal, and if the three-point reconstruction formulas in Appendix A1 are implemented with a phase convention that matches that sine form — the printed version fails a check on f(α)=cos(α).

What would settle it

Evaluate a single-qubit circuit with the observable |0⟩⟨0| and gate R_x(α), measure at α=0, π/2, −π/2, apply the paper's Eqs. (A1)-(A3) verbatim, and compare the reconstructed curve to the known value f(0)=1. If the printed formulas give f̂(0)≠1, then those formulas are not the ones used to produce the reported results; an implementation that reproduces the reported accuracies must be using a different, corrected reconstruction, and that correction should be published.

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

If this is right

  • Quantum circuit evaluation count per iteration drops from O(M d T_max) to O(M), making high-dimensional inputs up to 784 features trainable on near-term hardware.
  • Feature selection is integrated into the loss landscape, so raw high-dimensional data can be used without PCA pre-processing; the paper reports above 90% accuracy on MNIST 3 vs 5 and 99% on MNIST 0 vs 1 in simulation with a single data-dependent gate.
  • Learned circuits are shallow and topology-aware, which reduces the impact of hardware noise and keeps total runtime around four hours for a full MNIST experiment.
  • The same reconstruction trick applies to both QNN outputs and quantum-kernel entries under suitable gate-pool restrictions, so the method is not tied to a single classifier family.
  • Greedy gate-by-gate growth starts from a small circuit and adds resources only when loss improves, which the paper argues can delay barren-plateau concentration compared with fixed wide ansätze.

Where Pith is reading between the lines

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

  • The three-point sine reconstruction is a generic interpolation device: any variational algorithm whose candidate operations are rotations with square-identity generators could adopt the same classical offload, not just classifiers.
  • Once the feature-dependent search cost is removed, the practical bottleneck shifts to the number of data points and circuit depth; scaling to larger datasets will likely hinge on stochastic loss estimates or kernel-matrix approximations rather than on qubit count.
  • The paper's own proposed extension to K-th order Fourier reconstructions would allow generators with A²≠I; even at 2K+1 evaluations per gate, the cost would remain independent of d, so the resource decoupling would survive.
  • The formulas printed in Appendix A1 fail a trivial consistency check on f(α)=cos α, so the implementation must rely on a corrected phase convention; a published correction or independent re-derivation is needed before the reported experiments can be replicated from the paper alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Circularity Check

0 steps flagged

No circularity found; the central reconstruction is exact interpolation and reported accuracies are held-out test results.

full rationale

I find no significant circularity in the paper's derivation chain. The core resource-saving mechanism is the analytic reconstruction f(α)=a cos(α−b)+c of an expectation value with respect to a rotation angle α for generators satisfying A²=I. This is an exact functional identity for such gates, not a fitted model or a renamed empirical pattern. The algorithm then substitutes α=θ·x_k into this reconstructed function to classically evaluate candidate features and weights, so the O(MdT_max)→O(M) reduction in quantum circuit evaluations per iteration follows from a genuine mathematical decoupling, not from defining the reconstruction in terms of the target loss. Feature and gate selection are performed using training loss, while the reported benchmark accuracies are held-out test accuracies, so there is no fitted-input-called-prediction pattern. The self-citations (Refs. 11, 27, 41) are background context or related heuristics for adaptive ansaetze and are not load-bearing for the paper's principal claims. One correctness concern, which is not circularity, should be flagged: the printed reconstruction formulas (A1)-(A3) appear algebraically inconsistent (e.g., applied to f(α)=cos α at α0=0 they return a=−1, b=0, c=0, giving f̂=−cos α rather than cos α). This suggests a sign/phase typo in the appendix, but a wrong formula is an error, not a circular argument; the structural reasoning would be unchanged under the corrected convention. Overall, the derivation is self-contained and the central claims do not reduce to their inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical entities or fitted global constants. The central derivation is mathematical; however the printed reconstruction formulas contain a sign/phase error (Appendix A1), so a correct implementation depends on an unprinted correction. Hyperparameters listed above are hand-picked, not fitted to reproduce the target accuracy.

