Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Symmetry Constraints Regularize Neural Quantum State Learning

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 Symmetry constraints hard-wired into neural quantum states remove redundant parameters and concentrate the reachable state space around low-energy solutions, speeding up training without sacrificing accuracy.

desk verdict A clean analytic framework for symmetry-compiling NQS, with credible compression results, but the headline geometric mechanism rests on a coordinate-space proxy that does not measure Fubini–Study volume. read the letter →

arxiv 2608.08798 v1 pith:L7WDOQ6O submitted 2026-08-09 quant-ph

Turbasu Chatterjee , Manas Sajjan , Songbo Xie , Elliott Love , Vinit Singh , Bojko N. Bakalov , Sabre Kais This is my paper · ORCID
classification quant-ph
keywords neuralquantumstatessymmetrycompilationFubini-Studymetricvariationaloptimizationgeometrytransverse-fieldIsingmodelXXZspinchainparameterreductionmany-bodylearning
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 tries to establish that a neural quantum state (NQS) learns many-body ground states faster and more reliably when the Hamiltonian's symmetries are compiled directly into the variational parameters before training, rather than imposed softly or left to the optimizer. For Boltzmann-family NQS built from local Pauli-$Z$ generators, symmetry is enforced by tying coefficients that sit on the same physical orbit under translations, reflections, space groups, or global bitflip, which collapses the trainable coefficient space analytically before any optimization. The key claim is that this is not just parameter pruning: measured with the paper's new Fubini-Study based metrics $f_\epsilon$ and $R_\epsilon$, the symmetry-reduced ansatz concentrates a larger fraction of its reachable state space inside the target-accurate low-energy region, with the advantage growing with system size. On transverse-field Ising and XXZ spin chains, compressed ansätze keep ground-state energy errors in the $10^{-4}$ to $10^{-3}$ range while cutting the parameter count from 210 to 10--11 and delivering large wall-time speedups. A sympathetic reader should care because the paper offers a quantitative, coordinate-invariant way to judge where an ansatz's expressive power is spent.

What carries the argument

The carrying object is the symmetry-constrained coefficient subspace of the NQS ansatz: amplitude and phase generators expanded in $k$-local Pauli-$Z$ strings, with coefficient vectors $\vec{c}$ and $\vec{d}$. Symmetry operations act as permutations on the string supports, and requiring invariance yields a linear constraint matrix $V$ whose blocks are $I-P_g$ for translations, reflections, and point-group elements plus a bitflip parity block; the trainable coordinates are then the null-space coordinates $\vec{\xi}$ with $\vec{c}=U_V\vec{\xi}$. The geometric diagnostics are built on the pullback Fubini-Study metric $S=\mathrm{Re}[J^\dagger\Pi_\perp J]$ from the state Jacobian, the Fubini-Study-normalized curvature operator $K_* = S^{-1/2} M_* S^{-1/2}$, and the Theorem 1 volume formula $V_{\mathrm{good},+}^{\mathrm{FS}} = \frac{\pi^{r_+/2}}{\Gamma(r_+/2+1)}\frac{(2\epsilon)^{r_+/2}}{\sqrt{\det_+ K_*}}$, which turns local positive curvature into an $\epsilon$-good basin volume; $f_\epsilon$ divides that volume by the total reachable Fubini-Study volume and $R_\epsilon$ converts it to a rank-normalized linear scale.

What would settle it

Evaluate the exact Fubini-Study volumes for a small critical TFIM instance (e.g. $N=8$ or 12): compute the full Jacobian and Hessian without perturbation sampling, integrate Eq. (40) directly, and compare the resulting $\log_{10} f_\epsilon$ and $\log_{10} R_\epsilon$ ordering of unconstrained versus translation-constrained ansätze with the paper's sampled estimates. If the exact calculation shows the constrained manifold has no larger useful-volume fraction, or the ordering reverses at any tolerance, the geometric-regularization conclusion fails.

