⚛️ lattice

Sampling SU(N) gauge theory on a 2D lattice from independent plaquettes via holonomies and corner reweighting

This paper proposes a statistical sampling method for 2D SU(N) lattice gauge theory that bypasses the difficulty of mapping independent plaquettes to link variables by reformulating the problem in terms of holonomies and corner reweighting, achieving high acceptance rates and moderate effective sample sizes through a combination of normalizing flows and closed-form or trained samplers for the group commutator constraint.

Original authors: Javad Komijani

Published 2026-10-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Javad Komijani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand the universe at its smallest scales, physicists often turn to a method called lattice gauge theory. Imagine space not as a smooth, continuous fabric, but as a grid of tiny points, like the squares on a sheet of graph paper. On this grid, the fundamental forces of nature are represented by values attached to the lines connecting these points. The most famous version of this theory, known as the Wilson formulation, calculates the energy of the system by looking at the smallest possible loops on the grid. These loops, which the theory calls plaquettes, are the building blocks of the interaction. While it is relatively easy to understand the behavior of a single loop, the real challenge arises when scientists try to simulate the entire grid. The values on the lines must fit together perfectly to form a consistent picture of the physical world, much like tiles in a mosaic that must align to create a coherent image. If the pieces do not fit, the simulation fails to represent reality.

For decades, simulating these systems has relied on a slow, step-by-step process where the computer changes one piece at a time and checks if the new arrangement is valid. This method works, but it can be incredibly slow, especially when trying to generate a large number of independent examples of the system. A newer approach, using artificial intelligence to generate samples all at once, has shown promise, but it hits a wall when applied to these grid theories. The AI can easily create the small loops, but it struggles to stitch them together into a valid, consistent grid without breaking the rules of the physics. The pieces simply do not fit together in the way the theory demands.

A researcher named Javad Komijani has found a clever statistical way around this obstacle for two-dimensional grids. Instead of trying to force the pieces to fit perfectly from the start, the new method allows the computer to generate the small loops independently, ignoring the fact that they might not currently connect correctly. It then treats the problem of making them fit as a separate, manageable task. The key insight involves a mathematical trick where the researcher introduces a temporary, flexible version of the grid lines. These flexible lines can be adjusted to satisfy the connection rules, but only if a specific condition is met at the four corners of the grid. This condition is difficult to satisfy directly, so the researcher changed the rules of the game slightly. By allowing the edges of the grid to twist in a specific way, the difficult condition of making the pieces fit was transformed into a problem of finding two specific values that, when combined in a particular order, produce a known result.

This transformation turned a rigid, impossible-to-solve constraint into a statistical sampling problem. The computer can now generate the small loops freely, and then, for each set of loops, it calculates a specific value derived from the corners. The researcher then uses a specialized tool to find the two missing values that would make the corners fit, given that specific result. For the simplest version of the theory, this tool works perfectly and instantly, providing an exact solution. For a slightly more complex version, the tool uses a trained artificial intelligence model to find the best possible match. Because this matching process is not always perfect, the final results are adjusted with a weight, a number that tells the computer how much to trust each generated example.

The results of this new method are impressive. When tested on grids of different sizes, from tiny 2-by-2 squares to much larger 32-by-32 grids, the method produced results that matched the traditional, slower simulations with high precision. The acceptance rate, which measures how often the generated examples are valid, exceeded 99 percent. Perhaps most importantly, the efficiency of the method did not degrade as the grids got larger. In many other approaches, the difficulty of the problem grows rapidly with the size of the grid, but here, the quality of the samples remained steady. The final set of examples was statistically robust, with the effective number of useful samples staying well above half of the total generated, a strong sign that the method is reliable.

This work demonstrates that it is possible to bypass the traditional bottlenecks of simulating complex physical systems by rethinking how the constraints are applied. By turning a hard geometric puzzle into a statistical question, the researcher has shown that generative models can be successfully applied to lattice gauge theory without sacrificing accuracy. The method works for the simplest cases of the theory and provides a clear path forward for more complex versions. It offers a new way to explore the fundamental structure of matter, proving that with the right mathematical perspective, even the most stubborn constraints can be turned into solvable problems.

Technical Summary: Sampling SU(N) Gauge Theory on a 2D Lattice via Holonomies and Corner Reweighting

Problem Statement
In the Wilson formulation of lattice gauge theory, the fundamental degrees of freedom are group-valued link variables $U_\mu(x) \in SU(N)$ , while the action is a sum over the traces of plaquettes (smallest Wilson loops). For generative sampling methods, such as normalizing flows, a significant obstruction arises: while the distribution of an individual plaquette is relatively easy to model, mapping sampled plaquettes back to the underlying link variables is non-trivial. Standard approaches that exclude one plaquette to compute links and then reweight the configuration suffer from an effective sample size (ESS) that depends drastically on the gauge coupling, leading to poor efficiency.

