Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Divide-and-conquer neural surrogates speed up quantum MCMC by 20x

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 ↗

T0 review · glm-5.2

2026-07-05 02:08 UTC pith:J53UT7NP

load-bearing objection Paper mismatch and quantum MCMC assessment the 3 major comments →

arxiv 2604.20699 v2 pith:J53UT7NP submitted 2026-04-22 astro-ph.CO

Forecasts of CMB E-mode anomalies for AliCPT-1

Jiazheng Dou , Wen Zhao This is my paper
classification astro-ph.CO
keywords alicptanomaliesmodeanomalymodulationpolarizationsimulationsalignment
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 paper proposes a divide-and-conquer neural network surrogate framework to accelerate Markov chain Monte Carlo (MCMC) sampling for large-scale Ising optimization problems under fixed Hamming weight constraints. The core idea is to partition the interaction graph into smaller subgraphs, generate samples from QAOA circuits with an XY mixer on each subgraph, and train conditional MADE (Masked Autoencoder for Distribution Estimation) neural networks to approximate those sample distributions. These trained surrogates then serve as proposal distributions in Metropolis-Hastings MCMC, enabling non-local multi-spin updates that respect the constraint while avoiding the exponentially slow mixing of classical pair-flip (Kawasaki) dynamics. On 3-regular graphs, the method achieves speedup factors of about 20.3 and 7.6 over nearest-neighbor and non-nearest-neighbor Kawasaki dynamics respectively, with the advantage persisting as system size grows to N=512. Applied to MNIST feature selection with N=784, the method yields faster energy convergence and a 2.03% improvement in classification accuracy over classical approaches.

Core claim

The central discovery is that by dividing a large constrained Ising problem into fixed-size subgraphs, generating quantum samples via QAOA with an XY mixer on each subgraph, and training conditional MADE neural networks as surrogate proposal distributions, one can achieve MCMC mixing that is consistently faster than classical Kawasaki pair-flip dynamics and whose advantage scales with system size rather than degrading. The divide-and-conquer strategy makes the quantum circuit size fixed (as small as 16 qubits) while the classical neural network handles the combinatorial structure of the full problem, making the approach feasible on near-term quantum hardware.

What carries the argument

The framework combines four components: (1) graph partitioning into fixed-size blocks, (2) QAOA with an XY mixer that preserves Hamming weight on each subgraph, (3) conditional MADE neural networks trained per block with Hamming weight as a conditioning variable, and (4) Metropolis-Hastings MCMC using the MADE surrogates as proposal distributions. The Metropolis-Hastings acceptance step ensures correctness regardless of surrogate quality, while the surrogate quality determines the acceptance rate and thus mixing speed.

Load-bearing premise

The central assumption is that the conditional MADE neural network can accurately approximate the QAOA sample distribution for each subgraph well enough that the Metropolis-Hastings acceptance rate stays high. If the surrogate distribution diverges from the true low-temperature Boltzmann distribution of the subgraph, proposals are rejected and the speedup vanishes.

What would settle it

Run the framework on problem instances where the QAOA circuit's sample distribution for subgraphs diverges significantly from the constrained Boltzmann distribution, causing Metropolis-Hastings acceptance rates to drop near zero and mixing to slow below that of classical Kawasaki dynamics. Alternatively, test on graph topologies where subgraph partitioning destroys the problem structure that QAOA exploits.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This manuscript proposes a divide-and-conquer neural network surrogate framework to accelerate MCMC sampling for large-scale Ising problems under fixed Hamming weight constraints. The method divides the interaction graph into subgraphs (blocks), generates samples via QAOA with an XY mixer for each subproblem, trains conditional MADE neural networks as surrogate proposal distributions, and uses these surrogates within Metropolis-Hastings MCMC. Numerical experiments on 3-regular graphs (N=16 to 512) report speedup factors of ~20.3 and ~7.6 over Kawasaki pair-flip dynamics with nearest-neighbor and non-nearest-neighbor exchanges, respectively. An MNIST feature selection application (N=784) demonstrates faster energy convergence and a 2.03% improvement in classification accuracy. The central claim is that the quantum-inspired proposal distributions enable efficient and scalable MCMC on NISQ devices.

