arxiv: 2605.12488 · v1 · submitted 2026-05-12 · ✦ hep-th

Recognition: 2 theorem links

· Lean Theorem

Anomalies in Neural Network Field Theory

Christian Ferko, James Halverson, Samuel Frank, Vishnu Jejjala

Pith reviewed 2026-05-13 03:13 UTC · model grok-4.3

classification ✦ hep-th

keywords neural network field theoryanomaliesWard identitiesSchwinger-Dyson equationsparameter space currentWeyl anomalybosonic stringU(1) symmetry

0 comments

The pith

Anomalies in quantum field theories can be understood through conserved currents in neural network parameter space.

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

The paper formulates quantum field theories using neural network architectures and a density over their parameters. It derives Schwinger-Dyson equations and Ward identities that depend on a conserved parameter space current describing symmetries and how they break. This current applies even to non-local theories and recovers standard local currents through an appropriate average when the underlying Lagrangian is local. The approach is used to examine U(1) symmetry for a complex scalar, the scale anomaly in four-dimensional massless scalar theory, the Weyl anomaly for the bosonic string including a computation of the critical dimension, and discrete cases involving winding numbers and T-duality. A reader would care because the results live in network parameter space rather than ordinary field space, offering a distinct perspective on symmetries.

Core claim

Neural network field theory formulates field theory in terms of a network architecture and a density on its parameters. Schwinger-Dyson equations and Ward identities are derived and used to study anomalies. These equations depend on a conserved parameter space current that characterizes symmetries and how they break. The current remains relevant even in non-local NN-FTs but recovers local currents in the case of a local Lagrangian by fiber-wise average. Applications cover U(1) symmetry for a complex scalar, the scale anomaly in 4d massless phi^4 theory, the Weyl anomaly for the bosonic string with a new computation of the critical dimension, and examples with discrete topological data such a

What carries the argument

The conserved parameter space current that characterizes symmetries and how they break in the neural network formulation.

If this is right

Symmetries and anomalies can be analyzed directly in the space of network parameters for both local and non-local theories.
Standard local currents are recovered by fiber-wise averaging when the Lagrangian is local.
The critical dimension of the bosonic string emerges from the same parameter-space machinery.
Discrete topological features such as winding numbers and T-duality fit into the anomaly analysis.

Where Pith is reading between the lines

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

The same current formalism could be applied to gauge anomalies or gravitational anomalies by selecting suitable network architectures.
Numerical sampling over parameter densities might offer new computational routes to anomaly coefficients in theories where field-space methods are cumbersome.

Load-bearing premise

The neural-network formulation of field theory faithfully reproduces the anomaly structure of ordinary quantum field theory through the introduction and averaging of a parameter-space current.

What would settle it

A direct computation of the Weyl anomaly coefficient in the bosonic string using the NN-FT Ward identities that produces a critical dimension other than 26 would show the parameter-space current does not correctly capture the anomaly.

read the original abstract

Neural network field theory (NN-FT) formulates field theory in terms of a network architecture and a density on its parameters. We derive Schwinger--Dyson equations and Ward identities in NN-FT and utilize them to study anomalies. The equations depend on a conserved parameter space current that characterizes symmetries and how they break. It is relevant even in non-local NN-FTs, but can recover local currents in the case of a local Lagrangian by an appropriate fiber-wise average. In machine learning, this formalism is applied to feedforward networks and the attention mechanism. In physics, we use this machinery to study $U(1)$ symmetry for a complex scalar, the scale anomaly in $4d$ massless $\phi^4$ theory, the Weyl anomaly for the bosonic string (including a new computation of the critical dimension), and examples involving discrete topological data, such as winding numbers and T-duality. Since the results are obtained in network parameter space rather than the standard field space, they represent a new way to understand symmetries in quantum field theories.

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.

Desk Editor's Note private letter to a colleague

NN-FT gives a parameter-space current for deriving anomalies that recovers known results in examples, but the fiber averaging step needs explicit checks for exact coefficients.

read the letter

The main takeaway is that this paper sets up Schwinger-Dyson and Ward identities inside neural network field theory using a conserved current defined on the network parameters. That current lets them study anomalies even in non-local versions of the theory, and they show how to average over fibers to recover ordinary local currents when the underlying Lagrangian is local. They then apply the setup to several standard cases and claim it reproduces the expected physics, including a fresh calculation of the bosonic string critical dimension.