Watch

Extended reading notes

Core claim

Embedding the symmetry group of the Hamiltonian into the coefficient space of a Boltzmann-family neural quantum state removes symmetry-redundant training directions before optimization starts, and this removal changes the geometry of the loss landscape rather than merely shrinking it. The paper reports that translation-based symmetry families reduce the trainable parameters for $N=20$ TFIM and XXZ benchmarks from 210 to 10--11 while median endpoint energy-density errors remain on the same $10^{-4}$ scale as the unconstrained model; for $N=128$ TFIM the largest constraints compress 8256 parameters to 64 and cut VMC wall time by roughly 30 times. Using the target-aware useful expressibility $f_\epsilon$ (the Fubini-Study fraction of the reachable manifold lying within an energy tolerance of the target) and the characteristic basin scale $R_\epsilon$, the paper finds that constrained ansätze have larger $f_\epsilon$ than the unconstrained one at every system size and tolerance tested, with translation and space-group sectors highest and the gap widening with system size. It concludes that symmetry compilation regularizes learning by concentrating the reachable physical manifold around low-energy states while keeping the retained target-accurate basins broad.

Load-bearing premise

The numerical geometry claims depend on treating 128 random perturbations in a small parameter-space ball around the optimized endpoint, plus 16 outer samples, as an accurate proxy for the true Fubini-Study volume of the epsilon-good region; if that proxy is not faithful, the reported $f_\epsilon$ and $R_\epsilon$ values do not establish the geometric-regularization mechanism.

Editorial extensions

If this is right

  • For any Hamiltonian with an exact discrete symmetry, the parameter compression is known before training begins: Burnside orbit counting gives the dimension of the reduced coefficient space, so model size and training cost become predictable.
  • Translation and space-group constraints are the ones that matter most: they dominate the increase in $f_\epsilon$ and $R_\epsilon$, while bitflip alone gives the smallest gain, so symmetry choice can be guided by geometry rather than trial and error.
  • The geometric advantage of symmetry compilation grows with system size, meaning the method is most valuable precisely where NQS training is hardest.
  • Because the compressed models keep ground-state energy errors within the same range as unconstrained models while using roughly $20\times$ fewer parameters, larger systems can be studied with the same computational budget.
  • The $f_\epsilon$ and $R_\epsilon$ pair provides a coordinate-invariant comparison of different ansätze on the same physical footing, so it can be used to judge whether any proposed pruning or architectural change genuinely improves target alignment.

Reading between the lines

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

  • If the sampled perturbation proxy faithfully estimates Fubini-Study volume, the $f_\epsilon$ diagnostic could be computed on partially trained or even untrained ansätze to rank candidate symmetry constraints before running full optimization, something the paper does not claim.
  • The same orbit-tying logic should extend to other variational families with diagonal coefficient expansions, such as autoregressive or correlator-product states, and the geometric metrics would allow a fair comparison; this is a generalization the paper only gestures at.
  • A testable prediction beyond the paper: for a Hamiltonian whose symmetry is only approximate, soft tying should interpolate between the unconstrained and hard-constrained geometry, and $f_\epsilon$ should peak at the optimal tie strength rather than at maximal constraint.
  • The paper's geometry is evaluated at the optimized endpoint; tracking $f_\epsilon$ along a trajectory (as the appendices do for proxies) could reveal whether symmetry also shortens the transient phase of training, not just the final basin.
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 proposes to regularize neural quantum state (NQS) optimization by hard-wiring Hamiltonian symmetries into the variational parameterization. For a Boltzmann-family ansatz with diagonal Pauli- Z generators, symmetry constraints are imposed by tying coefficients along orbits of the physical symmetry group, yielding an analytic reduction of the trainable parameter space before optimization. The paper derives the unconstrained and constrained Jacobians, the energy Hessian, and a Fubini–Study-based geometric framework culminating in Theorem 1, which gives the local volume of the epsilon-good region from the positive-curvature spectrum of a Fubini–Study-normalized curvature operator. Numerically, the paper reports TFIM and XXZ ground-state energy benchmarks at N=20, parameter-compression and runtime data up to N=128, and geometric diagnostics (f_epsilon and R_epsilon) for the critical TFIM across N=8 to 64. The central claims are that symmetry compilation preserves variational accuracy while drastically reducing parameters, and that it improves both the global concentration and the local geometry of target-accurate states, with the geometric advantage growing with system size.

