Improved Ground State Estimation in Quantum Field Theories via Normalising Flow-Assisted Neural Quantum States

Michael Spannowsky; Timur Sypchenko; Vishal S. Ngairangbam

REVIEW 4 major objections 6 minor 67 references

Improved Ground State Estimation in Quantum Field Theories via Normalising Flow-Assisted Neural Quantum States

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Flow-assisted neural sampling reaches spin ground states rivals miss.

desk verdict The decoupled normalizing-flow sampler is a genuine new ingredient for NQS, and the long-range results are good, but the abstract's 'comparable to MPS' claim is contradicted by the paper's own short-range Table I, so the scaling story is not yet supported. read the letter →

arxiv 2506.12128 v1 pith:JLPKFKW4 submitted 2025-06-13 quant-ph cs.LGhep-lathep-ph

Vishal S. Ngairangbam , Michael Spannowsky , Timur Sypchenko This is my paper

classification quant-phcs.LGhep-lathep-ph

keywords neuralquantumstatesnormalisingflowssubspacesamplingvariationalMonteCarlotransverse-fieldIsingmodelgroundstateenergyestimationlong-rangeentanglementautoregressive

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 proposes a hybrid variational method that splits the ground-state problem for quantum many-body systems into two tasks: a normalising flow learns which basis states matter, and an independent neural network learns the amplitudes on that set. The central claim is that decoupling the sampler from the variational ansatz, and discretising the flow's continuous output into orthants of a hypercube, lets the method explore strongly entangled, non-local configuration subspaces that sequential or local samplers cannot reach. On the transverse-field Ising model with up to 50 spins, the method reports ground-state energies comparable to or better than matrix-product states, and it consistently avoids the collapse into trivial all-aligned configurations that the paper observes for autoregressive models. A sympathetic reader would take away a concrete recipe: learn the support of the wavefunction with a generative model, then learn the amplitudes with a separate network.

What carries the argument

The central machinery is a decoupled-subspace sampler: a normalising flow $f_\phi: \mathbb{R}^N \to [-1,1]^N$ built from affine coupling layers, whose output space is partitioned into orthants $R_x$, one per spin configuration $x$, via the sign function (or softmax and argmax for $d$-level sites). The discrete probability $p_\phi(x)$ assigned to a configuration is the integral of the continuous posterior density over $R_x$, estimated by Monte Carlo. An independent neural quantum state $\psi_\theta$ evaluates the variational energy on the sampled subspace and supplies the target probabilities $p_\theta(x) = |\psi_\theta(x)|^2 / \sum_{x'} |\psi_\theta(x')|^2$; the flow's loss is a cross-entropy between these targets and $p_\phi$, weighted by $|E[\psi_\theta]|$, while $\theta$ is updated by the energy itself. This lets the flow discover the effective subspace $\mathcal{H}_{\rm eff}$ while the neural network learns amplitudes on it.

What would settle it

Take the short-range $N=30$ Ising chain, where every method struggles, and compute a high-precision reference energy with a tensor-network method; if the flow-assisted energy error stays large even after increasing the sampled subspace size to, say, $10^5$ configurations, the effective-subspace assumption fails and the claimed advantage collapses. Equivalently, for a smaller exactly diagonalisable system, check whether the true ground state has significant weight outside the 5000-sample subspace the flow discovers.

Watch

Extended reading notes

Core claim

The discovery is that the bottleneck in neural quantum state simulation is not representational capacity but sampling, and that a continuous normalising flow can be aimed at a discrete Hilbert-space subspace by partitioning its output space into fixed regions. For spin-$1/2$ systems, each region is one orthant of $[-1,1]^N$, so the discrete configuration $x$ is obtained by the sign of each coordinate. The flow is trained by a loss that combines a cross-entropy term between the discrete flow probabilities and the variational probabilities, weighted by the magnitude of the variational energy, while the neural quantum state parameters are updated directly by the energy expectation. The flow's discretised posterior does not need to match the true ground-state probability exactly; it only needs to put mass on configurations in the effective support. In practice, the method keeps percentage energy errors below $1\%$ for exactly solvable systems of 10 to 20 spins across short-, intermediate-, and long-range interactions, and remains accurate at $N=30,40,50$ in the fully connected regime where exact diagonalisation is infeasible and autoregressive baselines converge to trivial spin-aligned states.

