pith. sign in

arxiv: 2606.28483 · v1 · pith:NCCD67ENnew · submitted 2026-06-26 · 🪐 quant-ph

Quantum Fourier Generative Models Trainable at Large Scale

Cenk T\"uys\"uz , Oleksandr Kyriienko , Michele Grossi This is my paper

Pith reviewed 2026-06-30 01:22 UTC · model grok-4.3

classification 🪐 quant-ph
keywords quantum generative modelsFourier feature mapsforrelation circuitslog-likelihood lossParseval's identityMonte Carlo estimatortotal variation distancequantum sampling
0
0 comments X

The pith

Quantum generative models train classically at over 1000 qubits using Fourier feature maps and a Parseval-based log-likelihood estimator, then deploy to quantum hardware for sampling.

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

The paper establishes a framework that embeds continuous variables through parallel Fourier feature maps, tunes coefficients with forrelation-type circuits, and trains via an unbiased Monte Carlo estimator of log-likelihood derived from Parseval's identity. This enables classical training at scales exceeding 1000 qubits on a single GPU while capturing non-trivial univariate and bivariate structures to low total variation distance. The trained model then converts via inverse quantum Fourier transform into a sampling circuit executable on superconducting hardware at roughly 300 microseconds per sample. A sympathetic reader would care because the method separates scalable classical training from fast quantum sampling and demonstrates better preservation of multi-modal structure than normalizing flows, diffusion models, or prior MMD-based quantum approaches.

Core claim

By combining parallel Fourier feature maps for embedding with forrelation-type circuits and training on an unbiased Monte Carlo estimator of log-likelihood loss obtained from Parseval's identity, quantum generative models can be trained at scales over 1000 qubits on classical hardware; once trained, inverse quantum Fourier transforms map the model to a separate sampling circuit in the computational basis that runs on quantum devices, achieving low total variation distance while avoiding the oversmoothing seen in classical baselines.

What carries the argument

The unbiased Monte Carlo estimator of log-likelihood loss derived from Parseval's identity applied to Fourier feature maps and forrelation circuits, which enables tractable classical training at large scale before inverse-QFT mapping to a sampling circuit.

If this is right

  • Univariate and bivariate models with highly non-trivial structure reach low total variation distance to the target.
  • The trained models avoid oversmoothing and preserve multi-modal structure better than normalizing flow or diffusion baselines.
  • Fine-tuned IQP circuits trained with MMD loss perform poorly by comparison.
  • Deployed models achieve per-sample execution times of approximately 300 microseconds on superconducting quantum hardware.

Where Pith is reading between the lines

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

  • The separation of classical training from quantum sampling phases could allow the same framework to target higher-dimensional distributions if the estimator's scaling holds.
  • Fast per-sample quantum execution times open the possibility of using these models for real-time inference tasks that require many draws from complex distributions.
  • If the Fourier embedding generalizes, the approach might serve as a template for other quantum generative tasks where moment-matching losses have previously limited expressivity.

Load-bearing premise

The log-likelihood loss estimator based on Parseval's identity remains unbiased and computationally tractable for the chosen Fourier feature maps and forrelation circuits at the reported qubit counts without hidden dependence on post-hoc parameter choices.

What would settle it

Running exact log-likelihood computation on a 10-qubit instance of the same Fourier feature map and forrelation circuit and finding that the Monte Carlo estimator deviates by more than statistical error would falsify the claim of unbiased scalability.

Figures

Figures reproduced from arXiv: 2606.28483 by Cenk T\"uys\"uz, Michele Grossi, Oleksandr Kyriienko.

