Sampling the Schwinger Model with Gauge-Equivariant Diffusion

Technical Summary: Sampling the Schwinger Model with Gauge-Equivariant Diffusion

Problem Statement
Lattice Field Theory (LFT) relies on numerical simulations to approximate path integrals for physical observables. However, as simulations approach the continuum limit, standard Markov chain Monte Carlo (MCMC) algorithms, such as Hybrid Monte Carlo (HMC), suffer from "critical slowing down." This phenomenon is characterized by exponentially growing autocorrelation times and low acceptance rates. Furthermore, in topologically non-trivial theories, MCMC chains often become "frozen" in specific topological sectors, failing to explore the full configuration space. While normalizing flows have shown promise in accelerating sampling for the lattice Schwinger model, the application of diffusion-based generative models to lattice gauge theories containing fermions remains unexplored.

Methodology
The authors propose a diffusion-based approach to sample gauge link configurations from the marginal distribution of the $N_f = 2$ lattice Schwinger model. The methodology involves the following components:

• Model Formulation: The authors utilize a score-based generative model formulated through Stochastic Differential Equations (SDEs). The forward process progressively corrupts initial gauge configurations $\phi_0$ with noise over time, while the reverse process uses a trained neural network (score network) to interpolate from noise back to data.
• Gauge Equivariance: To respect the $U(1)$ gauge symmetry of the model, the diffusion process is adapted to the Lie group $U(1)$. The forward SDE solution involves a "wrap" operation that folds the noisy configuration onto the unit circle modulo $2\pi$. The score network is constructed to be $U(1)$-equivariant by utilizing gauge-invariant inputs, specifically $2 \times 1$ rectangular Wilson loops and plaquettes.
• Training and Architecture: The score network employs a U-Net parameterization with Gaussian Fourier embeddings for time encoding. To address the non-locality introduced by the fermion determinant, the architecture incorporates dilated convolutions and axial attention layers to capture long-range correlations.
• Training Data: The model is trained on 8,192 gauge link configurations generated via HMC on $8 \times 8$ and $16 \times 16$ lattices with parameters $\beta = 2.0$ and $\kappa = 0.276$.
• Inference and Likelihood: New samples are generated by integrating the reverse ODE using an Euler integrator. Model likelihoods are estimated using the Skilling-Hutchinson trace estimator, allowing for the computation of unbiased observable estimates via reweighting.

Key Contributions

• First Application to Fermionic Lattice Gauge Theory: This work presents the first application of diffusion-based sampling to a lattice gauge theory that includes fermions.
• Gauge-Equivariant Score Network: The authors demonstrate the construction and training of a $U(1)$-equivariant score network capable of handling the marginal distribution of the Schwinger model after integrating out fermion fields.
• Unbiased Estimation: By computing model likelihoods, the authors establish a method to obtain unbiased estimates for physical observables directly from the diffusion-generated samples.

Results
The study evaluates the performance of the diffusion model against HMC on $8 \times 8$ and $16 \times 16$ lattices:

• Topological Exploration: Qualitative analysis of the topological charge ($Q$) and the Dirac sign ($\sigma$) reveals that the diffusion model exhibits significantly more transitions between even and odd topological sectors compared to HMC. This suggests a reduction in topological freezing and more efficient phase space exploration.
• Observable Accuracy: The diffusion-generated configurations yield unbiased estimates for key observables—the average plaquette $\langle P \rangle$, topological susceptibility $\chi_Q$, and chiral condensate $\langle \bar{\psi}\psi \rangle$—that closely match reference values obtained from HMC simulations.
• Efficiency Metrics: While the diffusion model shows lower acceptance rates (20–35%) and higher unnormalized Kullback-Leibler (KL) divergence compared to ideal HMC baselines, it achieves effective sample sizes (ESS) that indicate viable sampling capabilities. The integrated autocorrelation times for the topological charge are comparable to or slightly higher than HMC, but the qualitative improvement in topological tunneling is noted as a significant finding.

Significance and Claims
The authors position this work as a "first study" and an "exploratory" step forward for deep generative models in lattice gauge theory. They claim that their results demonstrate the feasibility of using diffusion models to overcome critical slowing down and topological freezing in fermionic theories. The paper modestly concludes that while the current approach is promising, further experimentation with model architectures is required to identify a scalable solution. Future work is identified as necessary to adapt these methods to different fermion handling schemes (such as pseudofermions) and to extend the approach to $SU(N)$ lattice gauge theories. The paper does not claim to have solved the scaling problem for large-scale QCD but rather establishes a proof-of-concept for diffusion-based sampling in the presence of fermions.