Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Functional Neural Wavefunction Optimization

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By viewing variational Monte Carlo optimizers as Galerkin projections of function-space algorithms, the paper derives projected inverse iteration, a method whose convergence stays fast when spectral gaps shrink.

desk verdict PII is a genuinely useful optimizer with a clean function-space story, but the paper's central claim rests on an unproved Galerkin-preservation step and on numerics that feed the target energy into the shift. read the letter →

arxiv 2507.10835 v1 pith:ILQ2MFLM submitted 2025-07-14 cond-mat.str-el cs.LGmath.OCphysics.comp-phquant-ph

Victor Armegioiu , Juan Carrasquilla , Siddhartha Mishra , Johannes Müller , Jannes Nys , Marius Zeinhofer , Hang Zhang This is my paper
classification cond-mat.str-elcs.LGmath.OCphysics.comp-phquant-ph MSC 68Q2568R1068U05
keywords variationalMonteCarloneuralquantumstatesstochasticreconfigurationprojectedinverseiterationRiemannianoptimizationGalerkinprojectionspinmodels
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 is trying to give variational quantum Monte Carlo a design principle: choose an optimization algorithm in the infinite-dimensional space of wavefunctions, then make it computable by projecting each update onto the tangent space of the neural-network ansatz. In this unified view, the standard optimizer stochastic reconfiguration is just Riemannian gradient descent on the energy sphere, and a second-order Newton variant is the classical Rayleigh quotient iteration. The paper's new algorithm, projected inverse iteration, is the projected version of shifted inverse iteration, with a shift that can encode prior knowledge of the ground-state energy. This matters because the shift lets the optimizer keep a fast linear convergence rate even when the gap to the first excited state shrinks, a regime where stochastic reconfiguration slows down.

What carries the argument

The load-bearing object is the Galerkin projection of a functional update onto the tangent space $V_\theta=\operatorname{span}\{\partial_{\theta_i}\hat\psi_\theta\}$ of the normalized neural-network ansatz. A functional algorithm is posed as a linear system $b(d,v)=f(v)$ on the sphere's tangent space; the Galerkin discretization replaces $b$ and $f$ by the overlap matrix $S(\theta)_{ij}=\langle \partial_{\theta_i}\hat\psi_\theta,\partial_{\theta_j}\hat\psi_\theta\rangle$, the Hamiltonian matrix $H(\theta)_{ij}=\langle \partial_{\theta_i}\hat\psi_\theta,\hat H\partial_{\theta_j}\hat\psi_\theta\rangle$, and the energy gradient $\nabla L(\theta)$, yielding preconditioned gradient updates $\theta_{k+1}=\theta_k-\eta Q(\theta_k)^{-1}\nabla L(\theta_k)$. Applying this to $L^2$ gradient descent gives stochastic reconfiguration; applying it to shifted Riemannian Newton (functional shifted inverse iteration) gives the PII update $\theta_{k+1}=\theta_k-\eta_k(H_k-\tau_k S_k+\varepsilon_k I)^{-1}\nabla L(\theta_k)$, with a Woodbury-based variant for $M\ll P$.

What would settle it

A deterministic full-sum run on a small transverse-field Ising model with a known narrow gap would falsify the central claim if the measured per-iteration energy-reduction factor of PII failed to match $|(E_0-\tau)/(E_1-\tau)|^2$ while SR's factor followed $\exp(-2|E_1-E_0|\eta)$, since the framework predicts both rates.

Watch

Extended reading notes

Core claim

The paper establishes that the main variational Monte Carlo optimizers are the same object seen at different levels: a function-space dynamics on the unit sphere, discretized by projecting onto the tangent space of the parameterized wavefunction. Stochastic reconfiguration is Riemannian $L^2$ gradient descent, whose linear convergence rate is set by the spectral gap $E_1-E_0$; the Riemannian Newton method is Rayleigh quotient iteration, locally cubic in the finite-dimensional case; and the globalized, shifted version is inverse iteration, whose linear rate for a shift $\tau$ is $|E_J-\tau|/\zeta$ toward the eigenstate closest to $\tau$. The new algorithm, projected inverse iteration, uses a shift $\tau=\alpha E_0$, and its per-iteration energy reduction factor $|(E_0-\tau)/(E_1-\tau)|^2$ stays favorable when $E_1-E_0$ is small, provided $\tau$ is chosen close to $E_0$. Numerically on transverse-field Ising and Heisenberg spin lattices with RBM and vision-transformer ans\"atze, PII outperforms SR and tolerates learning rates of order 1, while the theoretically best shift $\tau=E_0$ can fail under Monte Carlo noise and a perturbed shift restores stability.

