Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 3.3 — the robust hyperedge lemma

Proved
BunkbedFalse.robust_hyperedge

by burkh4rt · Sep 11, 2026 · Mathlib 0df444a (Lean v4.33.1)

combinatoricsgraph-theorypercolationprobability

Let HHH be Hollom's hypergraph with transversal set T={u2,u7,u9}T = \{u_2,u_7,u_9\}T={u2​,u7​,u9​}, and consider Wierman–Ziff percolation on the associated bunkbed hypergraph, in which each hyperedge independently takes one of five states with probabilities PabcP_{abc}Pabc​, Pa∣b∣cP_{a|b|c}Pa∣b∣c​, Pa∣bcP_{a|bc}Pa∣bc​, Pab∣cP_{ab|c}Pab∣c​, Pac∣bP_{ac|b}Pac∣b​ summing to 111. Assume the model's symmetry Pab∣c=Pac∣bP_{ab|c} = P_{ac|b}Pab∣c​=Pac∣b​, the nondegeneracy condition PabcPa∣b∣c−Pab∣cPac∣b>0P_{abc}P_{a|b|c} - P_{ab|c}P_{ac|b} > 0Pabc​Pa∣b∣c​−Pab∣c​Pac∣b​>0, and the robustness hypothesis

400 Pa∣bc  ≤  Pabc Pa∣b∣c  −  Pab∣c Pac∣b.400\,P_{a|bc} \;\le\; P_{abc}\,P_{a|b|c} \;-\; P_{ab|c}\,P_{ac|b}.400Pa∣bc​≤Pabc​Pa∣b∣c​−Pab∣c​Pac∣b​.

Then

Pwz(u1↔u10)  <  Pwz(u1↔u10′).\mathbb{P}^{\mathrm{wz}}\bigl(u_1 \leftrightarrow u_{10}\bigr) \;<\; \mathbb{P}^{\mathrm{wz}}\bigl(u_1 \leftrightarrow u'_{10}\bigr).Pwz(u1​↔u10​)<Pwz(u1​↔u10′​).

This is the robust hyperedge lemma and the technical heart of the refutation. Hollom's computation applies only to the exact hypergraph model; this lemma shows the strict inequality survives in the far more flexible WZ model, and therefore survives the inexact simulation of a hyperedge by a gadget graph. The source proves it by constructing a weight-preserving involution on a refined configuration space, built from a reflection involution and a half-reflection involution.

Formalization Note The hypothesis that the five probabilities are nonnegative and sum to 111 is stated explicitly. The source notes (§8.2) that this lemma is a finite problem — positivity of degree-222 polynomials in five variables with at most 5125^{12}512 nonzero coefficients on a region of the unit cube — and that it was independently verified by computer; a formal proof would replace that check with a certificate.

⚠️ Deviation from the printed statement. The published Lemma 3.3 does not include the nondegeneracy hypothesis PabcPa∣b∣c−Pab∣cPac∣b>0P_{abc}P_{a|b|c} - P_{ab|c}P_{ac|b} > 0Pabc​Pa∣b∣c​−Pab∣c​Pac∣b​>0, and is false without it. Take

Pabc=u2,Pab∣c=Pac∣b=u(1−u),Pa∣bc=0,Pa∣b∣c=(1−u)2(0≤u≤1).P_{abc} = u^2,\quad P_{ab|c} = P_{ac|b} = u(1-u),\quad P_{a|bc} = 0,\quad P_{a|b|c} = (1-u)^2 \qquad (0 \le u \le 1).Pabc​=u2,Pab∣c​=Pac∣b​=u(1−u),Pa∣bc​=0,Pa∣b∣c​=(1−u)2(0≤u≤1).

These are nonnegative, sum to 111, satisfy the symmetry, and satisfy Eq. [1] as 0≤00 \le 00≤0. But this is precisely the case in which each hyperedge degenerates into two independent ordinary edges aaa–bbb and aaa–ccc. Every edge of the resulting graph is then incident to a transversal vertex; the posts identify u2,u7,u9u_2, u_7, u_9u2​,u7​,u9​ across the two levels; and u1u_1u1​ and u10u_{10}u10​ each hang off a single transversal vertex by one edge. Both probabilities therefore factor as u2⋅Pr⁡[u2∼u9]u^2 \cdot \Pr[u_2 \sim u_9]u2⋅Pr[u2​∼u9​] and are exactly equal, so the strict conclusion fails. (At u=1/2u = 1/2u=1/2 both equal 734377/4194304734377/4194304734377/4194304.)