free parameters (4)
  • Convergence threshold ΔL = 10^-3
    Stopping criterion for QNN log-loss and QSVM target-alignment; chosen by the authors, not derived; controls circuit depth.
  • Measurement shots = 1024
    Used for all IBM hardware expectation/kernel estimates; finite-shot noise affects reconstructions but no shot-noise error bars are reported.
  • Mini-batch sizes = 64/200 (QNN), 32 (QSVM)
    Hardware runs; reduce quantum circuit evaluations at cost of loss-estimate variance; the QSVM mini-batching is later identified as a cause of underperformance.
  • QSVM regularization C = not reported
    Grid-searched classically; standard SVM hyperparameter, not a fitted physical parameter.
axioms (5)
  • standard math For any Hermitian A with A²=I, the expectation value of R(α)MR(α)† is a sinusoid a cos(α−b)+c (Eq. (13)-(14)).
    Mathematical property of rotation gates; used everywhere.
  • domain assumption Candidate gates are only appended at the end of the feature-map circuit, so the entire preceding circuit acts as a fixed state/observable and the sinusoid form is preserved.
    Sec. III A/III B; this restricts gate placement and excludes gates commuting with the observable for QNN.
  • domain assumption Features are pre-normalized to [-π,π] and weights to [-1,1], with rescaling inferred from the training data.
    Sec. III; needed for the substitution α=θ x_k to stay within gate periodicity.
  • domain assumption For QSVM, the fidelity kernel for pairs (xi,xi′) is expressible as the expectation value in Eq. (23)-(25) with M = |ψ(xi′)⟩⟨ψ(xi′)|, so the sinusoid reconstruction applies to kernels.
    Sec. III B; also requires data-dependent gates to avoid the gate erasure bug.
  • domain assumption Finite-shot estimates at the three reconstruction points are sufficient to recover the sinusoid coefficients accurately enough for greedy selection.
    Appendix A1/B2; no noise analysis of reconstruction errors is given.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum feature-map learning with reduced resource overhead." pith.science (2026). https://pith.science/paper/U7IP7ASZ

@misc{pith2026251003389,
  author       = {Pith},
  title        = {Pith review of: Quantum feature-map learning with reduced resource overhead},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7IP7ASZ}},
  note         = {Machine review of arXiv:2510.03389}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Current quantum computers require algorithms that use limited resources economically. In quantum machine learning, success hinges on quantum feature-maps, which embed classical data into the state space of qubits. We introduce Quantum Feature-Map Learning via Analytic Iterative Reconstructions (Q-FLAIR), an algorithm that reduces quantum resource overhead in iterative feature-map circuit construction. It shifts workloads to a classical computer via partial analytic reconstructions of the quantum model, using only a few evaluations. For each probed gate addition to the ansatz, the simultaneous selection and optimization of the data feature and weight parameter is then entirely classical. Integrated into quantum neural network and quantum kernel support vector classifiers, Q-FLAIR shows state-of-the-art benchmark performance. Since resource overhead decouples from feature dimension, we train a quantum model on a real IBM device in only four hours, surpassing 90% accuracy on the full-resolution MNIST dataset (784 features, digits 3 vs 5). Such results were previously unattainable, as the feature dimension prohibitively drives hardware demands for fixed and search costs for adaptive ans\"atze. Furthermore, Q-FLAIR demonstrates de-quantization robustness against direct classical modeling, satisfying a benchmark rare in the literature and a necessary condition for potential quantum advantage. By rethinking feature-map learning beyond black-box optimization, this work takes a concrete step toward enabling quantum machine learning for real-world problems and near-term quantum computers.

Figures

Figures reproduced from arXiv: 2510.03389 by Elham Torabian, Jonas J\"ager, Philipp Els\"asser.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic comparison of traditional quantum feature-map learning and Q-FLAIR. The figure illustrates the distribution of computa [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. QNN performance benchmark on four different datasets. Top panels: accuracy over the number of gates appended to the feature-map [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. QSVM performance benchmark on four different datasets. Top panels: accuracy over the number of gates appended to the feature-map [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Ablation study of Q-FLAIR optimizations for QNNs on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. NISQ benchmark of QNN performance on four datasets, including full-resolution [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. NISQ benchmark of QSVM performance on four datasets, including full-resolution [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Comparison between different post-optimization methods. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Minimal number of gates needed to reach an average accu [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

