arxiv: 2605.06134 · v1 · submitted 2026-05-07 · ✦ hep-lat · cs.LG

Recognition: unknown

Diffusion model for SU(N) gauge theories

Javad Komijani, Lara Turgut, Marina K. Marinkovic

Pith reviewed 2026-05-08 03:24 UTC · model grok-4.3

classification ✦ hep-lat cs.LG

keywords diffusion modelslattice gauge theorySU(3)Wilson actionscore matchingHybrid Monte Carloconfiguration sampling

0 comments

The pith

Diffusion models trained with score matching can generate SU(3) gauge configurations that match the Wilson action distribution.

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

The paper introduces a diffusion-model framework based on implicit score matching to produce lattice gauge-field samples for SU(N) theories. It trains the model on Hybrid Monte Carlo data for the Wilson action and tests the generated configurations in two and four dimensions by direct comparison with independent HMC runs. The central aim is to create an alternative sampling method that avoids the repeated cost of running full HMC chains for every new ensemble. If the learned reverse process reproduces the target Boltzmann weight, the approach supplies a new tool for lattice QCD simulations where configuration generation is a major bottleneck.

Core claim

We develop a score-matching framework for SU(N) lattice gauge theories that can be extended to other Lie groups. Applied to SU(3) with the Wilson gauge action in two and four dimensions, the diffusion models are trained on HMC-generated data and assessed by direct comparison with HMC simulations. For large values of the inverse coupling, accurate reverse-time integration requires predictor-corrector schemes, for which we introduce a corrector based on Hamiltonian molecular dynamics. While the corrector improves sampling quality, it also raises computational cost; several strategies for improving efficiency are outlined.

What carries the argument

Implicit score matching applied to diffusion models on SU(N) lattice gauge fields, augmented by a Hamiltonian molecular dynamics corrector for reverse-time sampling at strong coupling.

If this is right

Diffusion models supply an alternative to repeated HMC runs for producing gauge configurations.
Predictor-corrector integration with a Hamiltonian molecular dynamics corrector restores sampling accuracy at large inverse coupling.
The same score-matching construction extends in principle to other Lie groups and actions.
Outlined efficiency improvements can reduce the extra cost introduced by the corrector step.

Where Pith is reading between the lines

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

Once trained, the model could in principle be reused across multiple volumes without retraining, lowering the cost per configuration for large-scale studies.
The approach might be combined with existing multigrid or Fourier acceleration techniques to further speed up the reverse diffusion process.
Extension to dynamical fermions would require learning a score that also incorporates the fermion determinant, opening a route to full QCD sampling.

Load-bearing premise

The score function learned from a finite set of HMC configurations accurately encodes the probability gradient of the Wilson action for any lattice volume and coupling value.

What would settle it

Generate an ensemble at a new lattice size or coupling value with the trained diffusion model and compare the plaquette expectation value or Wilson-loop averages to an independent HMC run at the same parameters; statistically significant disagreement falsifies the claim.

Figures

Figures reproduced from arXiv: 2605.06134 by Javad Komijani, Lara Turgut, Marina K. Marinkovic.

**Figure 1.** Figure 1: FIG. 1: The evolution of the eigenangles of SU( view at source ↗

**Figure 2.** Figure 2: FIG. 2: Application of diffusion-based sampling to SU(2) and SU(3) matrix models. Shown are view at source ↗

**Figure 3.** Figure 3: FIG. 3 view at source ↗

read the original abstract

Implicit score matching provides a computationally efficient approach for training diffusion models and generating high-quality samples from complex distributions. In this work, we develop a score-matching framework for SU(N) lattice gauge theories, which can be extended to other Lie groups. We apply the method to SU(3) gauge configurations with the Wilson gauge action in two and four dimensions and assess the quality of the generated samples by comparison with Hybrid Monte Carlo (HMC) simulations. We show that the diffusion models can be successfully trained and applied for sampling the Wilson gauge action. For large values of inverse coupling, accurate reverse-time integration requires predictor-corrector schemes, for which we introduce a corrector based on Hamiltonian molecular dynamics. While the corrector significantly improves sampling quality, it also increases the computational cost. We outline several strategies for improving sampling efficiency.

