Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

The refined pair state space and the half-reflection involution

Definition
BunkbedWZRefined

by burkh4rt · Sep 12, 2026 · Mathlib 0df444a (Lean v4.33.1)

bunkbedcombinatoricsgraph-theorypercolationprobability

In the Wierman--Ziff model each hyperedge eee of a 333-uniform hypergraph carries a pair of states, one for each level of the bunkbed, so the per-hyperedge state space is Υ×Υ\Upsilon \times \UpsilonΥ×Υ with 252525 elements, where Υ={abc,  a∣b∣c,  a∣bc,  ab∣c,  ac∣b}\Upsilon = \{abc,\; a|b|c,\; a|bc,\; ab|c,\; ac|b\}Υ={abc,a∣b∣c,a∣bc,ab∣c,ac∣b} records which of the three vertices a hyperedge connects.

Following Gladkov--Pak--Zimin we refine this space by splitting each of the two pairs (abc, a∣b∣c)(abc,\,a|b|c)(abc,a∣b∣c) and (a∣b∣c, abc)(a|b|c,\,abc)(a∣b∣c,abc) into a +++ and a −-− variant, with weights

P[(abc,a∣b∣c)+]=Pab∣cPac∣b,P[(abc,a∣b∣c)−]=PabcPa∣b∣c−Pab∣cPac∣b,\mathbb{P}\bigl[(abc,a|b|c)^{+}\bigr] = P_{ab|c}P_{ac|b}, \qquad \mathbb{P}\bigl[(abc,a|b|c)^{-}\bigr] = P_{abc}P_{a|b|c} - P_{ab|c}P_{ac|b},P[(abc,a∣b∣c)+]=Pab∣c​Pac∣b​,P[(abc,a∣b∣c)−]=Pabc​Pa∣b∣c​−Pab∣c​Pac∣b​,

and similarly for (a∣b∣c, abc)(a|b|c,\,abc)(a∣b∣c,abc). The two weights sum to PabcPa∣b∣cP_{abc}P_{a|b|c}Pabc​Pa∣b∣c​, the weight of the unsplit pair, so the refinement changes no probability whatsoever. Its purpose is to provide a domain for the half-reflection involution: the −-− variants are exactly the configurations in which a hyperedge lies wholly on one level, and it is the +++ variants that the involution is able to move.

The refined space Υext2\Upsilon^2_{\mathrm{ext}}Υext2​ has 272727 elements — the 232323 unsplit pairs together with the four variants — and is modelled here as (Υ×Υ)⊔{0,1,2,3}(\Upsilon \times \Upsilon) \sqcup \{0,1,2,3\}(Υ×Υ)⊔{0,1,2,3}, with a predicate legal\texttt{legal}legal cutting out the two now-redundant copies of the split pairs on the left summand.

This file provides:

  • toPair\texttt{toPair}toPair, the underlying pair of states, which is what determines the realised graph (the +++ and −-− variants of a pair have the same underlying pair, hence the same graph);
  • wt\texttt{wt}wt, the refined weight;
  • inLamSq\texttt{inLamSq}inLamSq, membership in the 161616-element set Λ+2\Lambda^2_{+}Λ+2​, where Λ={abc,ab∣c,ac∣b,a∣b∣c}\Lambda = \{abc, ab|c, ac|b, a|b|c\}Λ={abc,ab∣c,ac∣b,a∣b∣c} is Υ\UpsilonΥ with the state a∣bca|bca∣bc removed, and the two split pairs are represented by their +++ variants only;
  • rho\texttt{rho}rho, the reflection involution exchanging the two levels;
  • eta\texttt{eta}eta, the half-reflection involution of Definition 3.6, which is the identity on Ω0\Omega_0Ω0​, coincides with rho\texttt{rho}rho on Ω1\Omega_1Ω1​, and on Ξ\XiΞ exchanges (abc,a∣b∣c)+↔(ab∣c,ac∣b)(abc,a|b|c)^{+} \leftrightarrow (ab|c,ac|b)(abc,a∣b∣c)+↔(ab∣c,ac∣b) and (a∣b∣c,abc)+↔(ac∣b,ab∣c)(a|b|c,abc)^{+} \leftrightarrow (ac|b,ab|c)(a∣b∣c,abc)+↔(ac∣b,ab∣c).

Mathematical role. These are the ingredients of the involution argument by which Gladkov--Pak--Zimin prove their robustness lemma. The composite involution takes the first hyperedge along a distinguished path whose state lies in Λ+2\Lambda^2_+Λ+2​, applies eta\texttt{eta}eta there and rho\texttt{rho}rho to every later hyperedge; it is weight-preserving and exchanges the two connection events, so the configurations meeting Λ+2\Lambda^2_+Λ+2​ contribute equally to both and cancel from the difference. What remains are the configurations avoiding Λ+2\Lambda^2_+Λ+2​ entirely, which is where the alternative model and Hollom's computation enter.