Load-bearing premise

The method assumes the ground state's probability mass is concentrated on a small effective subspace, so a bounded set of flow samples can discover the dominant configurations and the network can represent the state on that set; this is stated via Equation (2) and is not verified for $N \ge 30$.

Editorial extensions

If this is right

For spin chains with all-to-all interactions and strong coupling, the method should continue to find low-energy states where MCMC and autoregressive samplers converge to trivial spin-aligned states.
The discretisation via sign or softmax should generalise to $d$-level local sites with the latent dimension growing only linearly in system size.
Once the flow has converged, it can be reused to train a fresh amplitude network without further flow training, lowering the cost of repeating the optimisation.
Reported energies for $N = 30, 40, 50$ in the fully connected regime are competitive with or below those of fixed-bond-dimension matrix-product states, indicating the approach does not lose accuracy as entanglement grows.
The split between support discovery and amplitude learning implies that the final energy error is controlled mainly by whether the flow has covered the effective subspace, not by the local dynamics of a sampler.

Reading between the lines

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

The paper does not verify the effective-subspace assumption for $N \ge 30$; if the required sampled subspace size $|S|$ grows exponentially with $N$, the method would inherit the Hilbert-space curse it is designed to avoid.
The short-range ($L=1$) results are noticeably worse than the long-range ones, suggesting that in weakly entangled regimes the flow can miss configurations that matter; this is an inference from the reported tables, not a claim the paper makes.
The same decoupled-sampling scheme could be tested on fermionic or gauge-symmetric systems by choosing a discretisation that respects the constraints, since the flow only proposes configurations and does not impose any ordering on them.
A sharper test of the comparison with matrix-product states would be to raise the bond dimension of the matrix-product baseline before deciding which method is more accurate at large $N$.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The core idea is worth taking seriously. Decoupling the sampler from the amplitude network, mapping continuous flow outputs to orthants, and weighting the flow loss by the energy are all clean design choices. The small-system results (N ≤ 20) are checked against exact diagonalization and mostly show sub-percent errors, so the method does work in that regime. The long-range, large-N tables also look respectable, with energies close to MPS for L = N/2 at V = 0.5 and 1.0. Credit where due: the method is clearly described, the comparison against autoregressive models is fair, and the fact that those models collapse to trivial modes is an honest observation.

But the abstract is not supported by the data. The short-range rows of Table I are a mess. At N = 30, L = 1, NF+NQS gives -25.7 versus MPS -38.2, a relative error around 33%. At N = 50 it is -38.9 versus -63.6, roughly 39% off. The paper says 'comparable ground state energy errors' and 'consistently achieved lower or comparable' energies, but those rows say otherwise. The authors don't discuss this discrepancy at all. That is not a minor omission; it is the central claim failing in a whole interaction regime.

The deeper issue is the subspace circularity. For N ≥ 30 the flow samples a set S, the energy is evaluated on that set, and the flow is trained on that same projected energy. If S misses important configurations, the energy is a biased upper bound and the training objective is feeding that bias back into the sampler. The assumption that such a small H_eff exists is stated but never verified for large N. The reader’s stress-test note correctly identifies this. I would not go so far as to say the method is invalid—the ED-validated small-N results show it can work—but the large-N accuracy claims rest on an unverified assumption.

Minor issues: the title promises quantum field theories but the paper only treats Ising spin chains; the MCMC baseline is missing, which is a real gap because the whole pitch is that NF sampling beats MCMC; and the footnote about the 5-sigma truncation is fine but the NMC = 25 smoothing step deserves a sensitivity check.

This paper deserves a serious referee. The method is new, the small-system evidence is solid, and the long-range results are promising. But the abstract must be rewritten, an NQS+MCMC baseline should be added, and the authors should explicitly discuss the subspace bias for large N. I would not cite this in its current form, but I would watch the revision.

Referee Report

4 major / 6 minor

Summary. The paper proposes a hybrid variational method that combines a normalising-flow sampler with an independently parameterised neural quantum state (NQS) for estimating ground-state energies of spin systems. A continuous flow maps a latent prior to [−1,1]^N, and the sign of each coordinate discretises the output into orthants representing spin configurations. The NQS assigns amplitudes to the configurations in a sampled subspace, and the projected k×k Hamiltonian is diagonalised to obtain the energy. The flow is trained with the energy-weighted cross-entropy loss in Eq. (12), where the target distribution is derived from the NQS probabilities on the current subspace. The method is tested on the transverse-field Ising model for N=10–50, interaction lengths L=1, ⌈N/4⌉, ⌈N/2⌉, and coupling strengths V=0.1, 0.5, 1.0, with comparisons to MPS and two autoregressive models.

