Pith. sign in

REVIEW 3 major objections 4 minor 45 references

The learnability scaling of quantum states: restricted Boltzmann machines

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A restricted Boltzmann machine learning the critical transverse-field Ising ground state needs a number of weights that grows quadratically with the qubit count, while the required training data grows only linearly.

desk verdict First systematic empirical scaling study for RBM-based quantum state reconstruction, giving useful O(N^2) parameter and O(N) sample benchmarks, but the DMRG data fidelity and missing error bars keep the central scalings conditional. read the letter →

arxiv 1908.07532 v2 pith:OQ4JXTEU submitted 2019-08-20 quant-ph

Dan Sehayek , Anna Golubeva , Michael S. Albergo , Bohdan Kulchytskyy , Giacomo Torlai , Roger G. Melko This is my paper · ORCID
classification quant-ph
keywords restrictedBoltzmannmachinequantumstatereconstructiontransverse-fieldIsingmodelscalinganalysissamplecomplexitygenerativemodelingcriticalityweightpruning
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

This paper asks how the cost of reconstructing a quantum state from measurement data grows with the number of qubits, using a restricted Boltzmann machine (RBM) to learn the ground state of the one-dimensional transverse-field Ising model. It defines "learned" by a fixed relative error in the RBM's energy estimator and then measures the smallest hidden layer and fewest training samples that meet that bar. The central finding is that near the quantum critical point the required number of hidden units grows linearly in the number of qubits, which makes the number of weight parameters grow quadratically, while the minimal number of training samples grows only linearly. Away from criticality the cost is essentially constant. The paper also shows that many learned weights can be pruned after training, indicating that the RBM is deliberately over-parameterized to make learning easier.

What carries the argument

The load-bearing object is the restricted Boltzmann machine, a bipartite graphical model with energy $E_\lambda(v,h)=-\sum_{ij} W_{ij}v_i h_j - \sum_i b_i v_i - \sum_j c_j h_j$ over visible spin configurations $v$ and hidden units $h$, whose marginal $p_\lambda(v)$ is matched to the measurement distribution; for the positive-real TFIM ground state the wavefunction is $\psi_\lambda(v)=\sqrt{p_\lambda(v)}$ by the Perron-Frobenius theorem. The criterion that makes the scaling measurable is the relative observable error (ROE) $\epsilon$ defined on the energy estimator, including the 99% confidence interval. Counting parameters in the $N\times N_h$ weight matrix turns the observed linear hidden-unit growth $\mathcal{O}(N)$ into the quadratic parameter scaling $\mathcal{O}(N^2)$, and the pruning protocol (set small $|W_{ij}|$ to zero, fine-tune, repeat) is the tool that reveals over-parameterization.

What would settle it

Compare RBM training on the DMRG-generated data with training on data from a genuinely symmetric ground state, for example exact diagonalization up to $N=20$ or a DMRG run with bond dimension large enough that the spin-flip symmetry is restored to machine precision. If the hidden-unit count at $h/J\approx0.8$ drops and the parameter peak moves to $h/J=1$, the reported quadratic scaling describes learning an artifact of the corrupted training data; if the same $\mathcal{O}(N^2)$ scaling survives on faithful data, the claim stands.

Watch

Extended reading notes

Core claim

Under a fixed learning criterion—relative energy error below $\epsilon=0.002$ with 99% confidence on the RBM's own Monte Carlo samples—the minimal RBM needed to represent the positive-real transverse-field Ising model ground state has, near the quantum critical point $h/J=1$, a hidden layer of size $N_h \approx N/2$, so the weight matrix contains $\mathcal{O}(N^2)$ parameters; deep in the ferromagnetic or paramagnetic phases the same criterion is met with $\mathcal{O}(1)$ hidden units. With the hidden fraction fixed near $\alpha = N_h/N = 1/2$, the minimum number of independent projective measurements needed to train these weights grows linearly with $N$. After convergence, zeroing the smallest weights followed by brief fine-tuning cuts the required nonzero weights sharply (for $N=40$, from 760 to 119 at criticality), which the authors take as evidence that the RBM's over-parameterization exists to ease the contrastive-divergence optimization rather than to represent the state.

Load-bearing premise

That the simulated measurement data generated with a maximum tensor-network bond dimension of 2000 faithfully represents the true ground state at every system size and field strength studied, even where the expected spin-flip symmetry is absent from the data.

