Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Mixed-State Quantum Denoising Diffusion Probabilistic Model

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read A mixed-state quantum diffusion model generates ensembles using depolarizing noise instead of scrambling unitaries.

desk verdict Useful NISQ-oriented extension of QuDDPM that replaces scrambling unitaries with depolarizing channels; promising small-scale numerics, but the 'comparable performance' claim lacks a direct baseline and the trainability argument is imported from the pure-state setting. read the letter →

arxiv 2411.17608 v2 pith:6ZN4ALH4 submitted 2024-11-26 quant-ph cs.AIcs.LG

classification quant-phcs.AIcs.LG
keywords quantumgenerativemodeldenoisingdiffusiondepolarizingchannelmixedstateparameterizedcircuitsuperfidelitynoiseschedulingensemble
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

The paper proposes a quantum denoising diffusion model that replaces the high-fidelity scrambling unitaries of the original QuDDPM with depolarizing noise channels in the forward process and parameterized circuits with projective measurements in the backward process. The goal is to make quantum generative diffusion practical on near-term devices while extending generation from pure states to arbitrary mixed-state ensembles. The authors report that the model performs comparably to the scrambling-based approach on the tested tasks and successfully generates clustered, circular, and many-body phase ensembles. If the central claim holds, near-term quantum generative models could trade expensive scrambling operations for stepwise depolarization with shallow circuits.

What carries the argument

The forward depolarizing channel $\Phi_{t+1}(\rho)=(1-q_{t+1})\rho+q_{t+1}I/d$ is the noise provider that drives any ensemble toward the maximally mixed state, while the backward process is a chain of hardware-efficient parameterized circuits followed by projective measurements on ancilla qubits. Superfidelity, $G(\rho,\sigma)=\operatorname{Tr}(\rho\sigma)+\sqrt{[1-\operatorname{Tr}(\rho^2)][1-\operatorname{Tr}(\sigma^2)]}$, serves as the ensemble-similarity kernel for squared maximum mean discrepancy and Wasserstein costs. The cosine-exponent schedule $q_t=(1-\alpha_t/\alpha_{t-1})^k$, with $\alpha_t$ defined from a squared cosine, controls how slowly the initial state is depolarized. Together these components carry the argument that stepwise denoising can be learned without scrambling unitaries.

What would settle it

Train the same backward ansatz on the same task with the depolarizing forward process at a larger system size, such as eight or twelve qubits, and check whether the MMD or superfidelity loss to the target ensemble falls to the small values reported at four qubits; a plateau far above those values would show the shallow-circuit reverse maps are not expressive enough. As a smaller check, compare the learned one-step backward channel against the exact reverse depolarizing map for a single qubit, since a mismatch there would indicate the parameterized circuit cannot represent the required inverse.

Watch

Extended reading notes

Core claim

The central claim is that a forward diffusion process built from depolarizing channels, which push any initial ensemble toward the maximally mixed state, can be paired with a backward denoising process learned by a sequence of shallow parameterized circuits with projective measurements. This eliminates the need for high-fidelity scrambling unitaries that limited the original model to pure states and made implementation difficult. The paper further claims that a cosine-exponent noise schedule, a single-qubit Haar random ancilla, and superfidelity-based MMD or Wasserstein cost functions improve convergence, and it demonstrates the full model on one-qubit clustered and circular ensembles and on four-qubit transverse-field Ising ensembles.

Load-bearing premise

The load-bearing premise is that the reverse of the depolarizing forward process can be represented and learned by the chosen shallow parameterized circuits with projective measurements; if that representation fails, the backward circuits cannot undo the noise and the model's central claim collapses.

Editorial extensions

If this is right

  • Near-term implementations no longer require high-fidelity scrambling unitaries; depolarizing channels and shallow parameterized circuits suffice for the demonstrated tasks.
  • The model generates both pure and mixed quantum state ensembles, extending diffusion-based quantum generation beyond pure states.
  • Increasing the number of diffusion steps with shallow circuits outperforms increasing the number of ancilla qubits for the tested many-body phase task.
  • Cosine-exponent scheduling, especially cosine-square, preserves the target ensemble's purity longer and is needed for convergence on multi-qubit tasks.
  • Haar-random single-qubit ancillas introduce sample diversity and reach performance comparable to zero-state ancillas while using fewer parameters.

