\documentclass[11pt,letterpaper]{amsart} \usepackage[T1]{fontenc} \usepackage[utf8]{inputenc} \usepackage{amsmath,amssymb,amsthm} \usepackage{longtable} \usepackage[margin=1.25in]{geometry} \usepackage{microtype} \usepackage[colorlinks=true,allcolors=blue]{hyperref} \hypersetup{pdftitle={Flow reconfiguration with one modulus for all graphs, and Cereceda's conjecture for bipartite 2-degenerate graphs},pdfauthor={Dave Goel}, pdfkeywords={nowhere-zero flows, flow reconfiguration, abelian groups, graph recolouring, Cereceda's conjecture, degenerate graphs}} \def\InterfaceMacrosOnly{} \newcommand{\Z}{\mathbb{Z}} \newcommand{\Gam}{\Gamma} % \newcommand{\quo}[1]{\overline{#1}} % \newcommand{\Sig}{\Sigma} % \newcommand{\SigA}{\Sigma_A} % \newcommand{\SigB}{\Sigma_B} % \newcommand{\lift}{\gamma} % \newcommand{\Fl}{F} % \newcommand{\Ik}[1][k]{I_{#1}} % \newcommand{\Mk}[1][k]{M_{#1}} % \newcommand{\supp}{\operatorname{supp}} \newcommand{\eps}{\varepsilon} % \newcommand{\len}{\ell} % \newcommand{\thr}[1]{P_{#1}} % \newcommand{\HP}{H_P} % \newcommand{\Deps}{D_{\eps}} % \newcommand{\Te}{T^{e}} % \newcommand{\Tg}{T^{g}} % \newcommand{\red}{\mathrm{red}} % \newcommand{\Lam}{\Lambda} % \newcommand{\PL}[1]{P(#1)} % \newcommand{\RI}{\textup{(R1)}} \newcommand{\RIm}{\textup{(R1$^-$)}} \newcommand{\RII}{\textup{(R2)}} \newcommand{\RIII}{\textup{(R3$^*$)}} \newcommand{\AV}{\textup{(AV)}} \newcommand{\AVpp}{\textup{(AV$''$)}} \newcommand{\Hone}{\textup{(H1)}} \newcommand{\Htwo}{\textup{(H2)}} \newcommand{\Czero}{\textup{(C0)}} \newcommand{\Cone}{\textup{(C1)}} \newcommand{\Ctwo}{\textup{(C2)}} \newcommand{\Cthree}{\textup{(C3)}} \newcommand{\Cfour}{\textup{(C4)}} \newcommand{\condS}{\textup{(S)}} \newcommand{\condNR}{\textup{(NR)}} \newcommand{\condLone}{\textup{(L1)}} \newenvironment{condlist}[1]{\begin{list}{}{% \settowidth{\labelwidth}{\textup{#1}}\setlength{\labelsep}{0.5em}% \setlength{\leftmargin}{\labelwidth}\addtolength{\leftmargin}{\labelsep}% \setlength{\itemsep}{2pt}\setlength{\topsep}{4pt}\setlength{\parsep}{2pt}\setlength{\listparindent}{0pt}}}% {\end{list}} \theoremstyle{plain} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{criterion}[theorem]{Criterion} \newtheorem{conjecture}[theorem]{Conjecture} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}[theorem]{Remark} \raggedbottom \emergencystretch=1.5em \usepackage{mathtools} \usepackage{enumitem} \usepackage{booktabs} \newcommand{\F}{\mathbb F} \newcommand{\diam}{\operatorname{diam}} \newcommand{\mad}{\operatorname{mad}} \newcommand{\Rk}[1]{\mathcal{R}_{#1}} \newcommand{\ind}{\mathbf 1} \DeclareMathOperator{\im}{im} \setlist{} \renewcommand{\thepart}{\Roman{part}} \begin{document} \title[Flow reconfiguration and Cereceda's conjecture]{Flow reconfiguration with one modulus for all graphs, and Cereceda's conjecture for bipartite 2-degenerate graphs} \author{Dave Goel} \dedicatory{\upshape\footnotesize Version 2.30, October 23, 2026.} \begin{abstract} This paper answers a question of Esperet, Hendrey, Lagoutte, Marseloo, Norin and Steiner on flow reconfiguration and makes inroads into Cereceda's recolouring conjecture for $d=2$. Parts I and II are about flows. For a finite multigraph $G$ and an integer $k\ge 2$, let $M_k(G)$ be the graph on the nowhere-zero $\mathbb{Z}_k$-flows of $G$, and $I_k(G)$ the graph on the integer flows $F$ of $G$ with $1\le|F(e)|\le k-1$ on every edge; two flows are adjacent if the support of their difference is a cycle. We prove that $M_k(G)$ and $I_k(G)$ are connected for every finite bridgeless loopless multigraph $G$ if $k=40$, if $k=48$, or if $k$ is a multiple of $4$ with $k\ge 52$. This answers their Problem~1.2. The same holds for $k=66$, for every prime power $p^N$ with $N\ge 6$, and for a set of moduli of natural density one; these results do not cover every $k\ge 40$. A master theorem for finite abelian groups gives the connectivity of the flow graph $F(G,\Gamma)$ for every abelian group $\Gamma$ of order $40$, and a binary form of it for every abelian group of exponent dividing $4$ and order at least $32$, in particular $\mathbb{F}_2^5$ (Esperet et al.\ proved $\mathbb{F}_2^8$). Part III is about colourings: two proper $k$-colourings of a graph $G$ are adjacent if they differ at exactly one vertex. Cereceda conjectured that this graph has diameter $O(n^2)$ for every $d$-degenerate graph on $n$ vertices and $k=d+2$; the case $d=2$ is open. For four colours we prove the diameter bound $2n(n+1)$ for every bipartite graph of degeneracy at most $2$, with the order $n^2$ sharp in this class, also at girth $6$; the bound $5n(n+1)$ when an independent set of vertices of degree at most $2$ is added; and, with a computer-verified certificate, $4n(n+1)$ for every graph of cone width at most $2$, a class that contains all graphs of treewidth at most $2$ and some graphs of treewidth $3$ with triangles. \end{abstract} \maketitle \section*{Summary of results} This paper has three parts. Parts~I and~II are about the reconfiguration of nowhere-zero flows; Part~III is about the reconfiguration of proper colourings \textup{(}$\Rk{k}(G)$ is the graph on the proper $k$-colourings of $G$, two of them adjacent if they differ at exactly one vertex\textup{)}. On a plane graph $G$ the two objects are dual to each other~\cite{Tutte}: a proper $\Z_k$-colouring of the dual graph $G^*$ gives a nowhere-zero $\Z_k$-flow of $G$ \textup{(}on each edge, the colour of the face on its left minus the colour of the face on its right\textup{)}, every nowhere-zero $\Z_k$-flow of $G$ comes in this way from exactly $k$ colourings, and recolouring one vertex of $G^*$ changes the flow along the boundary of one face of $G$, which is a cycle if $G$ is $2$-connected. Part~III does not use Parts~I and~II, and Parts~I and~II do not use Part~III. The new results are the following. \begin{itemize} \item \emph{One modulus for all graphs} \textup{(}Theorem~\ref{thm:bases}\textup{)}. For every finite bridgeless loopless multigraph $G$, the flow graphs $\Mk(G)$ and $\Ik(G)$ are connected for $k=40$, $k=48$ and every multiple $k$ of $4$ with $k\ge 52$. This answers Problem~1.2 of~\cite{EHLMNS} positively. \item \emph{Further moduli.} The same for $k=66$ \textup{(}Corollary~\ref{cor:k66}\textup{)}, for every prime power $p^N$ with $N\ge 6$ \textup{(}Theorem~\ref{thm:pp}\textup{)}, and for every $k$ that is not a prime power and has at least four prime factors, counted with multiplicity, except $24$, $36$ and $54$ \textup{(}Theorem~\ref{thm:omega4}\textup{)}; the set of all moduli $k$ with this property has natural density~$1$ \textup{(}Corollary~\ref{cor:density}\textup{)}; for one fixed graph, every $k$ above a bound that depends on the graph, primes included \textup{(}Section~\ref{a10:sec}\textup{)}. \item \emph{Groups} \textup{(}Theorem~\ref{thm:G10}\textup{)}. Let $\Gam$ be a finite abelian group that is not cyclic of prime-power order, and let $|\Gam|$ have a divisor $s$ with $4\le s\le|\Gam|/10$. Then the flow graph $\Fl(G,\Gam)$ is connected for every finite bridgeless multigraph $G$, loops allowed; for example for every abelian group of order~$40$. This answers the question of~\cite{EHLMNS} whether the order $2^8$ of the group $\F_2^8$ in their positive answer to Problem~1.4 can be improved. \item \emph{Binary groups} \textup{(}Part~II, Theorem~\ref{thm:main} and Corollary~\ref{cor:exp4}\textup{)}. The same for every abelian group of exponent dividing $4$ and order at least $32$, in particular for $\F_2^5$. \item \emph{Edge connectivity, odd vertices and cubic graphs} \textup{(}Part~II, Theorems~\ref{thm:four}, \ref{thm:oddbound}, \ref{thm:order16bound} and~\ref{thm:cyclic11}\textup{)}. $\Fl(G,\F_2^m)$ is connected for every $4$-edge-connected multigraph $G$ and every $m\ge 3$; $\Fl(G,\F_2^3)$ and $\Fl(G,\F_2^4)$ are connected if every component of the bridgeless multigraph $G$ has at most $44$, respectively $19022$, vertices of odd degree; and $\Fl(G,\F_2^4)$ is nonempty and connected for every finite simple connected bridgeless cubic graph $G$ in which every cyclic edge cut has at least eleven edges. \item \emph{Recolouring bipartite $2$-degenerate graphs} \textup{(}Theorem~\ref{thm:bip}\textup{)}. Every bipartite graph $G$ on $n$ vertices of degeneracy at most $2$ has $\diam \Rk{4}(G) \le 2n(n+1)$. So the case $d=2$ of Cereceda's conjecture~\cite{Cer07} \textup{(}Conjecture~\ref{conj:cereceda}\textup{)} holds for bipartite graphs. The same bound holds when $G$ has an order with at most two later neighbours per vertex in which every cone \textup{(}Section~\ref{sec:setting}\textup{)} induces a bipartite graph; such graphs need not be bipartite \textup{(}Corollary~\ref{cor:par}\textup{)}. \item \emph{The order $n^2$ is sharp} \textup{(}Proposition~\ref{prop:lb}\textup{)} in this class, also at girth~$6$. \item \emph{An added independent set} \textup{(}Theorem~\ref{thm:terminal}\textup{)}. $\diam \Rk{4}(G) \le 5n(n+1)$ if $G$ has an independent set $Z$ of vertices of degree at most $2$ such that $G-Z$ is bipartite and $2$-degenerate; this allows triangles and many odd cycles. \item \emph{Cone width two} \textup{(}Theorem~\ref{thm:conewidth}, computer-assisted\textup{)}. $\diam \Rk{4}(G) \le 4n(n+1)$ for every graph $G$ of cone width at most $2$, a class that contains all graphs of treewidth at most~$2$ and some graphs of treewidth $3$ with triangles, and is closed under subgraphs. \textup{(}For treewidth at most $2$ alone, the bound $2n(n+1)$ of~\cite{BB18} is smaller.\textup{)} \end{itemize} \part{One modulus for all graphs} \setcounter{section}{-1} \section{Introduction}\label{a0:sec} \subsection*{The problem} Let $G$ be a finite multigraph and $k\ge 2$ an integer. A \emph{nowhere-zero $\Z_k$-flow} of $G$ gives every edge a non-zero element of $\Z_k$, with conservation at every vertex; an \emph{integer $k$-flow} gives every edge an integer $F(e)$ with $1\le|F(e)|\le k-1$, with conservation at every vertex. The reconfiguration graphs $\Mk(G)$ and $\Ik(G)$ have these flows as their vertices; two flows are adjacent if their difference is $\lambda\chi_C$ for a cycle $C$ of $G$ and a non-zero $\lambda$, that is, if the support of the difference is a cycle. For a finite abelian group $\Gam$ the graph $\Fl(G,\Gam)$ is defined in the same way, and $\Mk(G)=\Fl(G,\Z_k)$. The exact definitions are in Section~\ref{a1:sec}. A graph with a bridge has no nowhere-zero flow, so the graphs $G$ of interest are the bridgeless ones. These graphs are studied by Esperet, Hendrey, Lagoutte, Marseloo, Norin and Steiner~\cite{EHLMNS}. Their Problem~1.2 asks whether there is \emph{one} integer $k$ such that $\Ik(G)$ is connected for every $2$-edge-connected graph $G$; they prove that $\Ik(G)$ is connected for a modulus $k$ that depends on $G$. For groups they prove that $\Fl(G,\F_2^8)$ is connected for every $2$-edge-connected graph $G$ \textup{(}$\F_2^m$ is the elementary abelian group of order $2^m$\textup{)}, and that $\Fl(G,\Gam)$ is connected for every abelian group $\Gam$ whose order is at least an explicit bound of the order of $10^{694}$; they also show that the modulus $5$ fails. \textup{(}In~\cite{EHLMNS} the graphs $\Ik(G)$ and $\Mk(G)$ are $\mathcal{F}(G,k)$ and $\mathcal{F}(G,\Z_k)$.\textup{)} Their Problem~2.6 asks whether the connectivity of $\Mk(G)$ implies the connectivity of $\Ik(G)$; Cranston, Li, Su, Wang and Xu~\cite{CLSWX} conjecture that $\Ik(G)$ is connected if and only if $\Mk(G)$ is connected. \subsection*{The results} All results below are for every finite bridgeless loopless multigraph $G$; $G$ need not be connected. ``$k$ is reached'' means: one of the results of this paper applies to $k$ \textup{(}in Section~\ref{a9:sec}: $k$ is \emph{covered}, Definition~\ref{a9:def:covered}\textup{)}; then $\Mk(G)$ and $\Ik(G)$ are connected for every such $G$. \begin{condlist}{(6)} \item[\textup{(1)}] \emph{The bases} \textup{(}Theorem~\ref{thm:bases}\textup{)}. $k=40$, $k=48$ and every multiple $k$ of $4$ with $k\ge 52$ are reached. In particular $k=40$ gives a positive answer to Problem~1.2 of~\cite{EHLMNS}. \item[\textup{(2)}] \emph{Further moduli.} $k=66$ (Corollary~\ref{cor:k66}); every prime power $k=p^N$ with $N\ge 6$ (Theorem~\ref{thm:pp}); every product $k=ab$ of two coprime \emph{admitted factors}, that is integers $a$, $b$ each of which is at least $43$ or a product $de$ with $d\ge 4$, $e\ge 7$ (Corollary~\ref{cor:enlarged}): so every multiple of $1120$, every $pq$ with different primes $p,q\ge 43$, and every $p^2q$ with different primes $p\ge 7$, $q\ge 43$; every $k$ that is not a prime power and has at least four prime factors, counted with multiplicity, except $24$, $36$ and $54$ (Theorem~\ref{thm:omega4}); every $k=2p^3$ with a prime $p\ge 5$ (Proposition~\ref{prop:2p3}); and every $k$ to which Criterion~\ref{crit:F10} applies. \item[\textup{(3)}] \emph{Exactly which $k$ these results reach} \textup{(}Corollary~\ref{a9:cor:table}, Proposition~\ref{prop:omega3}\textup{)}, by the number $\Omega(k)$ of prime factors with multiplicity: see Table~\ref{a0:tab:omega}. The reached moduli have natural density~$1$ \textup{(}Corollary~\ref{cor:density}\textup{)}. Up to $100$ they are $40$, $48$, $52$, $56$, $60$, $64$, $66$, $68$, $72$, $76$, $78$, $80$, $84$, $88$, $90$, $92$, $96$, $100$. \item[\textup{(4)}] \emph{One graph} \textup{(}Section~\ref{a10:sec}\textup{)}. For a fixed $G$ the results go from a base $k_0$ to every $k\ge k_0$, prime $k$ included: see Table~\ref{a0:tab:graph}. For example, $\Mk(G)$ and $\Ik(G)$ are connected for every $k\ge 40$ if $G$ has at most $149$ vertices, or at most $189$ edges, or no cycle with more than $39$ edges, or if $G$ is connected with diameter at most~$3$. \item[\textup{(5)}] \emph{Groups} \textup{(}Theorem~\ref{thm:G10}\textup{)}. If $\Gam$ is not cyclic of prime-power order and $|\Gam|$ has a divisor $d$ with $4\le d\le|\Gam|/10$, then $\Fl(G,\Gam)$ is connected for every finite bridgeless multigraph $G$, loops allowed; the least such orders are $40$, $44$, $48$, $50$. The final remarks of~\cite{EHLMNS} name two main questions left open there: Problem~1.2, and whether the order $2^8=256$ of the group $\F_2^8$ in the positive answer to their Problem~1.4 can be improved. Item~\textup{(1)} answers the first, and this theorem answers the second: every abelian group of order $40$ works. By Corollary~\ref{cor:exp4} of Part~II, so does every abelian group of order $32$ whose exponent divides $4$. \item[\textup{(6)}] \emph{Loops.