Gibbs conditioning, atypical consensus and splitting Gibbs measures on random regular graphs

Given n independent Bernoulli(p) random variables X_i, i = 1, ..., n, representing the opinions of individuals connected by an underlying random k-regular graph G_n on {1, ..., n}, we show that when conditioned on an atypical empirical consensus, which is the normalized sum of X_i X_j over neighboring vertices i, j, the joint distribution of the random variables converges, as n goes to infinity, to an Ising measure on the infinite k-regular tree T^k with a specific external field that depends only on the bias parameter p, and a temperature that depends on both p and the atypical consensus value. In particular, we show that conditional on the empirical consensus being smaller (respy, larger) than typical, the limit is a translation-invariant splitting (TIS) antiferromagnetic (respy, ferromagnetic) Ising measure on T^k. Moreover, if the bias is zero, then there is a phase transition: when the consensus exceeds k/(k-1), the conditional limits could be either the plus or minus boundary condition Ising measures. Furthermore, when X_i, i = 1, ..., n, are i.i.d. on a finite space, we show that when conditioned on an atypical value of the scaled sum of h(X_i, X_j) over neighboring vertices i and j, for any symmetric edge potential h, the limiting joint distribution of {X_i} lies in the set of (possibly degenerate) TIS Gibbs measures on T^k. The proofs leverage a tractable form of the large deviation rate function for component empirical measures of random regular graphs with i.i.d. marks and Gibbs conditioning principles, and entail careful analyses of associated non-convex constrained optimization problems. As a by-product of our results, we also obtain an (asymptotic) analog of the maximum entropy principle for Gibbs measures on random regular graphs.