arxiv: 2604.07226 · v1 · submitted 2026-04-08 · ✦ hep-lat

Recognition: unknown

Neural network interpolators for Wilson loops

Julian Mayer-Steudte

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:30 UTC · model grok-4.3

classification ✦ hep-lat

keywords neural networksWilson loopslattice QCDstatic quark potentialexcited statesgauge-equivariantquenched approximationinterpolators

0 comments

The pith

Neural networks parametrize trial states to automatically obtain ground and excited state interpolators for Wilson loops in quenched lattice QCD.

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

The paper presents a neural network parametrization of trial states for Wilson loops, built from gauge-equivariant layers and trained via a loss function that favors overlaps with the ground and excited states. This targets the poor signal-to-noise ratio that arises when using straight Wilson lines at large Euclidean times for extracting the static quark-antiquark potential. The optimization is designed so that, in the quenched theory, the network directly produces the interpolators without manual construction of specific shapes. A sympathetic reader would care because the method could reduce reliance on ad-hoc smearing techniques and simplify access to excited states.

Core claim

By representing trial states as a neural network with gauge-equivariant layers and minimizing a loss function that promotes ground and excited state overlaps, the procedure automatically yields the corresponding interpolators for the static quark-antiquark potential in the quenched approximation.

What carries the argument

Neural-network parametrization of trial states with gauge-equivariant layers, optimized by a loss function that favors ground and excited state overlaps.

If this is right

Improved ground-state overlap reduces the noise in Wilson-loop measurements at large times.
Excited-state interpolators emerge automatically without hand-crafted shapes carrying specific quantum numbers.
The variational basis for the static potential is generated directly from the trained network.

Where Pith is reading between the lines

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

The same network architecture could be tested in dynamical QCD where smearing choices are more labor-intensive.
The loss function might be augmented with symmetry constraints to accelerate convergence to physical states.
This variational approach could be combined with existing multilevel algorithms to further suppress noise.

Load-bearing premise

Minimizing the chosen loss function over the network parameters will converge to physically meaningful interpolators rather than optimization artifacts.

What would settle it

Extracting the ground and first excited state energies from the neural-network interpolators on the same quenched ensembles and checking whether they match the values obtained from conventional smearing methods within statistical errors.

Figures

Figures reproduced from arXiv: 2604.07226 by Julian Mayer-Steudte.

**Figure 2.** Figure 2: The effective masses for 𝑟 = 3 and 𝑟 = 7 for the neural network with 𝑁 (𝑛hidden ) = 18 . weights, which cause unpredictable fluctuations leading to not-a-number operations. Additionally, we find that the neural network with 𝑁 (𝑛hidden ) = 18 performs slightly better than the smaller neural networks. Consequently, we conclude that while more hidden layers can enhance performance, they may also introduce tec… view at source ↗

**Figure 3.** Figure 3: The static energies as a function of𝑟 for the first three states. The left plot shows the static energies, the right plot the energy difference with respect to the ground state 𝑛 = 0. quantity is divergent in the continuum limit, the energy difference is a physical quantity and can be extrapolated to the continuum. We find that the ground state (𝑛 = 0) recovers the well-known static quark-antiquark potenti… view at source ↗

read the original abstract

The extraction of the static quark-antiquark potential from lattice QCD suffers from the poor signal-to-noise ratio of Wilson loops at large Euclidean times. To overcome this, smearing methods or the Coulomb gauge are used to improve the ground-state overlap with respect to the straight Wilson line trial state within the Wilson loop. To find excited states, complicated shapes are introduced to generate specific quantum numbers. Here, we introduce a neural-network parametrization of trial states, constructed with gauge-equivariant layers and optimized with a loss function that favors ground and excited states. In the quenched theory, we automatically obtain the interpolators for the ground and excited states.

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

The neural-network parametrization of Wilson-loop trial states is a fresh technical idea but the manuscript gives no solid evidence that the loss actually yields physical ground and excited states rather than optimizer artifacts.

read the letter

The main takeaway is that this paper puts forward a gauge-equivariant neural network to generate trial states for Wilson loops, trained with a loss aimed at both ground and excited states. That combination is not just a repackaging of smearing or Coulomb-gauge tricks, so the parametrization itself counts as new. The equivariant layers are a reasonable choice to keep the operators gauge-invariant, and the quenched-theory setup is straightforward enough to follow. If the method scales, it could cut down on the manual design of interpolators for potential calculations and spectroscopy. The paper does a clean job describing the architecture and the loss construction. The soft spots are more noticeable. The central claim that the procedure automatically produces the physical interpolators rests on the unproven assumption that the loss minimum aligns with the true eigenstates of the transfer matrix. Nothing in the text supplies an orthogonality constraint, a spectral argument, or even a numerical check that the resulting operators have maximal overlap with the ground and first excited states. There are also no direct comparisons to standard methods, no reported overlap factors with uncertainties, and no systematic study of how the results change with network depth or loss weights. The full results appear mostly qualitative. This work is aimed at lattice QCD people who already spend time tuning Wilson-loop operators. A reader in that niche might want to look at the code or try the loss on their own ensembles, but the current evidence is too thin for broader use. I would bring it to a reading group as maybe, to talk through the loss design. I would not cite it in the next year. It still deserves peer review because the technical setup is coherent and the idea is original; referees can ask for the missing benchmarks and convergence checks.

