pith. sign in

arxiv: 2606.20791 · v1 · pith:IBPY5EOWnew · submitted 2026-06-18 · ✦ hep-lat · hep-ph· hep-th· nucl-th· quant-ph

Neural Wavefunctions in Quantum Field Theory I: Asymptotic Freedom

Paulo F. Bedaque , Hersh Kumar , Suryansh Rajawat , Gregory Ridgway This is my paper

Pith reviewed 2026-06-26 14:27 UTC · model grok-4.3

classification ✦ hep-lat hep-phhep-thnucl-thquant-ph
keywords neural networksvariational methodsquantum field theorynonlinear sigma modelasymptotic freedomdynamical mass generationstep-scaling functionlattice field theory
0
0 comments X

The pith

Neural-network wavefunctions enable variational calculations that reproduce asymptotic freedom, dynamical mass generation, and the step-scaling function in the two-dimensional nonlinear sigma model.

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

The paper develops a variational method for quantum field theory in which the wavefunction is parameterized by a neural network. Modern machine-learning optimization is used to minimize the energy expectation value on a lattice discretization of the two-dimensional nonlinear sigma model. The resulting calculations recover the model's known non-perturbative signatures, including the running of the coupling toward zero at short distances, the generation of a mass gap, and the step-scaling function that describes how the coupling changes with scale. A sympathetic reader would care because the approach offers a route to variational studies of field theories whose non-perturbative physics has been difficult to access with conventional trial wavefunctions.

Core claim

We present a variational approach to quantum field theory based on wavefunctions parameterized by neural networks. We show that neural-network wavefunctions, combined with modern machine-learning techniques, enable competitive variational calculations in nontrivial field theories. As a demonstration, we reproduce the essential features of the two-dimensional nonlinear σ-model: asymptotic freedom, dynamical mass generation and the model's step-scaling function.

What carries the argument

Neural-network parameterization of the many-body wavefunction inside a variational Monte Carlo energy minimization on the lattice-regularized theory.

If this is right

  • Variational calculations become feasible for field theories whose ground states exhibit asymptotic freedom or dynamical mass generation.
  • The step-scaling function can be extracted directly from finite-volume energy minimizations without perturbative matching.
  • Dynamical mass generation appears as a finite correlation length in the optimized neural wavefunction.
  • The same parameterization can be used to test other lattice-regularized models that share the same symmetries.

Where Pith is reading between the lines

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

  • If the network architecture generalizes, the method could be tested on four-dimensional theories where traditional variational ansatze are intractable.
  • The approach may complement Monte Carlo methods in regimes where the sign problem appears, provided the wavefunction can be optimized without stochastic sampling of the action.
  • Systematic improvement could be obtained by increasing network depth or width and checking convergence of the step-scaling function against continuum extrapolations.

Load-bearing premise

The neural network is expressive enough to represent the true ground-state wavefunction of the field theory without missing essential non-perturbative structure.

What would settle it

A computed step-scaling function that deviates from established non-perturbative results for the two-dimensional nonlinear sigma model at multiple lattice spacings would falsify the claim that the neural wavefunction captures the essential physics.

Figures

Figures reproduced from arXiv: 2606.20791 by Gregory Ridgway, Hersh Kumar, Paulo F. Bedaque, Suryansh Rajawat.

Figure 1
Figure 1. Figure 1: FIG. 1: Top: Gaps ∆ as a function of volume [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Step scaling curve [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

We present a variational approach to quantum field theory based on wavefunctions parameterized by neural networks. While variational methods have a celebrated history across many fields, their application to quantum field theory has been limited by well-known challenges. We show that neural-network wavefunctions, combined with modern machine-learning techniques, enable competitive variational calculations in nontrivial field theories. As a demonstration, we reproduce the essential features of the two-dimensional nonlinear $\sigma$-model: asymptotic freedom, dynamical mass generation and the model's step-scaling function.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 0 minor

Summary. The manuscript introduces a variational approach to quantum field theory in which wavefunctions are parameterized by neural networks. It demonstrates the method on the two-dimensional nonlinear σ-model, claiming to reproduce the essential non-perturbative features of asymptotic freedom, dynamical mass generation, and the model's step-scaling function.