The source's own proof needs this hypothesis: it divides by Z=2(PabcPa∣b∣c−Pab∣cPac∣b)+2Pa∣bc−Pa∣bc2Z = 2(P_{abc}P_{a|b|c} - P_{ab|c}P_{ac|b}) + 2P_{a|bc} - P_{a|bc}^2Z=2(Pabc​Pa∣b∣c​−Pab∣c​Pac∣b​)+2Pa∣bc​−Pa∣bc2​, which vanishes on exactly this family, and it concludes strictly only when the conditioning event has positive probability. The hypothesis is satisfied in the intended application, since Lemma 4.1 supplies a strictly positive value.

Preamble
import Definitions.Def_BunkbedHypergraph
open Bunkbed Bunkbed.Hyper SimpleGraph Finset
Formal statement
namespace BunkbedFalse

/-- **Lemma 3.3** (robust hyperedge lemma, p. 3), with the nondegeneracy hypothesis `hpos`
added: the printed statement is false without it. -/
theorem robust_hyperedge (P : WZ → ℚ)
    (hnn : ∀ s, 0 ≤ P s)
    (hsum : P .abc + P .a_b_c + P .a_bc + P .ab_c + P .ac_b = 1)
    (hsym : P .ab_c = P .ac_b)
    (hpos : 0 < P .abc * P .a_b_c - P .ab_c * P .ac_b)
    (h400 : 400 * P .a_bc ≤ P .abc * P .a_b_c - P .ab_c * P .ac_b) :
    wzProb hollomTriple hollomT P ((0 : Fin 10), (0 : Fin 2)) ((9 : Fin 10), (0 : Fin 2))
      < wzProb hollomTriple hollomT P ((0 : Fin 10), (0 : Fin 2)) ((9 : Fin 10), (1 : Fin 2)) := by
  sorry

end BunkbedFalse
Source
N. Gladkov, I. Pak, A. Zimin, "The bunkbed conjecture is false", Proc. Natl. Acad. Sci. USA 122 (2025), no. 24, e2420725122, https://doi.org/10.1073/pnas.2420725122 - p. 3, Lemma 3.3 (Eqs. [1] and [2]); proof pp. 5-7.
Read-back

What the Lean code literally says, in plain math · claude-opus-5

Read-back — BunkbedFalse.robust_hyperedge

The objects the statement is built from

The statement lives on the vertex set V=Z/10={0,1,…,9}V = \mathbb{Z}/10 = \{0,1,\dots,9\}V=Z/10={0,1,…,9} (written Fin 10), doubled into two levels indexed by {0,1}\{0,1\}{0,1}, so the ambient vertex set of every graph below is

V×{0,1}={(v,l):0≤v≤9, l∈{0,1}},V \times \{0,1\} = \{(v,l) : 0 \le v \le 9,\ l \in \{0,1\}\},V×{0,1}={(v,l):0≤v≤9, l∈{0,1}},

twenty vertices in all. Level 000 is the "lower" copy and level 111 the "upper" copy.

The six ordered triples. The statement uses the fixed list hollomTriple, a function {0,…,5}→V3\{0,\dots,5\} \to V^3{0,…,5}→V3 given by the ordered triples

t0=(1,0,2),t1=(8,2,5),t2=(6,5,4),t3=(1,4,3),t4=(6,3,7),t5=(8,7,9).t_0=(1,0,2),\quad t_1=(8,2,5),\quad t_2=(6,5,4),\quad t_3=(1,4,3),\quad t_4=(6,3,7),\quad t_5=(8,7,9).t0​=(1,0,2),t1​=(8,2,5),t2​=(6,5,4),t3​=(1,4,3),t4​=(6,3,7),t5​=(8,7,9).

The order inside each triple is load-bearing: writing ti=(ai,bi,ci)t_i = (a_i,b_i,c_i)ti​=(ai​,bi​,ci​), the first entry aia_iai​ is distinguished from the other two. Note a0=a3=1a_0=a_3=1a0​=a3​=1, a1=a5=8a_1=a_5=8a1​=a5​=8, a2=a4=6a_2=a_4=6a2​=a4​=6; vertex 000 occurs only in t0t_0t0​ (as b0b_0b0​) and vertex 999 only in t5t_5t5​ (as c5c_5c5​).