Significance. If established, the paper would provide a useful framework for quantifying how symmetry constraints alter the optimization landscape of NQS: the algebraic symmetry-constraint construction is clean and gives explicit, testable parameter counts via orbit counting, and the analytic derivation of the local epsilon-good volume in Theorem 1 is a valuable contribution in its own right. The runtime and compression results (Tables I and Figure 5) are concrete and likely reproducible. However, the numerical support for the central geometric-regularization claim currently rests on coordinate-space Monte Carlo proxies rather than on the Fubini–Study volumes defined in the paper. The analytic framework is sound, but the quantitative conclusions drawn from Figures 6 and 7 are not established by the reported numerical evidence.

major comments (3)
  1. [Section IV; Definition 1; Eq. (38); Figs. 6–7]
  2. [Section IV; Theorem 1; Appendix E.3]
  3. [Section V.B; Figs. 6–7; Table I]
minor comments (4)
  1. [Section VI]
  2. [Section II.A; Ref. [50]]
  3. [Appendix E]
  4. [Section VII]

Circularity Check

1 steps flagged · score 6.0 of 10

The global-concentration claim reduces to coordinate-ball sampling: f_epsilon is defined as a Fubini–Study volume ratio but implemented with coordinate-space ball proxies that inject a fixed (0.05)^d factor, so the hierarchy tracks parameter count by construction.

  1. self definitional [Definition 1, Eq. (38) and Section IV, 'Numerical Implementation and Reproducibility']
    "Definition 1: 'fϵ(x∗) ≡ VolFS[Gϵ(x∗)] / VolFS(MΘ).' Section IV: 'The local good-volume proxy was estimated from 128 perturbations sampled uniformly from a parameter-space ball of radius 0.05 around the optimized endpoint. ... The global Fubini–Study normalization was estimated once for each system-size and symmetry-family pair using reference seed 0, with 16 outer samples drawn from a unit-radius parameter-space ball centered at the origin.'"

    The quantity reported as f_epsilon is not computed from the Fubini–Study volumes of Definition 1; it is replaced by coordinate-space ball sampling. In d-dimensional coordinate space the radius-0.05 ball has Lebesgue volume (0.05)^d times the unit-ball volume, so the surrogate log10 f_proxy contains the additive term d log10(0.05). For the N=20 TFIM, d_un=210 and d_con=10, giving a fixed separation 200 log10(0.05) ≈ −260 that is fixed by the parameter counts and the chosen radii before any state-space geometry is evaluated. The constrained–unconstrained gap in Fig. 6 therefore reduces by construction to the parameter reduction the paper aims to explain, rather than to a measured concentration of Fubini–Study volume around low-energy states.

full rationale

The parameter-compression and accuracy claims are independently supported: Burnside counting fixes the reduced dimensions, and the VMC energies are benchmarked against exact diagonalization, so those results are not circular. The factored-ansatz expressivity is attributed to the same-group preprint [50], but the factored form is also supported by [27] and by the empirical accuracy benchmarks, so that self-citation is not load-bearing. The central geometric claim, however, is partially circular. Definition 1 defines f_epsilon as a ratio of Fubini–Study volumes, but Section IV implements it with local and global coordinate-space ball proxies (radius 0.05 and unit radius). Under that implementation the result inherits a factor (0.05)^d, so the large-N separation between unconstrained (large d) and symmetry-constrained (small d) manifolds is dominated by the parameter-count difference itself. The R_epsilon basin-scale diagnostic and the runtime speedups remain independent evidence, which keeps the circularity partial rather than total. Because the global-concentration prediction reduces by construction to the input parameter reduction, the paper earns a 6 rather than a lower score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim depends on a few hand-set modeling constants (beta, theta_j, kmax, thresholds) and on two ad hoc approximations in the geometric estimator: the radius-0.05 coordinate-ball proxy for local Fubini-Study volume and the 16-sample global normalization. No new physical entities or forces are introduced. The constraints themselves are derived from standard group theory.