Reading between the lines

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

  • If the central claim generalizes, the practical bottleneck shifts from the gate fidelity of scrambling unitaries to the number of sequential circuit executions and classical optimization steps.
  • Because the forward depolarizing channel is unital and commutative, a testable extension would replace learned backward circuits with analytically constructed reverse channels and compare the generated ensembles.
  • The cosine-exponent schedule suggests that noise-schedule design, not architecture alone, determines multi-qubit diffusion performance; a principled rule for choosing the exponent $k$ could be derived from purity-decay targets.
  • Superfidelity-based costs avoid full tomography, so they may serve as a practical proxy for other quantum generative tasks, although their hardware sample overhead remains open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes MSQuDDPM, a mixed-state generalization of QuDDPM in which the forward diffusion process is implemented by repeated depolarizing channels and the backward process by a sequence of parameterized quantum circuits with projective measurements. The authors introduce a cosine-exponent noise schedule (Eqs. (2)-(3)), single-qubit Haar-random ancilla initialization, and superfidelity-based MMD and Wasserstein cost functions. The paper reports three numerical tasks -- 1-qubit clustered states, 1-qubit circular states, and a 4-qubit transverse-field Ising model ground-state ensemble -- and claims that MSQuDDPM eliminates the need for scrambling unitaries while achieving performance comparable to the original QuDDPM at lower implementation complexity.

Significance. The idea of replacing scrambling unitaries with depolarizing channels is natural and could be practically useful if the comparative performance claim holds. The paper's strengths are its explicit circuit construction (Fig. 7), the use of standard ensemble metrics, and the small-scale numerical evidence that the trained model outperforms the maximally-mixed-state baseline and that cosine-square scheduling markedly improves convergence in the 4-qubit task. However, the evidence currently supports only a qualitative proof-of-concept: the advertised comparison with QuDDPM is not benchmarked, and the trainability argument is imported from a pure-state scrambling analysis rather than analyzed for the mixed-state depolarizing setting.

major comments (4)
  1. [Section I and Table I] The abstract and Section I state that MSQuDDPM 'achieves comparable performance in the same tasks' as the original QuDDPM [25], but Table I contains no QuDDPM baseline. The reported numbers (e.g., F_gen=0.9873, Wass_gen=0.0151, Mx_gen=0.940) are absolute values whose effective reference is the maximally mixed state (e.g., Mx_data=0.951 vs Mx_gen=0.940). Since reducing implementation complexity while preserving performance is the paper's stated motivation, the missing direct comparison leaves the central claim unverified even at the demonstrated system sizes. Please add QuDDPM runs on the same tasks or rephrase the claim as one of qualitative generation quality.
  2. [Section II.B] The paragraph beginning 'Following the QuDDPM's trainability analysis [25]' transfers the poly-depth-to-log-depth splitting and barren-plateau avoidance argument from Ref. [25] to MSQuDDPM. The two settings are different in a load-bearing way: here the forward map is a non-unitary depolarizing channel, the intermediate target ensembles are mixed, and the exact inverse of the forward channel is not a CPTP map; the backward circuit must realize a stochastic Bayesian update through projective measurement branches. No expressibility result, convergence guarantee, or scaling test for this setting is given, and Section V concedes that simulations reach only four qubits. Because the central claim that scrambling unitaries can be eliminated depends on the backward circuits' ability to learn the denoising map, I ask for numerical scaling evidence on the mixed-state setting or an explicit reframing of this paragraph as an empirical observation.
  3. [Section IV.A, Fig. 4] The comparison in Fig. 4 between the proposed configuration (T=6, L=12, n_a=2, six actual qubits) and the benchmark (T=2, L=21, n_a=6, ten actual qubits) is used to conclude that more diffusion steps are preferable to more ancilla qubits. The two configurations differ simultaneously in T, L, n_a and circuit width, so the benchmark's failure (MMD=0.9325) may be due to the larger circuit width rather than to the smaller number of diffusion steps. A controlled study that varies one factor at a time is needed to support this design principle.
  4. [Section IV.B and Appendix C] The cosine-exponent schedule in Eqs. (2)-(3) is presented as a key improvement, but its parameters k and epsilon are fixed by hand (k=1 or k=2) and Appendix C states that 'a strategy for selecting an appropriate value of k' remains future work. Because the reported advantage of cosine-square scheduling over linear/cosine scheduling is based on these particular choices, a sensitivity analysis over k and epsilon is needed to show that the improvement is not task-specific tuning.
minor comments (4)
  1. [Reproducibility] No code repository or random seeds are provided, so the very small reported uncertainties (e.g., F_gen=0.9873 +/- 10^-5) are not independently reproducible; a methods paper of this kind would benefit from releasing the simulation code and seeds.
  2. [Table I] Several entries (e.g., Wass_data and Wass_gen for the circular task, and MMD entries for the many-body task) are reported without uncertainties, making it difficult to assess run-to-run variability.
  3. [Section IV.C and Fig. 6] The discussion around Fig. 6 claims that Haar ancillas achieve 'comparable' performance with 'fewer total parameters', but parameter counts for the clustered and circular tasks are not given; please report them or refer explicitly to the parameter counts in Fig. 4.
  4. [General] Minor typographical issues remain, e.g., 'accomodate' in Section III; a final proofreading pass is recommended.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; minor self-citation in the trainability argument.