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

Diffusion models for SU(3) gauge fields work after training on HMC data but need an HMC corrector to match the target distribution at large beta.

read the letter

The paper trains a diffusion model via implicit score matching on SU(3) lattice configurations generated by HMC, then uses the learned score to draw new samples for the Wilson action in two and four dimensions. They add a predictor-corrector step where the corrector runs short Hamiltonian molecular dynamics trajectories to clean up the output at strong coupling. That combination is the concrete new piece: prior diffusion work in physics did not target Lie-group valued fields with exact gauge symmetry preserved during sampling. The authors show the model can be trained and that the generated fields look reasonable when compared to HMC, which is a useful first demonstration. The HMC corrector improves quality but raises the cost, and they sketch a few ideas for cutting that overhead later. The main limitation is that the pure reverse diffusion process does not reproduce the Boltzmann weight accurately enough on its own once beta gets large; the paper is explicit that the corrector is required. Without the full set of quantitative metrics, error bars on observables, or scaling tests with volume and beta, it is hard to judge how close the samples actually get or how much the method saves over plain HMC. The training data come from HMC, so the claim reduces to whether the score network plus corrector can replace the expensive part of the chain. This is the sort of paper lattice gauge theorists working on new sampling algorithms should see. It is not yet a drop-in replacement, but the framework is worth checking in detail. I would send it to peer review so referees can look at the implementation, the gauge-invariance handling, and the actual cost-quality numbers.

Referee Report

3 major / 2 minor

Summary. The paper develops a score-matching framework for training diffusion models on SU(N) lattice gauge theories with the Wilson action. It applies the method to SU(3) configurations in two and four dimensions, compares the generated samples to Hybrid Monte Carlo (HMC) simulations, and introduces a Hamiltonian molecular dynamics (HMD) corrector within a predictor-corrector scheme to improve reverse-time integration at large inverse coupling beta. The authors claim successful training and sampling, while outlining strategies to enhance efficiency.

Significance. If the diffusion model (with or without the HMD corrector) produces samples whose distribution exactly matches the Boltzmann weight of the Wilson action, and if quantitative benchmarks demonstrate competitive or superior efficiency, this would constitute a novel machine-learning approach to configuration generation in lattice gauge theory. The framework's extensibility to other Lie groups and the explicit discussion of computational trade-offs are strengths. However, the reported need for an HMC-derived corrector at physically relevant couplings limits the claim of a standalone diffusion-based sampler.

major comments (3)

[Abstract] Abstract: The central claim that 'the diffusion models can be successfully trained and applied for sampling the Wilson gauge action' is undermined by the statement that 'accurate reverse-time integration at large inverse coupling requires predictor-corrector schemes' whose corrector is Hamiltonian molecular dynamics. This indicates that the base reverse diffusion process (trained via implicit score matching) does not reproduce the target distribution in the strong-coupling regime without additional HMC steps, making the method hybrid rather than purely diffusion-based.
[Abstract] Abstract and methods description: No details are supplied on how gauge invariance is enforced during the diffusion process, score estimation, or sampling steps. For SU(N) gauge theories this is load-bearing, as any violation would render the generated configurations unphysical; the manuscript must include an explicit mechanism (e.g., projection or invariant parameterization) with supporting equations.
[Results] Results/comparison section: The quality assessment against HMC lacks reported quantitative metrics such as plaquette expectation values with statistical errors, autocorrelation times, or distribution distances; without these, the assertion of 'high-quality samples' cannot be verified and the comparison remains qualitative.

minor comments (2)

[Abstract] Abstract: Consider adding one or two concrete numerical results (e.g., plaquette values or acceptance rates) to substantiate the comparison with HMC.
The outline of efficiency-improvement strategies is useful but would benefit from a brief table comparing wall-clock costs or effective sample sizes with and without the corrector.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments have prompted us to improve the clarity of our claims, add missing technical details, and strengthen the quantitative evidence in the manuscript. We address each major comment below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that 'the diffusion models can be successfully trained and applied for sampling the Wilson gauge action' is undermined by the statement that 'accurate reverse-time integration at large inverse coupling requires predictor-corrector schemes' whose corrector is Hamiltonian molecular dynamics. This indicates that the base reverse diffusion process (trained via implicit score matching) does not reproduce the target distribution in the strong-coupling regime without additional HMC steps, making the method hybrid rather than purely diffusion-based.