Significance. If the neural-network ansatz is shown to be sufficiently expressive and unbiased, the work would constitute a meaningful advance by extending modern machine-learning variational techniques to nontrivial continuum field theories, potentially offering a new computational route complementary to lattice Monte Carlo methods.

major comments (1)
  1. [Demonstration of the nonlinear σ-model (abstract and corresponding results section)] The central claim that the neural-network wavefunction reproduces the three key features of the 2D nonlinear σ-model rests on the unquantified assumption that the chosen parameterization spans a sufficiently dense and unbiased subset of the Hilbert space. No bound on the variational approximation error, comparison against an exactly solvable limit, or test of architectural bias (e.g., receptive-field limitations or incomplete symmetry enforcement) is supplied; failure of this assumption would invalidate all three reproduced features.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed review and for highlighting the importance of validating the expressiveness of the neural-network ansatz. We address the single major comment below.

read point-by-point responses

  1. Referee: [Demonstration of the nonlinear σ-model (abstract and corresponding results section)] The central claim that the neural-network wavefunction reproduces the three key features of the 2D nonlinear σ-model rests on the unquantified assumption that the chosen parameterization spans a sufficiently dense and unbiased subset of the Hilbert space. No bound on the variational approximation error, comparison against an exactly solvable limit, or test of architectural bias (e.g., receptive-field limitations or incomplete symmetry enforcement) is supplied; failure of this assumption would invalidate all three reproduced features.

    Authors: We agree that a rigorous a-priori bound on the variational error would strengthen the claims, but such bounds are generally unavailable for high-dimensional variational ansatzes in quantum field theory. Validation instead proceeds via systematic convergence with network depth/width, consistency across independent observables (mass gap, step-scaling function, and asymptotic-freedom signatures), and agreement with known perturbative and non-perturbative results. The manuscript already reports convergence tests with increasing network capacity and explicit enforcement of the O(3) symmetry; we will expand the results section with additional plots of these convergence diagnostics and a short discussion of possible architectural biases. We therefore view the central claim as supported by the existing numerical evidence, while acknowledging that a formal error bound remains out of reach. revision: partial

Circularity Check

0 steps flagged

No circularity: variational demonstration validated on independently known model features

full rationale

The paper introduces a neural-network variational ansatz for QFT wavefunctions and applies it to the 2D nonlinear σ-model to reproduce its established continuum features (asymptotic freedom, mass gap, step-scaling). These targets are taken from the existing literature on the model rather than being derived from the ansatz itself. No equation or procedure reduces a claimed prediction to a fitted input by construction, no self-citation supplies a load-bearing uniqueness theorem, and the central claim rests on numerical reproduction of external benchmarks. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to identify specific free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5623 in / 991 out tokens · 30848 ms · 2026-06-26T14:27:27.555335+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 1 canonical work pages

  1. [1]

    for MPS calculations. H.K. was supported by the DOE, Office of Science, Office of High Energy Physics (HEP), HEP Computing Traineeship program: Lattice Gauge Theory for HEP (under grant no. DE-SC0024053). The authors acknowl- edge the University of Maryland computing resources made available for conducting the research reported in this paper. 5

  2. [2]

    Bardeen, L

    J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Microscopic theory of superconductivity, Phys. Rev.106, 162 (1957)

    1957

  3. [3]

    Nambu and G

    Y. Nambu and G. Jona-Lasinio, Dynamical model of elemen- tary particles based on an analogy with superconductivity. i, Phys. Rev.122, 345 (1961)

    1961

  4. [4]

    Rovira, R

    M. Rovira, R. Perry, and A. Parreno, A Variational Ap- proach to Quantum Field Theory, PoSQNP2024, 036 (2025), arXiv:2409.17887 [hep-lat]

    arXiv 2025

  5. [5]

