Pith. sign in

REVIEW 2 major objections 4 minor 5 references

Neural quantum states match or beat the quantum processor on annealing once Monte Carlo noise is removed.

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

2026-07-13 06:29 UTC pith:4HYFRLUF

load-bearing objection Solid, limited-scope reanalysis showing that MC noise, not NQS failure, drove the reported correlation error on one L=8 instance. the 2 major comments →

arxiv 2607.08811 v1 pith:4HYFRLUF submitted 2026-07-09 quant-ph

Comment on "Beyond-classical computation in quantum simulation"

Wladislaw Krinitsin , Nikita Alert , Matteo Rizzi , Markus Schmitt This is my paper
classification quant-ph PACS 03.67.Ac05.10.Ln75.10.Nr
keywords neural quantum statesquantum annealingMonte Carlo samplingcorrelation errortime-dependent variational principlespin glassquantum simulation
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.

A recent Science paper reported that neural quantum states (NQS) lag a superconducting quantum processor when both are used to simulate quantum annealing of a spin-glass Hamiltonian. This comment re-examines that comparison for the same two-dimensional lattices and short annealing times. The authors show that the apparent failure of NQS is largely an artifact of finite Monte Carlo sampling: long-range correlations are drowned by statistical noise of size roughly 1 over square-root of sample size, and the final Born distribution is so peaked that samples are strongly autocorrelated. Once the measured correlation error is decomposed into a systematic piece plus a statistical piece that falls as 1 over N_MC and is extrapolated to infinite samples, the NQS result lies below the quantum-processor error for the instances examined. The practical message is that a fair classical-versus-quantum benchmark must treat sampling budget and autocorrelation as first-class experimental parameters rather than fixed overheads.

Core claim

For the first disorder realization on the L=8 cylinder and annealing times up to 7 ns, an NQS evolved by the time-dependent variational principle yields a correlation error that, after extrapolation to infinite Monte Carlo samples, is smaller than the error reported for the quantum processor.

What carries the argument

The two-parameter decomposition of the correlation error, ε_c(N_MC) = sqrt(ε_sys + ε_stat / N_MC), which isolates the method-specific systematic error from the statistical sampling contribution and permits a controlled infinite-sample limit.

Load-bearing premise

The two-parameter fit cleanly extracts the true systematic error of the neural quantum state, and the single disorder instance studied is representative enough to revise the original comparison.

What would settle it

Repeat the identical NQS protocol on the same L=8 instance with several independent Markov chains and sample sizes large enough that the measured correlation error saturates; if the saturated value still exceeds the quantum-processor error, the extrapolated claim fails.

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

2 major / 4 minor

Summary. This Comment re-examines the Neural Quantum State (NQS) results of King et al. (Science 388, 199, 2025) for quantum annealing of a spin-glass Hamiltonian on an L=8 cylinder. Using the same Restricted Boltzmann Machine ansatz and TDVP evolution, the authors show that the reported correlation error ε_c is dominated by Monte-Carlo sampling noise once the two-point correlators fall below ~10^{-3}. They decompose ε_c(N_MC)=sqrt(ε_sys+ε_stat/N_MC), fit the two parameters to data at both t_a=2 ns and 7 ns, and find that the extrapolated systematic error lies below the QPU value for the first disorder realization. The analysis is restricted to that single instance and to annealing times ≤7 ns; residual autocorrelation in the peaked Born distribution is acknowledged as a practical sampling issue.

Significance. If the extrapolated ε_sys is reliable, the Comment supplies a concrete, reproducible counter-example to the claim that NQS cannot match QPU accuracy on this protocol. The Monte-Carlo variance argument is standard, the distance-resolved diagnostics are transparent, and the public jVMC codebase plus the original MPS ground-truth data make the result independently checkable. Even though the scope is limited to one disorder instance and short annealing times, the work usefully clarifies how sampling budget and autocorrelation must be controlled before classical variational methods are declared inferior, and it motivates broader NQS benchmarks on the more challenging geometries and low-precision couplings studied in the original Science paper.

major comments (2)
  1. The central claim that NQS surpasses the QPU rests on the two-parameter fit of Eq. (7) performed on a single disorder realization (the first instance supplied in the Zenodo dataset). While the fit quality in Fig. 1c is good, no uncertainty on the extrapolated ε_sys is reported, nor is any second instance examined. Without at least a bootstrap error bar or a second independent disorder sample, it remains unclear whether the reported superiority is robust or instance-specific.
  2. The text correctly notes that long autocorrelation times appear for t_a=7 ns, yet the effective sample size that enters the fit of Eq. (7) is never quantified. If residual autocorrelation is absorbed into ε_sys rather than into an adjusted N_MC, the extrapolated systematic error is biased low. A short autocorrelation-time measurement (or a thinned-sample re-fit) for the t_a=7 ns data would remove this ambiguity.
minor comments (4)
  1. Fig. 1 caption refers to panels a)–c) while the main text later mentions “Fig. 1 d)”; the panel labels should be made consistent.
  2. The definition of the four distance classes D_k uses d_max=√65; a brief parenthetical that this is the maximum Euclidean distance on the (8,8) cylinder would aid readers.
  3. The phrase “the first disorder realization for the couplings J_ij on the L=8 lattice, as it can be found in the dataset [2]” should include the precise file or array index so that the calculation is fully reproducible.
  4. A short sentence clarifying that the same RBM architecture and TDVP hyperparameters as King et al. were used (or listing any deviations) would strengthen the direct comparison.