Referee Report

2 major / 2 minor

Summary. The paper proposes a gauge-equivariant neural-network parametrization of trial states for Wilson loops in lattice QCD. These states are optimized via a custom loss function designed to favor maximal overlap with the ground and excited states of the static quark-antiquark system. In the quenched approximation the authors claim that this procedure automatically yields the physical interpolators without manual construction of shapes or smearing.

Significance. If the central claim holds, the method would supply a systematic, symmetry-preserving route to high-overlap operators for both ground-state potential extraction and excited-state spectroscopy, potentially reducing reliance on hand-crafted trial states. The gauge-equivariant architecture is a clear technical asset that maintains the correct transformation properties under gauge transformations.

major comments (2)

[Abstract] Abstract: the assertion that the optimization 'automatically obtain[s] the interpolators for the ground and excited states' is load-bearing yet unsupported by any derivation, overlap factors, effective-mass plateaus, or comparison to existing smearing/Coulomb-gauge methods. No quantitative results, error analysis, or loss-function definition appear, so it is impossible to verify that the global minimum coincides with the true transfer-matrix eigenstates rather than linear combinations or gauge artifacts.
[Method / Loss function] Loss-function description (wherever presented): the manuscript supplies no explicit orthogonality constraint, Gram-Schmidt step, or variational penalty that would guarantee the network outputs are eigenstates. Without such a mechanism the skeptic's concern is valid: minimization can converge to non-physical configurations that merely extremize the chosen loss.

minor comments (2)

[Abstract] The abstract mentions 'a suitably designed loss function' but gives no functional form, weighting parameters, or training details; these must be supplied with explicit equations for reproducibility.
[Method] No mention of the neural-network architecture depth, width, or activation functions is provided; these parameters are listed as free in the axiom ledger and should be documented.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading of our manuscript and the constructive feedback. The comments highlight important points about clarity and supporting evidence that we will address in a revised version. We respond to each major comment below.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that the optimization 'automatically obtain[s] the interpolators for the ground and excited states' is load-bearing yet unsupported by any derivation, overlap factors, effective-mass plateaus, or comparison to existing smearing/Coulomb-gauge methods. No quantitative results, error analysis, or loss-function definition appear, so it is impossible to verify that the global minimum coincides with the true transfer-matrix eigenstates rather than linear combinations or gauge artifacts.

Authors: We acknowledge that the abstract is brief and the central claim requires explicit support. The full manuscript contains numerical results in the quenched theory, including effective-mass plateaus for the ground-state potential and first excited state, overlap factors extracted from correlators, and direct comparisons to unsmeared Wilson lines as well as standard smearing techniques. These demonstrate that the optimized states achieve higher ground-state overlap and correctly reproduce known spectroscopy. We will revise the abstract to include a concise reference to these quantitative findings and error analysis. In the methods section we will add an explicit definition of the loss function together with a short derivation showing why its global minimum aligns with transfer-matrix eigenstates (leveraging the gauge-equivariant architecture and the structure of the quenched theory to suppress gauge artifacts and linear combinations). revision: yes
Referee: [Method / Loss function] Loss-function description (wherever presented): the manuscript supplies no explicit orthogonality constraint, Gram-Schmidt step, or variational penalty that would guarantee the network outputs are eigenstates. Without such a mechanism the skeptic's concern is valid: minimization can converge to non-physical configurations that merely extremize the chosen loss.

Authors: The custom loss function contains an explicit variational penalty that enforces approximate orthogonality between the network outputs for the ground and excited states. This is implemented via cross-correlation terms that penalize non-zero overlaps between distinct trial states while maximizing the large-time Wilson-loop correlators. No post-hoc Gram-Schmidt orthogonalization is required because the penalty is active throughout training. We will insert the full mathematical expression of the loss into the revised manuscript and add a brief numerical check confirming that the converged states are orthogonal to within statistical errors and match the expected quantum numbers, thereby ruling out convergence to non-physical extrema. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is independent parametrization

full rationale

The paper introduces a gauge-equivariant neural-network parametrization of trial states for Wilson loops, optimized via a loss function designed to favor ground and excited states. The central claim that this yields the physical interpolators in the quenched theory rests on the convergence of the optimization process rather than any redefinition of the target states in terms of the network outputs or fitted parameters. No equations or steps in the provided abstract reduce the result to a tautological input by construction, nor is there evidence of self-citation load-bearing, uniqueness imported from authors, or renaming of known results. The derivation chain remains self-contained as a proposed computational method whose validity depends on empirical performance rather than circular equivalence to its inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method assumes standard neural-network optimization succeeds in this gauge-theory setting and that gauge-equivariant layers can be constructed without breaking lattice symmetries.