    Braga, N

    K. Braga, N. Sato, and A. Szczepaniak, Variational neural network approach to qft in the field basis, Physics Letters B 873, 140183 (2026)

    2026

  6. [6]

    J. M. Martyn, K. Najafi, and D. Luo, Variational neural- network ansatz for continuum quantum field theory, Phys. Rev. Lett.131, 081601 (2023)

    2023

  7. [7]

    R. P. Feynman, Difficulties in applying the variational princi- ple to quantum field theories, inVariational Calculations in Quantum Field Theory(World Scientific, 1988) pp. 28–40

    1988

  8. [8]

    Lovato, G

    A. Lovato, G. Carleo, B. Fore, M. Hjorth-Jensen, J. Kim, A. Rios, and N. Rocco, Neural-network quantum states for the nuclear many-body problem (2026), arXiv:2602.13826 [nucl-th]

    arXiv 2026

  9. [9]

    P. F. Bedaque, J. Cigliano, H. Kumar, S. Paul, and S. Rajawat, Kolmogorov-Arnold wavefunctions, Phys. Rev. C112, 054002 (2025), arXiv:2506.02171 [nucl-th]

    arXiv 2025

  10. [10]

    P. F. Bedaque, J. Cigliano, H. Kumar, S. Paul, and S. Rajawat, Trapped fermions through kolmogorov-arnold wavefunctions (2025), arXiv:2512.07800 [nucl-th]

    arXiv 2025

  11. [11]

    Gnech, B

    A. Gnech, B. Fore, A. J. Tropiano, and A. Lovato, Distilling the essential elements of nuclear binding via neural-network quantum states, Physical Review Letters133, 142501 (2024)

    2024

  12. [12]

    P. Wen, A. Gezerlis, and J. W. Holt, Neural quantum states for light nuclei with chiral two- and three-body interactions, Physical Review Letters136, 10.1103/ygkf-llyp (2026)

    work page doi:10.1103/ygkf-llyp 2026

  13. [13]

    Hermann, Z

    J. Hermann, Z. Sch¨ atzle, and F. No´ e, Deep-neural-network solution of the electronic schr¨ odinger equation, Nature Chem- istry12, 891–897 (2020)

    2020

  14. [14]

    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 research2, 033429 (2020)

    2020

  15. [15]

    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 B110, 035108 (2024)

    2024

  16. [16]

    Carleo and M

    G. Carleo and M. Troyer, Solving the quantum many-body problem with artificial neural networks, Science355, 602 (2017)

    2017

  17. [17]

    Saito, Solving the Bose–Hubbard Model with Ma- chine Learning, J

    H. Saito, Solving the Bose–Hubbard Model with Ma- chine Learning, J. Phys. Soc. Jap.86, 093001 (2017), arXiv:1707.09723 [cond-mat.dis-nn]

    Pith/arXiv arXiv 2017

  18. [18]

    A. M. Polyakov, Interaction of Goldstone Particles in Two- Dimensions. Applications to Ferromagnets and Massive Yang- Mills Fields, Phys. Lett. B59, 79 (1975)

    1975

  19. [19]

    P. B. Wiegmann, Exact solution of the o(3) nonlinear sigma- model, Physics Letters B152, 209 (1985)

    1985

  20. [20]

    Hasenfratz, M

    P. Hasenfratz, M. Maggiore, and F. Niedermayer, The exact mass gap of the o(3) and o(4) non-linear sigma-models in d= 2, Physics Letters B245, 522 (1990)

    1990

  21. [21]

    Luscher and U

    M. Luscher and U. Wolff, How to Calculate the Elastic Scat- tering Matrix in Two-dimensional Quantum Field Theories by Numerical Simulation, Nucl. Phys. B339, 222 (1990). [21]n without an index stands for the whole field configuration n1,· · ·,n L

    1990

  22. [22]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White, Multilayer feedfor- ward networks are universal approximators, Neural networks 2, 359 (1989)