Editorial extensions

If this is right

  • At the critical field $h/J=1$, an RBM with $N_h \approx N/2$ hidden units—hence $\mathcal{O}(N^2)$ weights—is needed to reach the energy-error threshold, so the practical cost of reconstructing the state grows quadratically with qubit count.
  • The sample complexity at criticality grows only linearly with $N$, matching the PAC-learning scaling for predicting measurement outcomes; full tomography's exponential cost is not needed for this energy-estimation task.
  • Deep inside the ferromagnetic and paramagnetic phases the required hidden-unit count stays $\mathcal{O}(1)$, so only states near criticality drive the quadratic resource demand.
  • Many trained weights are redundant: pruning plus brief fine-tuning cuts the required nonzero weights sharply (e.g., from 760 to 119 at $N=40$), so over-parameterization serves the optimization procedure rather than the representational capacity of the state.

Reading between the lines

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

  • If the DMRG bond-dimension limitation is the cause of the anomalous peak at $h/J \approx 0.8$, then increasing the bond dimension until the expected Z2 symmetry is restored should shift the hardest-to-learn region to the critical point $h/J=1$; this is my inference, not a claim of the paper.
  • The paper's energy-based criterion likely undercounts resources needed for stronger notions of accuracy: requiring fidelity or correlation-function convergence would plausibly push the required hidden-unit count above $N/2$; the paper itself lists this caveat.
  • Because pruning reduces the active weights substantially, an alternative parameterization with fewer explicit weights (e.g., an autoregressive architecture) may achieve the same energy accuracy with sub-quadratic parameters at criticality; the current data do not settle that.
  • The over-parameterization finding suggests RBM training difficulty, not state complexity, sets the observed $\mathcal{O}(N^2)$ scaling; a direct test would be to vary the optimizer or initialization and see whether the minimal hidden-unit count moves, which the paper does not do.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies how the resources required to reconstruct a positive-real quantum wavefunction with a restricted Boltzmann machine (RBM) scale with system size. Training data are projective measurements in the sigma_z basis sampled from DMRG-MPS simulations of the one-dimensional transverse-field Ising model, and the learning criterion is a fixed upper bound on the relative error of the RBM energy estimator. The authors report that, near the quantum critical point, the minimum number of hidden units grows linearly in the number of qubits N, implying an O(N^2) growth in the number of RBM weights, while the sample complexity grows linearly in N. They also show that pruning small weights after training can substantially reduce the number of parameters while retaining the same energy accuracy, which they interpret as evidence that over-parameterization aids stochastic optimization.

Significance. If the reported scaling laws are correct, they provide concrete guidance for practical quantum-state reconstruction with RBMs on near-term devices: the energy can be learned with O(N^2) parameters and O(N) samples at criticality, a much gentler resource growth than full tomography. The paper is commendably explicit in defining a learning criterion and in benchmarking against DMRG-computed energies rather than against the RBM's own training objective, and the pruning results are a useful empirical observation about the role of over-parameterization. The consistency with Aaronson's PAC-learning picture is suggestive, although the authors are appropriately careful in framing it as consistency rather than a proof. The main weaknesses are that the fidelity of the DMRG training data itself is not independently established, and the asymptotic scaling claims rest on fits whose statistical quality is not reported.

major comments (3)
  1. [Section III A and Appendix] The training target is a DMRG-MPS distribution with bond dimension 2000, but the paper does not establish that this distribution faithfully represents the true TFIM ground state for all N and h studied. This concern is raised by the authors themselves: the required hidden-unit count peaks at h/J about 0.8 rather than at h/J=1, and the Appendix states that the expected Z2 symmetry is not realized for large N at low transverse field, with both effects attributed to the DMRG data. Because the scaling fits in Fig. 2 include h/J=0.6 and 0.7, at least part of the claimed linear growth in Nh(N) may be the cost of fitting a symmetry-broken, truncated MPS distribution rather than the physics of the TFIM. The authors should add explicit bond-dimension convergence checks (for example, repeating the dataset generation with chi=500, 1000, and 2000 and comparing energies, magnetizations, and sampled sigma_z statistics), and either restrict the scaling analysis to parameter regions where the DMRG target is demonstrably converged or show that the scaling conclusions are insensitive to the truncation.
  2. [Section III A, Fig. 2] The central O(N^2) parameter-scaling claim is an extrapolation from fits over N up to about 100, yet the fit parameters, confidence intervals, and residuals are not reported, and the Nh values are shown without error bars. Without these details the reader cannot assess whether the apparent linear growth in Nh persists asymptotically or is contaminated by finite-size curvature. Please report the fitted slope and intercept with uncertainties for each h/J value, the exact range of N used in each fit, and a stability check as the maximum system size is varied.
  3. [Section III B, Fig. 4] The sample-complexity claim M ~ N is based on increasing M in steps of 2500 until the ROE criterion is met and then averaging over initial conditions, but the number of trials, the spread across initializations, and the fit to M(N) are not reported. The central linear-scaling result needs at least error bars and, ideally, the fits for each fixed alpha = Nh/N, so the reader can gauge the significance of the linear dependence and the role of the alpha = 1/2 choice.
minor comments (4)
  1. [Section II B, Eq. (5)] The definition of the ROE involves a Gaussian confidence interval, but the number of RBM samples n used in the estimator is never specified in the text; please state the value of n and check that the 99% confidence statement is robust to the choice.
  2. [Section III A, Fig. 2] The upper and lower panels both include h/J = 1, which makes the figure slightly redundant and may confuse readers; consider plotting all field values in one panel or clearly labeling the duplicated curve.
  3. [Section III C, Table I] The pruning results are reported only for N up to 40, and the text notes that rare pruning events can seriously alter the ROE; please state how many independent training runs were used and quote a range or standard deviation for the pruned weight counts.
  4. [Appendix] The derivation of the bias-to-weight ratios in Eqs. (7) and (8) assumes that the RBM actually realizes the Z2-symmetric parameterization, but Fig. 5 shows only approximate agreement near -2; a sentence quantifying the observed deviations would be helpful.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the scaling laws are empirical measurements against an external DMRG benchmark, not outputs of a fit or a self-citation chain.