free parameters (4)
  • beta (inverse temperature scaling factor) = Not stated
    Appears in Eq. (3) as the fixed exponent controlling the sharpness of the amplitude distribution. No value or tuning procedure is reported in the numerical section.
  • theta_j (reference local sigma-x rotation angles) = Not stated
    The reference product state in Eq. (3) uses local sigma-x rotations with angles theta_j; these set the baseline amplitudes that enter every benchmark, yet no values are reported.
  • kmax locality cutoff = 2
    All generators are restricted to at most two-body Pauli-Z strings. The compression counts, accuracy results, and expressibility of the ansatz depend on this truncation.
  • Diagnostic thresholds (tau_rank, tau_FS, positive-curvature cutoff) = tau_rank=1e-10, tau_FS=1e-12 and 1e-8, positive-curvature threshold not fully specified
    The reported values of r+, f_epsilon, and R_epsilon depend on these hand-selected numerical thresholds, and no sensitivity analysis is reported in the main text.
assumptions (5)
  • standard math The columns of the design matrix W are orthogonal characters of Z_2^n, so W^T W = 2^n I_{M_k} and the k-local coefficient-to-spectrum map is injective.
    Used in Section II.B to justify that the restricted coefficient vector is recoverable from the eigenvalue field. This is textbook Fourier analysis on the Boolean cube.
  • domain assumption The factored ansatz of Eq. (3) with k=2 diagonal generators is expressive enough to represent the TFIM and XXZ ground states within the claimed accuracy.
    The paper cites universal expressivity from the same-group preprint [50], but for the specific k=2 truncation only empirical energy and fidelity results support the claim.
  • domain assumption The optimized endpoints used for geometry calculations are stationary local minima with no resolved negative-curvature modes, as required by Theorem 1.
    Appendix E states that only endpoints satisfying this criterion should enter the reported local-volume statistics, but the main text does not demonstrate this for every reported point.
  • ad hoc to paper A radius-0.05 coordinate-space ball around the optimized endpoint is a faithful proxy for the Fubini-Study-local good volume in Definition 1.
    Introduced in Section IV as the local good-volume estimator. Coordinate-space perturbations are not metric-normalized, and the paper does not show that this proxy approximates the exact Fubini-Study volume.
  • ad hoc to paper The global Fubini-Study normalization Vol_FS(M) can be estimated from 16 outer samples in a unit ball centered at the origin.
    Section IV reports this 16-sample estimate for each system-size and symmetry-family pair; this is extremely sparse and could materially bias the f_epsilon denominator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetry Constraints Regularize Neural Quantum State Learning." pith.science (2026). https://pith.science/paper/L7WDOQ6O

@misc{pith2026260808798,
  author       = {Pith},
  title        = {Pith review of: Symmetry Constraints Regularize Neural Quantum State Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7WDOQ6O}},
  note         = {Machine review of arXiv:2608.08798}
}
abstract

Neural quantum states (NQS) offer highly expressive variational wavefunctions, but their optimization is frequently bottlenecked by redundant parameters and poorly conditioned landscapes. We demonstrate that embedding Hamiltonian symmetries directly into the variational parameterization geometrically regularizes this learning problem. For Boltzmann-family NQS, we enforce symmetries by tying local Pauli-$Z$ generators along physical geometric orbits, analytically collapsing the trainable coefficient space prior to optimization. To quantify the resulting optimization geometry, we introduce a geometric metric built on the Jacobian and Hessian of the optimization landscape. This framework evaluates the fraction of the physically accessible state space that corresponds to high-quality, low-energy solutions. Evaluating our approach on transverse-field Ising (TFIM) and XXZ spin chains shows that symmetry compilation excises the vast majority of parameters while maintaining ground-state accuracy within the resolution of the reported benchmarks. In large TFIM systems, strong spatial constraints compress thousands of parameters down to tens, delivering substantial runtime accelerations. Our geometric diagnostics indicate that symmetry produces a more favorable target-aware geometry by concentrating the reachable state space around low-energy solutions while retaining broad target basins. Together, our results indicate that symmetry compilation concentrates the expressive power of NQS on states relevant to the target problem, thereby reducing model size and training cost without sacrificing accuracy.