Figure 1
Figure 1. Figure 1: Train-on-classical deploy-on-quantum workflow. During training, data are encoded through a Fourier feature map and the model parameters are optimized classically in the latent Fourier representation. After training, the same parameters are deployed in a quantum sampling circuit, where inverse quantum Fourier transforms (QFT† ) produce samples in the computational basis. The construction combines a forrelat… view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy and scaling of the Monte Carlo estimator for estimating the cost function. (a) The top panel shows the mean absolute error |pˆ−pref| for estimating ˆp = Pbθ(0n | x), with pref obtained from exact statevector simulation. The dashed line marks the natural probability scale 2−n . The bottom panel shows the corresponding wall-clock evaluation time, compared with exact statevector evaluation. For each … view at source ↗
Figure 3
Figure 3. Figure 3: Learning and sampling from univariate benchmark distributions. (a) Learned model densities for four target distributions on the float16 grid over x ∈ [0, 1], containing 216 discrete points. The orange filled curves show the exact trained model densities on the discrete support, and the dark curves show the target densities. From top left to bottom right, the targets are a unimodal Gaussian, a two-component… view at source ↗
Figure 4
Figure 4. Figure 4: Bivariate benchmark against an IQP+MMD baseline. Comparison on a four-peak Gaussian target with an 8-bit discretization for each coordinate. The top row shows the target distribution, the IQP+MMD baseline, and the trained DQGM. The IQP baseline contains all one-qubit RZ rotations and all two-qubit RZZ rotations on the visible and hidden qubits (m = 4) , giving 210 trainable parameters. The DQGM uses m = 4 … view at source ↗
Figure 5
Figure 5. Figure 5: Bivariate benchmark against classical gener￾ative baselines. Comparison on a target distribution con￾sisting of separated Gaussian peaks arranged along a spiral. All models use the same discrete data access model with an 8-bit discretization for each coordinate. The panels show the target distribution, the trained DQGM, a normalizing-flow baseline, and a diffusion-model baseline. Distributions are plotted … view at source ↗
Figure 6
Figure 6. Figure 6: Pairwise correlation structure on the four￾feature finance dataset. (top) Two-dimensional marginals p(xi, xj ) for the feature pairs (x0, x3) and (x2, x3), comparing the training data (left) with samples from the trained DQGM (middle) and the IQP+MMD baseline (right). Each model panel reports its Kullback-Leibler divergence (KLD) and to￾tal variation distance (TVD) to the corresponding two-feature projecti… view at source ↗
Figure 7
Figure 7. Figure 7: Training and sampling circuits. Both circuits act on k features, each encoded with d-bit precision, for a total of n = kd visible qubits, together with m hidden qubits. In the training circuit (a), all visible qubits are measured and the hidden register is marginalized to obtain Pθ(0n | x). In the sampling circuit (b), an inverse QFT is applied to each feature block and the hidden register is discarded. Ea… view at source ↗
Figure 8
Figure 8. Figure 8: A detailed view of the training circuit. The encoding Uϕ(x) applies a Hadamard and an RZ(2πx/2 j ) to each visible qubit j. The inverse model unitary U † M then applies two diagonal layers D1, D2 separated by a fixed RY(π/2) on the visible qubits, conjugated by Hadamard layers on all N qubits. The training signal Pθ(0n | x) is the probability of the all-zero visible outcome with the hidden register traced … view at source ↗
Figure 9
Figure 9. Figure 9: Detailed Monte Carlo estimator accuracy and runtime benchmark. The three columns correspond to K = 102 , K = 103 , and K = 104 Monte Carlo samples. The horizontal axis shows the total number of qubits N = n + m, with n = m in all panels. Each (N, K) point uses 100 independently initialized training circuits. For each circuit, the same visible marginal probability is estimated 100 times with independent Mon… view at source ↗
Figure 10
Figure 10. Figure 10: Peak GPU memory for the Parseval Monte Carlo simulator. Measured and estimated peak memory are shown as a function of the total number of qubits for K = 102 , K = 103 , and K = 104 Monte Carlo samples. Measurements are per-process maxima obtained by polling nvidia-smi during execution on a single NVIDIA Tesla V100S GPU with 32 GB memory. The analytic estimate is given by Eq. (D5). The dashed horizontal li… view at source ↗
Figure 11
Figure 11. Figure 11: Learning curves for the univariate benchmark distributions. Columns correspond to the Gaussian, Gaussian mixture, cosine, and L´evy targets used in Fig. 3a. The rows show the mini-batch negative log-likelihood, the full-distribution KL divergence DKL(ptarget∥pmodel), and the total variation distance TVD(ptarget, pmodel). NLL is logged every 10 optimization steps on mini-batches of size 64. KL and TVD are … view at source ↗
Figure 12
Figure 12. Figure 12: Stage-by-stage comparison of the learned univariate densities on hardware. Each row is a separate target distribution (top to bottom: Gaussian, Gaussian mixture, cosine, and L´evy), and the columns trace the pipeline from data to device, left to right. (i) Training data, the empirical histogram of the samples the model was trained on. (ii) Ideal (exact), the trained DQGM circuit simulated exactly with a s… view at source ↗
Figure 13
Figure 13. Figure 13: Bivariate benchmark of exact and Monte Carlo IQP training. Comparison on a four-peak Gaussian target with an 8-bit discretization for each coordinate. The top row shows the target distribution (left), the best IQP trained with the exact MMD loss (middle), and the best IQP+MMD trained with a Monte Carlo (MC) estimator (right). Exact training does not scale to large numbers of qubits, and we include it as a… view at source ↗
Figure 14
Figure 14. Figure 14: Classical baseline samples on blobs-on-spiral. Learned joint probability mass over the 28 × 2 8 grid for the normalizing-flow (top row) and diffusion (bottom row) baselines at three model capacities, denoted (a), (b), and (c) in order of increasing size. Architectures and parameter counts are listed in Tables III and IV. Each panel shows the best of 10 seeds for the configuration selected by the hyperpara… view at source ↗
Figure 15
Figure 15. Figure 15: Pairwise two-dimensional marginals p(xi, xj ) for the multivariate benchmark on the finance dataset. The top row is the training-data (target) distribution and the lower rows are the trained DQGM and IQP+MMD models, and the columns are the unordered feature pairs. All panels share a common color scale (white at zero) and a single color bar, and each model panel is annotated with the Kullback–Leibler diver… view at source ↗
Figure 16
Figure 16. Figure 16: Preservation of the pairwise-correlation ordering on JetNet. For every feature pair we rank the Pearson correlations of the training data against those of one million samples drawn from each model, so points on the dashed diagonal indicate that the model reproduces the empirical ordering exactly. The legend reports the Spearman rank correlation ρ. The JetNet particle-physics dataset has k = 5 features and… view at source ↗
Figure 17
Figure 17. Figure 17: Pairwise two-dimensional marginals p(xi, xj ) on JetNet. The top row is the training-data (target) distribution and the lower rows are the trained DQGM and IQP+MMD models, and the columns are the unordered feature pairs. All panels share a common colour scale (white at zero) and a single colour bar, and each model panel is annotated with the Kullback– Leibler divergence (KLD) and total variation distance … view at source ↗
Figure 18
Figure 18. Figure 18: Sampling a trained circuit at higher precision by upsampling, with no retraining. Each row is a separate trained DQGM run, a cosine density in the top row and a Gaussian mixture in the bottom row, both trained at 8-bit output precision. The leftmost panel shows the trained 8-bit readout, and the two bracketed panels show the same learned parameters read out at 10 and 12 bits by appending ancilla qubits in… view at source ↗
read the original abstract

