Scaling of Stochastic Normalizing Flows in $\mathrm{SU}(3)$ lattice gauge theory

Andrea Bulgarelli, Elia Cellini, Alessandro Nada

Abstract

Non-equilibrium Markov Chain Monte Carlo (NE-MCMC) simulations provide a well-understood framework based on Jarzynski's equality to sample from a target probability distribution. By driving a base probability distribution out of equilibrium, observables are computed without the need to thermalize. If the base distribution is characterized by mild autocorrelations, this approach provides a way to mitigate critical slowing down. Out-of-equilibrium evolutions share the same framework of flow-based approaches and they can be naturally combined into a novel architecture called Stochastic Normalizing Flows (SNFs). In this work we present the first implementation of SNFs for $\mathrm{SU}(3)$ lattice gauge theory in 4 dimensions, defined by introducing gauge-equivariant layers between out-of-equilibrium Monte Carlo updates. The core of our analysis is focused on the promising scaling properties of this architecture with the degrees of freedom of the system, which are directly inherited from NE-MCMC. Finally, we discuss how systematic improvements of this approach can realistically lead to a general and yet efficient sampling strategy at fine lattice spacings for observables affected by long autocorrelation times.

Introduction

Simulating $\mathrm{SU}(3)$ lattice gauge theory—the mathematical framework for Quantum Chromodynamics (QCD)—is a cornerstone of modern particle physics. However, as researchers attempt to reach the "continuum limit" by using finer lattice spacings, standard Markov Chain Monte Carlo (MCMC) algorithms encounter a significant bottleneck known as critical slowing down. A particularly challenging aspect of this is topological freezing, where simulations become trapped in specific vacuum configurations, failing to explore the full physical landscape of the theory.

This paper presents the first implementation of Stochastic Normalizing Flows (SNFs) for 4D $\mathrm{SU}(3)$ gauge theory. By interleaving trainable neural network layers with stochastic Monte Carlo updates, the authors demonstrate a sampling strategy that inherits the robust scaling of non-equilibrium physics while significantly improving efficiency over traditional methods.

Methodology: Stochastic Normalizing Flows

The SNF architecture combines two powerful frameworks to navigate the configuration space of lattice gauge theory:

  1. Non-Equilibrium MCMC (NE-MCMC): Based on Jarzynski’s equality, this framework drives a system from a simple base distribution to a complex target distribution through a sequence of non-equilibrium steps, avoiding the need for costly thermalization at every stage.
  2. Normalizing Flows (NFs): These are trainable, deterministic layers that learn to map one distribution to another. To respect the symmetries of the theory, the authors utilize gauge-equivariant coupling layers based on "stout-smearing," which transforms lattice links while preserving underlying gauge invariance.

A key innovation is the step-by-step training strategy. Rather than training the entire flow at once, the authors minimize the "dissipated work" (mathematically equivalent to the Kullback-Leibler divergence) between the forward and reverse paths. This thermodynamic approach minimizes friction during the transition, maximizing the reversibility and efficiency of the sampling process.

Experimental Setup

The SNF framework was evaluated on $\mathrm{SU}(3)$ Yang-Mills theory across multiple lattice volumes and coupling regimes.

Table of ensembles used in the study, showing the range of inverse coupling $\beta$, lattice spacing ratios $a/r_0$, and lattice sizes $L/a$.

As detailed in the table above, the study utilized lattice sizes ranging from $10^4$ to $20^4$. The ensembles were designed to test the flow's performance as it transitions between different lattice spacings (controlled by the inverse coupling $\beta$).

Results: Scaling and Efficiency

The most significant finding is that SNFs inherit the favorable scaling properties of NE-MCMC. The computational cost required to maintain a specific sampling quality (measured by Effective Sample Size) grows linearly with the lattice volume ($L^4$).

Dissipation density $\rho^{(n)}$ as a function of the flow step $n$. The top panel shows raw values for different total steps $n_{step}$, while the bottom panel demonstrates a scaling collapse when normalized by $n_{step}$ and plotted against the fractional step $n/n_{step}$.

As shown in the figure above, the dissipation density profiles collapse onto a single curve when properly scaled, confirming that the algorithm's performance is a predictable function of the system's degrees of freedom. Key results include:

  • Efficiency Gains: SNFs were found to be roughly twice as efficient as pure NE-MCMC, achieving the same level of reversibility in half the number of steps.
  • Low Overhead: The neural network layers only increase the cost of a single update by about 25%, while providing a 100% gain in step-efficiency.
  • Transfer Learning: Parameters learned on a small $12^4$ lattice were successfully transferred to a $20^4$ lattice without retraining, suggesting that the learned "thermodynamic path" is robust across scales.

Conclusion and Future Outlook

Stochastic Normalizing Flows provide a stable and efficient hybrid architecture for high-dimensional lattice simulations. By finding thermodynamic paths that minimize dissipation, SNFs offer a scalable alternative to pure machine learning or traditional MCMC approaches. Future work will explore optimizing the coupling protocols and applying these techniques to boundary condition changes, which could directly mitigate the topological freezing that plagues simulations at fine lattice spacings.

Summary updated

Abstract

Non-equilibrium Markov Chain Monte Carlo (NE-MCMC) simulations provide a well-understood framework based on Jarzynski's equality to sample from a target probability distribution. By driving a base probability distribution out of equilibrium, observables are computed without the need to thermalize. If the base distribution is characterized by mild autocorrelations, this approach provides a way to mitigate critical slowing down. Out-of-equilibrium evolutions share the same framework of flow-based approaches and they can be naturally combined into a novel architecture called Stochastic Normalizing Flows (SNFs). In this work we present the first implementation of SNFs for $\mathrm{SU}(3)$ lattice gauge theory in 4 dimensions, defined by introducing gauge-equivariant layers between out-of-equilibrium Monte Carlo updates. The core of our analysis is focused on the promising scaling properties of this architecture with the degrees of freedom of the system, which are directly inherited from NE-MCMC. Finally, we discuss how systematic improvements of this approach can realistically lead to a general and yet efficient sampling strategy at fine lattice spacings for observables affected by long autocorrelation times.
BibTeX
@inproceedings{bulgarelli2025scaling,
  author = {Andrea Bulgarelli and Elia Cellini and Alessandro Nada},
  title = {{Scaling of Stochastic Normalizing Flows in $\textbackslash\{\}mathrm\{SU\}(3)$ lattice gauge theory}},
  booktitle = {Physical},
  year = {2025},
  doi = {10.1103/physrevd.111.074517},
}
Figure