\documentclass[11pt]{article} \pdfgentounicode=1 \input{glyphtounicode} \usepackage{cmap} \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage{lmodern} \usepackage{textcomp} \usepackage{amsmath} \usepackage{amssymb} \usepackage[letterpaper,top=0.8in,bottom=0.95in,left=0.85in,right=0.85in]{geometry} \usepackage{microtype} \usepackage{array} \usepackage{longtable} \usepackage{enumitem} \usepackage{pict2e} \usepackage{xcolor} \usepackage{accsupp} \usepackage{url} \usepackage[unicode,colorlinks=true,linkcolor=black,urlcolor=blue!60!black,bookmarksopen=true,pdftitle={Homological percolation in characteristic two: the Ising gauge case},pdfauthor={Tejasvi Singh Tomar}]{hyperref} \urlstyle{same} \renewcommand{\UrlBreaks}{} \renewcommand{\UrlBigBreaks}{} \newcommand{\cbrk}{\penalty5000\relax} \newcommand{\rowrule}{\noalign{\hrule height\arrayrulewidth\penalty0}} \newcommand{\lastrule}{\noalign{\hrule height\arrayrulewidth}} \frenchspacing \setlength{\parindent}{0pt} \setlength{\parskip}{5pt plus 2pt minus 1pt} \hyphenpenalty=10000 \sloppy \setlength{\emergencystretch}{3em} \DeclareFontFamily{U}{msa}{} \DeclareFontShape{U}{msa}{m}{n}{<-> msam10}{} \DeclareFontFamily{U}{msb}{} \DeclareFontShape{U}{msb}{m}{n}{<-> msbm10}{} \DeclareFontFamily{U}{euf}{} \DeclareFontShape{U}{euf}{m}{n}{<-> eufm10}{} \DeclareFontShape{U}{euf}{b}{n}{<-> eufb10}{} \makeatletter \newcommand{\Sup}[1]{\raisebox{\dimexpr\f@size pt/3+0.75pt\relax}{\fontsize{\strip@pt\dimexpr0.72\dimexpr\f@size pt\relax\relax}{\z@}\selectfont #1}} \newcommand{\Sub}[1]{\raisebox{-\dimexpr\f@size pt/5+0.75pt\relax}{\fontsize{\strip@pt\dimexpr0.72\dimexpr\f@size pt\relax\relax}{\z@}\selectfont #1}} \makeatother \newcommand{\uni}[2]{\leavevmode\BeginAccSupp{method=hex,unicode,ActualText=#1}\textcompwordmark#2\textcompwordmark\EndAccSupp{}} \newcommand{\discbullet}{\raisebox{0.15ex}{\setlength{\unitlength}{1pt}\begin{picture}(4.4,4.4)\put(2.2,2.2){\circle*{4.4}}\end{picture}}} \newcommand{\circbullet}{\raisebox{0.15ex}{\setlength{\unitlength}{1pt}\begin{picture}(4.4,4.4)\linethickness{0.5pt}\put(2.2,2.2){\circle{4}}\end{picture}}} \newcommand{\squarebullet}{\raisebox{0.15ex}{\rule{3.8pt}{3.8pt}}} \setlist{topsep=3pt,itemsep=1.5pt,parsep=3pt} \DeclareUnicodeCharacter{00A7}{\uni{00A7}{\S{}}} \DeclareUnicodeCharacter{00AC}{\uni{00AC}{\ensuremath{\neg}}} \DeclareUnicodeCharacter{00B0}{\uni{00B0}{\textdegree{}}} \DeclareUnicodeCharacter{00B1}{\uni{00B1}{\ensuremath{\pm}}} \DeclareUnicodeCharacter{00B2}{\uni{00B2}{\texttwosuperior{}}} \DeclareUnicodeCharacter{00B3}{\uni{00B3}{\textthreesuperior{}}} \DeclareUnicodeCharacter{00B7}{\uni{00B7}{\ensuremath{\cdot}}} \DeclareUnicodeCharacter{00B9}{\uni{00B9}{\textonesuperior{}}} \DeclareUnicodeCharacter{00BD}{\uni{00BD}{\textonehalf{}}} \DeclareUnicodeCharacter{00D7}{\uni{00D7}{\ensuremath{\times}}} \DeclareUnicodeCharacter{00DB}{\uni{00DB}{\^{U}}} \DeclareUnicodeCharacter{00E0}{\uni{00E0}{\`{a}}} \DeclareUnicodeCharacter{00E1}{\uni{00E1}{\'{a}}} \DeclareUnicodeCharacter{00E4}{\uni{00E4}{\"{a}}} \DeclareUnicodeCharacter{00E9}{\uni{00E9}{\'{e}}} \DeclareUnicodeCharacter{00F4}{\uni{00F4}{\^{o}}} \DeclareUnicodeCharacter{00F6}{\uni{00F6}{\"{o}}} \DeclareUnicodeCharacter{00FB}{\uni{00FB}{\^{u}}} \DeclareUnicodeCharacter{00FC}{\uni{00FC}{\"{u}}} \DeclareUnicodeCharacter{0177}{\uni{0177}{\^{y}}} \DeclareUnicodeCharacter{0393}{\uni{0393}{\ensuremath{\Gamma}}} \DeclareUnicodeCharacter{0394}{\uni{0394}{\ensuremath{\Delta}}} \DeclareUnicodeCharacter{0398}{\uni{0398}{\ensuremath{\Theta}}} \DeclareUnicodeCharacter{039B}{\uni{039B}{\ensuremath{\Lambda}}} \DeclareUnicodeCharacter{039E}{\uni{039E}{\ensuremath{\Xi}}} \DeclareUnicodeCharacter{03A0}{\uni{03A0}{\ensuremath{\Pi}}} \DeclareUnicodeCharacter{03A3}{\uni{03A3}{\ensuremath{\Sigma}}} \DeclareUnicodeCharacter{03A6}{\uni{03A6}{\ensuremath{\Phi}}} \DeclareUnicodeCharacter{03A8}{\uni{03A8}{\ensuremath{\Psi}}} \DeclareUnicodeCharacter{03A9}{\uni{03A9}{\ensuremath{\Omega}}} \DeclareUnicodeCharacter{03B1}{\uni{03B1}{\ensuremath{\alpha}}} \DeclareUnicodeCharacter{03B2}{\uni{03B2}{\ensuremath{\beta}}} \DeclareUnicodeCharacter{03B3}{\uni{03B3}{\ensuremath{\gamma}}} \DeclareUnicodeCharacter{03B4}{\uni{03B4}{\ensuremath{\delta}}} \DeclareUnicodeCharacter{03B5}{\uni{03B5}{\ensuremath{\varepsilon}}} \DeclareUnicodeCharacter{03B6}{\uni{03B6}{\ensuremath{\zeta}}} \DeclareUnicodeCharacter{03B7}{\uni{03B7}{\ensuremath{\eta}}} \DeclareUnicodeCharacter{03B8}{\uni{03B8}{\ensuremath{\theta}}} \DeclareUnicodeCharacter{03B9}{\uni{03B9}{\ensuremath{\iota}}} \DeclareUnicodeCharacter{03BA}{\uni{03BA}{\ensuremath{\kappa}}} \DeclareUnicodeCharacter{03BB}{\uni{03BB}{\ensuremath{\lambda}}} \DeclareUnicodeCharacter{03BC}{\uni{03BC}{\ensuremath{\mu}}} \DeclareUnicodeCharacter{03BD}{\uni{03BD}{\ensuremath{\nu}}} \DeclareUnicodeCharacter{03BE}{\uni{03BE}{\ensuremath{\xi}}} \DeclareUnicodeCharacter{03C0}{\uni{03C0}{\ensuremath{\pi}}} \DeclareUnicodeCharacter{03C1}{\uni{03C1}{\ensuremath{\rho}}} \DeclareUnicodeCharacter{03C2}{\uni{03C2}{\ensuremath{\varsigma}}} \DeclareUnicodeCharacter{03C3}{\uni{03C3}{\ensuremath{\sigma}}} \DeclareUnicodeCharacter{03C4}{\uni{03C4}{\ensuremath{\tau}}} \DeclareUnicodeCharacter{03C6}{\uni{03C6}{\ensuremath{\varphi}}} \DeclareUnicodeCharacter{03C7}{\uni{03C7}{\ensuremath{\chi}}} \DeclareUnicodeCharacter{03C8}{\uni{03C8}{\ensuremath{\psi}}} \DeclareUnicodeCharacter{03C9}{\uni{03C9}{\ensuremath{\omega}}} \DeclareUnicodeCharacter{03D6}{\uni{03D6}{\ensuremath{\varpi}}} \DeclareUnicodeCharacter{1D62}{\uni{1D62}{\Sub{i}}} \DeclareUnicodeCharacter{2013}{\uni{2013}{\textendash{}}} \DeclareUnicodeCharacter{2014}{\uni{2014}{\textemdash{}}} \DeclareUnicodeCharacter{2016}{\uni{2016}{\ensuremath{\|}}} \DeclareUnicodeCharacter{2020}{\uni{2020}{\textdagger{}}} \DeclareUnicodeCharacter{2022}{\uni{2022}{\textbullet{}}} \DeclareUnicodeCharacter{2026}{\uni{2026}{\ldots{}}} \DeclareUnicodeCharacter{2032}{\uni{2032}{\ensuremath{{}^{\prime}}}} \DeclareUnicodeCharacter{2033}{\uni{2033}{\ensuremath{{}^{\prime\prime}}}} \DeclareUnicodeCharacter{2070}{\uni{2070}{\Sup{0}}} \DeclareUnicodeCharacter{2074}{\uni{2074}{\Sup{4}}} \DeclareUnicodeCharacter{2075}{\uni{2075}{\Sup{5}}} \DeclareUnicodeCharacter{2076}{\uni{2076}{\Sup{6}}} \DeclareUnicodeCharacter{2077}{\uni{2077}{\Sup{7}}} \DeclareUnicodeCharacter{2078}{\uni{2078}{\Sup{8}}} \DeclareUnicodeCharacter{2079}{\uni{2079}{\Sup{9}}} \DeclareUnicodeCharacter{207B}{\uni{207B}{\Sup{\ensuremath{-}}}} \DeclareUnicodeCharacter{207F}{\uni{207F}{\Sup{n}}} \DeclareUnicodeCharacter{2080}{\uni{2080}{\Sub{0}}} \DeclareUnicodeCharacter{2081}{\uni{2081}{\Sub{1}}} \DeclareUnicodeCharacter{2082}{\uni{2082}{\Sub{2}}} \DeclareUnicodeCharacter{2083}{\uni{2083}{\Sub{3}}} \DeclareUnicodeCharacter{2084}{\uni{2084}{\Sub{4}}} \DeclareUnicodeCharacter{2085}{\uni{2085}{\Sub{5}}} \DeclareUnicodeCharacter{2086}{\uni{2086}{\Sub{6}}} \DeclareUnicodeCharacter{2087}{\uni{2087}{\Sub{7}}} \DeclareUnicodeCharacter{2088}{\uni{2088}{\Sub{8}}} \DeclareUnicodeCharacter{2089}{\uni{2089}{\Sub{9}}} \DeclareUnicodeCharacter{2102}{\uni{2102}{\ensuremath{\mathbb{C}}}} \DeclareUnicodeCharacter{210B}{\uni{210B}{\ensuremath{\mathcal{H}}}} \DeclareUnicodeCharacter{2113}{\uni{2113}{\ensuremath{\ell}}} \DeclareUnicodeCharacter{212C}{\uni{212C}{\ensuremath{\mathcal{B}}}} \DeclareUnicodeCharacter{2131}{\uni{2131}{\ensuremath{\mathcal{F}}}} \DeclareUnicodeCharacter{2190}{\uni{2190}{\ensuremath{\leftarrow}}} \DeclareUnicodeCharacter{2191}{\uni{2191}{\ensuremath{\uparrow}}} \DeclareUnicodeCharacter{2192}{\uni{2192}{\ensuremath{\rightarrow}}} \DeclareUnicodeCharacter{2193}{\uni{2193}{\ensuremath{\downarrow}}} \DeclareUnicodeCharacter{2194}{\uni{2194}{\ensuremath{\leftrightarrow}}} \DeclareUnicodeCharacter{21A6}{\uni{21A6}{\ensuremath{\mapsto}}} \DeclareUnicodeCharacter{2202}{\uni{2202}{\ensuremath{\partial}}} \DeclareUnicodeCharacter{2205}{\uni{2205}{\ensuremath{\emptyset}}} \DeclareUnicodeCharacter{2208}{\uni{2208}{\ensuremath{\in}}} \DeclareUnicodeCharacter{2209}{\uni{2209}{\ensuremath{\notin}}} \DeclareUnicodeCharacter{220B}{\uni{220B}{\ensuremath{\ni}}} \DeclareUnicodeCharacter{220C}{\uni{220C}{\ensuremath{\not\ni}}} \DeclareUnicodeCharacter{220E}{\uni{220E}{\ensuremath{\blacksquare}}} \DeclareUnicodeCharacter{2212}{\uni{2212}{\ensuremath{-}}} \DeclareUnicodeCharacter{2216}{\uni{2216}{\ensuremath{\setminus}}} \DeclareUnicodeCharacter{221A}{\uni{221A}{\ensuremath{\surd}}} \DeclareUnicodeCharacter{221D}{\uni{221D}{\ensuremath{\propto}}} \DeclareUnicodeCharacter{221E}{\uni{221E}{\ensuremath{\infty}}} \DeclareUnicodeCharacter{2227}{\uni{2227}{\ensuremath{\wedge}}} \DeclareUnicodeCharacter{2229}{\uni{2229}{\ensuremath{\cap}}} \DeclareUnicodeCharacter{222A}{\uni{222A}{\ensuremath{\cup}}} \DeclareUnicodeCharacter{2245}{\uni{2245}{\ensuremath{\cong}}} \DeclareUnicodeCharacter{2248}{\uni{2248}{\ensuremath{\approx}}} \DeclareUnicodeCharacter{2260}{\uni{2260}{\ensuremath{\neq}}} \DeclareUnicodeCharacter{2261}{\uni{2261}{\ensuremath{\equiv}}} \DeclareUnicodeCharacter{2262}{\uni{2262}{\ensuremath{\not\equiv}}} \DeclareUnicodeCharacter{2264}{\uni{2264}{\ensuremath{\leq}}} \DeclareUnicodeCharacter{2265}{\uni{2265}{\ensuremath{\geq}}} \DeclareUnicodeCharacter{2272}{\uni{2272}{\ensuremath{\lesssim}}} \DeclareUnicodeCharacter{2282}{\uni{2282}{\ensuremath{\subset}}} \DeclareUnicodeCharacter{2284}{\uni{2284}{\ensuremath{\not\subset}}} \DeclareUnicodeCharacter{2286}{\uni{2286}{\ensuremath{\subseteq}}} \DeclareUnicodeCharacter{2287}{\uni{2287}{\ensuremath{\supseteq}}} \DeclareUnicodeCharacter{2297}{\uni{2297}{\ensuremath{\otimes}}} \DeclareUnicodeCharacter{229B}{\uni{229B}{\ensuremath{\circledast}}} \DeclareUnicodeCharacter{22A5}{\uni{22A5}{\ensuremath{\perp}}} \DeclareUnicodeCharacter{22C2}{\uni{22C2}{\ensuremath{\bigcap}}} \DeclareUnicodeCharacter{22C3}{\uni{22C3}{\ensuremath{\bigcup}}} \DeclareUnicodeCharacter{22EF}{\uni{22EF}{\ensuremath{\cdots}}} \DeclareUnicodeCharacter{2308}{\uni{2308}{\ensuremath{\lceil}}} \DeclareUnicodeCharacter{2309}{\uni{2309}{\ensuremath{\rceil}}} \DeclareUnicodeCharacter{230A}{\uni{230A}{\ensuremath{\lfloor}}} \DeclareUnicodeCharacter{230B}{\uni{230B}{\ensuremath{\rfloor}}} \DeclareUnicodeCharacter{25A1}{\uni{25A1}{\ensuremath{\square}}} \DeclareUnicodeCharacter{266F}{\uni{266F}{\ensuremath{\sharp}}} \DeclareUnicodeCharacter{27E8}{\uni{27E8}{\ensuremath{\langle}}} \DeclareUnicodeCharacter{27E9}{\uni{27E9}{\ensuremath{\rangle}}} \DeclareUnicodeCharacter{27FA}{\uni{27FA}{\ensuremath{\Longleftrightarrow}}} \DeclareUnicodeCharacter{2C7C}{\uni{2C7C}{\Sub{j}}} \DeclareUnicodeCharacter{1D49C}{\uni{D835DC9C}{\ensuremath{\mathcal{A}}}} \DeclareUnicodeCharacter{1D4A6}{\uni{D835DCA6}{\ensuremath{\mathcal{K}}}} \DeclareUnicodeCharacter{1D4A9}{\uni{D835DCA9}{\ensuremath{\mathcal{N}}}} \DeclareUnicodeCharacter{1D4AB}{\uni{D835DCAB}{\ensuremath{\mathcal{P}}}} \DeclareUnicodeCharacter{1D4D4}{\uni{D835DCD4}{\ensuremath{\boldsymbol{\mathcal{E}}}}} \DeclareUnicodeCharacter{1D505}{\uni{D835DD05}{\ensuremath{\mathfrak{B}}}} \DeclareUnicodeCharacter{1D507}{\uni{D835DD07}{\ensuremath{\mathfrak{D}}}} \begin{document} \raggedright \noindent{\LARGE\bfseries Homological percolation in characteristic two: the Ising gauge case\par} \emph{Tejasvi Singh Tomar. Version 7, 2026-10-27.} \textbf{Abstract.} Duncan and Schweinhart proved that the i-dimensional plaquette random-cluster model with cluster weight q on the 2i-dimensional torus has a sharp homological percolation transition at the self-dual point, for coefficients in a field of characteristic other than 2. Their proof needs the point group of the lattice to act irreducibly on the i-th homology of the torus. In characteristic 2 it does not, and the most important case is excluded: q = 2 with F₂ coefficients, which represents the (i−1)-form Ising lattice gauge theory. We prove the theorem for q = 2\Sup{r} with r ≥ 1 and F₂ coefficients. The representation-theoretic step is replaced by a transfer-matrix inequality, the classical positivity of the partition functions of \textquotesingle{}t Hooft\textquotesingle{}s electric-flux sectors; read in the random-cluster representation, this inequality compares the probabilities of homological events. As a corollary, for the ferromagnetic (i−1)-form Ising gauge theory on the 2i-dimensional torus (Wegner\textquotesingle{}s model when i = 2), the free energy of every nontrivial \textquotesingle{}t Hooft twist tends to 0 below the self-dual coupling and to infinity above it as the torus grows. For Bernoulli plaquette percolation (q = 1) the characteristic-2 case remains open for i ≥ 2; we describe obstructions to proving it, and on the 4-torus we prove a computer-assisted divisibility theorem for the classes of integral giant cycles just above p = 1/2. \par\medskip\noindent{\color{black!25}\rule{\linewidth}{0.6pt}}\par\medskip \section*{\pdfbookmark[1]{1. Introduction and statement of results}{h1}1. Introduction and statement of results} Plaquette percolation and the plaquette random-cluster model extend bond percolation and the random-cluster model to i-dimensional cells [HS16, DKS20, DS22], and on a torus the natural question is homological: when do the open i-cells carry cycles that are nontrivial in the homology of the torus? For q = 2\Sup{r} and coefficients in F₂, the random-cluster model represents a lattice gauge theory of (i−1)-forms with gauge group F₂\Sup{r}, and the question is then one about \textquotesingle{}t Hooft\textquotesingle{}s twisted boundary conditions [\textquotesingle{}tH79]: the ratio of the gauge theory\textquotesingle{}s partition functions with and without a twist by a mod-2 cocycle is the probability that the twist vanishes on every cycle of open plaquettes. As the torus grows, a nontrivial twist therefore costs no free energy if, with probability tending to 1, these cycles are all trivial in homology, and infinite free energy if, with probability tending to 1, they fill every homology class. For q = 2 the gauge theory is the Ising lattice gauge theory, Wegner\textquotesingle{}s Z₂ gauge theory when i = 2. Duncan and Schweinhart [DS22] proved that on the 2i-dimensional torus the transition is sharp and sits at the self-dual point p\Sub{sd} = √q/(1+√q): with probability tending to 1, below p\Sub{sd} no cycle of open plaquettes is nontrivial in homology, and above p\Sub{sd} every homology class contains one. Their theorem is stated for coefficient fields F with char F ≠ 2, since the proof relies on the irreducibility of H\Sub{i}(T\Sup{2i}; F) under the point group of the lattice (the signed permutations of the coordinates), and in characteristic 2 this representation is reducible (§4). The most important case not covered was therefore the Ising gauge theory itself, q = 2 with F₂ coefficients; for Bernoulli plaquette percolation (q = 1), the question whether characteristic 2 can be included was raised by Duncan, Kahle and Schweinhart [DKS20, §9]. Theorem A below settles the cases q = 2\Sup{r}, r ≥ 1, with F₂ coefficients, and Corollary B states the consequence for the Ising gauge theory. \subsection*{\pdfbookmark[2]{1.1 Setting and main results}{h2}1.1 Setting and main results} Throughout, X = T\Sup{d}\Sub{N} is the cubical d-torus of side N ≥ 2, with d = 2i unless stated otherwise. For x ∈ (Z/NZ)\Sup{d} and I ⊂ [d] = \{1, …, d\}, the cell c(x, I) = x + Σ\Sub{a∈I}[0, 1]e\Sub{a} has \emph{type} I and dimension \textbar{}I\textbar{}, and X\Sup{(k)} is the set of k-cells. Unless stated otherwise, coefficient groups are F₂-vector spaces, so orientations play no role; Proposition 10 and §8 use other coefficients and fix an orientation of every cell. A configuration ω ⊂ X\Sup{(i)} is a set of open i-cells (plaquettes), and P = P(ω) is the full (i−1)-skeleton together with ω. For q ≥ 1 and p ∈ (0, 1), the plaquette random-cluster model with F₂ coefficients [HS16, DS22] is the measure μ(ω) = μ\Sub{N,p,q}(ω) ∝ p\Sup{\textbar{}ω\textbar{}} (1−p)\Sup{\textbar{}X\Sup{(i)}\textbar{}−\textbar{}ω\textbar{}} q\Sup{b\Sub{i−1}(P; F₂)}. We write H := H\Sub{i}(X; F₂) ≅ F₂\Sup{K} with K = C(d, i); its basis is e\Sub{I} = [T\Sub{I}], where T\Sub{I} is the coordinate i-torus in the directions I. Let φ: P → X be the inclusion. The giant cycles are the i-cycles of P whose classes in H are nonzero, and V(P) := im(φ\Sub{*}: H\Sub{i}(P; F₂) → H) is the subspace they span. Consider the events A := \{V(P) ≠ 0\} and S := \{V(P) = H\}. \textbf{Theorem A.} Let i ≥ 1, d = 2i, q = 2\Sup{r} with r ≥ 1, and p\Sub{sd} = √q/(1+√q). Then, as N → ∞, μ\Sub{N,p,q}(A) → 0 for p \textless{} p\Sub{sd}, μ\Sub{N,p,q}(S) → 1 for p \textgreater{} p\Sub{sd}. The analogue of Theorem A with coefficients in a field F of characteristic other than 2 is [DS22, Theorem 8]. For i = 1, Theorem A is the torus form of p\Sub{c} = p\Sub{sd} for the planar random-cluster model [BDC12]. The new case is F = F₂, i ≥ 2, and q = 2\Sup{r} with r ≥ 1. \textbf{Corollary B.} Consider the (i−1)-form Ising gauge theory on T\Sup{2i}\Sub{N}. Its spins are σ ∈ \{±1\}\Sup{X\Sup{(i−1)}}, and its weight is exp(J Σ\Sub{c} σ\Sub{c}), where σ\Sub{c} is the product of the spins on the boundary of the i-cell c. For a class n ∈ H\Sup{i}(X; F₂), let Z\Sub{J}(n) be the partition function with the couplings of the i-cells in a representative cocycle of n reversed. Then, with J\Sub{sd} = ½ ln(1+√2), for every n ≠ 0: Z\Sub{J}(n)/Z\Sub{J}(0) → 1 for 0 ≤ J \textless{} J\Sub{sd}, Z\Sub{J}(n)/Z\Sub{J}(0) → 0 for J \textgreater{} J\Sub{sd}. As N → ∞, then, the free energy −log(Z\Sub{J}(n)/Z\Sub{J}(0)) of every nontrivial twist tends to 0 for 0 ≤ J \textless{} J\Sub{sd} and to ∞ for J \textgreater{} J\Sub{sd}. \subsection*{\pdfbookmark[2]{1.2 Idea of the proof}{h3}1.2 Idea of the proof} For n ∈ H\Sup{i}(X; F₂) let R(n) be the probability that n vanishes on V(P); for q = 2, J ≥ 0 and p = 1 − e\Sup{−2J} it is the ratio Z\Sub{J}(n)/Z\Sub{J}(0) of Corollary B (Lemma 3). The proof of [DS22] uses irreducibility to bound μ(S) from below by a constant times a power of μ(A) [DS22, Lemma 41]. For q = 2\Sup{r} we obtain such a bound from one inequality instead, Lemma 4: if n is written in the basis dual to (e\Sub{I}), deleting its components on the coordinate tori that contain a fixed direction can only increase R(n). Taking that direction as time, Lemma 4 says that adding a temporal twist to a spatial one cannot increase the twisted partition function. In \textquotesingle{}t Hooft\textquotesingle{}s terms, the difference is, up to a factor 2, the partition function of the odd electric-flux sector, which is nonnegative; we prove this with a transfer matrix (§3), as in [BS83, TY85]. With the FKG inequality and the symmetry of the torus under coordinate permutations, Lemma 4 gives the spin-up bound of Lemma 5, μ(S) ≥ (μ(A)/K)\Sup{2\Sup{K}−1}, without representation theory. The duality and influence inequalities of [DS22], which hold in every characteristic, then complete the proof (§5). For q = 2\Sup{r} and F₂ coefficients the same argument, with duality between degrees i and d − i, gives a sharp threshold for A and S in every dimension d ≥ 2 and degree 1 ≤ i ≤ d − 1, around the value λ(N) of p at which μ(A) = 1/2 (Theorem C). \subsection*{\pdfbookmark[2]{1.3 Bernoulli plaquette percolation (q = 1)}{h4}1.3 Bernoulli plaquette percolation (q = 1)} For q = 1 there is no gauge theory and the inequality of Lemma 4 is not known, and for i ≥ 2 it remains open whether the conclusion of Theorem A holds with F₂ coefficients. Section 8 proves a theorem about integral coefficients instead. For Bernoulli plaquette percolation P on T⁴\Sub{N} (q = 1, i = 2), let Λ\Sub{P} ⊆ H₂(T⁴\Sub{N}; Z) ≅ Z⁶ be the lattice of classes of integral 2-cycles of P. Theorem W states that for every integer M ≥ 2 there is w\Sub{M} \textgreater{} 0 such that, for p ∈ (1/2, 1/2 + w\Sub{M}), with probability tending to 1, rank Λ\Sub{P} = 6 and Λ\Sub{P} ⊆ MZ⁶; the rank statement comes from [DKS20], and the new part is the divisibility. The proof is an Aizenman–Grimmett strict-shift argument [AG91, BBR14] whose local step rests on finite computer checks and on an induction (§§8.6–8.7). Sections 7 and 9 give bounds and obstructions for the F₂ question at q = 1. Among them is a family of product measures on 4-tori of side divisible by 8 that is invariant under coordinate permutations and exactly self-dual (§9.2). With high probability its giant-cycle space V equals one fixed subspace for all parameters in an interval around its self-dual point, and this subspace is Lagrangian for the intersection form. A proof at q = 1 therefore has to exploit a property that this family does not have, such as invariance under unit translations. \subsection*{\pdfbookmark[2]{1.4 Organization}{h5}1.4 Organization} Sections 2–5 prove Theorem A, Corollary B and Theorem C; with the remarks of §6 they can be read on their own. Section 7 treats q = 1 and other real q, Section 8 proves Theorem W, and Section 9 studies how, with probability tending to 1 along a subsequence of N, V could contain a fixed nonzero subspace for all p in an interval below 1/2. A dagger † marks a statement that rests on a computation; the Appendix lists, for each, the program and the log in the supplementary material. \section*{\pdfbookmark[1]{2. Twists and twist ratios}{h6}2. Twists and twist ratios} This section identifies twist ratios with probabilities of the events D\Sub{n} (Lemma 3). \textbf{Coordinate cocycles.} For an i-set I, let Σ\Sub{I} := \{c(x, I) : x\Sub{a} = 0 for all a ∈ I\} and η\Sub{I} := 1\Sub{Σ\Sub{I}} ∈ C\Sup{i}(X; F₂). Then η\Sub{I} is a cocycle: an (i+1)-cell of type I ∪ \{b\} has exactly two faces of type I, namely c(x, I) and c(x + e\Sub{b}, I), and they are both in Σ\Sub{I} or both outside it. Moreover ⟨η\Sub{I}, T\Sub{J}⟩ = δ\Sub{IJ}, so [η\Sub{I}] = e\Sub{I}\Sup{*}, the dual basis of H\Sup{i} := H\Sup{i}(X; F₂). We write n = Σ\Sub{I} n\Sub{I} e\Sub{I}\Sup{*} and \textbar{}n\textbar{} = \#\{I : n\Sub{I} = 1\}. \textbf{Events D\Sub{n}.} For n ∈ H\Sup{i} let D\Sub{n} := \{⟨n, V(P)⟩ = 0\}. \textbf{Lemma 1.} (a) D\Sub{n} is decreasing. (b) D\Sub{m} ∩ D\Sub{n} ⊆ D\Sub{m+n}. (c) A = ⋃\Sub{n≠0} D\Sub{n}\Sup{c} and S = ⋂\Sub{n≠0} D\Sub{n}\Sup{c}. \emph{Proof.} (a) and (b) are immediate. For (c), the evaluation pairing H\Sup{i} × H → F₂ is perfect, so V = H if and only if V\Sup{⊥} = 0, and V ≠ 0 if and only if V\Sup{⊥} ≠ H\Sup{i}. ∎ \textbf{Lemma 2 (FKG) [HS16; DS22, Thm 16].} For q ≥ 1, μ satisfies the FKG lattice condition. Hence increasing events are positively correlated, and so are decreasing events. ∎ \textbf{The twisted gauge theory.} Let G = F₂\Sup{r}, so \textbar{}G\textbar{} = q, let β ≥ 0, and let k(x) := exp(β·1\{x = 0\}) for x ∈ G. For a cocycle τ ∈ Z\Sup{i}(X; G) put Z(τ) := Σ\Sub{f ∈ C\Sup{i−1}(X; G)} Π\Sub{σ ∈ X\Sup{(i)}} k(δf(σ) + τ(σ)). \textbf{Lemma 3.} (a) Z(τ) depends only on the class [τ] ∈ H\Sup{i}(X; G). (b) Let p = 1 − e\Sup{−β}, let η ∈ Z\Sup{i}(X; F₂), and let g ∈ G ∖ \{0\}. Then Z(ηg)/Z(0) = μ\Sub{N,p,q}(D\Sub{[η]}). \emph{Proof.} (a) The substitution f ↦ f + h maps Z(τ) to Z(τ + δh). (b) Since k(x) = e\Sup{β}[(1−p) + p·1\{x = 0\}], expanding the product gives Z(τ) = e\Sup{β\textbar{}X\Sup{(i)}\textbar{}} Σ\Sub{ω} p\Sup{\textbar{}ω\textbar{}}(1−p)\Sup{\textbar{}X\Sup{(i)}\textbar{}−\textbar{}ω\textbar{}} · \#\{f : δf = τ on ω\}. Fix ω and P = P(ω). Since C\Sup{i−1}(X; G) = C\Sup{i−1}(P; G), the solution set of δf = τ on ω is either empty or a coset of Z\Sup{i−1}(P; G), and it is nonempty exactly when [τ\textbar{}\Sub{P}] = 0 in H\Sup{i}(P; G). The coset has \textbar{}Z\Sup{i−1}(P; G)\textbar{} = \textbar{}δC\Sup{i−2}(X; G)\textbar{} · \textbar{}H\Sup{i−1}(P; G)\textbar{} elements. The first factor does not depend on ω, and the second equals \textbar{}G\textbar{}\Sup{b\Sub{i−1}(P; F₂)} = q\Sup{b\Sub{i−1}}, coordinatewise in G = F₂\Sup{r}. When τ = ηg, we have [τ\textbar{}\Sub{P}] = 0 if and only if [η\textbar{}\Sub{P}] = 0 in H\Sup{i}(P; F₂): read off a coordinate j with g\Sub{j} = 1 for one direction, and use h = h\textquotesingle{}g for the other. Finally, by the universal coefficient theorem over a field, evaluation identifies H\Sup{i}(P; F₂) with Hom(H\Sub{i}(P; F₂), F₂), and ⟨η, z⟩ = ⟨[η], φ\Sub{*}[z]⟩ for a cocycle η on X and a cycle z in P. Hence [η\textbar{}\Sub{P}] = 0 if and only if D\Sub{[η]} holds, and Z(ηg) = C Σ\Sub{ω} p\Sup{\textbar{}ω\textbar{}}(1−p)\Sup{\textbar{}X\Sup{(i)}\textbar{}−\textbar{}ω\textbar{}} q\Sup{b\Sub{i−1}(P)} 1\Sub{D\Sub{[η]}} with C independent of ω and η. ∎ From now on write R(n) := μ(D\Sub{n}), which equals Z(ηg)/Z(0) for any representative η of n. \section*{\pdfbookmark[1]{3. The transfer-matrix inequality}{h7}3. The transfer-matrix inequality} For a direction μ ∈ [d] split n = s\Sub{μ}(n) + t\Sub{μ}(n), with s\Sub{μ}(n) := Σ\Sub{I∌μ} n\Sub{I} e\Sub{I}\Sup{*} (\textquotedbl{}spatial\textquotedbl{}) and t\Sub{μ}(n) := Σ\Sub{I∋μ} n\Sub{I} e\Sub{I}\Sup{*} (\textquotedbl{}temporal\textquotedbl{}). \textbf{Lemma 4.} Let q = 2\Sup{r} with r ≥ 1. For every n ∈ H\Sup{i} and every μ ∈ [d], R(n) ≤ R(s\Sub{μ}(n)). \emph{Proof.