Authors: We appreciate the referee highlighting the potential ambiguity in the abstract. The diffusion model is trained exclusively via implicit score matching on the Wilson action, and the forward/reverse processes are defined on the SU(N) manifold. At moderate and small β the reverse diffusion alone produces samples whose observables agree with HMC within statistical errors. At large β the numerical integration of the reverse SDE benefits from a small number of HMD corrector steps to reduce discretization error and ensure exact matching to the target measure. This corrector is not a replacement for the diffusion sampler but a refinement within the predictor-corrector integrator; it does not involve full HMC trajectories or Metropolis accept/reject. We will revise the abstract to state explicitly that the diffusion model forms the core sampler and that the HMD corrector is an optional, low-cost enhancement used only when higher precision is required at large β. We also note that large β corresponds to the weak-coupling regime, not the strong-coupling regime. revision: partial
Referee: [Abstract] Abstract and methods description: No details are supplied on how gauge invariance is enforced during the diffusion process, score estimation, or sampling steps. For SU(N) gauge theories this is load-bearing, as any violation would render the generated configurations unphysical; the manuscript must include an explicit mechanism (e.g., projection or invariant parameterization) with supporting equations.

Authors: We agree that an explicit description of gauge invariance is necessary. In the current implementation the diffusion is performed using a Lie-algebra parameterization: each update is generated via the exponential map from su(N) to SU(N), and after every diffusion step the resulting matrix is projected back onto the SU(N) manifold by polar decomposition (or QR-based retraction) to restore unitarity and determinant one. The score network is constructed to be gauge-equivariant by operating on plaquette-based or link-based invariant features. We will insert a new subsection in the Methods section containing the precise update rule, the projection operator, and the corresponding equations. revision: yes
Referee: [Results] Results/comparison section: The quality assessment against HMC lacks reported quantitative metrics such as plaquette expectation values with statistical errors, autocorrelation times, or distribution distances; without these, the assertion of 'high-quality samples' cannot be verified and the comparison remains qualitative.

Authors: We concur that quantitative benchmarks are required for a rigorous comparison. In the revised manuscript we will add tables and figures reporting: (i) plaquette expectation values with jackknife errors for both diffusion-generated and HMC ensembles at each β and volume studied; (ii) integrated autocorrelation times for the plaquette and for the topological charge; and (iii) a distribution-distance measure (e.g., sliced Wasserstein distance on the plaquette histogram or on a set of Wilson loops). These additions will allow direct verification of the sample quality. revision: yes

Circularity Check

0 steps flagged

No circularity: model trained on external HMC data and validated by independent comparison

full rationale

The paper generates training configurations with standard HMC, trains a score-based diffusion model via implicit score matching, then draws new samples and compares observables (e.g., plaquette, Polyakov loop) to separate HMC runs. This is ordinary generative modeling with external benchmark validation; no equation, ansatz, or self-citation reduces the claimed sampling distribution to the training data by construction. The predictor-corrector step that adds an HMC corrector is presented as an optional improvement for large β, not as a hidden premise that forces the base result.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The work rests on standard assumptions of diffusion models and lattice gauge theory; no new free parameters or invented entities are introduced in the abstract.

axioms (2)

domain assumption The target probability distribution is given by the Boltzmann weight of the Wilson gauge action.
Invoked when the diffusion model is trained to sample that distribution.
standard math Implicit score matching yields an unbiased estimator of the score for the chosen noise schedule.
Underlying the training procedure described in the abstract.

pith-pipeline@v0.9.0 · 5437 in / 1314 out tokens · 31878 ms · 2026-05-08T03:24:44.218464+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages · 4 internal anchors

[1]

The following discussion can be also extended to other Lie groups

Diffusion process and Fokker-Planck equation We consider a diffusion process acting on SU(N) gauge link variables on a lattice. The following discussion can be also extended to other Lie groups. For each lattice link variable located at sitexin directionµ, we denote its value at (fictitious) diffusion timetas Ut ≡U(t, x, µ).(B1) To simplify notation, we s...
[2]

(B2) is reversible in the sense that evolving the random field Ut1 forward in time fromt 1 tot 2 using Eq

Reverse Process The process defined in Eq. (B2) is reversible in the sense that evolving the random field Ut1 forward in time fromt 1 tot 2 using Eq. (B2), and subsequently evolving it backward fromt 2 tot 1, leaves the probability density function (PDF) ofU t1 unchanged. Equivalently, the forward- and reverse-time processes must satisfy the same Fokker–P...
[3]

Duane, A

S. Duane, A. D. Kennedy, B. J. Pendleton, and D. Roweth, Phys. Lett. B195, 216 (1987)

1987
[4]

Q. Zhu, G. Aarts, W. Wang, K. Zhou, and L. Wang, (2025), arXiv:2502.05504 [hep-lat]

work page arXiv 2025
[5]