We propose an algorithmic framework for building and training quantum generative models corresponding to multivariate probability distributions. Our model uses parallel Fourier feature maps for embedding continuous-valued variables combined with a forrelation-type quantum circuit for tuning Fourier coefficients of the quantum model. Crucially, we develop a distinct training strategy where training is enabled at large scale by log-likelihood loss with unbiased Monte Carlo estimator based on Parseval's identity. Unlike prior work that relied on maximal mean discrepancy (MMD) loss, our approach goes beyond matching just low frequency moments, while enabling efficient classical training. Once the model is trained, we use inverse quantum Fourier transforms to map it into a separate sampling circuit in the computational basis. We demonstrate the efficiency of the suggested framework by validating loss estimation at the scale of over 1000 qubits on a single GPU. We show that univariate and bivariate models with highly non-trivial structure can be trained to low total variation distance, while fine-tuned IQP models with MMD loss show poor performance. Comparing to classical baselines represented by normalizing flow and diffusion models, we show that our approach avoids oversmoothing and preserves multi-modal structure of the target. Finally, we have deployed the trained models on superconducting quantum devices, successfully sampling distributions with per-sample execution times of approximately $300\,\mu\mathrm{s}$. Our work shows that quantum generative models with the train-on-classical deploy-on-quantum approach can provide both high-quality structure at increased scale and fast sampling access needed for inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes an algorithmic framework for quantum generative models of multivariate distributions. It combines parallel Fourier feature maps for continuous variables with a forrelation-type circuit to tune Fourier coefficients. A key contribution is a log-likelihood training objective whose unbiased Monte Carlo estimator is derived from Parseval's identity, enabling classical training at scales exceeding 1000 qubits on a single GPU. After training, an inverse quantum Fourier transform produces a sampling circuit in the computational basis. Experiments report low total-variation distance for univariate and bivariate targets with non-trivial structure, superior performance relative to MMD-tuned IQP circuits and classical normalizing-flow/diffusion baselines (avoiding oversmoothing), and deployment on superconducting hardware with ~300 μs per-sample times.