full rationale

The paper's central scaling claims are empirical observations. The RBM is trained by minimizing KL divergence via contrastive divergence, and the hidden-unit count at which the energy ROE falls below epsilon = 0.002 is measured against external DMRG energies. This is not a self-definitional loop because the energy is not a training target; it is evaluated post hoc on the trained model. The sample-complexity scaling is likewise measured by increasing M until the same external energy criterion is met. The paper explicitly compares its linear sample complexity to Aaronson's independent PAC-learning theorem, and the DMRG data come from ITensor, an external library. The only self-references are the use of the authors' own QuCumber package and prior RBM-state-reconstruction papers (e.g., Ref. [2]), but these are implementation/tooling citations and are not load-bearing for the scaling conclusions. The acknowledged limitation that DMRG bond dimension 2000 may break Z2 symmetry below certain fields is a data-fidelity or correctness caveat, not a circularity: it concerns whether the learned distribution is the true ground state, not whether the derivation reduces to its inputs. No equation in the paper is shown to be equivalent by construction to a fitted parameter, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the observed scalings.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The study introduces no new physical entities. It relies on two hand-chosen numerical parameters (the ROE threshold and the pruning schedule) and on standard mathematical and domain assumptions about the TFIM ground state, the DMRG approximation, and the training-data sampling process. The central scaling results are empirical observations against an external DMRG benchmark.

free parameters (2)
  • Relative observable error threshold epsilon = 0.002
    Chosen as a value achievable by RBMs trained with conventional algorithms for N <= 100; the scaling results are defined with respect to this threshold, and the authors note that different criteria could change the scaling.
  • Pruning schedule fractions = 40% of nonzero weights in the first iteration, then 5% per iteration
    Ad hoc choice for the pruning procedure; affects the reported pruned weight counts in Table I.
assumptions (3)
  • standard math Perron-Frobenius theorem guarantees the TFIM ground state has a positive-real wavefunction in the computational basis, enabling a direct mapping to a probability distribution.
    Invoked in Section II to justify using an RBM to model the wavefunction as the square root of a probability distribution.
  • domain assumption DMRG simulations with maximum bond dimension 2000 produce training data and exact energies that faithfully represent the true ground state for all N and h studied.
    The training data and benchmark energies come from ITensor DMRG. The paper itself reports that the expected Z2 symmetry is not realized at large N and low h, indicating the assumption can fail in a regime relevant to the scaling claim (Section III A, Fig. 1 inset discussion).
  • domain assumption Training samples are drawn independently from the target measurement distribution, matching the setting of Aaronson's PAC-learning theorem.
    Stated in Section III B as the reason the observed linear sample complexity is expected to be consistent with the theorem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The learnability scaling of quantum states: restricted Boltzmann machines." pith.science (2026). https://pith.science/paper/OQ4JXTEU

@misc{pith2026190807532,
  author       = {Pith},
  title        = {Pith review of: The learnability scaling of quantum states: restricted Boltzmann machines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQ4JXTEU}},
  note         = {Machine review of arXiv:1908.07532}
}
read the original abstract

Generative modeling with machine learning has provided a new perspective on the data-driven task of reconstructing quantum states from a set of qubit measurements. As increasingly large experimental quantum devices are built in laboratories, the question of how these machine learning techniques scale with the number of qubits is becoming crucial. We empirically study the scaling of restricted Boltzmann machines (RBMs) applied to reconstruct ground-state wavefunctions of the one-dimensional transverse-field Ising model from projective measurement data. We define a learning criterion via a threshold on the relative error in the energy estimator of the machine. With this criterion, we observe that the number of RBM weight parameters required for accurate representation of the ground state in the worst case - near criticality - scales quadratically with the number of qubits. By pruning small parameters of the trained model, we find that the number of weights can be significantly reduced while still retaining an accurate reconstruction. This provides evidence that over-parametrization of the RBM is required to facilitate the learning process.