Circularity Check

0 steps flagged

No significant circularity; NQS error extrapolation is a standard statistical decomposition against independent MPS ground truth.

full rationale

The central claim rests on re-evolving an RBM NQS (same hyperparameters as the original work) with the public jVMC codebase on the first published disorder instance, estimating two-point correlators by Monte Carlo at varying sample budgets, and fitting the observed correlation error ε_c to the two-parameter form ε_c(N_MC)=sqrt(ε_sys+ε_stat/N_MC). That functional form follows directly from the definition of the relative L2 error together with the elementary sampling variance of Pauli correlators (Var[c_ij]≈1); it is not a tautology that forces the ranking versus the QPU. Ground-truth correlators are taken from the well-converged MPS data released with the Science paper, an independent external benchmark. No parameter is fitted to the target superiority claim, no uniqueness theorem is imported from prior author work, and the sole self-reference is the authors’ open-source code used merely as a computational tool. The derivation is therefore self-contained and exhibits no circular reduction by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on standard Monte-Carlo statistics, the published MPS ground truth, and a two-parameter empirical fit that separates systematic from statistical error. No new physical entities are introduced; the free parameters are simply the fitted intercepts and slopes of the error curves. Domain assumptions about the Born distribution and the form of the annealing Hamiltonian are taken unchanged from the original Science paper.

free parameters (2)
  • ε_sys (systematic correlation error) = extracted from Fig. 1c fits (numerical values not tabulated)
    Intercept of the two-parameter fit Eq. (7) to the measured ε_c versus N_MC; its value is read off from data for each annealing time and is the quantity claimed to be smaller than the QPU error.
  • ε_stat (statistical prefactor) = extracted from Fig. 1c fits
    Slope parameter of the same fit; absorbs the summed variances of all correlation pairs and any residual autocorrelation.
axioms (3)
  • standard math Monte-Carlo variance of each two-point correlator is approximately 1, so the sampling error scales as 1/sqrt(N_MC).
    Invoked explicitly in Eqs. (5)–(6) and used to explain the plateau in Fig. 1a.
  • ad hoc to paper The correlation error admits the additive decomposition ε_c^{2} = ε_sys + ε_stat / N_MC.
    Derived in the paragraph containing Eq. (7); assumes statistical and systematic contributions add under the square root after summing over all pairs.
  • domain assumption Well-converged MPS simulations supply the ground-truth correlators ~c_ij.
    Taken from the original Science paper and its Zenodo dataset; used as the reference in the definition of ε_c.

pith-pipeline@v1.1.0-grok45 · 9444 in / 2599 out tokens · 30306 ms · 2026-07-13T06:29:43.815170+00:00 · methodology

0 comments
read the original abstract

A recent article [Science 388, 199-204 (2025)] investigates the applicability of numerical methods and a quantum processor unit in simulating a quantum annealing protocol. One of the findings indicates that Neural Quantum States - a versatile variational ansatz for the many-body wave function based on artificial neural networks - fail to reach the same accuracy as the quantum processor. In this comment we revisit these concerns, demonstrating that NQS can provide competitive results in some of the cases when accounting for the Monte-Carlo noise and large autocorrelation times between samples obtained from the final state.

Figures

Figures reproduced from arXiv: 2607.08811 by Markus Schmitt, Matteo Rizzi, Nikita Alert, Wladislaw Krinitsin.

Figure 1
Figure 1. Figure 1: FIG. 1. a) Comparison between MPS and NQS of the distance-resolved magnitude of correlations (4) for [PITH_FULL_IMAGE:figures/full_fig_p002_1.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

5 extracted references · 1 canonical work pages

  1. [1]

    A. D. King, A. Nocera, M. M. Rams, J. Dziarmaga, R. Wiersema, W. Bernoudy, J. Raymond, N. Kaushal, N. Heinsdorf, R. Harris, K. Boothby, F. Altomare, M. Asad, A. J. Berkley, M. Boschnak, K. Chern, H. Chris- tiani, S. Cibere, J. Connor, M. H. Dehn, R. Deshpande, S. Ejtemaee, P. Farre, K. Hamer, E. Hoskinson, S. Huang, M. W. Johnson, S. Kortas, E. Ladizinsky...

    doi:10.1126/science.ado6285 2025

  2. [2]

    A. D. King, Beyond-classical computation in quantum simulation, doi:10.5281/zenodo.14063693 (2024)

    work page doi:10.5281/zenodo.14063693 2024

  3. [3]

    Nocera, J

    A. Nocera, J. Raymond, W. Bernoudy, M. H. Amin, and A. D. King, Evaluating classical simulations with a quan- tum processor (2025), arXiv:2508.15759 [quant-ph]

    Pith/arXiv arXiv 2025

  4. [4]

    Mauron and G

    L. Mauron and G. Carleo, Challenging the quantum ad- vantage frontier with large-scale classical simulations of annealing dynamics (2025), arXiv:2503.08247 [quant-ph]

    Pith/arXiv arXiv 2025

  5. [5]

    Schmitt and M

    M. Schmitt and M. Reh, jVMC: Versatile and performant variational Monte Carlo leveraging automated differentia- tion and GPU acceleration, SciPost Phys. Codebases , 2 (2022)

    2022