Random-order correction coupling from conditional partner bounds
ProvedFinitePermutationCoupling.random_order_switchingLet be finite and let partition it into classes. Let be the permutations preserving these classes, and designate some classes as active. Let be nonempty. Assume membership in depends only on the images of active inputs. For every exposed assignment, assume that an unexposed input in an active class with remaining inputs has each conditional partner probability at most under the uniform law on .
Let bound the size of every active class. There is a nonnegative joint weight on reference permutations and paths with reference marginal uniform on , terminal value on its support, and initial marginal uniform on . For every fixed reference , input , and ,
when the class is active; the left side is zero for inactive classes. The cap is stated by completion counts, including empty conditioning fibers and singleton residual classes.
This is the abstract finite stochastic construction in the random-order correction argument. Its hypotheses contain all source-specific combinatorics. Classes smaller than have identity padding. Stages may be recorded jointly across the disjoint active classes after their sequential histories have been sampled.
import Definitions.Def_FinitePermutationCoupling_BlockModel
namespace FinitePermutationCoupling
open Classical
open scoped BigOperators
/-- Abstract random-order residual-kernel construction from the proof of
Lemma 4.13, arXiv:2611.05912v1, Section 4.4, equations (76)--(77). -/
theorem random_order_switching
{α κ : Type*} [Fintype α] (block : α → κ) (active : κ → Prop)
(S : Finset (Equiv.Perm α)) (N : ℕ)
(hN : 0 < N) (hS : S.Nonempty) (hsub : S ⊆ blockSupport block)
(hsat : ActiveSaturated block active S)
(hcap : PartnerCap block active S)
(hsize : ∀ x, active (block x) → (blockFiber block x).card ≤ N) :
∃ v : Equiv.Perm α → (Fin (N + 1) → Equiv.Perm α) → ℝ,
(∀ h p, 0 ≤ v h p) ∧
(∀ h, (∑ p, v h p) = density S h) ∧
(∀ h p, v h p ≠ 0 → p (Fin.last N) = h) ∧
(∀ g, (∑ h, ∑ p, if p 0 = g then v h p else 0) =
density (blockSupport block) g) ∧
(∀ h (x : α) (i : Fin N),
(∑ p, if (p i.castSucc) x ≠ (p i.succ) x then v h p else 0) ≤
(if active (block x) then
2 / (((blockFiber block x).card : ℝ) * ((i.val : ℝ) + 1))
else 0) * density S h) := by
sorry
end FinitePermutationCoupling