} Every $k$ of \textup{(1)} and \textup{(2)} is reached through Theorem~\ref{thm:Z} or through Theorem~\ref{thm:twocoord}, and part~(a) of these two theorems gives the connectivity of $\Mk(G)$ also for every finite bridgeless multigraph $G$ with loops. \end{condlist} \noindent\emph{Open cases.} For all graphs, the moduli that are not reached include those from $6$ to $39$, those from $41$ to $47$, and the primes from $7$ on \textup{(}Remark~\ref{a9:rem:notcovered}\textup{)}; the modulus $5$ fails~\cite{EHLMNS}. Table~\ref{a0:tab:omega} gives exactly which moduli are reached. For the modular graph alone more is known: by the theorem of~\cite{EHLMNS} for large groups, $\Mk(G)$ is connected for every $k$ above a bound of the order of $10^{694}$, primes included. For $k=44$ the group theorem gives the connectivity of $\Mk[44](G)$ for every bridgeless $G$; the connectivity of $\Ik[44](G)$ is a case of Problem~2.6 of~\cite{EHLMNS}. Section~\ref{a11:sec} explains where the method stops, and Section~\ref{a12:sec} lists open problems. \begin{table}[ht] \centering\small \caption{The moduli $k$ for which this paper proves that $\Mk(G)$ and $\Ik(G)$ are connected for every finite bridgeless loopless multigraph $G$ \textup{(}Corollary~\ref{a9:cor:table}; ``covered'' in Definition~\ref{a9:def:covered}\textup{)}.}\label{a0:tab:omega} \begin{tabular}{@{}p{0.30\textwidth}p{0.62\textwidth}@{}} \hline \rule{0pt}{2.4ex}$k$ & reached exactly when\\[1pt] \hline \rule{0pt}{2.4ex}prime power $p^e$ & $e\ge 6$ \textup{(}no prime\textup{)}\\ $pq$, $p\ne q$ primes & $p\ge 43$ and $q\ge 43$\\ $\Omega(k)=3$, not a prime power & Proposition~\ref{prop:omega3}: $p^2q$ by the rules $(\alpha)$, $(\beta)$, $(\gamma)$ there; every odd $pqr$ except $105$; every $2pq$ except $30$, $42$, $70$, $110$, $182$, $286$, $506$, $646$, $1406$, $1798$\\ $\Omega(k)\ge 4$, not a prime power & $k\notin\{24,36,54\}$\\[1pt] \hline \end{tabular} \end{table} \begin{table}[ht] \centering\small \caption{One graph $G$ \textup{(}finite, bridgeless, loopless\textup{)}: the moduli $k$ for which this paper proves that $\Mk(G)$ and $\Ik(G)$ are connected. $c(G)$ is the number of edges of a longest cycle, $n$ the number of vertices.} \label{a0:tab:graph} \begin{tabular}{@{}p{0.50\textwidth}p{0.24\textwidth}p{0.16\textwidth}@{}} \hline \rule{0pt}{2.4ex}condition on $G$ & every $k$ with & where\\[1pt] \hline \rule{0pt}{2.4ex}none & $k\ge\max(48,c(G)+4)$ & Cor.~\ref{cor:pergraph}\\ $c(G)\le 39$ & $k\ge 40$ & Cor.~\ref{cor:pergraph}\\ at most $189$ edges \textup{(}at most $235$ edges\textup{)} & $k\ge 40$ \textup{(}$k\ge 48$\textup{)} & Cor.~\ref{cor:C189}\\ at most $149$ vertices \textup{(}at most $187$ vertices\textup{)} & $k\ge 40$ \textup{(}$k\ge 48$\textup{)} & Cor.~\ref{cor:N149}\\ $n\ge 150$ vertices & $k\ge 2n/\ln n+4$ & Cor.~\ref{cor:N149}\\ every edge on a cycle with at most $8$ edges & $k\ge 40$ & Cor.~\ref{cor:SC1}\\ connected, diameter at most $3$ & $k\ge 40$ & Cor.~\ref{cor:diam}\\ all edges but at most $40$ on cycles with at most $8$ edges & $k\ge 40$ & Cor.~\ref{cor:SC2}\\[1pt] \hline \end{tabular} \end{table} \subsection*{The method} \emph{From groups to integers.} A non-zero element $z$ of $\Z_k$ is \emph{liftable} if, for every loopless $G$ and every move $f\to f+z\chi_C$ of $\Mk(G)$, all integer lifts of the two flows lie in one component of $\Ik(G)$. Section~\ref{a1:sec} shows that $\pm 1$, $\pm 2$ and $\pm q$ for every divisor $q0$. \end{theorem} \noindent Theorems~\ref{q:euler} and~\ref{q:chebyshev} are used only in Corollary~\ref{cor:density}. \section{Conventions; lifting to integers; liftable steps}\label{a1:sec}\label{sec:flows} \subsection*{Graphs, flows and the reconfiguration graphs}\label{a1:conventions} Graphs are finite multigraphs. Loops are allowed, unless the graph is called loopless. A \emph{cycle} is a connected subgraph in which every vertex has degree~$2$; a loop is a cycle of length~$1$, and two parallel edges form a cycle of length~$2$. An edge that lies on no cycle is a \emph{bridge}; a graph is \emph{bridgeless} if it has no bridge, that is, if every edge lies on a cycle. A graph without edges is bridgeless. A multigraph is \emph{$j$-edge-connected} if, for every partition of its vertex set into two non-empty sets, at least $j$ edges have one end in each set. Every graph has a fixed reference orientation: every edge has a tail and a head \textup{(}for a loop they are the same vertex\textup{)}. Let $x$ be a map from $E(G)$ to an abelian group. Its \emph{boundary} at a vertex $v$ is \[ (\partial x)(v)\;=\sum_{e\,:\,v\text{ is the head of }e}x(e)\;-\sum_{e\,:\,v\text{ is the tail of }e}x(e). \] A loop at $v$ is in both sums, so it adds nothing. The map $x$ is a \emph{flow} if $\partial x=0$ \textup{(}conservation at every vertex\textup{)}. An \emph{integer flow} is a flow with values in $\Z$. Groups are finite abelian groups, if nothing else is said. For a group $\Gam$, a \emph{$\Gam$-flow} of $G$ is a flow $E(G)\to\Gam$; it is \emph{nowhere zero} if no value is $0$. The \emph{support} $\supp(x)$ is the set of the edges $e$ with $x(e)\ne 0$. For a cycle $C$ with a direction, $\chi_C$ is its signed indicator: $\chi_C(e)=1$ if $C$ passes the edge $e$ from the tail to the head \textup{(}\emph{forward}\textup{)}, $\chi_C(e)=-1$ if $C$ passes $e$ from the head to the tail, and $\chi_C(e)=0$ if $e$ is not on $C$. So $\chi_C$ is an integer flow, and $\lambda\chi_C$ is a $\Gam$-flow for every $\lambda$ in $\Gam$. If the reference orientation of one edge is reversed and the value of every flow on this edge is replaced by its negative, then flows stay flows, and nothing else changes. No statement of this paper depends on the reference orientation, and we change it freely inside proofs. $\Fl(G,\Gam)$ is the graph on the nowhere-zero $\Gam$-flows of $G$; two flows are adjacent if their difference is $\lambda\chi_C$ for a cycle $C$ of $G$ and $\lambda\ne 0$ \textup{(}a \emph{move}\textup{)}. $\Mk(G)=\Fl(G,\Z_k)$. $\Ik(G)$ is the graph on the integer flows $F$ of $G$ with $1\le|F(e)|\le k-1$ for every edge $e$; two of them are adjacent if their difference is $\lambda\chi_C$ with an integer $\lambda\ne 0$. An \emph{integer lift} of a $\Z_k$-flow $f$ is a vertex $F$ of $\Ik(G)$ with $F=f$ modulo $k$. These are the three \emph{reconfiguration graphs}. In each of the three graphs, two different vertices are adjacent if and only if the support of their difference is the edge set of a cycle: a flow whose support is a cycle $C$ has, by conservation at the vertices of $C$, the form $\lambda\chi_C$. A graph $G$ without edges has exactly one flow, the empty map; it is a vertex of each of the three reconfiguration graphs. A reconfiguration graph that has no vertex counts as connected. In this section $G$ is a finite multigraph, and loops are allowed. Only Definition~\ref{def:liftable}, Lemma~\ref{lem:liftable} and Corollary~\ref{a1:4p} speak of loopless graphs. \subsection*{The cycle decomposition}\label{a1:cycledec-sec} \begin{lemma}[Cycle decomposition]\label{lem:cycledec}\label{a1:cycledec} Let $G$ be a multigraph. \begin{condlist}{(a)} \item[\textup{(a)}] Let $A$ be an abelian group, finite or not, and let $T_0$ be a spanning forest of $G$ \textup{(}a spanning tree in every component of $G$\textup{)}. For an edge $e$ outside $T_0$ let $C_e$ be the fundamental cycle of $e$ in $T_0$, directed along $e$. Then every $A$-flow $h$ of $G$ is \[ h=\sum_{e\notin E(T_0)}h(e)\,\chi_{C_e}, \] and this is the only way to write $h$ as a sum $\sum_{e\notin E(T_0)}x_e\chi_{C_e}$ with $x_e$ in $A$. In particular every $A$-flow of $G$ is a sum of flows $x\chi_C$ with $x$ in $A$ and $C$ a cycle of $G$. \item[\textup{(b)}] \textup{(}The signed form.\textup{)} Every integer flow $h$ of $G$ is a sum \[ h=\chi_{O_1}+\dots+\chi_{O_j}\qquad(j\ge 0) \] of signed indicators of directed cycles $O_1,\dots,O_j$ of $G$ \textup{(}a cycle can occur several times\textup{)}, such that every $O_i$ passes each of its edges in the direction of the sign of $h$: $\chi_{O_i}(e)\,h(e)>0$ for every edge $e$ of $O_i$. Hence, for every edge $e$: exactly $|h(e)|$ of the cycles contain $e$; every sum of some of the numbers $\chi_{O_1}(e),\dots,\chi_{O_j}(e)$ lies between $0$ and $h(e)$; and all cycles lie in $\supp(h)$. If all values of $h$ are in $\{-1,0,1\}$, then $O_1,\dots,O_j$ are pairwise edge-disjoint. \end{condlist} \end{lemma} \begin{proof} (a) Put $h'=h-\sum_{e\notin E(T_0)}h(e)\chi_{C_e}$. This is an $A$-flow. The cycle $C_e$ contains $e$ and no other edge outside $T_0$, and $\chi_{C_e}(e)=1$. \textup{(}A loop is not in $T_0$, and it is its own fundamental cycle.\textup{)} So $h'$ is $0$ on every edge outside $T_0$. Suppose that $h'\ne 0$. The edges with $h'\ne 0$ form a forest with at least one edge. It has a vertex that meets exactly one of these edges. Conservation at this vertex says that the value of $h'$ on this edge is $0$, a contradiction. So $h'=0$. The sum $\sum x_e\chi_{C_e}$ has the value $x_e$ on the edge $e$ outside $T_0$; so the coefficients are determined by the flow. (b) Induction on $\sum_e|h(e)|$. If this sum is $0$, take $j=0$. Otherwise reverse the reference orientation of every edge $e$ with $h(e)<0$; now $h\ge 0$. If a loop has a positive value, let $O$ be this loop, passed forward. Otherwise use conservation: at every vertex, the sum of $h$ over the edges that enter is equal to the sum over the edges that leave \textup{(}loops are left out on both sides\textup{)}. So the head of an edge with a positive value is the tail of an edge with a positive value. Start with an edge with a positive value and go on in this way. The walk does not end, so it comes back to a vertex that it has seen. This gives a directed cycle $O$ that passes all its edges forward, and $h>0$ on all its edges. Now $h-\chi_O$ is an integer flow with $h-\chi_O\ge 0$ and with a smaller sum. By induction it is a sum of signed indicators of directed cycles that pass their edges forward. Add $\chi_O$. In the first reference orientation, ``forward'' is the direction of the sign of $h(e)$. For a fixed edge $e$, all non-zero numbers $\chi_{O_i}(e)$ have the sign of $h(e)$, and their sum is $h(e)$. This gives the three consequences. If $|h(e)|\le 1$ for all $e$, every edge is on at most one of the cycles. \end{proof} \subsection*{Rounding, integer lifts, and the fibres}\label{a1:lifting-sec} \begin{lemma}[Rounding]\label{lem:rounding}\label{a1:rounding} Let $x$ be a map from $E(G)$ to the real interval $[0,1]$ such that $(\partial x)(v)$ is an integer for every vertex $v$. Then there is a map $z\colon E(G)\to\{0,1\}$ with $\partial z=\partial x$ and with $z(e)=x(e)$ for every edge $e$ with $x(e)$ in $\{0,1\}$. \end{lemma} \begin{proof} Call an edge $e$ fractional if $00$ be the least number such that $x+\delta\chi_C$ has the value $0$ or $1$ on some edge of $C$; it exists, because all values of $x$ on $C$ are strictly between $0$ and $1$. The map $x+\delta\chi_C$ has values in $[0,1]$ and the same boundary as $x$, because $\partial\chi_C=0$. The values on the edges that are not fractional have not changed, and the number of fractional edges is smaller. Repeat this until no edge is fractional. \end{proof} \begin{corollary}[Integer lifts]\label{cor:lifting}\label{a1:lifting} Let $k\ge 2$ and let $f$ be a $\Z_k$-flow of $G$; zeros are allowed. Then there is an integer flow $F$ of $G$ with $F=f$ modulo $k$, with $|F(e)|\le k-1$ for every edge $e$, and with $F(e)=0$ exactly for the edges $e$ with $f(e)=0$. In particular every nowhere-zero $\Z_k$-flow of $G$ has an integer lift. \end{corollary} \begin{proof} For every edge $e$ let $r(e)$ in $\{0,\dots,k-1\}$ be the residue of $f(e)$. Then $\partial r=\partial f=0$ modulo $k$, so the map $x=r/k$ has values in $[0,1)$ and an integer boundary. Lemma~\ref{lem:rounding} gives $z\colon E(G)\to\{0,1\}$ with $\partial z=\partial x$ and with $z(e)=0$ if $r(e)=0$. Put $F=r-kz$. Then $\partial F=\partial r-k\,\partial x=0$, and $F=f$ modulo $k$. If $r(e)=0$, then $F(e)=0$. If $r(e)\ne 0$, then $F(e)$ is $r(e)$ or $r(e)-k$; the first number is in $[1,k-1]$ and the second one is in $[-(k-1),-1]$. \end{proof} Corollary~\ref{cor:lifting} is Tutte's lifting theorem in the form of \cite[Theorem~2.3]{EHLMNS}. \begin{lemma}[Connected fibres]\label{lem:fibres}\label{a1:fibres} Let $k\ge 2$ and let $f$ be a nowhere-zero $\Z_k$-flow of $G$. Then the integer lifts of $f$ induce a connected subgraph of $\Ik(G)$. \end{lemma} \begin{proof} Let $F$ and $F'$ be integer lifts of $f$. On every edge $e$, both $F(e)$ and $F'(e)$ are in $\{r(e),\,r(e)-k\}$, where $r(e)$ in $\{1,\dots,k-1\}$ is the residue of $f(e)$. So $h=(F'-F)/k$ is an integer flow with values in $\{-1,0,1\}$. By Lemma~\ref{lem:cycledec}~(b), $h=\chi_{O_1}+\dots+\chi_{O_j}$ with pairwise edge-disjoint directed cycles. Put $F_i=F+k\,(\chi_{O_1}+\dots+\chi_{O_i})$. On every edge, $F_i$ has the value of $F$ or the value of $F'$. So every $F_i$ is an integer lift of $f$, $F_0=F$, $F_j=F'$, and $F_i-F_{i-1}=k\chi_{O_i}$. This is a path from $F$ to $F'$ in $\Ik(G)$ through integer lifts of $f$. \end{proof} Lemma~\ref{lem:fibres} is \cite[Lemma~5.5]{EHLMNS}. \subsection*{The divisor extension}\label{a1:divext-sec} Let $k=mq$ with integers $m,q\ge 2$. Reduction modulo $q$ is a homomorphism from $\Z_k$ onto $\Z_q$; so the \emph{reduction modulo $q$} of a $\Z_k$-flow, taken edge by edge, is a $\Z_q$-flow. The multiples of $q$ in $\Z_k$ form the subgroup $q\Z_k$, and $qj\mapsto j$ modulo $m$ is an isomorphism from $q\Z_k$ onto $\Z_m$. For a $\Z_k$-flow $y$ with all values in $q\Z_k$ we write $y/q$ for its image, a $\Z_m$-flow. The next two proofs use the same setting. \smallskip \noindent\textbf{Setting.