Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory: Analytical Foundations and Numerical Scaling Tests

arXiv:2611.03530 · hep-lat · Submitted 2026-11-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Lattice. The day in lattice field theory: what the simulations put on the table.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory".

Mira: Here's the gist of the paper. Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories Here we focus on the underlying analytical framework The paper presents a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discusses the gauge invariance of the denoising loss function, and provides extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling Analytical Foundations The analytical framework begins by defining the lattice approximation of non-abelian Yang-Mills theory using a hypercubic lattice where the Euclidean Yang-Mills action is replaced with the Wilson action The continuum limit is approached by taking the lattice spacing to zero, which is achieved by letting L → ∞ and β → ∞ while keeping physical volume fixed Diffusion Processes on Non-Abelian Lie Groups The group-preserving diffusion process is defined via the Stratonovich stochastic differential equation (SDE) This SDE allows for multiple sampling methods, including the reverse SDE, which can be used to sample configurations independent and uncorrelated The Fokker-Planck equation (FPE) associated with this diffusion process is derived as d t p t(U) = −Da

Ka(U)pt(U): + 1/2 g 2(t)∆ˆ pt(U), where ∆ˆ is the Laplace-Beltrami operator Learning the Diffusion Process The reverse diffusion process requires knowledge of the derivative of the logarithm of the time-dependent probability density, which is approximated by a score function s aθ;x,µ(U, t) The optimal score function is identified as the negative gradient of a time-dependent action St(U), and it must be parametrized by a gauge-equivariant neural network The loss function used is denoising score matching (DSM) Where the Gains Come From The gains come from several key aspects of the method: 1.

Kai: Exact symmetries are implemented by using a gauge-equivariant neural network, such as the L-CNN.

Kai: The MAALA sampler connects the learned diffusion model with stochastic quantization (SQ) and includes Metropolis adjustment to correct for any bias or error in the machine-learned score function.

Paper summary: Kai: To summarize "Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory: Analytical Foundations and Numerical Scaling Tests", the paper lays out a way to use gauge equivariant diffusion models to generate ensembles of field configurations in non-abelian lattice gauge theories.

Mira: The thesis is that they provide a rigorous derivation of the reverse time stochastic differential equation on the group manifold, which underpins how you sample these configurations independently and uncorrelated.

Lev: They also discuss the gauge invariance of the denoising loss function, showing how that's crucial for making sure you get physically meaningful results regardless of what specific gauge you are in.

Kai: They also provide extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling, which is important for seeing if the method holds up when you go to larger systems.

Mira: So, they’ve shown that the analytical framework works by defining a group preserving diffusion process via the Stratonovich stochastic differential equation.

Lev: The Fokker-Planck equation associated with this diffusion process is derived as d t p t of U equals negative Da Ka of U pt of U plus half g squared t delta hat pt of U, where delta hat is the Laplace-Beltrami operator.

Kai: To learn this process, they need a score function s theta x mu t which approximates the derivative of the logarithm of the time dependent probability density.

Mira: The optimal score function they identify is actually the negative gradient of some time dependent action St U.

Lev: This action has to be parametrized by a gauge equivariant neural network, and they use denoising score matching as their loss function.

Kai: The method's gains come from using a gauge equivariant neural network, like an L CNN, to implement exact symmetries in the model architecture.

Mira: They also mention that the MAALA sampler connects this learned diffusion model with stochastic quantization and includes Metropolis adjustment to correct for any bias or error in the machine learned score function.

Lev: And they have score rescaling which allows sampling at inverse couplings different from the coupling of the training data, even if you only have a training ensemble at a single coupling.

Kai: So it’s about being able to apply this method to different physics regimes without having to retrain everything for each new parameter set.

Mira: And numerically, they tested two-dimensional U(two) lattice gauge theory, where the forward and backward paths of the reverse SDE agree very well with physical observables like the traced plaquette and topological charge susceptibility <ref:2611.03530#pg1>.

Conclusion: Kai: So looking at "Generalizable Equivariant Diffusion Models for Non-Abelian Lattice Gauge Theory: Analytical Foundations and Numerical Scaling Tests", what does this all actually mean?

