Pith. sign in

REVIEW 3 major objections 5 minor 35 references

Neural nets and self-attention predict bipartite Rényi-2 entropy from fewer random measurements than classical-shadow formulas.

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 · grok-4.5

2026-07-12 07:09 UTC pith:T2QY3U5E

load-bearing objection Clean empirical win: MLP and parameter-free attention beat Brydges Rényi-2 estimator on 2- and 4-qubit Haar pure states, with transparent heat-maps, but only inside Page-value concentration. the 3 major comments →

arxiv 2607.02767 v1 pith:T2QY3U5E submitted 2026-07-02 quant-ph

Self-Attention for Quantum Entanglement Prediction

Anuj Gore , Roopayan Ghosh , Dylan Lewis , Sougato Bose This is my paper
classification quant-ph
keywords quantum entanglementRényi entropyclassical shadowsself-attentionmachine learningrandomized measurementssample efficiency
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.

Quantifying entanglement on quantum devices is expensive: full tomography is impossible for large systems, and even randomized classical-shadow estimators need many unitaries and many shots per unitary. This paper trains two simple networks—a plain multilayer perceptron and the same network with a parameter-free self-attention layer—to map a modest collection of random unitaries plus their computational-basis outcome histograms directly to the bipartite second Rényi entropy. On Haar-random pure states of two and four qubits the learned estimators reach a given accuracy with substantially fewer unitaries and far fewer shots than the analytic formula of Brydges et al. The practical payoff is lower measurement overhead for entanglement characterization on near-term hardware, especially in the noisy, low-shot regime.

Core claim

When trained on finite classical-shadow data of Haar-random pure states, both a feed-forward network and a self-attention network predict the bipartite second Rényi entropy more accurately, and with tighter variance, than the standard analytic estimator, thereby reducing the number of random unitaries and measurement shots required for a target error on systems of two and four qubits.

What carries the argument

A two-stage architecture: a feed-forward map ϕ that embeds each random unitary into a real feature vector, optionally followed by a Gram-matrix self-attention layer built from those embeddings, whose output is concatenated with the empirical outcome probabilities and passed through a second network S that outputs the predicted Rényi-2 entropy.

Load-bearing premise

The networks are trained only on Haar-random pure states whose entanglement sits near the typical Page value; the same accuracy is assumed to carry over to the weakly entangled or mixed states that appear in real experiments.

What would settle it

Train the identical models on Haar data, then evaluate them on a balanced set of weakly entangled pure states (or on mixed states produced by a depolarizing channel) and check whether the mean-squared error stays below the analytic classical-shadow baseline for the same (N_U, N_M) budget.

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 / 5 minor

Summary. The paper proposes two neural models (a feed-forward MLP and a parameter-free self-attention layer built from a Gram matrix of vectorized unitaries) that take classical-shadow inputs—NU random unitaries together with NM computational-basis outcome probabilities—and regress the bipartite second Rényi entropy of pure Haar-random states. Targets are obtained independently via Schmidt decomposition. Performance is benchmarked against the analytical randomized-measurement estimator of Brydges et al. (Eqs. 2–3) on 2- and 4-qubit systems; correlation plots, error heat-maps with variance thresholds, and slope/intercept diagnostics are used to argue higher accuracy and lower sample complexity for the learned estimators.

Significance. If the sample-efficiency gains hold beyond the Haar ensemble, the approach would be a practical tool for low-shot entanglement estimation on near-term devices, where full tomography and large NU/NM budgets are prohibitive. Strengths include clean separation of training targets from the analytical baseline, thorough finite-sample diagnostics (Figs. 3–6, 10–11), an explicit construction of the attention matrix from the unitary ensemble (Sec. III B), and an honest Outlook that flags the Page-value bias and exponential Haar cost. The work therefore supplies a concrete, reproducible numerical baseline for data-driven classical-shadow post-processing.