} $w$ is a $\Z_q$-flow of $G$, and $E_0$ is the set of the edges $e$ with $w(e)=0$. By Corollary~\ref{cor:lifting} there is an integer flow $W$ of $G$ with $W=w$ modulo $q$, with $|W(e)|\le q-1$, and with $W(e)=0$ exactly on $E_0$. Reverse the reference orientation of every edge with $W(e)<0$. So from now on \[ 1\le W(e)\le q-1\quad\text{for every edge }e\notin E_0,\qquad W(e)=0\quad\text{for every edge }e\in E_0. \] A \emph{state} is a nowhere-zero $\Z_k$-flow of $G$ whose reduction modulo $q$ is $w$. \smallskip \noindent\textbf{Claim 1} \textup{(}the standard lift of a state\textup{)}. \emph{For every state $h$ there is an integer flow $L$ of $G$ with $|L(e)|\le m-1$ for every edge $e$, such that $F=W+qL$ is an integer lift of $h$. For every edge $e$ in $E_0$: $L(e)\ne 0$, and $h(e)=qL(e)$ modulo $k$.} \smallskip \noindent\emph{Proof.} The $\Z_k$-flow $h-W$ has the reduction $0$ modulo $q$, so all its values are in $q\Z_k$, and $y=(h-W)/q$ is a $\Z_m$-flow. By Corollary~\ref{cor:lifting} there is an integer flow $L$ with $L=y$ modulo $m$, with $|L(e)|\le m-1$, and with $L(e)=0$ exactly where $y(e)=0$. Put $F=W+qL$. This is an integer flow. The residue of $qL(e)$ modulo $k$ depends only on the residue of $L(e)$ modulo $m$; so $qL=h-W$ modulo $k$, and $F=h$ modulo $k$. Let $e\notin E_0$. Then $F(e)=W(e)$ modulo $q$, and this is not $0$; so $F(e)\ne 0$. And $1-q(m-1)\le F(e)\le q-1+q(m-1)$, so $|F(e)|\le k-1$. Let $e\in E_0$. Then $W(e)=0$ and $h(e)\ne 0$, so $y(e)\ne 0$ and $L(e)\ne 0$. So $F(e)=qL(e)$ is not $0$, and $|F(e)|\le q(m-1)4n_H-6$, that is $3m_H\ge 4n_H-5$. \smallskip \emph{\textup{(X4)}.} The graph $H-V'$ has $n_H-|V'|\ge 3$ vertices, fewer than $H$. So it is not dense: $3\,|E(H-V')|\le 4\,(n_H-|V'|)-6$. With \textup{(X3)}, the number of the edges of $H$ with an end in $V'$ is $m_H-|E(H-V')|\ge(4\,|V'|+1)/3$. \smallskip \emph{Step 1: $n_H\ge 6$, $m_H>n_H$, and every vertex of $H$ has degree at least $2$.} We have $m_H>(4/3)\,n_H-2\ge n_H-1$, so $m_H\ge n_H$, and $H$ contains a cycle. It has at least $6$ edges, so $n_H\ge 6$. Then $(4/3)\,n_H-2\ge n_H$, so $m_H>n_H$. By \textup{(X4)} for a set $V'$ with one vertex, more than $4/3$ edges meet this vertex. \smallskip \emph{Step 2: $H$ is connected.} Suppose that $H$ is the disjoint union of two graphs $H_1$, $H_2$ with $n_1$, $n_2$ vertices and $m_1$, $m_2$ edges. By Step~1 every vertex has degree at least $2$, so both graphs contain a cycle, and $n_1,n_2\ge 6$. Both have fewer vertices than $H$, so they are not dense: $m_i\le(4/3)\,n_i-2$. Then $m_H\le(4/3)\,n_H-4$, a contradiction. \smallskip \emph{Step 3: the partitions with at most two edges.} Let $V(H)$ be the disjoint union of two non-empty sets $V_1$, $V_2$ such that at most two edges of $H$ join $V_1$ and $V_2$. Let $H_i$ be the subgraph of $H$ induced on $V_i$, with $n_i$ vertices and $m_i$ edges. If $m_i\le(4/3)\,n_i-2$ for $i=1$ and for $i=2$, then $m_H\le(4/3)\,n_H-4+2$, a contradiction. So $m_i>(4/3)\,n_i-2$ for one $i$. This $H_i$ has fewer vertices than $H$, so it is not dense; hence $n_i\le 2$. A graph with at most $2$ vertices and without parallel edges has no cycle. So $n_i=1$ and $m_i=0$, or $n_i=2$ and $m_i=1$. Every vertex of $V_i$ has degree at least $2$ in $H$, and at most two edges leave $V_i$. So exactly two edges leave $V_i$, and every vertex of $V_i$ has degree exactly $2$ in $H$. Result: \emph{exactly two edges join $V_1$ and $V_2$, and one of the two sets has at most $2$ vertices, all of degree $2$ in $H$.} \smallskip \emph{Step 4: the threads.} Let $V_0$ be the set of the vertices of $H$ of degree at least $3$. It is not empty: $H$ is connected, every degree is at least $2$, and $m_H>n_H$. Because $H$ is connected and $V_0$ is not empty, $H$ is the union of paths and cycles of the following kind \textup{(}we call them \emph{threads} here\textup{)}: the two ends of a thread are in $V_0$ \textup{(}they are equal if the thread is a cycle\textup{)}, its interior vertices have degree $2$ in $H$, every edge of $H$ is on exactly one thread, and every vertex of degree $2$ is an interior vertex of exactly one thread. Let $I$ be the set of the interior vertices of a thread, and let $I$ be not empty. The first and the last edge of the thread are the only edges between $I$ and $V(H)\setminus I$. The set $V(H)\setminus I$ contains a vertex of degree at least $3$. So by Step~3, $|I|\le 2$. Hence every thread has at most $3$ edges. A thread that is a cycle would be a cycle with at most $3$ edges; there is none. So every thread is a path with two different ends in $V_0$. Let $H_0$ be the multigraph with the vertex set $V_0$ in which every thread is replaced by one edge between its ends. $H_0$ is loopless, and $H$ is a subdivision of $H_0$ in which every edge of $H_0$ is replaced by a path with at most $3$ edges. \smallskip \emph{Step 5: $H_0$ is $3$-edge-connected and has at least two vertices.} A thread has two different ends, so $|V_0|\ge 2$. Let $V_0$ be the disjoint union of two non-empty sets $V'_0$, $V''_0$ such that at most two edges of $H_0$ join $V'_0$ and $V''_0$ \textup{(}possibly none\textup{)}. Put the interior vertices of every thread on the side of the first end of the thread. This is a partition of $V(H)$ into two sets; at most two edges of $H$ join them \textup{(}for every thread between $V'_0$ and $V''_0$ its last edge\textup{)}; and both sets contain a vertex of degree at least $3$. This contradicts Step~3. \smallskip Steps~4 and~5 give \textup{(X1)}. \textup{(X2)}: a cycle of $H_0$ becomes a cycle of $H$ when its edges are replaced by their paths; it has at least $6$ edges; two parallel edges of $H_0$ are a cycle of $H_0$, and each of the two paths has at most $3$ edges. \textup{(X3)}, second part: let $j$ be the number of the vertices of degree $2$ of $H$. Then $n_H=n_0+j$ and $m_H=m_0+j$, and $3\,(m_0+j)\ge 4\,(n_0+j)-5$ gives $j\le 3m_0-4n_0+5$. \end{proof} Part~(a) of Lemma~\ref{lem:extraction} is the extraction lemma of \cite[Lemma~6.1]{EHLMNS}. The proof above treats the case of a cycle that is a thread explicitly. \begin{remark}[What the later sections use]\label{a2:rem-extraction} The proofs of the later sections use of a good subgraph $H$ only the two lines of Definition~\ref{def:good}. In the case \textup{(SUB)} this is \textup{(X1)}: $H_0$ is loopless and $3$-edge-connected, it has at least two vertices, and every edge of $H_0$ is replaced by a path with at most $3$ edges. So every good subgraph serves, and Lemma~\ref{lem:extraction}~(a) is all that the proofs need. The equivalent form of (a): a loopless multigraph with $n\ge 3$ vertices that contains no good subgraph has at most $(4/3)\,n-2$ edges. Part~(b) says more. Whenever the lemma is used for a graph $G$ without a cycle with at most $5$ edges, the good subgraph can be chosen as a dense subgraph with the least number of vertices, and then it has the properties \textup{(X2)}, \textup{(X3)} and \textup{(X4)} in addition. These three properties are not used in the proofs. They limit the edges of $H_0$ that are replaced by paths with $3$ edges, and the conditions \condS{} and \condNR{} of Definition~\ref{def:counts} speak of these edges. The bound in \textup{(X3)} is attained. Let $G$ consist of two vertices joined by three paths with $3$ edges each. It has $8$ vertices and $9>(4/3)\cdot 8-2$ edges, and no cycle with fewer than $6$ edges. A subgraph with $j\le 7$ vertices has at most $j$ edges if $j\ge 6$, and at most $j-1$ edges if $j\le 5$; so it is not dense. Hence $H=G$ in Lemma~\ref{lem:extraction}~(b), with $n_0=2$, $m_0=3$ and $6=3m_0-4n_0+5$ vertices of degree $2$. \end{remark} \subsection*{The truncated cubic reduction}\label{a2:reduction-sec} \begin{definition}[Truncated cubic graph]\label{a2:tc} A \emph{truncated cubic $3$-edge-connected graph} is a graph $Q=Q^+-v$, where $Q^+$ is a cubic $3$-edge-connected graph on at least four vertices and $v$ is a vertex of $Q^+$. \end{definition} Such a graph is simple and connected; three of its vertices have degree $2$ and all others have degree $3$ \textup{(}Lemma~\ref{app:count}~(a)\textup{)}. A \emph{cut vertex} of a connected graph is a vertex whose deletion leaves a graph that is not connected. A \emph{block} of a graph is a maximal connected subgraph without a cut vertex. A graph with two vertices and one or more edges between them has no cut vertex. We use the standard facts: every edge of a graph lies in exactly one block, and every cycle lies in one block. \begin{lemma}[Truncated cubic reduction]\label{lem:reduction}\label{a2:reduction} \begin{condlist}{(a)} \item[\textup{(a)}] Let $H_0$ be a loopless $3$-edge-connected multigraph on at least two vertices. Then every block $K$ of $H_0$ is loopless and $3$-edge-connected, has at least two vertices, and has no cut vertex. \item[\textup{(b)}] Let $K$ be a loopless $3$-edge-connected multigraph on at least two vertices without a cut vertex. Then there is a truncated cubic $3$-edge-connected graph $Q_K$ with these properties: $E(K)$ is a subset of $E(Q_K)$; the edges of $Q_K$ that are not in $E(K)$ form pairwise disjoint cycles; and $K$ is obtained from $Q_K$ by the contraction of each of these cycles to one vertex, that is, by the contraction of the edges of $Q_K$ that are not in $E(K)$. \end{condlist} \end{lemma} \begin{proof} (a) $H_0$ is connected and loopless and has an edge at every vertex; so a block has an edge, it has at least two vertices, and it is loopless. It has no cut vertex by definition. Let $V(K)$ be the disjoint union of two non-empty sets $V_1$, $V_2$. For a vertex $x$ of $K$ let $V[x]$ be the set of the vertices $y$ of $H_0$ that are joined to $x$ by a path without an edge of $K$ \textup{(}so $x$ is in $V[x]$\textup{)}. Every vertex of $H_0$ is in some $V[x]$: take a path to a vertex of $K$ and stop at the first vertex of $K$. No vertex is in $V[x]$ and in $V[x']$ with $x\ne x'$: otherwise there is a path without an edge of $K$ between two different vertices of $K$, and with no other vertex in $K$; the graph $K$ together with this path is connected, has no cut vertex, and is larger than $K$. An edge $yy'$ of $H_0$ that is not in $K$ has both ends in one set $V[x]$: a path from $x$ to $y$ without an edge of $K$, together with this edge, shows that $y'$ is in $V[x]$. So the edges of $H_0$ between the union of the $V[x]$ with $x$ in $V_1$ and the union of the $V[x]$ with $x$ in $V_2$ are exactly the edges of $K$ between $V_1$ and $V_2$. There are at least three of them, because $H_0$ is $3$-edge-connected. (b) \emph{Step 1.} Add to $K$ a new vertex $v$ and three new edges from $v$ to vertices of $K$, not all three to the same vertex. Call the result $K^+$. It is loopless. It has no cut vertex: for a vertex $u$ of $K$ the graph $K-u$ is connected, and $v$ has a neighbour other than $u$; so $K^+-u$ is connected; and $K^+-v=K$ is connected. It is $3$-edge-connected: let $V(K^+)$ be the disjoint union of two non-empty sets; if both contain vertices of $K$, at least three edges of $K$ join them; if not, one of them is $\{v\}$, and three edges join them. So every vertex of $K^+$ has degree at least $3$. A neighbour of $v$ has degree at least $3$ in $K$, so it has degree at least $4$ in $K^+$. \emph{Step 2.} For every vertex $u\ne v$ of $K^+$ of degree $j\ge 4$ take a cycle $O_u$ of length $j$ with new vertices and new edges. Replace $u$ by $O_u$: the $j$ edges at $u$ end now at the $j$ vertices of $O_u$, one edge at each vertex. Call the result $K^c$. It is cubic and loopless, the cycles $O_u$ are pairwise disjoint, and the contraction of every $O_u$ to one vertex gives $K^+$ back. Every vertex of $O_u$ has exactly one edge that is not on $O_u$, and the other end of this edge is not on $O_u$, because $K^+$ is loopless. \emph{Step 3: $K^c$ is $3$-edge-connected.