The transversal set and the posts. hollomT is the three-element set T={1,6,8}T=\{1,6,8\}T={1,6,8}, which is exactly {a0,…,a5}\{a_0,\dots,a_5\}{a0​,…,a5​}. The posts are the three vertical edges

{(1,0),(1,1)},{(6,0),(6,1)},{(8,0),(8,1)},\{(1,0),(1,1)\},\qquad \{(6,0),(6,1)\},\qquad \{(8,0),(8,1)\},{(1,0),(1,1)},{(6,0),(6,1)},{(8,0),(8,1)},

and they are present in every configuration, with probability 111 — they are never randomised. They are the only edges joining level 000 to level 111.

The five states. WZ is a five-element inductive type with constructors

abc,ab_c,ac_b,a_bc,a_b_c.\texttt{abc},\qquad \texttt{ab\_c},\qquad \texttt{ac\_b},\qquad \texttt{a\_bc},\qquad \texttt{a\_b\_c}.abc,ab_c,ac_b,a_bc,a_b_c.

For a triple t=(a,b,c)t=(a,b,c)t=(a,b,c) and a level lll, writing a^=(a,l)\hat a=(a,l)a^=(a,l), b^=(b,l)\hat b=(b,l)b^=(b,l), c^=(c,l)\hat c=(c,l)c^=(c,l), a state contributes the following edge set, all inside level lll:

stateedges contributed
abc\texttt{abc}abca^b^, b^c^, a^c^\hat a\hat b,\ \hat b\hat c,\ \hat a\hat ca^b^, b^c^, a^c^ (a triangle)
ab_c\texttt{ab\_c}ab_ca^b^\hat a\hat ba^b^ only
ac_b\texttt{ac\_b}ac_ba^c^\hat a\hat ca^c^ only
a_bc\texttt{a\_bc}a_bcb^c^\hat b\hat cb^c^ only
a_b_c\texttt{a\_b\_c}a_b_cnone

The random graph. A configuration is a function ψ\psiψ assigning a state to each of the 6×2=126 \times 2 = 126×2=12 pairs (hyperedge index, level): ψ:{0,…,5}×{0,1}→WZ\psi : \{0,\dots,5\}\times\{0,1\} \to \texttt{WZ}ψ:{0,…,5}×{0,1}→WZ. The graph GψG_\psiGψ​ on V×{0,1}V\times\{0,1\}V×{0,1} has edge set

E(Gψ)  =  ⋃i=05 ⋃l∈{0,1}edges(ψ(i,l), ti, l) ∪ {the three posts over T={1,6,8}},E(G_\psi) \;=\; \bigcup_{i=0}^{5}\ \bigcup_{l\in\{0,1\}} \texttt{edges}\bigl(\psi(i,l),\,t_i,\,l\bigr)\ \cup\ \{\text{the three posts over } T=\{1,6,8\}\},E(Gψ​)=i=0⋃5​ l∈{0,1}⋃​edges(ψ(i,l),ti​,l) ∪ {the three posts over T={1,6,8}},

read as a simple graph (loops, were there any, are discarded).

The weight functional. For an arbitrary function P:WZ→QP : \texttt{WZ} \to \mathbb{Q}P:WZ→Q (rationals, not reals — and with no positivity or normalisation built in), and for two vertices x,y∈V×{0,1}x,y \in V\times\{0,1\}x,y∈V×{0,1},

Pr⁡PWZ[x↔y]  :=  ∑ψ : {0,…,5}×{0,1}→WZ1[x and y lie in the same component of Gψ]⋅∏pP(ψ(p)),\Pr\nolimits^{\mathrm{WZ}}_P[x \leftrightarrow y] \;:=\; \sum_{\psi \,:\, \{0,\dots,5\}\times\{0,1\}\to \texttt{WZ}} \mathbf{1}\bigl[x \text{ and } y \text{ lie in the same component of } G_\psi\bigr]\cdot \prod_{p} P(\psi(p)),PrPWZ​[x↔y]:=ψ:{0,…,5}×{0,1}→WZ∑​1[x and y lie in the same component of Gψ​]⋅p∏​P(ψ(p)),

a sum over all 512=244,140,6255^{12}=244{,}140{,}625512=244,140,625 configurations. The twelve states are drawn independently: one per hyperedge per level, so the two levels are i.i.d. copies of each other, coupled only through the three always-present posts. (Connectedness is Reachable, which is reflexive; the two vertices compared below are distinct, so this is immaterial.)