Mira: It means they’ve shown that we can build a framework where you can sample complex field configurations in non abelian lattice gauge theories using diffusion models, with the key being the analytical foundation on the group manifold.

Lev: The authors are showing that this method has generalizability across lattice sizes and couplings, which is important because it means you don't just get stuck at one specific setting.

Kai: So when we talk about scaling behavior with volume and inverse coupling, they’re saying that the method doesn't suffer from increasing bias as you move to bigger systems or weaker couplings.

Mira: The title of this paper suggests a focus on both the analytical foundations and the numerical scaling tests, which means they want to show that it works not just theoretically but also practically.

Lev: It sets a clear path forward by suggesting that if SU three theory in four dimensions is feasible, then we can explore applying this technique to even more complicated physics.

Kai: So for someone listening who isn't deep in the weeds, it means we have a robust way to generate configurations that respects the symmetries of these theories without needing perfect knowledge of every single parameter upfront.

Mira: And they’re giving us a solid method for generating ensembles that is trainable at one setting but can then be used to explore many others reliably.

Gert Aarts, *, Diaa E. Habibi, ‡ Andreas Ipp, ↓ David I. Mäller, § Thomas R. Ranner, Υ Lingxiao Wang, ♅ Wei Wang, ↓ and Qianteng Zhu↓

Faculty of Physics, Bielefeld University · Centre for Quantum Fields and Gravity, Swansea University · Institute for Theoretical Physics, TU Wien · RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS) · Institute for Physics of Intelligence, Graduate School of Science, The University of Tokyo · State Key Laboratory of Dark Matter Physics, Key Laboratory for Particle Astrophysics and Cosmology (MOE) · Southern Center for Nuclear-Science Theory (SCNT)

hep-lat

Submitted: 2026-11-02

Updated: 2026-11-02

Comments: 28 pages, 18 figures, 1 table

Code: https://github.com/DMQFT/DM

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories Here we focus on the underlying analytical framework The paper…

Key concepts

Gauge-equivariant diffusion models
These are AI models designed to respect the symmetries of non-abelian gauge theories. They use diffusion processes, which gradually add noise to a configuration until it becomes pure randomness, and then learn how to reverse this process back into a valid physical configuration that respects the underlying mathematical symmetries of the theory.
Reverse-time stochastic differential equation (SDE)
This is a mathematical equation describing the path taken by a system in reverse time. In this context, it governs how to sample configurations from noise back into meaningful lattice gauge field configurations. It is derived on the group manifold, which is essential because gauge theories are defined on groups.
Score function
The score function acts as a guide for the diffusion process. It represents the gradient of the log probability density of a configuration at a given time step. The model learns this function to know exactly how to nudge noisy configurations back toward physically relevant field configurations.
Score rescaling
This technique allows the diffusion model, trained on data from one specific coupling and volume, to generate new ensembles for entirely different couplings or larger volumes. It enables generalization by adjusting the sampling process based on the desired physical parameters without retraining the entire model.

Terminology

Summary

Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories Here we focus on the underlying analytical framework The paper presents a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discusses the gauge invariance of the denoising loss function, and provides extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling

Analytical Foundations

The analytical framework begins by defining the lattice approximation of non-abelian Yang-Mills theory using a hypercubic lattice where the Euclidean Yang-Mills action is replaced with the Wilson action The continuum limit is approached by taking the lattice spacing to zero, which is achieved by letting L → ∞ and β → ∞ while keeping physical volume fixed

Diffusion Processes on Non-Abelian Lie Groups

The group-preserving diffusion process is defined via the Stratonovich stochastic differential equation (SDE) This SDE allows for multiple sampling methods, including the reverse SDE, which can be used to sample configurations independent and uncorrelated The Fokker-Planck equation (FPE) associated with this diffusion process is derived as d t p t(U) = −Da[Ka(U)pt(U)] + 1/2 g 2(t)∆ˆ pt(U), where ∆ˆ is the Laplace-Beltrami operator

