Exponential two-replica likelihood-density tail
OpenZeroFreeRRG.Hardcore.replica_likelihood_tailUnder Condition 1 for either positive-parameter model, for every there are a reveal probability and such that, for all sufficiently large admissible , the planted joint probability that the two pinned log-likelihoods sum to more than is at most . This is the density-tail estimate in Lemma 5.6, with .
Preamble
import Mathlib import Definitions.Def_ZeroFreeRRG_Hardcore_BetheFunctional import Definitions.Def_ZeroFreeRRG_Hardcore_Moments set_option autoImplicit false open Set Filter MeasureTheory AFIsingKS ZeroFreeRRG.Hardcore open scoped Classical BigOperators
Formal statement
theorem ZeroFreeRRG.Hardcore.replica_likelihood_tail (m : Model) (hm : m.Valid) (d : ℕ) (hd : 3 ≤ d) (t : ℝ) (ht : m.PosParam t)
(hC1 : Condition1 m d t) (η : ℝ) (hη : 0 < η) :
∃ s : ℝ, 0 < s ∧ s < 1 ∧ ∃ c : ℝ, 0 < c ∧ ∃ N : ℕ,
∀ n ≥ N, Even (n*d) →
(∑ M : Matching n d, ∑ σ : Fin n → m.Spin, ∑ σ' : Fin n → m.Spin, ∑ U : Finset (Fin n),
planted m t M σ * gibbs m t M σ' * revealProb s U *
(if pinnedLL m t M σ U + pinnedLL m t M σ' U > (s*η^2/24)*n then 1 else 0)) ≤
2*Real.exp (-c*n) := by sorrySource
Chen–Wang–Zhang, Zero-Freeness for Random Regular Graphs up to the Replica Symmetry Breaking Threshold, https://github.com/sigact-stoc/stoc2027/releases/download/papers-1/1461.pdf, p. 39, proof of Lemma 5.6.