Significance. If the central claims were fully supported, the decoupling of support discovery from amplitude learning would be a useful new tool for neural quantum states in long-range and volume-law-entangled regimes. The paper contains genuine strengths: exact-diagonalisation checks for N≤20 show low errors, the long-range results for N=30–50 are close to the MPS reference, and the autoregressive baselines collapse to trivial configurations while the proposed method does not. However, the broad claim of comparability with MPS is contradicted by the short-range rows of Table I, and the large-system energies are self-referential because the flow is trained on the NQS's own subspace-conditioned distribution with no exact reference. The significance therefore rests on a narrowed claim: the method appears promising for long-range interactions, but the scalable-accuracy claim for the method as stated is not established.

major comments (4)

[Table I, Abstract, Sec. V] The abstract and Sec. V claim that the method achieves 'comparable ground state energy errors with state-of-the-art matrix product states' and 'consistently achieved lower or comparable ground state energy errors.' Table I directly contradicts this for short-range interactions: for L=1 at N=30, 40, 50, the NF+NQS energies are -25.7±0.1, -36.22±0.09, and -38.89±0.08, whereas MPS gives -38.184±0.003, -50.914±0.003, and -63.643±0.004. These are relative errors of roughly 33%, 29%, and 39% relative to the MPS reference. The manuscript does not discuss this large discrepancy, and the claim of comparability with MPS must be either removed from the abstract and conclusions or restricted to the long-range regimes where the data are supportive.
[Eq. (12), Sec. IV.B, Sec. IV.C] The training objective in Eq. (12) uses pθ(x)=|ψθ(x)|^2 / Σ_{x'} |ψθ(x')|^2 and the projected energy E[ψθ] evaluated only on the currently sampled subspace S. The flow is therefore trained to reinforce the NQS's own subspace-conditioned distribution, and the energy used to drive the flow is an upper bound on the true ground-state energy that can be loose if S omits important configurations. For N≥30 there is no exact-diagonalisation reference, so the converged subspace can be self-consistent yet incomplete. The L=1 rows of Table I are concrete evidence of this failure mode. The authors should provide an independent check of subspace completeness for N≥30 (for example, convergence of the energy with |S|, or overlap with a high-quality MPS/DMRG reference) before claiming scalable accuracy.
[Sec. II.A, Eq. (2)] The method's core assumption is that the ground state has significant support only on a small subspace H_eff with dim H_eff ≪ dim H, as stated in Eq. (2). This assumption is not verified for N≥30. The poor short-range results suggest that either the assumption is violated for L=1 or the flow cannot discover the full support. Since the projected Hamiltonian is only a variational upper bound, a missing configuration can bias the energy upward substantially. The paper should either verify Eq. (2) for the large-system regimes studied or explicitly characterise the regimes in which the assumption is expected to hold.
[Sec. IV.B, Tables I and II] The comparison with MPS is not apples-to-apples. The MPS calculations use a MetropolisLocal sampler over the full Hilbert space and a variational energy estimator, while NF+NQS diagonalises a projected k×k Hamiltonian on the sampled subspace. The manuscript acknowledges that the projected estimator uses less information per iteration, but the conclusions do not account for the upward bias this introduces. In addition, Tables I and II show that even in long-range cases the NF+NQS energies are sometimes above the MPS energies by more than the quoted standard deviations (for example, N=50, L=13: -650.3±0.2 vs -650.956±0.004; and N=50, V=0.1: -126.15±0.04 vs -127.606±0.002). The claims should be recalibrated to the actual comparison, and the MPS bond dimension of five should be checked for convergence before using MPS as a reference.

minor comments (6)