Referee Report

2 major / 2 minor

Summary. The paper introduces Neural Network Field Theory (NN-FT), which reformulates field theory using a network architecture and a density over its parameters. It derives Schwinger-Dyson equations and Ward identities from a conserved parameter-space current that encodes symmetries and their breaking. This current remains relevant for non-local NN-FTs but recovers local currents via fiber-wise averaging when the underlying Lagrangian is local. The formalism is applied to machine-learning architectures (feedforward networks, attention) and to physics examples: U(1) symmetry for a complex scalar, the scale anomaly in 4d massless φ⁴ theory, the Weyl anomaly of the bosonic string (with a computation of the critical dimension), and discrete topological quantities such as winding numbers and T-duality. The central claim is that working in parameter space yields a new perspective on QFT symmetries and anomalies.

Significance. If the parameter-space current and fiber-wise averaging are shown to reproduce standard anomaly coefficients exactly, the framework would supply a genuinely new computational route to anomalies that applies equally to local, non-local, and discrete settings. The explicit treatment of the bosonic-string critical dimension constitutes a concrete, falsifiable test of the method. The approach also bridges machine-learning architectures with QFT Ward identities, which could prove useful for studying symmetries in models where the field-space formulation is cumbersome.

major comments (2)

[§3 and §4.3] §3 (derivation of Ward identities) and §4.3 (bosonic string): The fiber-wise averaging procedure is asserted to recover local currents and thereby the standard anomaly structures, yet no explicit calculation demonstrates that the numerical coefficients are preserved. In particular, the Weyl-anomaly computation that yields the critical dimension 26 must be shown to arise without extra or missing terms from the averaging; otherwise the reproduction of the known result is only qualitative and the central claim that NN-FT reproduces ordinary QFT anomalies is not yet load-bearing.
[§4.3] §4.3 (Weyl anomaly for bosonic string): The manuscript presents a “new computation” of the critical dimension. It is unclear whether this derivation is independent of the conventional Polyakov or light-cone methods or whether it implicitly imports the known coefficient through the choice of regularization or the definition of the parameter-space current. An explicit side-by-side comparison of the anomaly polynomial or central-charge term before and after averaging is required.

minor comments (2)

[Introduction] The notation for the parameter-space current J^μ and its conservation law should be introduced with an explicit definition in the main text rather than only in the abstract; readers unfamiliar with NN-FT will otherwise struggle to follow the transition from the Schwinger-Dyson equation to the Ward identity.
[§4.2] Figure captions and axis labels in the numerical examples (e.g., the φ⁴ scale anomaly plots) should state the precise network depth, width, and activation function used, so that the results can be reproduced.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments on our manuscript. We address each major comment point by point below.

read point-by-point responses

Referee: [§3 and §4.3] §3 (derivation of Ward identities) and §4.3 (bosonic string): The fiber-wise averaging procedure is asserted to recover local currents and thereby the standard anomaly structures, yet no explicit calculation demonstrates that the numerical coefficients are preserved. In particular, the Weyl-anomaly computation that yields the critical dimension 26 must be shown to arise without extra or missing terms from the averaging; otherwise the reproduction of the known result is only qualitative and the central claim that NN-FT reproduces ordinary QFT anomalies is not yet load-bearing.

Authors: We thank the referee for highlighting this important point. In §3 the fiber-wise averaging is derived as the operation that maps the conserved parameter-space current onto the standard local current for any local Lagrangian; the general Ward identity is obtained before averaging is applied. In §4.3 the Weyl anomaly for the bosonic string is first computed directly from the parameter-space current, producing the critical dimension. We agree that an explicit verification that the numerical coefficient is unchanged by the averaging step is necessary to make the reproduction quantitative rather than structural. We will add this side-by-side comparison (anomaly polynomial before and after averaging) in a new appendix or subsection of the revised manuscript. revision: yes
Referee: [§4.3] §4.3 (Weyl anomaly for bosonic string): The manuscript presents a “new computation” of the critical dimension. It is unclear whether this derivation is independent of the conventional Polyakov or light-cone methods or whether it implicitly imports the known coefficient through the choice of regularization or the definition of the parameter-space current. An explicit side-by-side comparison of the anomaly polynomial or central-charge term before and after averaging is required.