100 extracted references · 17 linked inside Pith

  1. [1]

    Reconstruction By obtaining estimates from the quantum computer for the weight parameter in parameter dependent gates at the three pointsθ j ∈ {α0,α 0 +π/2,α 0 −π/2 } yieldingz 0 =f(α 0), z+ =f(α 0 +π/2)andz − =f(α 0 −π/2), the analytic recon- struction of Eq. (14) can be determined by a=− p (2z0 −z + −z −)2 + (z+ −z −)2 2 (A1) b=−arctan2(2z 0 −z + −z −,z...

  2. [2]

    (5) is used for the empirical loss in Eq

    Hyperparameter settings For the training of the QNN, the logarithmic loss function defined in Eq. (5) is used for the empirical loss in Eq. (8). The stopping criterion is set to∆Llog <10 −3. In order to determine what threshold parameterbon the QNN output is required for the decision function as defined in Eq. (4), we select the best value from the receiv...

  3. [3]

    These measurements are im- plemented in the IBM hardware without basis transforma- tions

    Specifics for IBM benchmarks For both models, QNN and QSVM, for each quantum cir- cuit evaluated for a fixed parameter binding, the all-zero state 13 probability was estimated through computational basis mea- surements using 1024 shots. These measurements are im- plemented in the IBM hardware without basis transforma- tions. This probability directly corr...

  4. [4]

    This leads to local optima

    Effect of parameter re-optimization The number of evaluations in Q-FLAIR is kept low, because every parameter gets optimized only once. This leads to local optima. By re-optimizing the parameters, in theory better re- sults should be achievable. The re-optimization of parameters is done for the QNN of theMNIST PCAdataset. With the limited-memory Broyden-F...

  5. [5]

    While this is an important quantity, it is easier to compare the per- formance of different models if the minimal circuit size for a given accuracy is considered

    Accuracy comparison QSVM and QNN In section IV A, we concentrated mostly on the maximally achievable accuracy prior to the saturation of the loss. While this is an important quantity, it is easier to compare the per- formance of different models if the minimal circuit size for a given accuracy is considered. Such a comparison for the QSVMs and QNNs oftwo-...

  6. [6]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum machine learning, Nature549, 195 (2017)

    2017

  7. [7]

    Dunjko and H

    V . Dunjko and H. J. Briegel, Machine learning & artificial in- telligence in the quantum domain: a review of recent progress, Reports on Progress in Physics81, 074001 (2018)

    2018

  8. [8]

    Cerezo, G

    M. Cerezo, G. Verdon, H.-Y . Huang, L. Cincio, and P. J. Coles, Challenges and opportunities in quantum machine learning, Na- ture Computational Science2, 567 (2022)

    2022

  9. [9]

    Havlíˇcek, A

    V . Havlíˇcek, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kan- dala, J. M. Chow, and J. M. Gambetta, Supervised learning with quantum-enhanced feature spaces, Nature567, 209 (2019)

    2019

  10. [10]

    Schuld and N

    M. Schuld and N. Killoran, Quantum Machine Learning in Feature Hilbert Spaces, Physical Review Letters122, 040504 (2019)

    2019

  11. [11]

    Huang, Z.-A

    K. Huang, Z.-A. Wang, C. Song, K. Xu, H. Li, Z. Wang, Q. Guo, Z. Song, Z.-B. Liu, D. Zheng,et al., Quantum genera- tive adversarial networks with multiple superconducting qubits, npj Quantum Information7, 165 (2021)

    2021

  12. [12]

    Sakka, K

    K. Sakka, K. Mitarai, and K. Fujii, Automating quantum fea- ture map design via large language models, arXiv preprint arXiv:2504.07396 (2025)

    arXiv 2025

  13. [13]

    Ristè, M

    D. Ristè, M. P. Da Silva, C. A. Ryan, A. W. Cross, A. D. Córcoles, J. A. Smolin, J. M. Gambetta, J. M. Chow, and B. R. Johnson, Demonstration of quantum advantage in ma- chine learning, npj Quantum Information3, 16 (2017)