} Let Y be the (d−1)-dimensional cubical torus \{x\Sub{μ} = 0\}, and identify each slice \{x\Sub{μ} = t\} with Y. Spatial cells (of type I ∌ μ) correspond to Z\Sub{N} × Y\Sup{(\textbar{}I\textbar{})}, and temporal cells c × [t, t+1]\Sub{μ} to Z\Sub{N} × Y\Sup{(\textbar{}I\textbar{}−1)}. A cochain f ∈ C\Sup{i−1}(X; G) becomes (f\Sub{t}, g\Sub{t})\Sub{t∈Z\Sub{N}}, with f\Sub{t} ∈ C\Sup{i−1}(Y; G) and g\Sub{t} ∈ C\Sup{i−2}(Y; G); g\Sub{t} is absent if i = 1. The boundary of c × [t, t+1] is c × \{t\} + c × \{t+1\} + (∂c) × [t, t+1], so δf(t, c) = δ\Sub{Y} f\Sub{t}(c) for c ∈ Y\Sup{(i)}, δf(c × [t, t+1]) = f\Sub{t}(c) + f\Sub{t+1}(c) + δ\Sub{Y} g\Sub{t}(c) for c ∈ Y\Sup{(i−1)}. To represent the twist, we fix g ∈ G ∖ \{0\} and use τ := Σ\Sub{I} n\Sub{I} η\Sub{I} g. Its spatial part is the same in every slice: τ(t, c) = a(c) with a := Σ\Sub{I∌μ} n\Sub{I} η\Sup{Y}\Sub{I} g ∈ C\Sup{i}(Y; G). Its temporal part is supported on the cells c × [0, 1] and equals θ(c) there, with θ := Σ\Sub{I∋μ} n\Sub{I} η\Sup{Y}\Sub{I∖\{μ\}} g ∈ C\Sup{i−1}(Y; G). Here η\Sup{Y}\Sub{J} := 1\{cells of Y of type J with x\Sub{J} = 0\}. By the argument for η\Sub{I} applied inside Y, δ\Sub{Y} θ = 0. On ℋ := ℂ\Sup{C\Sup{i−1}(Y; G)} we use three operators: the multiplication operator M by S(φ) := Π\Sub{c∈Y\Sup{(i)}} k(δ\Sub{Y} φ(c) + a(c)) \textgreater{} 0, the operator 𝒦 with kernel 𝒦(φ, φ\textquotesingle{}) := Σ\Sub{γ∈C\Sup{i−2}(Y; G)} Π\Sub{c∈Y\Sup{(i−1)}} k(φ(c) + φ\textquotesingle{}(c) + δ\Sub{Y} γ(c)), and the shift (U\Sub{θ}ψ)(φ) := ψ(φ + θ). Summing over the temporal variables g\Sub{t} gives Z(τ) = Σ\Sub{(f\Sub{t})} Π\Sub{t} S(f\Sub{t}) Π\Sub{t≠0} 𝒦(f\Sub{t}, f\Sub{t+1}) · 𝒦(f\Sub{0}, f\Sub{1} + θ) = Tr(U\Sub{θ} (M𝒦)\Sup{N}). These operators have two properties. \begin{itemize}[label=\discbullet] \item{} (P1) 𝒦 ≥ 0. Indeed, 𝒦 = (⊗\Sub{c} κ) Σ\Sub{γ} U\Sub{δ\Sub{Y} γ}, where κ is the operator on ℂ\Sup{G} with kernel k(x + y), and κ is convolution by k because x + y = x − y in G. Its eigenvalues are \uni{006B0302}{\^{k}}(χ) = Σ\Sub{x} k(x)χ(x) = \textbar{}G\textbar{}·1\{χ = 1\} + (e\Sup{β} − 1) ≥ 0, so ⊗κ ≥ 0. The sum Σ\Sub{γ} U\Sub{δγ} is a positive multiple of the orthogonal projection onto vectors invariant under the subgroup δC\Sup{i−2}(Y; G), and the two factors commute, because translations commute with convolutions. Hence 𝒦 ≥ 0. \item{} (P2) U\Sub{θ} commutes with 𝒦 by translation invariance, and with M because δ\Sub{Y}(φ + θ) = δ\Sub{Y} φ. \end{itemize} We now put T := M\Sup{1/2} 𝒦 M\Sup{1/2} ≥ 0. By (P2), Tr(U\Sub{θ} (M𝒦)\Sup{N}) = Tr(U\Sub{θ} T\Sup{N}). The spatial twist s\Sub{μ}(n) has the same a and θ = 0, so Z(τ\Sub{sp}) = Tr(T\Sup{N}). Since U\Sub{θ} is unitary and T\Sup{N} ≥ 0, Z(τ) = Tr(U\Sub{θ} T\Sup{N}) ≤ ‖U\Sub{θ}‖ · ‖T\Sup{N}‖₁ = Tr(T\Sup{N}) = Z(τ\Sub{sp}). Dividing by Z(0) and using Lemma 3 gives R(n) ≤ R(s\Sub{μ}(n)). ∎ \emph{Remark.} Z(τ\Sub{sp}) − Z(τ) = 2 Tr(P\Sub{−} T\Sup{N}) ≥ 0, where P\Sub{−} = (1 − U\Sub{θ})/2: the difference is the partition function of the odd electric-flux sector. Lemma 4 is the classical positivity of \textquotesingle{}t Hooft\textquotesingle{}s electric-flux partition functions [\textquotesingle{}tH79], proved with transfer matrices in [BS83, TY85] for lattice gauge theories. What is new here is reading it in the random-cluster representation (Lemma 3), where it becomes an inequality between homological events. \section*{\pdfbookmark[1]{4. Spin-up}{h8}4. Spin-up} Lemma 4 and the FKG inequality give a lower bound on μ(S) in terms of μ(A). \textbf{Lemma 5.} Let q = 2\Sup{r} with r ≥ 1. For every N and p, with K = C(d, i): μ(S) ≥ ( μ(A) / K )\Sup{2\Sup{K} − 1}. \emph{Proof.} \begin{itemize}[label=\discbullet] \item{} (i) Coordinate permutations are automorphisms of X that preserve μ and map Σ\Sub{I} to Σ\Sub{π(I)}. Hence R(e\Sub{I}\Sup{*}) =: r is the same for all I. \item{} (ii) Let n ≠ 0 and choose I with n\Sub{I} = 1. Apply s\Sub{μ} successively for the i directions μ ∉ I. Each intermediate twist still contains the I-component, and the last one is e\Sub{I}\Sup{*}, so Lemma 4 gives R(n) ≤ r. \item{} (iii) By Lemma 1(b), D\Sub{n} ⊇ ⋂\Sub{I: n\Sub{I}=1} D\Sub{e\Sub{I}\Sup{*}}, and Lemma 2 for decreasing events gives R(n) ≥ r\Sup{\textbar{}n\textbar{}} ≥ r\Sup{K}. \item{} (iv) By Lemma 1(c), Lemma 2 for increasing events, and (ii), μ(S) = μ(⋂\Sub{n≠0} D\Sub{n}\Sup{c}) ≥ Π\Sub{n≠0}(1 − R(n)) ≥ (1 − r)\Sup{2\Sup{K}−1}. \item{} (v) If V ≠ 0, some nonzero v ∈ V has a coordinate v\Sub{I} = 1, so e\Sub{I}\Sup{*} does not vanish on V. Hence A = ⋃\Sub{I} D\Sup{c}\Sub{e\Sub{I}\Sup{*}} and μ(A) ≤ K(1 − r). \end{itemize} Combining (iv) and (v) proves the lemma. We do not need step (iii) for the bound, but it shows that every R(n) lies between r\Sup{K} and r. ∎ Lemma 5 replaces [DS22, Lemma 41] and [DKS20, Lemma 10 and Prop. 11], which give μ(S) ≥ c μ(A)\Sup{c\textquotesingle{}} only when H\Sub{i}(T\Sup{d}; F) is an irreducible representation of the point group. This fails for F = F₂: coordinate sign changes act trivially mod 2, so the point group acts on \{e\Sub{I}\} by permutations, and Σ\Sub{I} e\Sub{I} spans an invariant line [DKS20, Prop. 11]. In characteristic 2 the symmetry argument therefore cannot exclude an \textquotedbl{}oblique\textquotedbl{} phase. In such a phase V(P) would be a Lagrangian subspace such as span\{e\Sub{12} + e\Sub{34}, e\Sub{13} + e\Sub{24}, e\Sub{14} + e\Sub{23}\}, which is invariant under coordinate permutations and self-dual. Lemma 4 excludes it: there R(e\Sub{12}\Sup{*} + e\Sub{34}\Sup{*}) → 1 while R(e\Sub{12}\Sup{*}) → 0. \section*{\pdfbookmark[1]{5. Proof of Theorem A}{h9}5. Proof of Theorem A} We use two facts from [DS22] that hold for every q ≥ 1 and every field. \textbf{Fact D (duality) [DKS20, Lemma 9 (Duality Lemma); DS22, Thm 18].} Let P• be the dual complex in the dual torus X• ≅ X. \begin{itemize}[label=\discbullet] \item{} (a) rank φ\Sub{*} + rank ψ\Sub{*} = K for every configuration. Hence A(P) ⟺ ¬S(P•), and at least one of A(P), A(P•) holds. \item{} (b) The balanced measure \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p,q} ∝ q\Sup{−rank φ\Sub{*}/2} μ\Sub{N,p,q} satisfies: P• has law \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p*,q} when P has law \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p,q}, where p* = (1−p)q/((1−p)q + p). Here P• is the (d−i)-dimensional model on X•, which is i-dimensional because d = 2i. The translation X• → X by (½, …, ½) carries A and S to themselves. \item{} (c) d\uni{03BC0303}{\ensuremath{\tilde{\mu}}}/dμ ∈ [q\Sup{−K/2}, q\Sup{K/2}]. \end{itemize} \textbf{Fact T (sharp threshold) [GG06; DS22, Thm 39].} μ\Sub{N,p,q} is monotonic, and the automorphism group of X acts transitively on X\Sup{(i)}. Hence for every increasing automorphism-invariant event E, d/dp μ\Sub{p}(E) ≥ c(p) · min\{μ\Sub{p}(E), 1 − μ\Sub{p}(E)\} · log \textbar{}X\Sup{(i)}\textbar{}, with c(p) \textgreater{} 0 bounded below on compact subsets of (0, 1). Indeed, μ\Sub{p}(ω\Sub{σ} = 1 \textbar{} rest) ∈ [p/(p + q(1−p)), p], so the one-plaquette variance is bounded below by a constant depending only on p and q. \emph{Proof of Theorem A.} \begin{enumerate}[label=\arabic*.] \item{} At p = p\Sub{sd} = p\Sub{sd}*, Fact D(a) and (b) give 1 ≤ \uni{03BC0303}{\ensuremath{\tilde{\mu}}}(A(P)) + \uni{03BC0303}{\ensuremath{\tilde{\mu}}}(A(P•)) = 2\uni{03BC0303}{\ensuremath{\tilde{\mu}}}(A). By Fact D(c), μ\Sub{N,p\Sub{sd}}(A) ≥ q\Sup{−K/2}/2 for every N. \item{} By Lemma 5, μ\Sub{N,p\Sub{sd}}(S) ≥ δ₀ := (q\Sup{−K/2}/(2K))\Sup{2\Sup{K}−1} \textgreater{} 0, uniformly in N. \item{} S is increasing and automorphism-invariant, and μ\Sub{p}(S) is nondecreasing in p. Integrating Fact T from p\Sub{sd}, we see that μ\Sub{p}(S) reaches 1/2 within Δp ≤ log(1/(2δ₀))/(c log\textbar{}X\Sup{(i)}\textbar{}), and that 1 − μ\Sub{p}(S) then decays like exp(−c log\textbar{}X\Sup{(i)}\textbar{} · Δp). Hence μ\Sub{N,p}(S) → 1 for every p \textgreater{} p\Sub{sd}. \item{} For p \textless{} p\Sub{sd} we have p* \textgreater{} p\Sub{sd}, so μ\Sub{N,p*}(S) → 1 by step 3, and also \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p*}(S) → 1 by Fact D(c). By Fact D(b), \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p}(S(P•)) → 1. By Fact D(a), \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p}(A) = 1 − \uni{03BC0303}{\ensuremath{\tilde{\mu}}}\Sub{N,p}(S(P•)) → 0. By Fact D(c), μ\Sub{N,p}(A) → 0. ∎ \end{enumerate} \textbf{Theorem C (sharp threshold in every dimension).} Let d ≥ 2, 1 ≤ i ≤ d − 1, q = 2\Sup{r} with r ≥ 1, and F₂ coefficients. Define λ(N) by μ\Sub{N,λ(N),q}(A) = 1/2. Then for every ε \textgreater{} 0, μ\Sub{N,λ(N)−ε,q}(A) → 0 and μ\Sub{N,λ(N)+ε,q}(S) → 1. This is the characteristic-2 analogue of [DS22, Prop. 42 and Thm 11]. \emph{Proof.} Lemmas 3–5 do not use d = 2i. In Lemma 5(ii) the d − i projections s\Sub{μ}, μ ∉ I, leave only components I′ ⊆ I with \textbar{}I′\textbar{} = i, that is, e\Sub{I}\Sup{*}. Hence μ(S) ≥ (μ(A)/K)\Sup{2\Sup{K}−1} with K = C(d, i). At p = λ(N) this gives μ(S) ≥ (2K)\Sup{−(2\Sup{K}−1)}, and Fact T gives μ\Sub{λ+ε}(S) → 1. For the event A we pass to the dual complex P•, which is the (d−i)-dimensional plaquette random-cluster model with the same q on X• (balanced form of [DS22, Thm 18]). The duality lemma rank φ\Sub{*} + rank ψ\Sub{*} = K holds for every d and i, so A(P) ⟺ ¬S(P•). From μ\Sub{λ}(A) = 1/2 and Fact D(c), the dual model at λ* has μ(S•) ≥ q\Sup{−K}/2, hence μ(A•) ≥ q\Sup{−K}/2. Lemma 5 for the dual model (dimension d − i, same q) gives μ(S•) ≥ δ \textgreater{} 0 at λ*, and Fact T for the dual model then gives μ(S•) → 1 at (λ − ε)* \textgreater{} λ*. By duality, μ\Sub{λ−ε}(A) → 0. ∎ \emph{Proof of Corollary B.} Since 1\{σ = 1\} = (1+σ)/2, the Ising weight exp(J Σ σ\Sub{c}) is the gauge weight of §2 with q = 2 and β = 2J, so p = 1 − e\Sup{−2J} and p\Sub{sd} corresponds to J\Sub{sd} = ½ ln(1+√2). By Lemma 3, Z\Sub{J}(n)/Z\Sub{J}(0) = μ(D\Sub{n}). For 0 ≤ J \textless{} J\Sub{sd} we have μ(A) → 0 and A\Sup{c} ⊆ D\Sub{n}; for J \textgreater{} J\Sub{sd} we have μ(S) → 1 and S ⊆ D\Sub{n}\Sup{c}. ∎ \section*{\pdfbookmark[1]{6. Remarks}{h10}6. Remarks} \begin{enumerate}[label=\arabic*.] \item{} We indicate how the argument extends, with signs, to q = ℓ\Sup{r} and F = F\Sub{ℓ} for any prime ℓ, with gauge group G = F\Sub{ℓ}\Sup{r}. The slicing formula becomes δf(c × [t, t+1]) = ±(f\Sub{t+1} − f\Sub{t})(c) + δ\Sub{Y} g\Sub{t}(c), and η\Sub{I} needs a fixed orientation. The kernel k is even, so 𝒦 is again a convolution with nonnegative Fourier coefficients. U\Sub{θ} is unitary but no longer an involution, so the ± split of the remark after Lemma 4 becomes a Z\Sub{ℓ} Fourier split. Finally R(e\Sub{I}\Sup{*}) = R(−e\Sub{I}\Sup{*}), and 2\Sup{K} − 1 becomes ℓ\Sup{K} − 1. For odd ℓ this reproves [DS22, Thm 8] for these q without representation theory. \item{} Theorem A concerns tori only; the infinite-volume area/perimeter law [DS26, Conjecture 1] remains open for d = 4. \end{enumerate} \section*{\pdfbookmark[1]{7. Other values of q: reductions, partial results and obstructions}{h11}7. Other values of q: reductions, partial results and obstructions} For real q ≥ 1 with F₂ coefficients (in particular q = 1, Bernoulli plaquette percolation [DKS20]; whether the restriction on the characteristic can be removed there is asked in [DKS20, §9]) there is no gauge theory, and Lemma 4 is proved only for q = 2\Sup{r}. The rest of the argument does not use the gauge theory. In this section we reduce the theorem for such q to a single inequality and study that inequality. For q = 1 we also prove partial results with coefficients F₂ and Z/4; §8 continues with integral coefficients. \subsection*{\pdfbookmark[2]{7.1 Reduction to Lemma 4 and to a crossing estimate}{h12}7.1 Reduction to Lemma 4 and to a crossing estimate} \textbf{Theorem D (reduction).} Fix d, i, real q ≥ 1, and F₂ coefficients. Suppose R(n) ≤ R(s\Sub{μ}(n)) holds for μ\Sub{N,p,q}, for all N, p, n and μ. Then Lemma 5, Theorem C and, when d = 2i, Theorem A hold for this q. \emph{Proof.} Lemma 5 uses only Lemma 4, the FKG inequality and coordinate symmetry, all of which hold for every real q ≥ 1 [HS16]. Section 5 uses Facts D and T, which hold for every real q ≥ 1 [DS22]; for q = 1 they are exact self-duality at p = 1/2 and the Friedgut–Kalai theorem. ∎ \textbf{Proposition 6 (an unconditional window for Bernoulli).} For Bernoulli plaquette percolation with F₂ coefficients on T\Sup{2i}\Sub{N}: P\Sub{p}(A) → 0 for p \textless{} √2 − 1, P\Sub{p}(S) → 1 for p \textgreater{} 2 − √2. \emph{Proof.} Given the other plaquettes, a plaquette is open with conditional probability p if it closes an i-cycle and with probability p/(p + q(1−p)) if it kills an (i−1)-cycle. Holley\textquotesingle{}s inequality therefore gives Bernoulli(p\textquotesingle{}/(p\textquotesingle{} + 2(1−p\textquotesingle{}))) ≤\Sub{st} μ\Sub{p\textquotesingle{},2} ≤\Sub{st} Bernoulli(p\textquotesingle{}). Since A and S are increasing, for p \textgreater{} p\Sub{sd}(2) = 2 − √2 we get P\Sub{p}(S) ≥ μ\Sub{p,2}(S) → 1 by Theorem A. For p \textless{} √2 − 1, put p\textquotesingle{} = 2p/(1+p) \textless{} p\Sub{sd}(2); then P\Sub{p}(A) ≤ μ\Sub{p\textquotesingle{},2}(A) → 0. ∎ In the proof of Theorem A, Lemma 4 is used only to show that every R(n) stays bounded away from 1 at p = p\Sub{sd}. A lower bound on the probability of a coordinate giant cycle at the self-dual point gives the same conclusion, as we now show. For d = 2i we use the following duality. By Fact D(a), V(P) and V(P•) have complementary dimensions, and they pair to zero: a primal i-cell meets a dual i-cell transversally only when the two cells are dual to each other, and P• contains the duals of the closed i-cells only. Hence V(P•) = V(P)\Sup{⊥} under the intersection pairing, which after the translation of Fact D(b) is e\Sub{I} · e\Sub{J} = 1\{I ∩ J = ∅\}. Poincaré duality sends e\Sub{I}\Sup{*} to e\Sub{I\Sup{c}}, so for every n ∈ H\Sup{i} the event D\Sub{n} means exactly that PD(n) ∈ V(P•). For d = 4 we read a class as a set of edges of the complete graph on \{1, 2, 3, 4\}, where the plane \{a, b\} is the edge ab. Let u := Σ\Sub{I} e\Sub{I} be the full edge set, and let C be the span of the three 4-cycles, such as e\Sub{12} + e\Sub{14} + e\Sub{23} + e\Sub{34}. The three 4-cycles sum to 0, so dim C = 2. \textbf{Lemma 6′.} Let d = 4, i = 2. Every nonzero subspace of H that is invariant under coordinate permutations contains u or contains C. \emph{Proof.} Let U be such a subspace and take v ∈ U ∖ \{0, u\}. Some transposition τ = (ab) moves v: the permutations act transitively on the six planes, so 0 and u are the only vectors fixed by all transpositions. Let \{c, d\} = [4] ∖ \{a, b\}. Then τ fixes the edges ab and cd and swaps ac ↔ bc and ad ↔ bd. Hence w := v + τv is a nonzero element of U, equal to one of ac + bc, ad + bd, or ac + bc + ad + bd. In the last case w is a 4-cycle. In the first case w + (cd)w = ac + bc + ad + bd is a 4-cycle, and the second case is the same. Since coordinate permutations act transitively on the three 4-cycles, C ⊆ U. ∎ \textbf{Proposition 7 (reduction to a crossing estimate).} Let d = 4, i = 2, q ≥ 1, and F₂ coefficients. Suppose r₀ := inf\Sub{N} μ\Sub{N,p\Sub{sd},q}(e\Sub{12} ∈ V) \textgreater{} 0. Then Theorem A holds for this q. \emph{Proof.} Work at p = p\Sub{sd} and write \uni{03BC0303}{\ensuremath{\tilde{\mu}}} for the balanced measure of Fact D. \begin{itemize}[label=\discbullet] \item{} (i) If a · b = 1, then a ∈ V(P) and b ∈ V(P•) cannot both hold. P• has law \uni{03BC0303}{\ensuremath{\tilde{\mu}}} when P does, so \uni{03BC0303}{\ensuremath{\tilde{\mu}}}(a ∈ V) + \uni{03BC0303}{\ensuremath{\tilde{\mu}}}(b ∈ V) ≤ 1. \item{} (ii) The class e\Sub{12} meets u in the single disjoint plane 34. It also meets the 4-cycle e\Sub{12} + e\Sub{14} + e\Sub{23} + e\Sub{34} ∈ C in the plane 34. By (i) and Fact D(c), μ(u ∉ V) ≥ q\Sup{−K} r₀ and μ(C ⊄ V) ≥ q\Sup{−K} r₀. \item{} (iii) Let c ≠ 0 and ε := μ(c ∉ V). By symmetry and a union bound, the span U\Sub{c} of the 24 images of c satisfies μ(U\Sub{c} ⊆ V) ≥ 1 − 24ε. By Lemma 6′, U\Sub{c} contains u or C, so (ii) gives ε ≥ q\Sup{−K} r₀/24 =: δ. Thus μ(c ∈ V) ≤ 1 − δ for every c ≠ 0. \item{} (iv) For n ≠ 0, D\Sub{n}\Sup{c} = \{PD(n) ∉ V(P•)\}. Two uses of Fact D(c) and the self-duality of \uni{03BC0303}{\ensuremath{\tilde{\mu}}} give μ(D\Sub{n}\Sup{c}) ≥ q\Sup{−K} μ(PD(n) ∉ V) ≥ q\Sup{−K} δ. \item{} (v) By Lemma 1(c) and Lemma 2, μ(S) ≥ Π\Sub{n≠0} μ(D\Sub{n}\Sup{c}) ≥ (q\Sup{−K} δ)\Sup{2\Sup{K}−1}, uniformly in N. Steps 3 and 4 of the proof of Theorem A now apply verbatim. ∎ \end{itemize} For q = 1, step (i) and the symmetry e\Sub{12} ↔ e\Sub{34} give μ\Sub{1/2}(e\Sub{12} ∈ V) ≤ 1/2, while Monte Carlo at p = 1/2 (the samples of the d = 4 study in §7.5) gives μ(e\Sub{I} ∈ V) = 0.424, 0.436 and 0.452 for N = 4, 6 and 8, averaged over the six coordinate classes e\Sub{I}.† The hypothesis of Proposition 7 is a box-crossing (Russo–Seymour–Welsh) bound for giant surfaces at the self-dual point; its i = 1 analogue is the classical crossing bound that underlies Kesten\textquotesingle{}s proof of p\Sub{c} = 1/2. A crossing probability that decays slower than every power of N would suffice. If r₀(N) := μ\Sub{N,p\Sub{sd},q}(e\Sub{12} ∈ V) satisfies log(1/r₀(N)) = o(log N), steps (i)–(v) give μ\Sub{N,p\Sub{sd}}(S) ≥ (q\Sup{−2K} r₀(N)/24)\Sup{2\Sup{K}−1} = N\Sup{−o(1)}, and step 3 of the proof of Theorem A still reaches μ\Sub{p}(S) → 1 for every p \textgreater{} p\Sub{sd}, because it needs only log(1/μ\Sub{p\Sub{sd}}(S)) = o(log \textbar{}X\Sup{(i)}\textbar{}). \subsection*{\pdfbookmark[2]{7.2 Partnerless classes}{h13}7.2 Partnerless classes} In this subsection d = 4 and i = 2. Call a nonzero class z ∈ H \emph{partnered} if ⟨z, gz⟩ = 1 for some coordinate permutation g. At p = 1/2, partnered classes satisfy μ(z ∈ V) ≤ 1/2 by step (i) of the proof of Proposition 7, because μ(gz ∈ V) = μ(z ∈ V). \textbf{Proposition 8 (the partnerless classes).} A nonzero class is partnerless iff it lies in C\Sup{⊥}. Read as an edge set of K₄, this means that its four vertex degrees have equal parity. C\Sup{⊥} is the 4-dimensional permutation-invariant subspace spanned by the triangles and the stars; its 15 nonzero elements are the matchings, triangles, stars, 4-cycles and u. For Bernoulli percolation, P\Sub{p}(V ⊄ C\Sup{⊥}) → 0 for every p \textless{} 1/2 and P\Sub{p}(C ⊆ V) → 1 for every p \textgreater{} 1/2. Consequently, Theorem A at q = 1 holds iff P\Sub{p}(V ∩ C\Sup{⊥} ≠ 0) → 0 for every p \textless{} 1/2. \emph{Proof.} For the 4-cycle c with diagonals 13 and 24, ⟨z, c⟩ = deg\Sub{z}(1) + deg\Sub{z}(3) mod 2; the other two 4-cycles give deg\Sub{z}(1) + deg\Sub{z}(2) and deg\Sub{z}(1) + deg\Sub{z}(4). Thus z ∈ C\Sup{⊥} iff all degrees have equal parity, and a check of the ten permutation orbits of nonzero classes shows that this is exactly the set of partnerless classes.† Each event \{z ∈ V\} is increasing and invariant under translations, whose orbits on plaquettes have size N⁴. By the Friedgut–Kalai theorem in the form of [BR06, Lemma 7], μ\Sub{1/2}(z ∈ V) ≤ 1/2 forces μ\Sub{p}(z ∈ V) → 0 for every p \textless{} 1/2, and a union bound over the 48 partnered classes gives the first claim. For p \textgreater{} 1/2 the dual is Bernoulli(1−p), so V(P•) ⊆ C\Sup{⊥} whp, and V = V(P•)\Sup{⊥} ⊇ C. ∎ The next proposition shows that the obstruction is not an artefact of the cubic lattice. \textbf{Proposition 9 (periodic structures).} Let Λ ⊂ R⁴ be a lattice and G ⊂ GL(Λ) a finite group. Then Λ²(Λ/2Λ) ≅ H₂(R⁴/Λ; F₂) contains a nonzero z with ⟨z, gz⟩ = 0 for all g ∈ G, where ⟨·,·⟩ is the mod-2 intersection form. \emph{Proof.} G preserves a positive definite integral form B; divide B by a power of 2 until \uni{00420304}{\={B}} := B mod 2 ≠ 0. If \uni{00420304}{\={B}} is alternating, it is a nonzero G-fixed element of Λ²(Λ/2Λ)* ≅ Λ²(Λ/2Λ), and a fixed z has ⟨z, gz⟩ = ⟨z, z⟩ = 0. Otherwise ℓ(x) := \uni{00420304}{\={B}}(x, x) is a nonzero G-invariant linear form, and Λ²(ker ℓ) is G-invariant and isotropic, because x ∧ y ∧ x′ ∧ y′ ∈ Λ⁴(ker ℓ) = 0. Any z ∈ Λ²(ker ℓ) ∖ \{0\} works. ∎ For a periodic self-dual cell structure on R⁴/Λ, apply Proposition 9 to the group generated by its point group and the linear part of its duality isometry. The bound of Proposition 7(i) then leaves some nonzero class unconstrained, so changing the lattice does not remove the obstruction. For the permutohedral lattice A₄*, the second model of [DKS20], the point group S₅ fixes exactly one nonzero class u\Sub{Q} ∈ H₂(T⁴; F₂), and u\Sub{Q} is the only partnerless class.† [DKS20, Theorem 4] states that on permutohedral tori of even dimension the threshold λ\Sub{d/2}(N) tends to 1/2 in every characteristic. Its proof uses the irreducibility of H\Sub{d/2}(T\Sup{d}; F) under S\Sub{d+1} [DKS20, Proposition 22], which is proved there only when char F = 0 or char F is an odd prime not dividing d + 1. For d = 4 it fails over F₂, since S₅ fixes u\Sub{Q}. For F₂ and d = 4 that clause is therefore open, and the bound of Proposition 7(i) controls every class except u\Sub{Q}. \textbf{Proposition 10 (the threshold with Z/4 coefficients).} For Bernoulli percolation and every p \textless{} 1/2, P\Sub{p}(V\Sub{Z/4} ⊄ 2(Z/4)⁶) → 0, where V\Sub{Z/4} is the image of H₂(P; Z/4) in H₂(T⁴; Z/4) ≅ (Z/4)⁶. Equivalently, below 1/2 no giant mod-2 cycle lifts to a mod-4 cycle. \emph{Proof.} \begin{itemize}[label=\discbullet] \item{} (i) Enumerating all 4032 elements x ∈ (Z/4)⁶ of order 4 shows that each has a signed permutation g ∈ B₄ with ⟨x, gx⟩ ≢ 0 mod 4.† \item{} (ii) Disjoint Z/4-cycles have zero Z/4 intersection number, so V\Sub{Z/4}(P•) ⊥ V\Sub{Z/4}(P). At p = 1/2, V\Sub{Z/4}(P•) has the law of V\Sub{Z/4}(P), because the half-translation acts trivially on homology. Signed permutations are induced by lattice automorphisms, so μ(gx ∈ V\Sub{Z/4}) = μ(x ∈ V\Sub{Z/4}). Hence μ\Sub{1/2}(x ∈ V\Sub{Z/4}) ≤ 1/2 for every x of order 4. \item{} (iii) Sharp thresholds as in Proposition 8, and a union bound over the 4032 elements. ∎ \end{itemize} In Z/2\Sup{k} every element of order 2 is partnerless, so the obstruction of Proposition 8 reappears at the 2-torsion level. Proposition 10 does not settle the problem; it shows only that a putative oblique phase below 1/2 consists of mod-2 cycles whose Bockstein in H₁(P; Z) is not divisible by 2. Proposition 20 in §8 extends the argument of Propositions 8 and 10 to all coefficient rings Z/m, and Theorem W shows that these constraints persist slightly above 1/2: there no partnered class lifts to a Z/4-cycle and no nonzero mod-2 class lifts to a Z/8-cycle. \subsection*{\pdfbookmark[2]{7.3 Box crossings and crumpled cycles}{h14}7.3 Box crossings and crumpled cycles} \textbf{Proposition 11 (box crossings away from aspect ratio 1).} Let g\Sub{p}(a,b,c,d) be the probability that the box B with 12-sides (a,b) and 34-sides (c,d) has an open 12-crossing: a 2-chain of open plaquettes in B, with boundary in the 12-faces ∂₁₂B = \{x ∈ B : x₁ or x₂ extremal\}, that is nontrivial in H₂(B, ∂₁₂B; F₂) ≅ F₂. \begin{itemize}[label=\discbullet] \item{} For Bernoulli percolation exactly one of \{open 12-crossing, closed dual 34-crossing\} occurs. Rotating by (13)(24) gives g\Sub{p}(a,b,c,d) + g\Sub{1−p}(c+1, d+1, a−1, b−1) = 1, and in particular g\Sub{1/2}(c+1, d+1, c, d) = 1/2. \item{} Fix ε ∈ (0, 1/10). For p \textgreater{} 1/2, g\Sub{p}((1−ε)L, (1−ε)L, (1+ε)L, (1+ε)L) → 1. For p \textless{} 1/2, g\Sub{p}((1+ε)L, (1+ε)L, (1−ε)L, (1−ε)L) → 0. \end{itemize} \emph{Proof.} This is the argument of Bollobás and Riordan [BR06, Lemma 8], which they note extends to surfaces in higher dimensions. \begin{enumerate}[label=\arabic*.] \item{} On T⁴\Sub{4L} let U be the union, over all translates and coordinate permutations, of the crossing event of the self-dual box B\Sub{sd} with 12-sides L+1 and 34-sides L. U is increasing and invariant under a group acting transitively on plaquettes, and μ\Sub{1/2}(U) ≥ g\Sub{1/2}(B\Sub{sd}) = 1/2. By [BR06, Lemma 7], μ\Sub{p}(U) → 1 for every p \textgreater{} 1/2. \item{} If B′ has a smaller 12-square and a larger 34-square than B, suitably placed, a 12-crossing of B restricts to one of B′: cut along the 12-hyperplanes of B′; the mod-2 number of 12-plaquettes over an interior point is unchanged. \item{} Let s = ⌊εL/2⌋. Rounding the corner of each box in U down to the grid sZ⁴ gives U ⊆ ⋃\Sub{i ≤ k} A\Sub{i}, where each A\Sub{i} is a translate or permutation of the crossing event of the box B° with 12-sides L+1−2s and 34-sides L+2s, and k ≤ 6(8/ε)⁴. \item{} By Harris\textquotesingle{} inequality for the decreasing events A\Sub{i}\Sup{c}, μ\Sub{p}(A(B°)\Sup{c})\Sup{k} ≤ μ\Sub{p}(U\Sup{c}) → 0. Containment gives the first claim. \item{} The second claim is the first one for the dual, through the complementarity identity. ∎ \end{enumerate} Proposition 11 needs no gluing of sheets, whereas the self-dual box itself, and the passage from boxes to tori, both need a lower bound at p = 1/2 for boxes that are slightly harder than self-dual, i.e. sheet gluing. We explain in outline why Theorem A cannot be deduced from Proposition 11 by deterministic inclusions alone. Call a box δ-hard for I-crossings if each I-side exceeds a distinct I\Sup{c}-side by δL. A flat 2-plane T that is a graph x\Sub{I\Sup{c}} = A x\Sub{I} contains an I-crossing of the box with I-sides s and I\Sup{c}-sides t exactly when \textbar{}A\textbar{} s ≤ t entrywise (\textbar{}A\textbar{} the entrywise absolute value). Call T \emph{hardness-free} if no such box is δ-hard for any δ \textgreater{} 0 and any I; by Collatz–Wielandt this is a condition on spectral radii of 2×2 matrices built from the Plücker coordinates of T. The plane \{(x₁, x₂) = R\Sub{π/4}(x₃, x₄)\} is hardness-free with margin (each spectral radius is at least √2),† so hardness-free planes form a set with interior. Every nonzero class of H₂(T⁴; F₂) is the class of a cellular approximation of one hardness-free rational plane, or of the sum of two, for instance e\Sub{34} ← ⟨(2,2,1,0), (0,0,1,1)⟩ and u ← ⟨e₁+e₂, e₁+e₃⟩ + ⟨e₁+e₂+e₃, e₄⟩.† These configurations contain no δ-hard crossing at scales L ≥ C/δ. Hence, for every z ≠ 0, the event \{z ∈ V\} is contained in no monotone combination of the events that Proposition 11 controls. Similarly, primal gluing of two boxes fails exactly when a dual crossing of their union is contained in neither box. That dual crossing can be a cellular hardness-free plane, so gluing must be proved probabilistically. The obstruction just described is specific to the six coordinate orientations. Proposition 11 holds in tilted coordinates as well. \textbf{Lemma 12′.} For every 2-plane T ⊂ R⁴ there is an involution g ∈ B₄ with gT ∩ T = 0. \emph{Proof.} Let p = v₁ ∧ v₂ be the Plücker vector of T. Then det[v₁, v₂, gv₁, gv₂] = p ∧ (Λ²g)p, so gT ∩ T = 0 iff this number is nonzero. \begin{itemize}[label=\discbullet] \item{} For g = diag(1, 1, −1, −1), the Plücker relation p₁₂p₃₄ − p₁₃p₂₄ + p₁₄p₂₃ = 0 gives p ∧ (Λ²g)p = 4p₁₂p₃₄. The other two double sign changes give ±4p₁₃p₂₄ and ±4p₁₄p₂₃. So we are done unless p\Sub{I} p\Sub{I\Sup{c}} = 0 for all three pairs; then the support of p, read as edges of K₄, is contained in a star or in a triangle. \item{} \emph{Star.} After a coordinate permutation, T = span(e₁, w) with w = (0, a, b, c) ≠ 0. For j ∈ \{2, 3, 4\} let g swap e₁ ↔ e\Sub{j} and either fix one of the two remaining basis vectors and negate the other, or swap them. For j = 2 the determinant is ±2bc or ±(b² − c²), and both vanish only if b = c = 0. The cases j = 3, 4 are the same with (a, c) and (a, b). Since w ≠ 0, some j works. \item{} \emph{Triangle.