Authors: The computation begins from the definition of the Weyl current in parameter space and evaluates the associated Jacobian; the regularization is chosen only to be consistent with standard QFT schemes so that the result can be compared with the literature. The coefficient is not presupposed but obtained from the trace over the parameter-space measure. Nevertheless, we acknowledge that an explicit demonstration of independence is required. The revised manuscript will therefore contain the requested side-by-side comparison of the anomaly polynomial (or central-charge term) in parameter space and after fiber-wise averaging, showing that the coefficient 26 emerges identically from the averaged expression. revision: yes

Circularity Check

0 steps flagged

No circularity: derivations of SD equations, Ward identities, and anomaly coefficients proceed from the NN-FT parameter-space current without reduction to fitted inputs or self-citation chains.

full rationale

The paper defines the NN-FT setup, derives Schwinger-Dyson equations and Ward identities from a conserved parameter-space current, introduces fiber-wise averaging to recover local currents for local Lagrangians, and then applies the resulting identities to compute specific anomalies (U(1), scale, Weyl including critical dimension). These steps are presented as explicit derivations rather than tautological re-expressions of known results. No quoted equations show a 'prediction' that is statistically forced by a prior fit, nor does any load-bearing uniqueness theorem reduce to a self-citation. The central claim that results live in parameter space yet recover standard anomaly structures is therefore independent content, not circular by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper introduces NN-FT as a new language for field theory and postulates a conserved parameter-space current whose properties are used to derive anomalies. Standard mathematical identities (Schwinger-Dyson, Ward) are assumed to carry over. No free parameters are explicitly fitted in the abstract, but the mapping between network architecture and field theory is an unstated modeling choice.

axioms (1)

domain assumption Schwinger-Dyson equations and Ward identities hold in the NN-FT formulation
Invoked to derive the anomaly equations from the network architecture and parameter density.

invented entities (1)

conserved parameter space current no independent evidence
purpose: characterizes symmetries and how they break in NN-FT
New object introduced to encode symmetry information in parameter space; can be averaged to recover local currents.

pith-pipeline@v0.9.0 · 5483 in / 1381 out tokens · 118501 ms · 2026-05-13T03:13:07.491563+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes
The breaking function B(θ) := ∑ ξa sa + ∑ ∂a ξa ... p B = ∑ ∂a(ξa p) ... continuity equation
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking contradicts
Weyl anomaly ... mode counting ... Breg_anom = -1/3(D-26) ... D=26

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 3 internal anchors

[1]

Path-Integral Measure for Gauge-Invariant Fermion Theories

K. Fujikawa,“Path-Integral Measure for Gauge-Invariant Fermion Theories”, Phys. Rev. Lett.42, 1195 (1979)

work page 1979
[2]

Axial-Vector Vertex in Spinor Electrodynamics

S. L. Adler,“Axial-Vector Vertex in Spinor Electrodynamics”, Phys. Rev.177, 2426 (1969)

work page 1969
[3]

A PCAC Puzzle: π0 →γγ in the σ-Model

J. S. Bell & R. Jackiw,“A PCAC Puzzle: π0 →γγ in the σ-Model”, Nuovo Cim. A60, 47 (1969). 53

work page 1969
[4]

Halverson, A

J. Halverson, A. Maiti & K. Stoner,“Neural Networks and Quantum Field Theory”, Mach. Learn. Sci. Tech.2, 035002 (2021),arXiv:2008.08601 [cs.LG]

work page arXiv 2021
[7]

BAYESIAN LEARNING FOR NEURAL NETWORKS

R. M. Neal,“BAYESIAN LEARNING FOR NEURAL NETWORKS”, PhD thesis, Univer- sity of Toronto (1995)

work page 1995
[9]

Huang and K

G. Huang & K. Zhou,“The neural networks with tensor weights and emergent fermionic Wick rules in the large-width limit”, Phys. Lett. B873, 140146 (2026), arXiv:2507.05303 [hep-th]

work page arXiv 2026
[11]