    2017

  14. [14]

    Cho and D

    G. Cho and D. Kim, Machine learning on quantum experimen- tal data toward solving quantum many-body problems, Nature Communications15, 7552 (2024)

    2024

  15. [15]

    Y . Liu, S. Arunachalam, and K. Temme, A rigorous and ro- bust quantum speed-up in supervised machine learning, Nature Physics17, 1013 (2021)

    2021

  16. [16]

    Jäger and R

    J. Jäger and R. V . Krems, Universal expressiveness of varia- tional quantum classifiers and quantum kernels for support vec- tor machines, Nature Communications14, 576 (2023)

    2023

  17. [17]

    Schuld, R

    M. Schuld, R. Sweke, and J. J. Meyer, Effect of data encod- ing on the expressive power of variational quantum-machine- learning models, Physical Review A103, 032430 (2021)

    2021

  18. [18]

    Mitarai, M

    K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Quantum circuit learning, Physical Review A98, 032309 (2018)

    2018

  19. [19]

    Watabe, K

    M. Watabe, K. Shiba, M. Sogabe, K. Sakamoto, and T. Sogabe, Quantum circuit parameters learning with gradient descent us- ing backpropagation, arXiv preprint arXiv:1910.14266 (2019)

    Pith/arXiv arXiv 1910

  20. [20]

    Qi, C.-H

    J. Qi, C.-H. Yang, and P.-Y . Chen, Qtn-vqc: An end-to- end learning framework for quantum neural networks, Physica Scripta99, 015111 (2023)

    2023

  21. [21]

    Terashi, M

    K. Terashi, M. Kaneda, T. Kishimoto, M. Saito, R. Sawada, and J. Tanaka, Event classification with quantum machine learning in high-energy physics, Computing and Software for Big Sci- ence5, 1 (2021)

    2021

  22. [22]

    S. Y . C. Chen, C. M. Huang, C. W. Hsing, and Y . J. Kao, Hy- brid quantum-classical classifier based on tensor network and variational quantum circuit, arXiv preprint arXiv:2011.14651 (2020)

    Pith/arXiv arXiv 2011

  23. [23]

    Blance and M

    A. Blance and M. Spannowsky, Quantum machine learning for particle physics using a variational quantum classifier, Journal of High Energy Physics2021, 1 (2021)

    2021

  24. [24]

    Y . Kwak, W. J. Yun, S. Jung, J. K. Kim, and J. Kim, Introduc- tion to quantum reinforcement learning: Theory and pennylane- based implementation, International Conference on ICT Con- vergence , 416 (2021)

    2021

  25. [25]

    Sierra-Sosa, J

    D. Sierra-Sosa, J. Arcila-Moreno, B. Garcia-Zapirain, C. Castillo-Olea, and A. Elmaghraby, Dementia predic- tion applying variational quantum classifier, arXiv preprint arXiv:2007.08653 (2020)

    Pith/arXiv arXiv 2007

  26. [26]

    S. Y . C. Chen, C. H. H. Yang, J. Qi, P. Y . Chen, X. Ma, and H. S. Goan, Variational quantum circuits for deep reinforcement learning, IEEE Access8, 141007 (2020)

    2020

  27. [27]

    Lloyd, M

    S. Lloyd, M. Schuld, A. Ijaz, J. Izaac, and N. Killoran, Quantum embeddings for machine learning, arXiv preprint 16 arXiv:2001.03622 (2020)

    Pith/arXiv arXiv 2001

  28. [28]

    Hubregtsen, D

    T. Hubregtsen, D. Wierichs, E. Gil-Fuster, P.-J. H. S. Derks, P. K. Faehrmann, and J. J. Meyer, Training quantum embedding kernels on near-term quantum computers, Physical Review A 106, 042431 (2022)

    2022

  29. [29]

    L. P. Henry, S. Thabet, C. Dalyac, and L. Henriet, Quantum evo- lution kernel: Machine learning on graphs with programmable arrays of qubits, Physical Review A104, 032416 (2021)

    2021

  30. [30]

    Larocca, S

    M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Bia- monte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo, Barren plateaus in variational quantum computing, Nature Reviews Physics , 1 (2025)

