Researchers propose a machine‑learned way to make global updates in Monte Carlo simulations of SU(2) lattice gauge theory. The goal is to move beyond purely local updates, which explore configuration space slowly as the continuum limit is approached. The new construction is designed so that the learned global moves can be used inside a Metropolis‑Hastings algorithm without breaking the formal correctness of the sampler.

The core idea is a coupling‑flow transformation that updates only a selected subset of lattice links while keeping the rest fixed. The active links are multiplied on the left by SU(2) group elements produced by a neural network. Crucially, the network’s output for the active links depends only on the frozen links. That one‑way dependence makes the map explicitly invertible by construction. The authors also augment proposals with a simple auxiliary ±1 branch variable so that forward and reverse proposals are explicit and symmetric.

Because each active link is updated by left multiplication, the transformation preserves the product Haar measure on the lattice‑link manifold. The Haar measure is the natural uniform measure on the SU(2) group. Its left‑invariance means no Jacobian appears when changing variables. As a result, the Metropolis‑Hastings acceptance ratio reduces to the usual Boltzmann weight ratio, A = min(1, e^{−[S(U′)−S(U)]}), where S is the Wilson gauge action used in the study. In the implementation the network outputs three real numbers per active link, interpreted as an element of the su(2) Lie algebra and mapped to SU(2) by the matrix exponential. Links are represented in a quaternion (four‑component) form for numerical convenience.

The paper implements this scheme in two‑dimensional pure SU(2) lattice gauge theory and compares it to a standard local Metropolis algorithm used as a reference. In the testbed the learned proposal reproduces the target ensemble within statistical resolution across the tested configurations. In matched local‑step comparisons the learned proposal gives results of comparable quality to the baseline, but it does not outperform the pure local baseline in the conservative matched‑step case examined with seed‑level statistics. A mixed‑step hybrid arrangement, which alternates learned global proposals with local steps, produced a modest improvement in effective sample size per unit runtime in these experiments. The authors stress that their learned transformation stayed in a near‑identity regime during these tests.