[Title and Abstract] The title and abstract refer to 'Quantum Field Theories', but the numerical study is a transverse-field Ising spin chain. This overstates the scope; consider rephrasing to 'quantum many-body systems' or 'lattice spin systems' unless a QFT application is actually demonstrated.
[Sec. II.A, Eq. (2)] The parameter ε is introduced in Eq. (2) but never used quantitatively. Either specify how the support is defined or remove the threshold from the definition.
[Sec. IV.A] The text says 'we first consider exactly solvable systems with N∈{10,15,20}'. Exact diagonalisation is feasible for these sizes, but the long-range TFIM is not exactly solvable in the usual sense; the phrase 'exactly solvable' should be replaced by 'exactly diagonalisable' or similar.
[Sec. IV.B] The text refers to 'the k×k Hamiltonian' without defining k. It should state explicitly that k=|S|, the number of unique configurations in the sampled subspace.
[Footnote 1] The footnote says 'Due to the small standard deviation of 0.1', but the prior is described earlier as a mixture of Gaussians with standard deviation 0.33. Clarify which distribution has standard deviation 0.1 and how this relates to the smoothing procedure.
[Table II] The table heading contains a typo: 'interaction strenths' should be 'interaction strengths'. Additionally, the table captions should define ⌈N/2⌉ consistently with the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the energy estimate is a standard variational objective, externally benchmarked for N<=20; large-N values are unvalidated but not self-referential by construction.

full rationale

The derivation chain is a joint variational optimization. The flow loss in Eq. (12) depends on E[psi_theta] and p_theta(x) = |psi_theta(x)|^2 / sum |psi_theta(x')|^2, and the NQS loss is E[psi_theta] itself; the reported ground-state energies are the minimized value of this same objective. This is the standard definition of a variational method, not a circular reduction: the target quantity (the true ground-state energy of the Hamiltonian in Eq. (14)) is independent of the algorithm. For N=10,15,20 the paper explicitly anchors to exact diagonalization via Eq. (17) and reports percentage errors in Figs. 2-3, providing external falsifiability. The paper's self-citations (Refs. [11]-[14]) appear as background examples of many-body applications and carry no load-bearing argument; no uniqueness theorem or ansatz is imported from the authors' prior work, and the normalizing-flow universality invoked is an external result [56]. The limitation to flag is not circularity: in Sec. IV.C the authors state that for N in {30,40,50}, 'ED became infeasible, and we report energies obtained directly from each method.' Those large-N values are the unverified outcome of the variational loop, and the convergence criterion ('unique samples drop below 90% of |S|') can shrink the sampled subspace; that biases the upper-bound estimate and weakens the abstract's 'comparable ground state energy errors' claim, but it does not make the result equivalent to its inputs by definition.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption of a small effective support, on the universality of flows, and on the expressivity of the NQS. The method introduces several hand-chosen hyperparameters (subspace size, MC samples, convergence threshold, energy weighting) that affect the results but are not derived from first principles. No new physical entities are introduced.

free parameters (5)

Prior mixture standard deviation = 0.33
Chosen so the initial discretised posterior is approximately uniform over orthants; affects flow initialisation and training.
Monte Carlo samples per region N_MC = 25
Used to estimate p_phi(x) in Eq. (10); controls bias and variance of flow probability estimates.
Subspace size |S| = 5000 (150 for N=10)
Number of unique basis states per batch; determines the truncated Hilbert space for energy evaluation.
Convergence threshold = 90% unique samples
Training stops when unique samples fall below 90% of |S|, controlling the final subspace size.
Energy weighting factor |E| in NF loss = |E[psi_theta]|
Multiplicative factor in Eq. (12) that prioritises directions with higher energy magnitude; a heuristic choice.

assumptions (3)

domain assumption The ground state has significant support only on a small effective subspace Heff (Eq. 2).
Section II.A defines supp(psi0) with epsilon and dim Heff << dim H; the method relies on this to justify sampling a restricted subspace. It fails when the support is broad, e.g., weak-coupling paramagnetic states.
domain assumption Normalising flows are universal enough to learn the support of the ground state (citing Draxler et al. [56]).
Section III asserts distributional universality of flows; if the flow cannot concentrate mass on all important regions, the subspace will be incomplete.
domain assumption The NQS ansatz (a feedforward DNN) can express the ground state amplitudes on the discovered subspace.
Section IV.B describes the NQS as a DNN; expressivity is assumed, not proven for these systems.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Improved Ground State Estimation in Quantum Field Theories via Normalising Flow-Assisted Neural Quantum States." pith.science (2026). https://pith.science/paper/JLPKFKW4