} Let $V(K^c)$ be the disjoint union of two non-empty sets $V_1$, $V_2$ such that at most two edges join them. First case: every cycle $O_u$ is in one of the two sets. Then the contraction of the cycles gives two non-empty sets of vertices of $K^+$ with the same edges between them; this contradicts Step~1. Second case: some cycle $O_u$ has vertices in both sets. Then at least two edges of $O_u$ join the two sets. So exactly two edges join them, both are on $O_u$, and every other cycle $O_{u'}$ is in one of the two sets. A vertex of $O_u$ in $V_1$ has a neighbour outside $O_u$, and this neighbour is in $V_1$; the same for $V_2$. Now contract all cycles other than $O_u$ and delete $O_u$. The result is $K^+-u$. Its vertex set is split into two non-empty sets, which come from $V_1$ and from $V_2$, with no edge between them. So $K^+-u$ is not connected; this contradicts Step~1. \emph{Step 4.} By Step~1 the ends in $K$ of the three edges at $v$ have degree at least $4$ in $K^+$. So in $K^c$ the three edges at $v$ end at three different vertices, and $K^c$ has at least four vertices. Put $Q_K=K^c-v$. This is a truncated cubic $3$-edge-connected graph. Its edges are the edges of $K$ and the edges of the cycles $O_u$. The contraction of every $O_u$ to one vertex gives $K^+-v=K$. \end{proof} \subsection*{Threads and the auxiliary graph}\label{a2:threads-sec}\label{sec:frames} In the rest of this section $H$ is a subdivision as in Definition~\ref{def:good}~\textup{(SUB)}, and $A$ is an abelian group. \textup{(}In the later sections $A$ is the quotient $\Gam/B$ of a group $\Gam$ by a subgroup $B$.\textup{)} \begin{definition}[Threads and the auxiliary graph]\label{def:threads}\label{a2:threads} For an edge $t$ of $H_0$, the path $\thr{t}$ of $H$ that replaces $t$ is the \emph{thread} of $t$, and $\len(t)\in\{1,2,3\}$ is its number of edges. Every thread is directed as a path; its edges and the edge $t$ have this direction as reference orientation. Let $K_1,\dots,K_r$ be the blocks of $H_0$. For every block $K_b$ let $Q_b$ be a truncated cubic $3$-edge-connected graph \textup{(}a cubic $3$-edge-connected graph on at least four vertices minus one vertex\textup{)} such that $E(K_b)$ is a subset of $E(Q_b)$ and $K_b$ is obtained from $Q_b$ by the contraction of the other edges of $Q_b$. $Q$ is the disjoint union of the $Q_b$. The edges of $H_0$ are the \emph{original} edges of $Q$; the other edges of $Q$ are \emph{expansion} edges, and $\len(t)=0$ for an expansion edge $t$. For an $A$-flow $Y$ of $Q$, $\pi(Y)$ is the $A$-flow of $H$ that has the value $Y(t)$ on every edge of $\thr{t}$, for every original edge $t$. \end{definition} The graphs $Q_b$ exist by Lemma~\ref{lem:reduction}. We fix them, one for every block, with disjoint vertex sets. A vertex $x$ of $K_b$ is a set $V_x$ of vertices of $Q_b$: a component of the graph formed by the expansion edges of $Q_b$. The map $\pi(Y)$ is defined in the same way for every abelian group in the place of $A$, also for $\Z$; Lemma~\ref{a2:flows}~(c) says that it is a flow. \begin{lemma}[The flows of $H$, of $H_0$ and of $Q$]\label{a2:flows} Let $A$ be an abelian group, finite or not. \begin{condlist}{(a)} \item[\textup{(a)}] Every cycle of $H$ is a union of threads. If $O$ is a cycle of $H$ with a direction, then $\chi_O$ is constant, $1$ or $-1$, on the edges of each thread of $O$. Every $A$-flow of $H$ is constant on every thread. So the $A$-flows of $H$ are the maps that have the value $h_0(t)$ on every edge of $\thr{t}$, for an $A$-flow $h_0$ of $H_0$. \item[\textup{(b)}] The restriction of an $A$-flow of $H_0$ to a block $K_b$ is an $A$-flow of $K_b$. \item[\textup{(c)}] For every $A$-flow $Y$ of $Q$, the restriction of $Y$ to the original edges is an $A$-flow of $H_0$, and $\pi(Y)$ is an $A$-flow of $H$. Every $A$-flow of $H$ is $\pi(Y)$ for some $A$-flow $Y$ of $Q$. \item[\textup{(d)}] Let $C$ be a cycle of $Q$ with a direction. Then there are pairwise edge-disjoint directed cycles $O_1,\dots,O_j$ of $H$ \textup{(}$j\ge 0$\textup{)} such that $\pi(x\chi_C)=x\,(\chi_{O_1}+\dots+\chi_{O_j})$ for every $x$ in $A$. \end{condlist} \end{lemma} \begin{proof} (a) An interior vertex of a thread has degree $2$ in $H$, and the thread is directed as a path. So a cycle that contains one edge of a thread contains the whole thread and passes it in one direction; this gives the first two statements. And conservation at an interior vertex says that its two edges have the same value. For a map that is constant on every thread, conservation at a vertex of $H_0$ is the same condition in $H$ and in $H_0$. (b) By Lemma~\ref{lem:cycledec}~(a) an $A$-flow of $H_0$ is a sum of flows $x\chi_C$ with cycles $C$ of $H_0$. Every cycle lies in one block. So the restriction to $K_b$ is the sum of the terms with $C$ in $K_b$. (c) Let $x$ be a vertex of $K_b$. Add the conservation conditions of $Y$ at the vertices of $V_x$. Every edge with both ends in $V_x$ is in this sum with both signs. What is left is conservation at $x$ for the restriction of $Y$ to $E(K_b)$. So this restriction is an $A$-flow of $K_b$. At a vertex of $H_0$, conservation is the sum of the conservation conditions in the blocks that contain this vertex. So the restriction of $Y$ to the original edges is an $A$-flow of $H_0$, and by (a) $\pi(Y)$ is an $A$-flow of $H$. Now let an $A$-flow of $H$ be given. By (a) and (b) it comes from an $A$-flow $h_0$ of $H_0$ whose restriction to every block is a flow of the block. Let $x$ be a vertex of $K_b$. For a vertex $y$ in $V_x$ let $\nu(y)$ be the boundary at $y$ of the map that is $h_0$ on the original edges and $0$ on the expansion edges. The sum of $\nu(y)$ over $y$ in $V_x$ is the boundary of $h_0$ at $x$ in $K_b$, which is $0$. The expansion edges connect $V_x$; choose a spanning tree of $V_x$ that consists of expansion edges. There are values on the edges of this tree whose boundary at every $y$ in $V_x$ is $-\nu(y)$. \textup{(}Induction on $|V_x|$: for one vertex, $\nu=0$. Otherwise let $y$ be a vertex that meets only one edge $yy'$ of the tree; the value on this edge is fixed by the condition at $y$; delete $y$ and add $\nu(y)$ to $\nu(y')$.\textup{)} Give all other expansion edges the value $0$. Do this for every $x$ and every $b$. The result is an $A$-flow $Y$ of $Q$ that is equal to $h_0$ on the original edges; so $\pi(Y)$ is the given flow. (d) By (c), used for the group $\Z$, the map $\pi(\chi_C)$ is an integer flow of $H$. Its values are in $\{-1,0,1\}$. By Lemma~\ref{lem:cycledec}~(b) it is $\chi_{O_1}+\dots+\chi_{O_j}$ with pairwise edge-disjoint directed cycles of $H$. And $\pi(x\chi_C)=x\,\pi(\chi_C)$ edge by edge. \end{proof} \subsection*{Frames}\label{a2:frames-sec} \begin{definition}[$2$-joins and $2$-constructions]\label{a2:twoC} A \emph{$2$-join} of two graphs $U'$ and $U''$ with disjoint vertex sets is a graph that consists of $U'$, of $U''$ and of exactly two new edges, each with one end in $U'$ and one end in $U''$ \textup{(}the two new edges can be parallel\textup{)}. A graph is \emph{$2$-constructible} if it has one vertex and no edge, or if it is a $2$-join of two $2$-constructible graphs. A \emph{$2$-construction} of a $2$-constructible graph $U_0$ fixes one way to obtain $U_0$. If $U_0$ has more than one vertex, it names two graphs $U'_0$, $U''_0$ of which $U_0$ is a $2$-join, and it contains a $2$-construction of $U'_0$ and a $2$-construction of $U''_0$. The \emph{stages} of the $2$-construction are $U_0$ and, if $U_0$ has more than one vertex, the stages of the $2$-constructions of $U'_0$ and of $U''_0$. So the stages are subgraphs of $U_0$, and the stages with one vertex are the vertices of $U_0$. A stage $U$ with more than one vertex is a \emph{node}; it stands for the $2$-join that makes it. The two stages $U'$, $U''$ that it joins are its \emph{operands}, and its two new edges are the \emph{edges of the node}. \end{definition} \begin{definition}[Frame]\label{def:frame}\label{a2:frame} A \emph{frame} of $H$ consists of the following, for every component $Q_b$ of $Q$: \begin{condlist}{(F3)} \item[\textup{(F1)}] a \emph{well colouring} of $Q_b$: every edge is blue or red, the blue edges form a spanning tree $T_b$ of $Q_b$, and the graph $Q_b/\red$ \textup{(}its vertices are the components of the red edges, its edges are the blue edges\textup{)} is $2$-constructible; \item[\textup{(F2)}] a $2$-construction of $Q_b/\red$ from single vertices; its $2$-joins are the \emph{nodes}, and the two graphs that a node joins are its \emph{operands}; \item[\textup{(F3)}] for every node $i$, names $e_i$ and $f_i$ for its two new edges. The set $N_i=\{e_i,f_i\}$ is the \emph{pair} of the node $i$. \end{condlist} The nodes of all components are numbered $1,\dots,s$: component after component, and in every component every node after all nodes that are subgraphs of its two operands. $T$ is the union of the trees $T_b$; its edge set is the disjoint union of the pairs $N_1,\dots,N_s$. Every truncated cubic $3$-edge-connected graph has a well colouring \textup{(}Lemma~\ref{app:wellcolouring}\textup{)}, so every $H$ has a frame. The names in \textup{(F3)} are free: at every node the two names can be exchanged. \end{definition} In \textup{(F1)}, ``the components of the red edges'' are the \emph{red components}: the components of the graph with the vertex set $V(Q_b)$ and the red edges. The ends of a blue edge in $Q_b/\red$ are the red components of its two ends. \smallskip \noindent\textbf{Notation for a frame.} Let a frame of $H$ be given. For a stage $U$ of the $2$-construction of $Q_b/\red$, $V_U$ is the union of the red components that are the vertices of $U$, and $Q_U$ is the subgraph of $Q_b$ induced on $V_U$, the \emph{graph of $U$}. For a node $i$, $U_i$ is its stage, and $U'_i$, $U''_i$ are its operands. A node $h$ is \emph{below} the node $i$ if $h\ne i$ and $U_h$ is a subgraph of $U_i$; then $U_h$ lies in $U'_i$ or in $U''_i$ and is a stage of its $2$-construction \textup{(}two stages are disjoint or nested, Lemma~\ref{app:count}~(b), and $U_h\ne U_i$\textup{)}, and $hi$, then $C_l$ does not contain $f_i$. \textup{(}It can contain $e_i$.\textup{)} \item[\textup{(O5)}] For every abelian group $A$, every $A$-flow of $Q$ is a sum $x_1\chi_{C_1}+\dots+x_s\chi_{C_s}$ with $x_i$ in $A$, in exactly one way. \item[\textup{(O6)}] $\HP$ is connected and contains every vertex of $H$ that is not an interior vertex of a free thread; and the cycle $\Deps$ exists for every free edge $\eps$. \end{condlist} \end{lemma} In \textup{(O1)}, ``$\Tg$ is a spanning tree of the graph $Q_U$ of a stage $U$'' means: the edges of $\Tg$ in $Q_U$ form a spanning tree of $Q_U$. This holds for every stage, also for the stages with one vertex. \begin{proof} \textup{(O1)} The red edges form a forest by \textup{(G1)}. Let one edge $g_i$ in every pair $N_i$ be chosen. For a stage $U$ let $\Tg_U$ be the graph with the vertex set $V_U$ whose edges are the red edges with both ends in $V_U$ and the edges $g_h$ of the nodes $h$ with $U_h$ a subgraph of $U$. We show by induction on the $2$-construction that $\Tg_U$ is a spanning tree of $Q_U$. If $U$ has one vertex, then $V_U$ is one red component; it is connected by its red edges, and they form a forest; so it is a tree. Let $U=U_i$ be a node. By induction $\Tg_{U'_i}$ and $\Tg_{U''_i}$ are trees; by \textup{(G4)} their vertex sets are disjoint with the union $V_{U_i}$, and $g_i$ has one end in each of them. A red edge with both ends in $V_{U_i}$ has both ends in one red component, so in $V_{U'_i}$ or in $V_{U''_i}$. A node $h$ with $U_h$ a subgraph of $U_i$ is $i$, or $U_h$ is a subgraph of $U'_i$ or of $U''_i$. So $\Tg_{U_i}$ is the union of the two trees and the edge $g_i$: a spanning tree of $Q_{U_i}$. For $U=Q_b/\red$ this is the statement on $Q_b$. \emph{Consequence.} For every stage $U$ and two vertices of $V_U$, the path of $\Tg$ between them lies in $Q_U$: it is the path of the tree $\Tg_U$. \smallskip \textup{(O2)} By \textup{(G2)} the edges of $Q$ outside $\Te$ are $f_1,\dots,f_s$. So $C_i$ consists of $f_i$ and of the path of $\Te$ between the ends of $f_i$. Both ends are in $V_{U_i}$, so by the consequence of \textup{(O1)} this path lies in $Q_{U_i}$. By \textup{(G4)} one end of $f_i$ is in $V_{U'_i}$ and the other one is in $V_{U''_i}$, and the only edges of $Q_{U_i}$ between these two sets are $e_i$ and $f_i$. The path does not contain $f_i$. So it contains $e_i$. We note the result as \textup{($\ast$)}: \emph{$C_i$ is a cycle of $Q_{U_i}$; it consists of $f_i$, of $e_i$, of a path of $\Te$ in $Q_{U'_i}$ and of a path of $\Te$ in $Q_{U''_i}$.