Figures

Figures reproduced from arXiv: 2608.08798 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Critical-field correlation probe for the TFIM at [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. Fidelity and correlation diagnostics across the XXZ anisotropy sweep. (a) Final infidelity [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. Correlation diagnostics near the XXZ isotropic point ( [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12 [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: FIG. 13 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: FIG. 14 [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: FIG. 15 [PITH_FULL_IMAGE:figures/full_fig_p028_15.png]
Figure 16
Figure 16. Figure 16: FIG. 16 [PITH_FULL_IMAGE:figures/full_fig_p029_16.png]
Figure 17
Figure 17. Figure 17: FIG. 17 [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 45 canonical work pages

  1. [1]

    Carleo and M

    G. Carleo and M. Troyer, Science355, 602 (2017)

    2017

  2. [2]

    D.-L. Deng, X. Li, and S. Das Sarma, Phys. Rev. X7, 021021 (2017)

    2017

  3. [3]

    Lange, A

    H. Lange, A. V. de Walle, A. Abedinnia, and A. Bohrdt, From architectures to applications: A review of neural quantum states (2024), arXiv:2402.09402 [cond-mat.dis- nn]

    arXiv 2024

  4. [4]

    D. R. Vivas, J. Madroñero, V. Bucheli, L. O. Gómez, and J.H.Reina,Neural-networkquantumstates: Asystematic review (2022), arXiv:2204.12966 [quant-ph]. 17

    arXiv 2022

  5. [5]

    Medvidović and J

    M. Medvidović and J. R. Moreno, arXiv preprint arXiv:2402.11014 (2024)

    arXiv 2024

  6. [6]

    Sharir, A

    O. Sharir, A. Shashua, and G. Carleo, Physical Review B 106, 205136 (2022)

    work page 2022

  7. [7]

    K. Choo, A. Mezzacapo, and G. Carleo, Nature commu- nications11, 1 (2020)

    work page 2020

  8. [8]

    Hibat-Allah, M

    M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko, and J. Carrasquilla, Physical Review Research2, 023358 (2020)

    work page 2020

Show all 76 references
  1. [9]

    T. D. Barrett, A. Malyshev, and A. Lvovsky, Nature Machine Intelligence4, 351 (2022)

    2022

  2. [10]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature communications9, 4812 (2018)

    2018

  3. [11]

    Cerezo, A

    M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Nature communications12, 1791 (2021)

    2021

  4. [12]

    Holmes, K

    Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, PRX quantum3, 010313 (2022)

    2022

  5. [13]

    Larocca, P

    M. Larocca, P. Czarnik, K. Sharma, G. Muraleedharan, P. J. Coles, and M. Cerezo, Quantum6, 824 (2022)

    2022

  6. [14]

    Larocca, S

    M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Bia- monte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, and M. Cerezo, arXiv preprint arXiv:2405.00781 (2024)

    2024 arXiv

  7. [15]

    Ragone, B

    M. Ragone, B. N. Bakalov, F. Sauvage, A. F. Kemper, C. Ortiz Marrero, M. Larocca, and M. Cerezo, Nature Communications15, 7172 (2024)

    2024

  8. [16]

    Park and M

    C.-Y. Park and M. J. Kastoryano, Phys. Rev. Research 2, 023232 (2020)

    2020

  9. [17]

