Pith. sign in

REVIEW 2 major objections 4 minor 32 references

Realizing a quantum generative adversarial network using a programmable superconducting processor

T0 review · 2 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Programmable superconducting processor runs a multi-qubit quantum GAN, with the generator copying a target mixed state to 0.999 fidelity and learning the XOR gate to 0.927.

desk verdict First multi-qubit QGAN with a Hadamard-test quantum gradient on a superconducting processor, but the XOR benchmark as printed is internally inconsistent and needs correction. read the letter →

arxiv 2009.12827 v1 pith:3ARLZXGL submitted 2020-09-27 quant-ph

Kaixuan Huang , Zheng-An Wang , Chao Song , Kai Xu , Hekang Li , Zhen Wang , Qiujiang Guo , Zixuan Song
show 6 more authors
Zhi-Bo Liu Dongning Zheng Dong-Ling Deng H. Wang Jian-Guo Tian Heng Fan
This is my paper · ORCID
classification quant-ph
keywords quantumgenerativeadversarialnetworksuperconductingprocessorHadamardtestgradientNashequilibriumparameterizedcircuitsXORgatetransmonqubits
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

Generative adversarial networks pit a generator against a discriminator that tries to tell real from fake data; a quantum GAN replaces both players with tunable quantum circuits. This paper reports a multi-qubit implementation on a programmable superconducting processor in which the training gradient itself is computed by a quantum subroutine, not estimated on a classical computer. The generator learns to reproduce a target single-qubit mixed state with fidelity 0.999, and it learns the classical XOR gate with an average state fidelity of 0.927. The demonstration matters because it shows that near-term quantum hardware with entanglement can run the full adversarial learning loop automatically, a step toward practical quantum machine learning.

What carries the argument

The load-bearing mechanism is the Hadamard-test quantum gradient subroutine: to update a rotation angle $\theta$, the circuit inserts a controlled rotation after the single-qubit gate containing $\theta$ and a controlled-$Z$ gate at the end, then reads the auxiliary qubit $Q_0$. The identity $\partial\langle\sigma_1^z\rangle/\partial\theta = -\langle\sigma_0^z\rangle$ turns one ancilla measurement into an analytic gradient of the discriminator's score, halving the number of circuit executions compared with parameter-shift rules. The other component is the hardware-efficient ansatz: multi-qubit entangling gates $U_{\rm ENT}$ (fixed two- and three-qubit interactions mediated by the bus resonator) interleaved with $X$ and $Z$ rotations, so the circuits use native gates rather than exact textbook logic. The training alternates between maximizing the loss $V$ for the discriminator and minimizing $V^2$ for the generator.

What would settle it

Run the same QGAN procedure many times from fresh random initializations and compare the final fidelity distributions: if typical runs scatter well below 0.999 and 0.927, or if extending the iteration limits changes the results substantially, the reported values are not stable equilibrium outcomes. A simpler check is to continue training past the reported stopping point and see whether the loss keeps oscillating without further improvement, which would indicate the stopping rule, not convergence, produced the headline numbers.

Watch

Extended reading notes

Core claim

The paper's central claim is that a programmable superconducting processor can run a full quantum generative adversarial network in which both the generator and the discriminator are parameterized quantum circuits, trained alternately by gradients obtained from a Hadamard-test quantum subroutine. On the processor, five transmon qubits connected through a central bus provide the entanglement needed for hardware-efficient circuits made of multi-qubit entangling gates interleaved with single-qubit rotations. For a target mixed state the trained generator reaches a state fidelity of about 0.999, and for the classical XOR gate the average state fidelity reaches 0.927 with a truth table matching the gate. The loss function oscillates during training before settling near zero, which the paper interprets as the adversarial process reaching a Nash equilibrium. This extends previous single-qubit QGAN demonstrations by including entanglement and replacing finite-difference classical gradients with a quantum gradient.

Load-bearing premise

The load-bearing premise is that alternating GAN training converges to a stable equilibrium within the preset step limits, so the reported fidelities of 0.999 and 0.927 are equilibrium outcomes rather than transient points; GAN training in general carries no convergence guarantee.