Figures

Figures reproduced from arXiv: 1908.07532 by the authors.

Figure 1
Figure 1. FIG. 1. The procedure used to determine the RBM ex [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Minimum number of hidden units [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. FIG. 4. The minimum number of training examples [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: FIG. 5. The left and right figures show the values of the weight [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 26 canonical work pages

  1. [1]

    3, with Nh = 1 2N for h/J = 1)

    Start from the original, converged trained model (e.g., Fig. 3, with Nh = 1 2N for h/J = 1)

    work page

  2. [2]

    If a given|Wij| < δ, set Wij = 0, and freeze it for the following steps

    Set a threshold δ for the weight magnitudes. If a given|Wij| < δ, set Wij = 0, and freeze it for the following steps

    work page

  3. [3]

    Fine-tune the pruned model by running several more training iterations until the desired accuracy (as defined by the ROE learning criterion) is re- stored

    work page

  4. [4]

    original

    Repeat steps 2-4, pruning additional weights until the model fails to fulfill the learning criterion. We choose the pruning threshold such that 40% of the non-zero weights are pruned in the first iteration, and 5% of the non-zero weights in each following iteration. Note also that in this procedure we do not prune biases (see the Appendix for further commen...

    work page

  5. [5]

    Torlai and R

    G. Torlai and R. G. Melko, Physical Review B94, 165134 (2016)

    work page 2016

  6. [6]

    Torlai, G

    G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Nature Physics 14, 447 (2018)

    work page 2018

  7. [7]

    Carrasquilla, G

    J. Carrasquilla, G. Torlai, R. G. Melko, and L. Aolita, Nature Machine Intelligence 1, 155 (2019)

    work page 2019

  8. [8]

    J. Chen, S. Cheng, H. Xie, L. Wang, and T. Xiang, Phys. Rev. B 97, 085104 (2018)

    work page 2018

Show all 45 references
  1. [9]

    Cheng, J

    S. Cheng, J. Chen, and L. Wang, Entropy 20 (2018), 10.3390/e20080583

    2018 doi

  2. [10]

    Carleo, Y

    G. Carleo, Y. Nomura, and M. Imada, Nature Commu- nications 9, 5322 (2018)

    2018

  3. [11]

    Carleo and M

    G. Carleo and M. Troyer, Science 355, 602 (2017)

    2017

  4. [12]

    Torlai, B

    G. Torlai, B. Timar, E. P. van Nieuwenburg, H. Levine, A. Omran, A. Keesling, H. Bernien, M. Greiner, V. Vuleti, M. D. Lukin, R. G. Melko, and M. Endres, (2019), arXiv:1904.08441

    2019 arXiv

  5. [13]

    Torlai and R

    G. Torlai and R. G. Melko, (2019), arXiv:1905.04312

    2019 arXiv

  6. [14]

    R. G. Melko, G. Carleo, J. Carrasquilla, and J. I. Cirac, Nature Physics (2019)

    2019

  7. [15]

    Carleo, I

    G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto, and L. Zdeborova, (2019), arXiv:1903.10563

    2019 arXiv

  8. [16]

    Czischek, J

    S. Czischek, J. M. Pawlowski, T. Gasenzer, and M. Gart- tner, (2019), arXiv:1907.12844

    2019 arXiv

  9. [17]

    Paris and J

    M. Paris and J. Rehacek, eds., Quantum State Estima- tion, Lecture Notes in Physics, Vol. 649 (Springer-Verlag Berlin Heidelberg, 2004)

    2004

  10. [18]

    A. W. Sandvik, Phys. Rev. E 68, 056701 (2003)

    2003

  11. [19]

    E. M. Inack, G. Giudici, T. Parolini, G. Santoro, and S. Pilati, Phys. Rev. A 97, 032307 (2018)

    2018

  12. [20]

    Vidal, Phys

    G. Vidal, Phys. Rev. Lett. 99, 220405 (2007)

    2007

  13. [21]