Load-bearing premise

The argument rests on the assumption that projecting a functional algorithm onto the tangent space of a nonlinear neural-network ansatz keeps the algorithm's convergence rate intact; the theorems analyze the functional level, not the parameter-space update.

Editorial extensions

If this is right

  • PII gives ground-state energies of quantum spin models in far fewer iterations than SR when the spectral gap is narrow; on the tested 4x4 and 10x10 lattices it is the fastest optimizer across all hyperparameter settings.
  • PII can be used with learning rates around 0.5 to 1.0, while SR typically needs rates near 10^{-3}, because the functional counterpart of PII naturally uses unit step size.
  • Choosing the shift as $\tau=\alpha E_0$ with $\alpha\ge 1$ is the principled replacement for the ad-hoc regularization in Rayleigh-Gauss-Newton; $\tau=E_0$ is optimal but can be unstable with sampling, with $\tau=1.2E_0$ to $1.4E_0$ a robust substitute.
  • The minPII formulation via the Woodbury identity keeps the per-iteration cost linear in parameter count for $M\ll P$, making PII feasible for the large ans\"atze where SR already uses minSR.
  • The convergence-factor formulas $r_{\mathrm{SR}}=[\exp(-2|E_1-E_0|\eta)]^2$ and $r_{\mathrm{PII}}=|(E_0-\tau)/(E_1-\tau)|^2$ give a quantitative prediction for when each algorithm will win, so practitioners can decide which to use from a known or estimated gap.

Reading between the lines

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

  • Editorial extension: the same function-space-to-parameter-space pipeline should be able to turn trust-region or conjugate-gradient eigensolvers into new variational Monte Carlo optimizers with automatically selected hyperparameters, since the paper already sketches these as future directions.
  • Editorial extension: if the Galerkin projection preserves rates beyond linear ans\"atze, the shift $\tau$ could be updated on the fly from an estimator of $E_0$, removing the need to know $E_0$ in advance and turning the prior-knowledge requirement into an adaptive scheme.
  • Editorial extension: because PII's speedup is largest when $E_1-E_0$ is small, frustrated and fermionic systems, where small gaps are common, are the natural next test bed, but the paper only benchmarks stoquastic spin Hamiltonians and leaves sign-problematic cases out.
  • Editorial extension: identifying SR as $L^2$ gradient descent and PII as inverse iteration suggests analogous gap-independent convergence may be derivable for Wasserstein quantum Monte Carlo, which the paper classifies as a Galerkin discretization of Wasserstein gradient descent.
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.

Referee Report

3 major / 4 minor

Summary. The paper develops a geometric, function-space framework for optimizing neural-network variational wavefunctions. It shows that stochastic reconfiguration (SR) corresponds to Galerkin-projected L2 gradient descent on the unit sphere, that the Riemannian Newton method corresponds to Rayleigh quotient iteration, and that a diagonal-shift globalization corresponds to shifted inverse iteration. The shifted version, called projected inverse iteration (PII), is derived as a parameter-space algorithm and is claimed to converge faster and be more robust to small spectral gaps than SR. The theoretical core consists of functional-level convergence theorems (Theorem 3.1 for gradient flow, Theorem 3.8 for inverse iteration), and the numerical section compares SR, SGD, and PII on 4x4 and 10x10 spin models using restricted Boltzmann machines and vision transformers.