    Sehayek, A

    D. Sehayek, A. Golubeva, M. S. Albergo, B. Kulchytskyy, G. Torlai, and R. G. Melko, Phys. Rev. B100, 195125 (2019)

    2019

  10. [18]

    B. T. Gard, L. Zhu, G. S. Barron, N. J. Mayhall, S. E. Economou, and E. Barnes, npj Quantum Information6, 1 (2020)

    2020

  11. [19]

    Bravyi, J

    S. Bravyi, J. M. Gambetta, A. Mezzacapo, and K. Temme, arXiv preprint arXiv:1701.08213 (2017)

    2017 arXiv

  12. [20]

    Zhang, N

    F. Zhang, N. Gomes, N. F. Berthusen, P. P. Orth, C.-Z. Wang, K.-M. Ho, and Y.-X. Yao, Phys. Rev. Research3, 013039 (2021)

    2021

  13. [21]

    Sauvage, M

    F. Sauvage, M. Larocca, P. J. Coles, and M. Cerezo, Quantum Science and Technology9, 015029 (2024)

    2024

  14. [22]

    Wiersema, A

    R. Wiersema, A. F. Kemper, B. N. Bakalov, and N. Kil- loran, Physical Review Research7, 013148 (2025)

    2025

  15. [23]

    Wiersema, E

    R. Wiersema, E. Kökcü, A. F. Kemper, and B. N. Bakalov, npj Quantum Information10, 110 (2024)

    2024

  16. [24]

    K. Choo, G. Carleo, N. Regnault, and T. Neupert, Physi- cal Review Letters121, 167204 (2018)

    2018

  17. [25]

    Salakhutdinov and G

    R. Salakhutdinov and G. Hinton, inArtificial intelligence and statistics(PMLR, 2009) pp. 448–455

    2009

  18. [26]

    Salakhutdinov and H

    R. Salakhutdinov and H. Larochelle, inProceedings of the thirteenth international conference on artificial intel- ligence and statistics(JMLR Workshop and Conference Proceedings, 2010) pp. 693–700

    2010

  19. [27]

    R. G. Melko, G. Carleo, J. Carrasquilla, and J. I. Cirac, Nature Physics15, 887 (2019)

    2019

  20. [28]

    Fischer and C

    A. Fischer and C. Igel, Pattern Recognition47, 25 (2014)

    2014

  21. [29]

    M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, Physical Review X8, 021050 (2018)

    2018

  22. [30]

    Nomura, A

    Y. Nomura, A. S. Darmawan, Y. Yamaji, and M. Imada, Phys. Rev. B96, 205152 (2017)

    2017

  23. [31]

    Vieijra, C

    T. Vieijra, C. Casert, J. Nys, W. De Neve, J. Haegeman, J. Ryckebusch, and F. Verstraete, Phys. Rev. Lett.124, 097201 (2020)

    2020

  24. [32]

    S. Lu, X. Gao, and L.-M. Duan, Phys. Rev. B99, 155136 (2019)

    2019

  25. [33]

    Golubeva and R

    A. Golubeva and R. G. Melko, Phys. Rev. B105, 125124 (2022)

    2022

  26. [34]

    Demidik, C

    M. Demidik, C. Tüysüz, N. Piatkowski, M. Grossi, and K. Jansen, Expressive equivalence of classical and quantum restricted boltzmann machines (2025), arXiv:2502.17562 [quant-ph]

    2025

  27. [35]

    Nakaji and N

    K. Nakaji and N. Yamamoto, Quantum5, 434 (2021)

    2021

  28. [36]

    Jacot, F

    A. Jacot, F. Gabriel, and C. Hongler, Advances in neural information processing systems31(2018)

    2018

  29. [37]

    J. Lee, L. Xiao, S. Schoenholz, Y. Bahri, R. Novak, J. Sohl- Dickstein, and J. Pennington, Advances in neural infor- mation processing systems32(2019)

    2019

  30. [38]

    Novak, L

    R. Novak, L. Xiao, J. Hron, J. Lee, A. A. Alemi, J. Sohl-Dickstein, and S. S. Schoenholz, arXiv preprint arXiv:1912.02803 (2019)

