pith. sign in

arxiv: 2605.28861 · v1 · pith:BFPMLJHBnew · submitted 2026-05-21 · ❄️ cond-mat.str-el · cond-mat.dis-nn· cs.LG

Comment on "Spin-1/2 Kagome Heisenberg Antiferromagnet: Machine Learning Discovery of the Spinon Pair-Density-Wave Ground State"

Helia Kamal , Dominik Kufel , DinhDuy Vu , Chris R. Laumann , Norman Y. Yao This is my paper

Pith reviewed 2026-06-30 15:50 UTC · model grok-4.3

classification ❄️ cond-mat.str-el cond-mat.dis-nncs.LG
keywords kagome Heisenberg antiferromagnetneural network variational Monte CarloergodicityMetropolis-Hastings samplingspinon pair-density-waveDMRGspin liquidMarkov chain
0
0 comments X

The pith

Low energies reported for kagome magnet neural network arise from frozen Markov chains in sampling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This comment paper examines a recent neural-network variational Monte Carlo study that reported energies for the spin-1/2 kagome Heisenberg antiferromagnet lower than density matrix renormalization group benchmarks and identified a spinon pair-density-wave ground state on the N=108 cluster. The analysis shows that the single-spin-flip update rule in the Metropolis-Hastings sampler traps the Markov chains and prevents full exploration of spin configurations. Replacing that rule with spin-exchange updates restores ergodicity and drives the neural network to energies above the DMRG values. The discrepancy indicates that the reported low-energy state is an artifact rather than a true variational minimum.

Core claim

The variational energies reported in the original work are artifacts of broken ergodicity in the Metropolis-Hastings sampling with single-spin-flip updates; when ergodic sampling is enforced via spin-exchange updates, the neural network converges to energies significantly higher than existing DMRG results.

What carries the argument

Ergodicity of the Markov chain in Metropolis-Hastings sampling for neural-network wave functions, controlled by the choice between single-spin-flip and spin-exchange local updates.

If this is right

  • The spinon pair-density-wave state is not supported as the ground state of the neural-network ansatz under ergodic sampling.
  • Density matrix renormalization group calculations remain the lowest variational upper bounds obtained so far for this system size.
  • Neural-network variational Monte Carlo studies of quantum magnets require explicit checks that the chosen update rule produces ergodic chains.
  • Machine-learning claims of new ground-state phases must be cross-validated with multiple sampling schemes before they can be considered reliable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar ergodicity problems are likely to appear in other variational Monte Carlo applications to frustrated spin models where configuration space is constrained.
  • Studies using neural-network ansatzes should routinely compare results across independent update families to confirm that reported minima are not local.
  • The gap between neural-network and DMRG energies after ergodicity correction may reflect limits in the expressivity of the chosen network architecture or in the optimization procedure itself.
  • This example underscores the value of testing sampling procedures on benchmark clusters where exact or near-exact references exist.

Load-bearing premise

Spin-exchange updates achieve sufficient ergodicity on the N=108 cluster to reach the true variational minimum of the neural-network ansatz.

What would settle it

A side-by-side comparison, on smaller clusters with known exact energies, of the minimum reached by the identical neural network when using single-spin-flip updates versus spin-exchange updates.

Figures

Figures reproduced from arXiv: 2605.28861 by Chris R. Laumann, DinhDuy Vu, Dominik Kufel, Helia Kamal, Norman Y. Yao.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
read the original abstract

A recent article [Phys. Rev. X 15, 011047 (2025)] utilizes group-equivariant convolutional neural networks to study the ground state of the kagome Heisenberg antiferromagnet. On the largest finite-size cluster studied to date ($N=108$), the authors report variational energies significantly lower than other numerical methods, including state-of-the-art density matrix renormalization group (DMRG) calculations. In contrast to previous results suggesting a possible spin-liquid ground state, the authors observe a spinon pair-density-wave ground state. We find that: (i) the reported low energies are artifacts of broken ergodicity in the Metropolis--Hastings sampling, since the single-spin-flip update rule utilized by the authors effectively freezes the Markov chains; and (ii) when ergodic sampling is enforced via spin-exchange updates, the neural network converges to energies significantly higher than existing DMRG results, calling the paper's claims into question.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript is a comment arguing that the low variational energies obtained by group-equivariant convolutional neural networks for the spin-1/2 kagome Heisenberg antiferromagnet on the N=108 cluster (reported in Phys. Rev. X 15, 011047 (2025)) are artifacts of broken ergodicity under single-spin-flip Metropolis-Hastings updates. It further claims that replacing these with spin-exchange updates produces energies significantly above existing DMRG benchmarks, thereby undermining the original identification of a spinon pair-density-wave ground state.