} After a coordinate permutation, T = \{x₄ = 0, n·x = 0\} with n = (n₁, n₂, n₃, 0) ≠ 0. For i ∈ \{1, 2, 3\} let g swap e\Sub{i} ↔ e₄ and act on the other two basis vectors as above. Then gT ∩ T = 0 iff e₄, e\Sub{i}, n, gn are linearly independent. With \{j, k\} = \{1, 2, 3\} ∖ \{i\}, the relevant determinant is ±2n\Sub{j} n\Sub{k} or ±(n\Sub{j}² − n\Sub{k}²). Choose i with (n\Sub{j}, n\Sub{k}) ≠ 0. ∎ \end{itemize} By compactness of the Grassmannian, the transversality constant \textbar{}det[v₁, v₂, gv₁, gv₂]\textbar{} (orthonormal v₁, v₂) of the best g is bounded below by a constant c₀ \textgreater{} 0, uniformly in T. Numerically c₀ ≈ 0.62.† \textbf{Proposition 12 (tilted parallelepipeds).} Let T ⊂ R⁴ be a 2-plane and g ∈ B₄ an involution with gT ∩ T = 0 (Lemma 12′). Choose a basis v₁, v₂ of T and write x = s₁v₁ + s₂v₂ + t₁gv₁ + t₂gv₂, so that g acts by (s, t) ↦ (t, s). For a, b \textgreater{} 0 let Π(a, b) be the parallelepiped \{s ∈ [0, a]², t ∈ [0, b]²\}, with T-boundary ∂[0, a]² × [0, b]², and define a T-crossing of (a cubical approximation of) Π(a, b) as for boxes. Fix ε ∈ (0, 1/10). For p \textgreater{} 1/2, P\Sub{p}(T-crossing of Π((1−ε)L, (1+ε)L)) → 1, and for p \textless{} 1/2, P\Sub{p}(T-crossing of Π((1+ε)L, (1−ε)L)) → 0, uniformly in the position. \emph{Proof.} Let K be a union of lattice 4-cubes that is a 4-ball, with ∂K = A ∪ B, where A and B are unions of boundary 3-faces, each a solid torus. Lefschetz duality for (K; A, B) gives, exactly as for boxes, that exactly one of \{primal A-crossing, dual B-crossing\} occurs. The dual configuration shifted by h = (½, ½, ½, ½) has the law of the primal one at p = 1/2, and g maps T-crossings to gT-crossings and back. Therefore, for every such region, P\Sub{1/2}(T-crossing of K) + P\Sub{1/2}(T-crossing of K\textasciicircum{}♯) = 1, where K\textasciicircum{}♯ = g(K* + h) is the shifted dual region and lies within distance O(1) of g(Π(a, b) + h) = Π(b, a) + gh. Let U be the union, over all cubical approximations within distance C of translates of Π(L, L), of the T-crossing events. For C large this family is closed under ♯, so U is an increasing, translation-invariant event with μ\Sub{1/2}(U) ≥ 1/2. Steps 1–5 of the proof of Proposition 11 now apply: [BR06, Lemma 7], snapping to a grid of spacing εL/2, containment (a T-crossing of Q × Q′ restricts to one of Q° × Q′° when Q° ⊆ Q and Q′° ⊇ Q′), and Harris\textquotesingle{} inequality. ∎ A flat plane S crosses, in the S-direction, parallelepipeds Π(L, δL) that are as thin as we like. In its own coordinates every flat piece is thus maximally hard, and by Proposition 12 flat pieces of macroscopic size are improbable below 1/2 in every direction. Concretely, fix K, δ and a finite net of directions T; a union bound over the finitely many parallelepipeds of size at least N/K with these directions and with corners on a grid of spacing N/K² shows that, for p \textless{} 1/2, w.h.p. no surface in P contains a δ-hard crossing of any of them. Crumpled cycles, that is, cycles that at every macroscopic scale and position are not close to any flat plane, are the only remaining obstruction, in the following sense; we only sketch the proof. Fix p and a scale L, and suppose that for every grid box Q of side L there is an increasing event E\Sub{Q}, depending only on the plaquettes within distance 2L + 2 of Q. Suppose also that E\Sub{Q} occurs for every box Q visited by a connected cycle with nonzero class, and that sup\Sub{Q} P\Sub{p}(E\Sub{Q}) ≤ η₀, an absolute constant. Then P\Sub{p}(V ≠ 0) ≤ 2(N/L)⁴e\Sup{−6(N/L)²} → 0. Indeed, a connected giant cycle visits a *-connected set of at least (N/L)² boxes, there are at most (N/L)⁴(80e)\Sup{m} such sets of size m, and the events E\Sub{Q} of a 6-separated subset of size m/11⁴ are independent. The coarse-grained entropy (80e)\Sup{m} does not depend on L, so near 1/2 the difficulty is coverage rather than entropy. With the events of Proposition 12, whose proof gives a polynomial rate, the same counting, which we do not write out, shows the following: for each p \textless{} 1/2 and θ \textgreater{} 0, w.h.p. every giant cycle carries a hard crossing in fewer than a θ-fraction of its visited boxes, at every dyadic scale between a fixed L₀(p, θ) and N/16. What is still missing is a local event that occurs in every visited box of a crumpled cycle and is rare below 1/2. Its natural forms are subcritical arms of connected relative 2-cycles, or easy crossings (equivalently, by box duality, supercritical gluing). Numerically, in the ensemble in which exactly half of the plaquettes are open, the minimal representative of a coordinate giant class, computed by annealing, has area ≈ 1.06·N\Sup{3.35} for N = 4, …, 12.† For N ≤ 8 the annealing was checked against exact integer programs on 40 test cases from both ensembles (Bernoulli and fixed density, densities 0.485 to 0.5, several classes): it found the exact minimum in 36 of them and was at most 7.3\% above it in the other four.† These data indicate that critical giant cycles are far from flat. In runs at densities slightly below 1/2, the giant cycles that occur are flatter and more often straight.† For p \textless{} 1/2, Proposition 12 also excludes large graph pieces. Fix Λ, η \textgreater{} 0. We outline the argument. By compactness and quantitative differentiation (Arzelà–Ascoli plus Rademacher), a relative 2-cycle that lies within O(1) of a Λ-Lipschitz graph over a square of side r is, on some sub-square of side δ(Λ, η)r, within ηr of a plane. It then contains a hard crossing in that plane\textquotesingle{}s coordinates. Thus, for p \textless{} 1/2, w.h.p. no surface of P contains a piece that is a bounded-slope graph over a square of side ≥ ηN. Below 1/2, giant cycles, slab disks and gluing obstructions must fold back or be arbitrarily steep at every macroscopic scale. The following construction shows a limitation of these events. Drill small holes in a flat sheet and cap each hole by a long thin finger, a tube closed at its far end. The result is homologous to the sheet. A finger that leaves a box through its transverse faces puts boundary on those faces, so the restricted chain is no longer a crossing. Placing fingers of length ≈ L at density ≈ L⁻² at every dyadic scale L therefore destroys every cubic crossing at every scale above the shortest finger length ℓ₀. The area grows only by a factor 1 + O(1/ℓ₀). The same construction avoids hard tilted crossings below a fixed fraction of the size of the sheet, and an adversarial search in a continuum model of the fingered sheet finds no hard tilted crossing at the largest scales either.† Fingers can thus make a sheet invisible to the events controlled by Propositions 11 and 12, at negligible cost in area. We record one conditional route and sketch the implication. Call a pair of overlapping regions R₁, R₂ (two easy boxes, or two overlapping annular strips T¹ × [s, e] × □) \emph{uniquely crossed} if every 12-crossing of R₁ and every 12-crossing of R₂ have traces on the middle slice of the overlap that differ by a boundary in P ∩ (R₁ ∩ R₂). If, for every p \textgreater{} 1/2, these pairs are uniquely crossed w.h.p., then Theorem A holds. Indeed, cyclically overlapping easy boxes give a crossing of an annular strip, and three cyclically overlapping strips give a 12-torus. In Monte Carlo runs at p ≥ 0.54 the estimated probability of unique crossing increases towards 1 with size: for annular strips at p = 0.56 it is 0.22, 0.43, 0.66 and 0.86 for sides 4, 6, 8 and 10.† At p = 1/2 it is small and does not grow: 0.08, 0.03, 0.005 and 0.008 for the same sides.† \subsection*{\pdfbookmark[2]{7.4 Reflection positivity}{h15}7.4 Reflection positivity} We now show that Lemma 4 at q = 1 does not follow from reflection positivity. Reflect the torus in direction μ through two planes cutting the temporal cells of two opposite gaps. Write a configuration as X ∪ H ∪ θY, where X and Y are configurations of one half and H is the set of open plaquettes crossing the planes. Place the temporal twist e in one of the crossing layers. For q = 2 one reflection suffices, and Lemma 4 then holds for every reflection-symmetric product of couplings (β\Sub{θσ} = β\Sub{σ}). We sketch the argument. In the gauge theory, Z(τ) = Σ\Sub{a,b} F(a) M\Sub{τ}(a, b) F(b), where a and b are the gauge fields on the two boundary layers, F ≥ 0 is the partition function of one half with boundary value a, and M\Sub{τ} is the weight of the crossing layers, after summing over the gauge field on the crossing (i−1)-cells. Expanding the crossing weights in characters gives M\Sub{s} − M\Sub{s+e} = c Σ\Sub{S: e(S) = 1} (Π\Sub{h∈S} t\Sub{h}) χ\Sub{S} ⊗ χ\Sub{S}, with t\Sub{h} = p\Sub{h}/(2 − p\Sub{h}) ≥ 0 and c \textgreater{} 0. Here S runs over the sets of crossing plaquettes that are cycles relative to the crossing (i−1)-cells, and χ\Sub{S}(a) = (−1)\Sup{a(∂S)}. This kernel is positive semidefinite, which gives R(s+e) ≤ R(s). For q = 1 the analogous statement is false: \textbf{Proposition 13.} Take the torus Z\Sub{3} × Z\Sub{2} × Z\Sub{4} (d = 3, i = 2), with μ = 3 and the reflection x\Sub{3} ↦ 1 − x\Sub{3}. Its planes cut the temporal plaquettes of the gaps [0,1] and [2,3]. Let s = [η\Sub{12}] and e = [η\Sub{13}], and put η\Sub{13} on the gap [0,1]. Let H consist of the following 14 of the 24 crossing plaquettes, given as (type; base point (x\Sub{1}, x\Sub{2}, x\Sub{3})): \begin{itemize}[label=\discbullet] \item{} type \{1,3\}: (0,0,2), (0,1,0), (0,1,2), (1,0,0), (1,0,2), (1,1,2), (2,0,0), (2,0,2), (2,1,0); \item{} type \{2,3\}: (0,0,2), (0,1,0), (1,0,0), (2,0,2), (2,1,2). \end{itemize} Then Σ\Sub{X,Y} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(X ∪ H ∪ θY) = −9 711 396, the sum running over all 2\Sup{24} × 2\Sup{24} pairs of half-configurations. Consequently, take the product measure with probability 1/2 on non-crossing plaquettes, 1 − δ on H, and δ on the other crossing plaquettes. This measure is reflection-symmetric, and for all small δ \textgreater{} 0 it satisfies μ(D\Sub{s}) \textless{} μ(D\Sub{s+e}). The sum is computed exactly.† After scaling by 2\Sup{48}, every intermediate quantity is an integer of absolute value at most 2\Sup{48}. The computation reduces each half to its affine subspace of boundary values and works in F₂\Sup{H} modulo the crossing coboundaries. On 2×2×2 and 2×2×4, an independent brute-force computation reproduces its output to at least 10 significant digits, including a negative value on 2×2×4 for an inhomogeneous reflection-symmetric product measure.† A proof of Lemma 4 at q = 1 must therefore use the homogeneity of the measure, which the proof for q = 2\Sup{r} never uses. In d = 2 the per-H inequality holds in every case we tested: all crossing configurations on the tori 2×4, 3×4, 4×4, 3×6 and 4×6, for nine homogeneous and 501 random inhomogeneous reflection-symmetric product measures.† There the obstruction is subtler. On the tori Z\Sub{W} × Z\Sub{N} with W ∈ \{3, 4\} and p = 1/2, the twist gap is Δ\Sub{N} = R(s) − R(s+e) = Σ\Sub{j} c\Sub{j} λ\Sub{j}\Sup{N}, where all λ\Sub{j} lie in (0, 1) and the c\Sub{j} are integers in \{4, 2, 1, −3\}.† (These recurrences have orders 16 and 69 and are obtained by exact Berlekamp–Massey. For W = 3 at p = 1/4 and 3/4 the recurrences have order 17 and the same structure.) Some c\Sub{j} are negative, so (Δ\Sub{N}) is not a moment sequence. Hence Δ\Sub{N} ≠ ⟨v, T\Sup{N} v⟩ for any positive operator T, and Lemma 4 at q = 1 holds there because the leading positive mode dominates. \subsection*{\pdfbookmark[2]{7.5 A cell-wise form of Lemma 4}{h16}7.5 A cell-wise form of Lemma 4} Fix μ, a spatial twist s ≠ 0 and a temporal twist e ≠ 0. For a configuration ω let W(ω) := span\{s, e\} ∩ V(P(ω))\Sup{⊥}, the twists in span\{s, e\} that vanish on every cycle of P(ω). Since D\Sub{s} ∖ D\Sub{s+e} = \{W = ⟨s⟩\} and D\Sub{s+e} ∖ D\Sub{s} = \{W = ⟨s+e⟩\}, μ\Sub{N,p,q}(D\Sub{s}) − μ\Sub{N,p,q}(D\Sub{s+e}) = Z\Sup{−1} Σ\Sub{m,k} N\Sub{m,k} p\Sup{m} (1−p)\Sup{\textbar{}X\Sup{(i)}\textbar{}−m} q\Sup{k}, N\Sub{m,k} := \#\{ω : \textbar{}ω\textbar{} = m, b\Sub{i−1}(P(ω)) = k, W(ω) = ⟨s⟩\} − \#\{ω : \textbar{}ω\textbar{} = m, b\Sub{i−1}(P(ω)) = k, W(ω) = ⟨s+e⟩\}. \textbf{Conjecture E (cell-wise twist inequality).} N\Sub{m,k} ≥ 0 for all N, m, k, μ, s and e. Conjecture E implies Lemma 4 for every q \textgreater{} 0 simultaneously. By Theorem D, it therefore implies the characteristic-2 sharp threshold for every q ≥ 1, including the Bernoulli case. It has two equivalent forms. Within a cell, \textbar{}Z\Sup{i−1}(P(ω))\textbar{} = 2\Sup{k+c} is constant, so N\Sub{m,k} ≥ 0 says that among pairs (ω, f) with ω in the cell, solutions of δf = s on ω are at least as many as solutions of δf = s + e on ω. Exchanging the sums turns this into Σ\Sub{c∈[s]} N\Sup{(m,k)}(X ∖ supp c) ≥ Σ\Sub{c∈[s+e]} N\Sup{(m,k)}(X ∖ supp c), where c runs over all cocycles in the class, equivalently over dual cycles, and N\Sup{(m,k)}(S) counts the configurations ω ⊆ S in the cell. We enumerated all configurations on the following tori and found N\Sub{m,k} ≥ 0 in every cell, with no exceptions:† the cubical tori 2×2×2 and 2×2×3 (d = 3, i = 2; 2²⁴ and 2³⁶ configurations), for all nine admissible pairs (μ, n), and the tori 3×3, 3×4, 4×4, 3×5 and 4×5 (d = 2, i = 1; up to 2⁴⁰ configurations), for n = e\Sub{1}\Sup{*} + e\Sub{2}\Sup{*} and both transfer directions. For Lemma 4 itself at q = 1 on longer tori, we use an exact transfer operator. Its states are the affine relations between the gauge fields of the first and the current layer, taken modulo coboundaries. On Z\Sub{2} × Z\Sub{2} × Z\Sub{N} with the long direction as μ, it gives R(s) − R(s+e) \textgreater{} 0 for every N up to the limit of double precision: N ≤ 25 at p = 1/2, N ≤ 14 at p = 0.3, and N ≤ 30 at p = 0.7, for the temporal twist e\Sub{13}\Sup{*}, and N ≤ 25 at p = 1/2 for the temporal twist e\Sub{13}\Sup{*} + e\Sub{23}\Sup{*}.† The gap decays geometrically, with ratio 0.2396 at p = 1/2. In d = 4, Monte Carlo sampling of Bernoulli plaquette percolation on T⁴\Sub{N} (N = 4, 6, 8; up to 10⁵ samples per point) is consistent with R(n) ≤ R(s\Sub{μ}(n)) for all 63 twists: the smallest observed difference R(s\Sub{μ}(n)) − R(n) is −1·10⁻⁵, well within the statistical error.† It also finds a transition at p = 1/2 that is already sharp at N = 6: V = 0 in every sample at p = 0.45 and V = H in every sample at p = 0.55. The oblique Lagrangian span\{e\Sub{12} + e\Sub{34}, e\Sub{13} + e\Sub{24}, e\Sub{14} + e\Sub{23}\} occurs once in 2.2·10⁵ samples at p = 1/2 (N = 4, 6, 8 combined). \subsection*{\pdfbookmark[2]{7.6 Translation orbits of the layers}{h17}7.6 Translation orbits of the layers} Proposition 13 shows that reflection symmetry alone does not give Lemma 4 at q = 1. The experiments below indicate that Lemma 4 depends on spatial homogeneity, and they point to the orbit averages on which it may rest. Fix the transfer direction μ and let G ≅ (Z/NZ)\Sup{d−1} be the group of translations orthogonal to μ. A \emph{class} is the set of i-cells of a given type I in a given time layer: spatial cells of type I in the layer x\Sub{μ} = t, or temporal cells of type I in the gap [t, t+1]. Each class is a single free G-orbit. For g = (g\Sub{c}) ∈ G\Sup{C}, where C is the set of classes, g·ω translates the part of ω in each class c by g\Sub{c}. Let R be the group generated by G\Sup{C} and, in each class separately, the reflections of the spatial coordinates. These maps are not automorphisms of X, but they are bijections of the configuration space that preserve every class-wise homogeneous product measure. Our tests of Lemma 4 under random product measures are summarized in the following table.† In all tests with measures homogeneous within each class Lemma 4 at q = 1 held, and the Ising gauge theory (q = 2) behaved the same way. {\footnotesize \begin{longtable}{|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-10\tabcolsep-6\arrayrulewidth)*215/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-10\tabcolsep-6\arrayrulewidth)*5067/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-10\tabcolsep-6\arrayrulewidth)*802/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-10\tabcolsep-6\arrayrulewidth)*941/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-10\tabcolsep-6\arrayrulewidth)*2974/10000\relax}|} \hline \textbf{q} & \textbf{Product measure} & \textbf{Torus} & \textbf{Trials} & \textbf{Result} \\ \hline \endfirsthead \hline \textbf{q} & \textbf{Product measure} & \textbf{Torus} & \textbf{Trials} & \textbf{Result} \\ \hline \endhead 1 & homogeneous within each class, with arbitrary dependence on t and on the cell type; all three temporal twists & Z\Sub{2} × Z\Sub{2} × Z\Sub{N}, N = 2, 3 & 210 & no violation \\ \rowrule 1 & parameters varying from cell to cell & Z\Sub{2} × Z\Sub{2} × Z\Sub{2} & 240 & 2 violations, up to 3·10⁻² \\ \rowrule 2 & time-dependent, spatially homogeneous couplings & — & 800 & no violation (negative values only of order 10⁻¹⁵, i.e. round-off) \\ \rowrule 2 & couplings that also vary from plaquette to plaquette & — & — & violations of relative size up to 1.3·10⁻⁴ \\ \lastrule \end{longtable}} \textbf{Proposition 14.} There are configurations ω, in d = 2, 3 and 4, such that Σ\Sub{g ∈ G\Sup{C}} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(g·ω) \textless{} 0. \begin{itemize}[label=\discbullet] \item{} \emph{d = 3, i = 1 (bonds).} Take the torus Z\Sub{3} × Z\Sub{3} × Z\Sub{2}, with time t ∈ \{0, 1\}, s = e\Sub{1}\Sup{*} and e = e\Sub{t}\Sup{*}. In time slice 0, let the bonds of direction 1 be all sites off the diagonal \{(0,0), (1,1), (2,2)\} and the bonds of direction 2 the diagonal, with temporal bonds at \{(0,1), (1,0)\}. In slice 1, let the direction-1 bonds be at \{(0,0), (0,1), (0,2), (2,0), (2,2)\} and the direction-2 bonds at \{(0,1), (0,2), (1,1), (1,2), (2,0), (2,1)\}, with temporal bonds at \{(0,1), (1,0), (1,1), (2,0)\}. With one class pinned, the orbit sum over the remaining 9⁵ translations is −243 (486 positive and 729 negative terms).† \item{} \emph{d = 4, i = 2.} Thicken this example by a circle S¹\Sub{u} of length 2: replace every open bond b by the two plaquettes b × f, f an edge of S¹\Sub{u}. A 2-chain Σ c\Sub{b,f} b × f is a cycle iff c\Sub{b,f} = c\Sub{b} does not depend on f and \{b : c\Sub{b} = 1\} is a 1-cycle. Hence the 2-cycles are exactly γ × S¹\Sub{u} for 1-cycles γ, and D\Sub{n∧e\Sub{u}\Sup{*}} on the thickened torus equals D\Sub{n} on the original one. Translations along S¹\Sub{u} act trivially on the nonempty classes, so on Z\Sub{3} × Z\Sub{3} × Z\Sub{2} × Z\Sub{2} the sum over the translations of the nonempty classes, one of them pinned, is −243·2⁵ = −7776.† \item{} \emph{d = 2, i = 1.} On Z\Sub{6} × Z\Sub{39} (μ = 2), take 13 copies of the three-row block with rows \{1,3,5\}, \{0,1,3,4\}, \{1,3,5\} and gaps \{0,3,5\}, \{2,3,5\}, \{0\}. The uniform measure on its orbit has μ(D\Sub{s}) − μ(D\Sub{s+e}) = −6.98·10⁻⁹. Two programs compute this value exactly with big integers, and a third in floating point.† \end{itemize} The mechanism behind these examples is the following. Between two gaps with one open cell each, a block of layers contributes a displacement law ν on Z\Sub{2W}, where the extra factor 2 records the winding parity. The orbit sum of m stacked copies of a block is the sum of the antiperiodic Fourier coefficients of ν\Sup{*m}. Reflection sends ν(x) to ν(−x). If the class patterns are chiral, ν is in general not symmetric, its leading antiperiodic coefficients are r e\Sup{±iα} with α ≠ 0, and m copies contribute (2/W) r\Sup{m} cos(mα) + O(r′\Sup{m}) with r′ \textless{} r. For the d = 2 block above, α = 0.1286 and the sum first becomes negative at m = 13.† Every subset of Z\Sub{W} with W ≤ 5 is a translate of its reflection, so in d = 2 chirality first occurs at W = 6.† The same mechanism gives negative q-weighted orbit sums at q = 2 (on Z\Sub{7} × Z\Sub{232}) and q = 4 (on Z\Sub{7} × Z\Sub{378}).† Since at q = 2 Lemma 4 holds for every reflection-symmetric product measure (§7.4), the obstruction comes from chirality rather than from the value of q. The uniform measure on such an orbit is independent across classes and translation-invariant within each class, and it violates Lemma 4. Thus, at q = 1, Lemma 4 can fail for class-wise translation-invariant measures. Homogeneous product measures are also invariant under the reflections of each class, and the conjecture below keeps that symmetry. Conjecture E is not affected, because its sums run over all configurations with given \textbar{}ω\textbar{} and b\Sub{i−1}. \textbf{Conjecture F (reflection-orbit twist inequality).} Let q ≥ 1. For every configuration ω, Σ\Sub{r ∈ R} q\Sup{b\Sub{i−1}(P(r·ω))} ( 1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}} )(r·ω) ≥ 0. \textbf{Proposition 15.} Conjecture F implies Lemma 4 for the random-cluster weight q\Sup{b\Sub{i−1}} applied to any product measure that is homogeneous within each class. In particular it implies Lemma 4 for μ\Sub{N,p,q}, hence Theorem A at that q by Theorem D; at q = 1 this is the Bernoulli case. \emph{Proof.} A law on the subsets of a class that is invariant under the translations and reflections of the class is a mixture of uniform laws on single orbits of these maps, and a homogeneous product law is such a law. By independence across classes, the base measure is then a mixture of uniform measures on R-orbits. The weight q\Sup{b\Sub{i−1}} is applied configuration by configuration, so μ(D\Sub{s}) − μ(D\Sub{s+e}) is a nonnegative combination of the sums in Conjecture F. ∎ Every R-orbit is a union of translation orbits, so a nonnegative sum over every translation orbit implies Conjecture F on that torus. By exhaustive enumeration we found no negative translation-orbit sum on the following tori:† {\footnotesize \begin{longtable}{|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*2619/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*914/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*4311/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*2155/10000\relax}|} \hline \textbf{Torus (d)} & \textbf{Twists} & \textbf{q tested (no negative orbit)} & \textbf{Nonzero orbit sums at q = 1} \\ \hline \endfirsthead \hline \textbf{Torus (d)} & \textbf{Twists} & \textbf{q tested (no negative orbit)} & \textbf{Nonzero orbit sums at q = 1} \\ \hline \endhead 3×4 (d = 2) & — & 0.5, 1, 1.5, 2, 4 & 6 561 \\ \rowrule 4×4 (d = 2) & — & 0.1, 0.5, 1, 1.5, 2, 4, 8 & 390 625 \\ \rowrule 2×2×2 (d = 3) & e\Sub{13}\Sup{*} & 0.5, 1, 1.5, 2, 4 & 20 696 \\ \rowrule 2×2×3 (d = 3) & e\Sub{13}\Sup{*} & 0.1, 0.5, 1, 1.5, 2, 4, 8 & 2 868 660 \\ \rowrule 2×2×3 (d = 3) & e\Sub{23}\Sup{*} & 1 & 2 868 660 \\ \rowrule 2×3×2 (d = 3) & e\Sub{13}\Sup{*} & 0.1, 0.5, 1, 1.5, 2, 4, 8 & 2 762 976 \\ \rowrule 2×3×2 (d = 3) & e\Sub{23}\Sup{*} & 0.5, 1, 1.5, 2, 4, 8 & 988 624 \\ \lastrule \end{longtable}} Conjecture F cannot be refined to single values of b\Sub{i−1}, and it fails for every q \textless{} 1, even after reflection averaging. The smallest example has d = 2, i = 1, on Z\Sub{6} × Z\Sub{2}. Let every row and every gap be the alternating pattern \{0, 2, 4\}; a translation by 1 or a reflection maps it to \{1, 3, 5\}, so the R-orbit has 16 configurations. In 12 of them the open edges form three vertical cycles, with W = ⟨s⟩ and b\Sub{0} = 3. In the other 4 (rows different and gaps different) they form one Hamiltonian cycle of class (1, 1) mod 2, with W = ⟨s+e⟩ and b\Sub{0} = 1. The q-weighted orbit sum is therefore 12q³ − 4q, which is negative for q \textless{} 1/√3.† On Z\Sub{6} × Z\Sub{N} the same pattern gives a threshold value of q that tends to 1 as N → ∞.† Thickening by circles (as in Proposition 14) transfers these orbits to d = 3 and d = 4.† In d = 3, on 2×3×2 with e = e\Sub{23}\Sup{*}, exhaustive enumeration of the 714 136 R-orbits with nonzero sum finds none with a negative total, but 20 with a negative sum at a fixed value of b\Sub{i−1}; these 20 orbits are achiral.† With e = e\Sub{13}\Sup{*} on the same torus, and on 2×2×2 and 2×2×3, every rank part of every R-orbit is nonnegative.† There the rank failures of translation orbits appear to be effects of chirality: on 2×3×2 the translation orbit of the configuration 0x3f719edb6 (in the cell order of the program \texttt{orbtr.c} of the supplement) has rank sums +1584, +3264 and −192, and its R-orbit has rank sums +6336, +12672 and 0.† The following argument, given in outline, explains why q = 1 is the boundary case. Let k be odd, and lift ω to the k-fold cover of the torus in one spatial direction. The covering map and the transfer map act on H\Sub{i}(·; Z) by matrices that are the identity mod 2. Cycles project to cycles and preimages of cycles are cycles, so the lift has the same V, hence the same values of 1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}}. The R-orbit of the lift consists of the lifts of the R-orbit of ω, with equal multiplicities. Consequently the orbit sum at q = 1 is invariant under odd covers, while b\Sub{i−1} is not: in d = 2, a cluster whose cycles generate the horizontal winding a lifts to gcd(k, a) clusters. The orbits on Z\Sub{6} × Z\Sub{N} above are 3-fold covers of configurations on Z\Sub{2} × Z\Sub{N} with one open edge in every row and every gap, where Proposition 16 applies. The reflection averages of the counterexamples of Proposition 14 are positive. For the d = 3 example, the sum over the translations and the reflections of the first spatial coordinate of each class (one class pinned) is +69 984, and with the reflections of both coordinates it is +2 239 488.† In d = 4 the orbits are too large to enumerate directly (8¹⁸ translations on 2×2×2×3). On the spatial torus Z\Sub{2}³ every reflection acts as a translation, so there R-orbits are translation orbits, and we test random orbits and sum them exactly where we can. To sum orbits exactly in d = 4, fix the translations of the temporal classes. A 2-cycle of P is then z = Σ\Sub{t} c\Sub{t} + Σ\Sub{t} γ\Sub{t} × [t, t+1], where γ\Sub{t} is a 1-cycle of the slice formed by edges b for which b × [t, t+1] is open, and c\Sub{t} is a chain of open spatial plaquettes of the layer t with ∂c\Sub{t} = γ\Sub{t−1} + γ\Sub{t}. Moreover ⟨s, z⟩ = Σ\Sub{t} ⟨s, c\Sub{t}⟩, while ⟨e, z⟩ depends only on γ\Sub{0}. The layer t therefore enters only through the relation it induces between the cycle spaces of the two adjacent gaps, together with the values of s on the chains c\Sub{t}. Grouping the translations of each layer by this relation reduces the orbit sum to a sum over distinct relations. Our exact computation does this, and it also skips the translations of the temporal classes for which ⟨e, ·⟩ vanishes on every cycle that any choice of layers allows. It agrees with a direct cochain computation on every configuration tested, and with direct enumeration on a whole orbit of 2×2×2×2.