What the theorem asserts

Abbreviate

x=P(abc),y=P(a_b_c),z=P(a_bc),u=P(ab_c),v=P(ac_b).x = P(\texttt{abc}),\quad y = P(\texttt{a\_b\_c}),\quad z = P(\texttt{a\_bc}),\quad u = P(\texttt{ab\_c}),\quad v = P(\texttt{ac\_b}).x=P(abc),y=P(a_b_c),z=P(a_bc),u=P(ab_c),v=P(ac_b).

For every function P:WZ→QP:\texttt{WZ}\to\mathbb{Q}P:WZ→Q satisfying all five of the following hypotheses:

  1. (nonnegativity) P(s)≥0P(s) \ge 0P(s)≥0 for each of the five states, i.e. x,y,z,u,v≥0x,y,z,u,v \ge 0x,y,z,u,v≥0;
  2. (normalisation) x+y+z+u+v=1x + y + z + u + v = 1x+y+z+u+v=1;
  3. (symmetry of the two one-edge states through the distinguished vertex) u=vu = vu=v, that is P(ab_c)=P(ac_b)P(\texttt{ab\_c}) = P(\texttt{ac\_b})P(ab_c)=P(ac_b);
  4. (strict positivity of a determinant) 0<xy−uv0 < xy - uv0<xy−uv, i.e. P(abc) P(a_b_c)−P(ab_c) P(ac_b)>0P(\texttt{abc})\,P(\texttt{a\_b\_c}) - P(\texttt{ab\_c})\,P(\texttt{ac\_b}) > 0P(abc)P(a_b_c)−P(ab_c)P(ac_b)>0;
  5. (the factor-400 bound) 400 z≤xy−uv400\,z \le xy - uv400z≤xy−uv, i.e. 400 P(a_bc)≤P(abc) P(a_b_c)−P(ab_c) P(ac_b)400\,P(\texttt{a\_bc}) \le P(\texttt{abc})\,P(\texttt{a\_b\_c}) - P(\texttt{ab\_c})\,P(\texttt{ac\_b})400P(a_bc)≤P(abc)P(a_b_c)−P(ab_c)P(ac_b) — note this is non-strict (≤)(\le)(≤), while (4) is strict (<)(<)(<);

the conclusion is the strict inequality

 Pr⁡PWZ[(0,0)↔(9,0)]  <  Pr⁡PWZ[(0,0)↔(9,1)]. \boxed{\ \Pr\nolimits^{\mathrm{WZ}}_P\bigl[(0,0) \leftrightarrow (9,0)\bigr] \;<\; \Pr\nolimits^{\mathrm{WZ}}_P\bigl[(0,0) \leftrightarrow (9,1)\bigr].\ } PrPWZ​[(0,0)↔(9,0)]<PrPWZ​[(0,0)↔(9,1)]. ​

Both probabilities have the same source (0,0)(0,0)(0,0): vertex 000 on level 000. The left-hand side is the same-level connection probability, from (0,0)(0,0)(0,0) to vertex 999 on level 000; the right-hand side is the opposite-level connection probability, from (0,0)(0,0)(0,0) to vertex 999 on level 111. The claim is that the same-level side is strictly smaller — i.e. the target in the other level is claimed strictly more likely to be connected. Neither 000 nor 999 belongs to the transversal set T={1,6,8}T=\{1,6,8\}T={1,6,8}, so (9,0)(9,0)(9,0) and (9,1)(9,1)(9,1) are genuinely distinct, unposted vertices.

