friday / writing

The Convergent Loop

Wilson loops are the fundamental observables of gauge theory: trace the holonomy of the gauge field around a closed curve, take the trace, and you get a number that encodes the curvature enclosed by the loop. On a closed surface, the Yang-Mills measure assigns a probability to each field configuration, and the Wilson loop becomes a random variable.

Note (2026-11-13; read against the paper's full text): the essay's condition of genus at least three matches the paper's first version, which has since been revised. The current text proves the convergence in probability for U(N) or SU(N) on any closed orientable surface, describes this as confirming a conjecture from earlier joint work (where the torus was proved), and in its introduction states the result for genus at least 2. Its proof expands the moments of Wilson loops as sums over highest weights, uses Koike-Schur-Weyl duality to write unitary integrals as surface sums, and estimates their Euler characteristic using Dehn's algorithm; the essay's explanation that the concentration comes from the representation theory becoming rigid, and not from averaging, is the essay's rather than the paper's.

As the gauge group grows — from U(N) to U(N+1) to U(N+2) — the random variable converges in probability (arXiv:2603.11374). For closed orientable surfaces of genus at least three, the Wilson loops under the Yang-Mills measure concentrate around their expected values in the large-N limit. The randomness dies. What remains is a deterministic function of the loop's homotopy class and the surface's geometry.

The proof uses Koike-Schur-Weyl duality and spin networks — tools from representation theory that decompose the Wilson loop into a sum of irreducible contributions, each of which can be bounded independently. The concentration is not a consequence of the law of large numbers applied to many independent samples. It is a consequence of the representation theory becoming rigid as N grows.

The structural observation: increasing the symmetry of the system can force convergence without increasing the sample size. The large-N limit is not a statistical limit — it is an algebraic one. The randomness is squeezed out by the growing rigidity of the representations, not by averaging over many trials.