Definition code
import Definitions.Def_BunkbedHypergraph

/-!
# The refined pair state space and the half-reflection involution

Sections 3.3 and 3.4 of Gladkov-Pak-Zimin, *The bunkbed conjecture is false*,
PNAS 122 (2025) e2420725122.

In the Wierman-Ziff model each hyperedge `e` carries a pair of states, one for each level, so the
per-hyperedge state space is `WZ x WZ` with 25 elements. Following the source we *refine* it by
splitting each of the two pairs `(abc, a|b|c)` and `(a|b|c, abc)` into a `+` and a `-` variant,
whose weights are `P_ab|c * P_ac|b` and `P_abc * P_a|b|c - P_ab|c * P_ac|b`. These sum to the
original weight `P_abc * P_a|b|c`, so the refinement does not change any probability; it exists
only to give the half-reflection involution a domain.

The refined space has 27 elements: the 23 ordinary pairs other than the two split ones, plus the
four variants. It is modelled here as `(WZ x WZ) + Fin 4`, with `legal` cutting out the two
now-redundant copies of the split pairs on the left summand.
-/

namespace Bunkbed.Hyper

/-- The refined per-hyperedge state space. `inl (a, b)` is the ordinary pair `(a, b)`; the four
right-hand elements are `(abc, a|b|c)+`, `(abc, a|b|c)-`, `(a|b|c, abc)+`, `(a|b|c, abc)-`. -/
abbrev WZExt := (WZ × WZ) ⊕ Fin 4

namespace WZExt

/-- The two pairs that get split are represented only on the right summand, so their copies on
the left are excluded. Exactly 27 of the 29 elements are legal. -/
def legal : WZExt → Bool
  | .inl (.abc, .a_b_c) => false
  | .inl (.a_b_c, .abc) => false
  | _ => true

/-- The underlying pair of WZ states, which is what determines the graph. -/
def toPair : WZExt → WZ × WZ
  | .inl p => p
  | .inr 0 => (.abc, .a_b_c)
  | .inr 1 => (.abc, .a_b_c)
  | .inr 2 => (.a_b_c, .abc)
  | .inr 3 => (.a_b_c, .abc)

/-- The refined weight. The `+` variants carry `P_ab|c * P_ac|b`, the `-` variants the
complementary `P_abc * P_a|b|c - P_ab|c * P_ac|b`. -/
def wt (P : WZ → ℚ) : WZExt → ℚ
  | .inl (a, b) => P a * P b
  | .inr 0 => P .ab_c * P .ac_b
  | .inr 1 => P .abc * P .a_b_c - P .ab_c * P .ac_b
  | .inr 2 => P .ab_c * P .ac_b
  | .inr 3 => P .abc * P .a_b_c - P .ab_c * P .ac_b

/-- Membership in `Λ²₊`: both components lie in `Λ = {abc, ab|c, ac|b, a|b|c}` (equivalently,
neither is `a|bc`), with the two split pairs represented by their `+` variants only. -/
def inLamSq : WZExt → Bool
  | .inl (.abc, .a_b_c) => false
  | .inl (.a_b_c, .abc) => false
  | .inl (a, b) => a != .a_bc && b != .a_bc
  | .inr 0 => true
  | .inr 1 => false
  | .inr 2 => true
  | .inr 3 => false

/-- The reflection involution: swap the two levels. It fixes the sign of a refined state. -/
def rho : WZExt → WZExt
  | .inl (a, b) => .inl (b, a)
  | .inr 0 => .inr 2
  | .inr 1 => .inr 3
  | .inr 2 => .inr 0
  | .inr 3 => .inr 1

/-- The half-reflection involution on `Λ²₊`, Definition 3.6. It is the identity on `Ω₀`, the
reflection on `Ω₁`, and swaps the `+` variants with the mixed pairs on `Ξ`. Outside `Λ²₊` it is
the identity. -/
def eta : WZExt → WZExt
  | .inr 0 => .inl (.ab_c, .ac_b)
  | .inr 2 => .inl (.ac_b, .ab_c)
  | .inl (.ab_c, .ac_b) => .inr 0
  | .inl (.ac_b, .ab_c) => .inr 2
  | .inl (.abc, .ab_c) => .inl (.ab_c, .abc)
  | .inl (.ab_c, .abc) => .inl (.abc, .ab_c)
  | .inl (.a_b_c, .ac_b) => .inl (.ac_b, .a_b_c)
  | .inl (.ac_b, .a_b_c) => .inl (.a_b_c, .ac_b)
  | q => q

end WZExt

end Bunkbed.Hyper
Source
N. Gladkov, I. Pak, A. Zimin, The bunkbed conjecture is false, PNAS 122 (2025) no. 24, e2420725122, https://doi.org/10.1073/pnas.2420725122, Sections 3.3-3.4 (pp. 3-5): the extended state space Upsilon^2_ext and Eq. [4] (p. 4), Definition 3.4 (refined pair set Lambda^2_+), Definition 3.5 (reflection involution) and Definition 3.6 (half-reflection involution).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me