} \smallskip \textup{(O3)} Let $li$. \item[\textup{(c)}] $\Delta_i=\chi_{C_i^{f}}-\chi_{C_i^{e}}$ is an integer flow of $Q$. It is $0$ on $N_i$ and outside $Q_{U_i}$. So it is an integer flow of the graph $Q_{U'_i}\cup Q_{U''_i}$, whose edges are red edges and edges of pairs $N_h$ with $h$ below $i$. In particular $\Delta_i$ is $0$ on $N_i\cup N_{i+1}\cup\dots\cup N_s$. For every abelian group $A$ and every $x$ in $A$, the $A$-flow $x\Delta_i$ is a sum of flows $x'\chi_C$ with $x'$ in $A$ and $C$ a cycle of $Q_{U'_i}\cup Q_{U''_i}$. \end{condlist} \end{lemma} \begin{proof} (a) The proof of \textup{(O2)}, with the tree $T^{f}$ in the place of $\Te$ and with the names $e_i$ and $f_i$ exchanged, gives the first two sentences. By \textup{($\ast$)} in the proof of Lemma~\ref{lem:official}, the cycle $C_i^{e}$ has the same form, with paths of $\Te$. Let both cycles pass $f_i$ forward, say from its end in $V_{U'_i}$ to its end in $V_{U''_i}$ \textup{(}the other case is the same\textup{)}. After $f_i$ each of the two cycles stays in $Q_{U''_i}$ until it passes $e_i$, from the end of $e_i$ in $V_{U''_i}$ to its end in $V_{U'_i}$; then it goes back in $Q_{U'_i}$. So both pass $e_i$ in the same direction. (b) Let $h$ be below $i$. All edges of $C_i^{e}$ other than $f_i$ are in $\Te$, and $f_h$ is not in $\Te$. All edges of $C_i^{f}$ other than $e_i$ are in $T^{f}$, and $e_h$ is not in $T^{f}$. Let $h\ne i$ be not below $i$. Both cycles are cycles of $Q_{U_i}$, and by \textup{(G3)} the pair $N_h$ has no edge in $Q_{U_i}$. A node $h>i$ is not below $i$. (c) $\Delta_i$ is a difference of two integer flows. It is $0$ on $N_i$ by (a), and outside $Q_{U_i}$ because both cycles are in $Q_{U_i}$. By \textup{(G4)} it is therefore $0$ outside $Q_{U'_i}\cup Q_{U''_i}$; so it is an integer flow of this graph. By \textup{(G3)} the edges of this graph are red edges and edges of pairs $N_h$ with $U_h$ a subgraph of $U'_i$ or of $U''_i$, that is with $h$ below $i$. The last sentence is Lemma~\ref{lem:cycledec}~(a) for the $A$-flow $x\Delta_i$ of the graph $Q_{U'_i}\cup Q_{U''_i}$. \end{proof} \subsection*{Lists and counts}\label{a2:lists-sec} In this subsection $\Gam$ is a group, $B$ is a subgroup, $A=\Gam/B$, and $x\mapsto\quo{x}$ is the quotient map. The \emph{$A$-coordinate} of a $\Gam$-flow $f$ is the $A$-flow $u$ with the values $u(e)=\quo{f(e)}$. The \emph{$A$-part} of an element $x$ of $\Gam$ is $\quo{x}$. \begin{definition}[Lists; $H$-good flows]\label{def:lists}\label{a2:lists} Let a frame of $H$ be given. A \emph{list assignment} $\Lam$ gives every blue edge $t$ of $Q$ a set $\Lam(t)$ of elements of $A$, with $\Lam(t)$ empty for every expansion edge. $\PL{\Lam}$ is the set of the $A$-flows $Y$ of $Q$ with $Y(t)\notin\Lam(t)$ for every blue edge $t$. Let $G$ be a multigraph that contains $H$, and $f$ a nowhere-zero $\Gam$-flow of $G$ with $A$-coordinate $u$. In the case \textup{(SUB)}, $f$ is \emph{$H$-good} if $u(e)\ne 0$ for every edge $e$ of every protected thread; then \[ \Lam_f(t)=\{\,-u(e) : e\in E(\thr{t})\,\}\qquad\textup{for every original blue edge $t$.} \] So $|\Lam_f(t)|\le\len(t)$, $0$ is in $\PL{\Lam_f}$, and for an $A$-flow $Y$ of $Q$: $Y$ is in $\PL{\Lam_f}$ if and only if $u+\pi(Y)$ is non-zero on all protected threads. \textup{(}$u+\pi(Y)$ means: add $\pi(Y)$ on $E(H)$ and nothing outside $E(H)$.\textup{)} In the case \textup{(CYC)}, $f$ is \emph{$H$-good} if $u(e)\ne 0$ for every edge $e$ of $H$. \end{definition} The three claims after the display are clear: the thread $\thr{t}$ has $\len(t)$ edges; no element $-u(e)$ of a list is $0$; and $Y(t)\ne -u(e)$ for all edges $e$ of $\thr{t}$ says that $u+\pi(Y)$ is not $0$ on $\thr{t}$. \begin{definition}[The counts; counted frame]\label{def:counts}\label{a2:counts} Let a frame of $H$ be given. The \emph{counts} for $A$ and $B$ are the following inequalities. \begin{condlist}{(C4)} \item[\Cone] $2\,\len(e_i)+\len(f_i)<|A|$ for every node $i$ \textup{(}the walk lemma\textup{)}. \item[\Ctwo] $\len(e_i)+\len(f_i)+\max\{\len(e_i),\len(f_i)\}<|A|$ for every node $i$ \textup{(}the reverse choice for a step with non-zero $A$-part\textup{)}. \item[\Cthree] $\len(e_i)+\len(f_i)+2<|A|$ for every node $i$ \textup{(}the reverse choice for a step in $B$\textup{)}. \item[\Cfour] $\len(\eps)<|B|$ for every free edge $\eps$ \textup{(}the preparation in $B$\textup{)}. \end{condlist} The frame is \emph{counted} \textup{(}for $A$ and $B$\textup{)} if \Cone--\Cfour{} hold. Two conditions on a frame that do not use the groups: \begin{condlist}{(NR)} \item[\condS] no pair $N_i$ consists of two original edges whose threads both have $3$ edges; \item[\condNR] no free thread has $3$ edges \textup{(}every original edge whose thread has $3$ edges is blue\textup{)}. \end{condlist} \end{definition} \begin{remark}\label{rem:counts}\label{a2:rem-counts} The left sides of \Cone, \Ctwo, \Cthree, \Cfour{} are at most $9$, $9$, $8$, $3$. So every frame is counted if $|A|\ge 10$ and $|B|\ge 4$. If the names are chosen with $\len(e_i)\le\len(f_i)$ for every $i$, then \Ctwo{} implies \Cone. With these names: a frame with \condS{} is counted if $|A|\ge 9$ and $|B|\ge 4$ \textup{(}left sides at most $7$, $8$, $7$, $3$\textup{)}; a frame with \condNR{} is counted if $|A|\ge 10$ and $|B|\ge 3$ \textup{(}the left side of \Cfour{} is at most $2$\textup{)}. \end{remark} \section{Step systems and the master theorem}\label{a3:sec} This section proves the master theorem in its two forms (Theorems~\ref{thm:master} and~\ref{thm:masterframe}), three corollaries, and the integer form (Theorem~\ref{thm:Z}). All proofs are complete here. What is taken from other sections is cited at the place of use, with the property that is used: liftable steps (Definition~\ref{def:liftable}, Lemma~\ref{lem:liftable}); good subgraphs (Definition~\ref{def:good}) and the extraction lemma (Lemma~\ref{lem:extraction}); threads, the auxiliary graph, frames, official cycles, lists and counts (Definitions~\ref{def:threads}, \ref{def:frame}, \ref{def:official}, \ref{def:lists}, \ref{def:counts} and Lemma~\ref{lem:official}). The averaging conditions \AV{} and \AVpp{} are hypotheses in this section; Lemma~\ref{lem:AVG} proves them for products. Two rules hold in every proof of this section. \begin{condlist}{(b)} \item[\textup{(a)}] Every use of $|A|$ or of $|B|$, other than in the hypotheses of Lemmas~\ref{lem:reverse} and~\ref{lem:walk}, names the count that it uses: \Czero{} of Theorem~\ref{thm:master}, or one of \Cone--\Cfour{} of Definition~\ref{def:counts}; every use of the bound $\len(t)\le 3$ names \textup{(i)}. Remark~\ref{a3:rem:countuses} lists the places. \item[\textup{(b)}] \RI{} with six deleted elements, the count \Cone{} and the walk lemma (Lemma~\ref{lem:walk}) are used only in the proof of Lemma~\ref{lem:L1}. Every later proof uses Lemma~\ref{lem:L1} only through its conclusion. This is why the frame form of the theorem holds. \end{condlist} \subsection*{Step systems} Throughout, $\Gam$ is a finite abelian group, $B$ a subgroup, $A=\Gam/B$, and $x\mapsto\quo{x}$ is the quotient map. For a $\Gam$-flow $f$ of a multigraph the $A$-flow $u=\quo{f}$ (edge by edge) is the \emph{$A$-coordinate} of $f$. \begin{definition}[Step system]\label{def:stepsystem} A \emph{step system} is a set $\Sig$ of non-zero elements of $\Gam$ with $\Sig=-\Sig$. Put \[ \SigB=\Sig\cap B,\qquad \SigA=\{\,\quo{\lambda} : \lambda\in\Sig\,\}\setminus\{0\}, \] and for every $s$ in $\SigA$ fix one $\lift(s)$ in $\Sig$ with $\quo{\lift(s)}=s$. A \emph{$\Sig$-move} of a multigraph $G$ is a move $f\to f+\lambda\chi_C$ of $\Fl(G,\Gam)$ with $\lambda$ in $\Sig$. Two flows are \emph{$\Sig$-joined} if a sequence of $\Sig$-moves leads from one to the other. \end{definition} So a $\Sig$-move leads from a nowhere-zero flow to a nowhere-zero flow. The reverse of a $\Sig$-move is a $\Sig$-move, because $\Sig=-\Sig$. A flow is $\Sig$-joined to itself (the empty sequence). \begin{definition}[The conditions on a step system]\label{def:conditions} \begin{condlist}{(R3$^*$)} \item[\RI] For every set $X$ of at most $6$ elements of $A$, any two elements of $A\setminus X$ are joined by a walk in $A\setminus X$ all of whose steps are in $\SigA$. \item[\RIm] The same for every set $X$ of at most $5$ elements of $A$. \item[\RII] For every set $X$ of at most $2$ elements of $B$, any two elements of $B\setminus X$ are joined by a walk in $B\setminus X$ all of whose steps are in $\SigB$. \item[\RIII] Any two non-zero elements of $\Gam$ are joined by a walk in $\Gam\setminus\{0\}$ all of whose steps are in $\Sig$. \item[\AV] For every loopless $3$-edge-connected multigraph $G$ with $n\ge 2$ vertices and $m$ edges, every nowhere-zero $\Gam$-flow of $G$ is $\Sig$-joined to a nowhere-zero $\Gam$-flow whose $A$-coordinate has at most $m/9$ zeros. \item[\AVpp] For every loopless $3$-edge-connected multigraph $G$ with $n\ge 2$ vertices, every nowhere-zero $\Gam$-flow of $G$ is $\Sig$-joined to a nowhere-zero $\Gam$-flow $f'$ such that the multigraph $(V(G),\supp(\quo{f'}))$ contains a good subgraph. \end{condlist} \end{definition} A \emph{step} of a walk is the difference of two consecutive terms. \RI{} implies \RIm. \RI{} or \RIm{} with $X$ empty says: every element of $A$ is a sum of elements of $\SigA$. \RII{} with $X$ empty says: every element of $B$ is a sum of elements of $\SigB$. \medskip \noindent\textbf{Five elementary facts.} Here $G$ is a multigraph, $f$ a $\Gam$-flow of $G$ and $u=\quo{f}$. An \emph{integer flow} of $G$ is a map $E(G)\to\Z$ with conservation at every vertex. \begin{condlist}{(E5)} \item[\textup{(E1)}] $f(e)\ne 0$ if and only if $u(e)\ne 0$, or $u(e)=0$ and $f(e)$ is an element of $B$ other than $0$. \textup{(}If $u(e)=0$, then $f(e)$ is in $B$.\textup{)} \item[\textup{(E2)}] If $y$ is in $B$ and $C$ is a cycle of $G$, then $f+y\chi_C$ has the $A$-coordinate $u$. \item[\textup{(E3)}] \emph{Cycle decomposition} \textup{(}Lemma~\ref{lem:cycledec}\textup{)}. Every flow of $G$ with values in an abelian group is a sum of flows $x\chi_C$, where $C$ is a cycle of $G$ and $x$ is in the group. Every integer flow of $G$ with values in $\{0,1,-1\}$ is a sum $\chi_{O_1}+\dots+\chi_{O_\nu}$ with pairwise edge-disjoint cycles $O_1,\dots,O_\nu$ of $G$ \textup{(}$\nu=0$ for the zero flow\textup{)}. \item[\textup{(E4)}] Every $A$-flow of $G$ is the $A$-coordinate of some $\Gam$-flow of $G$. \item[\textup{(E5)}] \emph{Contraction.} Let $G'$ be obtained from $G$ in this way: the vertices of each class of a partition of $V(G)$ are identified to one vertex, and some edges that have both ends in one class are deleted. Then the restriction of a flow of $G$ to $E(G')$ is a flow of $G'$ \textup{(}values in any abelian group\textup{)}. \end{condlist} \noindent\emph{Proofs.} (E1) and (E2) hold because $B$ is the kernel of the quotient map. (E3) The first statement is Lemma~\ref{lem:cycledec}~(a). The second one is Lemma~\ref{lem:cycledec}~(b): for an integer flow with values in $\{0,1,-1\}$ the cycles of the signed form are pairwise edge-disjoint. (E4) By (E3) the $A$-flow is a sum $\sum_r a_r\chi_{C_r}$. Choose $x_r$ in $\Gam$ with $\quo{x_r}=a_r$. The $\Gam$-flow $\sum_r x_r\chi_{C_r}$ has this $A$-coordinate. (E5) Add the conservation equations of $G$ over the vertices of one class. An edge with both ends in the class enters the class once and leaves it once, so it drops out. What is left is the conservation equation of $G'$ at the vertex made from the class.\qed \begin{lemma}\label{lem:AVtoAVpp} \AV{} implies \AVpp. \textup{(}A loopless multigraph with $n\ge 3$ vertices and more than $(4/3)\,n-2$ edges contains a good subgraph; for $n=2$ the support has two parallel edges.\textup{)} \end{lemma} \begin{proof} Let $G$ be loopless and $3$-edge-connected with $n\ge 2$ vertices and $m$ edges. Every vertex has degree at least $3$, so $m\ge 3n/2$. Let $f'$ be the flow that \AV{} gives, with $A$-coordinate $u'$. Then \[ |\supp(u')|\ \ge\ m-m/9\ \ge\ (8/9)(3/2)\,n\ =\ (4/3)\,n\ >\ (4/3)\,n-2 . \] Let $n\ge 3$. The multigraph $(V(G),\supp(u'))$ is loopless and has $n$ vertices and more than $(4/3)\,n-2$ edges. By Lemma~\ref{lem:extraction}~(a) it contains a cycle with at most $5$ edges or a subdivision of a loopless $3$-edge-connected multigraph on at least two vertices in which every edge is replaced by a path with at most $3$ edges. The cycle has at least $2$ vertices, because the multigraph is loopless. This is a good subgraph (Definition~\ref{def:good}). Let $n=2$. Then $m\ge 3$, so $\supp(u')$ has at least $8m/9>2$ edges. $G$ is loopless, so two of them are parallel. They form a cycle with $2$ edges and $2$ vertices, a good subgraph of the kind (CYC). \end{proof} \subsection*{The subdivision: what is used} From here to the end of the proof of Lemma~\ref{lem:P}, $H$ is a subdivision as in Definition~\ref{def:good}~(SUB), with the graphs $H_0$, $K_b$, $Q_b$, $Q$ of Definition~\ref{def:threads}, and a frame of $H$ is given (Definition~\ref{def:frame}). The proofs use the following properties of these objects, and no other property. \begin{condlist}{(iii)} \item[\textup{(i)}] \emph{Threads} \textup{(Definition~\ref{def:threads})}. Every edge of $H$ lies on exactly one thread $\thr{t}$, where $t$ is an edge of $H_0$, and $1\le\len(t)\le 3$. The interior vertices of a thread have degree $2$ in $H$; the other vertices of $H$ are the vertices of $H_0$. $H_0$ is connected and loopless, every edge of $H_0$ lies in exactly one block $K_b$, and every cycle of $H_0$ lies in one block. \item[\textup{(ii)}] \emph{The auxiliary graph} \textup{(Definition~\ref{def:threads})}. $E(K_b)$ is a subset of $E(Q_b)$, and $K_b$ is obtained from $Q_b$ by the contraction of the expansion edges: the vertices of $K_b$ are the components of the subgraph of $Q_b$ formed by all vertices and the expansion edges, and an original edge joins the components of its two ends. $\len(t)=0$ for an expansion edge $t$. \item[\textup{(iii)}] \emph{The frame} \textup{(Definition~\ref{def:frame})}. The blue edges of $Q_b$ form a spanning tree $T_b$ of $Q_b$. Every blue edge of $Q$ lies in exactly one pair $N_i$. A \emph{stage} is a graph $U$ of the $2$-construction \textup{(}Definition~\ref{a2:twoC}\textup{)}; its vertices are components of the red edges. $V_U$ is the union of these components, and $Q_U$ is the subgraph of $Q$ induced on $V_U$. If $U'$ and $U''$ are the two operands of the node $i$, then $V_{U'}$ and $V_{U''}$ are disjoint, and each of the edges $e_i$, $f_i$ has one end in $V_{U'}$ and the other