Editorial extensions

If this is right

  • Quantum gradients can replace finite-difference estimation in experimental QGAN training, removing a known source of inaccuracy that slows convergence.
  • Multi-qubit entanglement can be present throughout adversarial learning, so QGANs are not limited to single-qubit data distributions.
  • The hardware-efficient ansatz, built from native entangling gates and rotations, transfers to other quantum platforms that support similar interactions.
  • Reported fidelities of 0.999 for a mixed state and 0.927 for XOR indicate that modest circuit depths suffice for these benchmarks.
  • Because the full training loop runs automatically on the processor, QGAN experiments on near-term devices can be treated as closed-loop optimizations rather than offline classical optimizations.

Reading between the lines

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

  • If 0.927 is an equilibrium value rather than a lucky trajectory, the same recipe should be testable on other small Boolean functions; a systematic scan over functions of two or three bits would show whether average fidelity degrades with input size.
  • The Hadamard-test gradient relies on an ancilla qubit and controlled gates staying coherent while the rest of the circuit runs; on larger devices, that overhead may grow faster than the benefit, so hybrid classical-quantum gradient estimators remain worth comparing.
  • The single training trajectory presented does not by itself demonstrate that alternating GAN updates converge from generic initializations; rerunning with many random seeds would convert the headline fidelities into a distribution rather than a point estimate.
  • The loss oscillation between turns matches the adversarial picture, but it also means early stopping rules are part of the algorithm; defining an equilibrium criterion independent of step limits would make the convergence claim stronger.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper reports an experimental implementation of a quantum generative adversarial network (QGAN) on a programmable five-transmon superconducting processor. Both the generator and the discriminator are parameterized quantum circuits, and gradients are obtained with a Hadamard-test quantum algorithm rather than classical finite differences. Two demonstrations are presented: learning a single-qubit mixed state with a reported state fidelity of 0.999, and learning the classical XOR gate with a reported average state fidelity of 0.927. The authors claim that the adversarial training reaches a Nash equilibrium and that the results demonstrate the feasibility of QGANs on noisy intermediate-scale quantum devices.

Significance. If the central claims hold, this is a useful experimental step beyond the earlier single-qubit QGAN demonstration in Ref. [7]: it uses multiqubit entangling gates in both generator and discriminator, implements an on-chip quantum gradient, and characterizes the learned states by quantum process tomography. The mixed-state result is supported by training curves and by an independent fidelity measurement, and the gradient verification against the analytic CLCU calculation is a valuable check. The XOR benchmark, however, is presented in a way that is internally inconsistent as printed, and the Nash-equilibrium claim is stronger than the evidence provided. The paper should be publishable after these issues are resolved.

major comments (2)
  1. [Fig. 3(a) inset and the XOR gate section] The XOR demonstration is internally inconsistent as printed. The text defines XOR by 00→0, 01→1, 10→1, 11→0 and states that classical 0 and 1 are encoded in |0> and |1>. The inset in Fig. 3(a) labels a row as 'P1 of Q3' with ideal truth table 'XOR 1 0 0 1' and trained generator probabilities 0.89, 0.09, 0.10, 0.93. With the stated encoding, the XOR target is '0 1 1 0' for P1 and '1 0 0 1' for P0; the printed 'XOR 1 0 0 1' therefore matches P0, not P1. Moreover, the score convention S = <σz>/2 + 1/2 defined in the main text equals P0, not P1. Under the literal P1 reading, the printed probabilities match XNOR, not XOR, and the reported average fidelity of 0.927 is inconsistent with the data; under the P0 reading, the data are consistent with XOR and the fidelity is approximately 0.91. The authors must correct the label, state the convention unambiguously, and reconcile the reported fidelity with the displayed numbers.
  2. [Section on QGAN training; Figs. 2(a) and 3(a)] The claim that the training 'arrives at a Nash equilibrium point' is not supported by the evidence presented. The loss function oscillates and approaches zero in a single training run with preset iteration caps (50 steps for D and 100 for G in the mixed-state case; 50 each for the XOR case), but GAN min-max training has no general convergence guarantee and can cycle without reaching an equilibrium. No convergence criterion, multiple random initializations, or quantitative equilibrium test is reported. The loss approaching zero only indicates that the discriminator cannot easily distinguish the generator output at the end of the run. Please either soften the conclusion to an approximate or empirical equilibrium or provide additional evidence for convergence.
minor comments (4)
  1. [Abstract and Section 1] The phrase 'experiment realization' should read 'experimental realization', and 'a even simpler method' should read 'an even simpler method'.
  2. [Fig. 2 caption] The word 'matirces' should be 'matrices'.
  3. [Fig. 3 caption] There is a duplicated 'of' in 'Tracking of of the loss function', and 'construcing' should be 'constructing'.
  4. [Introduction and Conclusion] The statements about exponential advantages and 'practical quantum supremacy' are forward-looking and should be explicitly framed as speculative possibilities, since the experiment does not claim or demonstrate quantum advantage over classical methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the experimental QGAN benchmarks are measured against independent targets and the gradient method is verified externally.