Significance. If the central claim is correct, PII would be a practically useful improvement over SR for neural quantum state optimization, especially in systems with small spectral gaps. The paper's functional-level analysis is a valuable contribution: it rigorously connects SR to L2 gradient flow, Rayleigh quotient iteration to Riemannian Newton, and shifted inverse iteration to a globalized Newton method, and Theorem 3.8 provides a precise convergence rate for functional inverse iteration. The Galerkin projection framework is conceptually clean and unifies several existing methods. However, the significance is tempered by the fact that the actual algorithm is a stochastic parameter-space recursion whose relation to the functional theorems is only heuristic, and the numerical evidence relies on oracle choices of the shift parameter and manually tuned hyperparameters.

major comments (3)
  1. [Section 2.1 and Section 5.1, Eq. (5.1)] The central claim that PII converges faster and is more gap-robust than SR is not established for the parameter-space algorithm. Theorems 3.1 and 3.8 analyze the functional-level gradient flow and shifted inverse iteration, while the practical method is the stochastic recursion (5.1). The paper asserts in Section 2.1 that 'the Galerkin projections ensure that key properties of the functional algorithm are preserved in the discretized form', but no proof is given for nonlinear ansatze, and Section 5.1 explicitly states that 'transferring learning rates from function space to the ansatz is not guaranteed'. The formula rPII = |(E0 - tau)/(E1 - tau)|^2 in Section 5.1 is the functional inverse-iteration rate, not a proven rate for the parameter-space update. Without a direct analysis of (5.1), the advertised speedup over SR remains conjectural.
  2. [Section 5.1 and Figures 1-5] The numerical validation relies on oracle information: the shift tau is set to E0 or a multiple of E0, where E0 is obtained from a Lanczos computation for the 4x4 system and supplied externally for the 10x10 system. The paper also notes that the theoretically optimal shift tau = E0 is frequently unstable and that tau = 1.2 E0 is needed for stabilization, which still requires knowing E0. All learning rates and diagonal shifts are manually tuned. Without a protocol that selects these hyperparameters without reference to the target ground-state energy, the reported superiority of PII over SR cannot be cleanly attributed to the algorithm rather than to the oracle input. This is a load-bearing limitation for the paper's empirical claims.
  3. [Remark 3.9 and Section 5.1] The gap-robustness claim is not supported for the shift values actually used in the experiments. The paper uses fixed multiples tau = 1.2 E0 and tau = 1.4 E0. For fixed alpha > 1, the convergence factor in Theorem 3.8 is |E0 - alpha E0| / |E1 - alpha E0|, which tends to 1 as E1 approaches E0. Hence the iteration slows down as the gap closes. The gap-independent rate described in Remark 3.9 requires tau = (1 - gamma)^{-1}(E1 - gamma E0), which depends on both E0 and E1, not merely on E0. Therefore the statement that PII is robust to narrow spectral gaps is not justified for the empirically used shift choices.
minor comments (4)
  1. [Title and abstract] There are typographical errors: 'W A VEFUNCTION' in the title and 'W a vefunction' in the running text, and 'OPTIMIZA TION' should be 'OPTIMIZATION'.
  2. [Figure 4 caption] The caption states 'sample size 36768' while the text and Figure 5 use M = 32768; this appears to be a typo and should be corrected.
  3. [General] The numerical figures do not include error bars or standard deviations over independent runs. Given the known stochasticity of VMC optimization, at least a brief statement on run-to-run variability would help the reader assess the robustness of the reported comparisons.
  4. [Section 5.1] The paragraph introducing rPII and rSR uses notation such as 'energy reduction factor' without a formal definition; the intended meaning is clear from context, but a short definition would improve precision.

Circularity Check

1 steps flagged · score 4.0 of 10

Core derivation is self-contained, but the numerical validation of PII is partly circular: the advertised speed and gap-robustness are demonstrated with the exact ground-state energy E0 fed into the shift τ, and the reported rate formula makes the convergence factor vanish by construction for τ=E0.

  1. fitted input called prediction [Abstract and Section 5.1 (Eq. (5.1), TFIM 4×4 experiments, rate formula for PII)]
    "We validate our framework with numerical experiments demonstrating its practical relevance through the accurate estimation of ground-state energies ... For PII, we test the shift values τk = E0 and τk = 1.4 × E0. ... In this example, we can compute the spectral gap exactly, resorting to a Lanczos iteration ... To be precise the energy reduction factor per iteration step for PII with unit learning rate is rPII = |(E0 − τ)/(E1 − τ)|^2."