    2025

  31. [31]

    Y . Du, T. Huang, S. You, M. Hsieh, and D. Tao, Quantum cir- cuit architecture search for variational quantum algorithms, npj Quantum Information8, 62 (2022)

    2022

  32. [32]

    Torabian and R

    E. Torabian and R. V . Krems, Compositional optimization of quantum circuits for quantum kernels of support vector ma- chines, Physical Review Research5, 013211 (2023)

    2023

  33. [33]

    Zahid and M

    S. Zahid and M. A. Tahir, Unlocking quantum svm potential: optimal feature map generation and feature selection, Physica Scripta100, 015120 (2024)

    2024

  34. [34]

    Fösel, M

    T. Fösel, M. Y . Niu, F. Marquardt, and L. Li, Quantum circuit optimization with deep reinforcement learning, arXiv preprint arXiv:2103.07585 (2021)

    Pith/arXiv arXiv 2021

  35. [35]

    Chakraborty, S

    S. Chakraborty, S. H. Shaikh, A. Chakrabarti, and R. Ghosh, A hybrid quantum feature selection algorithm using a quantum in- spired graph theoretic approach, Applied Intelligence50, 1775 (2020)

    2020

  36. [36]

    Z. Lu, P. Shen, and D. Deng, Markovian quantum neuroevolu- tion for machine learning, Physical Review Applied16, 044039 (2021)

    2021

  37. [37]

    H. Wang, Y . Ding, J. Gu, Y . Lin, D. Z. Pan, F. T. Chong, and S. Han, Quantumnas: Noise-adaptive search for robust quantum circuits, Proceedings of the IEEE International Symposium on High-Performance Computer Architecture , 69 (2022)

    2022

  38. [38]

    Altares-López, A

    S. Altares-López, A. Ribeiro, and J. J. García-Ripoll, Automatic design of quantum feature maps, Quantum Science and Tech- nology6, 045015 (2021)

    2021

  39. [39]

    Pellow-Jarman, A

    R. Pellow-Jarman, A. Pillay, I. Sinayskiy, and F. Petruccione, Hybrid genetic optimization for quantum feature map design, Quantum Machine Intelligence6, 45 (2024)

    2024

  40. [40]

    Y . Liu, F. Meng, L. Wang, Y . Hu, S. Li, Z. Zhang, and X. Yu, Hardware-aware quantum kernel design based on graph neural networks, arXiv preprint arXiv:2506.21161 (2025)

    Pith/arXiv arXiv 2025

  41. [41]

    Ostaszewski, E

    M. Ostaszewski, E. Grant, and M. Benedetti, Structure opti- mization for parameterized quantum circuits, Quantum5, 391 (2021)

    2021

  42. [42]

    K. M. Nakanishi, K. Fujii, and S. Todo, Sequential minimal optimization for quantum-classical hybrid algorithms, Physical Review Research2, 043158 (2020)

    2020

  43. [43]

    R. M. Parrish, J. T. Iosue, A. Ozaeta, and P. L. McMahon, A Jacobi Diagonalization and Anderson Acceleration Algorithm For Variational Quantum Algorithm Parameter Optimization, arXiv preprint arXiv:1904.03206 (2019)

    Pith/arXiv arXiv 1904

  44. [44]

    J. G. Vidal and D. O. Theis, Calculus on parameterized quantum circuits, arXiv preprint arXiv:1812.06323 (2018)

    Pith/arXiv arXiv 2018

  45. [45]

    Feniou, M

    C. Feniou, M. Hassan, B. Claudon, A. Courtat, O. Adjoua, Y . Maday, and J.-P. Piquemal, Greedy gradient-free adaptive variational quantum algorithms on a noisy intermediate scale quantum computer, Scientific Reports15, 18689 (2025)

    2025

  46. [46]

    Jäger, T

    J. Jäger, T. N. Kaldenbach, M. Haas, and E. Schultheis, Fast gradient-free optimization of excitations in variational quantum eigensolvers, arXiv preprint arXiv:2409.05939 (2025)

    arXiv 2025

  47. [47]

    H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, An adaptive variational algorithm for exact molecular simula- tions on a quantum computer, Nature Communications10, 3007 (2019)