free parameters (1)

Neural-network architecture and loss weights
Specific layer sizes, connectivity, and relative weighting of ground versus excited-state terms in the loss are chosen by the authors and not derived from first principles.

axioms (1)

domain assumption Gauge-equivariant layers preserve the required lattice symmetries
Invoked when the paper states the layers are gauge-equivariant; this must hold for the trial states to be valid interpolators.

pith-pipeline@v0.9.0 · 5387 in / 1164 out tokens · 92817 ms · 2026-05-10T17:30:15.711353+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages · 2 internal anchors

[1]

Wilson,Confinement of quarks,Phys

K.G. Wilson,Confinement of quarks,Phys. Rev. D10(1974) 2445

1974
[2]

Brambilla, A

N. Brambilla, A. Pineda, J. Soto and A. Vairo,Potential NRQCD: An effective theory for heavy quarkonium,Nucl. Phys. B566(2000) 275 [hep-ph/9907240]

work page arXiv 2000
[3]

Brambilla, A

N. Brambilla, A. Pineda, J. Soto and A. Vairo,Effective field theories for heavy quarkonium, Rev. Mod. Phys.77(2005) 1423 [hep-ph/0410047]

work page arXiv 2005
[4]

Hybrids, tetraquarks, pen- taquarks, doubly heavy baryons, and quarkonia in Born- Oppenheimer effective theory,

M. Berwein, N. Brambilla, A. Mohapatra and A. Vairo,Hybrids, tetraquarks, pentaquarks, doubly heavy baryons, and quarkonia in Born-Oppenheimer effective theory,Phys. Rev. D 110(2024) 094040 [2408.04719]. [5]TUMQCDcollaboration,Static energy in (2+1+1)-flavor lattice QCD: Scale setting and charm effects,Phys. Rev. D107(2023) 074503 [2206.03156]. [6]APEcoll...

work page arXiv 2024
[5]

Wilson loops with neural networks

V. Bellscheidt, N. Brambilla, A.S. Kronfeld and J. Mayer-Steudte,Wilson loops with neural networks,2602.02436

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Capitani, O

S. Capitani, O. Philipsen, C. Reisinger, C. Riehl and M. Wagner,Precision computation of hybrid static potentials in SU(3) lattice gauge theory,Phys. Rev. D99(2019) 034502 [1811.11046]

work page arXiv 2019
[7]

Schlosser and M

C. Schlosser and M. Wagner,Hybrid static potentials in SU(3) lattice gauge theory at small quark-antiquark separations,Phys. Rev. D105(2022) 054503 [2111.00741]

work page arXiv 2022
[8]

Favoni, A

M. Favoni, A. Ipp, D.I. Müller and D. Schuh,Lattice gauge equivariant convolutional neural networks,Phys. Rev. Lett.128(2022) 032003 [2012.12901]

work page arXiv 2022
[9]

Nagai and A

Y. Nagai and A. Tomiya,Gauge covariant neural network for quarks and gluons,Phys. Rev. D111(2025) 074501 [2103.11965]

work page arXiv 2025
[10]

Lehner and T

C. Lehner and T. Wettig,Gauge-equivariant neural networks as preconditioners in lattice QCD,Phys. Rev. D108(2023) 034503 [2302.05419]

work page arXiv 2023
[11]

Michael and I

C. Michael and I. Teasdale,Extracting glueball masses from lattice QCD,Nucl. Phys. B215 (1983) 433

1983
[12]

Kronfeld,Improved methods for computing masses from numerical simulations,Nucl

A.S. Kronfeld,Improved methods for computing masses from numerical simulations,Nucl. Phys. B Proc. Suppl.17(1990) 313

1990
[13]

Lüscher and U

M. Lüscher and U. Wolff,How to calculate the elastic scattering matrix in two-dimensional quantum field theories by numerical simulation,Nucl. Phys. B339(1990) 222

1990
[14]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter,Decoupled weight decay regularization, in2019 International Conference on Learning Representations, T. Sainath et al., eds., 2019, https://openreview.net/forum?id=Bkg6RiCqY7 [1711.05101]

work page internal anchor Pith review Pith/arXiv arXiv 2019
[15]

Lüscher and P

M. Lüscher and P. Weisz,Locality and exponential error reduction in numerical lattice gauge theory,JHEP09(2001) 010 [hep-lat/0108014]

work page arXiv 2001
[16]

Jay and E.T

W.I. Jay and E.T. Neil,Bayesian model averaging for analysis of lattice field theory results, Phys. Rev. D103(2021) 114502 [2008.01069]. 10

work page arXiv 2021