Significance. The paper addresses a meaningful problem: extending quantum-enhanced MCMC to large-scale constrained optimization. The divide-and-conquer strategy to make QAOA-based proposals tractable for large systems is a reasonable architectural contribution. The inclusion of an MNIST application at N=784 demonstrates practical scalability beyond toy models. The speedup factors reported are substantial if attributable to the proposal quality rather than the block structure alone. The framework is clearly described and the experiments are structured to test scaling with both system size and QAOA block size.

major comments (3)
  1. Missing classical block-update baseline (load-bearing for the central claim). The proposed method updates 16 spins per step (one block), while the Kawasaki baselines update only 2 spins (one pair). Block updates are a well-known classical technique that naturally accelerate mixing by making larger jumps in configuration space. The manuscript does not include a classical block-update baseline — e.g., a Gibbs sampler that approximately samples from each 16-spin block-conditional distribution — against which to isolate the contribution of the QAOA-derived proposal distribution. Without this control, the observed speedup could be largely attributable to the block structure (16-spin updates vs 2-spin updates) rather than to the quality of the quantum-inspired proposals. This is the single most load-bearing gap: if a classical block sampler with the same block size achieves comparable mixing,
  2. Abstract/manuscript mismatch. The abstract describes a paper on CMB E-mode polarization anomalies for AliCPT, but the full text is entirely about quantum-enhanced MCMC for Ising problems. This appears to be a submission error (wrong abstract or wrong manuscript text). This must be corrected before the paper can be properly evaluated. If the provided full text is the correct manuscript, then the abstract needs to be completely replaced. If the provided text is wrong, the correct manuscript needs to be submitted. Either way, this is a fundamental issue that prevents assessment of the paper as submitted.
  3. Incomplete manuscript text. The provided full text is truncated — it begins mid-sentence in Section II.A ('The Quantum Approximate Optimization Algorithm (QAOA) is a variational quantum algorithm that aims') and jumps to page 11 (Figure 9 and references). Sections II through IV are largely missing, including the core method description (Section III), the main numerical experiments (Section IV), and most figures and tables. Without the complete manuscript, a thorough assessment of the method's correctness, the experimental setup, and the validity of the speedup claims cannot be made. The assessment above is based on the abstract, introduction, and the partial results visible in the provided text.
minor comments (4)
  1. The parameter β̃ (appearing in Figure 9 and elsewhere) should be clearly defined in the text, including its relationship to the inverse temperature β and the role it plays in the MADE training.
  2. Reference [20] is cited as 'PRX Quantum 7, 010338 (2026)' — this appears to be a forthcoming reference. Please verify the citation details are correct at the time of resubmission.
  3. The introduction mentions that pair-flip dynamics 'may suffer from exponentially slow mixing' [32], but does not discuss whether the proposed method provably avoids this exponential slowdown or whether this is only empirically demonstrated.
  4. Figure 9 caption: 'fiß' appears to be a rendering artifact for 'β̃'. Please ensure consistent and correct rendering of mathematical symbols throughout.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful reading and for identifying two critical issues: (1) a submission error causing an abstract/manuscript mismatch and truncated text, and (2) the absence of a classical block-update baseline needed to isolate the contribution of QAOA-derived proposals from the block structure itself. We address both below.

read point-by-point responses

  1. Referee: Missing classical block-update baseline (load-bearing for the central claim). The proposed method updates 16 spins per step while Kawasaki baselines update only 2 spins. A classical block-update baseline is needed to isolate the contribution of the QAOA-derived proposal distribution from the block structure.

    Authors: The referee is correct that without a classical block-update baseline of matching block size, the observed speedup could be partially or largely attributable to the block structure rather than to the quality of the QAOA-derived proposals. This is a fair and important point. We will add a classical block Gibbs sampler that approximately samples from each 16-spin block-conditional distribution as an additional baseline in the revised manuscript. This will allow us to isolate the contribution of the quantum-inspired proposal quality from the effect of updating larger blocks. We note that the QAOA-MADE proposals are designed to approximate the constrained low-temperature Boltzmann distribution within each block, which is a stronger target than a generic block Gibbs update; the comparison will clarify how much of the speedup is due to proposal quality versus block size. We agree this control is essential for the central claim and will incorporate it. revision: yes

  2. Referee: Abstract/manuscript mismatch. The abstract describes CMB E-mode polarization anomalies for AliCPT, but the full text is about quantum-enhanced MCMC for Ising problems.