end in $V_{U''}$ \textup{(}a $2$-join adds two edges, each with one end in each of the two graphs; Lemma~\ref{app:operands}~(c)\textup{)}. The names $e_i$, $f_i$ can be exchanged at every node; the result is a frame again. \item[\textup{(iv)}] \emph{Official cycles} \textup{(Definition~\ref{def:official}, Lemma~\ref{lem:official})}. The statements \textup{(O1)--(O6)}. Of \textup{(O1)} we use, with $Q_U$ as in the notation for a frame after Definition~\ref{def:frame} and the paragraph after Lemma~\ref{lem:official}: for every choice of one edge $g_i$ in every pair, $\Tg$ is a spanning tree of every component $Q_b$; and for every stage $U$, the edges of $\Te$ in $Q_U$ form a spanning tree of $Q_U$ \textup{(}the proof of \textup{(O1)} shows this for every stage, by induction on the $2$-construction, from Lemmas~\ref{app:count} and~\ref{app:operands}\textup{)}. The blue edges of $\Tg$ are the edges $g_i$. The cycle $\Deps$ contains $\thr{\eps}$ and no other free thread, and it is directed along $\thr{\eps}$. \item[\textup{(v)}] \emph{Lists and counts} \textup{(Definitions~\ref{def:lists}, \ref{def:counts})}. A list assignment is empty on every expansion edge. An $H$-good flow is nowhere zero. For an $H$-good flow $f$ with $A$-coordinate $u$: $|\Lam_f(t)|\le\len(t)$; $0$ is in $\PL{\Lam_f}$; and an $A$-flow $Y$ of $Q$ is in $\PL{\Lam_f}$ if and only if $u+\pi(Y)$ is non-zero on every edge of every protected thread. Here and below, $u+\pi(Y)$ means: add $\pi(Y)$ on $E(H)$ and nothing outside $E(H)$. \end{condlist} Whenever $H$ is a subgraph of a multigraph $G$, the reference orientation of $G$ on $E(H)$ is the direction of the threads. This is no restriction: a change of the reference orientation of an edge changes the sign of all flow values on this edge and nothing else. For a flow $Y$ of $Q$, with values in $A$ or in $\Z$, $\pi(Y)$ has the value $Y(t)$ on every edge of $\thr{t}$, for every original edge $t$. Lemma~\ref{a2:flows} describes the flows of $H$ and of $Q$. Below we use its parts \textup{(a)}, \textup{(c)} and \textup{(d)}: every cycle of $H$ is a union of threads, and for a cycle $O$ of $H$ with a direction, $\chi_O$ is constant, $1$ or $-1$, on the edges of each thread of $O$; every flow of $H$ is constant on the edges of each thread; $\pi(Y)$ is an $A$-flow of $H$ for every $A$-flow $Y$ of $Q$, and every $A$-flow of $H$ is $\pi(Y)$ for some $A$-flow $Y$ of $Q$; and for every cycle $C$ of $Q$ with a direction there are pairwise edge-disjoint directed cycles $O_1,\dots,O_\nu$ of $H$ \textup{(}$\nu=0$ is possible\textup{)} with $\pi(x\chi_C)=x\,(\chi_{O_1}+\dots+\chi_{O_\nu})$ for every $x$ in $A$. \subsection*{The reverse choice and the walk lemma} \begin{lemma}[Reverse choice]\label{lem:reverse} Let $\Lam$ be a list assignment. \begin{condlist}{(a)} \item[\textup{(a)}] If $|\Lam(e_i)|+|\Lam(f_i)|<|A|$ for every node $i$, then for every $A$-flow $Y_0$ of $Q$ there are $x_1,\dots,x_s$ in $A$ with $Y_0+x_1\chi_{C_1}+\dots+x_s\chi_{C_s}$ in $\PL{\Lam}$. \item[\textup{(b)}] Let $E'$ be a set of red edges of $Q$, and for every $\eps$ in $E'$ let $\Phi(\eps)$ be a set of elements of $A$, not all of $A$. Let $i(\eps)$ be the least $i$ with $\eps$ on $C_i$, if there is one. If for every node $i$ \[ |\Lam(e_i)|+|\Lam(f_i)|+\sum_{\eps\in E',\ i(\eps)=i}|\Phi(\eps)|\ <\ |A|, \] then there is $Y$ in $\PL{\Lam}$ with $Y(\eps)\notin\Phi(\eps)$ for every $\eps$ in $E'$. \end{condlist} \end{lemma} \begin{proof} Every blue edge of $Q$ lies in exactly one pair \textup{(iii)}. By (O2) the cycle $C_i$ contains $e_i$ and $f_i$; so $\chi_{C_i}$ has the value $1$ or $-1$ on each of them. By (O3) no cycle $C_l$ with $li}x_l\chi_{C_l}$. For an edge $t$ of $N_i$ the flow $Y=Y_0+x_1\chi_{C_1}+\dots+x_s\chi_{C_s}$ has the value \[ Y(t)=Z_i(t)+x_i\,\chi_{C_i}(t), \] whatever $x_{i-1},\dots,x_1$ are (O3). Since $\chi_{C_i}(t)$ is $1$ or $-1$, the condition $Y(t)\notin\Lam(t)$ forbids exactly $|\Lam(t)|$ values of $x_i$. The two edges of $N_i$ together forbid at most $|\Lam(e_i)|+|\Lam(f_i)|<|A|$ values. So $x_i$ can be chosen. In the end $Y(t)\notin\Lam(t)$ for every blue edge $t$. (b) Let $E'_0$ be the set of the edges of $E'$ that lie on no official cycle. For $\eps$ in $E'_0$ choose $y_\eps$ in $A$ outside $\Phi(\eps)$. The blue edges form a spanning tree of the component of $Q$ that contains $\eps$ \textup{(iii)}. Let $F_\eps$ be the cycle formed by $\eps$ and the path of this tree between the ends of $\eps$, directed along $\eps$. Its only red edge is $\eps$. Put $Y_0=\sum_{\eps\in E'_0}y_\eps\chi_{F_\eps}$; so $Y_0(\eps)=y_\eps$ for every $\eps$ in $E'_0$. Choose $x_s,\dots,x_1$ as in (a), for the flow $Y=Y_0+x_1\chi_{C_1}+\dots+x_s\chi_{C_s}$, with one more condition on $x_i$: $Y(\eps)\notin\Phi(\eps)$ for every $\eps$ in $E'$ with $i(\eps)=i$. For such an $\eps$ the value $Y(\eps)=Z_i(\eps)+x_i\,\chi_{C_i}(\eps)$ does not depend on $x_{i-1},\dots,x_1$, because no cycle $C_l$ with $li$. By (O4), $C_{l_r}$ does not contain $f_i$. So $X_i^f(Y_r)=X_i^f(Y_{r-1})$, and \[ |X_i(Y_{r-1})\cup X_i(Y_r)|\ \le\ |X_i^e(Y_{r-1})|+|X_i^e(Y_r)|+|X_i^f(Y_{r-1})|\ =\ 2\,|\Lam(e_i)|+|\Lam(f_i)|\ <\ |A|. \] The last inequality is the second hypothesis of the lemma; for $\Lam=\Lam_f$ it is the count \Cone, see \textup{(5)}. Choose $c_r$ outside $X_i(Y_{r-1})\cup X_i(Y_r)$.\label{a3:use:walk:C1} \end{condlist} In \textup{(3b)} and \textup{(3c)}: $c_{r-1}$ and $c_r$ are outside $X_i(Y_{r-1})$, a set of at most $6$ elements by the first hypothesis of the lemma. By \RI{} there is a walk from $c_{r-1}$ to $c_r$ in $A\setminus X_i(Y_{r-1})$ with steps in $\SigA$.\label{a3:use:walk:R1} Add the flows $Y_{r-1}+c\chi$ for the terms $c$ of this walk after the first term. They are in $\PL{\Lam_i}$ by $(*)$, and two consecutive ones differ by $\sigma\chi_{C_i}$ with $\sigma$ in $\SigA$. Then add $Y_r+c_r\chi$. It is in $\PL{\Lam_i}$ by $(*)$, because $c_r\notin X_i(Y_r)$; it differs from the term before it, $Y_{r-1}+c_r\chi$, by $\sigma_r\chi_{C_{l_r}}$. \item[\textup{(4)}] \emph{The end of the step.} The sequence has reached $Y_n+c_n\chi$ with $c_n\notin X_i(Y_n)$. Also $0\notin X_i(Y_n)$ by $(*)$, because $Y_n=Y'$ is in $\PL{\Lam_i}$. By \RI{} there is a walk from $c_n$ to $0$ in $A\setminus X_i(Y_n)$ with steps in $\SigA$.\label{a3:use:walk:R1end} Add the flows $Y_n+c\chi$ for its terms after the first term. The last one is $Y'$. This proves Claim($i$). \item[\textup{(5)}] \emph{The case $\Lam=\Lam_f$.} $|\Lam_f(t)|\le\len(t)\le 3$ for every blue edge $t$ \textup{(v)}, \textup{(i)}. So $|\Lam_f(e_i)|+|\Lam_f(f_i)|\le 6$, and $2\,|\Lam_f(e_i)|+|\Lam_f(f_i)|\le 2\,\len(e_i)+\len(f_i)<|A|$ by \Cone.\label{a3:use:walk:C1f}\qedhere \end{condlist} \end{proof} \subsection*{A change of the \texorpdfstring{$A$}{A}-coordinate inside \texorpdfstring{$H$}{H}} Lemma~\ref{lem:L1} is put together from three parts: the walk lemma, which gives a sequence of steps $\sigma\chi_{C_l}$; Lemma~\ref{a3:lem:step}, which makes one such step in $G$; and Lemma~\ref{a3:lem:prepB}, which prepares the free threads for the step. Lemmas~\ref{a3:lem:prepB} and~\ref{a3:lem:step} do not use \RI, \Cone{} or the walk lemma, and Lemma~\ref{a3:lem:step} holds for every cycle $C$ of $Q$. \begin{lemma}[Preparation in $B$]\label{a3:lem:prepB} Assume \RII, and let the frame have \Cfour. Let $G$ be a multigraph that contains $H$, and $f$ an $H$-good $\Gam$-flow of $G$ with $A$-coordinate $u$. Let $E^{\circ}$ be a set of free edges. For every $\eps$ in $E^{\circ}$ put \[ W_\eps=\{\,-f(e) : e\in E(\thr{\eps}),\ u(e)=0\,\}, \] and let $T_\eps$ be a set of at most $\len(\eps)$ elements of $B$ that contains $W_\eps$. Then $f$ is $\Sig$-joined, by $\Sig$-moves $y\chi_{\Deps}$ with $y$ in $\SigB$ and $\eps$ in $E^{\circ}$, to an $H$-good flow $f^{\circ}$ with $A$-coordinate $u$ such that for every $\eps$ in $E^{\circ}$ \[ f^{\circ}=f+y_\eps\ \text{ on } E(\thr{\eps}),\qquad\text{with an element } y_\eps \text{ of } B \text{ outside } T_\eps . \] $f^{\circ}$ is equal to $f$ outside $E(H)$, and on every free thread $\thr{\eps}$ with $\eps$ not in $E^{\circ}$. \end{lemma} \begin{proof} $W_\eps$ is a subset of $B$ by (E1), and $0$ is not in $W_\eps$, because $f$ is nowhere zero. \begin{condlist}{(2)} \item[\textup{(1)}] \emph{One free edge $\eps$.} If $0$ is not in $T_\eps$, put $y_\eps=0$ and make no move. Let $0$ be in $T_\eps$. By \Cfour, $|T_\eps|\le\len(\eps)<|B|$.\label{a3:use:prepB:C4} So there is $y_\eps$ in $B$ outside $T_\eps$. $0$ is in $T_\eps$ and not in $W_\eps$, so $|W_\eps|\le|T_\eps|-1\le\len(\eps)-1\le 2$; the last step holds because a thread has at most $3$ edges \textup{(i)}. By \RII{} there is a walk $0=z_0,z_1,\dots,z_\nu=y_\eps$ in $B\setminus W_\eps$ with steps in $\SigB$.\label{a3:use:prepB:R2} Make the moves $(z_r-z_{r-1})\chi_{\Deps}$, $r=1,\dots,\nu$. After the $r$-th move the flow is $f+z_r\chi_{\Deps}$. Its $A$-coordinate is $u$ (E2). It is nowhere zero. Indeed, $\Deps$ is a union of threads (Lemma~\ref{a2:flows}~(a)); it contains $\thr{\eps}$ and no other free thread \textup{(iv)}; so its edges outside $\thr{\eps}$ lie on protected threads. An edge $e$ of a protected thread has $u(e)\ne 0$, because $f$ is $H$-good. An edge $e$ of $\thr{\eps}$ with $u(e)=0$ has the value $f(e)+z_r$, because $\Deps$ is directed along $\thr{\eps}$; this is not $0$, because $z_r$ is not in $W_\eps$. The other edges of $\thr{\eps}$ have $u(e)\ne 0$. The edges outside $\Deps$ do not change. With (E1): every move is a $\Sig$-move. The result $f+y_\eps\chi_{\Deps}$ is $H$-good, it is equal to $f+y_\eps$ on $E(\thr{\eps})$, and it is equal to $f$ outside $E(\Deps)$: on every other free thread and outside $E(H)$. \item[\textup{(2)}] \emph{All edges of $E^{\circ}$, one after the other.} Step~(1) for $\eps$ does not change the $A$-coordinate, and it changes no value on a free thread other than $\thr{\eps}$. So when the turn of $\eps$ comes, the flow at hand is $H$-good with $A$-coordinate $u$, it is still equal to $f$ on $\thr{\eps}$, its set $W_\eps$ is the set $W_\eps$ of the statement, and Step~(1) can be made with the same $T_\eps$. Later steps do not change the values on $\thr{\eps}$.\qedhere \end{condlist} \end{proof} \begin{lemma}[One step inside $H$]\label{a3:lem:step} Assume \RII, and let the frame have \Cfour. Let $G$ be a multigraph that contains $H$, and $f$ an $H$-good $\Gam$-flow of $G$ with $A$-coordinate $u$. Let $C$ be a cycle of $Q$ with a direction and $\sigma$ an element of $\SigA$ such that $u'=u+\pi(\sigma\chi_C)$ is non-zero on every edge of every protected thread. Then $f$ is $\Sig$-joined, by $\Sig$-moves on cycles of $H$, to an $H$-good flow with $A$-coordinate $u'$ that is equal to $f$ outside $E(H)$. \end{lemma} \begin{proof} Put $\gamma=\lift(\sigma)$; so $\gamma$ is in $\Sig$ and $\quo{\gamma}=\sigma$. By Lemma~\ref{a2:flows}~(d) and~(a), \[ u'=u+\sigma\,(\chi_{O_1}+\dots+\chi_{O_\nu})\quad\text{on } E(H), \] with pairwise edge-disjoint cycles $O_1,\dots,O_\nu$ of $H$; every $O_j$ is a union of threads, and $\chi_{O_j}$ is constant, $1$ or $-1$, on each of its threads. A thread lies on at most one $O_j$. For a free edge $\eps$ put \[ \delta_\eps=\gamma\cdot(\text{the value of }\chi_{O_j}\text{ on }\thr{\eps})\ \text{ if } \thr{\eps} \text{ lies on } O_j,\qquad \delta_\eps=0\ \text{ if } \thr{\eps} \text{ lies on no } O_j. \] So $\delta_\eps$ is an element of $\Gam$ with \[ \quo{\delta_\eps}=u'(e)-u(e)\quad\text{for every edge } e \text{ of } \thr{\eps}. \tag{$**$} \] \begin{condlist}{(ii)} \item[\textup{(i)}] \emph{Preparation in $B$.} For every free edge $\eps$ let $W_\eps$ be as in Lemma~\ref{a3:lem:prepB}, and \[ T_\eps=W_\eps\cup\{\,-f(e)-\delta_\eps : e\in E(\thr{\eps}),\ u'(e)=0\,\}. \] $T_\eps$ is a subset of $B$: if $u'(e)=0$, then $\quo{f(e)+\delta_\eps}=u(e)+u'(e)-u(e)=0$ by $(**)$. Every edge $e$ of $\thr{\eps}$ gives at most one element of $T_\eps$. Indeed, let $u(e)=0$ and $u'(e)=0$. Then $\quo{\delta_\eps}=0$ by $(**)$. If $\thr{\eps}$ were on some $O_j$, then $\quo{\delta_\eps}$ would be $\sigma$ or $-\sigma$, which is not $0$. So $\thr{\eps}$ lies on no $O_j$, $\delta_\eps=0$, and the two elements $-f(e)$ and $-f(e)-\delta_\eps$ are equal. Hence $|T_\eps|\le\len(\eps)$. Lemma~\ref{a3:lem:prepB}, with $E^{\circ}$ the set of all free edges, gives an $H$-good flow $f^{\circ}$ with $A$-coordinate $u$, equal to $f$ outside $E(H)$, with $f^{\circ}=f+y_\eps$ on $E(\thr{\eps})$ and $y_\eps\notin T_\eps$.\label{a3:use:step:prepB} So for every edge $e$ of a free thread $\thr{\eps}$: \[ \textup{(T)}\qquad \text{if } u'(e)=0,\ \text{ then } f^{\circ}(e)+\delta_\eps\ne 0 . \] \textup{(}$-f(e)-\delta_\eps$ is in $T_\eps$, so it is not $y_\eps$.\textup{)} \item[\textup{(ii)}] \emph{The step.} Make the moves $\gamma\chi_{O_1},\dots,\gamma\chi_{O_\nu}$, in this order. After the first $j$ of them an edge $e$ of $H$ carries the value $f^{\circ}(e)+\gamma\chi_{O}(e)$, with the $A$-coordinate $u(e)+\sigma\chi_O(e)=u'(e)$, if $e$ lies on a cycle $O$ among $O_1,\dots,O_j$; and it carries the value $f^{\circ}(e)$, with the $A$-coordinate $u(e)$, if not. These values are not $0$: \begin{condlist}{--} \item[--] on a protected thread, $u(e)\ne 0$ because $f$ is $H$-good, and $u'(e)\ne 0$ by hypothesis; use (E1); \item[--] on a free thread $\thr{\eps}$ that lies on one of $O_1,\dots,O_j$, the value is $f^{\circ}(e)+\delta_\eps$; it is not $0$ by (E1) if $u'(e)\ne 0$, and by \textup{(T)} if $u'(e)=0$; \item[--] on every other free thread the value is $f^{\circ}(e)$, and $f^{\circ}$ is nowhere zero. \end{condlist} The edges outside $E(H)$ do not change. So all these moves are $\Sig$-moves on cycles of $H$. After the last one every edge $e$ of $H$ has the $A$-coordinate $u'(e)$; for an edge on no $O_j$ note that $u'(e)=u(e)$. The result is nowhere zero and its $A$-coordinate $u'$ is non-zero on the protected threads: it is $H$-good. It is equal to $f$ outside $E(H)$.