@misc{pith2026250612128,
  author       = {Pith},
  title        = {Pith review of: Improved Ground State Estimation in Quantum Field Theories via Normalising Flow-Assisted Neural Quantum States},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JLPKFKW4}},
  note         = {Machine review of arXiv:2506.12128}
}

read the original abstract

We propose a hybrid variational framework that enhances Neural Quantum States (NQS) with a Normalising Flow-based sampler to improve the expressivity and trainability of quantum many-body wavefunctions. Our approach decouples the sampling task from the variational ansatz by learning a continuous flow model that targets a discretised, amplitude-supported subspace of the Hilbert space. This overcomes limitations of Markov Chain Monte Carlo (MCMC) and autoregressive methods, especially in regimes with long-range correlations and volume-law entanglement. Applied to the transverse-field Ising model with both short- and long-range interactions, our method achieves comparable ground state energy errors with state-of-the-art matrix product states and lower energies than autoregressive NQS. For systems up to 50 spins, we demonstrate high accuracy and robust convergence across a wide range of coupling strengths, including regimes where competing methods fail. Our results showcase the utility of flow-assisted sampling as a scalable tool for quantum simulation and offer a new approach toward learning expressive quantum states in high-dimensional Hilbert spaces.

Figures

Figures reproduced from arXiv: 2506.12128 by the authors.

**Figure 1.** FIG. 1. Schematic of the decoupled-subspace variational ansatz: a Normalising Flow (left) learns the support, while an [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** FIG. 2. Percentage error for different interaction length ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 52 canonical work pages

[1]

Sampling from the prior distribution pZ(z), we get discretized posteriors x via the continuous posterior pY (y)

work page
[2]

the discrete samples S = {x} is a set

Since it is not guaranteed that the batch contains only distinct basis states, we keep single unique val- ues of x, i.e. the discrete samples S = {x} is a set

work page
[3]

we define pθ(x) = |ψθ(x)|2/ P x′ |ψθ(x′)|2

For basis states x in S, we evaluate the dis- crete target probability of the NF model from the NQS model, i.e. we define pθ(x) = |ψθ(x)|2/ P x′ |ψθ(x′)|2

work page
[4]

In each region Rx that is contained in S, we in- dependently sample NMC values of y ∈ Rx from a multivariate normal distributions with a scalar co- variance matrix such that the mean is located at least five standard deviations away from any adja- cent regions. As long as we keep NMC not too large, it is highly unlikely for any set Mx to contain sam- ples...

work page
[5]

Thus, within such a procedure, the NQS energy expecta- tion E[ψθ] can drive the learningof the NF’s parameters

The continuous target probability of the Normalis- ing Flow’s posterior is defined as thepθ(x) weighted sum of the independent normal distributions in each region Rx. Thus, within such a procedure, the NQS energy expecta- tion E[ψθ] can drive the learningof the NF’s parameters. We define the loss function for the NF parameters as Lϕ[pY , S] = − |E[ψθ]| |S...

work page
[6]

J. A. Hertz, Phys. Rev. B 14, 1165 (1976)

1976
[7]

Amico, R

L. Amico, R. Fazio, A. Osterloh, and V. Vedral, Rev. Mod. Phys. 80, 517 (2008)

2008
[8]

LeBlond, K

T. LeBlond, K. Mallayya, L. Vidmar, and M. Rigol, Phys. Rev. E 100, 062134 (2019)

work page 2019

Show all 67 references

[9]

Hayden and J

P. Hayden and J. Preskill, JHEP 09, 120 (2007), arXiv:0708.4025 [hep-th]

2007 arXiv
[10]

Bhattacharya, P

A. Bhattacharya, P. P. Nath, and H. Sahu, Phys. Rev. D 109, 066010 (2024)

2024
[11]

Schiulaz, E

M. Schiulaz, E. J. Torres-Herrera, F. P´ erez-Bernal, and L. F. Santos, Phys. Rev. B 101, 174312 (2020)

2020
[12]

Belyansky, P

R. Belyansky, P. Bienias, Y. A. Kharkov, A. V. Gorshkov, and B. Swingle, Phys. Rev. Lett. 125, 130601 (2020)

2020
[13]