    Authors: The referee is correct. This is a submission error: the abstract from an unrelated manuscript (on CMB E-mode anomalies for AliCPT) was inadvertently attached to our manuscript on divide-and-conquer neural network surrogates for quantum sampling. The correct abstract for this manuscript describes our divide-and-conquer neural network surrogate framework for accelerating MCMC in large-scale constrained optimization problems. We will ensure the correct abstract is submitted and verify that no other metadata mismatches remain. We apologize for this error. revision: yes

  3. Referee: Incomplete manuscript text. The provided full text is truncated, beginning mid-sentence in Section II.A and jumping to page 11. Sections II through IV are largely missing, including the core method description, main numerical experiments, and most figures and tables.

    Authors: The referee is correct. The manuscript text provided to the referee was truncated, missing the core of Sections II through IV. This appears to be a file processing or submission error. The complete manuscript, including the full method description (Section III), numerical experiments (Section IV), and all figures and tables, will be resubmitted. We ask the referee to re-evaluate the complete manuscript once it is available, as the assessment of method correctness, experimental setup, and speedup claims depends on the full text. revision: yes

Circularity Check

0 steps flagged

No significant circularity; self-citations are methodological, not load-bearing for the central speedup claims

full rationale

The paper's central claims (speedup factors of ~20.3 and ~7.6 over Kawasaki dynamics) are established through independent numerical experiments comparing the proposed divide-and-conquer QAOA+MADE framework against classical pair-flip baselines on 3-regular graphs. The derivation chain does not reduce to its inputs by construction. The MADE neural network is trained to approximate QAOA circuit outputs, but the speedup is measured against classical methods, not against the QAOA itself, so there is no self-definitional loop. Self-citations to [20] (Nakano, Okada, Fujii) and [21] (Nakano, Hakoshima, Mitarai, Fujii) introduce the prior MADE-surrogate and quantum-enhanced MCMC ideas, but the present paper's contribution—the divide-and-conquer block decomposition with XY-mixer QAOA and Hamming-weight-conditioned MADE—is a methodological extension whose performance is evaluated independently. No uniqueness theorem is invoked, no fitted parameter is renamed as a prediction, and no ansatz is smuggled through self-citation. The skeptic's concern about missing classical block-update baselines is a valid experimental-design/correctness issue, but it is not circularity: the paper does not define its speedup in terms of a quantity that already encodes the answer. Score 1 reflects the minor self-citation to [20] for the MADE-surrogate concept, which is not load-bearing for the present paper's independent experimental claims.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The ledger captures the key design choices and assumptions. The block size and inverse temperature are free parameters tuned for the experiments. The core axioms are domain assumptions about the capabilities of QAOA and neural network surrogates.

free parameters (3)
  • QAOA block size = 16
    Fixed to 16 qubits for the block size in the divide-and-conquer decomposition, chosen as a practical NISQ limit.
  • Inverse temperature β̃ = 100
    Chosen for the MNIST feature selection experiment based on empirical performance tuning (Appendix Fig 9).
  • Graph sparsity threshold
    Connections below a threshold are removed for the MNIST interaction graph, but the specific threshold value is not stated.
axioms (2)
  • domain assumption QAOA with an XY mixer can approximate the constrained low-temperature Boltzmann distribution for subgraphs.
    The method relies on the assumption that the quantum circuit produces samples close enough to the target distribution to be useful for training the surrogate. This is a standard assumption in quantum-enhanced MCMC literature.
  • domain assumption Conditional MADE can learn the quantum sample distribution.
    The neural network surrogate's ability to approximate the quantum distribution is assumed and is central to the method's practicality.

pith-pipeline@v1.1.0-glm · 9573 in / 2123 out tokens · 209753 ms · 2026-07-05T02:08:38.069552+00:00 · methodology