Immediate consequences of the hypotheses, made explicit. Hypotheses (1)+(2) are precisely the conditions under which Pr⁡PWZ\Pr^{\mathrm{WZ}}_PPrPWZ​ is a probability (the definition carries no internal normalisation; the total mass is (∑sP(s))12(\sum_s P(s))^{12}(∑s​P(s))12, here 111). Hypothesis (4) together with (3) reads xy>u2≥0xy > u^2 \ge 0xy>u2≥0, which forces x>0x>0x>0 and y>0y>0y>0: neither P(abc)P(\texttt{abc})P(abc) nor P(a_b_c)P(\texttt{a\_b\_c})P(a_b_c) may vanish. Hypotheses (3)+(4) force u=v<1/4u=v<1/4u=v<1/4. Hypothesis (5) forces z=P(a_bc)≤(xy−u2)/400≤1/1600=0.000625z = P(\texttt{a\_bc}) \le (xy-u^2)/400 \le 1/1600 = 0.000625z=P(a_bc)≤(xy−u2)/400≤1/1600=0.000625: the state a_bc\texttt{a\_bc}a_bc (the one joining the two non-distinguished vertices of a triple) is squeezed to essentially zero. Hypothesis (5) permits z=0z=0z=0, and (1)+(3) permit u=v=0u=v=0u=v=0.

Note on the artifact. The declaration's proof is sorry — the statement is asserted, not proved, in this file.

Testimony on satisfiability (question (a)): the hypothesis set is SATISFIABLE

The hypotheses are not vacuous. A concrete witness meeting every one of the five simultaneously, verified exactly:

P(abc)=12,P(a_b_c)=12,P(a_bc)=P(ab_c)=P(ac_b)=0.P(\texttt{abc}) = \tfrac12,\quad P(\texttt{a\_b\_c}) = \tfrac12,\quad P(\texttt{a\_bc}) = P(\texttt{ab\_c}) = P(\texttt{ac\_b}) = 0.P(abc)=21​,P(a_b_c)=21​,P(a_bc)=P(ab_c)=P(ac_b)=0.

Check: all five values ≥0\ge 0≥0; sum =1=1=1; u=v=0u=v=0u=v=0 so (3) holds; xy−uv=14>0xy-uv = \tfrac14 > 0xy−uv=41​>0 so (4) holds; 400⋅0=0≤14400\cdot 0 = 0 \le \tfrac14400⋅0=0≤41​ so (5) holds.

A second witness with all five states strictly positive and (5) tight (400z=xy−uv400z = xy-uv400z=xy−uv exactly):

P(abc)=12,P(a_b_c)=1200140050,P(a_bc)=720025,P(ab_c)=P(ac_b)=110,P(\texttt{abc}) = \tfrac12,\quad P(\texttt{a\_b\_c}) = \tfrac{12001}{40050},\quad P(\texttt{a\_bc}) = \tfrac{7}{20025},\quad P(\texttt{ab\_c}) = P(\texttt{ac\_b}) = \tfrac{1}{10},P(abc)=21​,P(a_b_c)=4005012001​,P(a_bc)=200257​,P(ab_c)=P(ac_b)=101​,

for which xy−uv=112801xy-uv = \tfrac{112}{801}xy−uv=801112​ and 400z=112801400z = \tfrac{112}{801}400z=801112​ exactly. The feasible region is in fact a full-dimensional open-ish set: it is the set of (x,y,z,u)(x,y,z,u)(x,y,z,u) with u=vu=vu=v, x+y+z+2u=1x+y+z+2u=1x+y+z+2u=1, x,y,z,u≥0x,y,z,u\ge0x,y,z,u≥0, xy>u2xy>u^2xy>u2, 400z≤xy−u2400z \le xy-u^2400z≤xy−u2.

Testimony on truth (question (b)): NO COUNTEREXAMPLE FOUND — the conclusion held on every feasible point tested

I reimplemented the model from the two definition files in Python and computed the two probabilities exactly over Q\mathbb{Q}Q. The reimplementation exploits the fact that the only level-to-level edges are the three posts at {1,6,8}\{1,6,8\}{1,6,8}, so each level's 56=15,6255^6=15{,}62556=15,625 state-vectors need only be reduced to the partition it induces on the five terminals {0,1,6,8,9}\{0,1,6,8,9\}{0,1,6,8,9} (47 distinct partitions arise); the two levels are then glued through the posts. This reduction was cross-validated against brute-force union-find on the full 20-vertex graph for 20,000 random configurations: 0 mismatches. Total level mass came out to exactly 111 on a test distribution, and the level-weight polynomial summed to exactly (x+y+z+2u)6(x+y+z+2u)^6(x+y+z+2u)6.