major comments (3)
  1. [Abstract, Sec. IV, App. B] Abstract and Sec. IV claim “higher accuracy and improved sample efficiency across a range of system sizes” and “scalable \ldots estimation of quantum correlations.” All training and evaluation data are Haar-random pure states whose Rényi-2 values concentrate tightly about the Page value (App. B, Fig. 7). The analytical estimator (Eqs. 2–3) is unbiased for any pure state; the observed ML advantage is therefore demonstrated only inside this narrow measure. Without error heat-maps or slope diagnostics on any non-Page ensemble (weakly entangled, mixed, or the TFIM/XY circuits suggested in Sec. V), the reduction in NU and NM cannot be asserted for the states that appear in experiments or fault-tolerant protocols. This is load-bearing for the central claim.
  2. [Sec. III B, Figs. 4 and 6] Sec. III B asserts that the self-attention Gram matrix “plays a crucial role in the generalizability of the model to other forms of entanglement” and is “suited for entropy and purity based calculations.” No experiment tests any other entanglement measure, multipartite cut, or non-Haar ensemble; on the reported Haar bipartite Rényi-2 task the plain MLP consistently meets the error/variance thresholds with fewer resources than the attention model (Figs. 4 and 6). The generalizability claim is therefore unsupported and should be removed or substantiated.
  3. [Abstract, Introduction, App. C] Only N = 2 and N = 4 are studied, yet the abstract and introduction repeatedly invoke “a range of system sizes” and “scalable” methods. Appendix C already shows that the number of unitaries needed to cover the Haar measure grows as ~4^N; the numerical evidence therefore does not yet underwrite the scalability language used to frame the contribution.
minor comments (5)
  1. [Eq. (4)] Eq. (4) writes θj ∼ [π/2, π/2), an empty interval; the intended range is almost certainly [−π/2, π/2). Correct the typesetting.
  2. [passim] Rényi is spelled inconsistently (Renyi / R´enyi / Renyi-2). Standardize throughout, including figure captions.
  3. [App. D, Table I] Learning rates differ by three orders of magnitude (Attention 5e-8 vs MLP 1e-5, Table I) with no ablation or justification; a short sensitivity check would strengthen reproducibility.
  4. [Figs. 4 and 6] The highlighted heat-map thresholds (0.04±0.04 for 2 qubits, 0.06±0.1 for 4 qubits) appear chosen after inspection; state the selection criterion or report continuous error surfaces without binary highlighting.
  5. [App. D] Code and trained weights are not mentioned; releasing them would allow independent verification of the (NU, NM) heat-maps.

Circularity Check

0 steps flagged

No circularity: independent Schmidt targets, held-out evaluation vs. analytical estimator, attention Gram matrix built only from input unitaries.

full rationale

The paper's central claim is an empirical performance comparison: MLP and self-attention networks map finite classical-shadow inputs (N_U unitaries + N_M-shot probability vectors) to bipartite Rényi-2 entropy and achieve lower MSE / tighter variance than the Brydges et al. analytical estimator (Eqs. 2–3) on the same data. Training targets are obtained independently by Schmidt decomposition of the known pure Haar state (Appendix B), not from the measurement outcomes or from any fitted parameter that is later re-labeled a prediction. The self-attention Gram matrix A = ∑ |u⃗(s)⟩⟨u⃗(s)| is constructed solely from the input unitaries (Section III B) and is trained end-to-end against the independent entropy label; it does not encode the target by construction. Evaluation uses held-out (N_U, N_M) pairs and reports heat-map thresholds (0.04 ± 0.04 for 2 qubits, 0.06 ± 0.1 for 4 qubits) that are direct numerical comparisons, not tautologies. No uniqueness theorem, self-citation chain, or ansatz is load-bearing for the sample-efficiency claim. The acknowledged Haar/Page bias (Fig. 7, Section V) is a generalization limitation, not a circular reduction of the reported results. The derivation chain is therefore self-contained against its own external benchmark.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on standard quantum-information definitions, the classical-shadow protocol, the Brydges randomized-measurement formula, and the usual neural-network training assumptions. The only paper-specific constructions are the particular attention Gram matrix and the chosen network widths; no new physical entities are postulated.