Learning the Diffusion Process

The reverse diffusion process requires knowledge of the derivative of the logarithm of the time-dependent probability density, which is approximated by a score function s aθ;x,µ(U, t) The optimal score function is identified as the negative gradient of a time-dependent action St(U), and it must be parametrized by a gauge-equivariant neural network The loss function used is denoising score matching (DSM)

Where the Gains Come From

The gains come from several key aspects of the method:

  1. Exact symmetries are implemented by using a gauge-equivariant neural network, such as the L-CNN. p. 8

    Where the paper says it

    Furthermore, the score function needs to be compatible with all other global symmetries, such as translations, discrete rotations, and mirror symmetry. It is therefore essential that the score function is parametrized by a gauge-equivariant neural network.

  2. The MAALA sampler connects the learned diffusion model with stochastic quantization (SQ) and includes Metropolis adjustment to correct for any bias or error in the machine-learned score function.

  3. Score rescaling allows sampling at arbitrary couplings, even when only a training ensemble at a single coupling is available

Numerical Results and Benchmarking

Numerical experiments on two-dimensional U(2) lattice gauge theory show that the forward and backward paths of the reverse SDE agree very well, connecting the expectation value of the traced plaquette and topological charge susceptibility The MAALA generative process agrees with the forward diffusion process applied to HMC samples at β0 = 2 The method can generate new ensembles with couplings and volumes up to β = 18 and L = 256 without an increase in bias

Comparison to Stochastic Quantization

The comparison shows that MAALA performs significantly better for the topological susceptibility, with MAALA stabilizing quickly (bottom pane of Fig. 15) The methods can reach the analytic value with the chosen number of steps when increasing the total number of steps The score function is distinct from the gradient of the Wilson action multiplied by a time-dependent factor, establishing unambiguously that the learned timedependent effective action is more complicated than a simple modification of the Wilson action

Structure of the Time-Dependent Action

The forward diffusion process induces a more complicated lattice action than the Wilson action with a time-dependent coupling immediately, as shown in Fig. 17 The difference increases linearly in t and hence the time-dependent action deviates from a Wilson action immediately Links enclosed by a loop thus do not contribute, and only the links on Γ evolve The Creutz ratios are exactly conserved under the forward diffusion process, but they need to be rebuilt during the generative process

Conclusion and Outlook

The method can be trained at one value of the gauge coupling and on one lattice volume, but the generation of new ensembles can then be generalized to different gauge couplings and larger volumes, without an increase in bias The next step is therefore to investigate if practical application to the SU(3) theory in four dimensions is feasible It would also be interesting to include fermions, which has already been done for flow-based methods The next step is therefore to investigate if practical application to the SU(3) theory in four dimensions is feasible and if performance gains compared to standard methods can be achieved

Pruning redundant reasoning tokens

Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories The paper presents a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discusses the gauge invariance of the denoising loss function, and provides extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling

How the training signal is built

The training signal for the score function is constructed using denoising score matching (DSM) The loss function is rewritten as L = Z T 0 dt λ(t) E p(U(0)) p 0t(U(t)U(0)) X x,µ,a h s aθ;x,µ(U(t), t) − Da x,µ ln p 0t(U)(tU)(0) squared i squared

The primary advantage of MAALA lies in the ability to apply Metropolis adjustment to the Langevin trajectories to correct for any bias or error in the machine-learned score function The second advantage of MAALA is the ability to use score rescaling, that is, sampling at inverse couplings different from the coupling of the training data

The forward diffusion process induces a more complicated lattice action than the Wilson action with a time-dependent coupling immediately, as shown in Fig. 17 Links enclosed by a loop thus do not contribute, and only the links on Γ evolve The Creutz ratios are exactly conserved under the forward diffusion process, but they need to be rebuilt during the generative process

The method can be trained at one value of the gauge coupling and on one lattice volume, but the generation of new ensembles can then be generalized to different gauge couplings and larger volumes, without an increase in bias The next step is therefore to investigate if practical application to the SU(3) theory in four dimensions is feasible and if performance gains compared to standard methods can be achieved The next step is therefore to investigate if practical application to the SU(3) theory in four dimensions is feasible and if performance gains compared to standard methods can be achieved