    The numerical claim that PII estimates ground-state energies and converges much faster than SR is validated by supplying the exact target E0 as the shift τ (Section 5.1: 'we typically choose it in dependence of the ground state energy τk = αE0'). In the paper's own rate formula, τ = E0 gives rPII = 0, and any τ = αE0 gives a factor determined by the oracle-supplied E0. The observed fast convergence and apparent robustness to small spectral gaps are therefore forced by the input E0, not by an independent prediction from the algorithm. The abstract's 'accurate estimation' language makes this an input-called-prediction step. The functional derivation of PII itself is not circular, but this numerical validation is partly circular.

full rationale

The mathematical derivation chain is self-contained and does not reduce to its inputs: PII is derived by Galerkin-projecting shifted inverse iteration (Eqs. (3.18)-(3.20), (5.1)), Theorem 3.1 is a standard spectral-gap analysis of L2 gradient flow, and Theorem 3.8 is the classical convergence theorem for inverse iteration, proved in the paper with no reliance on the target energy. The Galerkin projection step is a genuine mathematical construction, and the claimed correspondence between SR and L2 gradient descent is an identity, not a tautology. Self-citations such as [28,29,30] are used for motivation and attribution, not as load-bearing justification. The only substantive circularity is in the empirical validation: PII's headline speed advantage is demonstrated with the exact ground-state energy E0 (obtained by Lanczos for the 4×4 system, and used as τ for the 10×10 systems) inserted into the shift, while the paper's own rate formula makes the convergence factor 0 for τ=E0. This is a partial, validation-level circularity; the central theoretical content remains independent. Separate limitations — the unproved transfer from functional to parameter-space dynamics (Section 2.1 assertion; Section 5.1 disclaimer about learning-rate transfer) and manual hyperparameter tuning — are correctness/support concerns, not circularity. Score 4 reflects a partially circular numerical demonstration rather than a circular derivation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central algorithm PII and its convergence analysis rest on four sets of premises: the Hamiltonian's spectral properties, the real/stoquastic restriction, the expressiveness of the neural-network tangent space, and the reliability of Monte Carlo estimates for the preconditioning matrices. The first is standard; the last two are not quantified in the paper. The free parameters are the manually tuned learning rate, diagonal regularization, and spectral shift, the last of which requires a priori knowledge of the ground-state energy.

free parameters (3)
  • Learning rate eta = SR eta=0.01, PII eta=0.5 (4x4 full sum); minPII eta=1.0 (10x10)
    Manually tuned for fastest convergence in each experiment (Section 5.1, Figure 1; Section 5.2). The theoretical optimal step size in Lemma 3.3 is not used in the numerics.
  • Diagonal regularization epsilon = epsilon=0.001 for SR; epsilon=0.1 or 0.01 for PII
    Added to preconditioning matrices to make linear systems solvable and stabilize estimates; chosen by hand.
  • Spectral shift tau (parameter alpha) = tau = E0, 1.2*E0, 1.4*E0 (alpha=1, 1.2, 1.4)
    Controls the globalization of Newton's method; requires a priori knowledge of the ground-state energy, which is the target of the optimization. The paper shows tau=E0 is unstable and larger alpha stabilizes.
assumptions (4)
  • domain assumption The Hamiltonian is a densely defined, self-adjoint operator, bounded below, with a spectral gap E1 > E0.
    Used in Theorems 3.1, 3.5, and 3.8 to analyze convergence of the flow and inverse iteration. Typical for VMC but not always easy to verify for realistic Hamiltonians.
  • domain assumption The wavefunction is real-valued; applications are restricted to stoquastic Hamiltonians.
    Stated at the start of Section 2 and in Section 5.2. Excludes fermionic sign problems and complex wavefunctions.
  • ad hoc to paper The neural-network ansatz is expressive enough that the Galerkin subspace V_theta approximates the functional update directions well.
    The analysis of SR uses the assumption 'Assuming expressive spaces V_theta' (Section 3.1). The convergence theorems do not quantify the projection error for nonlinear ansatze.
  • domain assumption Monte Carlo estimators of the overlap matrix, Hamiltonian matrix, and gradient are sufficiently accurate for the preconditioner.
    The paper shows the Hamiltonian estimator is only asymptotically symmetric (Lemma 2.1) and observes instability at small sample sizes or at tau=E0 (Figure 2), indicating the estimates can be poor.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Functional Neural Wavefunction Optimization." pith.science (2026). https://pith.science/paper/ILQ2MFLM