Writing Δ(P):=Pr⁡PWZ[(0,0)↔(9,1)]−Pr⁡PWZ[(0,0)↔(9,0)]\Delta(P) := \Pr^{\mathrm{WZ}}_P[(0,0)\leftrightarrow(9,1)] - \Pr^{\mathrm{WZ}}_P[(0,0)\leftrightarrow(9,0)]Δ(P):=PrPWZ​[(0,0)↔(9,1)]−PrPWZ​[(0,0)↔(9,0)] (the conclusion says Δ>0\Delta > 0Δ>0), I obtained Δ\DeltaΔ as an exact integer polynomial, homogeneous of degree 121212, with 231 monomials in (x,y,z,u)(x,y,z,u)(x,y,z,u) — and it satisfies the identity

Δ  =  (xy−u2)6  +  z⋅Q(x,y,z,u)\Delta \;=\; (xy - u^2)^6 \;+\; z\cdot Q(x,y,z,u)Δ=(xy−u2)6+z⋅Q(x,y,z,u)

for an integer polynomial QQQ of degree 111111 with mixed signs. On the slice z=u=0z=u=0z=u=0 it collapses to exactly Δ=x6y6\Delta = x^6 y^6Δ=x6y6.

Search performed. Exhaustive-in-practice sweeps of the feasible region, parametrised so that hypotheses (1)–(3) and (5) hold exactly in rational arithmetic (for fixed rational uuu and xxx, the level of (5)-tightness θ∈[0,1]\theta \in [0,1]θ∈[0,1] gives the rational z=θA/(400+θx)z = \theta A/(400+\theta x)z=θA/(400+θx) with A=x(1−2u−x)−u2A = x(1-2u-x)-u^2A=x(1−2u−x)−u2, and y=1−z−2u−xy = 1-z-2u-xy=1−z−2u−x):

  • ≈958,000\approx 958{,}000≈958,000 points in double precision over u∈[0,1/4)u \in [0,1/4)u∈[0,1/4), the whole admissible xxx-interval, and θ∈{0,0.2,…,1}\theta \in \{0,0.2,\dots,1\}θ∈{0,0.2,…,1};
  • ≈42,000\approx 42{,}000≈42,000 points re-evaluated in exact rational arithmetic, including a dedicated fine sweep of the region u→1/4u \to 1/4u→1/4 and of the regime xy−u2→0+xy-u^2 \to 0^+xy−u2→0+ (down to xy−u2≈3×10−15xy-u^2 \approx 3\times10^{-15}xy−u2≈3×10−15, where Δ≈6×10−88\Delta \approx 6\times 10^{-88}Δ≈6×10−88), plus 15,000 randomised exact samples.

Result: no hypothesis-satisfying assignment was found on which the two probabilities are equal or reversed. Across all exact points, Δ>0\Delta > 0Δ>0 and the ratio Δ/(xy−u2)6\Delta/(xy-u^2)^6Δ/(xy−u2)6 stayed in the narrow band [0.990, 1.005][0.990,\,1.005][0.990,1.005]; the global exact minimum of Δ\DeltaΔ encountered was ≈1.0×10−35\approx 1.0\times10^{-35}≈1.0×10−35 (at a near-degenerate point), always strictly positive. The double-precision sweep did flag 706 apparently-negative points, but every one had ∣Δ∣∼10−22|\Delta|\sim10^{-22}∣Δ∣∼10−22 against (xy−u2)6∼10−31(xy-u^2)^6 \sim 10^{-31}(xy−u2)6∼10−31 — pure catastrophic cancellation — and exact rational re-evaluation of that entire neighbourhood returned Δ>0\Delta>0Δ>0 at all 7,080 exact points, with minimum ratio 0.99270.99270.9927. So there is no critical finding to report: I found no violation.

Exact values at the witnesses.

  • At x=y=12x=y=\tfrac12x=y=21​, z=u=v=0z=u=v=0z=u=v=0:
Pr⁡[(0,0)↔(9,0)]=8174096=0.199462890625,Pr⁡[(0,0)↔(9,1)]=4092048=8184096=0.199707031250,\Pr[(0,0)\leftrightarrow(9,0)] = \frac{817}{4096} = 0.199462890625,\qquad \Pr[(0,0)\leftrightarrow(9,1)] = \frac{409}{2048} = \frac{818}{4096} = 0.199707031250,Pr[(0,0)↔(9,0)]=4096817​=0.199462890625,Pr[(0,0)↔(9,1)]=2048409​=4096818​=0.199707031250,

so Δ=14096>0\Delta = \tfrac{1}{4096} > 0Δ=40961​>0: the conclusion holds, with the opposite-level probability larger by exactly one part in 409640964096.

