Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Generative modeling assisted simulation of measurement-altered quantum criticality

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper proposes using a physics-preserving diffusion model to generate local density matrices from truncated measurement results, bypassing exponential post-selection in measurement-altered criticality.

desk verdict A clear proposal with one new numerical observation, but the load-bearing claim about a working generative model is untested and the paper admits it. read the letter →

arxiv 2412.01513 v1 pith:2B36OJSR submitted 2024-12-02 quant-ph cs.LG

Yuchen Zhu , Molei Tao , Yuebo Jin , Xie Chen This is my paper · ORCID
classification quant-phcs.LG
keywords diffusionmodelforquantumstatesconditionalgenerationpostselectionmeasurement-alteredcriticalitystructure-preservinglocalreduceddensitymatriceslocalitymany-bodysimulation
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 tackles the post-selection problem in simulating measurement-altered quantum criticality, where measuring an ancilla chain coupled to a one-dimensional critical Ising chain produces full measurement outcomes that are exponentially rare and costly to reconstruct. It argues that locality shrinks this task: the local reduced density matrix of the critical chain near a site is affected mainly by measurement outcomes at nearby sites, so the measurement string can be truncated to a short window. The paper then proposes that a physics-preserving conditional diffusion model can learn the resulting conditional distribution of local density matrices from a limited number of simulator runs and generate new samples from it, while enforcing the constraints every density matrix must satisfy. The claim is that this generative step bypasses the need to sample all exponentially many measurement outcomes.

What carries the argument

The carrying object is the conditional distribution $D(\rho_{s[i]})$ of local reduced density matrices given a truncated measurement string $s[i]$ around a site. The generative engine is the structure-preserving conditional diffusion model, which progressively noises density matrices and learns a score function to reverse the noising, with the Hermitian, positive-semidefinite, and unit-trace constraints hard-wired into the generation. The reduction that makes this feasible is locality: because the chain's interactions are short-range, the variance of a local RDM under changes of the measurement string decays as the changed site moves away, so a short window of measurement outcomes carries almost all the relevant information.

What would settle it

Train the proposed conditional diffusion model on the N=14 local reduced density matrices described in the paper, generate samples for each truncated measurement label, and compare the ensemble statistics of a nonlinear probe, such as the entanglement entropy of two-body reduced density matrices, against exact values obtained by direct simulation; a systematic mismatch would show that the generated samples do not reproduce $D(\rho_{s[i]})$.

Watch

Extended reading notes

Core claim

The central claim is that simulation of measurement-altered Ising criticality can be reformulated as conditional generative modeling of local reduced density matrices. For each site $i$, the post-measurement state is captured by RDMs $\rho_{s[i]}$ labeled by a truncated measurement string around that site, and because the truncation loses information, each label corresponds to a whole distribution $D(\rho_{s[i]})$ rather than a unique state. The paper asserts that a structure-preserving conditional diffusion model can learn these distributions from training data produced by a quantum simulator and sample from them, including for measurement strings not seen during training, while keeping every generated sample Hermitian, positive semidefinite, and trace one. It validates the locality premise on a 14-site chain by showing that one- and two-body RDM entries vary most under changes of nearby measurement outcomes and decay as the changed site moves away. The paper further argues that generating full local density matrices, rather than aggregated measurement statistics, is essential because nonlinear functionals such as entanglement entropy cannot be recovered from mixed statistics.

Load-bearing premise

The load-bearing premise is that a structure-preserving diffusion model trained on local reduced density matrices labeled by truncated measurement strings will faithfully reproduce the true conditional distribution $D(\rho_{s[i]})$, including nonlinear properties such as entanglement, even though the paper does not test the learning and sampling step.

Editorial extensions

If this is right

  • Simulating measurement-altered criticality on a quantum simulator no longer requires accumulating many copies of each exponentially rare measurement outcome; the generative model supplies additional samples from the learned distribution.
  • Because the measurement label is truncated to a local window, training data collected on a small system can generate local reduced density matrices for larger systems, as long as the window still captures the relevant physics.
  • The conditional model can propose samples for measurement strings absent from the training set, extending a finite set of simulator runs to unseen local outcomes.
  • Since the generated objects are full local density matrices, nonlinear quantities such as entanglement entropy remain computable from the generated ensemble, unlike approaches that only reproduce measurement statistics of operators.