Halverson, J

J. Halverson, J. Naskar & J. Tian,“Conformal fields from neural networks”, JHEP2510, 039 (2025),arXiv:2409.12222 [hep-th]

work page arXiv 2025
[12]

Robinson,Virasoro Symmetry in Neural Network Field Theories,2512.24420

B. Robinson,“Virasoro Symmetry in Neural Network Field Theories”, arXiv:2512.24420 [hep-th]

work page arXiv
[13]

Capuozzo, B

P. Capuozzo, B. Robinson & B. Suzzoni,“Conformal Defects in Neural Network Field Theories”,arXiv:2512.07946 [hep-th]

work page arXiv
[15]

Excited String States and D-branes from Infinite Width Neural Networks

D. S. Ageev & Y. A. Ageeva,“Excited String States and D-branes from Infinite Width Neural Networks”,arXiv:2602.10214 [hep-th]

work page arXiv
[16]

Topological Effects in Neural Network Field Theory

C. Ferko, J. Halverson, V. Jejjala & B. Robinson,“Topological Effects in Neural Network Field Theory”,arXiv:2604.02313 [hep-th]

work page internal anchor Pith review Pith/arXiv arXiv
[17]

Ferko and J

C. Ferko & J. Halverson,“Quantum Mechanics and Neural Networks”, Mach. Learn. Sci. Tech.7, 015002 (2026),arXiv:2504.05462 [hep-th]

work page arXiv 2026
[18]

Optimal Architecture and Fundamental Bounds in Neural Network Field Theory

Z. Zhang,“Optimal Architecture and Fundamental Bounds in Neural Network Field Theory”, arXiv:2604.27050 [hep-th]

work page internal anchor Pith review Pith/arXiv arXiv
[19]

Neural network representa- tion of quantum systems

K. Hashimoto, Y. Hirono, J. Maeda & J. Totsuka-Yoshinaka,“Neural network representa- tion of quantum systems”, Mach. Learn. Sci. Tech.5, 045039 (2024), arXiv:2403.11420 [hep-th]

work page arXiv 2024
[20]

Sen and V

S. Sen & V. Vaidya,“Viability of perturbative expansion for quantum field theories on neurons”,arXiv:2508.03810 [hep-th]. 54

work page arXiv
[21]

Non-perturbative renormalization for the neural network-QFT correspondence

H. Erbin, V. Lahoche & D. O. Samary,“Non-perturbative renormalization for the neural network-QFT correspondence”, Mach. Learn. Sci. Tech.3, 015027 (2022), arXiv:2108.01403 [hep-th]

work page arXiv 2022
[22]

Halverson,TASI Lectures on Physics for Machine Learning,2408.00082

J. Halverson,“TASI Lectures on Physics for Machine Learning”, arXiv:2408.00082 [hep-th]

work page arXiv
[23]

Maiti, K

A. Maiti, K. Stoner & J. Halverson,“Symmetry-via-Duality: Invariant Neural Network Densities from Parameter-Space Correlators”,arXiv:2106.00694 [cs.LG]

work page arXiv
[24]

Ferko, J

C. Ferko, J. Halverson & A. Mutchler,“Universality of Neural Network Field Theory”, arXiv:2601.14453 [hep-th]

work page arXiv
[25]

Random Features for Large-Scale Kernel Machines

A. Rahimi & B. Recht,“Random Features for Large-Scale Kernel Machines”, in“Advances in Neural Information Processing Systems 20 (NeurIPS)”, p. 1177–1184

work page
[26]

Neural Network Quantum Field Theory from Transformer Architectures

D. S. Ageev & Y. A. Ageeva,“Neural Network Quantum Field Theory from Transformer Architectures”,arXiv:2602.10209 [cs.LG]

work page arXiv
[27]

Attention is All You Need

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser & I. Polosukhin,“Attention is All You Need”, in“Advances in Neural Information Processing Systems”, p. 5998–6008

work page
[28]

Symmetry Breaking in Transformers for Efficient and Interpretable Training

E. Silverstein, D. Kunin & V. Shyam,“Symmetry Breaking in Transformers for Efficient and Interpretable Training”,arXiv:2601.22257 [cs.LG]

work page arXiv
[29]

Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion

B. Zhang, Z. Zheng, Z. Chen & J. Li,“Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion”,arXiv:2502.00264 [cs.LG]

work page arXiv
[30]