    2019

  48. [48]

    Grossi, N

    M. Grossi, N. Ibrahim, V . Radescu, R. Loredo, K. V oigt, C. V on Altrock, and A. Rudnik, Mixed quantum–classical method for fraud detection with quantum feature selection, IEEE Transactions on Quantum Engineering3, 1 (2022)

    2022

  49. [49]

    A. S. Albino, O. M. Pires, M. Q. Nooblath, and E. G. S. Nasci- mento, Evolutionary quantum feature selection, arXiv preprint arXiv:2303.07131 (2023)

    Pith/arXiv arXiv 2023

  50. [50]

    L. Wang, Z. Y . Chen, F. Y . Le, Z. Q. Yu, C. Xue, X. N. Zhuang, Q. Yan, Y . Yang, Y . C. Wu, and G. P. Guo, A quantum fea- ture selection framework via ground state preparation, Physica Scripta98, 115121 (2023)

    2023

  51. [51]

    Mücke, R

    S. Mücke, R. Heese, S. Müller, M. Wolter, and N. Piatkowski, Feature selection on quantum computers, Quantum Machine In- telligence5, 11 (2023)

    2023

  52. [52]

    Battiti, Using mutual information for selecting features in supervised neural net learning, IEEE Transactions on Neural Networks and Learning Systems5, 537 (1994)

    R. Battiti, Using mutual information for selecting features in supervised neural net learning, IEEE Transactions on Neural Networks and Learning Systems5, 537 (1994)

    1994

  53. [53]

    M. A. F. A. Fida, T. Ahmad, and M. Ntahobari, Variance thresh- old as early screening to boruta feature selection for intrusion detection system, Proceedings of the 13th International Confer- ence on Information and Communication Technology and Sys- tems , 46 (2021)

    2021

  54. [54]

    Guyon, J

    I. Guyon, J. Weston, S. Barnhill, and V . Vapnik, Gene selec- tion for cancer classification using support vector machines, Machine Learning46, 389 (2002)

    2002

  55. [55]

    Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology58, 267 (1996)

    R. Tibshirani, Regression shrinkage and selection via the lasso, Journal of the Royal Statistical Society Series B: Statistical Methodology58, 267 (1996)

    1996

  56. [56]

    Bowles, S

    J. Bowles, S. Ahmed, and M. Schuld, Better than classical? The subtle art of benchmarking quantum machine learning models, arXiv preprint arXiv:2403.07059 (2024)

    Pith/arXiv arXiv 2024

  57. [57]

    Farhi and H

    E. Farhi and H. Neven, Classification with Quantum Neu- ral Networks on Near Term Processors, arXiv preprint arXiv:1802.06002 (2018)

    Pith/arXiv arXiv 2018

  58. [58]

    Schuld, A

    M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, Circuit- centric quantum classifiers, Physical Review A101, 032308 (2020)

    2020

  59. [59]

    [66, 94, 95]

    This loss function draws motivation from statistics and infor- mation theory, also known as log loss, logistic loss or cross- entropy loss, and has been established as the standard choice in the classical machine learning literature – including extensions to multi-class – classification. [66, 94, 95]

  60. [60]

    Huang, M

    H.-Y . Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, Power of data in quantum machine learning, Nature Communications12, 2631 (2021)

    2021

  61. [61]

    B. E. Boser, I. M. Guyon, and V . N. Vapnik, A training algo- rithm for optimal margin classifiers, inProceedings of the fifth annual workshop on Computational learning theory, COLT ’92 (Association for Computing Machinery, New York, NY , USA,

  62. [62]

    Note that given the evaluated quantum kernel values, commonly aggregated in the so-called kernel matrix or Gram matrix, the training of the QSVM does not differ from that of classical SVMs [66, 94, 95]

  63. [63]

    Cristianini, J

    N. Cristianini, J. Shawe-Taylor, A. Elisseeff, and J. Kandola, On kernel-target alignment, Advances in neural information processing systems14(2001)

    2001

  64. [64]

    A. E. Paine, V . E. Elfving, and O. Kyriienko, Quantum kernel methods for solving regression problems and differential equa- tions, Physical Review A107, 032428 (2023). 17