free parameters (4)
  • learning rates (Attention 5e-8, MLP 1e-5)
    Hand-tuned values listed in Table I; different rates for the two models affect convergence and final accuracy.
  • hidden-layer dimensions of ϕ and S
    Architecture widths ([128,512,128] for 2 qubits; larger for 4 qubits) chosen by the authors and not derived from first principles.
  • embedding dimension d = 2^N
    Ad-hoc choice stated to ‘work well in practice’; could be any value and is not optimized.
  • error/variance thresholds (0.04±0.04 for 2q, 0.06±0.1 for 4q)
    Highlighting criteria used to declare ‘success’ in the heat-maps; chosen after inspection of the data.
axioms (4)
  • domain assumption Bipartite second Rényi entropy fully quantifies entanglement for pure states via the spectrum of the reduced density matrix.
    Standard quantum-information fact invoked in Section II and Appendix A; used to justify the target quantity.
  • domain assumption The Brydges et al. formula (Eqs. 2–3) correctly estimates H_2 from randomized unitaries and computational-basis probabilities.
    Taken as the analytical baseline against which ML models are compared (Section II, Ref. [22]).
  • ad hoc to paper Haar-random pure states are a representative ensemble for testing entanglement estimators.
    Explicitly chosen for data generation (Section IV, Appendix B); the paper later acknowledges the strong Page-value bias.
  • domain assumption Standard neural-network training (Adam, ReLU, residual connections, L2 loss) yields models that generalize across (N_U, N_M) pairs.
    Implicit throughout Section III and Appendix D; no theoretical generalization bound is supplied.
invented entities (1)
  • Self-attention Gram matrix A formed from outer products of vectorized unitaries after the ϕ embedding no independent evidence
    purpose: To inject pairwise unitary overlaps into the network without adding trainable parameters, claimed to encode geometry relevant to purity/entropy.
    Constructed in Section III B (Eqs. 6–7); no independent experimental verification outside the present simulations.

pith-pipeline@v1.1.0-grok45 · 41342 in / 2993 out tokens · 26970 ms · 2026-07-12T07:09:30.441834+00:00 · methodology

0 comments
read the original abstract

Quantum entanglement is a powerful resource for quantum-enhanced technologies. However, its reliable quantification remains challenging due to the exponential scaling of the Hilbert space with system size, which renders full state tomography infeasible. Moreover, experimentally estimating entanglement typically requires a large number of measurement samples leading to a significant overhead. In this work, we present two models, a feed-forward neural network and an attention-based model, to accurately predict the bipartite second Renyi from projective measurements of quantum states. We benchmark their performance against standard classical shadow estimators and find that the machine-learning approaches achieve higher accuracy and improved sample efficiency across a range of system sizes. Our results demonstrate the potential of machine learning for scalable and efficient estimation of quantum correlations.

Figures

Figures reproduced from arXiv: 2607.02767 by Anuj Gore, Dylan Lewis, Roopayan Ghosh, Sougato Bose.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Correlation between True and Predicted entropy for 2 qubits for different [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Accuracy of entropy prediction for 2 qubits for varying [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Correlation between True and Predicted entropy for 4 qubits for different [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Accuracy of entropy prediction for 4 qubits for varying [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Distribution of entropy values during simulation as a result of Haar randomizing. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Loss curves of both models. Hot pink corresponds to the self-attention model while the blue curve corresponds to the [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Progression of slope and intercept in the line of best fit of Fig 3 and Fig 5. In the figures of the main text, the [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11. 3D extrusions of Fig 4 and Fig 6. We extrude the figures by plotting the error on the z-axis. Note that the wireframe [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 4 canonical work pages

  1. [1]

    Bouwmeester, J.-W

    D. Bouwmeester, J.-W. Pan, K. Mattle, M. Eibl, H. We- infurter, and A. Zeilinger, Experimental quantum tele- portation, Nature390, 575–579 (1997)

    1997

  2. [2]

    Jozsa and N

    R. Jozsa and N. Linden, On the role of entanglement in quantum-computational speed-up, Proceedings of the Royal Society of London. Series A: Mathematical, Phys- ical and Engineering Sciences459, 2011–2032 (2003)

    2011

  3. [3]

    J. Yin, Y. Cao, Y.-H. Li, S.-K. Liao, L. Zhang, J.-G. Ren, W.-Q. Cai, W.-Y. Liu, B. Li, H. Dai, G.-B. Li, Q.-M. Lu, Y.-H. Gong, Y. Xu, S.-L. Li, F.-Z. Li, Y.-Y. Yin, Z.-Q. Jiang, M. Li, J.-J. Jia, G. Ren, D. He, Y.-L. Zhou, X.-X. Zhang, N. Wang, X. Chang, Z.-C. Zhu, N.-L. Liu, Y.-A. Chen, C.-Y. Lu, R. Shu, C.-Z. Peng, J.-Y. Wang, and J.- W. Pan, Satellite...