Lashkari, D

N. Lashkari, D. Stanford, M. Hastings, T. Osborne, and P. Hayden, Journal of High Energy Physics 2013, 22 (2013)

2013
[14]

Swingle, G

B. Swingle, G. Bentsen, M. Schleier-Smith, and P. Hay- den, Phys. Rev. A 94, 040302 (2016)

2016
[15]

Yin and A

C. Yin and A. Lucas, Phys. Rev. A 102, 022402 (2020)

2020
[16]

J. C. Criado, S. Schenk, M. Spannowsky, P. D. Hat- ton, and L. A. Turnbull, Sci. Rep. 12, 19179 (2022), arXiv:2109.15020 [cond-mat.str-el]

2022 arXiv
[17]

J. C. Criado, P. D. Hatton, A. Lanza, S. Schenk, and M. Spannowsky, Phys. Rev. B 109, 195114 (2024), arXiv:2303.04099 [cond-mat.str-el]

2024 arXiv
[18]

Schenk and M

S. Schenk and M. Spannowsky, Phys. Rev. B103, 144436 (2021), arXiv:2012.10325 [hep-th]

2021 arXiv
[19]

J. Y. Araz, S. Schenk, and M. Spannowsky, Phys. Rev. A 107, 032619 (2023), arXiv:2210.03679 [quant-ph]

2023 arXiv
[20]

J. Lin, D. Luo, X. Yao, and P. E. Shanahan, JHEP 06, 211 (2024), arXiv:2402.06607 [hep-ph]

2024 arXiv
[21]

S. Lu, G. Giudice, and J. I. Cirac, Phys. Rev. B 111, 075102 (2025), arXiv:2401.14243 [quant-ph]

2025 arXiv
[22]

Carleo and M

G. Carleo and M. Troyer, Science 355, 602 (2017), https://www.science.org/doi/pdf/10.1126/science.aag2302

2017 doi
[23]

Passetti, D

G. Passetti, D. Hofmann, P. Neitemeier, L. Grunwald, M. A. Sentef, and D. M. Kennes, Phys. Rev. Lett. 131, 036502 (2023)

2023
[24]

J. Kim, D. Kim, and D.-H. Kim, Phys. Rev. E 109, 064123 (2024)

2024
[25]

Denis, A

Z. Denis, A. Sinibaldi, and G. Carleo, Phys. Rev. Lett. 134, 079701 (2025)

2025
[26]

Chen and M

A. Chen and M. Heyl, Nature Physics 20, 1476 (2024)

2024
[27]

Roca-Jerat, M

S. Roca-Jerat, M. Gallego, F. Luis, J. Carrete, and D. Zueco, Phys. Rev. B 110, 205147 (2024)

2024
[28]

P. M. Vecsei, C. Flindt, and J. L. Lado, Phys. Rev. Res. 5, 033116 (2023)

2023
[29]

Bukov, M

M. Bukov, M. Schmitt, and M. Dupont, SciPost Phys. 10, 147 (2021)

2021
[30]

Roca-Jerat, M

S. Roca-Jerat, M. Gallego, F. Luis, J. Carrete, and D. Zueco, Phys. Rev. B 110, 205147 (2024), arXiv:2407.04773 [quant-ph]

2024 arXiv
[31]

Romero, J

I. Romero, J. Nys, and G. Carleo, Commun. Phys. 8, 46 (2025), arXiv:2406.09077 [cond-mat.str-el]

2025 arXiv
[32]

Rende and L

R. Rende and L. L. Viteritti, Mach. Learn. Sci. Tech. 6, 010501 (2024), arXiv:2405.18874 [cond-mat.dis-nn]

2024 arXiv
[33]

Rende, S

R. Rende, S. Goldt, F. Becca, and L. L. Viteritti, Phys. Rev. Res. 6, 043280 (2024), arXiv:2403.07795 [cond- mat.dis-nn]

2024 arXiv
[34]

Joshi, R

A. Joshi, R. Peters, and T. Posske, Phys. Rev. B 110, 104411 (2024), arXiv:2403.08184 [cond-mat.dis-nn]

2024 arXiv
[35]

J. Beck, J. Bodky, J. Motruk, T. M¨ uller, R. Thomale, and P. Ghosh, Phys. Rev. B 109, 184422 (2024), arXiv:2401.04995 [cond-mat.str-el]

2024 arXiv
[36]