Significance. If the Parseval-based estimator is confirmed to be unbiased and to possess tractable variance independent of the fitted model, the work would demonstrate a concrete route to classically trainable quantum generative models that scale beyond current MMD-limited approaches while preserving multi-modal structure. The train-on-classical/deploy-on-quantum separation and the reported hardware sampling times would constitute a practical advantage over purely classical or purely quantum generative methods at the claimed qubit counts.

major comments (2)
  1. [Methods / loss estimator] The central scalability claim (loss estimation and training at >1000 qubits) rests on the Monte Carlo estimator derived from Parseval's identity remaining unbiased and having variance that does not grow prohibitively with the number of Fourier modes or the forrelation circuit depth. The abstract provides no explicit variance bound, circuit-depth scaling, or proof that the estimator is independent of post-hoc parameter choices; this must be supplied with a concrete derivation or numerical verification in the methods section before the efficiency result can be accepted.
  2. [Experiments] The comparison to fine-tuned IQP models with MMD loss reports poor performance, yet the paper does not specify the circuit depth, number of Fourier modes, or optimization hyperparameters used for the IQP baseline. Without these controls it is impossible to determine whether the reported advantage is due to the loss function, the feature-map architecture, or differences in model capacity.
minor comments (2)
  1. [Abstract] The abstract states that models are trained to low total variation distance, but no numerical values, error bars, or dataset sizes are given; these should appear in the results tables or figures.
  2. [Model definition] Notation for the parallel Fourier feature maps and the forrelation unitary should be introduced with explicit definitions of the feature dimension and the circuit's action on the Fourier coefficients.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive comments. We respond to each major comment below, indicating where we will revise the manuscript to address the concerns.

read point-by-point responses

  1. Referee: [Methods / loss estimator] The central scalability claim (loss estimation and training at >1000 qubits) rests on the Monte Carlo estimator derived from Parseval's identity remaining unbiased and having variance that does not grow prohibitively with the number of Fourier modes or the forrelation circuit depth. The abstract provides no explicit variance bound, circuit-depth scaling, or proof that the estimator is independent of post-hoc parameter choices; this must be supplied with a concrete derivation or numerical verification in the methods section before the efficiency result can be accepted.

    Authors: The methods section already derives the unbiased estimator from Parseval's identity and demonstrates numerical stability at >1000 qubits. We agree, however, that an explicit variance analysis is needed to fully support the scalability claim. In the revised manuscript we will add a derivation showing that the estimator variance depends only on the number of Monte Carlo samples (and is independent of forrelation depth and Fourier-mode count) together with additional numerical verification of variance scaling. revision: yes

  2. Referee: [Experiments] The comparison to fine-tuned IQP models with MMD loss reports poor performance, yet the paper does not specify the circuit depth, number of Fourier modes, or optimization hyperparameters used for the IQP baseline. Without these controls it is impossible to determine whether the reported advantage is due to the loss function, the feature-map architecture, or differences in model capacity.

    Authors: We acknowledge the omission of these controls. The revised experiments section will explicitly state the circuit depth, number of Fourier modes, and optimization hyperparameters employed for the MMD-tuned IQP baselines, enabling a clear assessment that the observed performance gap arises from the loss function and architecture rather than unequal model capacity. revision: yes

Circularity Check

0 steps flagged

No significant circularity; estimator grounded in external theorem

full rationale

The paper's training strategy relies on a log-likelihood estimator derived via Parseval's identity (a standard external theorem) applied to the Fourier feature map model. This is used to enable classical training of parameters in the forrelation circuit, followed by separate inverse QFT sampling. No load-bearing step reduces by construction to fitted inputs, self-definition, or self-citation chains. The scalability demonstration at >1000 qubits is presented as empirical validation rather than a derived equivalence. The derivation chain remains self-contained against external mathematical facts and does not exhibit the enumerated circular patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The framework rests on standard quantum circuit assumptions and the applicability of Parseval's identity to the chosen embedding; no independent evidence for the new model class is supplied beyond the abstract claims.