Reading between the lines

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

  • Editorial inference: the same conditional-generation strategy could transfer to other measurement-induced phenomena, such as measurement-induced entanglement transitions, wherever a light-cone or locality structure makes local post-measurement states depend on local measurement outcomes.
  • Editorial inference: the paper presents the locality reduction but not the generative model's performance, so the immediate check is to train the structure-preserving diffusion model on the N=14 local RDMs and compare generated entanglement statistics with exact simulation.
  • Editorial inference: the truncation window length sets a bias-variance tradeoff, and in a critical system the RDM variance likely decays polynomially rather than exponentially with distance, which would bound how much the measurement label can be shortened before the conditional distribution becomes too broad to learn.
  • Editorial inference: a natural scaling test is to run the protocol at increasing chain sizes where exact simulation is still possible and check whether the number of training samples required by the generative model grows slowly enough to remain practical as the post-selection probability shrinks.
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

3 major / 5 minor

Summary. The paper proposes a machine-learning-assisted simulation strategy for the measurement-altered Ising quantum criticality protocol of Ref. [7]. It argues that, because the post-measurement reduced density matrices of a local region depend mostly on nearby measurement outcomes, the full exponentially long measurement string can be truncated to a local window s[i]. This turns the problem into conditional generative modeling of a distribution D(ρ_{s[i]}) of local reduced density matrices. The authors propose to use the Structure-Preserving Diffusion Model (SPDM) of Ref. [37] for this task, pointing out that SPDM exactly enforces Hermiticity, positive semidefiniteness, and trace one. The paper reports a locality analysis on a single N=14 chain (Figures 2 and 3) and concludes by stating that implementing the generative model is future work.

Significance. The idea of using generative models to bypass exponential post-selection in measurement-induced quantum phenomena is timely and potentially important. If the proposed pipeline were demonstrated, it would provide a practical way to estimate nonlinear properties such as entanglement from local RDMs without sampling exponentially many global measurement outcomes. The manuscript has clear strengths: it identifies a concrete and physically motivated simplification (locality of RDMs), it correctly emphasizes that truncating the measurement string makes the problem generative rather than deterministic, and it leverages a principled structure-preserving diffusion framework whose constraints are exactly enforced. However, the central claim of the paper, that SPDM can learn and faithfully sample from D(ρ_{s[i]}), is not tested anywhere in the manuscript. The paper is best read as a proposal; as written, the abstract and title assert a capability that the body does not establish.

major comments (3)
  1. [Section 3.2] The central claim is unverified. The abstract states that generation of local RDMs is 'enabled by a physics-preserving conditional diffusion generative model,' and Section 3.2 asserts that SPDM can learn D(ρ_{s[i]}) and sample from it, even for unseen labels. However, Section 4 explicitly says: 'We plan to implement that method to demonstrate the effectiveness of the machine learning methodology.' No implementation, training data size, hyperparameters, convergence diagnostics, or comparison against exact RDMs is provided. The only numerical evidence is the locality analysis, which concerns the physical simplification, not the generative model. Because the paper's title and abstract present the generative modeling approach as a working solution, this is a load-bearing gap that cannot be fixed by rephrasing; it requires either a demonstration or a substantial reframing of the paper as a proposal.
  2. [Section 3.1] The locality analysis, which is the sole quantitative support for truncating s to s[i], is based on a single N=14 chain with no error bars or statistical methodology. Figure 3 plots average variances of one- and two-body RDM entries as a function of the site at which the measurement outcome is changed, but the text does not report the number of independent measurement samples, the values of the protocol parameters (u, Δt, k, K), or any measure of sampling error. Without this information, the reader cannot assess whether the decay of the variance is significant or whether the truncation window of five sites is adequate. Since the entire reduction from full-state sampling to local RDM generation rests on this locality assumption, this point needs a quantitative and statistically grounded validation, ideally with finite-size scaling.
  3. [Section 3.2] The conditional generation capability of SPDM is asserted rather than demonstrated for the specific distribution D(ρ_{s[i]}). The paper does not address whether finite training data from N=14 suffice to learn a distribution over density matrices whose most important downstream quantities, such as Rényi entropies, are nonlinear functions of ρ. Generative models can suffer from mode collapse and over-smoothing, which would systematically bias these nonlinear functionals. The paper also claims that SPDM 'can even leverage the extrapolation capability of neural network to generate samples from D(ρ_{s[i]}) even when s[i] is a new label,' but no evidence is provided for extrapolation in this setting, and the mechanism by which the conditional diffusion model generalizes to unseen measurement strings is not explained. This is not a minor omission; it is the load-bearing assumption of the entire method.
minor comments (5)
  1. [Abstract] The phrase 'physics-preserving conditional diffusion generative model' is used without definition in the abstract; the body only cites Ref. [37]. Please clarify in the abstract or introduce the term explicitly in Section 3.2.
  2. [Appendix A] There is a notation inconsistency in the definition of the effective Hamiltonian: the text introduces 'a(j, k)' and 'a(j)', but the displayed formula for a(j,k) uses ⟨s̃| X̃_j X̃_k |ψ_a⟩/⟨s̃|ψ_a⟩, while the formula for a(j) is missing. Please align the notation and define all coefficients explicitly.
  3. [Appendix A] The parameter C in the inter-chain unitary U_j = exp(iu(Z_j - C) X̃_j) is set to C = -1 without explanation. Please provide the physical motivation or reference for this choice.
  4. [Section 3.1] The text says 'the zero-one sequence refers to the measurement outcome observed at site i-2 to site i+2, where 0 stands for observing |0⟩ and 1 stands for observing |1⟩.' It would be clearer to state explicitly which basis (X or Z) is used for these outcomes, since the protocol measures the ancilla chain in either X or Z basis.
  5. [General] The manuscript does not mention availability of code or data for reproducing Figures 2 and 3. For a computational study, this is a reproducibility concern that should be addressed in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an explicit proposal and the generative model is neither fitted nor used to define the physics it discusses.