Aarts, D

G. Aarts, D. E. Habibi, A. Ipp, D. I. M¨ uller, T. R. Ranner, L. Wang, W. Wang, and Q. Zhu, (2026), arXiv:2601.19552 [hep-lat]

work page arXiv 2026
[6]

Kanwar, M

G. Kanwar, M. S. Albergo, D. Boyda, K. Cranmer, D. C. Hackett, S. Racani` ere, D. J. Rezende, and P. E. Shanahan, Phys. Rev. Lett.125, 121601 (2020), arXiv:2003.06413

work page arXiv 2020
[7]

Boyda, G

D. Boyda, G. Kanwar, S. Racani` ere, D. J. Rezende, M. S. Albergo, K. Cranmer, D. C. Hackett, and P. E. Shanahan, Phys. Rev. D103, 074504 (2021), arXiv:2008.05456

work page arXiv 2021
[8]

Kanwar,Flow-based sampling for lattice field theories,PoSLATTICE2023(2024) 114 [2401.01297]

G. Kanwar, in40th International Symposium on Lattice Field Theory(2024) arXiv:2401.01297 [hep-lat]

work page arXiv 2024
[9]

Bacchio, P

S. Bacchio, P. Kessel, S. Schaefer, and L. Vaitl, Phys. Rev. D107, L051504 (2023), arXiv:2212.08469 [hep-lat]

work page arXiv 2023
[10]

Deep Unsupervised Learning using Nonequilibrium Thermodynamics

J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, CoRR abs/1503.03585(2015), 1503.03585

work page internal anchor Pith review arXiv 2015
[11]

Song and S

Y. Song and S. Ermon, CoRRabs/1907.05600(2019), 1907.05600

work page arXiv 1907
[12]

J. Ho, A. Jain, and P. Abbeel, CoRRabs/2006.11239(2020), 2006.11239

work page internal anchor Pith review arXiv 2006
[13]

O. Vega, J. Komijani, A. El-Khadra, and M. Marinkovic, (2025), arXiv:2510.26081

work page arXiv 2025
[14]

L. Wang, G. Aarts, and K. Zhou, JHEP05, 060 (2024), arXiv:2309.17082 [hep-lat]

work page arXiv 2024
[15]

Aarts, D

G. Aarts, D. E. Habibi, L. Wang, and K. Zhou, JHEP12, 160 (2025), arXiv:2510.01328

work page arXiv 2025
[16]

S. A. Gottlieb, W. Liu, D. Toussaint, R. L. Renken, and R. L. Sugar, Phys. Rev. D35, 2531 (1987)

1987
[17]

R. M. Neal,Handbook of Markov Chain Monte Carlo(2011) arXiv:1206.1901 [stat.CO]

work page arXiv 2011
[18]

U-Net: Convolutional Networks for Biomedical Image Segmentation

O. Ronneberger, P. Fischer, and T. Brox, CoRRabs/1505.04597(2015), 1505.04597

work page internal anchor Pith review arXiv 2015
[19]

normflow: Normalizing flow for generating lattice field configurations,

“normflow: Normalizing flow for generating lattice field configurations,”https://github. com/jkomijani/normflow(2026)

2026
[20]

Komijani and M

J. Komijani and M. K. Marinkovic, PoSLATTICE2024, 050 (2025), arXiv:2501.18288

work page arXiv 2025
[21]

Hyv¨ arinen and P

A. Hyv¨ arinen and P. Dayan, Journal of Machine Learning Research6(2005)

2005
[22]

Vincent, Neural Computation23, 1661 (2011)

P. Vincent, Neural Computation23, 1661 (2011)

2011
[23]

Favoni, A

M. Favoni, A. Ipp, D. I. M¨ uller, and D. Schuh, Phys. Rev. Lett.128, 032003 (2022), arXiv:2012.12901 [hep-lat]

work page arXiv 2022
[24]

Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, CoRR abs/2011.13456(2020), 2011.13456

work page internal anchor Pith review arXiv 2011
[25]

Parisi, Nucl

G. Parisi, Nucl. Phys. B180, 378 (1981). 23

1981
[26]

C. M. Bender, A. Fring, and J. Komijani, J. Phys. A: Math. Theor.47, 235204 (2014), arXiv:1401.6161 [math-ph]

work page arXiv 2014
[27]

Komijani, J

J. Komijani, J. Phys. A: Math. Theor.54, 465202 (2021), arXiv:2107.02475 [math-ph]

work page arXiv 2021
[28]

I. T. Drummond, S. Duane, and R. R. Horgan, Nucl. Phys. B220, 119 (1983)

1983