@misc{pith2026250710835,
  author       = {Pith},
  title        = {Pith review of: Functional Neural Wavefunction Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILQ2MFLM}},
  note         = {Machine review of arXiv:2507.10835}
}
read the original abstract

We propose a framework for the design and analysis of optimization algorithms in variational quantum Monte Carlo, drawing on geometric insights into the corresponding function space. The framework translates infinite-dimensional optimization dynamics into tractable parameter-space algorithms through a Galerkin projection onto the tangent space of the variational ansatz. This perspective unifies existing methods such as stochastic reconfiguration and Rayleigh-Gauss-Newton, provides connections to classic function-space algorithms, and motivates the derivation of novel algorithms with geometrically principled hyperparameter choices. We validate our framework with numerical experiments demonstrating its practical relevance through the accurate estimation of ground-state energies for several prototypical models in condensed matter physics modeled with neural network wavefunctions.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Is Variational Monte Carlo Robust? Sharp Moment Thresholds and Heavy-tailed Stochastic Optimization

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    VMC's gradient estimators are generically heavy-tailed (no 3/2 moment for Slater–Jastrow); PS-Clip-VMC, which clips energies and per-sample gradients, is provably convergent under weak moments and stabilizes FermiNet ...

  2. A Sketch-and-Project Analysis of Subsampled Natural Gradient Algorithms

    cs.LG 2025-08 conditional novelty 6.0 of 10

    For linear least squares, SNGD and SPRING are proved equivalent to accelerated regularized Kaczmarz methods, yielding the first fast rates and first SPRING guarantee; the general quadratic analysis holds under strong ...

Reference graph

Works this paper leans on

55 extracted references · 51 canonical work pages · cited by 2 Pith papers

  1. [1]

    Astrakhantsev, T

    N. Astrakhantsev, T. Westerhout, A. Tiwari, K. Choo, A. Chen, M. H. Fischer, G. Carleo, and T. Neupert , Broken-Symmetry Ground States of the Heisenberg Model on the Pyrochlore Lattice, Physical Review X, 11 (2021), p. 041021

    work page 2021

  2. [2]

    N. Ay, J. Jost, H. Vˆan Lˆe, and L. Schwachh¨ofer, Information Geometry, vol. 64, Springer, 2017

    work page 2017

  3. [3]

    Becca and S

    F. Becca and S. Sorella , Quantum Monte Carlo Approaches for Correlated Systems , Cam- bridge University Press, 2017

    work page 2017

  4. [4]

    Boumal, An introduction to optimization on smooth manifolds, Cambridge University Press, 2023

    N. Boumal, An introduction to optimization on smooth manifolds, Cambridge University Press, 2023

    2023

  5. [5]

    Carleo and M

    G. Carleo and M. Troyer , Solving the quantum many-body problem with artificial neural networks, Science, 355 (2017), pp. 602–606

    work page 2017

  6. [6]

    Chen and M

    A. Chen and M. Heyl, Empowering deep neural quantum states through efficient optimization, Nature Physics, 20 (2024), pp. 1476–1481

    work page 2024

  7. [7]

    Cuzzocrea, A

    A. Cuzzocrea, A. Scemama, W. J. Briels, S. Moroni, and C. Filippi , Variational Princi- ples in Quantum Monte Carlo: The Troubled Story of Variance Minimization , Journal of chemical theory and computation, 16 (2020), pp. 4203–4212

    work page 2020

  8. [8]

    Dangel, J

    F. Dangel, J. M ¨uller, and M. Zeinhofer , Kronecker-Factored Approximate Curvature for Physics-Informed Neural Networks , in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024, https://openreview.net/forum?id=jrNlWfor7q

    work page 2024

Show all 55 references
  1. [9]