† The orbits we tested in d = 4 are listed in the following table.† No orbit sum was found to be negative. {\footnotesize \begin{longtable}{|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*914/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*1874/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*914/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-8\tabcolsep-5\arrayrulewidth)*6297/10000\relax}|} \hline \textbf{Torus} & \textbf{Twist pairs} & \textbf{Random orbits per pair} & \textbf{Orbit sums} \\ \hline \endfirsthead \hline \textbf{Torus} & \textbf{Twist pairs} & \textbf{Random orbits per pair} & \textbf{Orbit sums} \\ \hline \endhead 2×2×2×3 & (e\Sub{12}\Sup{*}, e\Sub{34}\Sup{*}), (e\Sub{12}\Sup{*}, e\Sub{14}\Sup{*}), and the spatial and temporal halves of Σ\Sub{I} e\Sub{I}\Sup{*} & 1500 & 3705 of the 4500 computed exactly: 3605 are 0, 100 positive, none negative; of the other 795, 792 positive by Monte Carlo (by more than 5σ) and 3 zero in every one of 2·10⁸ samples \\ \rowrule 2×2×2×2 & the three above and (e\Sub{12}\Sup{*} + e\Sub{13}\Sup{*}, e\Sub{24}\Sup{*}) & 3000 & Monte Carlo with 3·10⁵ samples per orbit: more than 3σ above zero or zero in every sample, except for 3 orbits, which are 0 when summed exactly \\ \rowrule 3×2×2×2 & the first three above & 1500 & Monte Carlo with 4·10⁵ samples per orbit: more than 3σ above zero or zero in every sample, except for 12 orbits, which are positive when summed exactly \\ \lastrule \end{longtable}} One of the exact zeros on 2×2×2×3 is the orbit whose Monte Carlo mean is −1.2·10⁻⁷ ± 0.9·10⁻⁷ after 10¹⁰ samples.† The coarser groupings that we tested fail. Translating all classes of a time step together gives negative orbit sums on 2×2×2 and 3×4. In d = 4, translating all spatial classes of a layer together and all temporal classes of a gap together gives negative orbit sums on 2×2×2×2, for 19 to 40 of 20 000 random patterns per twist pair; the exact per-class sums of eight of these patterns are positive.† On the 3×4 torus, which has eight classes, averaging over the translations of any six classes (with the other two fixed) gives a nonnegative sum for every configuration. For every choice of five classes, some configuration gives a negative sum.† \textbf{Proposition 16 (one temporal cell per gap, d = 2).} Let d = 2, i = 1, and work on Z\Sub{W} × Z\Sub{N} with μ = 2. Fix arbitrary patterns H\Sub{t} of open edges in the rows and exactly one open vertical edge in each gap. Translate each row and each gap by independent uniform elements of Z\Sub{W}. Then E[1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}}] ≥ 0. \emph{Proof.} \begin{enumerate}[label=\arabic*.] \item{} If some row is full, it is an s-odd, e-even cycle, so both events fail. Assume no row is full. \item{} Every cycle crosses each gap exactly once, so there is at most one cycle. It exists exactly when, for every t, the columns c\Sub{t−1} and c\Sub{t} of the vertical edges below and above row t lie in a common arc of row t. \item{} Given the columns, the rows are independent. With r\Sub{t} = (c\Sub{t} − c\Sub{t−1}) mod W, row t joins them to the right with probability f\Sub{t}(r\Sub{t}) and to the left with probability f\Sub{t}(W − r\Sub{t}). Here f\Sub{t}(ℓ) is the probability that ℓ consecutive edges are open after the random translation, with f\Sub{t}(0) = 1 and f\Sub{t}(W) = 0. If Σ\Sub{t} r\Sub{t} = mW, the winding of the cycle in the first direction is m − \#\{left\}. So E[1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}} \textbar{} c] = (−1)\Sup{m} Π\Sub{t} u\Sub{t}(r\Sub{t}), where u\Sub{t}(r) = f\Sub{t}(r) − f\Sub{t}(W − r). \item{} The columns are independent and uniform, and (−1)\Sup{m} 1\{W \textbar{} Σ r\Sub{t}\} = W\Sup{−1} Σ\Sub{k=0}\Sup{W−1} exp(iπ(2k+1)Σ\Sub{t} r\Sub{t}/W). Hence E[1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}}] = W\Sup{−N} Σ\Sub{k=0}\Sup{W−1} Π\Sub{t} Û\Sub{t}(k), where Û\Sub{t}(k) = Σ\Sub{\textbar{}r\textbar{}\textless{}W} f\Sub{t}(\textbar{}r\textbar{}) e\Sup{iπ(2k+1)r/W}. \item{} Each factor is nonnegative. For r ≥ 1, f\Sub{t}(r) = W\Sup{−1} Σ\Sub{A} (\textbar{}A\textbar{} + 1 − r)\Sub{+}, the sum running over the arcs A of row t. Write F\Sub{n}(θ) = \textbar{}Σ\Sub{j=0}\Sup{n} e\Sup{ijθ}\textbar{}² ≥ 0 for the Fejér kernel. Then Û\Sub{t}(k) = 1 − W\Sup{−1} Σ\Sub{A} (\textbar{}A\textbar{} + 1) + W\Sup{−1} Σ\Sub{A} F\Sub{\textbar{}A\textbar{}}(π(2k+1)/W) ≥ 0, because each arc is followed by a closed edge, so Σ\Sub{A} (\textbar{}A\textbar{} + 1) ≤ W. ∎ \end{enumerate} The proof shows how the assumption of one temporal cell per gap is used. After averaging over the translations of each class, every row contributes a translation-invariant kernel whose Fourier transform is nonnegative at every frequency, in particular at the antiperiodic modes selected by the temporal twist. The same proof applies, with changes we do not write out, in d = 3, i = 2 when the open temporal cells of every gap form one full line of type \{1, 3\}; the role of the arcs is then played by maximal runs of fully open rows of spatial plaquettes. It also gives the inequality rank by rank in a larger class. \textbf{Proposition 17 (thin or full gaps, d = 2).} Let d = 2, i = 1, μ = 2, and suppose every gap of ω is thin (one open edge) or full (all W edges open), and at least one gap is thin. Then for every k, Σ\Sub{r ∈ R : b\Sub{0}(P(rω)) = k} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(rω) ≥ 0, and the same holds with R replaced by the translations G\Sup{C}. \emph{Proof.} \begin{enumerate}[label=\arabic*.] \item{} A full row is a horizontal cycle, and then both events fail. Cut the time circle at the thin gaps into m blocks of consecutive rows. Inside a block the full gaps join (x, t) to (x, t+1) for every x, so the components of the block are \{α\} × (rows of the block), where α runs over the arcs, isolated vertices included, of the union U of the row patterns of the block. If U = Z\Sub{W} there is a horizontal cycle and both events fail again. Assume neither happens. \item{} Every cycle with odd vertical winding uses the edge of each thin gap exactly once. So there is a cycle through all thin edges, its class (h, 1) mod 2 spans V, and the events give +1 if h = 0 and −1 if h = 1. Such a cycle exists exactly when, in every block, the columns of the thin edges below and above lie in one arc of U. Its horizontal winding is D/W, where D is the sum over the blocks of the displacement inside the arc. On these configurations, b\Sub{0} = Σ\Sub{j} c\Sub{j} − m + 1, where c\Sub{j} is the number of arcs of the union of block j. \item{} For each block j and each c, let ν\Sub{j}\Sup{(c)}(d), d ∈ Z\Sub{2W}, count the images of the rows of block j with c arcs in which the columns 0 and d mod W lie in one arc with lifted displacement ≡ d mod 2W. Summing over the translations of the thin gaps gives Σ\Sub{r : b\Sub{0} = k} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(rω) = W Σ\Sub{c\Sub{1}+…+c\Sub{m} = k+m−1} [(ν\Sub{1}\Sup{(c\Sub{1})} ⊛ … ⊛ ν\Sub{m}\Sup{(c\Sub{m})})(0) − (ν\Sub{1}\Sup{(c\Sub{1})} ⊛ … ⊛ ν\Sub{m}\Sup{(c\Sub{m})})(W)], with ⊛ the cyclic convolution on Z\Sub{2W}. \item{} Group the images of the rows of block j into orbits of the diagonal translation, which moves all rows of the block together. On such an orbit the arcs of U are translated, so c is constant, and summing over the orbit gives Σ\Sub{α} (\textbar{}α\textbar{} − \textbar{}d\textbar{})\Sub{+} for \textbar{}d\textbar{} \textless{} W, where \textbar{}α\textbar{} is the number of vertices of the arc α. Every ν\Sub{j}\Sup{(c)} is thus a nonnegative combination of tents supported in \textbar{}d\textbar{} ≤ W − 1, hence symmetric and nonincreasing in \textbar{}d\textbar{} on Z\Sub{2W}. \item{} Cyclic convolution preserves this property, so each bracket in step 3 is ≥ 0. ∎ \end{enumerate} A gap is thick if it has at least two open edges. Full gaps are thick, but Proposition 17 does not extend to the other thick gaps: the rank-refined inequality fails there (the example on Z\Sub{6} × Z\Sub{2} above, which has no thin gap), and the single-block argument fails too, because the averaged displacement law of a block between two thin gaps need not be nonincreasing in \textbar{}d\textbar{}. For the two-row block on Z\Sub{8} with rows \{0, 1, 4, 5\} and internal gap \{0, 2, 4\}, the law averaged over translations and reflections takes the values 64 and 96 at \textbar{}d\textbar{} = 5 and 6.† Its antiperiodic Fourier coefficients are positive, but other blocks have negative ones.† For an arbitrary configuration with a thin gap, the proof of Proposition 17 still reduces the orbit sum to a convolution of block laws. Cut the time circle at the thin gaps into blocks of consecutive rows. For the j-th block, let \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}(d), d ∈ Z\Sub{2W}, be the number of elements of the product of the dihedral groups of its rows and internal gaps for which: \begin{itemize}[label=\discbullet] \item{} the block has no cycle of odd horizontal winding; and \item{} the column-0 vertex of its bottom row is joined, inside the block, to its top row with lifted horizontal displacement ≡ d mod 2W. \end{itemize} Reflecting all classes of the block at once shows that \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}(−d) = \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}(d). Put h\Sub{j}(d) := \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}(d) − \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}(d + W). This function is symmetric and antiperiodic (h\Sub{j}(d + W) = −h\Sub{j}(d)). \textbf{Proposition 17′.} Let d = 2, i = 1, μ = 2, and let ω have m ≥ 1 thin gaps. \begin{itemize}[label=\discbullet] \item{} (i) Σ\Sub{r∈R} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(rω) = 2W · (h\Sub{1} ⊛ ⋯ ⊛ h\Sub{m})(0), where ⊛ is the cyclic convolution on Z\Sub{2W}. \item{} (ii) For W ≥ 3 and a symmetric antiperiodic h on Z\Sub{2W}, put F\Sub{j}(h) := Σ\Sub{d∈Z\Sub{2W}} h(d) cos(π(2j + 1)d/W) for j = 0, …, W − 1. Let K\Sub{W} be the set of such h with F\Sub{0} \textgreater{} 0, \textbar{}F\Sub{j}\textbar{} ≤ F\Sub{0} and F\Sub{j} ≥ −2F\Sub{0}/(W − 2) for every j. Then K\Sub{W} is closed under ⊛, and h(0) ≥ 0 for every h ∈ K\Sub{W}. \item{} (iii) For W ≥ 3 and every block with n rows, h\Sub{j} = 0 or h\Sub{j} ∈ K\Sub{W}, whenever n = 1 and W ≤ 16, n = 2 and W ≤ 12, n = 3 and W ≤ 7, or n = 4 and W ≤ 5.† \end{itemize} Consequently, Conjecture F at q = 1 holds on Z\Sub{W} × Z\Sub{N} for every configuration with a thin gap whose blocks all have at most n\Sub{W} rows. Here n\Sub{W} = 4 for 3 ≤ W ≤ 5, n\Sub{W} = 3 for W = 6, 7, n\Sub{W} = 2 for 8 ≤ W ≤ 12, and n\Sub{W} = 1 for 13 ≤ W ≤ 16. \emph{Proof.} \begin{enumerate}[label=\arabic*.] \item{} As in step 1 of the proof of Proposition 17, a cycle of odd horizontal and even vertical winding makes both events fail. Such a cycle exists if and only if some block contains a cycle of odd horizontal winding. \item{} Assume no block contains one. As in step 2, every cycle with odd vertical winding uses the edge of each thin gap exactly once, and cycles with even vertical winding lie inside the blocks. Every cycle with odd vertical winding therefore has horizontal winding ≡ Σ\Sub{j} D\Sub{j}/W mod 2, where D\Sub{j} is the lifted displacement across block j, and the events give +1 or −1 according to this parity. \item{} Fix the positions of the thin edges below and above block j. The images of the block in which these columns are joined with displacement d are counted by ν\Sub{j}(d)·1\{d ≡ c′ − c mod W\}, where c and c′ are the two columns and ν\Sub{j} is the law over translations only. \item{} Summing over the translations of the thin gaps, as in step 3, gives Σ\Sub{r∈G\Sup{C}} (1\Sub{D\Sub{s}} − 1\Sub{D\Sub{s+e}})(rω) = W · (ν\Sub{1} ⊛ ⋯ ⊛ ν\Sub{m})\Sup{−}(0), where f\Sup{−}(d) := f(d) − f(d + W). \item{} Adding the reflections of the classes replaces each ν\Sub{j} by \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{j}, by multilinearity. A reflection of a thin gap is a translation of its single edge, which gives a factor 2 for each of the m thin gaps. \item{} Finally, (f ⊛ g)\Sup{−} = ½ f\Sup{−} ⊛ g\Sup{−}. To see this, write f = f\Sub{+} + ½ f\Sup{−}, where f\Sub{+}(d) := ½(f(d) + f(d + W)) is W-periodic; then f\Sub{+} ⊛ g\Sup{−} = 0, because g\Sup{−} is antiperiodic. Applying this identity repeatedly gives 2\Sup{m} W (\uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{1} ⊛ ⋯ ⊛ \uni{03BD0304}{\ensuremath{\bar{\nu}}}\Sub{m})\Sup{−}(0) = 2W (h\Sub{1} ⊛ ⋯ ⊛ h\Sub{m})(0). This proves (i). \end{enumerate} For (ii), convolution multiplies the coefficients F\Sub{j}, which are real. \begin{itemize}[label=\discbullet] \item{} If F\Sub{j}(h) \textless{} 0 \textless{} F\Sub{j}(g), then F\Sub{j}(h)F\Sub{j}(g) ≥ −2F\Sub{0}(h)F\Sub{j}(g)/(W − 2) ≥ −2F\Sub{0}(h)F\Sub{0}(g)/(W − 2). The other sign cases are immediate, so K\Sub{W} is closed under ⊛. \item{} By Fourier inversion, 2W·h(0) = Σ\Sub{j=0}\Sup{W−1} F\Sub{j}, since only the antiperiodic frequencies contribute. The terms j = 0 and j = W − 1 both equal F\Sub{0}, and each of the other W − 2 terms is at least −2F\Sub{0}/(W − 2). Hence h(0) ≥ 0. \end{itemize} Part (iii) is a finite computation over all block types up to the symmetries of each class, in floating point.† Apart from the blocks with h(d) = 0 for 0 \textless{} d \textless{} W, for which \textbar{}F\Sub{j}\textbar{} = F\Sub{0} for every j, the largest ratio \textbar{}F\Sub{j}\textbar{}/F\Sub{0} for j ≠ 0, W − 1 is 0.983. It is attained by a row with one edge and tends to 1 as W grows. The smallest ratio F\Sub{j}/F\Sub{0} is −0.022, against a threshold −2/(W − 2) ≤ −1/7. The consequence follows from (i)–(iii). If some h\Sub{j} = 0, the orbit sum is 0. ∎ Neither of the two natural cones contains the antiperiodic part of every block law, which is why K\Sub{W} is needed. \begin{itemize}[label=\discbullet] \item{} Nonnegative antiperiodic Fourier coefficients: this cone fails for the blocks with some F\Sub{j} \textless{} 0, such as the two-row block on Z\Sub{10} with ratio −0.022. \item{} Antiperiodic parts that are nonincreasing on \{0, …, W\}: this cone fails for 17 two-row block types on Z\Sub{11}, all at the middle step d = 5.† \end{itemize} Proposition 17′ covers only blocks of at most n\Sub{W} rows, and the cone K\Sub{W} shrinks as W grows, so it does not settle the thin-gap case. For thick gaps one can ask for positivity frequency by frequency. Call [N − 1, N] the closing gap. Let E be an event that depends only on which gaps are thick and on the pattern of the closing gap up to translations and reflections. For a ∈ Z\Sub{2W}, let X\Sub{a} be the torus obtained from Z\Sub{W} × [0, N] by gluing (x, N) to (x + a, 0), let s\Sub{a} ∈ H\Sup{1}(X\Sub{a}; F₂) be the class of the double cover obtained in the same way from Z\Sub{2W} × [0, N], and let Ψ\Sup{E}(a) be the probability of E ∩ D\Sub{s\Sub{a}} for Bernoulli bond percolation P\Sub{p} on X\Sub{a}. Then X\Sub{W} = X\Sub{0} = Z\Sub{W} × Z\Sub{N}, s\Sub{0} = s and s\Sub{W} = s + e, so D\Sub{s\Sub{a}} is D\Sub{s} for a = 0 and D\Sub{s+e} for a = W. For odd k ∈ Z\Sub{2W} put Δ\Sup{E}\Sub{k} := W\Sup{−1} Σ\Sub{a∈Z\Sub{2W}} Ψ\Sup{E}(a) e\Sup{−iπka/W}. Reflecting every class maps X\Sub{a} to X\Sub{−a} and preserves E, so Ψ\Sup{E}(−a) = Ψ\Sup{E}(a) and Δ\Sup{E}\Sub{k} is real. Then P\Sub{p}(D\Sub{s} ∩ E) − P\Sub{p}(D\Sub{s+e} ∩ E) = Σ\Sub{k odd} Δ\Sup{E}\Sub{k}, where k runs over the W odd residues mod 2W. Moreover Δ\Sup{E}\Sub{2j+1} = F\Sub{j}(Ψ\Sup{E}\Sub{−})/(2W), where Ψ\Sup{E}\Sub{−}(a) := Ψ\Sup{E}(a) − Ψ\Sup{E}(a + W) and F\Sub{j} is as in Proposition 17′(ii). Convolution multiplies the F\Sub{j}, so positivity of every Δ\Sup{E}\Sub{k} is preserved under the convolutions of the block decomposition. When the closing gap is thin, every Δ\Sup{E}\Sub{k} was positive in all the cases we computed,† although single block laws can have negative coefficients F\Sub{j} (see above). When the gaps are thick, it can fail: \begin{itemize}[label=\discbullet] \item{} On Z\Sub{9} × Z\Sub{2} at p = 1/2, Δ\Sup{E}\Sub{7} \textless{} 0 for each of three events: E requiring all gaps to be thick, E requiring the closing gap to be thick (the other gap is free), and E the whole space (the total over all closing patterns). For the total, Δ\Sup{E}\Sub{7} ≈ −1.17·10\Sup{−3}.† \item{} For these events the sum Σ\Sub{k} Δ\Sup{E}\Sub{k} = P\Sub{p}(D\Sub{s} ∩ E) − P\Sub{p}(D\Sub{s+e} ∩ E) stays positive in every case we computed.† \item{} Among the tori with W ≥ 3 and N ≥ 2, Z\Sub{9} × Z\Sub{2} is the smallest such example by area: for these three events and for E requiring the closing gap to be thin, every other such torus with WN ≤ 18 has all Δ\Sup{E}\Sub{k} \textgreater{} 0 for every p ∈ (0, 1), and every value we computed with W ≤ 8 and N ≥ 2 is positive (for W ≤ 7 in floating point, at p = 0.1, 0.3, 0.5, 0.7, 0.9 and N ≤ 30, or N ≤ 25 in some runs at W = 7).† \end{itemize} The failures are not confined to very wide tori. At p = 1/2, Δ\Sup{E}\Sub{7} \textless{} 0 for each of the three events also on Z\Sub{11} × Z\Sub{3}, Z\Sub{12} × Z\Sub{3}, Z\Sub{13} × Z\Sub{4} and Z\Sub{14} × Z\Sub{5}, where in the last case N \textgreater{} W/3.† On these tori every other Δ\Sup{E}\Sub{k} is positive, except Δ\Sup{E}\Sub{5} on Z\Sub{12} × Z\Sub{3} for E requiring all gaps to be thick.† At p = 1/2 and for W ≥ 3, the smallest failing width is W = 6, 9, 11, 13 and 14 for N = 1, …, 5: for N ≤ 4 every narrower torus is positive, and for N = 5 so is every torus with W ≤ 12, while the positivity of Z\Sub{13} × Z\Sub{5} rests on a Monte Carlo estimate.† On the square tori Z\Sub{N} × Z\Sub{N} with 3 ≤ N ≤ 11, every Δ\Sup{E}\Sub{k} is positive for the three events at p = 1/2.† Apart from the Monte Carlo estimate and the whole space on Z\Sub{11} × Z\Sub{3} and Z\Sub{12} × Z\Sub{3}, which we computed exactly, these values were computed in floating point with a first-order bound on the rounding error. Since P\Sub{p} restricted to such an E is a mixture of uniform measures on R-orbits (proof of Proposition 15), on each failing torus the failure for E requiring all gaps to be thick gives an R-orbit of configurations with all gaps thick that has a negative coefficient. So, on the failing tori, positivity frequency by frequency does not yield the thick-gap case of Conjecture F. In the scaling limit there are failures on wide, square and long tori. Assume that at p = 1/2, as W → ∞ with N/W → ρ, the probabilities of the homology classes of the open clusters on X\Sub{a} converge, uniformly in a, to their values in the Coulomb-gas description of critical percolation on the torus [DSZ87, Pin94] with modular parameter τ = (−a + iN)/W. This convergence is not proved, and no rate is known. In this limit each Δ\Sup{E}\Sub{k} converges, and the three events have the same limits Δ\Sub{k}, since they differ only on configurations with a gap of at most one open edge, which have probability at most N(W + 1)2\Sup{−W}. Then some limiting sector is negative at each of the following ratios: \begin{itemize}[label=\discbullet] \item{} Δ\Sub{7} on a grid of step 0.01 in 0.03 ≤ ρ \textless{} 0.9768, and Δ\Sub{11} and Δ\Sub{13} at ρ = 0.02;† \item{} Δ\Sub{5} on a grid of step 0.02 in 2.8082 ≤ ρ ≤ 12;† \item{} a negative tail at ρ = 0.98, 1, 1.2, 1.5, 2, 2.5, 3 and 4: every computed Δ\Sub{k} from some k\Sub{0}(ρ) on is negative, with k\Sub{0} = 1031, 987, 665, 405, 209, 123, 77 and 35.† On the square ρ = 1 every computed Δ\Sub{k} with k ≥ 987 is negative (all odd k up to 1201, and k = 1501, 1503, 2001, 3001 and 4001), beginning with Δ\Sub{987} ≈ −5.2·10\Sup{−1333}.† \end{itemize} We have no proof that every ratio fails. The lattice and the limit agree in sign only in part: on Z\Sub{9} × Z\Sub{3} and Z\Sub{12} × Z\Sub{4} every lattice sector is positive, although the limiting Δ\Sub{5} and Δ\Sub{7} at ρ = 1/3 are negative, and no lattice failure with N ≥ W is seen at computable sizes.† On square tori, then, positivity frequency by frequency fails in the limit. But the negative sectors we found on the square have k ≥ 987, and the sectors satisfy k ≤ W, so this failure needs W ≥ 987, and possibly much larger widths. The odd-cover argument sketched above gives Conjecture F at q = 1 for the odd covers of the configurations of Propositions 16 and 17. These covers have thick gaps. \section*{\pdfbookmark[1]{8. Divisibility of the classes of giant cycles near p = 1/2}{h18}8. Divisibility of the classes of giant cycles near p = 1/2} In this section P is Bernoulli(p) plaquette percolation on X = T⁴\Sub{N} (q = 1, i = 2), and P• is its dual complex, which is Bernoulli(1 − p) plaquette percolation on the dual torus X•. Propositions 8 and 10 constrain the giant cycles below 1/2 with coefficients F₂ and Z/4. Here we study integral giant cycles. Above 1/2 they exist in every direction [DKS20, Theorem 1 with F = Q]. We show that just above 1/2 the classes of all integral giant cycles are divisible by any fixed integer M. As in §1, we use the strict-shift argument of [AG91] in the finite-volume form of [BBR14]. Its input is the local conversion lemma of §8.3, which §§8.4–8.7 prove for all configurations. \textbf{Conventions.} \begin{itemize}[label=\discbullet] \item{} All cells carry fixed orientations. For an abelian group Γ put V\Sub{Γ}(P) := im(H₂(P; Γ) → H₂(X; Γ)), so V = V\Sub{F₂} and Λ\Sub{P} := V\Sub{Z}(P) ⊆ H₂(X; Z) = Z⁶. For m = 2\Sup{k} put L\Sub{m}(P) := red₂ V\Sub{Z/m}(P) ⊆ F₂⁶. Here red\Sub{r} is reduction mod r of coefficients and of classes, and v\Sub{ℓ} is the ℓ-adic valuation. Reduction of coefficients maps cycles to cycles, so red₂Λ\Sub{P} ⊆ L₈(P) ⊆ L₄(P) ⊆ V(P). \item{} Through the translation by (½, ½, ½, ½) we identify H₂(X•; Γ) with H₂(X; Γ) = Λ²Γ⁴ = Γ⁶. The translation acts trivially on homology. The cellular intersection number of a 2-cycle of X with a 2-cycle of X• is then the intersection form ⟨z, w⟩, defined by z ∧ w = ⟨z, w⟩ e₁∧e₂∧e₃∧e₄. It is unimodular, and ⟨e\Sub{I}, e\Sub{J}⟩ = ±1\{I ∩ J = ∅\}. \item{} The signed permutations g ∈ B₄ are automorphisms of the lattice Z⁴. They act on X, on Γ⁶ = Λ²Γ⁴ by Λ²g, and they preserve the law of P. An element x ∈ (Z/m)⁶ is \emph{partnered} if ⟨x, gx⟩ ≠ 0 in Z/m for some g ∈ B₄. For m = 2 this is the notion of §7, because signs act trivially mod 2. \item{} B\Sub{t}(y) is the set of plaquettes whose centres lie within l∞ distance t of the centre of the plaquette y, and \textbar{}B\Sub{t}\textbar{} is its cardinality. \item{} In this section the letter d denotes an ℓ-adic exponent, not the dimension. \end{itemize} \subsection*{\pdfbookmark[2]{8.1 Statement}{h19}8.1 Statement} \textbf{Theorem W (divisibility near p = 1/2).} For every integer M ≥ 2 there is w\Sub{M} \textgreater{} 0 such that, as N → ∞: \begin{itemize}[label=\discbullet] \item{} (i) P\Sub{p}(Λ\Sub{P} ⊄ MZ⁶) → 0 for all p \textless{} 1/2 + w\Sub{M}. More precisely, for each prime power ℓ\Sup{d} dividing M, P\Sub{p}(V\Sub{Z/ℓ\Sup{n}}(P) ⊄ ℓ\Sup{d}(Z/ℓ\Sup{n})⁶) → 0, where n := 3d if ℓ = 2 and n := 3d − 1 if ℓ is odd. \item{} (ii) For p ∈ (1/2, 1/2 + w\Sub{M}), with probability tending to 1, Λ\Sub{P} ⊆ MZ⁶ and rank Λ\Sub{P} = 6. In particular M⁶ divides the index [Z⁶ : Λ\Sub{P}]. \item{} (iii) If M is even, P\Sub{p}(L₄(P) ⊄ C\Sup{⊥}) → 0 for all p \textless{} 1/2 + w\Sub{M}. \item{} (iv) For p \textgreater{} 1/2 − w\Sub{M} and each prime power ℓ\Sup{d} dividing M, with probability tending to 1, V\Sub{Z/ℓ\Sup{n}}(P) ⊇ ℓ\Sup{n−d}(Z/ℓ\Sup{n})⁶. \end{itemize} Since 1/2 \textless{} 1/2 + w\Sub{M}, statements (i) and (iii) hold at p = 1/2. For M = 2, part (i) says that L₈(P) = 0 with probability tending to 1 for p \textless{} 1/2 + w₂: no giant mod-2 cycle lifts to a Z/8-cycle. Part (ii) says that just above 1/2 the integral giant lattice has full rank, while all integral giant cycles have even classes. Part (iii) says that no partnered class lifts to a Z/4-cycle, and part (iv) says that for p \textgreater{} 1/2 − w₂, in particular just below 1/2, each class 4\uni{00730303}{\~{s}} (s ∈ F₂⁶, \uni{00730303}{\~{s}} a lift) is the class of a giant Z/8-cycle. For each M the index of Λ\Sub{P} is a multiple of M⁶ for p ∈ (1/2, 1/2 + w\Sub{M}), with probability tending to 1, so it tends to infinity as p ↓ 1/2 in this sense. However, Theorem W does not settle the case q = 1 of Theorem A, which concerns V(P) below 1/2. The constant w\Sub{M} is explicit (Remark 1 in §8.8) and very small. \textbf{Corollary W1 (an F\Sub{ℓ}-only window).} Let ℓ be an odd prime. For p ∈ (1/2 − w\Sub{ℓ}, 1/2 + w\Sub{ℓ}), with probability tending to 1, V\Sub{Z/ℓ²}(P) = ℓ(Z/ℓ²)⁶. For p ∈ (1/2, 1/2 + w\Sub{ℓ}), moreover V\Sub{F\Sub{ℓ}}(P) = F\Sub{ℓ}⁶ with probability tending to 1. In particular, just above 1/2 there are giant F\Sub{ℓ}-cycles in all classes, and none of them lifts to a Z/ℓ²-cycle. \textbf{Corollary W2 (torsion just below 1/2).} For p ∈ (1/2 − w\Sub{M}, 1/2), with probability tending to 1, the torsion subgroup T of H₁(P; Z) has order at least M⁶ and exponent divisible by M, and the linking map of Proposition 21 maps T onto a subgroup of Q⁶/Z⁶ that contains (M⁻¹Z/Z)⁶. \subsection*{\pdfbookmark[2]{8.2 Algebraic inputs}{h20}8.2 Algebraic inputs} We collect the facts about coefficients that the proof uses: the partner rule at every level, exact duality, the subcritical bound, and the linking map of the torsion. \textbf{Lemma 18 (partner rule).} \begin{itemize}[label=\discbullet] \item{} (a) Let ℓ be an odd prime. Every nonzero z ∈ (Z/ℓ)⁶ is partnered. Hence each x ∈ (Z/ℓ\Sup{a})⁶ of order ℓ\Sup{b} with 2b \textgreater{} a is partnered. \item{} (b) Each x ∈ (Z/2\Sup{a})⁶ of order 2\Sup{b} with 2b ≥ a + 2 is partnered. \end{itemize} \emph{Proof.