free parameters (1)
  • Fourier coefficients
    Tuned during classical training; exact count and initialization not specified in abstract.
axioms (1)
  • domain assumption Parseval's identity applies directly to the quantum Fourier feature map and forrelation circuit to yield an unbiased Monte Carlo estimator of log-likelihood.
    Invoked to justify the training loss in the abstract.
invented entities (1)
  • Quantum Fourier generative model with parallel feature maps and forrelation circuit no independent evidence
    purpose: To represent and sample multivariate probability distributions
    New model class introduced in the abstract; no external falsifiable prediction supplied.

pith-pipeline@v0.9.1-grok · 5798 in / 1403 out tokens · 32653 ms · 2026-06-30T01:22:48.763701+00:00 · methodology

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

103 extracted references · 28 canonical work pages · 6 internal anchors

  1. [1]

    Generalized Denoising Auto-Encoders as Generative Models

    Y. Bengio, L. Yao, G. Alain, and P. Vincent, General- ized denoising auto-encoders as generative models (2013), arXiv:1305.6663 [cs.LG]

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  2. [2]

    Rezende and S

    D. Rezende and S. Mohamed, inProceedings of the 32nd International Conference on Machine Learning, Proceed- ings of Machine Learning Research, Vol. 37, edited by F. Bach and D. Blei (PMLR, Lille, France, 2015) pp. 1530–1538

    2015

  3. [3]

    J. Song, C. Meng, and S. Ermon, inInternational Con- ference on Learning Representations(2021)

    2021

  4. [4]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, inInternational Conference on Learning Representations(2021)

    2021

  5. [5]

    Liu and L

    J.-G. Liu and L. Wang, Phys. Rev. A98, 062324 (2018)

    2018

  6. [6]

    Zoufal, A

    C. Zoufal, A. Lucchi, and S. Woerner, npj Quantum In- formation5, 103 (2019)

    2019

  7. [7]

    Coyle, D

    B. Coyle, D. Mills, V. Danos, and E. Kashefi, npj Quan- tum Information6, 60 (2020)

    2020

  8. [8]

    M. J. Bremner, R. Jozsa, and D. J. Shepherd, Proceedings of the Royal Society A: Mathemati- cal, Physical and Engineering Sciences467, 459 (2010), https://royalsocietypublishing.org/rspa/article- pdf/467/2126/459/789839/rspa.2010.0301.pdf

    work page arXiv 2010

  9. [9]

    Aaronson and A

    S. Aaronson and A. Arkhipov, Theory of Computing9, 143 (2013)

    2013

  10. [10]

    Arute, K

    F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. S. L. Brandao, D. A. Buell, and et al., Nature574, 505 (2019)

    2019

  11. [11]

    Hangleiter and J

    D. Hangleiter and J. Eisert, Rev. Mod. Phys.95, 035001 (2023)

    2023

  12. [12]

    Huang, M

    H.-Y. Huang, M. Broughton, N. Eassa, H. Neven, R. Babbush, and J. R. McClean, Generative quantum advantage for classical and quantum problems (2025), arXiv:2509.09033 [quant-ph]

    work page arXiv 2025

  13. [13]

    M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, Physical Review X8, 021050 (2018)

    2018

  14. [14]

    Coopmans and M

    L. Coopmans and M. Benedetti, Communications Physics7, 274 (2024)

    2024

  15. [15]

    T¨ uys¨ uz, M

    C. T¨ uys¨ uz, M. Demidik, L. Coopmans, E. Rinaldi, V. Croft, Y. Haddad, M. Rosenkranz, and K. Jansen, Learning to generate high-dimensional distributions with low-dimensional quantum boltzmann machines (2024), arXiv:2410.16363 [quant-ph]

    work page arXiv 2024

  16. [16]

    Demidik, C

    M. Demidik, C. T¨ uys¨ uz, N. Piatkowski, M. Grossi, and K. Jansen, Communications Physics8, 413 (2025)

    2025

  17. [17]

    Demidik, C

    M. Demidik, C. T¨ uys¨ uz, M. Grossi, and K. Jansen, Sample-based training of quantum generative models (2025), arXiv:2511.11802 [quant-ph]

    work page arXiv 2025

  18. [18]

    Kyriienko, A

    O. Kyriienko, A. E. Paine, and V. E. Elfving, Physical Review Research6, 033291 (2024)

    2024

  19. [19]