    W. W. Ho and T. H. Hsieh, SciPost Phys. 6, 29 (2019). 7

    2019

  14. [22]

    W. W. Ho, C. Jonay, and T. H. Hsieh, Phys. Rev. A 99, 052332 (2019)

    2019

  15. [23]

    M. J. S. Beach, R. G. Melko, T. Grover, and T. H. Hsieh, (2019), arXiv:1904.00019

    2019 arXiv

  16. [24]

    A. J. Ferris and G. Vidal, Phys. Rev. B 85, 165146 (2012)

    2012

  17. [25]

    ITensor Library (version 2.0.11) http://itensor.org

  18. [26]

    D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, Cog- nitive Science 9, 147 (1985)

    1985

  19. [27]

    G. E. Hinton, Neural Computation 14, 1771 (2002), https://doi.org/10.1162/089976602760128018

    2002 doi

  20. [28]

    M. J. S. Beach, I. D. Vlugt, A. Golubeva, P. Huembeli, B. Kulchytskyy, X. Luo, R. G. Melko, E. Merali, and G. Torlai, SciPost Phys. 7, 9 (2019)

    2019

  21. [29]

    Aaronson, Proc

    S. Aaronson, Proc. R. Soc. A 463 (2007)

    2007

  22. [30]

    Livni, S

    R. Livni, S. Shalev-Shwartz, and O. Shamir, CoRR abs/1410.1141 (2014), arXiv:1410.1141

    2014 arXiv

  23. [31]

    Arora, N

    S. Arora, N. Cohen, and E. Hazan, CoRR abs/1802.06509 (2018), arXiv:1802.06509

    2018 arXiv

  24. [32]

    Allen-Zhu, Y

    Z. Allen-Zhu, Y. Li, and Z. Song, (2018), arXiv:1811.03962

    2018 arXiv

  25. [33]

    K. A. Sankararaman, S. De, Z. Xu, W. R. Huang, and T. Goldstein, (2019), arXiv:1904.06963

    2019 arXiv

  26. [34]

    Easing non-convex opti- mization with neural networks,

    D. Lopez-Paz and L. Sagun, “Easing non-convex opti- mization with neural networks,” (2018)

    2018

  27. [35]

    S. Han, J. Pool, J. Tran, and W. Dally, in Advances in Neural Information Processing Systems 28 , edited by C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett (Curran Associates, Inc., 2015) pp. 1135– 1143

    2015

  28. [36]

    D. C. Mocanu, E. Mocanu, P. H. Nguyen, M. Gibescu, and A. Liotta, Machine Learning 104, 243 (2016)

    2016

  29. [37]

    Gao and L.-M

    X. Gao and L.-M. Duan, Nature Communications 8, 662 (2017)

    2017

  30. [38]

    Glasser, N

    I. Glasser, N. Pancotti, M. August, I. D. Rodriguez, and J. I. Cirac, Phys. Rev. X 8, 011006 (2018)

    2018

  31. [39]

    Frankle and M

    J. Frankle and M. Carbin, (2018), arXiv:1803.03635

    2018 arXiv

  32. [40]

    H. Zhou, J. Lan, R. Liu, and J. Yosinski, (2019), arXiv:1905.01067

    2019 arXiv

  33. [41]

    N. Lee, T. Ajanthan, and P. H. S. Torr, (2018), arXiv:1810.02340

    2018 arXiv

  34. [42]

    Mazurenko, C

    A. Mazurenko, C. S. Chiu, G. Ji, M. F. Parsons, M. Kan´ asz-Nagy, R. Schmidt, F. Grusdt, E. Demler, D. Greif, and M. Greiner, Nature 545, 462 EP (2017)

    2017

  35. [43]

    B. P. Lanyon, C. Maier, M. Holz¨ apfel, T. Baumgratz, C. Hempel, P. Jurcevic, I. Dhand, A. S. Buyskikh, A. J. Daley, M. Cramer, M. B. Plenio, R. Blatt, and C. F. Roos, Nature Physics 13, 1158 (2017)

    2017

  36. [44]

    Keesling, A

    A. Keesling, A. Omran, H. Levine, H. Bernien, H. Pich- ler, S. Choi, R. Samajdar, S. Schwartz, P. Silvi, S. Sachdev, P. Zoller, M. Endres, M. Greiner, V. Vuletic, and M. D. Lukin, (2018), arXiv:1809.05540

    2018 arXiv

  37. [45]

    occupation

    A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, Nature 549, 242 EP (2017). A. Appendix Projective measurement data in the σz basis for the TFIM will generally obey Z2 symmetry in the absence of symmetry breaking, which may occur due to ...

    2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.