} (a) Write z = Σ z\Sub{ij} e\Sub{i}∧e\Sub{j} and put α₁ := z₁₂z₃₄, α₂ := z₁₃z₂₄, α₃ := z₁₄z₂₃. Let g\Sub{jk} be the diagonal sign change of coordinates j and k. A direct computation gives ⟨z, z⟩ = 2(α₁ − α₂ + α₃), ⟨z, g₃₄z⟩ = 2(α₁ + α₂ − α₃), ⟨z, g₂₄z⟩ = −2(α₁ + α₂ + α₃), ⟨z, g₂₃z⟩ = 2(−α₁ + α₂ + α₃). For odd ℓ these four linear forms span all linear forms in (α₁, α₂, α₃) mod ℓ (the 3×3 minors of the coefficient matrix of α₁ − α₂ + α₃, α₁ + α₂ − α₃, α₁ + α₂ + α₃, −α₁ + α₂ + α₃ are ±4). So z is partnered unless α₁ ≡ α₂ ≡ α₃ ≡ 0. In that case the support of z, read as a set of edges of K₄, contains no two disjoint edges, so it lies in a star or in a triangle. Then z = v₁ ∧ v₂ is decomposable over F\Sub{ℓ}. With T := span(v₁, v₂), the number ⟨z, gz⟩ = det[v₁, v₂, gv₁, gv₂] is nonzero if and only if gT ∩ T = 0. After a coordinate permutation, T = span(e₁, w) with w = (0, w₂, w₃, w₄) ≠ 0 (star), or T = \{x₄ = 0, n·x = 0\} with n = (n₁, n₂, n₃, 0) ≠ 0 (triangle). The proof of Lemma 12′ uses only linear algebra and polynomial identities, so it holds over F\Sub{ℓ}. For a suitable pair (x, y) ≠ (0, 0) of coordinates of w, resp. of n, it gives g, g′ ∈ B₄ with gT ∩ T = 0 if and only if 2xy ≠ 0, and g′T ∩ T = 0 if and only if x² − y² ≠ 0. For odd ℓ these two numbers do not both vanish. For the second claim of (a), write x = ℓ\Sup{a−b}y with y ≢ 0 mod ℓ. Then ⟨x, gx⟩ = ℓ\Sup{2(a−b)}⟨y, gy⟩. Choosing g with ⟨y, gy⟩ ≢ 0 mod ℓ gives ⟨x, gx⟩ ≢ 0 mod ℓ\Sup{2(a−b)+1}, and 2(a − b) + 1 ≤ a. For (b), write x = 2\Sup{a−b}y with y ≢ 0 mod 2. The reduction of y mod 4 has order 4, so Proposition 10(i) gives g with ⟨y, gy⟩ ≢ 0 mod 4. Hence ⟨x, gx⟩ ≢ 0 mod 2\Sup{2(a−b)+2}, and 2(a − b) + 2 ≤ a. ∎ The identities in the proof, and the conclusion of (a) for ℓ = 3, 5, 7, are also checked by computer.† \textbf{Lemma 19 (exact duality).} Let m ≥ 2 and A := Z/m. For any configuration, V\Sub{A}(P•) = V\Sub{A}(P)\Sup{⊥} with respect to ⟨·,·⟩ on A⁶. \emph{Proof.} The cells of X that are not in P are the closed plaquettes and all 3- and 4-cells. Their duals are exactly the cells of P•. Sending a cell σ to the cochain dual to σ* identifies C\Sub{k}(X; A)/C\Sub{k}(P; A) with C\Sup{4−k}(P•; A), and the boundary map with the coboundary map up to signs. Hence H₂(X, P; A) ≅ H²(P•; A). The ring A is self-injective, so Hom\Sub{A}(·, A) is exact and H²(P•; A) ≅ Hom\Sub{A}(H₂(P•; A), A). The composite H₂(X; A) → H₂(X, P; A) → Hom\Sub{A}(H₂(P•; A), A) sends x to y ↦ ⟨x, y⟩. Exactness of H₂(P; A) → H₂(X; A) → H₂(X, P; A) gives V\Sub{A}(P) = V\Sub{A}(P•)\Sup{⊥}. The dual complex of P• is P, up to the translation, so the same argument gives V\Sub{A}(P•) = V\Sub{A}(P)\Sup{⊥}. ∎ With A = Q in place of Z/m, the same argument gives dim V\Sub{Q}(P) + dim V\Sub{Q}(P•) = 6. \textbf{Proposition 20 (the subcritical bound at all levels).} Let m ≥ 2 and let x ∈ (Z/m)⁶ be partnered. Then P\Sub{p}(x ∈ V\Sub{Z/m}(P)) → 0 for all p \textless{} 1/2. \emph{Proof.} Choose g ∈ B₄ with ⟨x, gx⟩ ≠ 0. Disjoint cycles have intersection number 0, so x ∈ V\Sub{Z/m}(P) and gx ∈ V\Sub{Z/m}(P•) cannot both hold. At p = 1/2, V\Sub{Z/m}(P•) has the law of V\Sub{Z/m}(P), and g preserves this law. Hence P\Sub{1/2}(x ∈ V\Sub{Z/m}(P)) ≤ 1/2. The event \{x ∈ V\Sub{Z/m}(P)\} is increasing and invariant under translations, whose orbits on plaquettes have size N⁴. By [BR06, Lemma 7], P\Sub{p}(x ∈ V\Sub{Z/m}(P)) → 0 for all p \textless{} 1/2. ∎ Propositions 8 and 10 are the cases m = 2 and m = 4 of Proposition 20, together with the classification of the partnered elements. \textbf{Proposition 21 (torsion linking).} Let P be a configuration with rank Λ\Sub{P} = 6, and let T• be the torsion subgroup of H₁(P•; Z). For [γ] ∈ T• choose m ≥ 1 and an integral 2-chain W on P• with ∂W = mγ, and put Ψ([γ]) := m⁻¹[W mod m] ∈ Q⁶/Z⁶, where [W mod m] ∈ (Z/m)⁶ is the class of the Z/m-cycle W mod m and m⁻¹ maps (Z/m)⁶ onto (m⁻¹Z/Z)⁶. Then Ψ is a well-defined homomorphism T• → Q⁶/Z⁶ with image Λ\Sub{P}\Sup{\#}/Z⁶, where Λ\Sub{P}\Sup{\#} := \{y ∈ Q⁶ : ⟨y, Λ\Sub{P}⟩ ⊆ Z\}. In particular \textbar{}T•\textbar{} ≥ [Z⁶ : Λ\Sub{P}]. If Λ\Sub{P} ⊆ MZ⁶, the image contains (M⁻¹Z/Z)⁶ and M divides the exponent of T•. \emph{Proof.} \begin{enumerate}[label=\arabic*.] \item{} By the remark after Lemma 19, V\Sub{Q}(P•) = 0. Since H₂(X•; Z) is torsion-free, all integral 2-cycles of P• are null-homologous in X•. Given (γ, m), the chain W is unique up to such cycles, which do not change [W mod m]. Replacing γ by γ + ∂c replaces W by W + mc, and replacing m by km replaces W by kW. Ψ is additive. \item{} As in the proof of Lemma 19, H₂(X, P; Z) ≅ H²(P•; Z). The class x ∈ H₂(X; Z) maps to the class of a cochain f\Sub{x} with f\Sub{x}(W) = ⟨x, W⟩ for integral 2-chains W on P•. The kernel of H₂(X; Z) → H₂(X, P; Z) is Λ\Sub{P}. \item{} The cochain f\Sub{x} vanishes on the cycles of P•, by step 1. Its class therefore lies in the subgroup Ext(H₁(P•; Z), Z) of H²(P•; Z), which is identified with Hom(T•, Q/Z) through χ\Sub{f}([γ]) := f(W)/m mod 1, for mγ = ∂W. \item{} Hence x ↦ χ\Sub{f\Sub{x}} is injective on Z⁶/Λ\Sub{P}. Since ⟨x, W⟩ mod m is the Z/m intersection number of x with the cycle W mod m, χ\Sub{f\Sub{x}}([γ]) = ⟨x, Ψ([γ])⟩ mod 1. For x ∈ Λ\Sub{P} the cochain f\Sub{x} is a coboundary, so ⟨x, Ψ(T•)⟩ ⊆ Z and Ψ(T•) ⊆ Λ\Sub{P}\Sup{\#}/Z⁶. \item{} The group Z⁶/Λ\Sub{P} is finite, and ⟨·,·⟩ induces a perfect pairing between Z⁶/Λ\Sub{P} and Λ\Sub{P}\Sup{\#}/Z⁶, because the form is unimodular. By step 4 the annihilator of Ψ(T•) in Z⁶/Λ\Sub{P} is trivial, so Ψ(T•) = Λ\Sub{P}\Sup{\#}/Z⁶. \end{enumerate} Finally \textbar{}Λ\Sub{P}\Sup{\#}/Z⁶\textbar{} = [Z⁶ : Λ\Sub{P}]. If Λ\Sub{P} ⊆ MZ⁶ then Λ\Sub{P}\Sup{\#} ⊇ M⁻¹Z⁶, and any preimage of an element of order M has order divisible by M. ∎ \subsection*{\pdfbookmark[2]{8.3 The strict-shift argument}{h21}8.3 The strict-shift argument} We first show that a local conversion lemma rules out certain giant cycles slightly above 1/2. \textbf{Setting.} Fix a prime ℓ, integers a, d ≥ 1, n := a + d, the ring A := Z/ℓ\Sup{n}, and a submodule Θ ⊆ A⁶ such that \begin{itemize}[label=\discbullet] \item{} (Θ1) ℓ\Sup{d} A⁶ ⊆ Θ; \item{} (Θ2) for each v ∈ A⁶ ∖ Θ, the reduction red\Sub{ℓ\Sup{a}}(v) ∈ (Z/ℓ\Sup{a})⁶ is partnered. \end{itemize} An A-cycle is \emph{Θ-odd} if its class does not lie in Θ. Let E\Sub{Θ}(ω) be the event that the plaquette set ω carries a Θ-odd A-cycle, that is, V\Sub{A}(ω) ⊄ Θ. It is increasing. Theorem W uses two instances. \begin{itemize}[label=\discbullet] \item{} (I1) Given a prime power ℓ\Sup{d}, take a := 2d if ℓ = 2 and a := 2d − 1 if ℓ is odd, so n = 3d or 3d − 1, and Θ := ℓ\Sup{d} A⁶. Then E\Sub{Θ} = \{V\Sub{Z/ℓ\Sup{n}}(P) ⊄ ℓ\Sup{d}(Z/ℓ\Sup{n})⁶\}. Condition (Θ1) is trivial. For (Θ2), let v ∉ ℓ\Sup{d} A⁶, so v\Sub{ℓ}(v) = j \textless{} d. Then red\Sub{ℓ\Sup{a}}(v) has order ℓ\Sup{b} with b = a − j ≥ a − d + 1. For ℓ = 2 this gives 2b ≥ 2d + 2 = a + 2, and for odd ℓ it gives 2b ≥ 2d = a + 1. In both cases Lemma 18 applies. \item{} (I2) Take ℓ = 2, a = d = 1 and Θ := red₂⁻¹(C\Sup{⊥}) ⊆ (Z/4)⁶. Then E\Sub{Θ} = \{L₄(P) ⊄ C\Sup{⊥}\}. Condition (Θ1) holds because red₂(2A⁶) = 0, and (Θ2) is Proposition 8. \end{itemize} \textbf{Enhancement.} Fix ρ ≥ 1. Give each plaquette y an independent bit ξ\Sub{y} of mean s ∈ [0, 1], independent of P. For a plaquette set ω let ω\Sup{e} := ω ∪ \{y : ξ\Sub{y} = 1 and ∂y ∈ ∂C₂(ω ∩ B\Sub{ρ}(y); Z/ℓ\Sup{a})\}. The condition says that ∂y bounds a Z/ℓ\Sup{a}-chain on the plaquettes of ω within distance ρ of y. The configuration ω\Sup{e} is increasing in ω and in ξ. \textbf{Lemma 22.} Let N \textgreater{} 2ρ + 2. Then V\Sub{Z/ℓ\Sup{a}}(ω\Sup{e}) = V\Sub{Z/ℓ\Sup{a}}(ω). Hence, if E\Sub{Θ}(ω\Sup{e}) holds, then V\Sub{Z/ℓ\Sup{a}}(ω) contains a partnered element. \emph{Proof.} For each added y choose a Z/ℓ\Sup{a}-chain g\Sub{y} on ω ∩ B\Sub{ρ}(y) with ∂g\Sub{y} = ∂y. Then y − g\Sub{y} is a Z/ℓ\Sup{a}-cycle in a box of side less than N, hence null-homologous. For a Z/ℓ\Sup{a}-cycle z of ω\Sup{e}, the cycle z − Σ\Sub{y} z(y)(y − g\Sub{y}) lies on ω and has the class of z. For the second claim, let v ∈ V\Sub{A}(ω\Sup{e}) ∖ Θ. Then red\Sub{ℓ\Sup{a}}(v) ∈ V\Sub{Z/ℓ\Sup{a}}(ω\Sup{e}) = V\Sub{Z/ℓ\Sup{a}}(ω), and it is partnered by (Θ2). ∎ A plaquette τ is \emph{E\Sub{Θ}-pivotal} for ω if E\Sub{Θ}(ω ∪ τ) holds and E\Sub{Θ}(ω ∖ τ) fails. \textbf{Local conversion lemma LCL\Sub{Θ}(ρ, κ).} There is N₀ such that the following holds for all N ≥ N₀. Let ω be any set of plaquettes of T⁴\Sub{N} and let τ be E\Sub{Θ}-pivotal for ω. Then there are a plaquette set ω′ with ω′ Δ ω ⊆ B\Sub{κ}(τ) and a plaquette x ∈ B\Sub{κ}(τ) ∖ ω′ such that: \begin{itemize}[label=\discbullet] \item{} E\Sub{Θ}(ω′ ∪ x) holds; \item{} E\Sub{Θ}(ω′) fails; \item{} ∂x ∈ ∂C₂(ω′ ∩ B\Sub{ρ}(x); Z/ℓ\Sup{a}). \end{itemize} \textbf{Proposition 23 (strict shift).} Assume (Θ2) and LCL\Sub{Θ}(ρ, κ). Put κ′ := κ + 2ρ + 1 and δ := 1/(\textbar{}B\Sub{κ′}\textbar{}·16\Sup{\textbar{}B\Sub{κ′}\textbar{}}). Then P\Sub{p}(E\Sub{Θ}(P)) → 0 for all p \textless{} 1/2 + δ/2. \emph{Proof.} We follow [AG91] in the finite-volume form of [BBR14]. Let N ≥ max(N₀, 2κ′ + 2). \begin{enumerate}[label=\arabic*.] \item{} \emph{Russo.} Put θ\Sub{N}(p, s) := P\Sub{p,s}(E\Sub{Θ}(P\Sup{e})). A plaquette is \emph{p-pivotal} in (ω, ξ) if changing its state changes E\Sub{Θ}(ω\Sup{e}), and a bit is \emph{s-pivotal} if changing it changes E\Sub{Θ}(ω\Sup{e}). Let Piv\Sub{p} and Piv\Sub{s} count them. The event is increasing in finitely many independent bits, so ∂\Sub{p}θ\Sub{N} = E Piv\Sub{p} and ∂\Sub{s}θ\Sub{N} = E Piv\Sub{s}. \item{} \emph{Surgery.} Let τ be p-pivotal in (ω, ξ). We keep τ open and switch off the bits in B\Sub{κ′}(τ) one at a time, in a fixed order. \begin{itemize}[label=\circbullet] \item{} Suppose E\Sub{Θ} first fails when the bit of y is switched off. We output y, together with the configuration just before that step. The bit of y is s-pivotal there. \item{} Otherwise let \uni{03BE0303}{\ensuremath{\tilde{\xi}}} be ξ with all bits in B\Sub{κ′}(τ) switched off. Then τ is still p-pivotal in (ω, \uni{03BE0303}{\ensuremath{\tilde{\xi}}}), because E\Sub{Θ}((ω ∖ τ)\Sup{e}) is increasing in the bits. Let Y\Sub{far} be the set of plaquettes outside ω that are added by bits at distance more than κ′ from τ. Their balls B\Sub{ρ} avoid B\Sub{κ+ρ}(τ), so Y\Sub{far} does not change when ω changes inside B\Sub{κ}(τ). Now τ is E\Sub{Θ}-pivotal for ω ∪ Y\Sub{far}, and LCL\Sub{Θ} gives \uni{03C90302}{\ensuremath{\hat{\omega}}}′ and x. Put ω′ := \uni{03C90302}{\ensuremath{\hat{\omega}}}′ ∖ Y\Sub{far}, which agrees with ω outside B\Sub{κ}(τ). In (ω′, \uni{03BE0303}{\ensuremath{\tilde{\xi}}}) the enhanced configuration is \uni{03C90302}{\ensuremath{\hat{\omega}}}′ ∪ x if the bit of x is on and \uni{03C90302}{\ensuremath{\hat{\omega}}}′ if it is off, because \uni{03C90302}{\ensuremath{\hat{\omega}}}′ ∩ B\Sub{ρ}(x) = ω′ ∩ B\Sub{ρ}(x). So the bit of x is s-pivotal. We output (ω′, \uni{03BE0303}{\ensuremath{\tilde{\xi}}}) and x. Pivotality of a bit does not depend on its own value, and \uni{03BE0303}{\ensuremath{\tilde{\xi}}}\Sub{x} = 0. \end{itemize} \item{} \emph{Distortion.} An output differs from (ω, ξ) in at most \textbar{}B\Sub{κ′}\textbar{} plaquette states and \textbar{}B\Sub{κ′}\textbar{} bits, all in B\Sub{κ′}(τ), and bits change only from 1 to 0. For s ≤ 1/2 its probability is therefore at least min(p, 1 − p)\Sup{\textbar{}B\Sub{κ′}\textbar{}} times that of (ω, ξ). The output site lies in B\Sub{κ′}(τ), so an output arises from at most \textbar{}B\Sub{κ′}\textbar{} choices of τ and 4\Sup{\textbar{}B\Sub{κ′}\textbar{}} original values in the ball. Hence E Piv\Sub{s} ≥ δ E Piv\Sub{p} for p ∈ [1/4, 3/4] and s ∈ [0, 1/2]. \item{} \emph{Integration.} By steps 1 and 3, t ↦ θ\Sub{N}(p₀ − δt, t) is nondecreasing on [0, 1/2] for p₀ ∈ [1/2, 1/2 + δ/2). Hence P\Sub{p₀}(E\Sub{Θ}(P)) = θ\Sub{N}(p₀, 0) ≤ θ\Sub{N}(p₀ − δ/2, 1/2). By Lemma 22, the right side is at most the probability that V\Sub{Z/ℓ\Sup{a}}(P) contains a partnered element, for P with parameter p₀ − δ/2 \textless{} 1/2. This tends to 0 by Proposition 20 and a union bound over the finitely many partnered elements. For p₀ \textless{} 1/2 the same bound applies directly, with s = 0. ∎ \end{enumerate} \subsection*{\pdfbookmark[2]{8.4 Descending funnels}{h22}8.4 Descending funnels} We now reduce LCL\Sub{Θ} to the existence of a design inside a fixed region. Section 8.5 will reduce the existence of designs to three finite conditions on the region, and §8.7 will verify them. Fix (ℓ, a, d, Θ) as in §8.3, satisfying (Θ1). \textbf{Regions.} A \emph{region} is a finite union 𝔇 of closed unit 4-cubes of the lattice that is star-shaped about one of its vertices o, hence contractible. It is placed with o a vertex of the pivotal plaquette τ, and N is large enough that 𝔇 embeds in X and that all A-cycles supported in 𝔇 are null-homologous in X. Let 𝒫(𝔇) be the set of plaquettes that are faces of cubes of 𝔇. We assume τ ∈ 𝒫(𝔇). \begin{itemize}[label=\discbullet] \item{} An edge of 𝔇 is \emph{safe} if all six plaquettes containing it lie in 𝒫(𝔇). Let NS be the set of edges of 𝔇 that are not safe. A plaquette outside 𝒫(𝔇) can contain an edge of 𝔇 only if that edge lies in NS. \item{} NS carries a fixed total order. \item{} Let O := ω ∖ 𝒫(𝔇) be the \emph{outside}. Since O ⊆ ω ∖ τ and τ is pivotal, O carries no Θ-odd cycle. \end{itemize} \textbf{Relative cycles.} A \emph{relative cycle} on O is an A-chain U on O whose boundary lies in 𝔇. Its \emph{trace} γ := −∂U is a 1-cycle on NS. Since 𝔇 is contractible, γ = ∂G for a chain G in 𝒫(𝔇), and two such G differ by a cycle in 𝔇, which is null-homologous. We call U \emph{Θ-odd} if the cycle U + G is Θ-odd for one, equivalently every, such G. The restriction to O of a Θ-odd cycle of ω ∪ τ is a Θ-odd relative cycle. \textbf{Lemma 24 (the min-max edge).} Let e* be the least edge e ∈ NS for which some Θ-odd relative cycle on O has trace supported on \{f ∈ NS : f ≤ e\}. Fix such a U and put γ := −∂U. Then: \begin{itemize}[label=\discbullet] \item{} (a) no Θ-odd relative cycle on O has trace supported on \{f ∈ NS : f \textless{} e*\}; \item{} (b) e* ∈ supp γ; \item{} (c) each endpoint of e* lies on an edge of supp γ other than e*, and that edge is \textless{} e*. \end{itemize} \emph{Proof.} (a) A nonzero such trace contradicts the minimality of e*, and a zero trace gives a Θ-odd cycle on O. (b) follows from (a). For (c), ∂γ = 0 and γ(e*) ≠ 0, so each endpoint of e* lies on another edge of supp γ, and all edges of supp γ are ≤ e*. ∎ \textbf{Port networks.} Let t ∈ \{0, …, n − 1\}. A \emph{port network of type t} at a plaquette π is a set 𝒩 of plaquettes with π ∉ 𝒩, together with a plaquette x ∉ 𝒩 ∪ \{π\}, such that: \begin{itemize}[label=\discbullet] \item{} (N1) each A-chain on 𝒩 whose boundary lies on the four edges of ∂π has boundary b∂π with b ∈ ℓ\Sup{t+d}A; \item{} (N2) for every b ∈ ℓ\Sup{t} A there is an A-chain on 𝒩 ∪ \{x\} with boundary b∂π; \item{} (N3) ∂x ∈ ∂C₂(𝒩 ∩ B\Sub{ρ}(x); Z/ℓ\Sup{a}). \end{itemize} Here ℓ\Sup{t+d}A = 0 when t + d ≥ n. \textbf{Funnel designs.} Let t := v\Sub{ℓ}(γ(e*)). A \emph{funnel design} for (γ, e*) is a 2-chain F in 𝒫(𝔇) together with a port network (𝒩, x) of type t at a plaquette π of supp F, such that: \begin{itemize}[label=\discbullet] \item{} (F1) ∂F = γ; \item{} (F2) the NS edges of the plaquettes of supp F are all ≤ e*; \item{} (F3) e* lies in exactly one plaquette Q₀ of supp F; \item{} (F4) there are pairwise distinct plaquettes Q₀, Q₁, …, Q\Sub{J} = π of supp F such that Q\Sub{k−1} and Q\Sub{k} share a safe edge σ\Sub{k}, and σ\Sub{k} lies in no other plaquette of supp F; \item{} (F5) 𝒩 ∪ \{x\} ⊆ 𝒫(𝔇), and no edge of 𝒩 ∪ \{x\} other than the four edges of ∂π lies in a plaquette of O ∪ (supp F ∖ π). \end{itemize} \textbf{Lemma 25.} Assume (Θ1) and 𝒫(𝔇) ⊆ B\Sub{κ}(τ). If a funnel design exists, then ω′ := O ∪ (supp F ∖ π) ∪ 𝒩 and x satisfy the three conditions of LCL\Sub{Θ}(ρ, κ) at τ. \emph{Proof.} Each plaquette of 𝒩 has an edge outside ∂π, so by (F5) 𝒩 is disjoint from O ∪ (supp F ∖ π). For the same reason x ∉ ω′. Moreover ω′ Δ ω ⊆ 𝒫(𝔇) ⊆ B\Sub{κ}(τ). \begin{itemize}[label=\discbullet] \item{} \emph{E\Sub{Θ}(ω′ ∪ x).} Ξ := U + F is a Θ-odd cycle on O ∪ supp F. By (F3), F(Q₀) = ±γ(e*). At σ\Sub{k} only Q\Sub{k−1} and Q\Sub{k} carry F, by (F4), and ∂F(σ\Sub{k}) = γ(σ\Sub{k}) = 0 because σ\Sub{k} is safe. Hence F(π) = ±γ(e*), which has valuation t. By (N2) there is a chain n* on 𝒩 ∪ \{x\} with ∂n* = Ξ(π)∂π. Then Ξ − Ξ(π)π + n* is a cycle on ω′ ∪ x. It differs from Ξ by a cycle in 𝒫(𝔇), which is null-homologous, so it is Θ-odd. \item{} \emph{The enhancement condition} is (N3), since 𝒩 ⊆ ω′. \item{} \emph{¬E\Sub{Θ}(ω′).} Suppose Ω is a Θ-odd cycle on ω′. \begin{enumerate}[label=\arabic*.] \item{} Write Ω = Ω₁ + n with n on 𝒩. By (F5), ∂n lies on the edges of ∂π, and a 1-cycle on these four edges is a multiple of ∂π. So ∂n = b∂π with b ∈ ℓ\Sup{t+d}A by (N1). Then Ω′ := Ω₁ + bπ is a cycle on O ∪ supp F with the class of Ω, because n − bπ is a cycle in 𝒫(𝔇). Hence Ω′ is Θ-odd, and Ω′(π) = b. \item{} Write F(π) = ℓ\Sup{t} ν with ν a unit, and b = ℓ\Sup{t+d}μ. Put ζ := ℓ\Sup{d} μ ν⁻¹ ∈ ℓ\Sup{d} A and Ω″ := Ω′ − ζΞ. Then Ω″(π) = 0. Its class differs from that of Ω′ by an element of ℓ\Sup{d} A⁶ ⊆ Θ, by (Θ1), so Ω″ is Θ-odd. \item{} Write Ω″ = U″ + Y with U″ on O and Y on supp F. No plaquette of O contains a safe edge, so ∂Y vanishes on safe edges. By (F4), Y(Q\Sub{k−1}) = ±Y(Q\Sub{k}) for all k. Hence Y(Q₀) = ±Y(π) = 0. \item{} The trace of U″ is ∂Y. It lies on NS edges of plaquettes of supp F, which are ≤ e* by (F2), and it vanishes at e* by (F3), since Y(Q₀) = 0. Also U″ is Θ-odd, since U″ + Y = Ω″. This contradicts Lemma 24(a). ∎ \end{enumerate} \end{itemize} \subsection*{\pdfbookmark[2]{8.5 Existence of designs}{h23}8.5 Existence of designs} \textbf{Holes and sublevel complexes.} Fix an axis box 𝔅 ⊆ 𝔇 with 𝔅 ± 1 ⊆ 𝔇 (the box enlarged by 1 in each coordinate), and a plaquette π\Sub{𝔅} on a face of 𝔅. A port network (𝒩, x) at π\Sub{𝔅} \emph{fits} in 𝔅 if all vertices of 𝒩 ∪ \{x\} other than the four vertices of π\Sub{𝔅} lie in the interior of 𝔅. Call a vertex, an edge or a plaquette \emph{available} if none of its vertices lies in the interior of 𝔅. All NS edges are available, since an edge with a vertex in the interior of 𝔅 has all its plaquettes in 𝒫(𝔅 ± 1) and is therefore safe. For h ∈ NS ∪ \{∞\}, let 𝒦\Sub{\textless{}h} be the 2-complex whose vertices are the available vertices, whose edges are the available safe edges and the NS edges below h, and whose 2-cells are the available plaquettes of 𝒫(𝔇) all of whose edges are edges of 𝒦\Sub{\textless{}h}. Removing an edge from a 2-complex means removing it together with the 2-cells that contain it. \textbf{Lemma 26 (expansions).} Suppose the available safe edges, followed by NS in its order, can be listed so that each edge either joins two components of the graph formed by the available vertices and the earlier edges, or lies in an available plaquette of 𝒫(𝔇) whose other three edges come earlier. Then H₁(𝒦\Sub{\textless{}h}; Z) = 0 for all h. \emph{Proof.} Start from the available vertices and no edges. A joining edge does not change H₁. An edge that lies in a plaquette whose other three edges come earlier is attached together with that plaquette, which has it as a free face, and this does not change the homotopy type. The remaining 2-cells only kill H₁. The 1-skeleton of 𝒦\Sub{\textless{}h} is an initial segment of the list. ∎ \textbf{Lemma 27 (integral link criterion).} Let 𝒦′ be a 2-complex with H₁(𝒦′; Z) = 0, let σ be an edge of 𝒦′, and let π₁, …, π\Sub{r} be the 2-cells of 𝒦′ that contain σ, with incidence numbers ε\Sub{i} ∈ \{±1\}. If each 1-cycle ∂(π₁ − ε₁ε\Sub{j}π\Sub{j}), j = 2, …, r, bounds an integral 2-chain in 𝒦′ ∖ σ, then H₁(𝒦′ ∖ σ; Z) = 0. The hypothesis holds if the π\Sub{j} are connected to π₁ through lattice 3-cubes whose two faces at σ are 2-cells of 𝒦′ and whose other four faces are 2-cells of 𝒦′ ∖ σ. \emph{Proof.} Let z be a 1-cycle of 𝒦′ ∖ σ, and write z = ∂(c′ + Σ a\Sub{i}π\Sub{i}) in 𝒦′, with c′ supported away from σ. The coefficient of σ gives Σ a\Sub{i}ε\Sub{i} = 0, so Σ a\Sub{i}π\Sub{i} = Σ\Sub{j≥2} a\Sub{j}(π\Sub{j} − ε\Sub{j}ε₁π₁) = −Σ\Sub{j≥2} a\Sub{j}ε\Sub{j}ε₁(π₁ − ε₁ε\Sub{j}π\Sub{j}). Each ∂(π₁ − ε₁ε\Sub{j}π\Sub{j}) avoids σ and bounds in 𝒦′ ∖ σ, so z bounds there. For the second claim, a 3-cube with faces π\Sub{i}, π\Sub{j} at σ has a boundary 2-cycle whose coefficients at π\Sub{i} and π\Sub{j} are proportional to 1 and −ε\Sub{i}ε\Sub{j}, so ∂(π\Sub{i} − ε\Sub{i}ε\Sub{j}π\Sub{j}) is ± the boundary of its other four faces. These relations compose along chains of cubes. ∎ \textbf{Proposition 28 (existence of designs).} Let 𝔇, 𝔅, π\Sub{𝔅} be as above, with 𝔅 ± 1 ⊆ 𝔇, and suppose: \begin{itemize}[label=\discbullet] \item{} (C1) the hypothesis of Lemma 26 holds; \item{} (C2) each NS edge e that is not a joining edge of the list in (C1), and whose endpoints both lie on NS edges below e, lies in an available plaquette Q₀ of 𝒫(𝔇), the \emph{collar plaquette}, whose other three edges are edges of 𝒦\Sub{\textless{}e}, at least one of them safe; \item{} (C3) for each such e, with Q₀ as in (C2), there are pairwise distinct plaquettes Q₀, Q₁, …, Q\Sub{J} = π\Sub{𝔅} (the \emph{strip}) with Q₁, …, Q\Sub{J} 2-cells of 𝒦\Sub{\textless{}e}, and pairwise distinct safe edges σ\Sub{k} ⊆ Q\Sub{k−1} ∩ Q\Sub{k} (the \emph{spine edges}) lying in no other Q\Sub{j}, such that removing σ\Sub{J}, σ\Sub{J−1}, …, σ₁ one at a time from 𝒦\Sub{\textless{}e} keeps H₁(·; Z) = 0. \end{itemize} Then for each port network (𝒩, x) of type t that fits in 𝔅 at π\Sub{𝔅}, and each (γ, e*) of Lemma 24 with v\Sub{ℓ}(γ(e*)) = t, a funnel design exists. \emph{Proof.} \begin{itemize}[label=\discbullet] \item{} \emph{The edge e* is not a joining edge.} All edges of supp γ are NS edges ≤ e*. If e* joined two components of the earlier graph, then γ − γ(e*)e* would be a chain on that graph with boundary −γ(e*)∂e* ≠ 0, which is impossible. \item{} By Lemma 24(c), both endpoints of e* lie on NS edges below e*. So (C2) gives the collar plaquette Q₀, and (C3) gives the strip. Put F := Σ\Sub{k=0}\Sup{J} α\Sub{k} Q\Sub{k} + F′, where α₀ = ±γ(e*) is chosen so that ∂(α₀Q₀)(e*) = γ(e*), and α\Sub{k} = ±α\Sub{k−1} is chosen so that α\Sub{k−1}Q\Sub{k−1} + α\Sub{k} Q\Sub{k} has no boundary at σ\Sub{k}. \item{} Let 𝒦″ be 𝒦\Sub{\textless{}e*} with σ₁, …, σ\Sub{J} removed. Then γ′ := γ − ∂(Σ α\Sub{k} Q\Sub{k}) is a 1-cycle of 𝒦″: it vanishes at e* and at each σ\Sub{k}, the other edges of Q₀ are edges of 𝒦\Sub{\textless{}e*} by (C2), and the Q\Sub{k} with k ≥ 1 are 2-cells of 𝒦\Sub{\textless{}e*}. \item{} By Lemma 26 and (C3), H₁(𝒦″; Z) = 0. Its H₀ is free, so H₁(𝒦″; A) = 0 by the universal coefficient theorem. Hence γ′ = ∂F′ for an A-chain F′ on 𝒦″. \item{} (F1) holds by construction. \item{} (F2): Q₀ satisfies it by (C2), and all other plaquettes used are 2-cells of 𝒦\Sub{\textless{}e*}. \item{} (F3): no 2-cell of 𝒦\Sub{\textless{}e*} contains e*, so Q₀ is the only plaquette of supp F at e*, and F(Q₀) = α₀ ≠ 0. \item{} (F4): 𝒦″ has no 2-cell containing a spine edge σ\Sub{k}, and Q\Sub{k} contains σ\Sub{k}. Therefore F(Q\Sub{k}) = α\Sub{k} ≠ 0, and σ\Sub{k} lies in no plaquette of supp F other than Q\Sub{k−1} and Q\Sub{k}. \item{} (F5): 𝒩 ∪ \{x\} ⊆ 𝒫(𝔅) ⊆ 𝒫(𝔇). Each edge of 𝒩 ∪ \{x\} other than the edges of ∂π\Sub{𝔅} has a vertex v in the interior of 𝔅, because the only lattice edges spanned by the vertices of π\Sub{𝔅} are those of ∂π\Sub{𝔅}. All sixteen unit cubes at v lie in 𝔅 ⊆ 𝔇, so all plaquettes containing the edge lie in 𝒫(𝔇), and none lies in O. The plaquettes of supp F are available, so none of them contains the edge either. ∎ \end{itemize} By Lemma 25, (C1)–(C3) for a region with 𝒫(𝔇) ⊆ B\Sub{κ}(τ) and a family of fitting port networks of all types t ∈ \{0, …, n − 1\} with 𝒩 ⊆ B\Sub{ρ}(x) prove LCL\Sub{Θ}(ρ, κ), with N₀ := 2κ′ + 2. The conditions (C1)–(C3) concern only 𝔇, 𝔅 and π\Sub{𝔅}: one certificate serves all networks that fit, and all (ℓ, a, d, Θ). \subsection*{\pdfbookmark[2]{8.6 Comb networks}{h24}8.6 Comb networks} All port networks we use share one cross-section and differ only in length. We describe them in a local frame (u₁, u₂, u₃, u₄) with long axis u₁. Let sq(r) be the unit square spanning u₁, u₂ with lower corner (r, 0, 0, 0). \begin{itemize}[label=\discbullet] \item{} A \emph{sheet box} over [lo, hi] with holes at z₁ \textless{} … \textless{} z\Sub{s} in [lo, hi] is the boundary of the 3-dimensional box [lo, hi + 1] × [−2, 3] × [0, 1] × \{0\}, minus the squares sq(z\Sub{i}), which lie in its face u₃ = 0. It is a 2-sphere with s holes, oriented so that its boundary is −Σ\Sub{i} ∂sq(z\Sub{i}). \item{} An \emph{around-tube} from sq(r₁) to sq(r₂), r₂ ≥ r₁ + 3, is the lateral surface of a column of unit 3-cubes. The column goes from sq(r₁) down to u₄ = −2, out to u₃ = −2, along u₁ by 2, up to u₄ = 2, along u₁ at height 1 ≤ u₄ ≤ 2 to r₂, back to u₃ = 0, and down into sq(r₂). A \emph{mirrored} tube is its image under u₄ ↦ −u₄. It leaves upwards and arrives from below. \item{} The \emph{tail} at sq(r₁) is the lateral surface of the column of three unit cubes below sq(r₁). Its bottom face, at u₄ = −3, is the port π. \end{itemize} Fix L\Sub{c} ≥ 0 and L\Sub{g} ≥ 2. The network 𝒩(L\Sub{c}, L\Sub{g}) consists of the following pieces. \begin{itemize}[label=\discbullet] \item{} \emph{Converter} (only if L\Sub{c} ≥ 1): a sheet box over [−2, 4L\Sub{c} + 2] with holes h\Sub{j} := sq(4j) for 0 ≤ j \textless{} L\Sub{c} and x := sq(4L\Sub{c}), mirrored around-tubes from h\Sub{j} to h\Sub{j+1} for 0 ≤ j ≤ L\Sub{c} − 2, and the tail at h₀. \item{} If L\Sub{c} = 0 there is no converter: x := sq(0), and the tail hangs below x. \item{} \emph{Gadget}: with g₀ := 4L\Sub{c} + 8, a sheet box over [g₀ − 2, g₀ + 4L\Sub{g} + 2] with holes y₀ := sq(g₀) and y\Sub{k} := sq(g₀ + 4k) for 1 ≤ k ≤ L\Sub{g}, and around-tubes from y\Sub{k} to y\Sub{k+1} for 1 ≤ k ≤ L\Sub{g} − 1. \item{} The \emph{X-tube}: an around-tube from x to y₀, mirrored if L\Sub{c} = 0. \end{itemize} The enhanced plaquette is x, and x, π ∉ 𝒩. Each piece carries its fundamental chain with the orientation induced by the cubes, written [Π] for the piece Π. The \emph{hole loops} are the boundaries of the squares h\Sub{j}, x, y\Sub{k} and of π. \textbf{Lemma 29 (structure).