full rationale

MSQuDDPM's derivation chain is a standard generative training loop: the forward depolarizing channels in Eq. (1) produce intermediate ensembles {ρ_m}; the backward PQC is trained to match those ensembles via the MMD/Wasserstein losses in Eqs. (5)-(7); and performance is assessed on fresh maximally-mixed test inputs. The generated states are not defined as the forward samples, and the numerical results are not forced: the linear-schedule model fails (M_x,gen=0.427 vs 0.940 for cosine-square in Table I), which shows that the successful runs carry information beyond the construction. The only self-referential element is the trainability argument in Sec. II.B, which imports QuDDPM [25] ('Following the QuDDPM's trainability analysis [25], suppose the model requires at least a polynomial circuit depth...'). This is an explicit assumption/analogy from the same group's prior work, not an equation-level reduction of the target result, so it raises a self-citation concern but not circularity. No prediction in the paper is equivalent to its input by construction; score 2 reflects the minor self-citation.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The model's success rests on standard quantum-information facts (depolarizing channels, superfidelity) plus several untested assumptions: the existence of a shallow PQC that reverses each depolarizing step and the transfer of QuDDPM's trainability analysis to this setting. The manually chosen hyperparameters (k, T, L, na, epsilon) are part of the contribution and are not derived, so they are recorded as free parameters. No new physical entities are introduced.

free parameters (5)
  • cosine-exponent exponent k = 1 or 2; k=2 for final tasks
    Eq. (2) defines q_t = (1 - alpha_t/alpha_{t-1})^k; k=2 (cosine square) was chosen after comparing linear, cosine, and cosine-square schedules on the same benchmark tasks (Section IV B).
  • schedule offset epsilon = small; exact value not reported
    Offset inside f(t) in Eq. (3); introduced to avoid numerical singularities, not derived from data.
  • diffusion steps T = 4 or 6 per task
    Manual hyperparameter; Table I uses T=6 for clustered and many-body tasks and T=4 for one circular-state run.
  • PQC layers L = 4, 8, or 12 per task
    Circuit depth per backward step is chosen by hand; Table I lists L values for each task.
  • ancilla count na = 2 (6 in benchmark)
    Number of auxiliary qubits is manually chosen; Section IV A argues more diffusion steps are better than more ancillas.
assumptions (5)
  • standard math Depolarizing channels with parameter q map rho to (1-q)rho + q I/d and converge to the maximally mixed state.
    Eq. (1) defines the forward process; this is a standard property of the depolarizing channel.
  • domain assumption Superfidelity is a valid, efficiently estimable upper bound to fidelity.
    Section II C relies on Ref. [37]; the paper assumes it can be estimated from fewer copies than full tomography without providing a sample-complexity analysis.
  • domain assumption MMD and Wasserstein losses with superfidelity kernel are appropriate training objectives for quantum state ensembles.
    Section II C defines these losses; no proof is given that their minima correspond to correct ensemble generation.
  • ad hoc to paper The reverse map of the depolarizing forward process can be represented by the chosen parameterized circuits with projective measurements.
    Section II B constructs the backward process as a VQC and assumes it can learn each denoising step; no expressibility or convergence theorem is provided.
  • ad hoc to paper The trainability analysis from QuDDPM, including polynomial depth splitting to log(n) and avoidance of barren plateaus, carries over to the mixed-state depolarizing setting.
    Section II B imports this from Ref. [25] without derivation for depolarizing channels and mixed states.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixed-State Quantum Denoising Diffusion Probabilistic Model." pith.science (2026). https://pith.science/paper/6ZN4ALH4