Significance. If the central claims hold, the comment would establish that standard single-spin-flip sampling is inadequate for neural-network variational Monte Carlo on frustrated magnets and would restore consistency with DMRG results, thereby affecting the interpretation of the kagome ground state. The work draws on textbook Markov-chain properties and external DMRG benchmarks without introducing fitted parameters.

major comments (2)
  1. [Abstract] Abstract: the claim that spin-exchange updates render the Metropolis-Hastings chain ergodic enough for the neural-network ansatz to reach its true variational minimum on N=108 is not accompanied by quantitative diagnostics (autocorrelation times, integrated autocorrelation lengths, or convergence statistics from independent initializations). Without these, the reported energies higher than DMRG could still reflect residual sampling bias rather than an intrinsic limitation of the ansatz.
  2. [Abstract] The manuscript contrasts the two update rules and reports the resulting energy discrepancy, but does not demonstrate that the single-spin-flip chains are demonstrably frozen (e.g., via acceptance-rate collapse or trapped magnetization sectors) on the specific N=108 cluster; this leaves the attribution of the original low energies as purely an ergodicity artifact partially unanchored.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading of our comment. We address the two major comments point by point below, indicating where we agree that additional material would strengthen the presentation and where we maintain that the existing evidence is sufficient.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that spin-exchange updates render the Metropolis-Hastings chain ergodic enough for the neural-network ansatz to reach its true variational minimum on N=108 is not accompanied by quantitative diagnostics (autocorrelation times, integrated autocorrelation lengths, or convergence statistics from independent initializations). Without these, the reported energies higher than DMRG could still reflect residual sampling bias rather than an intrinsic limitation of the ansatz.

    Authors: We agree that explicit autocorrelation diagnostics would add quantitative support. The energy difference we report (well above DMRG benchmarks) is already inconsistent with residual bias under spin-exchange updates, given that the neural-network ansatz is variationally bounded from below. In revision we will add a short paragraph reporting integrated autocorrelation lengths and results from independent initializations on the N=108 cluster to make this explicit. revision: partial

  2. Referee: [Abstract] The manuscript contrasts the two update rules and reports the resulting energy discrepancy, but does not demonstrate that the single-spin-flip chains are demonstrably frozen (e.g., via acceptance-rate collapse or trapped magnetization sectors) on the specific N=108 cluster; this leaves the attribution of the original low energies as purely an ergodicity artifact partially unanchored.

    Authors: The primary evidence for freezing remains the unphysical result that single-spin-flip sampling produces variational energies below all published DMRG values on the same cluster; a correctly sampled variational wave function cannot do so. While direct acceptance-rate or sector-trapping data for the original N=108 runs are not reproduced in our comment (as they were not reported in the target paper), the systematic energy lowering with single-spin-flip updates across multiple neural-network architectures is the expected signature of broken ergodicity in frustrated magnets. We will add a concise paragraph recalling this textbook diagnostic in the revised manuscript. revision: partial

Circularity Check

0 steps flagged

No circularity: argument rests on standard MCMC properties and external DMRG benchmarks

full rationale

The comment identifies broken ergodicity from single-spin-flip updates and reports higher energies under spin-exchange updates, directly comparing to independent DMRG results. No self-definitional steps, fitted parameters renamed as predictions, or load-bearing self-citations appear; the central claims invoke textbook Markov-chain mixing and externally published DMRG energies rather than reducing to the paper's own inputs or prior author work. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the standard requirement that Markov-chain Monte Carlo sampling must be ergodic to converge to the target distribution, plus the assumption that DMRG supplies a reliable external benchmark; both are drawn from established literature rather than introduced ad hoc.

axioms (1)
  • standard math Metropolis-Hastings sampling with a given update rule must be ergodic to produce unbiased estimates of variational energies.
    Standard convergence requirement for MCMC methods invoked when contrasting single-spin-flip and spin-exchange updates.

pith-pipeline@v0.9.1-grok · 5730 in / 1389 out tokens · 68053 ms · 2026-06-30T15:50:31.670485+00:00 · methodology

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

21 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    This results in non- ergodic sampling

    The single-spin-flip update rule used for the 108- site cluster is incompatible with an important sym- metry of the system, namely, the conservation of total magnetization,S z tot. This results in non- ergodic sampling. As the neural network learns to concentrate around the physicalS z tot = 0sec- tor, the acceptance probability for single-spin-flip updat...