0 comments
read the original abstract

The standard $\Lambda$CDM model has been highly successful in describing cosmic microwave background (CMB) observations. Nevertheless, a set of large-scale statistical anomalies persists in temperature anisotropies across WMAP and Planck. CMB $E$-mode polarization offers an independent probe of these anomalies, circumventing the look-elsewhere effect inherent in temperature-only analyses. In this paper, we forecast the capability of the Ali CMB Polarization Telescope (AliCPT), a ground-based CMB experiment in the Northern Hemisphere, to detect such anomalies in large-scale $E$-mode polarization. Using 1000 unconstrained simulations processed with the NILC component separation method, we evaluate four anomaly estimators: dipole modulation, lack of large-angle correlations, quadrupole-octopole alignment, and point-parity asymmetry. Our analysis considers two noise levels for AliCPT, the goal configuration of the Simons Observatory (SO) Large Aperture Telescope (LAT) alone, and a joint AliCPT+SO configuration. For dipole modulation, we validate the local variance estimator on modulated simulations with an input amplitude $A_d = 0.07$, and find that the combined AliCPT+SO dataset is likely to detect the injected $E$-mode modulation at a 99% confidence level. Tests of the full suite of anomaly statistics on unconstrained isotropic simulations indicate that AliCPT alone, owing to its limited sky coverage, might introduce systematic biases or enlarged uncertainties, especially for quadrupole-octopole alignment and point-parity asymmetry. The combination with SO largely restores the statistical distributions to those expected in an ideal full-sky scenario, thereby establishing a near-cosmic-variance benchmark for upcoming anomaly investigations.

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Modifications of CMB Temperature and Polarization Quadrupole Signals in Thurston Spacetimes

    gr-qc 2026-05 unverdicted novelty 6.0

    Thurston spacetimes generate distinct evolving temperature and polarization patterns in the CMB that can be tracked via Stokes parameters and potentially isolated per geometry.

  2. Modifications of CMB Temperature and Polarization Quadrupole Signals in Thurston Spacetimes

    gr-qc 2026-05 unverdicted novelty 4.0

    The authors introduce Thurston spacetimes as cosmological backgrounds, solve transfer equations for temperature and polarization patterns, and analyze symmetries in Stokes parameters to attempt isolation of individual...

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    P. W. Shor, inProceedings 35th annual symposium on foundations of computer science(Ieee, 1994) pp. 124– 134

    work page 1994

  2. [2]

    L. K. Grover, Physical review letters79, 325 (1997)

    work page 1997

  3. [3]

    D. S. Abrams and S. Lloyd, Physical Review Letters83, 5162 (1999)

    work page 1999

  4. [4]

    A. W. Harrow, A. Hassidim, and S. Lloyd, Physical re- view letters103, 150502 (2009)

    work page 2009

  5. [5]

    Preskill, Quantum2, 79 (2018)

    J. Preskill, Quantum2, 79 (2018)

    work page 2018

  6. [6]

    Arute, K

    F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A. Buell,et al., nature574, 505 (2019)

    work page 2019

  7. [7]

    Zhong, H

    H.-S. Zhong, H. Wang, Y.-H. Deng, M.-C. Chen, L.-C. Peng, Y.-H. Luo, J. Qin, D. Wu, X. Ding, Y. Hu,et al., Science370, 1460 (2020)

    work page 2020

  8. [8]

    L. S. Madsen, F. Laudenbach, M. F. Askarani, F. Rortais, T. Vincent, J. F. Bulmer, F. M. Miatto, L. Neuhaus, L. G. Helt, M. J. Collins,et al., Nature606, 75 (2022)

    work page 2022

  9. [9]

    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., Nature Reviews Physics3, 625 (2021)

    work page 2021

  10. [10]

    A Quantum Approximate Optimization Algorithm

    E. Farhi, J. Goldstone, and S. Gutmann, arXiv preprint arXiv:1411.4028 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  11. [11]

    Mitarai, M

    K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Phys- ical Review A98, 032309 (2018)

    work page 2018

  12. [12]

    Classification with Quantum Neural Networks on Near Term Processors

    E. Farhi and H. Neven, arXiv preprint arXiv:1802.06002 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  13. [13]

    Peruzzo, J

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’brien, Nature communications5, 4213 (2014)

    work page 2014

  14. [14]

    J. F. Gonthier, M. D. Radin, C. Buda, E. J. Doskocil, C. M. Abuan, and J. Romero, Physical Review Research 4, 033154 (2022)

    work page 2022

  15. [15]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature communications9, 4812 (2018)

    work page 2018

  16. [16]

    Hangleiter and J

    D. Hangleiter and J. Eisert, Reviews of Modern Physics 95, 035001 (2023)

    work page 2023

  17. [17]

    Quantum-Selected Configuration Interaction: classical diagonalization of Hamiltonians in subspaces selected by quantum computers

    K. Kanno, M. Kohda, R. Imai, S. Koh, K. Mitarai, W. Mizukami, and Y. O. Nakagawa, arXiv preprint arXiv:2302.11320 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Robledo-Moreno, M