} For all L\Sub{c} ≥ 0 and L\Sub{g} ≥ 2: \begin{itemize}[label=\discbullet] \item{} (S0) the fundamental chain of a tube with entry square □ and exit square □′ (the port π for the tail) has boundary ε(∂□′ − ∂□) for a sign ε, and each sheet box has boundary −Σ\Sub{i} ∂sq(z\Sub{i}); \item{} (U) the fundamental chains have coefficients ±1; \item{} (S1) the pieces are pairwise plaquette-disjoint; \item{} (S2) the edges lying in exactly one plaquette of 𝒩 are the four edges of ∂π, and each edge that lies in plaquettes of two different pieces lies on a hole loop; \item{} (S3) each piece is connected through its \emph{interior} edges, the edges that lie in exactly two plaquettes of 𝒩, both in that piece; \item{} (S4) the hole loops are pairwise edge-disjoint; \item{} (S5) all vertices of 𝒩 ∪ \{x\} other than the four vertices of π lie in the open box (−3, 4L\Sub{c} + 4L\Sub{g} + 15) × (−3, 4) × (−3, 2) × (−3, 3). \end{itemize} \emph{Proof.} These are finite checks for L\Sub{c} ≤ 7 and 2 ≤ L\Sub{g} ≤ 7.† Each condition concerns a single piece, or the pieces that meet a single edge or a single hole loop. Within a comb the tubes are translates of each other by multiples of 4 along u₁, the X-tube and the tail have fixed shapes, and the sheet boxes are periodic in u₁ away from their ends. Tubes T\Sub{j} and T\Sub{j′} with \textbar{}j − j′\textbar{} ≥ 2 have disjoint u₁-ranges separated by a gap, and a tube meets its sheet box only near its two holes. Hence each configuration of pieces near an edge that occurs for some (L\Sub{c}, L\Sub{g}) already occurs, up to translation, for L\Sub{c}, L\Sub{g} ≤ 4. ∎ \textbf{Lemma 30 (comb modules).} Let A = Z/ℓ\Sup{n} with n = a + d, let t ∈ \{0, …, n − 1\}, and put L\Sub{c} := ℓ\Sup{t} if t ≥ 1, L\Sub{c} := 0 if t = 0, and L\Sub{g} := ℓ\Sup{a}. Then (𝒩(L\Sub{c}, L\Sub{g}), x) is a port network of type t at π in the sense of §8.4, for all ρ ≥ 4L\Sub{c} + 4L\Sub{g} + 18. More precisely: \begin{itemize}[label=\discbullet] \item{} the A-chains on 𝒩 whose boundary lies on ∂π have boundaries exactly ℓ\Sup{t+d}A·∂π; \item{} the A-chains on 𝒩 ∪ \{x\} whose boundary lies on ∂π have boundaries exactly ℓ\Sup{t}A·∂π; \item{} ∂x bounds a Z/ℓ\Sup{a}-chain on the gadget and the X-tube. \end{itemize} \emph{Proof.} \begin{enumerate}[label=\arabic*.] \item{} \emph{Chains are piecewise constant.} Let c be an A-chain on 𝒩 ∪ \{x\} with ∂c supported on ∂π. Let e be an interior edge of a piece Π, lying in the plaquettes ϖ, ϖ′ of Π. Then e ∉ ∂π by (S2), and e ∉ ∂x, because by (S0) each edge of ∂x lies in plaquettes of two different pieces (the X-tube, and the converter sheet box or the tail). Thus c and [Π] both have zero boundary at e. By (U) this gives c(ϖ)/[Π](ϖ) = c(ϖ′)/[Π](ϖ′). By (S3), c = β\Sub{Π}[Π] on Π for some β\Sub{Π} ∈ A. \item{} \emph{Loop equations.} So c = Σ\Sub{Π} β\Sub{Π}[Π] + β\Sub{x} x. By (S0), ∂c is a combination of hole loops, and by (S4) ∂c = b∂π holds if and only if the coefficients of all hole loops vanish and that of ∂π equals b. Replace the coefficient of each tube by its product with the sign ε of (S0). Write β\Sub{c} and β\Sub{g} for the coefficients of the converter and gadget sheet boxes, β\Sup{c}\Sub{j} for the converter tube from h\Sub{j} to h\Sub{j+1}, β\Sup{g}\Sub{k} for the gadget tube from y\Sub{k} to y\Sub{k+1}, β\Sub{π} for the tail and β\Sub{X} for the X-tube. The equations read: \begin{itemize}[label=\circbullet] \item{} gadget: −β\Sub{g} + β\Sub{X} = 0 at y₀; −β\Sub{g} − β\Sup{g}\Sub{1} = 0 at y₁; −β\Sub{g} + β\Sup{g}\Sub{k−1} − β\Sup{g}\Sub{k} = 0 at y\Sub{k} for 2 ≤ k ≤ L\Sub{g} − 1; and −β\Sub{g} + β\Sup{g}\Sub{L\Sub{g}−1} = 0 at y\Sub{L\Sub{g}}; \item{} converter (L\Sub{c} ≥ 1): −β\Sub{c} − β\Sup{c}\Sub{0} − β\Sub{π} = 0 at h₀; −β\Sub{c} + β\Sup{c}\Sub{j−1} − β\Sup{c}\Sub{j} = 0 at h\Sub{j} for 1 ≤ j ≤ L\Sub{c} − 2; −β\Sub{c} + β\Sup{c}\Sub{L\Sub{c}−2} = 0 at h\Sub{L\Sub{c}−1} (for L\Sub{c} = 1 the equation at h₀ is −β\Sub{c} − β\Sub{π} = 0); and −β\Sub{c} − β\Sub{X} + β\Sub{x} = 0 at x; \item{} without converter: −β\Sub{π} − β\Sub{X} + β\Sub{x} = 0 at x; \item{} at ∂π: β\Sub{π} = b. \end{itemize} \item{} The gadget equations give β\Sup{g}\Sub{k} = −kβ\Sub{g} and L\Sub{g}β\Sub{g} = 0, so β\Sub{X} = β\Sub{g} satisfies ℓ\Sup{a} β\Sub{X} = 0, that is, β\Sub{X} ∈ ℓ\Sup{d}A. Each β\Sub{X} ∈ ℓ\Sup{d}A occurs. \item{} The converter equations give β\Sup{c}\Sub{j} = (L\Sub{c} − 1 − j)β\Sub{c} and β\Sub{π} = −L\Sub{c}β\Sub{c}, with β\Sub{c} = β\Sub{x} − β\Sub{X}. Hence b = −L\Sub{c}(β\Sub{x} − β\Sub{X}). Without converter, b = β\Sub{x} − β\Sub{X}. \item{} For t ≥ 1, β\Sub{x} = 0 gives b = L\Sub{c}β\Sub{X} ∈ L\Sub{c}ℓ\Sup{d}A = ℓ\Sup{t+d}A, and arbitrary β\Sub{x} gives b ∈ L\Sub{cA} = ℓ\Sup{t}A. For t = 0, b = β\Sub{x} − β\Sub{X} lies in ℓ\Sup{d}A when β\Sub{x} = 0 and ranges over A otherwise. All such b occur. \item{} Over Z/ℓ\Sup{a}, L\Sub{g}β\Sub{g} = ℓ\Sup{a}β\Sub{g} = 0 for all β\Sub{g}. So β\Sub{g} = β\Sub{X} = 1, β\Sup{g}\Sub{k} = −k and all other coefficients 0 solve all equations except the one at x, and this chain on the gadget and the X-tube has boundary −∂x. \end{enumerate} Finally, by (S5) the network lies in a box of l∞ diameter 4L\Sub{c} + 4L\Sub{g} + 18, so 𝒩 ⊆ B\Sub{ρ}(x). ∎ The modules of Lemma 30 are also computed directly, by exact elimination, for several (ℓ, a, d) and all types.† For (ℓ, a, d) = (2, 1, 1), (2, 2, 1), (3, 1, 1), (5, 1, 1) and every type t, the first two module statements, and the third with the chain allowed anywhere on the network, are also proved in Lean 4, with Mathlib [Mat20], for the networks 𝒩(L\Sub{c}, L\Sub{g}) of the supplement. The proof checks explicit chain and cochain certificates with compiled code (\texttt{native\_\cbrk{}decide}). In \texttt{lean\_\cbrk{}kernel\_\cbrk{}nets/\cbrk{}} the same checks run in Lean\textquotesingle{}s kernel alone (\texttt{decide +kernel}) on an explicit copy of the networks and certificates, which is matched to the original data by a generator and a compiled comparison.† \subsection*{\pdfbookmark[2]{8.7 Capsule regions}{h25}8.7 Capsule regions} We now define the regions of the proof and verify (C1)–(C3) for them. Fix o ∈ Z⁴ and L ≥ 1, and put s\Sub{L}(y) := max(0, \textbar{}y₃ − o₃\textbar{} − L). For a cell or an edge e, x₃(e) is the third coordinate of its lower corner; the \emph{axial distance} of two sets is the distance of their projections to the x₃-axis; and the \emph{window of radius r} of a cell is the set of cells within l∞ distance r. \begin{itemize}[label=\discbullet] \item{} The \emph{capsule} 𝔇\Sub{L} is the union of the closed unit 4-cubes all of whose corners y satisfy \textbar{}y₁ − o₁\textbar{} + \textbar{}y₂ − o₂\textbar{} + s\Sub{L}(y) + \textbar{}y₄ − o₄\textbar{} ≤ 19 and max(\textbar{}y₁ − o₁\textbar{}, \textbar{}y₂ − o₂\textbar{}, s\Sub{L}(y), \textbar{}y₄ − o₄\textbar{}) ≤ 10. The set of points satisfying these inequalities is convex and invariant under the reflections fixing o, so with each point it contains the box spanned by that point\textquotesingle{}s reflections. Hence all lattice cubes between o and a cube of 𝔇\Sub{L} lie in 𝔇\Sub{L}, and 𝔇\Sub{L} is star-shaped about o. It consists of a prism of length 2L in the direction e₃ and two end caps, each a translate of half of the region obtained for L = 0. \item{} The \emph{hole} is 𝔅\Sub{L} := [o₁ − 3, o₁ + 4] × [o₂ − 3, o₂ + 2] × [o₃ − L, o₃ + L] × [o₄ − 3, o₄ + 3], and the \emph{port} is π\Sub{L} := c((o₁, o₂, o₃ − L + 3, o₄ − 3), \{1, 3\}), on the face x₄ = o₄ − 3 of 𝔅\Sub{L}. \item{} The corners of 𝔅\Sub{L} ± 1 satisfy the defining inequalities (the left sides are at most 14 and 5), so 𝔅\Sub{L} ± 1 ⊆ 𝔇\Sub{L} by convexity. The vertex o lies in the interior of 𝔅\Sub{L}, so the pivotal plaquette τ lies in 𝒫(𝔇\Sub{L}). \item{} \emph{Placement.} The map (u₁, u₂, u₃, u₄) ↦ (o₁ + u₂, o₂ + u₃, o₃ − L + 3 + u₁, o₄ + u₄) sends the port of §8.6 to π\Sub{L}. By (S5), 𝒩(L\Sub{c}, L\Sub{g}) then fits in 𝔅\Sub{L} at π\Sub{L} whenever 4L\Sub{c} + 4L\Sub{g} + 15 ≤ 2L − 3. \item{} \emph{Radii.} 𝔇\Sub{L} lies in the l∞ ball of radius L + 10 about o, so 𝒫(𝔇\Sub{L}) ⊆ B\Sub{L+11}(τ). A fitting network lies in 𝔅\Sub{L}, whose l∞ diameter is 2L, so 𝒩 ⊆ B\Sub{2L+1}(x). \end{itemize} \textbf{Orders and strips.} The orders and the strips are given by explicit rules, implemented in the supplement. For the list of (C1), available vertices y are ordered by their l∞ distance to 𝔅\Sub{L}, then by the number of coordinates attaining that distance, then by ⟨G, y − o⟩, where G := (0.618034, 0.414214, 0.302776, 0.236068). Each vertex then contributes its safe edges to earlier vertices, in an explicit order. NS is ordered lexicographically by the key of the endpoint with the larger key, then by the key of the midpoint, where key(y) := (‖ς(y − o)‖₁, ‖ς(y − o)‖₂², ⟨ν₀, y − o⟩), ν₀ := (1, 1/√2, 1/√3, 1/√5), and ς replaces the third coordinate z₃ by sign(z₃)·max(0, \textbar{}z₃\textbar{} − L). Since the entries of ν₀ are linearly independent over Q, ⟨ν₀, y − o⟩ separates the points of (½Z)⁴. Hence distinct NS edges have distinct keys, and the NS order is the exact lexicographic order of the keys. The floating-point sort of the computation produces exactly this order: for all 284,752 NS edges of 𝔇₂₄ it agrees with a 50-digit evaluation of the keys, in which no two keys are equal; locally the keys are also well separated, since the minimum of \textbar{}⟨ν₀, Δy⟩\textbar{} over nonzero Δy ∈ (½Z)⁴ of l∞ norm at most 17 is 1.86·10⁻⁶, far above double-precision rounding.† The vertex key may tie for distant vertices, but the checks of the local form of (C1) in the proof below compare keys only of vertices at l∞ distance at most 2, whose difference Δy satisfies \textbar{}⟨G, Δy⟩\textbar{} ≥ 0.012;† so the floating-point order and every total order refining the vertex key give the same checks. For the strips of (C3), let the \emph{zone} be the set of plaquettes with o₃ − L + 8 ≤ x₃ ≤ o₃ + L − 8, and let the \emph{lane plaquettes} be λ\Sub{m} := c((o₁, o₂, m, o₄ − 4), \{1, 3\}). The strip begins with the collar plaquette Q₀ of (C2) and continues by a depth-first search through 2-cells of 𝒦\Sub{\textless{}e}, until it reaches a safe edge σ such that all faces of the lattice 3-cubes containing σ are available plaquettes of 𝒫(𝔇\Sub{L}) with only safe edges. From there it follows the parent pointers of a breadth-first search tree rooted at π\Sub{L}. The tree is built on the available plaquettes of 𝒫(𝔇\Sub{L}) all of whose edges are safe, from which the plaquettes of the zone in planes containing the direction 3, other than the lane plaquettes, are excluded. The \emph{start part} of the strip is its initial segment before its first lane plaquette in the zone, the \emph{lane run} is the following run of lane plaquettes, and the \emph{end path} is the rest. \textbf{Proposition 31 (capsule certificate).} For all L ≥ 24, the triple (𝔇\Sub{L}, 𝔅\Sub{L}, π\Sub{L}) satisfies (C1)–(C3) of Proposition 28. \emph{Proof.} We prove by induction on L ≥ 24 that (C1) holds in the following local form, and that each edge e of (C2) has a certificate of the following shape (H\Sub{L}). \begin{itemize}[label=\discbullet] \item{} \emph{(C1), local form.} Each edge of the list is either an expansion (it lies in an available plaquette of 𝒫(𝔇\Sub{L}) whose other three edges come earlier) or has an endpoint that lies on no earlier edge of the list. Such an edge joins two components, so this implies (C1). \item{} \emph{(H\Sub{L}), route structure and window trichotomy.} The strip of e consists of its start part, followed, unless the whole strip lies in the start window, by a contiguous lane run λ\Sub{m₀}, λ\Sub{m₀−1}, …, λ\Sub{o₃−L+8} for some m₀ and by the end path from λ\Sub{o₃−L+8} to π\Sub{L}, which is the same for all such e and has a fixed position relative to the lower end of 𝔇\Sub{L}. The \emph{start window} W(e) is the axial range of the start part, or of the whole strip if the strip uses no lane plaquette of the zone. It contains the slice of e, because the collar plaquette does, and it is of one of three types: \begin{itemize}[label=\circbullet] \item{} \emph{zone type}: W(e) has axial extent ≤ 2; \item{} \emph{lower type}: W(e) ⊆ \{x₃ ≤ o₃ − L + 10\}; \item{} \emph{upper type}: W(e) ⊆ \{x₃ ≥ o₃ + L − 8\}. \end{itemize} \item{} \emph{(F) Far windows.} No NS edge lies within l∞ distance 3 of the lane plaquettes in the zone or of the end path. \end{itemize} \emph{Base case L = 24.}† (C1) holds with no failure. Apart from the first edge, each joining edge of the list has exactly one endpoint that lies on no earlier edge, so the local form holds. (C2) and (C3), with the link criterion of Lemma 27, hold at all 168,712 edges e as in (C2). The route structure holds, and all start windows are typed: 107,378, 45,544 and 39,876 of them satisfy the zone, lower and upper conditions (the types overlap). Finally (F) holds. \emph{Formal verification.} The base case is also checked in Lean 4 [dMU21] by the kernel alone.† The formalization defines 𝔇₂₄, 𝔅₂₄, π₂₄, the safe and NS edges, availability, the complexes 𝒦\Sub{\textless{}h}, the lane, the end path and (H₂₄) as above. Its NS order is the lexicographic order of an integer form of the key, with doubled coordinates and ν₀ replaced by round(2⁵⁰ν₀). On the NS edges of 𝔇₂₄ this is the order of the keys above.† Certificates produced by an untrusted program are checked by the Lean kernel: (C1) in the local form; (C2) and (C3) at all 168,712 edges, with an explicit integral 2-chain for every relation of Lemma 27 that is not a 3-cube relation; (H₂₄) for these certificates; and (F). From these checks Lean proves, for every commutative ring A: \begin{itemize}[label=\discbullet] \item{} H₁(𝒦\Sub{\textless{}h}; A) = 0 for every h; \item{} each edge e as in (C2) has a certificate for (C2) and (C3) of the shape (H₂₄), and H₁(𝒦″; A) = 0, where 𝒦″ is 𝒦\Sub{\textless{}e} with the spine edges removed; \item{} for every NS edge e* and every A-cycle γ on NS edges ≤ e* with γ(e*) ≠ 0, there is a 2-chain F on available plaquettes of 𝒫(𝔇₂₄) with (F1)–(F4) and Q\Sub{J} = π₂₄. \end{itemize} These proofs use only the axioms propext, Classical.choice and Quot.sound of Lean. No step uses \texttt{native\_\cbrk{}decide} or other code generated by Lean\textquotesingle{}s compiler; the trusted base is Lean\textquotesingle{}s kernel, including its built-in arithmetic on natural numbers. The induction step, Lemmas 24, 25 and 29, the placement of the networks and §8.3 are not formalized. \emph{Locality.} On \textbar{}x₃ − o₃\textbar{} ≤ L − 3 the region, the hole, the safe and NS classes and availability are invariant under translation by e₃. A comparison of keys between two cells there is unchanged when both are translated by the same multiple of e₃, because the keys are functions that do not depend on x₃ there, plus linear forms. The two end caps of 𝔇\Sub{L+1}, with their keys up to an additive constant in the linear forms, are the translates by ±e₃ of those of 𝔇\Sub{L}, because ς measures the third coordinate from the end. For a slice \uni{00740302}{\^{t}} with \textbar{}\uni{00740302}{\^{t}} − o₃\textbar{} ≤ L − 11, let ι\Sub{\uni{00740302}{\^{t}}} map the cells of 𝔇\Sub{L} below \uni{00740302}{\^{t}} by −e₃ and those above \uni{00740302}{\^{t}} by +e₃. This inserts a slab of width 2 at \uni{00740302}{\^{t}} and identifies 𝔇\Sub{L}, away from the slab, with 𝔇\Sub{L+1}, away from the inserted slab. Consequently ι\Sub{\uni{00740302}{\^{t}}} preserves the safe and NS classes, availability, and every comparison between two cells on the same side of the slab. In particular each window of radius 3 of 𝔇\Sub{L+1} is a translate of a window of 𝔇\Sub{L}, with the same comparisons among its cells. Comparisons between cells on opposite sides of the slab may change. \emph{(C1) for L + 1.} The local form at an edge is decided by the window of radius 2 of the edge, so it transfers from 𝔇\Sub{L} to 𝔇\Sub{L+1}. The same holds for (F). \emph{(C2), (C3) for L + 1: choice of \uni{00740302}{\^{t}}.} Let e₀ be an edge of 𝔇\Sub{L+1} as in (C2), at slice m. We need an edge e₀′ of 𝔇\Sub{L} as in (C2) and a slice \uni{00740302}{\^{t}} with \textbar{}\uni{00740302}{\^{t}} − o₃\textbar{} ≤ L − 11 such that either e₀ = e₀′ + e₃ and W(e₀′) lies above \uni{00740302}{\^{t}} + 3 (case A), or e₀ = e₀′ − e₃ and W(e₀′) lies below \uni{00740302}{\^{t}} − 3 (case B). We call such a slice \uni{00740302}{\^{t}} admissible. Assume L ≥ 18. \begin{itemize}[label=\discbullet] \item{} If m ≥ o₃ − L + 18, put \uni{00740302}{\^{t}} := o₃ − L + 11 and e₀′ := e₀ − e₃. The window of radius 3 of e₀ lies above a slab inserted at \uni{00740302}{\^{t}} and is the translate by +e₃ of that of e₀′, so e₀′ is an edge of 𝔇\Sub{L} as in (C2), by locality. W(e₀′) contains the slice m − 1 ≥ o₃ − L + 17 \textgreater{} o₃ − L + 10, so it is of zone type (then W(e₀′) ⊆ [m − 3, m + 1]) or of upper type. In both cases it lies above \uni{00740302}{\^{t}} + 3: case A. \item{} Otherwise m ≤ o₃ + L − 18, since L ≥ 18. Put \uni{00740302}{\^{t}} := o₃ + L − 11 and e₀′ := e₀ + e₃. Symmetrically, e₀′ is an edge of 𝔇\Sub{L} as in (C2), and W(e₀′) contains the slice m + 1 ≤ o₃ + L − 17 \textless{} o₃ + L − 8, so it is of zone type (W(e₀′) ⊆ [m − 1, m + 3]) or of lower type. In both cases it lies below \uni{00740302}{\^{t}} − 3: case B. \end{itemize} The other translate need not correspond: near the lower end e₀ − e₃, and near the upper end e₀ + e₃, is in general not an edge of 𝔇\Sub{L} as in (C2) (17,316 and 17,356 edges of 𝔇₂₅).† In case A the strip of e₀′ uses the lane, because its window does not contain the slice of π\Sub{L}. \emph{Transport.} In case A, translate the start part by +e₃, extend the lane run by two lane plaquettes, and translate the end path by −e₃. In case B, translate the whole certificate of e₀′ by −e₃. It lies below \uni{00740302}{\^{t}} − 3: its window does, and its lane run descends from the window to the lower end, followed by the end path. We check that the result is a certificate for e₀ in 𝔇\Sub{L+1}. Each condition of (C2) and (C3) is local. The conditions on Q₀ and the statement that Q\Sub{k} is a 2-cell of 𝒦\Sub{\textless{}e₀} are read in windows of radius 1. The link criterion at σ\Sub{k} is read in the box of radius 2 about σ\Sub{k}, in the complex from which the later spine edges have been removed. Near the start part, these windows and their comparisons with e₀ are images under ι\Sub{\uni{00740302}{\^{t}}}, together with e₀, of those of e₀′. By (F), the windows along the lane run and the end path contain no NS edge, so their content does not depend on e₀. Lane windows are translates of each other, and the windows at the two ends of the lane run are translates of those for e₀′. The start part, the lane run and the end path occupy disjoint stretches of the strip, so its plaquettes and spine edges remain pairwise distinct, and each σ\Sub{k} lies in no strip plaquette other than Q\Sub{k−1} and Q\Sub{k}. Finally the transported certificates have the route structure and the window types of (H\Sub{L+1}): lower windows move by −e₃ with the lower end, upper windows by +e₃ with the upper end, zone windows keep their extent, and the lane runs end at o₃ − (L + 1) + 8. Thus the induction hypothesis holds at L = 24, and by the induction step, which needs L ≥ 18, Proposition 31 holds for all L ≥ 24. ∎ The induction step transports certificates and makes no use of the search that produced them; we have also tested it directly. Every edge of 𝔇₂₅ and of 𝔇₂₆ as in (C2) is reached by the rule above. Transporting a complete set of certificates for 𝔇₂₄ to 𝔇₂₅ and then to 𝔇₂₆ (with case A whenever it holds for some admissible \uni{00740302}{\^{t}}) gives certificates that pass an independent checker, and transported certificates are checked in 𝔇₂₅, 𝔇₂₆ and 𝔇₃₀ by a second independent implementation.† For L = 16 the coverage fails even with any admissible \uni{00740302}{\^{t}}: 56 edges of 𝔇₁₇ are reached by no transport.† \subsection*{\pdfbookmark[2]{8.8 Proofs of Theorem W and its corollaries}{h26}8.8 Proofs of Theorem W and its corollaries} \emph{Proof of Theorem W.} Let ℓ\Sup{d} be a prime power dividing M, and take (a, n, Θ) as in (I1). Put L := max(24, 2(ℓ\Sup{n−1} + ℓ\Sup{a}) + 9), κ := L + 11 and ρ := 2L + 1. For t ∈ \{0, …, n − 1\} the network of Lemma 30 has L\Sub{c} ≤ ℓ\Sup{n−1}, so it fits in 𝔅\Sub{L} at π\Sub{L} (§8.7), lies in B\Sub{ρ}(x), and is a port network of type t. By Proposition 31, (C1)–(C3) hold for (𝔇\Sub{L}, 𝔅\Sub{L}, π\Sub{L}), and 𝒫(𝔇\Sub{L}) ⊆ B\Sub{κ}(τ). By Proposition 28 and Lemma 25, LCL\Sub{Θ}(ρ, κ) holds, and Proposition 23 gives P\Sub{p}(E\Sub{Θ}(P)) → 0 for p \textless{} 1/2 + δ\Sub{ℓ,d}/2, where δ\Sub{ℓ,d} is the constant of Proposition 23 for these radii. Put w\Sub{M} := ½ min δ\Sub{ℓ,d}, the minimum over all prime powers ℓ\Sup{d} dividing M. \begin{itemize}[label=\discbullet] \item{} (i) The precise form is the statement just proved. If Λ\Sub{P} ⊄ MZ⁶, then Λ\Sub{P} ⊄ ℓ\Sup{d}Z⁶ for some prime power ℓ\Sup{d} dividing M. The reduction mod ℓ\Sup{n} of an integral giant cycle whose class is not in ℓ\Sup{d}Z⁶ is a giant Z/ℓ\Sup{n}-cycle whose class is not in ℓ\Sup{d}(Z/ℓ\Sup{n})⁶, since d ≤ n. \item{} (ii) rank Λ\Sub{P} = dim V\Sub{Q}(P), because H₂(P; Q) = H₂(P; Z) ⊗ Q. For p \textgreater{} 1/2 this dimension is 6 with probability tending to 1, by [DKS20, Theorem 1] with F = Q. A sublattice of MZ⁶ of rank 6 has index divisible by M⁶. \item{} (iii) Instance (I2) with the networks of Lemma 30 for (ℓ, a, d) = (2, 1, 1) needs only L = 24, and the constant δ of Proposition 23 decreases as L grows. The argument above therefore gives (iii) for p \textless{} 1/2 + w\Sub{M} when M is even. \item{} (iv) P• is Bernoulli(1 − p) plaquette percolation on the cubical torus X•, and 1 − p \textless{} 1/2 + w\Sub{M}. By (i) for P•, V\Sub{Z/ℓ\Sup{n}}(P•) ⊆ ℓ\Sup{d}(Z/ℓ\Sup{n})⁶ with probability tending to 1. By Lemma 19 and unimodularity, V\Sub{Z/ℓ\Sup{n}}(P) = V\Sub{Z/ℓ\Sup{n}}(P•)\Sup{⊥} ⊇ (ℓ\Sup{d}(Z/ℓ\Sup{n})⁶)\Sup{⊥} = ℓ\Sup{n−d}(Z/ℓ\Sup{n})⁶. ∎ \end{itemize} \emph{Proof of Corollary W1.} Take M = ℓ, so n = 2 and d = 1. The inclusion V\Sub{Z/ℓ²}(P) ⊆ ℓ(Z/ℓ²)⁶ for p \textless{} 1/2 + w\Sub{ℓ} is Theorem W(i). The reverse inclusion for p \textgreater{} 1/2 − w\Sub{ℓ} is Theorem W(iv). For p \textgreater{} 1/2, since all nonzero classes mod ℓ are partnered, V\Sub{F\Sub{ℓ}}(P•) = 0 with probability tending to 1 by Proposition 20, and V\Sub{F\Sub{ℓ}}(P) = V\Sub{F\Sub{ℓ}}(P•)\Sup{⊥} = F\Sub{ℓ}⁶ by Lemma 19. ∎ \emph{Proof of Corollary W2.} P• is Bernoulli(1 − p) plaquette percolation with 1 − p ∈ (1/2, 1/2 + w\Sub{M}). By Theorem W(ii) for P•, with probability tending to 1, Λ\Sub{P•} ⊆ MZ⁶ and rank Λ\Sub{P•} = 6. The dual complex of P• is P, up to the translation. Proposition 21, applied to P• in place of P, gives the claim for T = Tors H₁(P; Z). ∎ \textbf{Remarks.} \begin{enumerate}[label=\arabic*.] \item{} \emph{Constants.} For M = 2 the proof uses L = 25, κ = 36, ρ = 51 and κ′ = 139, and \textbar{}B\Sub{139}\textbar{} = 36,095,493,313, so log₂(1/w₂) ≈ 1.44·10¹¹;† in this sense Theorem W is qualitative. Theorem W(iii) and the prime power 3 use L = 24, so for them the region certificate is exactly the one verified in Lean. \item{} \emph{A second certificate for M = 2.} Parts (i) and (iii) for M = 2 also follow from a different region and different networks. The region is the union of the unit 4-cubes all of whose corners satisfy \textbar{}y − o\textbar{}₁ ≤ 25 and \textbar{}y − o\textbar{}\Sub{∞} ≤ 13. The networks are built from a tube, lattice Möbius bands and converters of degree 2 and 4. Since only Z/8 and Z/4 coefficients occur, the F₂ form of the link criterion suffices there: H₁(𝒦; F₂) = 0 implies that H₁(𝒦; Z) is finite of odd order, so H₁(𝒦; Z/2\Sup{k}) = 0. (C1)–(C3) hold at all 77,712 relevant edges for each of the five networks (Z/8-types 0, 1, 2 and Z/4-types 0, 1).† (C1)–(C3) for this region are also checked in Lean 4 by the kernel alone, for the hole of each of the five networks and with the integral link criterion, on certificates produced by an untrusted program. From these checks Lean proves Lemma 26 and the 2-chain of Proposition 28 for this region, over every commutative ring and using only the standard axioms.† As for 𝔇₂₄, the formalization orders the NS edges by an integer form of the key; on the NS edges of this region this order agrees with the floating-point order of the certificate above, and that agreement is checked outside Lean.† This certificate gives κ = 14 and ρ = 12, hence log₂(1/w₂) ≈ 9.1·10⁸.† \item{} \emph{End-to-end tests.} The construction is also run on explicit configurations with Z/9 coefficients, on T⁴₅₂ and T⁴₆₄, with networks of types 0 and 1 placed in a capsule. An exact computation confirms that ω′ carries no Θ-odd cycle and that ω′ ∪ x does;† these tests are not used in the proofs. \item{} \emph{What Theorem W excludes.} A natural route to Theorem A at q = 1 is to show that every mod-2 class lifts to a giant Z/4-cycle above 1/2, that is, L₄(P) = F₂⁶. Indeed, by Lemma 19 the set \{b ∈ F₂⁶ : 2\uni{00620303}{\~{b}} ∈ V\Sub{Z/4}(P•)\} is L₄(P)\Sup{⊥}, and it contains V(P•). Theorem W(iii) shows that this route fails near 1/2, since there L₄(P) ⊆ C\Sup{⊥}. \end{enumerate} \section*{\pdfbookmark[1]{9. Dualities and locks at q = 1}{h27}9. Dualities and locks at q = 1} In this section d = 4 and plaquettes are 2-cells. By Proposition 8, Theorem A at q = 1 fails exactly when, for some p₀ \textless{} 1/2, P\Sub{p₀}(V ∩ C\Sup{⊥} ≠ 0) does not tend to 0. Fix a subsequence of N and a class z₀ ∈ C\Sup{⊥} ∖ 0 such that P\Sub{p₀}(z₀ ∈ V) stays bounded away from 0 along it. For p ∈ (0, 1) let F\Sub{p} be the set of classes z with P\Sub{p}(z ∈ V) → 1 along the subsequence. By sharp thresholds, z₀ ∈ F\Sub{p} for every p ∈ (p₀, 1/2). F\Sub{p} is a subspace invariant under coordinate permutations, and for p \textless{} 1/2 it is isotropic: F\Sub{p} ⊆ F\Sub{1−p} by monotonicity, while F\Sub{1−p} ⊆ F\Sub{p}\Sup{⊥} since V(P) = V(P•)\Sup{⊥} and P• has law P\Sub{p} when P has law P\Sub{1−p}. By Lemma 6′, F\Sub{p} contains u or C. We call this a \emph{lock} onto F\Sub{p}. This section records three facts about locks (§§9.1–9.3). For z ∈ H write Pf(z) := z\Sub{12} z\Sub{34} + z\Sub{13} z\Sub{24} + z\Sub{14} z\Sub{23} ∈ F₂. Since e\Sub{I} · e\Sub{J} = 1\{I ∩ J = ∅\}, Pf is a quadratic form whose polar form is the intersection form: Pf(z + z′) = Pf(z) + Pf(z′) + ⟨z, z′⟩. A subspace on which Pf vanishes identically is \emph{totally singular}; it is isotropic. Put W₀ := \{w ∈ F₂⁴ : Σ wᵢ = 0\} and 1 := (1, 1, 1, 1), and let Cyc := Λ²(W₀), Cut := 1 ∧ F₂⁴, L₀ := span\{e\Sub{12} + e\Sub{34}, e\Sub{13} + e\Sub{24}, e\Sub{14} + e\Sub{23}\}, E := u\Sup{⊥}. Cyc is spanned by triangles and 4-cycles, Cut by the four stars, and L₀ is the oblique Lagrangian of §7.5; L₀ = C + ⟨u⟩. The permutation-invariant subspaces of H are 0, ⟨u⟩, C, L₀, Cyc, Cut, C\Sup{⊥}, E and H, and Pf vanishes identically exactly on 0, C, Cyc and Cut.