  2. [2]

    Comment on "Spin-1/2 Kagome Heisenberg Antiferromagnet: Machine Learning Discovery of the Spinon Pair-Density-Wave Ground State"

    For the largestN= 108cluster, when the neural quantum state ansatz is properly optimized using a sampler based on spin-exchange updates, the varia- tional energy converges stably but remains≈3.5% higher than the DMRG benchmark [Fig. 1(b)]. We note that this behavior is consistent with the method’s performance on smaller clusters (also us- ing the spin-exc...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Duric, J

    T. Duric, J. H. Chung, B. Yang, and P. Sengupta, Phys. Rev. X15, 011047 (2025)

    2025

  4. [4]

    A. M. Läuchli, J. Sudan, and R. Moessner, Phys. Rev. B 100, 155142 (2019)

    2019

  5. [5]

    Y. Ran, M. Hermele, P. A. Lee, and X.-G. Wen, Phys. Rev. Lett.98, 117205 (2007)

    2007

  6. [6]

    Iqbal, F

    Y. Iqbal, F. Becca, and D. Poilblanc, Phys. Rev. B84, 020407 (2011)

    2011

  7. [7]

    Iqbal, D

    Y. Iqbal, D. Poilblanc, and F. Becca, Phys. Rev. B89, 020407 (2014)

    2014

  8. [8]

    Jiang, Z.-Y

    H.-C. Jiang, Z.-Y. Weng, and D. N. Sheng, Phys. Rev. Lett.101, 117203 (2008)

    2008

  9. [9]

    Jiang, Z

    H.-C. Jiang, Z. Wang, and L. Balents, Nat. Phys.8, 902 (2012)

    2012

  10. [10]

    S. Yan, D. A. Huse, and S. R. White, Science332, 1173 (2011)

    2011

  11. [11]

    Depenbrock, I

    S. Depenbrock, I. P. McCulloch, and U. Schollwöck, Phys. Rev. Lett.109, 067201 (2012)

    2012

  12. [12]

    Y.-C. He, M. P. Zaletel, M. Oshikawa, and F. Pollmann, Phys. Rev. X7, 031020 (2017)

    2017

  13. [13]

    Z.-Y. Xie, J. Chen, J.-F. Yu, X. Kong, B. Normand, and T. Xiang, Phys. Rev. X4, 011025 (2014)

    2014

  14. [14]

    Mei, J.-Y

    J.-W. Mei, J.-Y. Chen, H. He, and X.-G. Wen, Phys. Rev. B95, 235107 (2017)

    2017

  15. [15]

    Jiang, P

    S. Jiang, P. Kim, J. H. Han, and Y. Ran, SciPost Phys. 7, 006 (2019)

    2019

  16. [16]

    Roth and A

    C. Roth and A. H. MacDonald, arXiv:2104.05085 (2021)

    work page arXiv 2021

  17. [17]

    C. Roth, A. Szabó, and A. H. MacDonald, Phys. Rev. B 108, 054410 (2023)

    2023

  18. [18]

    and Bagrov, A.A., 2020

    Westerhout, T., Astrakhantsev, N., Tikhonov, K.S., Kat- snelson, M.I. and Bagrov, A.A., 2020. Generalization properties of neural network approximations to frus- trated magnet ground states.Nature communications, 11(1), p.1593

    2020

  19. [19]

    and Heyl, M., 2024

    Chen, A. and Heyl, M., 2024. Empowering deep neu- ral quantum states through efficient optimization.Nature Physics,20(9), pp.1476-1481

    2024

  20. [20]

    and Carleo, G., 2022

    Vicentini, F., Hofmann, D., Szabó, A., Wu, D., Roth, C., Giuliani, C., Pescia, G., Nys, J., Vargas-Calderón, V., Astrakhantsev, N. and Carleo, G., 2022. NetKet 3: Machine learning toolbox for many-body quantum sys- tems. SciPost Physics Codebases, p.007

    2022

  21. [21]

    and Sengupta, P., 2025

    Ðurić, T. and Sengupta, P., 2025. 2k F instability and chiral spin density wave at the 1/9 magnetization plateau in the kagome antiferromagnets.arXiv preprint arXiv:2512.11670

    work page arXiv 2025