@misc{pith2026241117608,
  author       = {Pith},
  title        = {Pith review of: Mixed-State Quantum Denoising Diffusion Probabilistic Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZN4ALH4}},
  note         = {Machine review of arXiv:2411.17608}
}
read the original abstract

Generative quantum machine learning has gained significant attention for its ability to produce quantum states with desired distributions. Among various quantum generative models, quantum denoising diffusion probabilistic models (QuDDPMs) [Phys. Rev. Lett. 132, 100602 (2024)] provide a promising approach with stepwise learning that resolves the training issues. However, the requirement of high-fidelity scrambling unitaries in QuDDPM poses a challenge in near-term implementation. We propose the \textit{mixed-state quantum denoising diffusion probabilistic model} (MSQuDDPM) to eliminate the need for scrambling unitaries. Our approach focuses on adapting the quantum noise channels to the model architecture, which integrates depolarizing noise channels in the forward diffusion process and parameterized quantum circuits with projective measurements in the backward denoising steps. We also introduce several techniques to improve MSQuDDPM, including a cosine-exponent schedule of noise interpolation, the use of single-qubit random ancilla, and superfidelity-based cost functions to enhance the convergence. We evaluate MSQuDDPM on quantum ensemble generation tasks, demonstrating its successful performance.

Figures

Figures reproduced from arXiv: 2411.17608 by the authors.

Figure 1
Figure 1. FIG. 1. The model architecture and training strategy of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Bloch sphere representations of (a) forward and (b) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. FIG. 4. The MMD distance decay of models with different dif [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: FIG. 5. Comparison of average purity decay for forward sam [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. MSQuDDPM test samples using different ancilla [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The detailed quantum circuits for the forward and [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. The mean purity decay of samples throughout the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Quantum Reversibility Meets Classical Reverse Diffusion

    quant-ph 2025-10 conditional novelty 4.0 of 10

    The semiclassical limit of the Petz-reversed Lindblad equation reproduces the Bayes-rule reverse-time diffusion equation, with the reference state's Wigner function playing the role of the classical score distribution.

Reference graph

Works this paper leans on

54 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [25]

    Khoshaman, W

    A. Khoshaman, W. Vinci, B. Denis, E. Andriyash, H. Sadeghi, and M. H. Amin, Quantum variational au- toencoder, Quantum Sci. Technol. 4, 014001 (2018)

  2. [1]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum machine learning, Na- ture 549, 195 (2017)

  3. [2]

    Farhi, J

    E. Farhi, J. Goldstone, and S. Gutmann, A quantum approximate optimization algorithm, arXiv:1411.4028 (2014)

  4. [3]

    M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, Quantum boltzmann machine, Phys. Rev. X 9 8, 10.1103/physrevx.8.021050 (2018)

  5. [4]

    Rebentrost, M

    P. Rebentrost, M. Mohseni, and S. Lloyd, Quantum sup- port vector machine for big data classification, Phys. Rev. Lett. 113, 130503 (2014)

  6. [5]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, et al. , Variational quantum algorithms, Nat. Rev. Phys. 3, 625 (2021)

  7. [6]

    Woerner and D

    S. Woerner and D. J. Egger, Quantum risk analysis, npj Quantum Inf. 5, 15 (2019)

  8. [7]

    Maheshwari, B

    D. Maheshwari, B. Garcia-Zapirain, and D. Sierra-Sosa, Quantum machine learning applications in the biomedi- cal domain: A systematic review, Ieee Access 10, 80463 (2022)

Show all 54 references
  1. [8]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, Generative adversarial networks, Commun. ACM63, 139 (2020)

  2. [9]

    D. P. Kingma and M. Welling, Auto-encoding variational bayes, arXiv:1312.6114 (2013)

  3. [10]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, Attention is all you need, in Advances in Neural Infor- mation Processing Systems , Vol. 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vish- w...