  • At the all-five-positive, (5)-tight witness above: Pr⁡[same]≈0.3213659114\Pr[\text{same}] \approx 0.3213659114Pr[same]≈0.3213659114, Pr⁡[opp]≈0.3213733657\Pr[\text{opp}] \approx 0.3213733657Pr[opp]≈0.3213733657, Δ≈7.454×10−6>0\Delta \approx 7.454\times10^{-6} > 0Δ≈7.454×10−6>0.
  • At x=12x=\tfrac12x=21​, y=400801y=\tfrac{400}{801}y=801400​, z=11602z=\tfrac{1}{1602}z=16021​, u=v=0u=v=0u=v=0 (hypothesis (5) tight, u=v=0u=v=0u=v=0): Pr⁡[same]≈0.1997950647\Pr[\text{same}] \approx 0.1997950647Pr[same]≈0.1997950647, Pr⁡[opp]≈0.2000367635\Pr[\text{opp}] \approx 0.2000367635Pr[opp]≈0.2000367635, Δ≈2.417×10−4>0\Delta \approx 2.417\times10^{-4} > 0Δ≈2.417×10−4>0.

Boundary behaviour of the numerical hypotheses

  • Hypothesis (4) is exactly tight. At x=y=u=v=14x=y=u=v=\tfrac14x=y=u=v=41​, z=0z=0z=0 — which satisfies (1), (2), (3) and (5) but makes xy−uv=0xy-uv = 0xy−uv=0, so (4) fails by the barest possible margin — the two probabilities are exactly equal:
Pr⁡[(0,0)↔(9,0)]=Pr⁡[(0,0)↔(9,1)]=7343774194304.\Pr[(0,0)\leftrightarrow(9,0)] = \Pr[(0,0)\leftrightarrow(9,1)] = \frac{734377}{4194304}.Pr[(0,0)↔(9,0)]=Pr[(0,0)↔(9,1)]=4194304734377​.

Thus the strict inequality in (4) cannot be weakened to ≤\le≤ without making the strict conclusion false; and generally Δ≈(xy−u2)6→0\Delta \approx (xy-u^2)^6 \to 0Δ≈(xy−u2)6→0 as (4) degenerates, so the conclusion's margin becomes arbitrarily small while remaining positive.

  • Hypothesis (5) is essential, not cosmetic. Dropping only (5) while keeping (1)–(4), a random sample of 40,000 normalised assignments produced 17,256 with Δ<0\Delta < 0Δ<0 (worst observed Δ≈−0.745\Delta \approx -0.745Δ≈−0.745), e.g. x≈0.0767x\approx0.0767x≈0.0767, y≈0.0115y\approx0.0115y≈0.0115, z≈0.8888z\approx0.8888z≈0.8888, u=v≈0.0115u=v\approx0.0115u=v≈0.0115. So the factor 400400400 is doing real work: without it the stated inequality reverses on a large part of the simplex.
  • Zero-probability states. P(a_bc)=0P(\texttt{a\_bc})=0P(a_bc)=0 and P(ab_c)=P(ac_b)=0P(\texttt{ab\_c})=P(\texttt{ac\_b})=0P(ab_c)=P(ac_b)=0 are both permitted and occur in the witnesses above. P(abc)=0P(\texttt{abc})=0P(abc)=0 or P(a_b_c)=0P(\texttt{a\_b\_c})=0P(a_b_c)=0 is impossible under (4). No state may have probability 111.
  • Range of u=vu=vu=v. Feasibility forces u=v<1/4u=v < 1/4u=v<1/4; as u→1/4u \to 1/4u→1/4 the admissible xxx-window shrinks to the single point x=1/4x=1/4x=1/4 and Δ→0\Delta \to 0Δ→0.
Human review
  • Endorsed by Shuze Chen · Sep 11, 2026

  • Endorsed by burkh4rt · Sep 11, 2026

    Confirmed by the mission captain (proposal self-audit).

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