Duric, J

T. Duric, J. H. Chung, B. Yang, and P. Sengupta, Phys. Rev. X 15, 011047 (2025), arXiv:2401.02866 [cond- mat.str-el]

2025 arXiv
[37]

Z. Wu, R. Zen, H. P. Casagrande, D. Poletti, and S. Bres- san, Comput. Phys. Commun. 300, 109169 (2024), arXiv:2305.03394 [cond-mat.stat-mech]

2024 arXiv
[38]

Joshi, R

A. Joshi, R. Peters, and T. Posske, Phys. Rev. B 108, 094410 (2023), arXiv:2304.09504 [cond-mat.mes-hall]

2023 arXiv
[39]

Mezera, J

M. Mezera, J. Menˇ s ´ ıkov´ a, P. Bal´ aˇ z, and M.ˇZonda, Sci- Post Phys. Core 6, 088 (2023), arXiv:2303.14108 [cond- mat.dis-nn]

2023 arXiv
[40]

Metropolis, A

N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, J. Chem. Phys. 21, 1087 (1953)

1953
[41]

W. K. Hastings, Biometrika 57, 97 (1970)

1970
[42]

W. M. C. Foulkes, L. Mitas, R. J. Needs, and G. Ra- jagopal, Rev. Mod. Phys. 73, 33 (2001)

2001
[43]

Zhang, Z.-Q

S.-X. Zhang, Z.-Q. Wan, and H. Yao, Phys. Rev. Res. 5, 033041 (2023)

2023
[44]

Carleo, K

G. Carleo, K. Choo, D. Hofmann, J. E. T. Smith, T. Westerhout, F. Alet, E. J. Davis, S. Efthymiou, I. Glasser, S.-H. Lin, M. Mauri, G. Mazzola, C. B. Mendl, E. van Nieuwenburg, O. O’Reilly, H. Th´ eveniaut, G. Tor- lai, F. Vicentini, and A. Wietek, SoftwareX , 100311 (2019)

2019
[45]

Vicentini, D

F. Vicentini, D. Hofmann, A. Szab´ o, D. Wu, C. Roth, C. Giuliani, G. Pescia, J. Nys, V. Vargas-Calder´ on, N. Astrakhantsev, and G. Carleo, SciPost Phys. Code- bases , 7 (2022)

2022
[46]

Levine, O

Y. Levine, O. Sharir, N. Cohen, and A. Shashua, Phys. Rev. Lett. 122, 065301 (2019)

2019
[47]

Gao and L.-M

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

2017
[48]

T.-H. Yang, M. Soleimanifar, T. Bergamaschi, and J. Preskill, (2024), arXiv:2410.23152 [quant-ph]

2024 arXiv
[49]

Uria, M.-A

B. Uria, M.-A. Cˆ ot´ e, K. Gregor, I. Murray, and H. Larochelle, Journal of Machine Learning Research 17, 1 (2016)

2016
[50]

Sharir, Y

O. Sharir, Y. Levine, N. Wies, G. Carleo, and A. Shashua, Phys. Rev. Lett. 124, 020503 (2020)

2020
[51]

D. Luo, Z. Chen, K. Hu, Z. Zhao, V. M. Hur, and B. K. Clark, Phys. Rev. Res. 5, 013216 (2023)

2023
[52]

Humeniuk, Y

S. Humeniuk, Y. Wan, and L. Wang, SciPost Phys. 14, 171 (2023)

2023
[53]

Zhang and M

Y.-H. Zhang and M. Di Ventra, Phys. Rev. B107, 075147 (2023)

2023
[54]

D. Luo, Z. Chen, J. Carrasquilla, and B. K. Clark, Phys. Rev. Lett. 128, 090501 (2022)

2022
[55]

Bortone, Y

M. Bortone, Y. Rath, and G. H. Booth, Quantum 8, 1245 (2024), arXiv:2306.05917 [quant-ph]

2024 arXiv
[56]

Rezende and S

D. Rezende and S. Mohamed, in Proceedings of the 32nd International Conference on Machine Learning, Proceed- ings of Machine Learning Research, Vol. 37, edited by F. Bach and D. Blei (PMLR, Lille, France, 2015) pp. 1530–1538

2015
[57]

Stokes, B

J. Stokes, B. Chen, and S. Veerapaneni, Machine Learn- ing: Science and Technology 4, 021001 (2023)

2023
[58]