    A. E. Paine, V. E. Elfving, and O. Kyriienko, Advanced Quantum Technologies6, 2300065 (2023)

    2023

  20. [20]

    Bak´ o, D

    B. Bak´ o, D. T. R. Nagy, P. H´ aga, Z. Kallus, and Z. Zim- bor´ as, Quantum Science and Technology11, 035012 (2026)

    2026

  21. [21]

    Oszmaniec, N

    M. Oszmaniec, N. Dangniam, M. E. Morales, and Z. Zim- bor´ as, PRX Quantum3, 020328 (2022)

    2022

  22. [22]

    Y. Wang, S. Xue, Y. Wang, Y. Liu, J. Ding, W. Shi, D. Wang, Y. Liu, X. Fu, G. Huang, A. Huang, M. Deng, and J. Wu, Opt. Lett.48, 5197 (2023)

    2023

  23. [23]

    Kailasanathan, W

    R. Kailasanathan, W. R. Clements, M. R. Bosk- abadi, S. M. Gibford, E. Papadakis, C. J. Savoie, and S. S. Mansouri, Quantum enhanced ensemble gans for anomaly detection in continuous biomanufacturing (2026), arXiv:2508.21438 [cs.LG]

    work page arXiv 2026

  24. [24]

    Quantum latent distributions in deep generative models

    O. Bacarreza, T. Farnsworth, A. Makarovskiy, H. Wall- ner, T. Hicks, S. Sempere-Llagostera, J. Price, R. J. A. Francis-Jones, and W. R. Clements, Quantum la- tent distributions in deep generative models (2026), arXiv:2508.19857 [cs.LG]

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Herbst, I

    S. Herbst, I. Brandi´ c, and A. P´ erez-Salinas, Limits of quantum generative models with classical sampling hard- ness (2025), arXiv:2512.24801 [quant-ph]

    work page arXiv 2025

  26. [26]

    Holmes, K

    Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, PRX Quantum3, 010313 (2022)

    2022

  27. [27]

    Cerezo, M

    M. Cerezo, M. Larocca, D. Garc´ ıa-Mart´ ın, N. L. Diaz, P. Braccia, E. Fontana, M. S. Rudolph, P. Bermejo, A. Ijaz, S. Thanasilp, E. R. Anschuetz, and Z. Holmes, Nature Communications16, 7907 (2025)

    2025

  28. [28]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications9, 4812 (2018)

    2018

  29. [29]

    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, Nature Reviews Physics7, 174 (2025)

    2025

  30. [30]

    You \ and\ author X

    X. You and X. Wu, Exponentially many local minima in quantum neural networks (2021), arXiv:2110.02479 [quant-ph]

    work page arXiv 2021

  31. [31]

    E. R. Anschuetz, Critical points in quantum generative models (2023), arXiv:2109.06957 [quant-ph]

    work page arXiv 2023

  32. [32]

    Schuld, V

    M. Schuld, V. Bergholm, C. Gogolin, J. Izaac, and N. Kil- loran, Phys. Rev. A99, 032331 (2019)

    2019

  33. [33]

    Kyriienko and V

    O. Kyriienko and V. E. Elfving, Phys. Rev. A104, 052417 (2021)

    2021

  34. [34]

    Wierichs, J

    D. Wierichs, J. Izaac, C. Wang, and C. Y.-Y. Lin, Quan- tum6, 677 (2022)

    2022

  35. [35]

    Kasture, O

    S. Kasture, O. Kyriienko, and V. E. Elfving, Phys. Rev. A108, 042406 (2023)

    2023

  36. [36]

    Train on classical, deploy on quantum: scaling generative quantum machine learn- ing to a thousand qubits

    E. Recio-Armengol, S. Ahmed, and J. Bowles, Train on classical, deploy on quantum: scaling generative quantum machine learning to a thousand qubits (2026), arXiv:2503.02934 [quant-ph]

    work page arXiv 2026

  37. [37]

    Kurkin, K

    A. Kurkin, K. Shen, S. Pielawa, H. Wang, and V. Dunjko, Universality and kernel-adaptive training for classically trained, quantum-deployed generative models (2025), arXiv:2510.08476 [quant-ph]

    work page arXiv 2025

  38. [38]

    Bak´ o, Z

    B. Bak´ o, Z. Kolarovszki, and Z. Zimbor´ as, Fermionic born machines: Classical training of quantum gen- erative models based on fermion sampling (2025), arXiv:2511.13844 [quant-ph]

    work page arXiv 2025

  39. [39]

