Negative Correlations for Forests and the $q<1$ Random Cluster Model

Abstract

We study negative edge correlation for two well-known models in statistical physics, the arboreal gas and the $q < 1$ random-cluster model. We show that after the edges $e$ and $f$ are removed, their Rayleigh difference is a crossing contribution minus the covariance of two endpoint-connectivity events. For the arboreal gas on an arbitrary finite graph, we analyze the two leading coefficients of this difference. The first is the classical transfer-current square. If it vanishes, the second has an electrical sum-of-squares formula. This proves negative edge correlation at sufficiently large fugacity whenever one of the two leading coefficients of the Rayleigh difference is nonzero, and characterizes simultaneous equality by an equipotential decomposition. For the random-cluster model on the complete graph $K_n$, we prove negative edge correlation throughout $q<1$ when all edge weights are at least $2$. In the uniform case, we prove something stronger:negative correlation holds when all the weights are equal and at least $1$. In particular, distinct edges in a connected spanning subgraph of $K_n$, weighted by a fixed fugacity, are negatively correlated for every $n$. We also explore what seems to be a general phenomenon: positive correlation of connectivity events. For the arboreal gas, we prove positive correlation for connectivity events at high fugacities. On lattices, the conjectured inequality would make the two-point function supermultiplicative and produce a convex inverse correlation length. Exact computations support the graph and matroid conjectures, including the Seymour--Welsh matroid $\mathcal S_8$, where ordinary edge-negative correlation fails. The results in this paper were derived by the authors without the use of Large Language models. The authors did benefit from using GPT-5 Pro for generating code to test out hypotheses as well as for simplifying the arguments.