† In particular Cyc and Cut are totally singular Lagrangians, and Pf(u) = 1. The nonzero invariant isotropic subspaces are ⟨u⟩, C, L₀, Cyc and Cut, so a lock is onto one of these. \subsection*{\pdfbookmark[2]{9.1 The Pfaffian counts dual pairs}{h28}9.1 The Pfaffian counts dual pairs} For a plaquette σ let σ* be its dual plaquette (same centre, complementary directions), and h := (½, ½, ½, ½). A \emph{duality} is a map D(σ) = g(σ*), where g is an isometry of R⁴ with diagonal linear part (so that g descends to T⁴\Sub{N} and acts trivially on H₂(X; F₂)) that maps the dual lattice h + Z⁴ onto Z⁴. Two families are relevant: \begin{itemize}[label=\discbullet] \item{} translation dualities, g(y) = y + h − t with t ∈ Z⁴; D² is the translation by 1 − 2t, so D is an involution iff 1 − 2t ∈ NZ⁴, which forces N odd; \item{} point-reflection dualities, g(y) = m + h − y with m ∈ Z⁴, so D(c(x, I)) = c(m − x, I\Sup{c}); these are involutions on every T⁴\Sub{N}. \end{itemize} D maps a plaquette of type I to one of type I\Sup{c}, so it has no fixed plaquette, and an involutive D splits the plaquettes into \emph{D-pairs} \{σ, Dσ\}. Since g maps the dual complex onto the primal one and acts trivially on H₂(X; F₂), Fact D(a) and the orthogonality V(P•) = V(P)\Sup{⊥} noted after Proposition 6 give V(D(ω\Sup{c})) = V(ω)\Sup{⊥} for every configuration ω. \textbf{Theorem 32 (dual pairs).} Let D be an involutive duality. For every mod-2 2-cycle S, Pf([S]) ≡ \#\{D-pairs \{σ, Dσ\} ⊆ S\} (mod 2). \emph{Proof.} For x ∈ F₂\Sup{X\Sup{(2)}} put Q(x) := Σ\Sub{pairs} x\Sub{σ} x\Sub{Dσ}. Its polar form is B(x, y) := Q(x + y) − Q(x) − Q(y) = Σ\Sub{σ} x\Sub{σ} y\Sub{Dσ} = \textbar{}x ∩ D(y)\textbar{} mod 2. \begin{itemize}[label=\discbullet] \item{} (a) \emph{On cycles, B is the intersection form.} A primal plaquette fixes its two non-type coordinates to integers and a dual plaquette fixes its two non-type coordinates to half-integers, so they meet only if their types are complementary, and then only if they have the same centre, i.e. are dual to each other; they then cross transversally in one point. A primal edge and a dual plaquette (or a dual edge and a primal plaquette) never meet: together they fix five coordinates, so some coordinate is fixed to an integer by one and to a half-integer by the other. Since g commutes with σ ↦ σ*, D² = g² on plaquettes, so g² is the identity on X and g also maps primal cells to dual cells. Hence σ meets g(σ′) iff g(σ′) = σ*, i.e. σ′ = g(σ*) = D(σ). So for cycles S, S′ the set \textbar{}S\textbar{} ∩ g\textbar{}S′\textbar{} consists of transverse crossings, one for each σ ∈ S ∩ D(S′), and B(S, S′) = ⟨[S], [gS′]⟩ = ⟨[S], [S′]⟩. \item{} (b) \emph{Q vanishes on boundaries.} The faces of a 3-cube have types inside the cube\textquotesingle{}s three directions, and D maps each type to its complement, which is not among them; so Q(∂c) = 0. For a sum of cube boundaries, Q(Σ ∂cᵢ) = Σ Q(∂cᵢ) + Σ\Sub{i\textless{}j} B(∂cᵢ, ∂cⱼ) = 0 by (a). \item{} (c) Hence Q(S + ∂b) = Q(S) + Q(∂b) + B(S, ∂b) = Q(S), so Q induces \uni{00710302}{\^{q}}: H → F₂ with \uni{00710302}{\^{q}}(z + z′) = \uni{00710302}{\^{q}}(z) + \uni{00710302}{\^{q}}(z′) + ⟨z, z′⟩. \item{} (d) \uni{00710302}{\^{q}} − Pf is therefore linear. A coordinate torus of type I contains no D-pair, since D changes the type, so \uni{00710302}{\^{q}}(e\Sub{I}) = 0 = Pf(e\Sub{I}). Hence \uni{00710302}{\^{q}} = Pf. ∎ \end{itemize} \textbf{Corollary 33.} Let D be an involutive duality. For a set Y of plaquettes put V(Y) := V(X\Sup{(1)} ∪ Y). \begin{itemize}[label=\discbullet] \item{} (a) If Y ∩ D(Y) = ∅, then V(Y) is totally singular. \item{} (b) (\emph{Frozen structures.}) Let ω be random with ω ⊇ Y and ω ∩ D(Y) = ∅ almost surely, where Y ∩ D(Y) = ∅. Then V(Y) ⊆ V(ω) ⊆ V(Y)\Sup{⊥}. If Y is invariant under coordinate permutations, V(Y) is one of 0, C, Cyc, Cut; in particular a frozen permutation-invariant structure never forces u ∈ V. \item{} (c) (\emph{Self-dual product families.}) Let (ν\Sub{s}) be a monotone family of product measures, P\Sub{s}(σ open) = π\Sub{σ}(s), with π\Sub{Dσ}(1 − s) = 1 − π\Sub{σ}(s) (so that D(ω\Sub{s}\Sup{c}) has law ν\Sub{1−s}). For s ≤ 1/2 the majority set \{σ : π\Sub{σ}(s) \textgreater{} 1/2\} is D-free, and every cycle of ω\Sub{s} with Pf = 1 contains a doubly open pair \{σ, Dσ\}. Since π\Sub{σ}(s) + π\Sub{Dσ}(s) ≤ 1, it contains an open plaquette σ with π\Sub{σ}(s) ≤ 1/2. \end{itemize} \emph{Proof.} (a) is Theorem 32. (b) V(Y) ⊆ V(ω) since Y ⊆ ω, and V(ω)\Sup{⊥} = V(D(ω\Sup{c})) ⊇ V(D(D(Y))) = V(Y) since D(Y) ⊆ ω\Sup{c}. The list of permutation-invariant totally singular subspaces gives the rest, as u ∉ Cyc ∪ Cut. (c) For s ≤ 1/2, π\Sub{σ}(s) + π\Sub{Dσ}(s) ≤ π\Sub{σ}(1 − s) + π\Sub{Dσ}(s) = 1 by monotonicity and the hypothesis for Dσ, so the majority set is D-free; the rest is Theorem 32. ∎ So under an involutive duality a lock onto a subspace on which Pf does not vanish, such as ⟨u⟩ or L₀, cannot be produced by plaquettes that are open with probability one: every cycle whose class has Pf = 1 contains a doubly open pair. For even N the translation duality is not an involution, and the analogous count of ordered pairs is even on every cycle, so it carries no information. Point-reflection dualities are involutions on every T⁴\Sub{N}, so Theorem 32 still applies on even tori. \subsection*{\pdfbookmark[2]{9.2 A symmetric, self-dual product family that locks}{h29}9.2 A symmetric, self-dual product family that locks} The following family shows which properties of Bernoulli percolation a proof of Theorem A at q = 1 must use. Let 8 \textbar{} N. Call Σⱼ xⱼ ∈ Z/N the \emph{exact level} of a cell c = c(x, I), and let λ(c) be its residue mod 8. Split the plaquettes into the classes 𝒜 := \{λ ∈ \{0, 1\}\}, ℬ := \{λ ∈ \{4, 5\}\} and ℱ := \{λ ∈ \{2, 3, 6, 7\}\}. Fix ε ∈ (0, 1/5), δ ∈ (0, 1/2) and a continuous, piecewise C¹ map φ: [0, 1] → [0, 1] with inf φ′ \textgreater{} 0, φ(0) = 0, φ(1) = 1 and 1 − ε ≤ φ ≤ 1 − ε/2 on [δ, 1 − δ]. Let ν\Sub{s} be the product measure with P(σ open) = φ(s) on 𝒜, 1 − φ(1 − s) on ℬ, s on ℱ, and let ω\Sub{s} := \{σ : U\Sub{σ} ≤ P\Sub{s}(σ open)\} for independent uniform labels U\Sub{σ}. \textbf{Theorem 34 (a robust lock onto Cyc).} For every N ≡ 0 mod 8, P(V(ω\Sub{s}) = Cyc for all s ∈ [δ, 1 − δ]) ≥ 1 − 48 N³ (5ε)\Sup{2N} / (5(1 − 5ε)). The family (ν\Sub{s}) is a monotone family of product measures, invariant under coordinate permutations and under translations by v with Σ vᵢ ≡ 0 mod 8, it has finite energy, and it is exactly self-dual: for τ(c(x, I)) := c(x + 1\_I, I\Sup{c}) and t := (6, 0, 0, 0), the configuration ω\textasciicircum{}♯ := \{τσ − t : σ ∉ ω\Sub{s}\} has law ν\Sub{1−s}, and V(ω\textasciicircum{}♯) = V(ω\Sub{s})\Sup{⊥}. It is also self-dual under the involutive point-reflection duality D(c(x, I)) = c(m − x, I\Sup{c}) with Σ mᵢ ≡ 5 mod 8. \emph{Proof.} \begin{itemize}[label=\discbullet] \item{} \emph{Self-duality.} λ(τσ − t) = λ(σ) + \textbar{}I\textbar{} − 6 = λ(σ) − 4, so σ ↦ τσ − t exchanges 𝒜 and ℬ and preserves ℱ, and the densities match: P(τσ − t ∈ ω\textasciicircum{}♯) = 1 − φ(s) = 1 − φ(1 − (1 − s)) for σ ∈ 𝒜, 1 − (1 − φ(1 − s)) = φ(1 − s) for σ ∈ ℬ, and 1 − s for σ ∈ ℱ. The set τ(ω\Sup{c}) is the dual complex translated by h, so V(ω\textasciicircum{}♯) = V(ω)\Sup{⊥} by Fact D. The point reflection maps λ to 5 − λ, so it too exchanges 𝒜 and ℬ and preserves ℱ, and the same densities match. \item{} \emph{Staircases.} For an integer ℓ let S\Sub{ℓ} be the union of the 3-cells y + [0, 1]\Sup{[4]∖\{k\}}, k ∈ [4], with Σⱼ yⱼ ≡ ℓ mod N, and their faces. A plaquette c(z, J) with J\Sup{c} = \{a, b\} lies in the four 3-cells z + [0,1]\Sup{J∪\{a\}}, z + [0,1]\Sup{J∪\{b\}}, (z − e\Sub{a}) + [0,1]\Sup{J∪\{a\}}, (z − e\Sub{b}) + [0,1]\Sup{J∪\{b\}}, of base levels Σz, Σz, Σz − 1, Σz − 1. So the plaquettes of S\Sub{ℓ} are those of exact level ℓ or ℓ + 1, each lies in exactly two 3-cells of S\Sub{ℓ}, and for ℓ ≡ 0 mod 8 they all lie in 𝒜. \item{} \emph{S\Sub{ℓ} is a 3-torus carrying Cyc.} In R⁴ let R\Sub{ℓ} := ∪\{x + [0,1]⁴ : Σx ≤ ℓ − 1\}. A point w lies in R\Sub{ℓ} iff Φ(w) := Σᵢ ⌈wᵢ⌉ ≤ ℓ + 3, and Φ is nondecreasing along 1. So ∂R\Sub{ℓ} is the graph of a continuous function over \{Σ = 0\}, and it is the lift of S\Sub{ℓ}. Its stabilizer in NZ⁴ is Γ := NZ⁴ ∩ \{Σ = 0\} ≅ Z³, so S\Sub{ℓ} ≅ R³/Γ. Γ is saturated in NZ⁴, so H₁(S\Sub{ℓ}; F₂) → H₁(X; F₂) is injective with image W₀; since the inclusion is homotopic to a linear map of tori, the map on H₂ is Λ² of the map on H₁, hence injective with image Λ²(W₀) = Cyc. \item{} \emph{Robustness.} Fix ℓ ≡ 0 mod 8. For s ∈ [δ, 1 − δ] the plaquettes of S\Sub{ℓ} are open independently with probability ≥ 1 − ε. Let G\Sub{ℓ} be the 6-regular graph whose vertices are the 4N³ 3-cells of S\Sub{ℓ} and whose edges are its plaquettes. By Poincaré–Lefschetz duality on the closed 3-manifold S\Sub{ℓ}, if the closed plaquettes contain no cycle of G\Sub{ℓ} that is essential in H₁(S\Sub{ℓ}; F₂), the open plaquettes carry H₂(S\Sub{ℓ}; F₂), hence Cyc ⊆ V. An essential cycle lifts to a path between translates by some v ∈ Γ ∖ \{0\}, and \textbar{}v\textbar{}\Sub{∞} ≥ N; adjacent 3-cells have centres differing by ½(e\Sub{a} − e\Sub{b}), so the cycle has length ≥ 2N. Counting non-backtracking closed walks (at most 6·5\Sup{k−1} of length k from a vertex) gives P(Cyc ⊄ V(ω\Sub{δ})) ≤ 24 N³ (5ε)\Sup{2N} / (5(1 − 5ε)). \item{} \emph{The lock.} The same bound applied to ω\textasciicircum{}♯\Sub{1−δ}, which has law ν\Sub{δ}, gives Cyc ⊆ V(ω\textasciicircum{}♯\Sub{1−δ}) = V(ω\Sub{1−δ})\Sup{⊥}, that is V(ω\Sub{1−δ}) ⊆ Cyc\Sup{⊥} = Cyc, off an event of the same probability. By monotonicity, Cyc ⊆ V(ω\Sub{δ}) ⊆ V(ω\Sub{s}) ⊆ V(ω\Sub{1−δ}) ⊆ Cyc for s ∈ [δ, 1 − δ]. ∎ \end{itemize} In particular ν\Sub{1/2}(V = Cyc) → 1. Theorem A fails for this family, since V = Cyc ≠ 0 with high probability for every s ∈ [δ, 1/2), and the hypothesis of Proposition 7 fails too, since e\Sub{12} ∉ Cyc. So does Lemma 4: for n = e\Sub{12}\Sup{*} + e\Sub{14}\Sup{*} + e\Sub{23}\Sup{*} + e\Sub{34}\Sup{*}, PD(n) is a 4-cycle, so it lies in Cyc, while PD(s\Sub{1}(n)) = e\Sub{12} + e\Sub{14} ∉ Cyc; since D\Sub{n} means PD(n) ∈ V(P•) = Cyc on the lock, at s = 1/2 we get R(n) → 1 and R(s\Sub{1}(n)) → 0. The point-reflection duality maps 𝒜 onto ℬ, so 𝒜, which is the majority set \{π\Sub{σ}(s) \textgreater{} 1/2\} for s ∈ [δ, 1/2], is D-free. By Corollary 33(a) and the list above, V(𝒜) is a permutation-invariant totally singular subspace. It contains the image Cyc of the staircases, so V(𝒜) = Cyc: the lock is onto the subspace carried by the D-free majority set. \textbf{Corollary 35.} The conclusion of Theorem A at q = 1 does not follow from the following properties of a family of measures on configurations: product structure (hence the Harris–FKG, van den Berg–Kesten and Reimer inequalities [vdBK85, Rei00], the OSSS inequality [OSSS05] and Russo\textquotesingle{}s formula), finite energy, exchangeability of the labels within each orbit of the symmetry group, invariance under coordinate permutations, exact self-duality V(ω\textasciicircum{}♯) = V(ω)\Sup{⊥}, also under an involutive duality, and sharp thresholds with polynomial rates for invariant increasing events [BKKKL92]. A proof must therefore use a property that this family lacks, such as homogeneity (invariance under unit translations) or the sign changes of the hyperoctahedral group B₄ other than x ↦ −x. The family is invariant under x ↦ m − x for Σ mᵢ ≡ 3 mod 8, which acts trivially on H₂. \emph{Proof.} The family of Theorem 34 has all the listed properties. For the sharp thresholds, let 𝓔 be an invariant increasing event with ν\Sub{s}(𝓔) = t, and let J(σ) := ν\Sub{s}(𝓔 \textbar{} σ open) − ν\Sub{s}(𝓔 \textbar{} σ closed). J is constant on each symmetry orbit, and each orbit has at least 6N⁴/8 plaquettes. So the Bourgain–Kahn–Kalai–Katznelson–Linial inequality on the uniform labels gives Σ\Sub{σ} J(σ) ≥ c t(1 − t) log N. Every P\Sub{s}(σ open) has derivative ≥ min(1, inf φ′) \textgreater{} 0, so Russo\textquotesingle{}s formula gives dν\Sub{s}(𝓔)/ds ≥ c′ t(1 − t) log N. This yields windows of width O(1/log N) and polynomial rates. ∎ The staircases are orientable, and the proof of Theorem 34 applies verbatim with coefficients in Q, Z/4 and F\Sub{ℓ′} (ℓ′ an odd prime). The robustness bound excludes every closed dual cycle whose lift is not closed, and Poincaré–Lefschetz duality on the orientable S\Sub{ℓ} holds with any coefficients. The upper bound uses V\Sub{A}(P•) ⊥ V\Sub{A}(P) for the coefficient ring A (Lemma 19, or [DKS20, Lemma 9] over Q), and Λ²(ker Σ) is a Lagrangian direct summand of H₂(T⁴; Z). So the family locks onto Λ²(ker Σ) with these coefficients as well. This does not contradict [DKS20] or Proposition 10. [DKS20] uses the irreducibility of H₂(T⁴; F) under the point group, and Proposition 10 uses partners in B₄. Both need sign changes other than x ↦ −x, and this family has only the central one, which acts trivially on H₂. In a B₄-invariant self-dual family with sharp thresholds, the argument of Proposition 10 excludes, below the self-dual point, every lock in which V\Sub{Z/4} contains an element of order 4, in particular every lock carried by orientable cycles. So a B₄-invariant counterexample would have to lock onto non-orientable cycles. Whether one exists is open. \subsection*{\pdfbookmark[2]{9.3 An anisotropic window}{h30}9.3 An anisotropic window} Call a plaquette \emph{vertical} if its type contains 1 and \emph{horizontal} otherwise. Consider the product measure in which vertical plaquettes are open with probability 1 − a and horizontal ones with probability b, 0 \textless{} a, b ≤ 1. Every duality maps type I to I\Sup{c} and so exchanges vertical and horizontal plaquettes. Hence D(ω\Sup{c}) has the law of the model with a and b exchanged, and V(D(ω\Sup{c})) = V(ω)\Sup{⊥}. On the self-dual line a = b, P(V = 0) = P(V = H) ≤ 1/2, as in step 1 of the proof of Theorem A. The isotropic model is a = 1 − p, b = p. \textbf{Lemma 36 (slices).} Let S be a mod-2 2-cycle, i ∈ [4] and r ∈ Z/N. Send each plaquette c(x, \{i, j\}) of S with xᵢ = r to the edge of T³\Sub{N} (coordinate i deleted) from \uni{00780302}{\^{x}} in direction j. The result σ\Sub{i,r}(S) is a mod-2 1-cycle, and for j ≠ i its winding parity in direction j is the coefficient z\Sub{ij} of [S]. \emph{Proof.} The edge of X from y to y + eᵢ (yᵢ = r) lies in the six plaquettes c(y, \{i, j\}) and c(y − e\Sub{j}, \{i, j\}), j ≠ i, which are sent to the six edges of T³\Sub{N} at ŷ; since ∂S = 0, every vertex of σ\Sub{i,r}(S) has even degree. Its edges in direction j crossing \{x\Sub{j} = 1/2\} are the plaquettes of S of type \{i, j\} with xᵢ = r and x\Sub{j} = 0; their number is the value on S of a cocycle Poincaré dual to the coordinate torus of type \{i, j\}\Sup{c}, so its parity is z\Sub{ij}. ∎ \textbf{Theorem 37.} In the anisotropic model put κ := 4b(2 − a)/a. If κ \textless{} 1, then P(V ≠ 0) → 0 exponentially fast in N. More precisely, for θ ∈ (0, 1) with 4bA\Sub{θ} \textless{} 1, where A\Sub{θ} := (1 + (1 − a)\Sup{θ})/(1 − (1 − a)\Sup{θ}), P(V ≠ 0) ≤ 3N³ Σ\Sub{n ≥ N} κⁿ + 3N²(1 − a)\Sup{N} + 3N³ (1 − a)\Sup{(1−θ)N} · 4bA\Sub{θ}/(1 − 4bA\Sub{θ}), and such a θ exists exactly when κ \textless{} 1, because A\Sub{θ} decreases to (2 − a)/a as θ ↑ 1 (strictly when a \textless{} 1). \emph{Proof.} Let S ⊆ ω be a cycle. \begin{itemize}[label=\discbullet] \item{} \emph{Horizontal coordinates.} Suppose z\Sub{34}([S]) = 1. By Lemma 36 (i = 3, r = 0), σ\Sub{3,0}(S) is a 1-cycle of the torus with coordinates x₁, x₂, x₄ with odd winding in direction 4, so it contains a simple cycle γ with odd winding in direction 4, hence with at least N edges in direction 4. Its edges are distinct open plaquettes: direction 4 ↔ type \{3, 4\} and direction 2 ↔ type \{2, 3\} (probability b), direction 1 ↔ type \{1, 3\} (probability 1 − a). Write γ as a closed walk (s₁, R₁, …, s\Sub{n}, R\Sub{n}) starting at a non-time step, where each s\Sub{m} is one of the 4 steps ±e₂, ±e₄ and each R\Sub{m} is a maximal run of time steps; a simple cycle does not reverse, so R\Sub{m} is monotone, with a length t\Sub{m} ≥ 0 and one of 2 directions if t\Sub{m} ≥ 1; and n ≥ N. Summing Π\Sub{m} b(1 − a)\Sup{t\Sub{m}} over the walks from a fixed vertex gives at most Σ\Sub{n ≥ N} (4b(1 + 2Σ\Sub{t≥1}(1 − a)\Sup{t}))ⁿ = Σ\Sub{n≥N} κⁿ, and there are N³ starting vertices. The coefficients z\Sub{23}, z\Sub{24} are handled with the slice i = 2, r = 0 in the same way. This gives the first term. \item{} \emph{Vertical coordinates.} Suppose z\Sub{1j}([S]) = 1 for some j ∈ \{2, 3, 4\}. By Lemma 36 with i = j and r = 0, σ\Sub{j,0}(S) is a 1-cycle of the torus with coordinates x₁, x\Sub{k}, x\Sub{l}, where \{k, l\} = [4] ∖ \{1, j\}, with odd winding in direction 1. It contains a simple cycle γ with odd winding in direction 1. Every hyperplane \{x₁ = r′ + 1/2\}, r′ ∈ Z/N, is crossed by an odd number of its time edges, so γ has at least N of them, and each is a vertical plaquette (type \{1, j\}), open with probability 1 − a. Its other edges are horizontal plaquettes (types \{j, k\}, \{j, l\}), open with probability b. If γ has no spatial edge it is a straight time circle: N² choices, probability (1 − a)\Sup{N}. Otherwise write it as (s₁, R₁, …, s\Sub{n}, R\Sub{n}) with n ≥ 1 and monotone time runs of lengths t\Sub{m}, Σ t\Sub{m} ≥ N. Using (1 − a)\Sup{Σt} ≤ (1 − a)\Sup{(1−θ)N}(1 − a)\Sup{θΣt} and Σ\Sub{t≥0}(1 − a)\Sup{θt} 2\Sup{[t≥1]} = A\Sub{θ}, the expected number of open such γ is at most N³(1 − a)\Sup{(1−θ)N} Σ\Sub{n≥1}(4bA\Sub{θ})ⁿ. Sum over j. ∎ \end{itemize} \textbf{Corollary 38 (anisotropic window).} In the anisotropic model, V = 0 w.h.p. if 4b(2 − a) \textless{} a, and V = H w.h.p. if 4a(2 − b) \textless{} b. So for each a every threshold of V, in particular a sharp transition curve if there is one, lies in the closed wedge a/(4(2 − a)) ≤ b ≤ 8a/(1 + 4a). The self-dual line b = a lies in its interior for every a ∈ (0, 1]; for small a and b the wedge is a/8 ≲ b ≲ 8a. \emph{Proof.} The first statement is Theorem 37. For the second, apply Theorem 37 to D(ω\Sup{c}), whose parameters are (b, a), and use V(ω) = V(D(ω\Sup{c}))\Sup{⊥}. ∎ For the isotropic model Theorem 37 gives only p \textless{} (√41 − 5)/8 ≈ 0.175, weaker than Proposition 6. Corollary 38 is useful in the anisotropic regime, where monotone comparison as in Proposition 6 gives nothing: it needs max(1 − a, b) \textless{} √2 − 1. The bounds are asymptotic. At a = 0.2, for instance, the bound of Theorem 37 drops below 1 only from N = 93, 134 and 384 for b = 0.005, 0.01 and 0.02 (and from N = 38 as b → 0). The method uses one slice, so it cannot reach the self-dual line: Theorem 37 itself stops at b = a/(4(2 − a)) ≈ 0.028 for a = 0.2, and at N = 20 the slices already contain cycles winding in a spatial direction at b ≈ 0.07 (at N = 10 from b ≈ 0.09), against the line b = 0.2.† Between that point and the line, the statement is again an F₂ sharpness problem for surfaces in four dimensions, now without a symmetry exchanging vertical and horizontal classes; its two-dimensional analogue, the critical line p\Sub{h} + p\Sub{v} = 1 of inhomogeneous bond percolation, rests on planar duality and crossing estimates [Kes82, §3.4], [Gri99, §11.9]. \section*{\pdfbookmark[1]{Appendix: verification table}{h31}Appendix: verification table} Every statement marked † rests on the computation listed below. The supplementary material contains each program and each log, together with a README that gives the exact command for every row, and SHA-256 hashes of all files. The C and C++ programs are compiled with gcc or g++ and OpenMP (\texttt{capgen.cpp} also needs GMP); the Python programs need numpy, scipy, sympy, mpmath and, for one row, cvxpy; the Lean projects need Lean 4 (v4.35.0-rc3) and Mathlib (commit c55e6e78). Exact computations are deterministic. The multithreaded Monte Carlo rows use one random stream per thread, so a rerun reproduces their statistics but not their digits. Rows marked \emph{cross-check} re-verify a statement that is proved by hand, or verify a computation independently; no proof depends on them. {\footnotesize \begin{longtable}{|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-6\tabcolsep-4\arrayrulewidth)*4159/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-6\tabcolsep-4\arrayrulewidth)*2509/10000\relax}|>{\raggedright\arraybackslash}p{\dimexpr(\linewidth-6\tabcolsep-4\arrayrulewidth)*3332/10000\relax}|} \hline \textbf{Statement} & \textbf{Programs} & \textbf{Logs} \\ \hline \endfirsthead \hline \textbf{Statement} & \textbf{Programs} & \textbf{Logs} \\ \hline \endhead §7, after Prop. 7: μ(e\Sub{I} ∈ V) = 0.424, 0.436, 0.452 & {\scriptsize\texttt{bern4d\_\cbrk{}h.\cbrk{}c}}, {\scriptsize\texttt{vhist\_\cbrk{}summary.\cbrk{}py}} & {\scriptsize\texttt{vhist\_\cbrk{}N*\_\cbrk{}p0.\cbrk{}500.\cbrk{}txt}}, {\scriptsize\texttt{vhist\_\cbrk{}summary.\cbrk{}txt}} \\ \rowrule Prop. 8: the partnerless classes are the 15 nonzero classes of C\Sup{⊥} & {\scriptsize\texttt{z4partners.\cbrk{}py}} & {\scriptsize\texttt{z4partners.\cbrk{}txt}} \\ \rowrule After Prop. 9: S₅ fixes only u\Sub{Q}, the only partnerless class of A₄* & {\scriptsize\texttt{perm\_\cbrk{}fixed.\cbrk{}py}} & {\scriptsize\texttt{perm\_\cbrk{}fixed.\cbrk{}txt}} \\ \rowrule Prop. 10(i): all 4032 elements of order 4 are partnered & {\scriptsize\texttt{z4partners.\cbrk{}py}} & {\scriptsize\texttt{z4partners.\cbrk{}txt}} \\ \rowrule After Prop. 11: spectral radii ≥ √2; every class from one or two hardness-free planes & {\scriptsize\texttt{balanced.\cbrk{}py}} & {\scriptsize\texttt{balanced.\cbrk{}log}} \\ \rowrule Lemma 12′: the determinant formulas (cross-check) & {\scriptsize\texttt{lemma12\_\cbrk{}dets.\cbrk{}py}} & {\scriptsize\texttt{lemma12\_\cbrk{}dets.\cbrk{}txt}} \\ \rowrule After Lemma 12′: c₀ ≈ 0.62 & {\scriptsize\texttt{good\_\cbrk{}involutions.\cbrk{}py}} & {\scriptsize\texttt{good\_\cbrk{}involutions.\cbrk{}log}} \\ \rowrule §7: area ≈ 1.06·N\Sup{3.35}; 36 of 40 exact; flatter below 1/2 & {\scriptsize\texttt{minrep.\cbrk{}c}}, {\scriptsize\texttt{run\_\cbrk{}probe*.\cbrk{}sh}}, {\scriptsize\texttt{ilp\_\cbrk{}check.\cbrk{}py}}, {\scriptsize\texttt{fit\_\cbrk{}area.\cbrk{}py}}, {\scriptsize\texttt{crumple\_\cbrk{}analyze.\cbrk{}py}} & {\scriptsize\texttt{fit\_\cbrk{}area.\cbrk{}txt}}, {\scriptsize\texttt{crumple\_\cbrk{}tables\_\cbrk{}regen.\cbrk{}txt}}, {\scriptsize\texttt{crumple\_\cbrk{}data/\cbrk{}}} \\ \rowrule §7: fingered sheets, adversarial search at the largest scales & {\scriptsize\texttt{fingers.\cbrk{}py}}, {\scriptsize\texttt{fingers\_\cbrk{}opt.\cbrk{}py}}, {\scriptsize\texttt{fingers\_\cbrk{}opt2.\cbrk{}py}} & {\scriptsize\texttt{slabgm\_\cbrk{}data/\cbrk{}fingers\_\cbrk{}*.\cbrk{}txt}}, {\scriptsize\texttt{fingers\_\cbrk{}run1\_\cbrk{}regen.\cbrk{}txt}} \\ \rowrule §7: unique crossing at p ≥ 0.54 and at p = 1/2 & {\scriptsize\texttt{pairs.\cbrk{}c}}, {\scriptsize\texttt{run\_\cbrk{}pairs.\cbrk{}sh}} & {\scriptsize\texttt{slabgm\_\cbrk{}data/\cbrk{}EP*.\cbrk{}txt}}, {\scriptsize\texttt{pairs\_\cbrk{}p0.\cbrk{}5\_\cbrk{}c*.\cbrk{}txt}} \\ \rowrule Prop. 13: the sum −9 711 396; brute-force agreement & {\scriptsize\texttt{prp.\cbrk{}c}}, {\scriptsize\texttt{brute2.\cbrk{}c}}, {\scriptsize\texttt{run\_\cbrk{}prp\_\cbrk{}checks.\cbrk{}sh}} & {\scriptsize\texttt{prp\_\cbrk{}checks.\cbrk{}txt}} \\ \rowrule §7: per-H inequality in d = 2 & {\scriptsize\texttt{prp.\cbrk{}c}} & {\scriptsize\texttt{prp\_\cbrk{}2d.\cbrk{}txt}} \\ \rowrule §7: twist-gap recurrences of orders 16, 69 and 17 & {\scriptsize\texttt{tmrel.\cbrk{}c}}, {\scriptsize\texttt{bm3.\cbrk{}py}} & {\scriptsize\texttt{bm3\_\cbrk{}w3.\cbrk{}txt}}, {\scriptsize\texttt{bm3\_\cbrk{}w4.\cbrk{}txt}}, {\scriptsize\texttt{w3\_\cbrk{}*.\cbrk{}txt}} \\ \rowrule §7: Conjecture E on small tori & {\scriptsize\texttt{cell3d.\cbrk{}c}}, {\scriptsize\texttt{bond2d.\cbrk{}c}}, {\scriptsize\texttt{bond2d\_\cbrk{}x.\cbrk{}c}} & {\scriptsize\texttt{cell3d.\cbrk{}txt}}, {\scriptsize\texttt{bond2d.\cbrk{}txt}}, {\scriptsize\texttt{bond2d\_\cbrk{}4x5\_\cbrk{}record.\cbrk{}txt}} \\ \rowrule §7: R(s) − R(s+e) \textgreater{} 0 on Z₂ × Z₂ × Z\Sub{N}; ratio 0.2396 & {\scriptsize\texttt{tm3.\cbrk{}c}}, {\scriptsize\texttt{bf2d.\cbrk{}py}} & {\scriptsize\texttt{long224\_\cbrk{}*.\cbrk{}out}}, {\scriptsize\texttt{tm3\_\cbrk{}vs\_\cbrk{}brute.\cbrk{}txt}} \\ \rowrule §7: Monte Carlo in d = 4 & {\scriptsize\texttt{bern4d.\cbrk{}c}}, {\scriptsize\texttt{bern4d\_\cbrk{}h.\cbrk{}c}}, {\scriptsize\texttt{vhist\_\cbrk{}summary.\cbrk{}py}} & {\scriptsize\texttt{bern4d\_\cbrk{}scan.\cbrk{}txt}}, {\scriptsize\texttt{vhist\_\cbrk{}summary.\cbrk{}txt}} \\ \rowrule §7: class-wise and cell-wise measures; q = 2 controls & {\scriptsize\texttt{tm4.\cbrk{}c}}, {\scriptsize\texttt{prof\_\cbrk{}scan2.\cbrk{}sh}}, {\scriptsize\texttt{q2time3.\cbrk{}py}}, {\scriptsize\texttt{q2cells.\cbrk{}py}} & {\scriptsize\texttt{scan2\_\cbrk{}N*.\cbrk{}txt}}, {\scriptsize\texttt{profs2\_\cbrk{}*.\cbrk{}txt}}, {\scriptsize\texttt{tm4\_\cbrk{}percell\_\cbrk{}q1.\cbrk{}txt}}, {\scriptsize\texttt{q2\_\cbrk{}time\_\cbrk{}inhom.\cbrk{}txt}}, {\scriptsize\texttt{q2\_\cbrk{}cells.\cbrk{}txt}} \\ \rowrule Prop. 14, d = 3: −243 (486/729), by three programs & {\scriptsize\texttt{orbd.\cbrk{}c}}, {\scriptsize\texttt{verify3d.\cbrk{}py}}, {\scriptsize\texttt{v3d.\cbrk{}py}} & {\scriptsize\texttt{orbd\_\cbrk{}prop14\_\cbrk{}d3.\cbrk{}txt}}, {\scriptsize\texttt{verify3d.\cbrk{}txt}}, {\scriptsize\texttt{v3d.\cbrk{}txt}} \\ \rowrule Prop. 14, d = 4: −7776 & {\scriptsize\texttt{t\_\cbrk{}4d\_\cbrk{}embed.\cbrk{}py}}, {\scriptsize\texttt{orbp.\cbrk{}c}} & {\scriptsize\texttt{t\_\cbrk{}4d\_\cbrk{}embed.\cbrk{}txt}} \\ \rowrule Prop. 14, d = 2: −6.98·10⁻⁹, two exact programs and one floating-point program & {\scriptsize\texttt{blocknu.\cbrk{}c}}, {\scriptsize\texttt{blockdisp.\cbrk{}py}}, {\scriptsize\texttt{exact\_\cbrk{}mstar.\cbrk{}py}}; {\scriptsize\texttt{exact\_\cbrk{}reldp.\cbrk{}py}}, {\scriptsize\texttt{reldp.\cbrk{}py}}; {\scriptsize\texttt{reldp\_\cbrk{}odd.\cbrk{}c}}, {\scriptsize\texttt{t\_\cbrk{}rdp\_\cbrk{}W6.\cbrk{}py}}, {\scriptsize\texttt{rdp.