    Denis and G

    Z. Denis and G. Carleo, Accurate neural quantum states for interacting lattice bosons, Quan- tum, 9 (2025), p. 1772

    2025

  2. [10]

    Drissi, J

    M. Drissi, J. T. Keeble, J. Rozal ´en Sarmiento, and A. Rios , Second-order optimization strategies for neural network quantum states , Philosophical Transactions A, 382 (2024), p. 20240057

    2024

  3. [11]

    Emonts, M

    P. Emonts, M. C. Ba˜nuls, I. Cirac, and E. Zohar, Variational Monte Carlo simulation with tensor networks of a pure Z3 gauge theory in (2 + 1)D, Physical Review D, 102 (2020), p. 074501

    2020

  4. [12]

    Engel, R

    K.-J. Engel, R. Nagel, and S. Brendle , One-Parameter Semigroups for Linear Evolution Equations, vol. 194, Springer, 2000

    2000

  5. [13]

    Foster, Z

    A. Foster, Z. Sch ¨atzle, P. B. Szab ´o, L. Cheng, J. K ¨ohler, G. Cassella, N. Gao, J. Li, F. No´e, and J. Hermann, An ab initio foundation model of wavefunctions that accurately describes chemical bond breaking, arXiv preprint arXiv:2506.19960, (2025)

    2025 arXiv

  6. [14]

    Goldshlager, N

    G. Goldshlager, N. Abrahamsen, and L. Lin , A Kaczmarz-inspired approach to accelerate the optimization of neural network wavefunctions , Journal of Computational Physics, 516 (2024), p. 113351

    2024

  7. [15]

    Hermann, Z

    J. Hermann, Z. Sch ¨atzle, and F. No ´e, Deep-neural-network solution of the electronic Schr¨ odinger equation, Nature Chemistry, 12 (2020), pp. 891–897

    2020

  8. [16]

    Hermann, J

    J. Hermann, J. Spencer, K. Choo, A. Mezzacapo, W. M. C. Foulkes, D. Pfau, G. Carleo, and F. No ´e, Ab initio quantum chemistry with neural-network wavefunctions , Nature Reviews Chemistry, 7 (2023), pp. 692–709

    2023

  9. [17]

    Hibat-Allah, M

    M. Hibat-Allah, M. Ganahl, L. E. Hayward, R. G. Melko, and J. Carrasquilla , Recur- rent neural network wave functions , Phys. Rev. Res., 2 (2020), p. 023358, https://doi.org/ 10.1103/PhysRevResearch.2.023358, https://link.aps.org/doi/10.1103/PhysRevResearch. 2.023358

    2020 doi

  10. [18]

    J. Kim, G. Pescia, B. Fore, J. Nys, G. Carleo, S. Gandolfi, M. Hjorth-Jensen, and FUNCTIONAL NEURAL W A VEFUNCTION OPTIMIZATION 25 A. Lovato, Neural-network quantum states for ultra-cold Fermi gases , Communications Physics, 7 (2024), p. 148

    2024

  11. [19]

    Lange, A

    H. Lange, A. V an de W alle, A. Abedinnia, and A. Bohrdt, From architectures to applica- tions: A review of neural quantum states , Quantum Science and Technology, (2024)

    2024

  12. [20]

    R. Li, H. Ye, D. Jiang, X. Wen, C. W ang, Z. Li, X. Li, D. He, J. Chen, W. Ren, et al., A computational framework for neural network-based variational Monte Carlo with Forward Laplacian, Nature Machine Intelligence, 6 (2024), pp. 209–219

    2024

  13. [21]

    Linteau, G

    D. Linteau, G. Pescia, J. Nys, G. Carleo, and M. Holzmann , Phase Diagram and Crystal Melting of Helium-4 in Two Dimensions , Physical Review Letters, 134 (2025), p. 246001

    2025

  14. [22]

    W.-Y. Liu, H. Zhai, R. Peng, Z.-C. Gu, and G. K.-L. Chan , Accurate simulation of the hub- bard model with finite fermionic projected entangled pair states , Physical Review Letters, 134 (2025), p. 256502

    2025

  15. [23]