full rationale

The paper reports an experimental implementation rather than a derivation, and its central claims do not reduce to their inputs by construction. The mixed-state benchmark defines a target density matrix rho_R by a fixed circuit (Fig. 2 inset) and measures the fidelity F(rho_R, rho_G) after training; the success metric is independent of the loss function and of the fitted rotation angles. The XOR benchmark compares the trained generator's output probabilities against the stated classical truth table 00->0, 01->1, 10->1, 11->0 with |0>,|1> encoding, and reports an average state fidelity; this is a measured benchmark, not a quantity fitted into the model. The quantum-gradient subroutine is validated against the independent CLCU parameter-shift method (Fig. S5), so the gradient is not assumed circularly. The Nash-equilibrium language is an interpretation of the observed loss oscillation, not a fitted parameter or a derived uniqueness claim. The paper cites prior work by overlapping authors for device characteristics and for the earlier single-qubit QGAN, but these citations are background and hardware characterization, not load-bearing derivations of the present results. The apparent inconsistency in the Fig. 3 inset label ('P1 of Q3' versus the score convention S = <sigma_z>/2 + 1/2 = P0) is an internal consistency/correctness issue about which bit probability the plotted numbers denote, not a circularity: even on the reading that would invalidate the XOR claim, the claim is contradicted by the data rather than forced by the model's construction. No fitted input is renamed as a prediction, no self-citation is invoked to forbid alternatives, and no ansatz is smuggled in via citation. The experimental conclusions are therefore self-contained with respect to the circularity concerns in scope.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The free parameters are training hyperparameters (learning rates, circuit depths) chosen by hand or simulation. The axioms are standard background in quantum computing, the Hadamard test gradient formula, and assumptions about ansatz expressibility and GAN convergence.

free parameters (4)
  • Discriminator learning rate alpha_D = 0.8 (mixed state), 1.0 (XOR)
    Chosen by hand; not optimized. Affects training dynamics and convergence.
  • Generator learning rate alpha_G = 0.6 (mixed state), 1.5 (XOR)
    Chosen by hand; not optimized. Affects training dynamics and convergence.
  • Generator circuit depth = 1 layer (mixed state), 2 layers (XOR)
    Selected via numerical simulations to balance learning fidelity and efficiency. A structural hyperparameter.
  • Discriminator circuit depth = 3 layers (both tasks)
    Selected via numerical simulations to balance learning fidelity and efficiency. A structural hyperparameter.
assumptions (4)
  • standard math The Hadamard test gradient formula (Eq. S2 in the supplementary) correctly gives the partial derivative of the expectation value with respect to a circuit parameter.
    Adopted from Dallaire-Demers and Killoran (2018) and Mitarai and Fujii (2019). The paper verifies it experimentally on a simple circuit, but the proof is not in this paper.
  • domain assumption The parameterized quantum circuits for G and D are expressive enough to represent the target data distributions at the chosen depths.
    The paper uses numerical simulation to decide circuit depths, implicitly assuming the hardware-efficient ansatz covers the target states and functions.
  • domain assumption The alternating training converges to a Nash equilibrium within the preset iteration limits.
    GAN training lacks a general convergence guarantee, and the paper offers no proof; the loss trajectory is the only evidence.
  • domain assumption The device operation, including gate fidelities, all-to-all connectivity, and readout, is stable and accurate enough for the reported results.
    The supplementary provides stability data, but no error bars are given for the final fidelities, and single runs are shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Realizing a quantum generative adversarial network using a programmable superconducting processor." pith.science (2026). https://pith.science/paper/3ARLZXGL