full rationale

The paper makes no circular reduction. It is an explicit research proposal: Section 4 states, "We plan to implement that method to demonstrate the effectiveness of the machine learning methodology of generative modeling in assisting with quantum simulation protocols," so no generative model is fitted, tested, or used to produce a predicted physical quantity. The locality simplification in Section 3.1 is supported by direct N=14 simulation data shown in Figures 2 and 3, not by the same quantities that the proposed generative model would later target. The reliance on the authors' own Structure-Preserving Diffusion Model (Ref. [37]) is a citation of a tool, rather than an imported uniqueness theorem or ansatz that defines the measurement-altered criticality result. The untested assumption that SPDM can learn D(ρ_{s[i]}) and extrapolate to unseen labels is a verification gap and a correctness risk, not a circularity. No fitted parameter is renamed as a prediction, and the target protocol is not defined in terms of the generative model's output.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. The key assumptions are the validity of the effective measurement protocol, the locality truncation, and the learnability of the distribution by the diffusion model. The latter is unverified and is the main reason the central claim is not supported.

free parameters (3)
  • coupling strength u = small positive number, not specified
    Controls the strength of inter-chain interaction in the measurement-altered criticality protocol. Chosen by hand and not fitted to data.
  • imaginary time step Δt and number of steps k for ancilla state = not specified
    Used in Trotterized imaginary time evolution to prepare the ancilla state. The final results depend on these choices.
  • truncation length of measurement string s[i] = 2 sites on each side (5 sites total)
    The paper chooses to truncate s to a window around the probed site. This is a key modeling assumption that determines the distribution D(ρ_{s[i]}).
assumptions (3)
  • domain assumption The measurement-altered criticality protocol as described in Ref [7] produces the post-measurement state |ψ̃_s⟩ = (1/√N) U' e^{-Hm/2} |ψ_c⟩.
    The paper takes this effective description from Ref [7] as the starting point for simulation.
  • domain assumption Locality: local reduced density matrices depend mostly on nearby measurement outcomes.
    Used to truncate s to s[i]. The paper verifies this numerically for N=14 but does not prove it generally.
  • domain assumption The structure-preserving diffusion model (SPDM) from Ref [37] can learn and sample the conditional distribution D(ρ_{s[i]}) faithfully.
    This is the central unverified assumption. The paper plans to implement it but does not demonstrate it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative modeling assisted simulation of measurement-altered quantum criticality." pith.science (2026). https://pith.science/paper/2B36OJSR

@misc{pith2026241201513,
  author       = {Pith},
  title        = {Pith review of: Generative modeling assisted simulation of measurement-altered quantum criticality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2B36OJSR}},
  note         = {Machine review of arXiv:2412.01513}
}
read the original abstract