The paper presents a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discusses the gauge invariance of the denoising loss function, and provides extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling The analytical framework begins by defining the lattice approximation of non-abelian Yang-Mills theory using a hypercubic lattice where the Euclidean Yang-Mills action is replaced with the Wilson action The continuum limit is approached by taking the lattice spacing to zero, which is achieved by letting L → ∞ and β → ∞ while keeping physical volume fixed The group-preserving diffusion process is defined via the Stratonovich stochastic differential equation (SDE) The Fokker-Planck equation (FPE) associated with this diffusion process is derived as d t p t(U) = −Da[Ka(U)pt(U)] + 1/2 g 2(t)∆ˆ pt(U), where ∆ˆ is the Laplace-Beltrami operator The analytical framework begins by defining the lattice approximation of non-abelian Yang-Mills theory using a hypercubic lattice where the Euclidean Yang-Mills action is replaced with the Wilson action

Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories The paper presents a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discusses the gauge invariance of the denoising loss function, and provides extensive numerical benchmarks regarding scaling behavior with respect to lattice volume and inverse coupling The analytical framework begins by defining the lattice approximation of non-abelian Yang-Mills theory using a hypercubic lattice where the Euclidean Yang-Mills action is replaced with the Wilson action The continuum limit is approached by taking the lattice spacing to zero, which is achieved by letting L → ∞ and β → ∞ while keeping physical volume fixed The group-preserving diffusion process is defined via the Stratonovich stochastic differential equation (SDE) The Fokker-Planck equation (FPE) associated with this diffusion process is derived as d t p t(U) = −Da[Ka(U)pt(U)] + 1/2 g 2(t)∆ˆ pt(U), where ∆ˆ is the Laplace-Beltrami operator The reverse diffusion process requires knowledge of the derivative of the logarithm of the time-dependent probability density, which is approximated by a score function s aθ;x,µ(U, t) The optimal score function is identified as the negative gradient of a time-dependent action St(U), and it must be parametrized by a gauge-equivariant neural network

The forward diffusion process induces a more complicated lattice action than the Wilson action with a time-dependent coupling immediately, as shown in Fig.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Implement score matching loss functions using lattice gauge-equivariant convolutional neural networks (L-CNN) to ensure that generated samples are representative of the target distribution while maintaining exact symmetries like local gauge symmetry.

  2. Develop a reverse diffusion process that utilizes a learned score function s theta;x,mu(U, t) to sample the target distribution p0(U), enabling sampling at arbitrary couplings and volumes by rescaling the score function according to score rescaling techniques.

  3. Integrate Metropolis-adjusted annealed Langevin algorithm (MAALA) with diffusion models to correct for any bias or error in the machine-learned score function during the sampling process, ensuring that generated ensembles asymptotically follow the same timedependent probability distribution as those from reverse SDE.

  4. Train diffusion models using time-dependent L-CNN architectures, specifically using a structure involving bilinear and convolutional layers with a random Fourier feature embedding to model the time dependence of the score function.

  5. Utilize analytical solutions for two-dimensional U(Nc) lattice gauge theory to benchmark trained models against known ground truth, verifying the implementation of the forward process and confirming that observables like Creutz ratios are exactly conserved under forward diffusion.

Abstract

Gauge-equivariant diffusion models can be used to generate ensembles of field configurations in non-abelian lattice gauge theories. Previously, we demonstrated the feasibility of modeling these theories by combining L-CNN-based diffusion models with the Metropolis-adjusted annealed Langevin algorithm (MAALA). Here we focus on the underlying analytical framework. In particular, we present a rigorous derivation of the reverse-time stochastic differential equation on the group manifold, discuss the gauge invariance of the denoising loss function, and provide extensive numerical benchmarks regarding the scaling behavior with respect to the lattice volume and the inverse coupling β. Full analytical expressions for arbitrarily sized Wilson loops and the topological charge susceptibility during the forward process are given as well.

Sources

Related papers