Halverson,Building Quantum Field Theories Out of Neurons,2112.04527

J. Halverson,“Building Quantum Field Theories Out of Neurons”, arXiv:2112.04527 [hep-th]

work page arXiv
[31]

Demirtas, J

M. Demirtas, J. Halverson, A. Maiti, M. D. Schwartz & K. Stoner,“Neural Network Field Theories: Non-Gaussianity, Actions, and Locality”,arXiv:2307.03223 [hep-th]

work page arXiv
[32]

Conformal invariance, supersymmetry and string theory

D. Friedan, E. J. Martinec & S. H. Shenker,“Conformal invariance, supersymmetry and string theory”, Nucl. Phys. B271, 93 (1986)

work page 1986
[33]

Note on the Bondi-Metzner-Sachs Group

E. T. Newman & R. Penrose,“Note on the Bondi-Metzner-Sachs Group”, J. Math. Phys. 7, 863 (1966)

work page 1966
[34]

Spin-sSpherical Harmonics andð

J. N. Goldberg, A. J. Macfarlane, E. T. Newman, F. Rohrlich & E. C. G. Sudarshan, “Spin-sSpherical Harmonics andð”, J. Math. Phys.8, 2155 (1967)

work page 1967
[35]

Frank, J

S. Frank, J. Halverson, A. Maiti & F. Ruehle,“Fermions and Supersymmetry in Neural Network Field Theories”,arXiv:2511.16741 [hep-th]

work page arXiv
[36]

Frank and J

S. Frank & J. Halverson,“String Theory from Infinite Width Neural Networks”, arXiv:2601.06249 [hep-th]

work page arXiv
[37]

Generalized Global Symmetries

D. Gaiotto, A. Kapustin, N. Seiberg & B. Willett,“Generalized Global Symmetries”, JHEP 1502, 172 (2015),arXiv:1412.5148 [hep-th]

work page internal anchor Pith review Pith/arXiv arXiv 2015
[38]

Notes on generalized global symmetries in QFT

E. Sharpe,“Notes on generalized global symmetries in QFT”, Fortsch. Phys.63, 659 (2015), arXiv:1508.04770 [hep-th]. 55

work page arXiv 2015
[39]

An introduction to higher-form symmetries

P. R. S. Gomes,“An introduction to higher-form symmetries”, SciPost Phys. Lect. Notes 74, 1 (2023),arXiv:2303.01817 [hep-th]

work page arXiv 2023
[40]

Schafer-Nameki, ICTP lectures on (non-)invertible general- ized symmetries, Phys

S. Schafer-Nameki,“ICTP lectures on (non-)invertible generalized symmetries”, Phys. Rept. 1063, 1 (2024),arXiv:2305.18296 [hep-th]

work page arXiv 2024
[41]

Introduction to Generalized Global Symmetries in QFT and Particle Physics

T. D. Brennan & S. Hong,“Introduction to Generalized Global Symmetries in QFT and Particle Physics”,arXiv:2306.00912 [hep-ph]

work page arXiv
[42]

Bhardwaj, L

L. Bhardwaj, L. E. Bottini, L. Fraser-Taliente, L. Gladden, D. S. W. Gould, A. Platschorre & H. Tillim,“Lectures on generalized symmetries”, Phys. Rept.1051, 1 (2024), arXiv:2307.07547 [hep-th]

work page arXiv 2024
[43]

What’s Done Cannot Be Undone: TASI Lectures on Non-Invertible Sym- metries

S.-H. Shao,“What’s Done Cannot Be Undone: TASI Lectures on Non-Invertible Sym- metries”, in“Theoretical Advanced Study Institute in Elementary Particle Physics 2023: Aspects of Symmetry”,arXiv:2308.00747 [hep-th]

work page arXiv 2023
[44]

Naturalness, chiral symmetry, and spontaneous chiral symmetry breaking

G. ’t Hooft,“Naturalness, chiral symmetry, and spontaneous chiral symmetry breaking”, NATO Sci. Ser. B59, 135 (1980)

work page 1980
[45]

Introduction to Analytic Number Theory

T. M. Apostol,“Introduction to Analytic Number Theory”, Springer-Verlag (1976), New York. 56

work page 1976