    2019 arXiv

  31. [39]

    Koczor and S

    B. Koczor and S. C. Benjamin, arXiv preprint arXiv:1912.08660 (2019)

    2019 arXiv

  32. [40]

    T. Haug, K. Bharti, and M. Kim, PRX Quantum2, 040309 (2021)

    2021

  33. [41]

    J. L. Beckey, M. Cerezo, A. Sone, and P. J. Coles, Physical Review Research4, 013083 (2022)

    2022

  34. [42]

    J. J. Meyer, Quantum5, 539 (2021)

    2021

  35. [43]

    Abbas, D

    A. Abbas, D. Sutter, A. Figalli, and S. Woerner, arXiv preprint arXiv:2112.04807 (2021)

    2021 arXiv

  36. [44]

    Abbas, D

    A. Abbas, D. Sutter, C. Zoufal, A. Lucchi, A. Figalli, and S. Woerner, Nature Computational Science1, 403 (2021)

    2021

  37. [45]

    Koffel, M

    T. Koffel, M. Lewenstein, and L. Tagliacozzo, Physical review letters109, 267203 (2012)

    2012

  38. [46]

    Vitagliano, A

    G. Vitagliano, A. Riera, and J. I. Latorre, New Journal of Physics12, 113049 (2010)

    2010

  39. [47]

    Scheie, N

    A. Scheie, N. Sherman, M. Dupont, S. Nagler, M. Stone, G. Granroth, J. Moore, and D. Tennant, Nature Physics 17, 726 (2021)

    2021

  40. [48]

    Rehet al., Phys

    M. Rehet al., Phys. Rev. B107, 195115 (2023)

    2023

  41. [49]

    O’Donnell,Analysis of Boolean Functions(Cambridge University Press, 2014)

    R. O’Donnell,Analysis of Boolean Functions(Cambridge University Press, 2014)

    2014

  42. [50]

    Sajjan, V

    M. Sajjan, V. Singh, and S. Kais, arXiv preprint arXiv:2412.12398 (2024)

    2024 arXiv

  43. [51]

    S. Dash, L. Gravina, F. Vicentini, M. Ferrero, and A. Georges, Communications Physics8, 92 (2025)

    2025

  44. [52]

    Burnside,Theory of Groups of Finite Order, Cam- bridge Library Collection - Mathematics (Cambridge Uni- versity Press, 1911)

    W. Burnside,Theory of Groups of Finite Order, Cam- bridge Library Collection - Mathematics (Cambridge Uni- versity Press, 1911)

    1911

  45. [53]

    J. H. Redfield, American Journal of Mathematics49, 433 (1927)

    1927

  46. [54]

    Pólya, Acta Mathematica68, 145 (1937)

    G. Pólya, Acta Mathematica68, 145 (1937)

    1937

  47. [55]

    Harary and E

    F. Harary and E. Palmer,Graphical Enumeration(Aca- demic Press, 1973)

    1973

  48. [56]

    J. P. Provost and G. Vallee, Communications in Mathe- matical Physics76, 289 (1980)

    1980

  49. [57]

    Stokes, J

    J. Stokes, J. Izaac, N. Killoran, and G. Carleo, Quantum 4, 269 (2020)

    2020

  50. [58]

    Larocca, N

    M. Larocca, N. Ju, D. García-Martín, P. J. Coles, and M. Cerezo, Nature Computational Science3, 542–551 (2023)

    2023

  51. [59]

    S. L. Braunstein and C. M. Caves, Phys. Rev. Lett.72, 3439 (1994)

    1994

  52. [60]

    Bradbury, R

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

    2018

  53. [61]

    Carleo, K

    G. Carleo, K. Choo, D. Hofmann, J. E. Smith, T. West- erhout, F. Alet, E. J. Davis, S. Efthymiou, I. Glasser, 18 S.-H. Lin, M. Mauri, G. Mazzola, C. B. Pereira, and F. Vicentini, SoftwareX10, 100311 (2019)