    Kolarovszki, B

    Z. Kolarovszki, B. Bak´ o, M. Oszmaniec, C. Oh, and Z. Zimbor´ as, Generative modeling with gaussian boson sampling: classically trainable bosonic born machines (2026), arXiv:2603.11195 [quant-ph]

    work page arXiv 2026

  40. [40]

    Gottlieb, R

    F. Gottlieb, R. Mezher, B. Ventura, S. Mansfield, and A. Salavrakos, Efficient training of photonic quantum generative models (2026), arXiv:2603.08793 [quant-ph]

    work page arXiv 2026

  41. [41]

    Kurkin, U

    A. Kurkin, U. Chabaud, Z. Kolarovszki, B. Bak´ o, Z. Zim- bor´ as, and V. Dunjko, Universality of classically train- able, quantum-deployed boson-sampling generative mod- els (2026), arXiv:2603.11014 [quant-ph]

    work page arXiv 2026

  42. [42]

    M. V. den Nest, Simulating quantum computers with probabilistic methods (2010), arXiv:0911.1624 [quant- 12 ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  43. [43]

    M. S. Rudolph, S. Lerch, S. Thanasilp, O. Kiss, O. Shaya, S. Vallecorsa, M. Grossi, and Z. Holmes, npj Quantum Information10, 116 (2024)

    2024

  44. [44]

    Spectral methods: crucial for machine learning, natural for quantum computers?

    V. Belis, J. Bowles, R. Gupta, E. Peters, and M. Schuld, Spectral methods: crucial for machine learning, natu- ral for quantum computers? (2026), arXiv:2603.24654 [quant-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  45. [45]

    Herrero-Gonzalez, B

    M. Herrero-Gonzalez, B. Coyle, K. McDowall, R. Grassie, S. Beentjes, A. Khamseh, and E. Kashefi, The born ultimatum: Conditions for classical surrogation of quantum generative models with correlators (2025), arXiv:2511.01845 [quant-ph]

    work page arXiv 2025

  46. [46]

    K. Shen, S. Pielawa, V. Dunjko, and H. Wang, Character- izing trainability of instantaneous quantum polynomial circuit born machines (2026), arXiv:2602.11042 [quant- ph]

    work page arXiv 2026

  47. [47]

    Wakeham and M

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

    work page arXiv 2024

  48. [48]

    Di Meglio, K

    A. Di Meglio, K. Jansen, I. Tavernelli,et al., PRX Quan- tum5, 037001 (2024)

    2024

  49. [49]

    Aaronson and A

    S. Aaronson and A. Ambainis, SIAM Journal on Com- puting47, 982 (2018)

    2018

  50. [50]

    Umeano, S

    C. Umeano, S. Scali, and O. Kyriienko, Phys. Rev. A 113, 052425 (2026)

    2026

  51. [51]

    Bravyi, D

    S. Bravyi, D. Gosset, D. Grier, and L. Schaeffer, Clas- sical algorithms for Forrelation (2021), arXiv:2102.06963 [quant-ph]

    work page arXiv 2021

  52. [52]

    G. E. Hinton, A practical guide to training restricted boltzmann machines, inNeural Networks: Tricks of the Trade: Second Edition, edited by G. Montavon, G. B. Orr, and K.-R. M¨ uller (Springer Berlin Heidelberg, Berlin, Heidelberg, 2012) pp. 599–619

    2012

  53. [53]

    Demonstrating Record Fidelity for the Quantum Fourier Transform

    P. Aumann, M. Fellner, D. Alber, M. Cykiert, C. Flecken- stein, R. ter Hoeven, L. Stenzel, R. J. Valencia-Tortora, and W. Lechner, Demonstrating record fidelity for the quantum fourier transform (2026), arXiv:2604.12465 [quant-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  54. [54]

    T¨ uys¨ uzet al.,https://github.com/cnktysz/DQGM (2026), GitHub repository

    C. T¨ uys¨ uzet al.,https://github.com/cnktysz/DQGM (2026), GitHub repository

    2026

  55. [55]

    Revised MD17 dataset (rMD17),

    C. T¨ uys¨ uzet al.,https://doi.org/10.6084/m9. figshare.32789952(2026), Public dataset repository

    work page doi:10.6084/m9 2026

  56. [56]