    Pith/arXiv arXiv 2017

  4. [5]

    Huang, R

    H.-Y. Huang, R. Kueng, and J. Preskill, Predicting many properties of a quantum system from very few measure- ments, Nature Physics16, 1050–1057 (2020)

    2020

  5. [6]

    Z. Cai, R. Babbush, S. C. Benjamin, S. Endo, W. J. Hug- gins, Y. Li, J. R. McClean, and T. E. O’Brien, Quan- tum error mitigation, Reviews of Modern Physics95, 10.1103/revmodphys.95.045005 (2023)

    doi:10.1103/revmodphys.95.045005 2023

  6. [7]

    Huang, M

    H.-Y. Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, Power of data in quantum machine learning, Nature Communications12, 10.1038/s41467-021-22539-9 (2021)

    doi:10.1038/s41467-021-22539-9 2021

  7. [8]

    Huang, R

    H.-Y. Huang, R. Kueng, G. Torlai, V. V. Albert, and J. Preskill, Provably efficient machine learning for quan- tum many-body problems, Science377, 10.1126/sci- ence.abk3333 (2022)

    doi:10.1126/sci- 2022

  8. [9]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, At- tention is all you need (2023), arXiv:1706.03762 [cs.CL]

    Pith/arXiv arXiv 2023

  9. [10]

    Havl´ ıˇ cek, A

    V. Havl´ ıˇ cek, A. D. C´ orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Super- vised learning with quantum-enhanced feature spaces, Nature567, 209–212 (2019)

    2019

  10. [11]

    P. R. Hegde, G. Passarelli, A. Scocco, and P. Lucignano, Genetic optimization of quantum annealing, Physical Re- view A105, 10.1103/physreva.105.012612 (2022)

    work page doi:10.1103/physreva.105.012612 2022

  11. [12]

    Bausch, A

    J. Bausch, A. W. Senior, F. J. H. Heras, T. Edlich, A. Davies, M. Newman, C. Jones, K. Satzinger, M. Y. Niu, S. Blackwell, G. Holland, D. Kafri, J. Atalaya, C. Gidney, D. Hassabis, S. Boixo, H. Neven, and P. Kohli, Learning high-accuracy error decoding for quantum pro- cessors, Nature635, 834–840 (2024)

    2024

  12. [13]

    Russo, M

    E. Russo, M. Palesi, D. Patti, G. Ascia, and V. Catania, Attention-based deep reinforcement learning for qubit allocation in modular quantum architectures (2024), arXiv:2406.11452 [quant-ph]

    Pith/arXiv arXiv 2024

  13. [14]

    P. Cha, P. Ginsparg, F. Wu, J. Carrasquilla, P. L. McMa- hon, and E.-A. Kim, Attention-based quantum tomog- raphy, Machine Learning: Science and Technology3, 01LT01 (2021)

    2021

  14. [15]

    Zhang and M

    Y.-H. Zhang and M. Di Ventra, Transformer quan- tum state: A multipurpose model for quantum many- body problems, Physical Review B107, 10.1103/phys- revb.107.075147 (2023)

    doi:10.1103/phys- 2023

  15. [16]

    D. Luo, Z. Chen, J. Carrasquilla, and B. K. Clark, Au- toregressive neural network for simulating open quantum systems via a probabilistic formulation, Physical Review Letters128, 10.1103/physrevlett.128.090501 (2022)

    doi:10.1103/physrevlett.128.090501 2022

  16. [17]

    N. Asif, U. Khalid, A. Khan, T. Q. Duong, and H. Shin, Entanglement detection with artificial neural networks, Scientific Reports13, 1562 (2023)

    2023

  17. [18]

    Ure˜ na, A

    J. Ure˜ na, A. Sojo, J. Bermejo-Vega, and D. Manzano, Entanglement detection with classical deep neural net- works, Scientific Reports14, 10.1038/s41598-024-68213- 0 (2024)

    work page doi:10.1038/s41598-024-68213- 2024

  18. [19]

    J. Gray, L. Banchi, A. Bayat, and S. Bose, Machine- learning-assisted many-body entanglement measure- ment, Phys. Rev. Lett.121, 150503 (2018)

    2018

  19. [20]

    Huang, L

    Y. Huang, L. Che, C. Wei, F. Xu, X. Nie, J. Li, D. Lu, and T. Xin, Measuring quantum entanglement from local information by machine learning (2022), arXiv:2209.08501 [quant-ph]

    Pith/arXiv arXiv 2022

  20. [21]

    M. B. Plenio and S. Virmani, An introduction to en- tanglement measures (2006), arXiv:quant-ph/0504163 [quant-ph]

    Pith/arXiv arXiv 2006

  21. [22]