    2019

  54. [62]

    Vicentini, D

    F. Vicentini, D. Hofmann, A. Szabó, D. Wu, C. Roth, C. Giuliani, G. Pescia, J. Nys, V. Vargas-Calderón, N. As- trakhantsev, and G. Carleo, SciPost Phys. Codebases , 7 (2022)

    2022

  55. [63]

    Marshall, Proceedings of the Royal Society of Lon- don

    W. Marshall, Proceedings of the Royal Society of Lon- don. A. Mathematical and Physical Sciences232, 48 (1955), https://royalsocietypublishing.org/rspa/article- pdf/232/1188/48/49224/rspa.1955.0200.pdf

    1955

  56. [64]

    Sorella, Physical Review B71, 10.1103/phys- revb.71.241103 (2005)

    S. Sorella, Physical Review B71, 10.1103/phys- revb.71.241103 (2005)

    2005 doi

  57. [65]

    Metropolis, A

    N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, The Journal of Chemical Physics 21, 1087 (1953)

    1953

  58. [66]

    Pfeuty, Annals of Physics57, 79 (1970)

    P. Pfeuty, Annals of Physics57, 79 (1970)

    1970

  59. [67]

    Hutchinson, Communications in Statistics - Simulation and Computation19, 433 (1990), https://doi.org/10.1080/03610919008812866

    M. Hutchinson, Communications in Statistics - Simulation and Computation19, 433 (1990), https://doi.org/10.1080/03610919008812866

    1990 doi

  60. [68]

    Ubaru, J

    S. Ubaru, J. Chen, and Y. Saad, SIAM Journal on Matrix Analysis and Applications38, 1075 (2017)

    2017

  61. [69]

    C. N. Yang and C. P. Yang, Phys. Rev.150, 321 (1966)

    1966

  62. [70]

    C. N. Yang and C. P. Yang, Phys. Rev.151, 258 (1966)

    1966

  63. [71]

    Roy and M

    O. Roy and M. Vetterli, in2007 15th European Signal Processing Conference(2007) pp. 606–610

    2007

  64. [72]

    I. C. F. Ipsen and A. K. Saibaba, Stable rank and in- trinsic dimension of real and complex matrices (2024), arXiv:2407.21594 [math.NA]. Appendix A: Proofs of Theorems in Section III Proof of Lemma 1.The unconstrained amplitude is ψ(⃗ z; ⃗ c,⃗d) = Z(⃗ c)−1/2e− β 2 h(⃗ c,⃗ z)e−...

    2024 arXiv

  65. [73]

    At a stationary physical minimum, the local good volume depends on both the positive-curvature rankr+ and the physical pseudodeterminantdet+ K∗

    Decomposition of the local positive-curvature geometry The first diagnostic resolves the two spectral contributions entering Theorem 1. At a stationary physical minimum, the local good volume depends on both the positive-curvature rankr+ and the physical pseudodeterminantdet+ ...

  66. [74]

    The pullback Fubini–Study metricS = Re[J †Π⊥J]defined in Eq

    Compression of the physical tangent space The next diagnostic tests whether the parameter reduction survives projection onto physically distinguishable state variations. The pullback Fubini–Study metricS = Re[J †Π⊥J]defined in Eq. (20) vanishes along parameter directions that ...

  67. [75]

    We therefore examine the geometric estimators at multiple points along the optimization trajectory

    Stability of the geometric diagnostics during optimization The construction in Theorem 1 is local and assumes a stationary physical minimum with no resolved negative- curvature modes. We therefore examine the geometric estimators at multiple points along the optimization traje...

  68. [76]

    We therefore compare several lower-cost estimators of the physical dimension and sampled local good fraction

    Scalable proxy and estimator diagnostics The exact construction, storage, and spectral decomposition of the Fubini–Study metric and curvature operators become increasingly expensive as the number of variational coordinates grows, motivating scalable stochastic or reduced spect...

Pith tools

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