    M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information, 10th ed. (Cambridge Univer- sity Press, 2010)

    2010

  57. [57]

    C. A. Williams, A. E. Paine, H.-Y. Wu, V. E. Elfving, and O. Kyriienko, Quantum chebyshev transform: Mapping, embedding, learning and sampling distributions (2023), arXiv:2306.17026 [quant-ph]

    work page arXiv 2023

  58. [58]

    H.-Y. Wu, V. E. Elfving, and O. Kyriienko, Advanced Quantum Technologies8, 2400337 (2025)

    2025

  59. [59]

    J. J. Mart´ ınez de Lejarza, H.-Y. Wu, O. Kyriienko, G. Rodrigo, and M. Grossi, Communications Physics8, 448 (2025)

    2025

  60. [60]

    Shepherd and M

    D. Shepherd and M. J. Bremner, Proceedings of the Royal Society A: Mathematical, Physi- cal and Engineering Sciences465, 1413 (2009), https://royalsocietypublishing.org/rspa/article- pdf/465/2105/1413/753599/rspa.2008.0443.pdf

    work page arXiv 2009

  61. [61]

    M. J. Bremner, A. Montanaro, and D. J. Shepherd, Phys. Rev. Lett.117, 080501 (2016)

    2016

  62. [62]

    Aaronson, inProceedings of the Forty-Second ACM Symposium on Theory of Computing, STOC ’10 (Asso- ciation for Computing Machinery, New York, NY, USA,

    S. Aaronson, inProceedings of the Forty-Second ACM Symposium on Theory of Computing, STOC ’10 (Asso- ciation for Computing Machinery, New York, NY, USA,

  63. [63]

    O’Donnell,Analysis of Boolean Functions(Cambridge University Press, USA, 2014)

    R. O’Donnell,Analysis of Boolean Functions(Cambridge University Press, USA, 2014)

    2014

  64. [64]

    Quantum computing with Qiskit

    A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. Johnson, and J. M. Gambetta, Quantum computing with Qiskit (2024), arXiv:2405.08810 [quant-ph]

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  65. [65]

    Akiba, S

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, inProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2019)

    2019

  66. [66]

    Durkan, A

    C. Durkan, A. Bekasov, I. Murray, and G. Papamakar- ios, inProceedings of the 33rd International Conference on Neural Information Processing Systems(Curran As- sociates Inc., Red Hook, NY, USA, 2019)

    2019

  67. [67]

    Rozet and others, Zuko: Normalizing flows in pytorch (2024)

    F. Rozet and others, Zuko: Normalizing flows in pytorch (2024)

    2024

  68. [68]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨ opf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, inProceedings of the 33rd In- ternational Conference on Neural Information Processing Systems(Curran Associates In...

    2019

  69. [69]

    J. Ho, A. Jain, and P. Abbeel, inProceedings of the 34th International Conference on Neural Information Pro- cessing Systems, NIPS ’20 (Curran Associates Inc., Red Hook, NY, USA, 2020)

    2020

  70. [70]

    Kansal, J

    R. Kansal, J. Duarte, H. Su, B. Orzari, T. Tomei, M. Pierini, M. Touranakou, J.-R. Vlimant, and D. Gunopulos, 10.5281/zenodo.6975118 (2022). 13 CONTENTS I. Introduction 1 II. Framework 2 A. Model design 2 B. Classical training algorithm 3 III. Numerical results 4 A. Estimator accuracy and scaling 4 B. Learning univariate benchmark distributions 5 C. Sampl...

    work page doi:10.5281/zenodo.6975118 2022

  71. [71]

    Differentiable quantum generative model 14

  72. [72]

    IQP circuits and forrelation 15

  73. [73]

    Additional details of the DQGM model 18 C

    MMD loss, Pauli-Z expectations, and the Walsh–Hadamard transform 16 B. Additional details of the DQGM model 18 C. Derivation of the marginal estimator 19

  74. [74]

    Problem and circuit 19

  75. [75]

    Proof of Proposition 1 21

  76. [76]

    Additional details of numerical results 23

    Toy example 22 D. Additional details of numerical results 23

  77. [77]

    Hardware and software information 23

  78. [78]

    Estimator accuracy and resource scaling 23

  79. [79]

    Univariate benchmark training details 25

  80. [80]

    Hardware run details 26

Showing first 80 references.