\qedhere \end{condlist} \end{proof} \begin{lemma}[A change of the $A$-coordinate inside $H$]\label{lem:L1} Assume \RI{} and \RII, and let the frame have \Cone{} and \Cfour. Let $G$ be a multigraph that contains $H$, $f$ an $H$-good $\Gam$-flow of $G$ with $A$-coordinate $u$, and $Y$ a flow in $\PL{\Lam_f}$. Then $f$ is $\Sig$-joined, by $\Sig$-moves on cycles of $H$, to an $H$-good flow with $A$-coordinate $u+\pi(Y)$ that is equal to $f$ outside $E(H)$. \end{lemma} \begin{proof} $0$ and $Y$ are in $\PL{\Lam_f}$ \textup{(v)}. The frame has \Cone, so the walk lemma (Lemma~\ref{lem:walk}, its last sentence) gives a sequence $0=Y_0,Y_1,\dots,Y_n=Y$ in $\PL{\Lam_f}$ with $Y_r-Y_{r-1}=\sigma_r\chi_{C_{l_r}}$, $\sigma_r$ in $\SigA$.\label{a3:use:L1:walk} We build $H$-good flows $f=f_0,f_1,\dots,f_n$ of $G$, each $\Sig$-joined to the one before it by $\Sig$-moves on cycles of $H$ and equal to $f$ outside $E(H)$, such that $f_r$ has the $A$-coordinate $u_r=u+\pi(Y_r)$. Let $f_{r-1}$ be built. Then \[ u_{r-1}+\pi(\sigma_r\chi_{C_{l_r}})=u+\pi(Y_{r-1})+\pi(Y_r-Y_{r-1})=u_r , \] and $u_r$ is non-zero on every edge of every protected thread, because $Y_r$ is in $\PL{\Lam_f}$ \textup{(v)}. So Lemma~\ref{a3:lem:step}, for the flow $f_{r-1}$, the cycle $C_{l_r}$ and $\sigma_r$, gives $f_r$; its hypotheses \RII{} and \Cfour{} are assumed.\label{a3:use:L1:step} The flow $f_n$ has the $A$-coordinate $u+\pi(Y)$. \end{proof} \noindent In the rest of this section we write \condLone{} for the conclusion of Lemma~\ref{lem:L1}, for a given subdivision $H$ with a given frame: \emph{for every multigraph $G$ that contains $H$, every $H$-good $\Gam$-flow $f$ of $G$ with $A$-coordinate $u$, and every $Y$ in $\PL{\Lam_f}$, the flow $f$ is $\Sig$-joined, by $\Sig$-moves on cycles of $H$, to an $H$-good flow with $A$-coordinate $u+\pi(Y)$ that is equal to $f$ outside $E(H)$.} By Lemma~\ref{lem:L1}, \condLone{} holds if \RI{} and \RII{} hold and the frame has \Cone{} and \Cfour. \subsection*{The path} \begin{lemma}[The path]\label{lem:P} Let $p\ne p'$ be vertices of $H$. There is a path $J$ of $H$ from $p$ to $p'$ such that, for every node $i$, at most $\max\{\len(e_i),\len(f_i)\}$ edges of $J$ lie on the threads of the edges of $N_i$ \textup{(}both threads together; so at most $3$\textup{)}. $J$ can contain any number of free threads. \end{lemma} \begin{proof} An expansion edge has no thread. So ``the threads of the edges of $N_i$'' are the threads of the original edges of $N_i$; there are two, one or none of them. \begin{condlist}{(4)} \item[\textup{(1)}] \emph{The subgraph $H_g$.} Let $g=(g_1,\dots,g_s)$ be a choice of one edge $g_i$ in every pair $N_i$. Let $H_g$ be the subgraph of $H$ formed by the vertices of $H_0$ and the threads of all original edges of the trees $\Tg$ of all components $Q_b$. Then: \begin{condlist}{(1c)} \item[\textup{(1a)}] $H_g$ is connected. Reason: $\Tg$ is a spanning tree of $Q_b$ \textup{(iv)}. Under the contraction of the expansion edges \textup{(ii)} it becomes a connected subgraph of $K_b$ that contains all vertices of $K_b$; its edges are the original edges of $\Tg$. Any two vertices of $H_0$ are joined by a path of $H_0$; the two ends of an edge of this path lie in one block $K_b$ \textup{(i)}, and they are joined by original edges of $\Tg$. So the original edges of the trees $\Tg$ form a connected subgraph of $H_0$ on all vertices of $H_0$. Replace these edges by their threads. \item[\textup{(1b)}] A vertex of $H$ is outside $H_g$ only if it is an interior vertex of the thread of an original edge that is in some pair $N_i$ and is not $g_i$. Reason: $\Tg$ contains all red edges and all edges $g_i$, and every blue edge is in a pair \textup{(iii)}. \item[\textup{(1c)}] A path of $H_g$ has, on the threads of the edges of $N_i$, only edges of the thread of $g_i$: at most $\len(g_i)\le\max\{\len(e_i),\len(f_i)\}$ edges. Reason: $g_i$ is the only edge of $N_i$ in $\Tg$ \textup{(iv)}. \end{condlist} \item[\textup{(2)}] \emph{Case 1: $p$ and $p'$ are interior vertices of the same thread $\thr{t}$.} Let $J$ be the part of $\thr{t}$ between them. A thread has at most $3$ edges \textup{(i)}, so at most $2$ interior vertices; hence $J$ has one edge. If $t$ is in a pair $N_i$, then $J$ has $1\le\len(t)\le\max\{\len(e_i),\len(f_i)\}$ edge on the threads of $N_i$. $J$ has no edge on the threads of the other pairs. \item[\textup{(3)}] \emph{Case 2: not Case 1, and there is no pair $N_i$ such that $p$ is an interior vertex of the thread of one edge of $N_i$ and $p'$ an interior vertex of the thread of the other edge.} A vertex is an interior vertex of at most one thread. So, by the case, every pair $N_i$ has at most one edge whose thread has $p$ or $p'$ as an interior vertex. If there is such an edge, let it be $g_i$; if not, put $g_i=e_i$. By \textup{(1b)}, $p$ and $p'$ are in $H_g$. By \textup{(1a)} there is a path $J$ of $H_g$ from $p$ to $p'$. By \textup{(1c)} it is as stated. \item[\textup{(4)}] \emph{Case 3: $p$ is an interior vertex of $\thr{t}$ and $p'$ an interior vertex of $\thr{t'}$, where $t\ne t'$ are the two edges of one pair $N_i$.} Let $U'$, $U''$ be the two operands of the node $i$. By \textup{(iii)}, $t$ has an end $x_1$ in $V_{U'}$ and an end in $V_{U''}$, and $t'$ has an end $x_2$ in $V_{U'}$ and an end in $V_{U''}$. So each of the threads $\thr{t}$, $\thr{t'}$ has one end \emph{on the side of $U'$} \textup{(}the vertex of $H_0$ that comes from $x_1$, from $x_2$\textup{)} and one end on the side of $U''$. Let $c$ be the number of edges of $\thr{t}$ between $p$ and the end on the side of $U'$, and $c'$ the same for $\thr{t'}$ and $p'$. The numbers of edges to the ends on the side of $U''$ are $\len(t)-c$ and $\len(t')-c'$. The two sums $c+c'$ and $(\len(t)-c)+(\len(t')-c')$ add up to $\len(t)+\len(t')$. So one of them is at most $(\len(t)+\len(t'))/2\le\max\{\len(t),\len(t')\}$. Let it be $c+c'$; the other case is the same with $U'$ and $U''$ exchanged. By \textup{(iv)} the edges of $\Te$ in $Q_{U'}$ form a spanning tree of $Q_{U'}$. It contains a path $R$ from $x_1$ to $x_2$. The edges $t$, $t'$ are not edges of $Q_{U'}$: each has an end outside $V_{U'}$. So $R$ contains neither $t$ nor $t'$, and every blue edge of $R$ is an edge $e_l$ with $l\ne i$. Contract the expansion edges and replace the original edges of $R$ by their threads. The result $L$ is a connected subgraph of $H$ \textup{(}one vertex, if $R$ has no original edge\textup{)}. It contains the ends of $\thr{t}$ and of $\thr{t'}$ on the side of $U'$. It has no edge on $\thr{t}$ or $\thr{t'}$. For a pair $N_l$ with $l\ne i$ it has edges on the thread of $e_l$ only. Let $J$ be a path from $p$ to $p'$ in the union of $L$, the part of $\thr{t}$ from $p$ to its end on the side of $U'$, and the part of $\thr{t'}$ from $p'$ to its end on the side of $U'$. This union is connected, so $J$ exists. On the threads of $N_i$ the path $J$ has at most $c+c'\le\max\{\len(e_i),\len(f_i)\}$ edges. On the threads of a pair $N_l$, $l\ne i$, it has at most $\len(e_l)\le\max\{\len(e_l),\len(f_l)\}$ edges. \end{condlist} The three cases cover all pairs $p\ne p'$. The bound ``at most $3$'' holds because $\len(t)\le 3$ \textup{(i)}. \end{proof} \subsection*{The lift of one move of the contracted graph} Now $H$ is a cycle or a subdivision. If $H$ is a cycle of a multigraph $G$, a nowhere-zero $\Gam$-flow $f$ of $G$ is \emph{$H$-good} if $\quo{f}(e)\ne 0$ for every edge $e$ of $H$ (Definition~\ref{def:lists}). \begin{lemma}[The lift of one move of the contracted graph]\label{lem:L2} Assume $|A|\ge 8$, \RI{} and \RII. Let $G$ be a bridgeless multigraph and $H$ a subgraph of $G$ that is a cycle with at most $5$ edges \textup{(}case \textup{CYC)}, or a subdivision with a counted frame \textup{(}case \textup{SUB)}. Let $G^*=G/H$ be the multigraph obtained from $G$ by the identification of $V(H)$ to one vertex and the deletion of $E(H)$. Let $f$ be an $H$-good $\Gam$-flow of $G$ with restriction $f^*$ to $G^*$, and let $g^*=f^*+\lambda\chi_{C^*}$ be the result of a $\Sig$-move of $G^*$. Then $f$ is $\Sig$-joined to an $H$-good $\Gam$-flow $g$ of $G$ whose restriction to $G^*$ is $g^*$. \textup{(}The frame form: the same holds with \RIm{} in the place of \RI{} and, in the place of \Cone, the conclusion of Lemma~\textup{\ref{lem:L1}} for this $H$ and all $f$, $Y$.\textup{)} \end{lemma} \begin{proof} We prove the frame form. It contains the first form: \RI{} implies \RIm, and for a subdivision with a counted frame \condLone{} holds by Lemma~\ref{lem:L1}. So we assume: $|A|\ge 8$; \RIm; \RII; and in the case SUB the frame has \Ctwo, \Cthree, \Cfour, and \condLone{} holds. Let $u$ be the $A$-coordinate of $f$. \begin{condlist}{(5)} \item[\textup{(1)}] \emph{The contracted graph.} $E(G^*)=E(G)\setminus E(H)$. Let $v^*$ be the vertex of $G^*$ made from $V(H)$. By (E5) the restriction of a flow of $G$ to $E(G^*)$ is a flow of $G^*$. $C^*$ is a cycle of $G^*$ with a direction, $\lambda$ is in $\Sig$, and $f^*$ and $g^*$ are nowhere zero. \item[\textup{(2)}] \emph{The cycle $C$.} If $C^*$ does not contain $v^*$, its edges form a cycle of $G$. If $C^*$ contains $v^*$, its edges form in $G$ a path between two vertices $p$, $p'$ of $H$ whose other vertices are outside $V(H)$, or, if $p=p'$, a cycle of $G$. \textup{(}$C^*$ is a loop at $v^*$, or it has exactly two edges at $v^*$.\textup{)} If the edges of $C^*$ form a cycle of $G$, direct it as $C^*$ and put $g=f+\lambda\chi_{C^*}$. This flow is equal to $g^*$ on $E(G^*)$ and to $f$ on $E(H)$. So it is nowhere zero, the move is a $\Sig$-move, $g$ is $H$-good, and its restriction is $g^*$. From now on the edges of $C^*$ form a path between $p\ne p'$. For every path $J$ of $H$ from $p$ to $p'$ the union $C=C^*\cup J$ is a cycle of $G$, because all vertices of $J$ are in $V(H)$. Direct $C$ so that $\chi_C=\chi_{C^*}$ on $E(G^*)$. The path $J$ is chosen below. In each case we reach an $H$-good flow $f'$ that is equal to $f$ outside $E(H)$, and then make the move $g=f'+\lambda\chi_C$. On $E(G^*)$ the flow $g$ is equal to $f^*+\lambda\chi_{C^*}=g^*$; so it is not $0$ there, and the restriction of $g$ is $g^*$. It remains to look at the edges of $H$. \item[\textup{(3)}] \emph{Case \textup{CYC}.} Direct the cycle $H$. Let $J$ be an arc of $H$ between $p$ and $p'$ with at most $2$ edges; it exists because $H$ has at most $5$ edges. $\quo{\lambda}=0$. Put $f'=f$. By (E2) the move $\lambda\chi_C$ does not change the $A$-coordinate, which is non-zero on $E(H)$. So $g$ is nowhere zero (E1) and $H$-good. $\quo{\lambda}\ne 0$. Put \begin{align*} X &= \{\,c\in A : u+c\chi_H \text{ has a zero on } E(H)\,\},\\ X' &= \{\,c\in A : u+c\chi_H+\quo{\lambda}\chi_C \text{ has a zero on } E(J)\,\}. \end{align*} Every edge of $H$ excludes one value of $c$, because $\chi_H$ is $1$ or $-1$ on it. So $|X|\le 5$, $|X'|\le 2$, and $0$ is not in $X$, because $f$ is $H$-good. By the hypothesis $|A|\ge 8$ \textup{(}it is \Czero{} of Theorem~\ref{thm:master}\textup{)}, $|X\cup X'|\le 5+2<|A|$; choose $c$ outside $X\cup X'$.\label{a3:use:L2:C0} By \RIm{} there is a walk $0=c_0,c_1,\dots,c_\nu=c$ in $A\setminus X$ with steps $d_r=c_r-c_{r-1}$ in $\SigA$.\label{a3:use:L2:R1m} Make the moves $\lift(d_r)\chi_H$, $r=1,\dots,\nu$. After the $r$-th move the $A$-coordinate is $u+c_r\chi_H$; it is non-zero on $E(H)$, because $c_r$ is not in $X$; outside $E(H)$ nothing changes. So these are $\Sig$-moves (E1). The result $f'$ is $H$-good and equal to $f$ outside $E(H)$. Now $g=f'+\lambda\chi_C$ has, on $E(H)$, the $A$-coordinate $u+c\chi_H$ outside $J$ and $u+c\chi_H+\quo{\lambda}\chi_C$ on $J$. Both are non-zero, because $c$ is not in $X$ and not in $X'$. So $g$ is nowhere zero and $H$-good. \item[\textup{(4)}] \emph{Case \textup{SUB}, $\quo{\lambda}\ne 0$.} Let $J$ be the path of Lemma~\ref{lem:P}, and let $j_i$ be the number of edges of $J$ on the threads of the edges of $N_i$. So $j_i\le\max\{\len(e_i),\len(f_i)\}$. \begin{condlist}{(4c)} \item[\textup{(4a)}] \emph{The protected threads.} For an original blue edge $t$ put \[ \Lam_1(t)=\Lam_f(t)\cup\{\,-u(e)-\quo{\lambda}\chi_C(e) : e\in E(\thr{t})\cap E(J)\,\}, \] and let $\Lam_1$ be empty on the expansion edges. Then, for every node $i$, \[ |\Lam_1(e_i)|+|\Lam_1(f_i)|\ \le\ \len(e_i)+\len(f_i)+j_i\ \le\ \len(e_i)+\len(f_i)+\max\{\len(e_i),\len(f_i)\}\ <\ |A| \] by \textup{(v)} and \Ctwo.\label{a3:use:L2:C2} By Lemma~\ref{lem:reverse}~(a), with $Y_0=0$, there is $Y$ in $\PL{\Lam_1}$. It is in $\PL{\Lam_f}$, because $\Lam_1$ contains $\Lam_f$. By \condLone{}, $f$ is $\Sig$-joined to an $H$-good flow $f_1$ with $A$-coordinate $u_1=u+\pi(Y)$ that is equal to $f$ outside $E(H)$.\label{a3:use:L2:L1a} For an edge $e$ of $J$ on a protected thread $\thr{t}$: $u_1(e)+\quo{\lambda}\chi_C(e)=u(e)+Y(t)+\quo{\lambda}\chi_C(e)\ne 0$, because $Y(t)$ is not in $\Lam_1(t)$. \item[\textup{(4b)}] \emph{The free threads.