    J. Robledo-Moreno, M. Motta, H. Haas, A. Javadi- Abhari, P. Jurcevic, W. Kirby, S. Martiel, K. Sharma, S. Sharma, T. Shirakawa,et al., Science Advances11, eadu9991 (2025)

    work page 2025

  19. [19]

    Layden, G

    D. Layden, G. Mazzola, R. V. Mishmash, M. Motta, P. Wocjan, J.-S. Kim, and S. Sheldon, Nature619, 282 (2023)

    work page 2023

  20. [20]

    Nakano, K

    Y. Nakano, K. N. Okada, and K. Fujii, PRX Quantum 7, 010338 (2026)

    work page 2026

  21. [21]

    Nakano, H

    Y. Nakano, H. Hakoshima, K. Mitarai, and K. Fujii, Physical Review Research6, 033105 (2024)

    work page 2024

  22. [22]

    P. C. Lotshaw, G. Siopsis, J. Ostrowski, R. Herrman, 12 R. Alam, S. Powers, and T. S. Humble, Physical Review A108, 042411 (2023)

    work page 2023

  23. [23]

    D´ ıez-Valle, D

    P. D´ ıez-Valle, D. Porras, and J. J. Garc´ ıa-Ripoll, Physical review letters130, 050601 (2023)

    work page 2023

  24. [24]

    Germain, K

    M. Germain, K. Gregor, I. Murray, and H. Larochelle, in International conference on machine learning(PMLR,

    work page

  25. [25]

    Uria, M.-A

    B. Uria, M.-A. Cˆ ot´ e, K. Gregor, I. Murray, and H. Larochelle, Journal of Machine Learning Research17, 1 (2016)

    work page 2016

  26. [26]

    Larochelle and I

    H. Larochelle and I. Murray, inProceedings of the four- teenth international conference on artificial intelligence and statistics(JMLR Workshop and Conference Pro- ceedings, 2011) pp. 29–37

    work page 2011

  27. [27]

    Neuhaus and J

    T. Neuhaus and J. S. Hager, Journal of statistical physics 113, 47 (2003)

    work page 2003

  28. [28]

    Nußbaumer, E

    A. Nußbaumer, E. Bittner, and W. Janke, Progress of Theoretical Physics Supplement184, 400 (2010)

    work page 2010

  29. [29]

    M. A. Nau, L. A. Nutricati, B. Camino, P. A. Warburton, and A. K. Maier, Scientific Reports15, 28937 (2025)

    work page 2025

  30. [30]

    M¨ ucke, R

    S. M¨ ucke, R. Heese, S. M¨ uller, M. Wolter, and N. Pi- atkowski, Quantum Machine Intelligence5, 11 (2023)

    work page 2023

  31. [31]

    P. C. Hohenberg and B. I. Halperin, Reviews of Modern Physics49, 435 (1977)

    work page 1977

  32. [32]

    Kuchukova, M

    A. Kuchukova, M. Pappik, W. Perkins, and C. Yap, Ran- dom Structures & Algorithms67, e70038 (2025)

    work page 2025

  33. [33]

    Hadfield, Z

    S. Hadfield, Z. Wang, B. O’gorman, E. G. Rieffel, D. Ven- turelli, and R. Biswas, Algorithms12, 34 (2019)

    work page 2019

  34. [34]

    Z. Wang, N. C. Rubin, J. M. Dominy, and E. G. Rieffel, Physical Review A101, 012320 (2020)

    work page 2020

  35. [35]

    Papamakarios, T

    G. Papamakarios, T. Pavlakou, and I. Murray, Advances in neural information processing systems30(2017)

    work page 2017

  36. [36]

    Metropolis, A

    N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, The journal of chemical physics21, 1087 (1953)

    work page 1953

  37. [37]

    W. K. Hastings, Biometrika57, 97 (1970)

    work page 1970

  38. [38]

    Kawasaki, Physical Review145, 224 (1966)

    K. Kawasaki, Physical Review145, 224 (1966)

    work page 1966

  39. [39]

    Hukushima and K

    K. Hukushima and K. Nemoto, Journal of the Physical Society of Japan65, 1604 (1996)

    work page 1996

  40. [40]

    H. G. Katzgraber, M. K¨ orner, and A. P. Young, Physical Review B—Condensed Matter and Materials Physics73, 224432 (2006)

    work page 2006

  41. [41]

    H. Peng, F. Long, and C. Ding, IEEE Transactions on pattern analysis and machine intelligence27, 1226 (2005)

    work page 2005