  4. [11]

    Z. Wang, Q. She, and T. E. Ward, Generative adversarial networks in computer vision: A survey and taxonomy, ACM Comput. Surv. 54, 1 (2021)

  5. [12]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Alt- man, S. Anadkat, et al. , Gpt-4 technical report, arXiv:2303.08774 (2023)

  6. [13]

    X. Tan, T. Qin, F. Soong, and T.-Y. Liu, A sur- vey on neural speech synthesis (2021), arXiv:2106.15561 [eess.AS]

  7. [14]

    Dhariwal, H

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, Jukebox: A generative model for music (2020), arXiv:2005.00341 [eess.AS]

  8. [15]

    Combining these ele- ments with backward VQC through their quantum coun- terparts could enable even more complex quantum state generation

    and latent diffusion model [43]. Combining these ele- ments with backward VQC through their quantum coun- terparts could enable even more complex quantum state generation. Furthermore, while Haar random states have shown their potential as auxiliary qubits in MSQuDDPM training...

  9. [16]

    Zhang, M

    Q. Zhang, M. Tao, and Y. Chen, gddim: Generalized denoising diffusion implicit models, arXiv:2206.05564 (2022)

  10. [17]

    J. Ho, A. Jain, and P. Abbeel, Denoising diffusion prob- abilistic models, Adv. Neural Inf. Process. Syst. 33, 6840 (2020)

  11. [18]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, Deep unsupervised learning using nonequi- librium thermodynamics, in International conference on machine learning (PMLR, 2015) pp. 2256–2265

  12. [19]

    Kingma, T

    D. Kingma, T. Salimans, B. Poole, and J. Ho, Varia- tional diffusion models, Adv. Neural Inf. Process. Syst. 34, 21696 (2021)

  13. [20]

    Lloyd and C

    S. Lloyd and C. Weedbrook, Quantum generative adver- sarial learning, Phys. Rev. Lett. 121, 040502 (2018)

  14. [21]

    L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, and M.-H. Yang, Diffusion models: A comprehensive survey of methods and applications, ACM Comput. Surv. 56, 1 (2023)

  15. [22]

    Huang, Y

    H.-L. Huang, Y. Du, M. Gong, Y. Zhao, Y. Wu, C. Wang, S. Li, F. Liang, J. Lin, Y. Xu, et al., Experimental quan- tum generative adversarial networks for image genera- tion, Phys. Rev. Applied 16, 024051 (2021)

  16. [23]

    Zoufal, A

    C. Zoufal, A. Lucchi, and S. Woerner, Quantum genera- tive adversarial networks for learning and loading random distributions, npj Quantum Inf. 5, 103 (2019)

  17. [24]

    M. L. Wall, M. R. Abernathy, and G. Quiroz, Generative machine learning with tensor networks: Benchmarks on near-term quantum computers, Phys. Rev. Research 3, 023010 (2021)

  18. [26]

    Cacioppo, L

    A. Cacioppo, L. Colantonio, S. Bordoni, and S. Giagu, Quantum diffusion models, arXiv:2311.15444 (2023)

  19. [27]

    Zhang, P

    B. Zhang, P. Xu, X. Chen, and Q. Zhuang, Generative quantum machine learning via denoising diffusion prob- abilistic models, Phys. Rev. Lett. 132, 100602 (2024)

  20. [28]

    Chen and Q

    C. Chen and Q. Zhao, Quantum generative diffusion model, arXiv:2401.07039 (2024)

  21. [29]

    Parigi, S

    M. Parigi, S. Martina, and F. Caruso, Quantum-noise- driven generative diffusion models, Adv. Quantum Tech- nol. , 2300401 (2024)

  22. [30]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Barren plateaus in quantum neural net- work training landscapes, Nat. Commun. 9, 4812 (2018)

  23. [31]

    Y. Zhu, T. Chen, E. A. Theodorou, X. Chen, and M. Tao, Quantum state generation with structure-preserving dif- fusion model, arXiv:2404.06336 (2024)

  24. [32]

    King, The capacity of the quantum depolarizing chan- nel, IEEE Trans

    C. King, The capacity of the quantum depolarizing chan- nel, IEEE Trans. Inf. Theory 49, 221 (2003)

  25. [33]

    Cerezo, A

    M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Cost function dependent barren plateaus in shal- low parametrized quantum circuits, Nat. Commun. 12, 10.1038/s41467-021-21728-w (2021)

  26. [34]

    A. Q. Nichol and P. Dhariwal, Improved denoising diffu- sion probabilistic models, in International conference on machine learning (PMLR, 2021) pp. 8162–8171

  27. [35]