    W. T. Lou, H. Sutterud, G. Cassella, W. M. C. Foulkes, J. Knolle, D. Pfau, and J. S. Spencer, Neural Wave Functions for Superfluids, Physical Review X, 14 (2024), p. 021030

    2024

  16. [24]

    Martens and R

    J. Martens and R. Grosse , Optimizing Neural Networks with Kronecker-factored Approxi- mate Curvature, in International conference on machine learning, PMLR, 2015, pp. 2408– 2417

    2015

  17. [25]

    Medvidovi´c and J

    M. Medvidovi´c and J. R. Moreno , Neural-network quantum states for many-body physics , The European Physical Journal Plus, 139 (2024), pp. 1–26

    2024

  18. [26]

    M. S. Moss, R. Wiersema, M. Hibat-Allah, J. Carrasquilla, and R. G. Melko , Leverag- ing recurrence in neural network wavefunctions for large-scale simulations of Heisenberg antiferromagnets: The triangular lattice , June 2025, https://doi.org/10.48550/arXiv.2505. 20406, https...

    2025 doi

  19. [27]

    Motta, K

    M. Motta, K. J. Sung, and J. Shee , Quantum Algorithms for the Variational Optimization of Correlated Electronic States with Stochastic Reconfiguration and the Linear Method , The Journal of Physical Chemistry A, 128 (2024), pp. 8762–8776

    2024

  20. [28]

    M¨uller and M

    J. M¨uller and M. Zeinhofer, Achieving High Accuracy with PINNs via Energy Natural Gra- dient Descent, in International Conference on Machine Learning, PMLR, 2023, pp. 25471– 25485

    2023

  21. [29]

    M¨uller and M

    J. M¨uller and M. Zeinhofer, Position: Optimization in SciML Should Employ the Function Space Geometry, in Forty-first International Conference on Machine Learning, 2024

    2024

  22. [30]

    Neklyudov, J

    K. Neklyudov, J. Nys, L. Thiede, J. Carrasquilla, Q. Liu, M. Welling, and A. Makhzani , Wasserstein Quantum Monte Carlo: A Novel Approach for Solving the Quantum Many-Body Schr¨ odinger Equation, Advances in Neural Information Processing Systems, 36 (2023), pp. 63461–63482

    2023

  23. [31]

    Neuscamman, C

    E. Neuscamman, C. Umrigar, and G. K.-L. Chan , Optimizing large parameter sets in variational quantum Monte Carlo , Physical Review B—Condensed Matter and Materi- als Physics, 85 (2012), p. 045103

    2012

  24. [32]

    Nightingale and V

    M. Nightingale and V. Melik-Alaverdian , Optimization of Ground- and Excited-State Wave Functions and van der Waals Clusters , Physical review letters, 87 (2001), p. 043401

    2001

  25. [33]

    Nocedal and S

    J. Nocedal and S. J. Wright , Numerical Optimization, Springer, 1999

    1999

  26. [34]

    Peng and G

    R. Peng and G. K. Chan , An analysis of first-and second-order optimization algorithms in variational monte carlo , arXiv preprint arXiv:2502.19576, (2025)

    2025 arXiv

  27. [35]

    Pescia, J

    G. Pescia, J. Nys, J. Kim, A. Lovato, and G. Carleo , Message-passing neural quantum states for the homogeneous electron gas , Physical Review B, 110 (2024), p. 035108

    2024

  28. [36]

    D. Pfau, S. Axelrod, H. Sutterud, I. von Glehn, and J. S. Spencer, Accurate computation of quantum excited states with neural networks , Science, 385 (2024), p. eadn0137

    2024

  29. [37]

    D. Pfau, J. S. Spencer, A. G. Matthews, and W. M. C. Foulkes , Ab initio solution of the many-electron Schr¨ odinger equation with deep neural networks, Physical review research, 2 (2020), p. 033429

    2020

  30. [38]

    Rende, L

    R. Rende, L. L. Viteritti, L. Bardone, F. Becca, and S. Goldt , A simple linear algebra identity to optimize large-scale neural network quantum states , Communications Physics, 7 (2024), p. 260

    2024

  31. [39]