    1989

  23. [23]

    Blum-Smith and S

    B. Blum-Smith and S. Villar, Machine learning and invariant theory, Notices of the American Mathematical Society70, 1 (2023)

    2023

  24. [24]

    Sorella, Green function monte carlo with stochastic recon- figuration, Phys

    S. Sorella, Green function monte carlo with stochastic recon- figuration, Phys. Rev. Lett.80, 4558 (1998)

    1998

  25. [25]

    Neuscamman, C

    E. Neuscamman, C. J. Umrigar, and G. K.-L. Chan, Optimiz- ing large parameter sets in variational quantum monte carlo, Physical Review B85, 045103 (2012)

    2012

  26. [26]

    Wolff, Collective monte carlo updating for spin systems, Physical Review Letters62, 361 (1989)

    U. Wolff, Collective monte carlo updating for spin systems, Physical Review Letters62, 361 (1989)

    1989

  27. [27]

    For L = 28, training on 2 NVIDIA A100s was done in 25 hours

  28. [28]

    Bruckmann, K

    F. Bruckmann, K. Jansen, and S. K¨ uhn, O (3) nonlinear sigma model in 1+ 1 dimensions with matrix product states, Physical Review D99, 074501 (2019)

    2019

  29. [29]

    Caracciolo, R

    S. Caracciolo, R. G. Edwards, A. Pelissetto, and A. D. Sokal, Asymptotic scaling in the two-dimensionalo(3) σ model at cor- relation length 10 5, Physical Review Letters75, 1891 (1995)

    1995

  30. [30]

    Luscher, Volume Dependence of the Energy Spectrum in Massive Quantum Field Theories

    M. Luscher, Volume Dependence of the Energy Spectrum in Massive Quantum Field Theories. 1. Stable Particle States, Commun. Math. Phys.104, 177 (1986)

    1986

  31. [31]

    Luscher, Volume Dependence of the Energy Spectrum in Massive Quantum Field Theories

    M. Luscher, Volume Dependence of the Energy Spectrum in Massive Quantum Field Theories. 2. Scattering States, Commun. Math. Phys.105, 153 (1986)

    1986

  32. [32]

    Bhattacharya, A

    T. Bhattacharya, A. J. Buser, S. Chandrasekharan, R. Gupta, and H. Singh, Qubit regularization of asymptotic freedom, Phys. Rev. Lett.126, 172001 (2021)

    2021

  33. [33]

    Caracciolo, R

    S. Caracciolo, R. G. Edwards, S. J. Ferreira, A. Pelissetto, and A. D. Sokal, Extrapolating monte carlo simulations to infinite volume: Finite-size scaling at ξ/L ≫ 1, Phys. Rev. Lett.74, 2969 (1995)

    1995

  34. [34]

    I. L. Guti´ errez and C. B. Mendl, Real time evolution with neural-network quantum states, Quantum6, 627 (2022)

    2022

  35. [35]

    W. Fu, Z. Li, Y. Qian, R. Li, W. Ren, and J. Chen, To- wards stable and accurate electron dynamics via neural net- work based time-dependent variational monte carlo (2026), arXiv:2606.05850 [physics.comp-ph]

    Pith/arXiv arXiv 2026

  36. [36]

    Bradbury, R

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

    2018

  37. [37]

    N. Gao, J. K¨ ohler, and A. Foster, folx - forward laplacian for jax (2023). 6

    2023

  38. [38]

    R. Li, H. Ye, D. Jiang, X. Wen, C. Wang, Z. Li, X. Li, D. He, J. Chen, W. Ren, and L. Wang, Forward laplacian: A new computational framework for neural network-based variational monte carlo, arXiv preprint arXiv:2307.08214 (2023)

    arXiv 2023

  39. [39]

    Fishman, S

    M. Fishman, S. R. White, and E. M. Stoudenmire, The ITen- sor Software Library for Tensor Network Calculations, SciPost Phys. Codebases , 4 (2022)

    2022