In quantum many-body systems, measurements can induce qualitative new features, but their simulation is hindered by the exponential complexity involved in sampling the measurement results. We propose to use machine learning to assist the simulation of measurement-induced quantum phenomena. In particular, we focus on the measurement-altered quantum criticality protocol and generate local reduced density matrices of the critical chain given random measurement results. Such generation is enabled by a physics-preserving conditional diffusion generative model, which learns an observation-indexed probability distribution of an ensemble of quantum states, and then samples from that distribution given an observation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [37]

    A., Chen, X

    Zhu, Y., Chen, T., Theodorou, E. A., Chen, X. & Tao, M. Quantum state generation with structure-preserving diffusion model (2024). URL https://arxiv.org/abs/2404. 06336. 2404.06336. 10

    arXiv 2024

  2. [7]

    Murciano, S., Sala, P., Liu, Y., Mong, R. S. K. & Alicea, J. Measurement-altered ising quantum criticality. Phys. Rev. X 13, 041042 (2023). URL https://link.aps.org/ doi/10.1103/PhysRevX.13.041042

    doi:10.1103/physrevx.13.041042 2023

  3. [1]

    Nielsen, M. A. & Chuang, I. L. Quan- tum computation and quantum information (Cambridge university press, 2010)

    work page 2010

  4. [2]

    & Fisher, M

    Li, Y., Chen, X. & Fisher, M. P. A. Quan- tum zeno effect and the many-body entan- glement transition. Phys. Rev. B98, 205136 (2018). URL https://link.aps.org/doi/10. 1103/PhysRevB.98.205136

    work page 2018

  5. [3]

    & Nahum, A

    Skinner, B., Ruhman, J. & Nahum, A. Measurement-induced phase transitions in the dynamics of entanglement. Phys. Rev. X 9, 031009 (2019). URL https://link.aps.org/ doi/10.1103/PhysRevX.9.031009

    doi:10.1103/physrevx.9.031009 2019

  6. [4]

    & Fisher, M

    Li, Y., Chen, X. & Fisher, M. P. A. Measurement-driven entanglement transition in hybrid quantum circuits. Phys. Rev. B 100, 134306 (2019). URL https://link.aps. org/doi/10.1103/PhysRevB.100.134306

    doi:10.1103/physrevb.100.134306 2019

  7. [5]

    & Lud- wig, A

    Jian, C.-M., You, Y.-Z., Vasseur, R. & Lud- wig, A. W. W. Measurement-induced critical- ity in random quantum circuits. Phys. Rev. B 101, 104302 (2020). URL https://link.aps. org/doi/10.1103/PhysRevB.101.104302

    doi:10.1103/physrevb.101.104302 2020

  8. [6]

    Y., Jian, C.-M

    Lee, J. Y., Jian, C.-M. & Xu, C. Quan- tum criticality under decoherence or weak measurement. PRX Quantum 4, 030317 (2023). URL https://link.aps.org/doi/10. 1103/PRXQuantum.4.030317

    work page 2023

Show all 37 references
  1. [8]

    & Jian, C.-M

    Yang, Z., Mao, D. & Jian, C.-M. Entan- glement in a one-dimensional critical state after measurements. Phys. Rev. B 108, 165120 (2023). URL https://link.aps.org/ doi/10.1103/PhysRevB.108.165120

    2023 doi

  2. [9]

    & Garratt, S

    Weinstein, Z., Sajith, R., Altman, E. & Garratt, S. J. Nonlocality and entangle- ment in measured critical quantum ising chains. Phys. Rev. B 107, 245132 (2023). URL https://link.aps.org/doi/10. 1103/PhysRevB.107.245132

    2023

  3. [10]

    Y., Ji, W., Bi, Z

    Lee, J. Y., Ji, W., Bi, Z. & Fisher, M. P. A. Decoding measurement-prepared quantum phases and transitions: from ising model to gauge theory, and beyond (2022). URL https: //arxiv.org/abs/2208.11699. 2208.11699

    2022 arXiv

  4. [11]

    J., Weinstein, Z

    Garratt, S. J., Weinstein, Z. & Altman, E. Measurements conspire nonlocally to restruc- ture critical quantum states. Phys. Rev. X 13, 021026 (2023). URL https://link.aps. org/doi/10.1103/PhysRevX.13.021026