    Brydges, A

    T. Brydges, A. Elben, P. Jurcevic, B. Vermersch, C. Maier, B. P. Lanyon, P. Zoller, R. Blatt, and C. F. Roos, Probing r´ enyi entanglement entropy via random- ized measurements, Science364, 260–263 (2019)

    2019

  22. [23]

    Lewis, R

    D. Lewis, R. Wiersema, J. Carrasquilla, and S. Bose, Geodesic algorithm for unitary gate design with time- independent hamiltonians, Physical Review A111, 10.1103/physreva.111.052618 (2025)

    work page doi:10.1103/physreva.111.052618 2025

  23. [24]

    L. P. Hughston, R. Jozsa, and W. K. Wootters, A com- plete classification of quantum ensembles having a given density matrix, Physics Letters A183, 14 (1993)

    1993

  24. [25]

    L. Chen, H. Zhu, and T.-C. Wei, Connections of geomet- ric measure of entanglement of pure symmetric states to quantum state estimation, Physical Review A83, 10.1103/physreva.83.012305 (2011)

    work page doi:10.1103/physreva.83.012305 2011

  25. [26]

    Gielerak and M

    R. Gielerak and M. Sawerwain, Maximally entangled states of (d,∞) quantum systems: some numerical stud- ies (2019), arXiv:1912.03738 [quant-ph]

    Pith/arXiv arXiv 2019

  26. [27]

    G. K. Brennen, An observable measure of entangle- ment for pure states of multi-qubit systems (2003), arXiv:quant-ph/0305094 [quant-ph]

    Pith/arXiv arXiv 2003

  27. [28]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Brad- bury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨ opf, E. Yang, Z. De- Vito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, Pytorch: An impera- tive style, high-performance deep learning library (2019), arXiv:1912.01703 [cs.LG]

    Pith/arXiv arXiv 2019

  28. [29]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Van- derPlas, S. Wanderman-Milne, and Q. Zhang, JAX: com- posable transformations of Python+NumPy programs (2018)

    2018

  29. [30]

    J. Heek, A. Levskaya, A. Oliver, M. Ritter, B. Ronde- pierre, A. Steiner, and M. van Zee, Flax: A neural net- work library and ecosystem for JAX (2024)

    2024

  30. [31]

    LeCun, L

    Y. LeCun, L. Bottou, G. B. Orr, and K. R. M¨ uller, Effi- cient backprop, inNeural Networks: Tricks of the Trade, edited by G. B. Orr and K.-R. M¨ uller (Springer Berlin 10 Heidelberg, Berlin, Heidelberg, 1998) pp. 9–50

    1998

  31. [32]

    D. N. Page, Average entropy of a subsystem, Physical Review Letters71, 1291–1294 (1993)

    1993

  32. [33]

    Schuster, J

    T. Schuster, J. Haferkamp, and H.-Y. Huang, Random unitaries in extremely low depth (2025), arXiv:2407.07754 [quant-ph]. Appendix A: Preliminaries on Quantum Entanglement As stated in Section II, we define entanglement as en- tanglement can be defined as the inseparability of a func- tion in a Hilbert space into two or more constituent sub- spaces. Let us...

    Pith/arXiv arXiv 2025

  33. [34]

    Training Configuration Our training configuration is summarized in Table I. TABLE I. Training hyperparameters. Model architecture is described in the main text. Hyperparameter Value Optimizer Adam Learning rate (Attention) 5×10 −8 Learning rate (MLP) 1×10 −5 Batch size 32 Number of epochs 5000 Loss function L2-loss

  34. [35]

    TABLE II

    Compute Environment Our compute details is summarized in Table II. TABLE II. Compute and software environment. Component Details GPU NVIDIA Tesla V100 (32 GB) Number of GPUs 4 RAM 512 GB Framework Jax 0.4.28 Training time (2 qubits, MLP)∼10 hours Training time (2 qubits, Attention)∼10 hours Training time (4 qubits, MLP)∼46 hours Training time (4 qubits, A...

  35. [36]

    Both these models had the same number of parame- ters per qubit number - the 2 qubit model had 401,797 parameters while the 4 qubit model had 3,232,529 parameters

    Model details As mentioned there are two neural networks - one for processing the unitaryϕand another for processing the outcome probabilities and the processed unitary together S. Both these models had the same number of parame- ters per qubit number - the 2 qubit model had 401,797 parameters while the 4 qubit model had 3,232,529 parameters. For 2 qubits...