@misc{pith2026200912827,
  author       = {Pith},
  title        = {Pith review of: Realizing a quantum generative adversarial network using a programmable superconducting processor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ARLZXGL}},
  note         = {Machine review of arXiv:2009.12827}
}
read the original abstract

Generative adversarial networks are an emerging technique with wide applications in machine learning, which have achieved dramatic success in a number of challenging tasks including image and video generation. When equipped with quantum processors, their quantum counterparts--called quantum generative adversarial networks (QGANs)--may even exhibit exponential advantages in certain machine learning applications. Here, we report an experimental implementation of a QGAN using a programmable superconducting processor, in which both the generator and the discriminator are parameterized via layers of single- and multi-qubit quantum gates. The programmed QGAN runs automatically several rounds of adversarial learning with quantum gradients to achieve a Nash equilibrium point, where the generator can replicate data samples that mimic the ones from the training set. Our implementation is promising to scale up to noisy intermediate-scale quantum devices, thus paving the way for experimental explorations of quantum advantages in practical applications with near-term quantum technologies.

Figures

Figures reproduced from arXiv: 2009.12827 by the authors.

Figure 1
Figure 1. QGAN algorithm and its implementation. a, Overview of the QGAN logic. b, Sketch of the superconducting processor used to implement the QGAN algorithm, where the five qubits, Q0 to Q4, are interconnected by the central bus resonator. c, An instance of the experimental sequences for fulfilling the QGAN algorithm to learn the classical XOR gate. Both G and D are parameterized quantum circuits consisting of layers of th… view at source ↗
Figure 2
Figure 2. QGAN performance in learning a mixed state. a, Tracking of the loss function V , the output scores S D,R/G, and the state fidelity between R/G’s output states F during the adversarial training procedure with learning rates αD = 0.8 and αG = 0.6. The alternate training stages of D and G are marked by red and white regions, respectively. In each stage, the maximum step number is limited to 50 for D and 100 for G. Inse… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [7]

    Hu, S.-H

    L. Hu, S.-H. Wu, W. Cai, Y. Ma, X. Mu, Y. Xu, H. Wang, Y. Song, D.-L. Deng, C.-L. Zou, and L. Sun, Science ad- vances 5, eaav2761 (2019)

    work page 2019

  2. [1]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, in Advances in neural information processing systems (2014) pp. 26722680

    work page 2014

  3. [2]

    Creswell, T

    A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, IEEE Signal Processing Magazine 35, 53 (2018)

    work page 2018

  4. [3]

    Lloyd and C

    S. Lloyd and C. Weedbrook, Phys. Rev. Lett. 121, 040502(2018)

    work page 2018

  5. [4]

    Dallaire-Demers and N

    P.-L. Dallaire-Demers and N. Killoran, Phys. Rev. A 98, 012324 (2018)

    work page 2018

  6. [5]

    J. Zeng, Y. Wu, J.-G. Liu, L. Wang, and J. Hu, Phys. Rev. A 99, 052306 (2019)

    work page 2019

  7. [6]

    Zoufal, A

    C. Zoufal, A. Lucchi, and S. Woerner, npj Quantum In- formation 5, 103 (2019)

    work page 2019

  8. [8]

    Variational quantum generators: Generative adversarial quantum machine learning for continuous distributions

    J. Romero and A. Aspuru-Guzik, arXiv:1901.00848 (2019)

    work page Pith review arXiv 2019

Show all 32 references
  1. [9]

    Anand, J

    A. Anand, J. Romero, M. Degroote, and A. Aspuru-Guzik, arXiv:2006.01976 (2020)

    2020 arXiv

  2. [10]

    Carleo, I

    G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto, and L. Zdeborov, Rev. Mod. Phys. 91, 045002 (2019)

    2019

  3. [11]

    S. D. Sarma, D.-L. Deng, and L.-M. Duan, Physics Today 72, 48 (2019)

    2019

  4. [12]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Nature 549, 195 (2017)

    2017

  5. [13]

    Dunjko and H

    V. Dunjko and H. J. Briegel, Rep. Prog. Phys. 81, 074001 (2018)

    2018

  6. [14]