    M. A. Nielsen and I. L. Chuang, Quantum Computa- tion and Quantum Information: 10th Anniversary Edi- tion (Cambridge University Press, 2010)

  28. [36]

    Schmale, M

    T. Schmale, M. Reh, and M. G¨ arttner, Efficient quantum state tomography with convolutional neural networks, npj Quantum Inf. 8, 115 (2022)

  29. [37]

    Jozsa, Fidelity for mixed quantum states, J

    R. Jozsa, Fidelity for mixed quantum states, J. Mod. Optic. 41, 2315 (1994)

  30. [38]

    In the learning cycle at t = m + 1 in Fig

    and Wasserstein distance [39] with superfidelity as a kernel function and the distance measure, respectively. In the learning cycle at t = m + 1 in Fig. 1 (a), the model constructs two ensembles, one from the backward model outputs {˜ρm} and the other from the target states {ρ...

  31. [39]

    J. A. Miszczak, Z. Pucha la, P. Horodecki, A. Uhlmann, and K. ˙Zyczkowski, Sub– and super–fidelity as bounds for quantum fidelity (2008), arXiv:0805.2037 [quant-ph]

  32. [40]

    Gretton, K

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch¨ olkopf, and A. Smola, A kernel two-sample test, J. Mach. Learn. Res. 13, 723 (2012)

  33. [41]

    Oreshkov and J

    O. Oreshkov and J. Calsamiglia, Distinguishability mea- sures between ensembles of quantum states, Phys. Rev. A 79, 032336 (2009)

  34. [42]

    Pfeuty, The one-dimensional ising model with a trans- verse field, ANNALS of Physics 57, 79 (1970)

    P. Pfeuty, The one-dimensional ising model with a trans- verse field, ANNALS of Physics 57, 79 (1970)

  35. [43]

    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, A review of barren plateaus in variational quantum computing, arXiv:2405.00781 (2024)

  36. [44]

    Chen, On the importance of noise scheduling for dif- fusion models, arXiv:2301.10972 (2023)

    T. Chen, On the importance of noise scheduling for dif- fusion models, arXiv:2301.10972 (2023)

  37. [45]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, High-resolution image synthesis with latent diffusion models, in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition(2022) 10 pp. 10684–10695

  38. [46]

    Kandala, A

    A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, Hardware- efficient variational quantum eigensolver for small molecules and quantum magnets, Nature 549, 242 (2017)

  39. [47]

    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. Kopf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, Pytorch: An imperative style, high-p...

  40. [48]

    Zhang, J

    S.-X. Zhang, J. Allcock, Z.-Q. Wan, S. Liu, J. Sun, H. Yu, X.-H. Yang, J. Qiu, Z. Ye, Y.-Q. Chen, C.-K. Lee, Y.-C. Zheng, S.-K. Jian, H. Yao, C.-Y. Hsieh, and S. Zhang, TensorCircuit: a Quantum Software Framework for the NISQ Era, Quantum 7, 912 (2023)

  41. [49]

    Flamary, N

    R. Flamary, N. Courty, A. Gramfort, M. Z. Alaya, A. Boisbunon, S. Chambon, L. Chapel, A. Corenflos, K. Fatras, N. Fournier, L. Gautheron, N. T. Gayraud, H. Janati, A. Rakotomamonjy, I. Redko, A. Rolet, A. Schutz, V. Seguy, D. J. Sutherland, R. Tavenard, A. Tong, and T. Vayer, ...

  42. [50]

    Johansson, P

    J. Johansson, P. Nation, and F. Nori, Qutip: An open- source python framework for the dynamics of open quan- tum systems, Comput. Phys. Commun. 183, 1760–1772 (2012)

  43. [51]

    Kingma and J

    D. Kingma and J. Ba, Adam: A method for stochastic optimization, arXiv:1412.6980 (2014)

  44. [52]

    A. C. Wilson, R. Roelofs, M. Stern, N. Srebro, and B. Recht, The marginal value of adaptive gradient meth- ods in machine learning, Adv. Neural Inf. Process. Syst. 30 (2017)

  45. [53]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Decoupled weight decay reg- ularization, arXiv:1711.05101 (2017)

  46. [54]

    Glorot and Y

    X. Glorot and Y. Bengio, Understanding the difficulty of training deep feedforward neural networks, in Pro- ceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , Proceedings of Ma- chine Learning Research, Vol. 9, edited by Y. W. Teh a...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.