    Robledo Moreno, G

    J. Robledo Moreno, G. Carleo, A. Georges, and J. Stokes , Fermionic wave functions from neural-network constrained hidden states , Proceedings of the National Academy of Sciences, 119 (2022), p. e2122059119

    2022

  32. [40]

    Scherbela, L

    M. Scherbela, L. Gerard, and P. Grohs , Towards a transferable fermionic neural wave- function for molecules , Nature Communications, 15 (2024), p. 120

    2024

  33. [41]

    Sorella , Green function Monte Carlo with stochastic reconfiguration , Physical Review Letters, 80 (1998), p

    S. Sorella , Green function Monte Carlo with stochastic reconfiguration , Physical Review Letters, 80 (1998), p. 4558

    1998

  34. [42]

    Sorella , Generalized Lanczos algorithm for variational quantum Monte Carlo , Physical Review B, 64 (2001), p

    S. Sorella , Generalized Lanczos algorithm for variational quantum Monte Carlo , Physical Review B, 64 (2001), p. 024512. 26 ARMEGIOIU, CARRASQUILLA, MISHRA, M ¨ULLER, NYS, ZEINHOFER, ZHANG

    2001

  35. [43]

    Stokes, J

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

    2020

  36. [44]

    Toulouse and C

    J. Toulouse and C. J. Umrigar , Optimization of quantum Monte Carlo wave functions by energy minimization, The Journal of chemical physics, 126 (2007)

    2007

  37. [45]

    L. N. Trefethen and D. Bau , Numerical Linear Algebra, SIAM, 2022

    2022

  38. [46]

    Umrigar, J

    C. Umrigar, J. Toulouse, C. Filippi, S. Sorella, and R. G. Hennig , Alleviation of the Fermion-Sign Problem by Optimization of Many-Body Wave Functions , Physical review letters, 98 (2007), p. 110201

    2007

  39. [47]

    Vicentini, D

    F. Vicentini, D. Hofmann, A. Szab ´o, D. Wu, C. Roth, C. Giuliani, G. Pescia, J. Nys, V. V argas-Calder´on, N. Astrakhantsev, et al., NetKet 3: Machine Learning Toolbox for Many-Body Quantum Systems , SciPost Physics Codebases, (2022), p. 007

    2022

  40. [48]

    Vieijra and J

    T. Vieijra and J. Nys , Many-body quantum states with exact conservation of non-abelian and lattice symmetries through variational Monte Carlo , Physical Review B, 104 (2021), p. 045123

    2021

  41. [49]

    L. L. Viteritti, R. Rende, and F. Becca , Transformer Variational Wave Functions for Frustrated Quantum Spin Systems , Physical Review Letters, 130 (2023), p. 236401

    2023

  42. [50]

    von Glehn, J

    I. von Glehn, J. S. Spencer, and D. Pfau , A self-attention ansatz for ab-initio quantum chemistry, in The Eleventh International Conference on Learning Representations, 2023, https://openreview.net/forum?id=xveTeHVlF7j

    2023

  43. [51]

    R. J. Webber and M. Lindsey, Rayleigh-Gauss-Newton optimization with enhanced sampling for variational Monte Carlo , Physical Review Research, 4 (2022), p. 033099

    2022

  44. [52]

    D. Wu, R. Rossi, F. Vicentini, N. Astrakhantsev, F. Becca, X. Cao, J. Carrasquilla, F. Ferrari, A. Georges, M. Hibat-Allah, et al. , Variational benchmarks for quantum many-body problems, Science, 386 (2024), pp. 296–301

    2024

  45. [53]

    Wu and Z

    Y. Wu and Z. Dai , Algorithms for variational Monte Carlo calculations of fermion PEPS in the swap gates formulation , arXiv preprint arXiv:2506.20106, (2025)

    2025 arXiv

  46. [54]

    Xu and L

    J. Xu and L. Zikatanov , Some observations on babuska and brezzi theories , Numerische Mathematik, 94 (2003), pp. 195–202

    2003

  47. [55]

    X. Yuan, S. Endo, Q. Zhao, Y. Li, and S. C. Benjamin , Theory of variational quantum simulation, Quantum, 3 (2019), p. 191

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.