The refined model computes the same expectations
ProvedBunkbedAux.refined_sum_eqThe refinement of the Wierman--Ziff pair state space splits each of the two pairs and into a and a variant, of weights and . Hyperedge by hyperedge this is clearly harmless, since the two weights sum to the weight of the unsplit pair. This theorem upgrades that observation to whole configurations.
Let index the hyperedges. A refined configuration is a map , and it is legal when no entry is one of the two redundant copies of a split pair. Each refined configuration determines an ordinary pair configuration , which is what fixes the realised graph. The claim is that for every function of the pair configuration,
Taking to be a connectivity indicator shows that no probability whatsoever is changed by passing to the refined model; the refinement is pure bookkeeping.
Mathematical role. This is what makes the involution argument of Gladkov--Pak--Zimin legitimate. The involution lives on the refined configuration space — it has to, because it exchanges a sub-event of of weight with the pair , and no such weight-preserving exchange exists before refining. This theorem is the statement that a cancellation established in the refined model is a cancellation of genuine Wierman--Ziff probabilities.
Formalization note. The proof groups refined configurations by the pair configuration they induce. Over a fixed the fibre is a product of per-hyperedge fibres, so the sum of products factors as a product of sums; each per-hyperedge fibre then contributes exactly , which for the two split pairs is where the two variant weights recombine. Note the statement holds for an arbitrary finite index type and arbitrary ; no hypothesis that is a probability distribution is needed.
import Definitions.Def_BunkbedWZRefined open Bunkbed.Hyper Finset
namespace BunkbedAux
/-- **The refined model computes the same expectations as the Wierman-Ziff model.** Summing any
function of the underlying pair configuration against the refined weights over all *legal*
refined configurations gives the same value as summing it against the ordinary product weights
over all pair configurations. So the refinement is measure-preserving at the level of whole
configurations, not merely hyperedge by hyperedge. -/
theorem refined_sum_eq {ι : Type*} [Fintype ι] [DecidableEq ι] (P : WZ → ℚ)
(g : (ι → WZ × WZ) → ℚ) :
∑ Ψ ∈ Finset.univ.filter (fun Ψ : ι → WZExt => ∀ i, WZExt.legal (Ψ i) = true),
(∏ i, WZExt.wt P (Ψ i)) * g (fun i => WZExt.toPair (Ψ i))
= ∑ φ : ι → WZ × WZ, (∏ i, P (φ i).1 * P (φ i).2) * g φ := by
sorry
end BunkbedAux