Methodology
The paper proposes a statistical construction in two dimensions to circumvent the mapping obstruction by utilizing a formulation based on prelinks and holonomies on an extended lattice.

Prelinks and Holonomies:
The author introduces prelink variables $V_\mu(x)$ defined on a minimally extended lattice $\tilde{\Lambda}$ . Unlike standard links, prelinks possess a "semi-global" freedom (initial-value symmetry) along each lattice line. By defining holonomies $h_{\mu\nu}(x) = V_\mu(x)V^\dagger_\nu(x)$ , the plaquette action can be rewritten entirely in terms of these variables.
The Corner Constraint:
On a periodic lattice, holonomy variables are not fully independent; they must satisfy a consistency condition at the four corners of the extended lattice: $C_{00} C^\dagger_{01} C_{11} C^\dagger_{10} = I$ . This hard, group-valued delta-function constraint prevents direct independent sampling of holonomies.
Twisted Boundary Conditions and Group Commutator:
To relax this constraint, the author employs a twisted boundary condition on the prelinks. This transforms the hard corner constraint into a conditional sampling problem. Specifically, the consistency condition is traded for the requirement to sample two group elements $X, Y \in SU(N)$ given their group commutator $Z = XYX^\dagger Y^\dagger$ , where $Z$ is determined by the holonomies at the four corners.
Factorization of the Measure:
The total measure factorizes into:
- An open-boundary holonomy theory where plaquettes can be sampled independently.
- A conditional sampling term for the corner variables $(X, Y)$ given $Z$ .
- A weight factor $\rho(Z)$ , representing the marginal density of the group commutator $Z$ .
Sampling Workflow:
- Plaquettes: A normalizing flow (spectral-flow model) is trained to sample single plaquettes according to the Wilson action. These are drawn independently.
- Holonomies: Holonomies are constructed deterministically from the plaquettes via cumulative matrix products.
- Corner Correction: Given the computed $Z$ $Z$ from the holonomies, the author samples the remaining degrees of freedom $(D, \Omega)$ $(D, Ω)$ required to reconstruct $X$ $X$ and $Y$ $Y$ .
  - For SU(2), this is solved in closed form using an encoder-decoder scheme and Euler angles, yielding an exact sampler with unit weight.
  - For SU(3), a trained normalizing flow is used to approximate the conditional distribution of $(D, \Omega)$ given $Z$ .
- Reweighting: The final configurations are reweighted by the density $\rho(Z)$ (and the conditional sampler weight for SU(3)) to recover the exact measure.

Key Contributions

Closed-Form Densities: The paper derives the closed-form density $\rho(Z)$ for the group commutator $Z = XYX^\dagger Y^\dagger$ for both $SU(2)$ and $SU(3)$ . For $SU(3)$ , this is expressed as a ratio of Vandermonde determinants of eigen-angles and eigenvalues.
Exact SU(2) Sampler: An exact, closed-form conditional sampler for the corner problem in $SU(2)$ is constructed, eliminating the need for approximate inference in this case.
Normalizing Flow for SU(3) Corners: A trained normalizing flow is implemented to handle the conditional sampling for $SU(3)$ , demonstrating that the approach scales to larger groups.
Statistical Construction: The method establishes a weighted map between independent plaquette samples and valid link configurations, avoiding the drastic coupling-dependent ESS degradation seen in exclusion-based baselines.

Results
The method was tested on $2 \times 2$ and $32 \times 32$ lattices for both $SU(2)$ and $SU(3)$ at various couplings ( $\beta$ ).

Accuracy: The reweighted normalizing flow results ("NF:weighted") agree well with Hybrid Monte Carlo (HMC) benchmarks. The normalized deviation $z$ between the flow results and HMC is typically $|z| \lesssim 2.1$ across 64 different Wilson loop observables.
Efficiency: The per-plaquette acceptance rates for the independent sampling step exceed 99%.
Effective Sample Size (ESS): The ESS of the final reweighted configurations is moderate and stable, typically ranging between 0.56 and 0.75. Crucially, this ESS does not degrade significantly as the lattice size increases from $2 \times 2$ to $32 \times 32$ , nor does it depend directly on the gauge coupling (only indirectly through $Z$ ). For $SU(2)$ , the ESS is determined solely by the density of $Z$ , approaching $\approx 0.61$ for asymptotically large lattices.

Significance
The paper claims that this approach provides a viable statistical construction for sampling 2D $SU(N)$ gauge theories using generative models. By trading a hard geometric constraint for a conditional sampling problem with a known density, the method allows for the independent sampling of plaquettes while maintaining a moderate effective sample size that is robust against lattice size and coupling variations. The author notes that while the holonomy parametrization is applicable in higher dimensions, the additional constraints there make the resolution of corner conditions non-trivial and are left for future work.

Technical Summary: Sampling SU(N) Gauge Theory on a 2D Lattice via Holonomies and Corner Reweighting

More like this