    2023 doi

  5. [12]

    Garratt, S. J. & Altman, E. Prob- ing postmeasurement entanglement without postselection. PRX Quantum 5, 030311 (2024). URL https://link.aps.org/doi/10. 1103/PRXQuantum.5.030311. 8

    2024

  6. [13]

    & Ganguli, S

    Sohl-Dickstein, J., Weiss, E., Mah- eswaranathan, N. & Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. ICML (2015)

    2015

  7. [14]

    & Abbeel, P

    Ho, J., Jain, A. & Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020)

    2020

  8. [15]

    Song, Y. et al. Score-based generative model- ing through stochastic differential equations. ICLR (2021)

    2021

  9. [16]

    & Ommer, B

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P. & Ommer, B. High-resolution image synthesis with latent diffusion models. IEEE/CVF conference on computer vision and pattern recognition (CVPR)(2022)

    2022

  10. [17]

    Ho, J. et al. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research23, 2249–2281 (2022)

    2022

  11. [18]

    Anderson, B. D. Reverse-time diffusion equation models. Stochastic Processes and their Applications 12, 313–326 (1982)

    1982

  12. [19]

    Haussmann, U. G. & Pardoux, E. Time rever- sal of diffusions. The Annals of Probability 1188–1205 (1986)

    1986

  13. [20]

    Ge, S. et al. Preserve your own correla- tion: A noise prior for video diffusion mod- els. IEEE/CVF International Conference on Computer Vision (ICCV)(2023)

    2023

  14. [21]

    et al.Align your latents: High- resolution video synthesis with latent diffu- sion models

    Blattmann, A. et al.Align your latents: High- resolution video synthesis with latent diffu- sion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)

    2023

  15. [22]

    D., Ho, J., Tarlow, D

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D. & Van Den Berg, R. Structured denois- ing diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems 34, 17981–17993 (2021)

    2021

  16. [23]

    & Ermon, S

    Lou, A., Meng, C. & Ermon, S. Discrete diffusion language modeling by estimating the ratios of the data distribution. NeurIPS (2023)

    2023

  17. [24]

    Zeng, X. et al. Lion: Latent point diffu- sion models for 3d shape generation. arXiv preprint arXiv:2210.06978 (2022)

    2022 arXiv

  18. [25]

    Luo, S. & Hu, W. Diffusion probabilis- tic models for 3d point cloud generation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)(2021)

    2021

  19. [26]

    Jumper, J. et al. Highly accurate protein structure prediction with alphafold. Nature 596, 583–589 (2021)

    2021

  20. [27]

    Watson, J. L. et al. Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models. BioRxiv 2022–12 (2022)

    2022

  21. [28]

    & Kulik, H

    Duan, C., Du, Y., Jia, H. & Kulik, H. J. Accurate transition state generation with an object-aware equivariant elementary reaction diffusion model. Nature Computational Sci- ence (2023)

    2023

  22. [29]

    G., Vignac, C

    Hoogeboom, E., Satorras, V. G., Vignac, C. & Welling, M. Equivariant diffusion for molecule generation in 3d. International con- ference on machine learning (ICML)(2022)

    2022

  23. [30]

    Mardani, M. et al. Residual diffusion mod- eling for km-scale atmospheric downscaling (2024)

    2024

  24. [31]

    Liu, G.-H., Chen, T., Theodorou, E. A. & Tao, M. Mirror diffusion models for constrained and watermarked generation. NeurIPS (2023)

    2023

  25. [32]

    & De Bortoli, V

    Fishman, N., Klarner, L., Mathieu, E., Hutchinson, M. & De Bortoli, V. Metropo- lis sampling for constrained diffusion models. NeurIPS (2023)

    2023

  26. [33]

    & Hutchinson, M

    Fishman, N., Klarner, L., De Bortoli, V., Mathieu, E. & Hutchinson, M. Diffusion models for constrained domains. TMLR (2023). 9

    2023

  27. [34]

    & Ermon, S

    Lou, A. & Ermon, S. Reflected diffusion models. ICML (2023)

    2023

  28. [35]

    & Nichol, A

    Dhariwal, P. & Nichol, A. Diffusion mod- els beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  29. [36]

    & Salimans, T

    Ho, J. & Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    2022 arXiv

Pith tools

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