    Gao, Z.-Y

    X. Gao, Z.-Y. Zhang, and L.-M. Duan, Science advances 4, eaat9004 (2018)

    2018

  7. [15]

    Harrow and J

    A. Harrow and J. Napp, arXiv:1901.05374 (2019)

    2019 arXiv

  8. [16]

    Havlicek, A

    V. Havlicek, A. D. Corcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Nature 567, 209 (2019)

    2019

  9. [17]

    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, B. Burkett, Y. Chen, Z. Chen, B. Chiaro, R. Collins, W. Courtney, A. Dunsworth, E. Farhi, B. Foxen, A. Fowler, C. Gidney, M. Giustina, R. Graff, K. Guerin, ...

    2019

  10. [18]

    Jarrod, R., B

    M. Jarrod, R., B. Sergio, S. Vadim, N., B. Ryan, and N. Hartmut, Nat. Commun. 9 (2018), 10.1038/s41467-018- 07090-4

    2018 doi

  11. [19]

    Cerezo, A

    M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, arXiv:2001.00550 (2020)

    2020 arXiv

  12. [20]

    Skolik, J

    A. Skolik, J. R. McClean, M. Mohseni, P. Smagt, and M. Leib, arXiv:2006.14904 (2020)

    2020 arXiv

  13. [21]

    Huembeli and A

    P. Huembeli and A. Dauphin, arXiv:2008.02785 (2020)

    2020 arXiv

  14. [22]

    Schuld, V

    M. Schuld, V. Bergholm, C. Gogolin, J. Izaac, and N. Killoran, Phys. Rev. A 99, 032331 (2019)

    2019

  15. [23]

    Mitarai and K

    K. Mitarai and K. Fujii, Physical Review Research 1, 013006 (2019)

    2019

  16. [24]

    Supplementary-Materials

  17. [25]

    C. Song, K. Xu, H. Li, Y.-R. Zhang, X. Zhang, W. Liu, Q. Guo, Z. Wang, W. Ren, J. Hao, H. Feng, H. Fan, D. Zheng, D.-W. Wang, H. Wang, and S.-Y. Zhu, Science 365, 574 (2019)

    2019

  18. [26]

    Realizing a quantum generative adversarial network using a programmable superconducting processor

    J. Preskill, Quantum 2, 79 (2018). 6 Supplementary Material for “Realizing a quantum generative adversarial network using a programmable superconducting processor” DETAILS ABOUT THE QUANTUM DEVICE Our experimental device is a superconducting circuit consisting of 20 transmon q...

    2018

  19. [27]

    C. Song, K. Xu, H. Li, Y.-R. Zhang, X. Zhang, W. Liu, Q. Guo, Z. Wang, W. Ren, J. Hao, H. Feng, H. Fan, D. Zheng, D.-W. Wang, H. Wang, and S.-Y. ZhuScience 365, 574-577 (2019)

    2019

  20. [28]

    C. Song, K. Xu, W. Liu, C.-P Yang, S.-B. Zheng, H. Deng, Q. Xie, K. Huang, Guo, L. Zhang, P. Zhang, D. Xu, D. Zheng, X. Zhu, H. Wang, Y.-A. Chen, C.-Y. Lu, S. Han, and J-W. Pan, Phys. Rev. Lett. 119, 180511 (2017)

    2017

  21. [29]

    Motzoi, J

    F. Motzoi, J. M. Gambetta, P. Rebentrost, and F. K. Wil- helm.Phys. Rev. Lett. 103, 110501 (2009)

    2009

  22. [30]

    Guo, S.-B

    Q. Guo, S.-B. Zheng, J. Wang, C. Song, P. Zhang, K. Li, W. Liu, H. Deng, K. Huang, D. Zheng, X. Zhu, H. Wang, C.-Y. Lu, and J.-W. Pan, Phys. Rev. Lett. 121, 130501 (2018)

    2018

  23. [31]

    Dallaire-Demers, N

    P.-L. Dallaire-Demers, N. Killoran. Phys. Rev. A 98, 012324 (2018)

    2018

  24. [32]

    Schuld, V

    M. Schuld, V. Bergholm, C. Gogolin, J. Izaac and N. Kil- loran. Phys. Rev. A 99, 032331 (2019)

    2019

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.