    2023

  65. [65]

    Salmenperä, I

    I. Salmenperä, I. Kuhtarskis, A. M. van de Griend, and J. K. Nurminen, The impact of feature embedding placement in the ansatz of a quantum kernel in qsvms, in2024 IEEE Interna- tional Conference on Quantum Computing and Engineering (QCE), V ol. 1 (IEEE, 2024) pp. 1663–1671

    2024

  66. [66]

    The required form to apply Q-FLAIR is still maintained when appending a fixed gate along with a (weight-) data-dependent gate because the fixed gate is absorbed by the state and observ- able in Eq. (25)

  67. [67]

    Bowles, S

    J. Bowles, S. Ahmed, and M. Schuld, Pennylane datasets for better than classical? the subtle art of bench- marking quantum machine learning models,https: //pennylane.ai/datasets/linearly-separable, https://pennylane.ai/datasets/two-curves, https://pennylane.ai/datasets/bars-and-stripes, https://pennylane.ai/datasets/downscaled-mnist (2024), accessed: 2024-09-22

    2024

  68. [68]

    LeCun, The MNIST database of handwritten digits, http://yan.lecun.com/exdb/mnist/ (1998)

    Y . LeCun, The MNIST database of handwritten digits, http://yan.lecun.com/exdb/mnist/ (1998)

    1998

  69. [69]

    Buchanan, D

    S. Buchanan, D. Gilboa, and J. Wright, Deep networks and the multiple manifold problem, arXiv preprint arXiv:2008.11245 (2020)

    Pith/arXiv arXiv 2008

  70. [70]

    A. C. Lorena, A. I. Maciel, P. B. de Miranda, I. G. Costa, and R. B. Prudêncio, Data complexity meta-features for regression problems, Machine Learning107, 209 (2018)

    2018

  71. [71]

    K. P. Murphy,Machine learning: a probabilistic perspective, 2nd ed. (MIT Press, Cambridge, MA, 2012)

    2012

  72. [72]

    Coelho, G

    R. Coelho, G. Kruse, and A. Rosskopf, Quantum-Efficient Ker- nel Target Alignment, arXiv preprint arXiv:2502.08225 (2025)

    Pith/arXiv arXiv 2025

  73. [73]

    Due to qubit demands, we do not directly compare Q-FLAIR feature maps with fixed-ansatz maps matching the qubit count to the feature count, e.g., angle embedding, Z-feature-map, and ZZ-feature-map [4]

  74. [74]

    Kerenidis and A

    I. Kerenidis and A. Luongo, Classification of the mnist data set with quantum slow feature analysis, Physical Review A101, 062327 (2020)

    2020

  75. [75]

    Slysz, K

    M. Slysz, K. Kurowski, G. Waligóra, and J. W˛ eglarz, Exploring the capabilities of quantum support vector machines for image classification on the mnist benchmark, inInternational Confer- ence on Computational Science(Springer, 2023) pp. 193–200

    2023

  76. [76]

    Senokosov, A

    A. Senokosov, A. Sedykh, A. Sagingalieva, B. Kyriacou, and A. Melnikov, Quantum machine learning for image classifica- tion, Machine Learning: Science and Technology5, 015040 (2024)

    2024

  77. [77]

    R. Zhou, S. Sarkar, S. Bose, and A. Bayat, Enhanced image classification via hybridizing quantum dynamics with classical neural networks, arXiv preprint arXiv:2507.13587 (2025)

    Pith/arXiv arXiv 2025

  78. [78]

    Erkan, J

    Z. Erkan, J. Rahebi, and A. Yelghi, Quantum image dataset transform (qidt) for image processing, Quantum Information Processing24, 156 (2025)

    2025

  79. [79]

    K. Chen, J. Liu, and F. Yan, Exploring quantum neural networks for binary classification on mnist dataset: A swap test approach, Neural Networks188, 107442 (2025)

    2025

  80. [80]

    P. A. X. Tognini, L. Banchi, and G. De Palma, Solving mnist with a globally trained mixture of quantum experts, arXiv preprint arXiv:2505.14789 (2025)

    Pith/arXiv arXiv 2025

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.