\cbrk{}py}} & {\scriptsize\texttt{exact\_\cbrk{}mstar\_\cbrk{}Z6xZ39.\cbrk{}txt}}, {\scriptsize\texttt{blockdisp\_\cbrk{}Z6xZ39.\cbrk{}txt}}, {\scriptsize\texttt{exact\_\cbrk{}reldp\_\cbrk{}Z6xZ39.\cbrk{}txt}}, {\scriptsize\texttt{t\_\cbrk{}rdp\_\cbrk{}W6.\cbrk{}txt}} \\ \rowrule §7, mechanism: α = 0.1286, first negative at m = 13; W ≤ 5 achiral & {\scriptsize\texttt{blocknu.\cbrk{}c}}, {\scriptsize\texttt{exact\_\cbrk{}mstar.\cbrk{}py}} & {\scriptsize\texttt{alpha\_\cbrk{}Z6xZ39.\cbrk{}txt}}, {\scriptsize\texttt{exact\_\cbrk{}mstar\_\cbrk{}Z6xZ39.\cbrk{}txt}}, {\scriptsize\texttt{chirality\_\cbrk{}small\_\cbrk{}W.\cbrk{}txt}} \\ \rowrule §7, mechanism: q = 2 on Z₇ × Z₂₃₂, q = 4 on Z₇ × Z₃₇₈ & {\scriptsize\texttt{blocknu\_\cbrk{}q.\cbrk{}c}}, {\scriptsize\texttt{t\_\cbrk{}q2exact.\cbrk{}py}} & {\scriptsize\texttt{t\_\cbrk{}q2exact.\cbrk{}txt}} \\ \rowrule §7: table of translation-orbit sums & {\scriptsize\texttt{orbtr\_\cbrk{}q.\cbrk{}c}}, {\scriptsize\texttt{orbtr.\cbrk{}c}} & {\scriptsize\texttt{orbtr\_\cbrk{}q\_\cbrk{}small.\cbrk{}txt}}, {\scriptsize\texttt{q44.\cbrk{}out}}, {\scriptsize\texttt{q223\_\cbrk{}ie5.\cbrk{}out}}, {\scriptsize\texttt{t223\_\cbrk{}ie6.\cbrk{}out}}, {\scriptsize\texttt{q232\_\cbrk{}ie5.\cbrk{}out}}, {\scriptsize\texttt{q232\_\cbrk{}ie6.\cbrk{}out}} \\ \rowrule §7: 12q³ − 4q on Z₆ × Z₂; threshold → 1; thickening to d = 3, 4 & {\scriptsize\texttt{verify2d\_\cbrk{}orbit.\cbrk{}py}}, {\scriptsize\texttt{orb2d\_\cbrk{}rk.\cbrk{}c}}, {\scriptsize\texttt{orb2d\_\cbrk{}rkq.\cbrk{}c}}, {\scriptsize\texttt{altfam.\cbrk{}py}}, {\scriptsize\texttt{altfam\_\cbrk{}formula.\cbrk{}py}}, {\scriptsize\texttt{verify\_\cbrk{}embed.\cbrk{}py}} & {\scriptsize\texttt{verify2d\_\cbrk{}orbit.\cbrk{}txt}}, {\scriptsize\texttt{orb2d\_\cbrk{}rk\_\cbrk{}6x2.\cbrk{}txt}}, {\scriptsize\texttt{orb2d\_\cbrk{}rkq\_\cbrk{}6x2.\cbrk{}txt}}, {\scriptsize\texttt{altfam.\cbrk{}txt}}, {\scriptsize\texttt{altfam\_\cbrk{}formula\_\cbrk{}M3.\cbrk{}txt}}, {\scriptsize\texttt{verify\_\cbrk{}embed.\cbrk{}txt}} \\ \rowrule §7: 714 136 R-orbits, 20 negative rank parts, all achiral; rank parts for e₁₃* & {\scriptsize\texttt{orbtr\_\cbrk{}R.\cbrk{}c}}, {\scriptsize\texttt{verify\_\cbrk{}orbtr\_\cbrk{}witness.\cbrk{}py}} & {\scriptsize\texttt{orbtrR\_\cbrk{}*\_\cbrk{}rank.\cbrk{}txt}}, {\scriptsize\texttt{R232\_\cbrk{}ie6\_\cbrk{}wit.\cbrk{}out}}, {\scriptsize\texttt{verify\_\cbrk{}orbtr\_\cbrk{}witness\_\cbrk{}R232\_\cbrk{}ie6.\cbrk{}txt}} \\ \rowrule Prop. 17′(iii): for every block with at most n\Sub{W} rows, h\Sub{j} = 0 or h\Sub{j} ∈ K\Sub{W}; the extreme ratios & {\scriptsize\texttt{spec\_\cbrk{}exh.\cbrk{}c}} & {\scriptsize\texttt{spec\_\cbrk{}W*\_\cbrk{}n*.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}spec.\cbrk{}log}} \\ \rowrule §7.6: 17 two-row block types on Z₁₁ whose antiperiodic part is not nonincreasing, all at d = 5 & {\scriptsize\texttt{uminus\_\cbrk{}exh.\cbrk{}c}}, {\scriptsize\texttt{ufour\_\cbrk{}exh.\cbrk{}c}}, {\scriptsize\texttt{indep\_\cbrk{}block.\cbrk{}py}} & {\scriptsize\texttt{W10\_\cbrk{}n2.\cbrk{}log}}, {\scriptsize\texttt{W11\_\cbrk{}n2.\cbrk{}log}}, {\scriptsize\texttt{W12\_\cbrk{}n2.\cbrk{}log}}, {\scriptsize\texttt{ufour\_\cbrk{}W*\_\cbrk{}n*.\cbrk{}log}}, {\scriptsize\texttt{indep\_\cbrk{}block\_\cbrk{}W11.\cbrk{}txt}} \\ \rowrule §7.6: Δ\Sup{E}\Sub{7} \textless{} 0 on Z₉ × Z₂ at p = 1/2 for E = all gaps thick, closing gap thick, and the total (two exact programs; interval arithmetic); Σ\Sub{k} Δ\Sup{E}\Sub{k} \textgreater{} 0 in the computed cases; for these events and a thin closing gap, Δ\Sup{E}\Sub{k} \textgreater{} 0 on (0, 1) on every other torus with W ≥ 3, N ≥ 2, WN ≤ 18 (exact counts; the sign on (0, 1) certified in interval arithmetic) and at every computed value with W ≤ 8, N ≥ 2 (floating point for W ≤ 7); the failures on Z₁₀ × Z₂ and at N = 1 & {\scriptsize\texttt{bf2\_\cbrk{}helical.\cbrk{}c}}, {\scriptsize\texttt{bf\_\cbrk{}helical.\cbrk{}c}}, {\scriptsize\texttt{tgm3.\cbrk{}c}} (with {\scriptsize\texttt{-\cbrk{}DEXACT}}), {\scriptsize\texttt{witness\_\cbrk{}check.\cbrk{}py}}, {\scriptsize\texttt{tgx\_\cbrk{}sectors.\cbrk{}py}}, {\scriptsize\texttt{sign\_\cbrk{}cert.\cbrk{}py}}, {\scriptsize\texttt{poly\_\cbrk{}positive.\cbrk{}py}}, {\scriptsize\texttt{check\_\cbrk{}twist\_\cbrk{}positive.\cbrk{}py}}, {\scriptsize\texttt{tgm.\cbrk{}c}}, {\scriptsize\texttt{tgm2.\cbrk{}c}}, {\scriptsize\texttt{make\_\cbrk{}jobs.\cbrk{}py}}, {\scriptsize\texttt{analyse\_\cbrk{}tgm.\cbrk{}py}}, {\scriptsize\texttt{make\_\cbrk{}tables.\cbrk{}py}} & {\scriptsize\texttt{bf2\_\cbrk{}W*\_\cbrk{}N*.\cbrk{}txt}}, {\scriptsize\texttt{bf\_\cbrk{}W*\_\cbrk{}N1.\cbrk{}txt}}, {\scriptsize\texttt{W9x/\cbrk{}}}, {\scriptsize\texttt{witness\_\cbrk{}W9\_\cbrk{}N2.\cbrk{}log}}, {\scriptsize\texttt{sign\_\cbrk{}cert\_\cbrk{}WN18.\cbrk{}log}}, {\scriptsize\texttt{sign\_\cbrk{}cert\_\cbrk{}N1\_\cbrk{}W10.\cbrk{}log}}, {\scriptsize\texttt{poly\_\cbrk{}positive\_\cbrk{}small.\cbrk{}log}}, {\scriptsize\texttt{poly\_\cbrk{}test.\cbrk{}log}}, {\scriptsize\texttt{poly\_\cbrk{}W10\_\cbrk{}N2.\cbrk{}log}}, {\scriptsize\texttt{check\_\cbrk{}twist\_\cbrk{}positive.\cbrk{}log}}, {\scriptsize\texttt{tgx\_\cbrk{}W8\_\cbrk{}N1to4\_\cbrk{}p05.\cbrk{}log}}, {\scriptsize\texttt{tableA.\cbrk{}md}}, {\scriptsize\texttt{full\_\cbrk{}W*.\cbrk{}tsv}} \\ \rowrule §7.6: Δ\Sup{E}\Sub{7} \textless{} 0 at p = 1/2 on Z₁₁ × Z₃, Z₁₂ × Z₃, Z₁₃ × Z₄ and Z₁₄ × Z₅ for the three events, and the other sectors there (exact for the whole space on Z₁₁ × Z₃ and Z₁₂ × Z₃, otherwise floating point with a rounding bound; Monte Carlo agrees); the smallest failing widths for N = 1, …, 5; Z₁₃ × Z₅ positive by Monte Carlo; the square tori with 3 ≤ N ≤ 11 & {\scriptsize\texttt{tlt2.\cbrk{}c}}, {\scriptsize\texttt{tlt4.\cbrk{}c}}, {\scriptsize\texttt{tls.\cbrk{}c}}, {\scriptsize\texttt{tls2.\cbrk{}c}}, {\scriptsize\texttt{tls3.\cbrk{}c}}, {\scriptsize\texttt{tls2x.\cbrk{}c}}, {\scriptsize\texttt{tlt\_\cbrk{}thin.\cbrk{}c}}, {\scriptsize\texttt{combine.\cbrk{}py}}, {\scriptsize\texttt{exact\_\cbrk{}tls.\cbrk{}py}}, {\scriptsize\texttt{merge\_\cbrk{}chunks.\cbrk{}py}}, {\scriptsize\texttt{w14n5\_\cbrk{}analyse2.\cbrk{}py}}, {\scriptsize\texttt{map\_\cbrk{}table.\cbrk{}py}}, {\scriptsize\texttt{square\_\cbrk{}table.\cbrk{}py}}, {\scriptsize\texttt{mc\_\cbrk{}check.\cbrk{}c}}, {\scriptsize\texttt{mc\_\cbrk{}table2.\cbrk{}py}} & {\scriptsize\texttt{p05\_\cbrk{}*.\cbrk{}json}}, {\scriptsize\texttt{exact\_\cbrk{}*.\cbrk{}txt}}, {\scriptsize\texttt{tables\_\cbrk{}final.\cbrk{}md}}, {\scriptsize\texttt{mc\_\cbrk{}table.\cbrk{}md}}, {\scriptsize\texttt{validation.\cbrk{}log}}, {\scriptsize\texttt{w14n5\_\cbrk{}analyse2.\cbrk{}log}}, {\scriptsize\texttt{mc\_\cbrk{}*.\cbrk{}log}} \\ \rowrule §7.6, scaling limit (under the stated assumption): Δ\Sub{7} \textless{} 0 on the grid 0.03 ≤ ρ ≤ 0.97, Δ\Sub{11} and Δ\Sub{13} \textless{} 0 at ρ = 0.02, Δ\Sub{5} \textless{} 0 on the grid 2.82 ≤ ρ ≤ 12, the sign changes; the negative tails at ρ = 0.98, …, 4 and on the square from k = 987; the limit at ρ = 1/3 against the lattice & {\scriptsize\texttt{cont\_\cbrk{}limit2.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}bisect.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}bisect2.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}dense\_\cbrk{}scan.\cbrk{}sh}}, {\scriptsize\texttt{dense\_\cbrk{}summary.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}hik.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}real\_\cbrk{}hik.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}real\_\cbrk{}combine.\cbrk{}py}}, {\scriptsize\texttt{cont\_\cbrk{}check.\cbrk{}py}} & {\scriptsize\texttt{cont2/\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{hik/\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{cont\_\cbrk{}bisect\_\cbrk{}k*.\cbrk{}log}}, {\scriptsize\texttt{cont\_\cbrk{}check.\cbrk{}log}}, {\scriptsize\texttt{continuum.\cbrk{}md}} \\ \rowrule §7: the orbit of 0x3f719edb6 & {\scriptsize\texttt{orbtr\_\cbrk{}r.\cbrk{}c}}, {\scriptsize\texttt{vgfail.\cbrk{}py}}, {\scriptsize\texttt{rorbit\_\cbrk{}rank.\cbrk{}py}}, {\scriptsize\texttt{torbit\_\cbrk{}rank.\cbrk{}py}}, {\scriptsize\texttt{verify\_\cbrk{}orbtr\_\cbrk{}witness.\cbrk{}py}} & {\scriptsize\texttt{g232\_\cbrk{}ie5\_\cbrk{}wit.\cbrk{}out}}, {\scriptsize\texttt{vgfail.\cbrk{}txt}}, {\scriptsize\texttt{rorbit\_\cbrk{}rank.\cbrk{}txt}}, {\scriptsize\texttt{orbtrT\_\cbrk{}232\_\cbrk{}ie5\_\cbrk{}rank.\cbrk{}txt}}, {\scriptsize\texttt{verify\_\cbrk{}orbtr\_\cbrk{}witness\_\cbrk{}3f719edb6.\cbrk{}txt}} \\ \rowrule §7: reflection averages +69 984 and +2 239 488 & {\scriptsize\texttt{v3dref.\cbrk{}c}} & {\scriptsize\texttt{v3dref.\cbrk{}txt}} \\ \rowrule §7: validation of the exact program in d = 4 & {\scriptsize\texttt{orbhom2.\cbrk{}c}}, {\scriptsize\texttt{orbhom3.\cbrk{}c}}, {\scriptsize\texttt{orb4mc.\cbrk{}c}}, {\scriptsize\texttt{orb4ex.\cbrk{}c}}, {\scriptsize\texttt{validate\_\cbrk{}orbhom.\cbrk{}sh}} & {\scriptsize\texttt{orbhom\_\cbrk{}validation.\cbrk{}txt}}, {\scriptsize\texttt{orb4ex\_\cbrk{}2x2x2x2\_\cbrk{}s7.\cbrk{}out}} \\ \rowrule §7: exact and Monte Carlo orbit sums in d = 4 & {\scriptsize\texttt{orbhom2.\cbrk{}c}}, {\scriptsize\texttt{orbhom3.\cbrk{}c}}, {\scriptsize\texttt{orb4mc.\cbrk{}c}}, {\scriptsize\texttt{orb4scan.\cbrk{}c}}, {\scriptsize\texttt{psweep*.\cbrk{}sh}}, {\scriptsize\texttt{susp\_\cbrk{}exact.\cbrk{}sh}}, {\scriptsize\texttt{summarize\_\cbrk{}4d.\cbrk{}py}} & {\scriptsize\texttt{summary\_\cbrk{}4d.\cbrk{}txt}}, {\scriptsize\texttt{exact\_\cbrk{}*.\cbrk{}out}}, {\scriptsize\texttt{mc\_\cbrk{}*.\cbrk{}out}}, {\scriptsize\texttt{s1019\_\cbrk{}*.\cbrk{}out}}, {\scriptsize\texttt{scan4d\_\cbrk{}*.\cbrk{}out}}, {\scriptsize\texttt{exact\_\cbrk{}borderline.\cbrk{}out}} \\ \rowrule §7: coarser groupings fail; subsets of classes on 3×4 & {\scriptsize\texttt{orbtr.\cbrk{}c}}, {\scriptsize\texttt{orb4.\cbrk{}c}}, {\scriptsize\texttt{orbhom2.\cbrk{}c}}, {\scriptsize\texttt{orbtr\_\cbrk{}v2.\cbrk{}c}}, {\scriptsize\texttt{run\_\cbrk{}subsets\_\cbrk{}3x4.\cbrk{}sh}} & {\scriptsize\texttt{orbtr\_\cbrk{}grouping.\cbrk{}txt}}, {\scriptsize\texttt{orb4\_\cbrk{}mode2.\cbrk{}txt}}, {\scriptsize\texttt{exact\_\cbrk{}mode2\_\cbrk{}negatives\_\cbrk{}perclass.\cbrk{}txt}}, {\scriptsize\texttt{subsets\_\cbrk{}3x4.\cbrk{}txt}} \\ \rowrule After Prop. 17: values 64, 96; antiperiodic coefficients & {\scriptsize\texttt{blocknu.\cbrk{}c}}, {\scriptsize\texttt{blockrk.\cbrk{}c}}, {\scriptsize\texttt{urk\_\cbrk{}exh.\cbrk{}c}}, {\scriptsize\texttt{spec\_\cbrk{}search.\cbrk{}py}} & {\scriptsize\texttt{blocknu\_\cbrk{}W8\_\cbrk{}reflsum.\cbrk{}txt}}, {\scriptsize\texttt{blockrk\_\cbrk{}W8\_\cbrk{}33\_\cbrk{}33\_\cbrk{}15.\cbrk{}txt}}, {\scriptsize\texttt{antiperiodic\_\cbrk{}W8\_\cbrk{}block.\cbrk{}txt}}, {\scriptsize\texttt{urk\_\cbrk{}exh\_\cbrk{}W8\_\cbrk{}n2\_\cbrk{}dump.\cbrk{}txt}}, {\scriptsize\texttt{spec\_\cbrk{}search\_\cbrk{}W8.\cbrk{}txt}} \\ \rowrule Lemmas 3, 4 and 5(iv) on small tori (cross-check) & {\scriptsize\texttt{check3d.\cbrk{}c}}, {\scriptsize\texttt{checkq.\cbrk{}c}}, {\scriptsize\texttt{check2d\_\cbrk{}q4.\cbrk{}c}} & {\scriptsize\texttt{check3d.\cbrk{}txt}}, {\scriptsize\texttt{checkq.\cbrk{}txt}}, {\scriptsize\texttt{check2d\_\cbrk{}q4.\cbrk{}txt}} \\ \rowrule Lemma 18(a): identities, and ℓ = 3, 5, 7 exhaustively (cross-check); Smith form of the span of the forms ⟨z, gz⟩ & {\scriptsize\texttt{partner\_\cbrk{}rule.\cbrk{}py}}, {\scriptsize\texttt{qspan\_\cbrk{}ref.\cbrk{}py}} & {\scriptsize\texttt{partner\_\cbrk{}rule.\cbrk{}txt}}, {\scriptsize\texttt{qspan\_\cbrk{}ref.\cbrk{}log}} \\ \rowrule Lemma 29: (S0)–(S5) for L\Sub{c} ≤ 7, 2 ≤ L\Sub{g} ≤ 7; (S0)–(S5) and (U) for 255 pairs up to 32 & {\scriptsize\texttt{netstruct.\cbrk{}py}}, {\scriptsize\texttt{net\_\cbrk{}ref.\cbrk{}py}} & {\scriptsize\texttt{netstruct.\cbrk{}log}}, {\scriptsize\texttt{net\_\cbrk{}ref.\cbrk{}log}} \\ \rowrule After Lemma 30: modules computed directly; placements at the minimal L & {\scriptsize\texttt{netcert.\cbrk{}py}}, {\scriptsize\texttt{net\_\cbrk{}ref.\cbrk{}py}}, {\scriptsize\texttt{placement\_\cbrk{}ref.\cbrk{}py}} & {\scriptsize\texttt{netcert.\cbrk{}log}}, {\scriptsize\texttt{net\_\cbrk{}ref.\cbrk{}log}}, {\scriptsize\texttt{placement\_\cbrk{}ref.\cbrk{}log}} \\ \rowrule §8.7, orders: the floating-point NS order of 𝔇₂₄ is the exact order (284,752 edges, no equal keys); gaps of ⟨ν₀, ·⟩ up to l∞ norm 17 and of ⟨G, ·⟩ up to l∞ norm 2 (with a cross-check) & {\scriptsize\texttt{nsorder\_\cbrk{}exact.\cbrk{}py}}, {\scriptsize\texttt{keygap.\cbrk{}py}}, {\scriptsize\texttt{vkey\_\cbrk{}gap.\cbrk{}py}}, {\scriptsize\texttt{key\_\cbrk{}gaps.\cbrk{}py}} & {\scriptsize\texttt{nsorder\_\cbrk{}exact\_\cbrk{}L24.\cbrk{}log}}, {\scriptsize\texttt{keygap\_\cbrk{}R17.\cbrk{}log}}, {\scriptsize\texttt{vkey\_\cbrk{}gap.\cbrk{}log}}, {\scriptsize\texttt{key\_\cbrk{}gaps.\cbrk{}log}} \\ \rowrule Prop. 31, base case L = 24: (C1) with locality, (C2), (C3) at all 168,712 edges & {\scriptsize\texttt{capcert.\cbrk{}py}}, {\scriptsize\texttt{caplib.\cbrk{}py}}, {\scriptsize\texttt{intlink.\cbrk{}py}}, {\scriptsize\texttt{funnel.\cbrk{}py}}, {\scriptsize\texttt{cfg.\cbrk{}py}}, {\scriptsize\texttt{runcap.\cbrk{}sh}} & {\scriptsize\texttt{cap\_\cbrk{}19\_\cbrk{}10\_\cbrk{}24\_\cbrk{}*of12.\cbrk{}log}}, {\scriptsize\texttt{SCRIPT\_\cbrk{}MD5\_\cbrk{}L24int.\cbrk{}txt}} \\ \rowrule Prop. 31, base case: route structure and window types at L = 24; (F) at L = 24, 25, 30, 60 & {\scriptsize\texttt{analyze\_\cbrk{}routes.\cbrk{}py}}, {\scriptsize\texttt{routes\_\cbrk{}ref.\cbrk{}py}}, {\scriptsize\texttt{transport\_\cbrk{}reach.\cbrk{}py}}, {\scriptsize\texttt{check\_\cbrk{}far\_\cbrk{}windows.\cbrk{}py}} & {\scriptsize\texttt{analyze\_\cbrk{}routes\_\cbrk{}L24.\cbrk{}log}}, {\scriptsize\texttt{routes\_\cbrk{}ref.\cbrk{}log}}, {\scriptsize\texttt{transport\_\cbrk{}reach.\cbrk{}log}}, {\scriptsize\texttt{check\_\cbrk{}far\_\cbrk{}windows.\cbrk{}log}} \\ \rowrule Prop. 31: the other translate need not correspond (17,316 and 17,356 edges of 𝔇₂₅) & {\scriptsize\texttt{make\_\cbrk{}maxable.\cbrk{}py}}, {\scriptsize\texttt{preamble\_\cbrk{}check.\cbrk{}py}} & {\scriptsize\texttt{make\_\cbrk{}maxable\_\cbrk{}24.\cbrk{}log}}, {\scriptsize\texttt{make\_\cbrk{}maxable\_\cbrk{}25.\cbrk{}log}}, {\scriptsize\texttt{preamble\_\cbrk{}check.\cbrk{}log}} \\ \rowrule Prop. 31: integral link steps re-proved by explicit chains; exact integral test against a certified solver; independent replay of (C1) at L = 24, 25; an independent implementation of the search reproduces the base-case statistics chunk by chunk (cross-checks) & {\scriptsize\texttt{certify\_\cbrk{}int\_\cbrk{}steps.\cbrk{}py}}, {\scriptsize\texttt{test\_\cbrk{}bounds\_\cbrk{}torsion.\cbrk{}py}}, {\scriptsize\texttt{test\_\cbrk{}bounds\_\cbrk{}real.\cbrk{}py}}, {\scriptsize\texttt{zcert.\cbrk{}py}}, {\scriptsize\texttt{c1\_\cbrk{}verify.\cbrk{}py}}, {\scriptsize\texttt{refgeom.\cbrk{}py}}, {\scriptsize\texttt{capgen.\cbrk{}cpp}} & {\scriptsize\texttt{certify\_\cbrk{}int\_\cbrk{}steps.\cbrk{}log}}, {\scriptsize\texttt{test\_\cbrk{}bounds\_\cbrk{}torsion.\cbrk{}log}}, {\scriptsize\texttt{tbr\_\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{c1v\_\cbrk{}L24.\cbrk{}log}}, {\scriptsize\texttt{c1v\_\cbrk{}L25.\cbrk{}log}}, {\scriptsize\texttt{c1\_\cbrk{}L25.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}capgen\_\cbrk{}19\_\cbrk{}10\_\cbrk{}24.\cbrk{}log}} \\ \rowrule After Prop. 31: every edge of 𝔇₂₅, 𝔇₂₆ as in (C2) reached by the rule of the proof & {\scriptsize\texttt{make\_\cbrk{}maxable.\cbrk{}py}}, {\scriptsize\texttt{transport\_\cbrk{}reach\_\cbrk{}simple.\cbrk{}py}}, {\scriptsize\texttt{transport\_\cbrk{}reach.\cbrk{}py}} & {\scriptsize\texttt{make\_\cbrk{}maxable\_\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{transport\_\cbrk{}reach\_\cbrk{}simple.\cbrk{}log}}, {\scriptsize\texttt{transport\_\cbrk{}reach.\cbrk{}log}} \\ \rowrule After Prop. 31: full transport 𝔇₂₄ → 𝔇₂₅ → 𝔇₂₆ passes an independent checker & {\scriptsize\texttt{capgen.\cbrk{}cpp}}, {\scriptsize\texttt{transport\_\cbrk{}cert.\cbrk{}py}}, {\scriptsize\texttt{capcheck.\cbrk{}cpp}}, {\scriptsize\texttt{run\_\cbrk{}chain.\cbrk{}sh}} & {\scriptsize\texttt{chain\_\cbrk{}capgen\_\cbrk{}19\_\cbrk{}10\_\cbrk{}2*.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}tr\_\cbrk{}19\_\cbrk{}10\_\cbrk{}2*.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}check\_\cbrk{}19\_\cbrk{}10\_\cbrk{}2*.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}19\_\cbrk{}10\_\cbrk{}24\_\cbrk{}to34.\cbrk{}out}} \\ \rowrule After Prop. 31: 1,416 designs at L = 24 and their transports to L = 25, 26, 30, verified by a second implementation & {\scriptsize\texttt{verify\_\cbrk{}designs.\cbrk{}py}}, {\scriptsize\texttt{verify\_\cbrk{}int\_\cbrk{}edges.\cbrk{}py}}, {\scriptsize\texttt{refgeom.\cbrk{}py}} & {\scriptsize\texttt{vd\_\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{vi\_\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{agg\_\cbrk{}vd.\cbrk{}log}}, {\scriptsize\texttt{agg\_\cbrk{}vi.\cbrk{}log}} \\ \rowrule After Prop. 31: coverage fails from L = 16 to 17 (56 edges) & {\scriptsize\texttt{capgen.\cbrk{}cpp}}, {\scriptsize\texttt{transport\_\cbrk{}cert.\cbrk{}py}}, {\scriptsize\texttt{capcheck.\cbrk{}cpp}}, {\scriptsize\texttt{run\_\cbrk{}chain.\cbrk{}sh}} & {\scriptsize\texttt{chain\_\cbrk{}check\_\cbrk{}19\_\cbrk{}10\_\cbrk{}17.\cbrk{}log}}, {\scriptsize\texttt{chain\_\cbrk{}batch.\cbrk{}out}} \\ \rowrule §8.8, Remarks 1, 2: radii, ball sizes and log₂(1/w₂) & {\scriptsize\texttt{constants.\cbrk{}py}} & {\scriptsize\texttt{constants.\cbrk{}txt}} \\ \rowrule §8.8, Remark 2: second certificate for M = 2 & {\scriptsize\texttt{verify\_\cbrk{}funnel.\cbrk{}py}}, {\scriptsize\texttt{funnel.\cbrk{}py}}, {\scriptsize\texttt{colsearch.\cbrk{}py}}, {\scriptsize\texttt{devices.\cbrk{}py}}, {\scriptsize\texttt{netcheck.\cbrk{}py}}, {\scriptsize\texttt{chunkcount.\cbrk{}py}} & {\scriptsize\texttt{vf\_\cbrk{}*.\cbrk{}log}}, {\scriptsize\texttt{vf\_\cbrk{}totals.\cbrk{}log}}, {\scriptsize\texttt{SCRIPT\_\cbrk{}MD5.\cbrk{}txt}}, {\scriptsize\texttt{netcheck\_\cbrk{}ctrunc25.\cbrk{}log}}, {\scriptsize\texttt{chunkcount.\cbrk{}log}} \\ \rowrule §8.8, Remark 3: end-to-end tests with Z/9 & {\scriptsize\texttt{e2e\_\cbrk{}cap.\cbrk{}py}}, {\scriptsize\texttt{e2e\_\cbrk{}cap\_\cbrk{}cc.\cbrk{}py}}, {\scriptsize\texttt{vimg\_\cbrk{}huge.\cbrk{}c}}, {\scriptsize\texttt{vimg\_\cbrk{}h27.\cbrk{}c}} & {\scriptsize\texttt{z9\_\cbrk{}t0.\cbrk{}log}}, {\scriptsize\texttt{z9\_\cbrk{}cc\_\cbrk{}t1.\cbrk{}log}} \\ \rowrule Prop. 31, base case L = 24: (C1) in local form, (C2), (C3) at all 168,712 edges, (H₂₄), (F); Lemma 26 and the 2-chain of Prop. 28 for (𝔇₂₄, 𝔅₂₄, π₂₄) over every commutative ring; Lean kernel only, standard axioms & Lean 4 project {\scriptsize\texttt{lean\_\cbrk{}kernel/\cbrk{}}} ({\scriptsize\texttt{KernelV3/\cbrk{}Main.\cbrk{}lean}}; leaves {\scriptsize\texttt{KernelV3/\cbrk{}Leaf/\cbrk{}}}, by {\scriptsize\texttt{decide +kernel}}) & {\scriptsize\texttt{logs/\cbrk{}axioms\_\cbrk{}v3.\cbrk{}log}}, build log \\ \rowrule Lemma 30 for (ℓ, a, d) = (2,1,1), (2,2,1), (3,1,1), (5,1,1), all types (the third module statement with the chain anywhere on 𝒩) & Lean 4 project {\scriptsize\texttt{lean/\cbrk{}}} ({\scriptsize\texttt{Char2Proof/\cbrk{}NetsMain.\cbrk{}lean}}, {\scriptsize\texttt{native\_\cbrk{}decide}}) & {\scriptsize\texttt{logs/\cbrk{}axioms.\cbrk{}log}} \\ \rowrule The same statements of Lemma 30 by the Lean kernel alone, on an explicit copy of the networks and certificates; the copy matched to the original data by a generator and a compiled comparison & {\scriptsize\texttt{lean\_\cbrk{}kernel\_\cbrk{}nets/\cbrk{}}} ({\scriptsize\texttt{KernelNets/\cbrk{}}}, by {\scriptsize\texttt{decide +kernel}}; {\scriptsize\texttt{checks/\cbrk{}}}) & {\scriptsize\texttt{logs/\cbrk{}Axioms.\cbrk{}log}}, {\scriptsize\texttt{logs/\cbrk{}EvalEq.\cbrk{}log}}, {\scriptsize\texttt{logs/\cbrk{}dump\_\cbrk{}compare.\cbrk{}log}}, {\scriptsize\texttt{logs/\cbrk{}checks.\cbrk{}log}} \\ \rowrule §8.8, Remark 2: (C1), and (C2) and (C3) with the integral link criterion at all 77,712 relevant edges, for the hole of each of the five networks; Lemma 26 and the 2-chain of Prop. 28 for this region over every commutative ring; Lean kernel only, standard axioms & Lean 4 project {\scriptsize\texttt{lean\_\cbrk{}kernel\_\cbrk{}ct25/\cbrk{}}} ({\scriptsize\texttt{KernelCT25/\cbrk{}MainCT25.\cbrk{}lean}}; leaves {\scriptsize\texttt{KernelCT25/\cbrk{}*/\cbrk{}Leaf/\cbrk{}}}, by {\scriptsize\texttt{decide +kernel}}) & {\scriptsize\texttt{logs/\cbrk{}axioms.\cbrk{}log}}, build and replay logs \\ \rowrule §8.7 and §8.8, Remark 2, orders: the integer key of the Lean projects induces the order of §8.7 on the NS edges of 𝔇₂₄, and the floating-point order of the certificate of Remark 2 on the NS edges of its region & {\scriptsize\texttt{test\_\cbrk{}nskey24.\cbrk{}py}} (with {\scriptsize\texttt{nsorder\_\cbrk{}exact.\cbrk{}py}}), or the direct 60-digit check {\scriptsize\texttt{ns\_\cbrk{}order\_\cbrk{}check.\cbrk{}py}}; {\scriptsize\texttt{test\_\cbrk{}nskey\_\cbrk{}ct25.\cbrk{}py}} & {\scriptsize\texttt{test\_\cbrk{}nskey24\_\cbrk{}D24.\cbrk{}log}}, {\scriptsize\texttt{ns\_\cbrk{}order\_\cbrk{}D24.\cbrk{}log}}, {\scriptsize\texttt{test\_\cbrk{}nskey\_\cbrk{}ct25\_\cbrk{}CT25.\cbrk{}log}} \\ \rowrule §9, before §9.1: the nine permutation-invariant subspaces of H, which of them are isotropic, and which are spanned by vectors with Pf = 0 (an isotropic subspace spanned by such vectors is totally singular); the log calls Cyc \textquotedbl{}Z\textquotedbl{} & {\scriptsize\texttt{dichotomy\_\cbrk{}check.\cbrk{}py}} & {\scriptsize\texttt{dichotomy\_\cbrk{}check.\cbrk{}txt}} \\ \rowrule After Cor. 38: at a = 0.2 a slice contains a cycle with odd spatial winding with probability ≥ 1/2 from b ≈ 0.07 at N = 20 and from b ≈ 0.09 at N = 10 (40 samples per point) & {\scriptsize\texttt{slice\_\cbrk{}wrap.\cbrk{}py}} & {\scriptsize\texttt{slice\_\cbrk{}a0.\cbrk{}2.\cbrk{}txt}} \\ \lastrule \end{longtable}} \section*{\pdfbookmark[1]{Data and code availability}{h32}Data and code availability} The programs, logs and certificates of the Appendix and the Lean projects, with the README and the hash list, are archived at \mbox{\url{https://doi.org/10.5281/zenodo.23631702}}, and \mbox{\url{https://doi.org/10.5281/zenodo.23477694}} points to the latest version of that archive. This paper is archived at \mbox{\url{https://doi.org/10.5281/zenodo.23716219}}. What the Lean projects check, and what is left outside Lean, is stated in §8.6, §8.7 and Remark 2 of §8.8; the README gives the commands. \section*{\pdfbookmark[1]{Use of AI tools}{h33}Use of AI tools} Generative AI tools were used to find and prove the results, to write the programs, to review the proofs and to draft the text. The author is responsible for all statements in the paper. \section*{\pdfbookmark[1]{References}{h34}References} \begin{itemize}[label=\discbullet] \item{} [AG91] M. Aizenman, G. Grimmett, Strict monotonicity for critical points in percolation and ferromagnetic models, J. Stat. Phys. 63 (1991) 817–835. \item{} [BBR14] P. Balister, B. Bollobás, O. Riordan, Essential enhancements revisited, arXiv:1402.0834. \item{} [BDC12] V. Beffara, H. Duminil-Copin, The self-dual point of the two-dimensional random-cluster model is critical for q ≥ 1, Probab. Theory Related Fields 153 (2012) 511–542. \item{} [BKKKL92] J. Bourgain, J. Kahn, G. Kalai, Y. Katznelson, N. Linial, The influence of variables in product spaces, Israel J. Math. 77 (1992) 55–64. \item{} [BR06] B. Bollobás, O. Riordan, Sharp thresholds and percolation in the plane, Random Structures Algorithms 29 (2006) 524–548. \item{} [BS83] C. Borgs, E. Seiler, Lattice Yang–Mills theory at nonzero temperature and the confinement problem, Comm. Math. Phys. 91 (1983) 329. \item{} [DKS20] P. Duncan, M. Kahle, B. Schweinhart, Homological percolation on a torus: plaquettes and permutohedra, Ann. Inst. Henri Poincaré Probab. Stat. 61 (2025) 2235–2261; arXiv:2011.11903. \item{} [dMU21] L. de Moura, S. Ullrich, The Lean 4 theorem prover and programming language, in: Automated Deduction – CADE 28, Lecture Notes in Comput. Sci. 12699, Springer, 2021, 625–635. \item{} [DS22] P. Duncan, B. Schweinhart, Topological phases in the plaquette random-cluster model and Potts lattice gauge theory, Comm. Math. Phys. 406 (2025), Paper No. 145; arXiv:2207.08339. \item{} [DS26] P. Duncan, B. Schweinhart, A topological formula for Potts lattice gauge theory correlations, arXiv:2607.02434. \item{} [DSZ87] P. di Francesco, H. Saleur, J.-B. Zuber, Relations between the Coulomb gas picture and conformal invariance of two-dimensional critical models, J. Stat. Phys. 49 (1987) 57–79. \item{} [GG06] B. T. Graham, G. R. Grimmett, Influence and sharp-threshold theorems for monotonic measures, Ann. Probab. 34 (2006) 1726–1745. \item{} [Gri99] G. Grimmett, Percolation, 2nd ed., Grundlehren der mathematischen Wissenschaften 321, Springer, 1999. \item{} [HS16] Y. Hiraoka, T. Shirai, Tutte polynomials and random-cluster models in Bernoulli cell complexes, RIMS Kôkyûroku Bessatsu B59 (2016) 289–304. \item{} [Kes82] H. Kesten, Percolation Theory for Mathematicians, Progress in Probability and Statistics 2, Birkhäuser, 1982. \item{} [Mat20] The mathlib Community, The Lean mathematical library, in: Proceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2020), ACM, 2020, 367–381. \item{} [OSSS05] R. O\textquotesingle{}Donnell, M. Saks, O. Schramm, R. A. Servedio, Every decision tree has an influential variable, in: Proc. 46th IEEE Symposium on Foundations of Computer Science (FOCS 2005), 31–39. \item{} [Pin94] H. T. Pinson, Critical percolation on the torus, J. Stat. Phys. 75 (1994) 1167–1177. \item{} [Rei00] D. Reimer, Proof of the van den Berg–Kesten conjecture, Combin. Probab. Comput. 9 (2000) 27–32. \item{} [\textquotesingle{}tH79] G. \textquotesingle{}t Hooft, A property of electric and magnetic flux in non-abelian gauge theories, Nucl. Phys. B 153 (1979) 141. \item{} [TY85] E. T. Tomboulis, L. G. Yaffe, Finite temperature SU(2) lattice gauge theory, Comm. Math. Phys. 100 (1985) 313. \item{} [vdBK85] J. van den Berg, H. Kesten, Inequalities with applications to percolation and reliability, J. Appl. Probab. 22 (1985) 556–569. \end{itemize} \end{document}