Lawrence, A

S. Lawrence, A. Shelby, and Y. Yamauchi, (2024), 10 arXiv:2406.02451 [quant-ph]

2024 arXiv
[59]

D. Tran, K. Vafa, K. Agrawal, L. Dinh, and B. Poole, in Advances in Neural Information Processing Systems, Vol. 32, edited by H. Wallach, H. Larochelle, A. Beygelz- imer, F. d 'Alch´ e-Buc, E. Fox, and R. Garnett (Curran Associates, Inc., 2019)

2019
[60]

Bengio, N

Y. Bengio, N. L´ eonard, and A. C. Courville, CoRR abs/1308.3432 (2013), 1308.3432

2013 arXiv
[61]

Draxler, S

F. Draxler, S. Wahl, C. Schnoerr, and U. Koethe, in Proceedings of the 41st International Conference on Ma- chine Learning, Proceedings of Machine Learning Re- search, Vol. 235, edited by R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and F. Berkenk...

2024
[62]

Pfeuty, Annals of Physics 57, 79 (1970)

P. Pfeuty, Annals of Physics 57, 79 (1970)

1970
[63]

Pytorch: an imperative style, high-performance deep learning library,

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨ opf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: an imperative style, hig...

2019
[64]

Stimper, D

V. Stimper, D. Liu, A. Campbell, V. Berenz, L. Ryll, B. Sch¨ olkopf, and J. M. Hern´ andez-Lobato, Journal of Open Source Software 8, 5361 (2023)

2023
[65]

L. Dinh, J. Sohl-Dickstein, and S. Bengio, in Interna- tional Conference on Learning Representations(2017)

2017
[66]

S. K. Lam, A. Pitrou, and S. Seibert, in Proceedings of the Second Workshop on the LL VM Compiler Infrastruc- ture in HPC(2015) pp. 1–6

2015
[67]

The DeepMind JAX Ecosystem,

DeepMind, I. Babuschkin, K. Baumli, A. Bell, S. Bhu- patiraju, J. Bruce, P. Buchlovsky, D. Budden, T. Cai, A. Clark, I. Danihelka, A. Dedieu, C. Fantacci, J. God- win, C. Jones, R. Hemsley, T. Hennigan, M. Hes- sel, S. Hou, S. Kapturowski, T. Keck, I. Kemaev, M. King, M. Kunes...

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Sampling from the prior distribution pZ(z), we get discretized posteriors x via the continuous posterior pY (y)

work page

[2] [2]

the discrete samples S = {x} is a set

Since it is not guaranteed that the batch contains only distinct basis states, we keep single unique val- ues of x, i.e. the discrete samples S = {x} is a set

work page

[3] [3]

we define pθ(x) = |ψθ(x)|2/ P x′ |ψθ(x′)|2

For basis states x in S, we evaluate the dis- crete target probability of the NF model from the NQS model, i.e. we define pθ(x) = |ψθ(x)|2/ P x′ |ψθ(x′)|2

work page

[4] [4]

In each region Rx that is contained in S, we in- dependently sample NMC values of y ∈ Rx from a multivariate normal distributions with a scalar co- variance matrix such that the mean is located at least five standard deviations away from any adja- cent regions. As long as we keep NMC not too large, it is highly unlikely for any set Mx to contain sam- ples...

work page

[5] [5]

Thus, within such a procedure, the NQS energy expecta- tion E[ψθ] can drive the learningof the NF’s parameters

The continuous target probability of the Normalis- ing Flow’s posterior is defined as thepθ(x) weighted sum of the independent normal distributions in each region Rx. Thus, within such a procedure, the NQS energy expecta- tion E[ψθ] can drive the learningof the NF’s parameters. We define the loss function for the NF parameters as Lϕ[pY , S] = − |E[ψθ]| |S...

work page

[6] [6]

J. A. Hertz, Phys. Rev. B 14, 1165 (1976)

1976

[7] [7]

Amico, R

L. Amico, R. Fazio, A. Osterloh, and V. Vedral, Rev. Mod. Phys. 80, 517 (2008)

2008

[8] [8]

LeBlond, K

T. LeBlond, K. Mallayya, L. Vidmar, and M. Rigol, Phys. Rev. E 100, 062134 (2019)

work page 2019