} Let $E^{\circ}$ be the set of the free edges $\eps$ such that $\thr{\eps}$ has an edge on $J$. Their number is not bounded. For $\eps$ in $E^{\circ}$ put \begin{align*} W_\eps &= \{\,-f_1(e) : e\in E(\thr{\eps}),\ u_1(e)=0\,\},\\ T_\eps &= W_\eps\cup\{\,-f_1(e)-\lambda\chi_C(e) : e\in E(\thr{\eps})\cap E(J),\ u_1(e)+\quo{\lambda}\chi_C(e)=0\,\}. \end{align*} $T_\eps$ is a subset of $B$: the $A$-coordinate of $f_1(e)+\lambda\chi_C(e)$ is $u_1(e)+\quo{\lambda}\chi_C(e)$. Every edge $e$ of $\thr{\eps}$ gives at most one element of $T_\eps$: for an edge of $J$, $\chi_C(e)$ is $1$ or $-1$ and $\quo{\lambda}\ne 0$, so $u_1(e)=0$ and $u_1(e)+\quo{\lambda}\chi_C(e)=0$ exclude each other. Hence $|T_\eps|\le\len(\eps)$. By Lemma~\ref{a3:lem:prepB} \textup{(}it uses \RII{} and \Cfour\textup{)}, $f_1$ is $\Sig$-joined to an $H$-good flow $f_2$ with $A$-coordinate $u_1$, equal to $f_1$ outside $E(H)$, with $f_2=f_1+y_\eps$ on $E(\thr{\eps})$ and $y_\eps\notin T_\eps$, for every $\eps$ in $E^{\circ}$.\label{a3:use:L2:prepB} \item[\textup{(4c)}] \emph{The move.} Put $f'=f_2$ and $g=f_2+\lambda\chi_C$. On an edge $e$ of $J$ the new $A$-coordinate is $u_1(e)+\quo{\lambda}\chi_C(e)$. If $e$ is on a protected thread, this is not $0$ by \textup{(4a)}. If $e$ is on a free thread $\thr{\eps}$ and this is $0$, then $-f_1(e)-\lambda\chi_C(e)$ is in $T_\eps$, so it is not $y_\eps$, and $g(e)=f_1(e)+y_\eps+\lambda\chi_C(e)\ne 0$. The edges of $H$ outside $J$ do not change. So $g$ is nowhere zero (E1), the move is a $\Sig$-move, and $g$ is $H$-good: on the protected threads its $A$-coordinate is $u_1$ outside $J$ and $u_1+\quo{\lambda}\chi_C$ on $J$. \end{condlist} \item[\textup{(5)}] \emph{Case \textup{SUB}, $\quo{\lambda}=0$.} So $\lambda$ is in $\SigB$. \begin{condlist}{(5c)} \item[\textup{(5a)}] \emph{The path $J$.} A thread has at most $3$ edges \textup{(i)}. So every interior vertex of a thread is next to an end of the thread, and two interior vertices of one thread are next to each other. If $p$ and $p'$ are interior vertices of the same free thread, $J$ is the edge of this thread between them. Otherwise: if $p$ is an interior vertex of a free thread, let $q$ be an end of this thread that is next to $p$, and let $J$ start with the edge from $p$ to $q$; if not, put $q=p$. In the same way $p'$ gives $q'$, and $J$ ends with the edge from $q'$ to $p'$ if $q'\ne p'$. By (O6), $q$ and $q'$ are in $\HP$, and $\HP$ is connected; between $q$ and $q'$ the path $J$ runs inside $\HP$. \textup{(}An interior vertex of a free thread is not in $\HP$; so $J$ is a path.\textup{)} A \emph{free piece} of $J$ is the part of $J$ on one free thread. Hence: $J$ has at most two free pieces; they lie on different free threads; \emph{each has exactly one edge}; and all other edges of $J$ are on protected threads. \item[\textup{(5b)}] \emph{The $A$-coordinate on $J$.} Let $E'$ be the set of the free edges $\eps$ such that $\thr{\eps}$ has a free piece of $J$; so $|E'|\le 2$. For $\eps$ in $E'$ let $e_\eps$ be the edge of this free piece, and $\Phi(\eps)=\{-u(e_\eps)\}$: one element, not all of $A$ by the hypothesis $|A|\ge 8$ \textup{(}\Czero\textup{)}.\label{a3:use:L2:C0phi} For every node $i$, \[ |\Lam_f(e_i)|+|\Lam_f(f_i)|+\sum_{\eps\in E',\ i(\eps)=i}|\Phi(\eps)|\ \le\ \len(e_i)+\len(f_i)+2\ <\ |A| \] by \textup{(v)} and \Cthree.\label{a3:use:L2:C3} By Lemma~\ref{lem:reverse}~(b) there is $Y$ in $\PL{\Lam_f}$ with $Y(\eps)\ne -u(e_\eps)$ for every $\eps$ in $E'$. By \condLone{}, $f$ is $\Sig$-joined to an $H$-good flow $f_1$ with $A$-coordinate $u_1=u+\pi(Y)$ that is equal to $f$ outside $E(H)$.\label{a3:use:L2:L1b} $u_1$ is non-zero on the protected threads, and on a free piece $u_1(e_\eps)=u(e_\eps)+Y(\eps)\ne 0$. So $u_1$ is non-zero on every edge of $J$. \item[\textup{(5c)}] \emph{The move.} Put $f'=f_1$ and $g=f_1+\lambda\chi_C$. By (E2) the $A$-coordinate does not change; it is non-zero on $J$. So $g$ is nowhere zero (E1) and $H$-good. The group $B$ and the condition \RII{} are not used in this case, except inside \condLone.\qedhere \end{condlist} \end{condlist} \end{proof} \subsection*{The master theorem} \begin{theorem}[The master theorem]\label{thm:master} Let $\Gam$ be a finite abelian group, $B$ a subgroup, $A=\Gam/B$, and $\Sig$ a step system. Assume: \begin{condlist}{(H2)} \item[\Hone] the step conditions \RI, \RII, \RIII, and the averaging condition \AVpp; \item[\Htwo] the counts: \begin{condlist}{(C0)} \item[\Czero] $|A|\ge 8$; \item[] and in \AVpp{} the good subgraph can always be chosen to be a cycle with at most $5$ edges, or a subdivision that has a counted frame, that is a frame with \Cone, \Ctwo, \Cthree, \Cfour. \end{condlist} \end{condlist} Then on every connected bridgeless multigraph $G$ \textup{(}loops allowed\textup{)} every nowhere-zero $\Gam$-flow is $\Sig$-joined to every $\Gam$-flow whose $A$-coordinate is nowhere zero. Hence any two nowhere-zero $\Gam$-flows of a bridgeless multigraph are $\Sig$-joined, and $\Fl(G,\Gam)$ is connected for every finite bridgeless multigraph $G$. \end{theorem} \begin{theorem}[The master theorem, frame form]\label{thm:masterframe} The conclusion of Theorem~\textup{\ref{thm:master}} holds also if, in \Hone, \RI{} is replaced by \RIm, and, in \Htwo, \Cone{} is replaced by \begin{condlist}{(L1)} \item[\condLone] the conclusion of Lemma~\textup{\ref{lem:L1}} holds for the subdivisions $H$ and the frames of \Htwo, for every multigraph $G$ that contains $H$. \end{condlist} \textup{(}\RI{} for six deleted elements and \Cone{} are used only in the walk lemma, and the walk lemma is used only in Lemma~\textup{\ref{lem:L1}}.\textup{)} \end{theorem} \begin{proof}[Proof of Theorems~\ref{thm:master} and~\ref{thm:masterframe}] The hypotheses of Theorem~\ref{thm:master} imply those of Theorem~\ref{thm:masterframe}: \RI{} implies \RIm, and \condLone{} holds by Lemma~\ref{lem:L1}, which uses \RI, \RII, \Cone{} and \Cfour.\label{a3:use:M:L1} So let the hypotheses of the frame form hold: \RIm, \RII, \RIII, \AVpp, \Czero, and the good subgraph in \AVpp{} can be chosen \emph{admissible}: a cycle with at most $5$ edges, or a subdivision with a frame that has \Ctwo, \Cthree, \Cfour{} and \condLone. For an admissible $H$ Lemma~\ref{lem:L2} holds in its frame form. A \emph{target} of a multigraph is a $\Gam$-flow whose $A$-coordinate is nowhere zero. A target is nowhere zero (E1). \smallskip \noindent\textsc{Claim.} On every connected bridgeless multigraph $G$, every nowhere-zero $\Gam$-flow is $\Sig$-joined to every target. \begin{condlist}{(4)} \item[\textup{(1)}] \emph{The claim gives the theorem.} The first statement is the claim. Let $G$ be a bridgeless multigraph. It has a nowhere-zero $A$-flow, because $|A|\ge 6$ by \Czero\label{a3:use:M:C0target} \textup{(}Theorem~\ref{q:flows}, the theorems of Seymour and of Tutte: every bridgeless multigraph, loops allowed, has a nowhere-zero flow with values in any abelian group of order at least $6$\textup{)}. By (E4) it is the $A$-coordinate of a $\Gam$-flow $h$: a target. A cycle of $G$ lies in one component of $G$, and a $\Sig$-move of a component is a $\Sig$-move of $G$. So, component after component, the claim joins every nowhere-zero $\Gam$-flow of $G$ to $h$. The reverse of a $\Sig$-move is a $\Sig$-move; so any two nowhere-zero $\Gam$-flows are $\Sig$-joined through $h$. $\Sig$-moves are moves of $\Fl(G,\Gam)$; so $\Fl(G,\Gam)$ is connected. \item[\textup{(2)}] \emph{The induction.} We prove the claim by induction on the number $n$ of vertices of $G$; for each $n$ first for the loopless graphs, then for the graphs with loops. A connected loopless multigraph with one vertex has no edge and only the empty flow. A connected bridgeless loopless multigraph with $n\ge 2$ has a cut with exactly $2$ edges or is $3$-edge-connected. So three cases remain: \textup{(a)}, \textup{(b)}, \textup{(c)}. In all of them $f$ is a nowhere-zero $\Gam$-flow and $h$ a target of $G$. \item[\textup{(a)}] \emph{$G$ has a loop.} Let $G^-$ be $G$ without its loops. It is connected, bridgeless and loopless, and it has $n$ vertices; so the claim holds for $G^-$ \textup{(}the loopless graphs with $n$ vertices come first\textup{)}. The restrictions of $f$ and $h$ to $G^-$ are a nowhere-zero flow and a target. A $\Sig$-move of $G^-$ is a $\Sig$-move of $G$: the values on the loops do not change. So $f$ is $\Sig$-joined to the flow that is equal to $h$ on $E(G^-)$ and to $f$ on the loops. Now let $e$ be a loop, with the values $x=f(e)$ and $x'=h(e)$. Both are non-zero; $x'$ because $\quo{x'}\ne 0$. By \RIII{} there is a walk from $x$ to $x'$ in $\Gam\setminus\{0\}$ with steps in $\Sig$.\label{a3:use:M:R3} A loop is a cycle with one edge. The moves $\lambda\chi_e$, for the steps $\lambda$ of the walk, are $\Sig$-moves: the value on $e$ is never $0$, and no other edge changes. Do this for every loop. The result is $h$. \item[\textup{(b)}] \emph{$G$ is loopless and has a cut with exactly two edges $e_1$, $e_2$.} Let $G/e_1$ be obtained from $G$ by the identification of the two ends of $e_1$ and the deletion of $e_1$. It is connected, it has $n-1$ vertices, and it is bridgeless \textup{(}a cycle of $G$ through an edge $b\ne e_1$ becomes a closed trail of $G/e_1$ through $b$, and a closed trail through $b$ contains a cycle through $b$\textup{)}; it can have loops. By induction the claim holds for $G/e_1$. By (E5) the restriction of a flow of $G$ to $E(G/e_1)$ is a flow of $G/e_1$. Every flow of $G/e_1$ is the restriction of exactly one flow of $G$: the value on $e_1$ is fixed by conservation at one end of $e_1$, and conservation at the other end then follows from conservation at the vertex of $G/e_1$ made from the two ends. In a flow of $G$ the value on $e_1$ is the value on $e_2$ or its negative \textup{(}conservation over one side of the cut\textup{)}. So a flow of $G$ is nowhere zero if and only if its restriction is, and it is a target if and only if its restriction is. Let $g'\to g'+\lambda\chi_{C'}$ be a $\Sig$-move of $G/e_1$. The edges of $C'$ form a cycle of $G$, or a path of $G$ between the two ends of $e_1$; in the second case add $e_1$. This gives a cycle $C$ of $G$ with $\chi_C=\chi_{C'}$ on $E(G/e_1)$. If $g$ is the flow of $G$ with restriction $g'$, then $g+\lambda\chi_C$ is the flow of $G$ with restriction $g'+\lambda\chi_{C'}$. Both are nowhere zero. So every $\Sig$-move of $G/e_1$ lifts to a $\Sig$-move of $G$, and the claim for $G/e_1$ gives the claim for $G$. \item[\textup{(c)}] \emph{$G$ is loopless and $3$-edge-connected, $n\ge 2$.} By \AVpp{} and \Htwo, $f$ is $\Sig$-joined to a nowhere-zero flow $f'$ with $A$-coordinate $u'$ such that $(V(G),\supp(u'))$ contains an admissible good subgraph $H$.\label{a3:use:M:AVpp} If $H$ is a subdivision, fix a frame of $H$ as in \Htwo. $u'$ is non-zero on $E(H)$, so $f'$ is $H$-good. $H$ has at least $2$ vertices (Definition~\ref{def:good}). So $G^*=G/H$ has fewer vertices than $G$. It is connected. It is bridgeless: the edges outside $E(H)$ of a cycle of $G$ through an edge $b$ of $G^*$ form a closed trail of $G^*$ through $b$, and it contains a cycle through $b$. It can have loops. By induction the claim holds for $G^*$. The restriction of $f'$ to $G^*$ is nowhere zero, and the restriction $h^*$ of $h$ is a target of $G^*$. So there are $\Sig$-moves of $G^*$ that lead from the restriction of $f'$ through flows $g_0,g_1,\dots,g_\nu=h^*$. Lemma~\ref{lem:L2}, used $\nu$ times, gives $H$-good flows $f'=f_0,f_1,\dots,f_\nu$ of $G$, each $\Sig$-joined to the one before it, such that the restriction of $f_r$ to $G^*$ is $g_r$.\label{a3:use:M:L2} So $f_\nu$ is $H$-good and equal to $h$ outside $E(H)$. Let $u_\nu$ be its $A$-coordinate. \emph{End, case \textup{CYC}.} Direct the cycle $H$. The $A$-flow $\quo{h}-u_\nu$ is $0$ outside $E(H)$, so it is $c\chi_H$ with $c$ in $A$. The set $X=\{\,c'\in A : u_\nu+c'\chi_H \text{ has a zero on } E(H)\,\}$ has at most $5$ elements. $0$ is not in $X$, because $f_\nu$ is $H$-good; $c$ is not in $X$, because $\quo{h}$ is nowhere zero. By \RIm{} there is a walk from $0$ to $c$ in $A\setminus X$ with steps $d_r$ in $\SigA$.\label{a3:use:M:R1m} The moves $\lift(d_r)\chi_H$ are $\Sig$-moves: after each of them the $A$-coordinate is non-zero on $E(H)$, and nothing changes outside $E(H)$. The result $f''$ has the $A$-coordinate $\quo{h}$ and is equal to $h$ outside $E(H)$. So $h-f''=y\chi_H$ with $\quo{y}=0$, that is with $y$ in $B$. By \RII{} with $X$ empty, $y$ is a sum of elements of $\SigB$.\label{a3:use:M:R2cyc} The moves on $H$ with these elements are $\Sig$-moves: the $A$-coordinate stays $\quo{h}$ (E2), which is nowhere zero (E1). The result is $h$. \emph{End, case \textup{SUB}.} The $A$-flow $\quo{h}-u_\nu$ is $0$ outside $E(H)$, so it is an $A$-flow of $H$. By Lemma~\ref{a2:flows}~(c) there is an $A$-flow $Y$ of $Q$ with $\pi(Y)=\quo{h}-u_\nu$. $Y$ is in $\PL{\Lam_{f_\nu}}$, because $u_\nu+\pi(Y)=\quo{h}$ has no zero \textup{(v)}. By \condLone{}, $f_\nu$ is $\Sig$-joined to a flow $f''$ with $A$-coordinate $\quo{h}$ that is equal to $f_\nu$, hence to $h$, outside $E(H)$.\label{a3:use:M:L1end} So $h-f''$ is a flow of $H$ with values in $B$. By (E3) it is a sum of flows $y\chi_O$ with $O$ a cycle of $H$ and $y$ in $B$, and by \RII{} with $X$ empty every $y$ is a sum of elements of $\SigB$.\label{a3:use:M:R2sub} Make these moves one by one. They are $\Sig$-moves: the $A$-coordinate stays $\quo{h}$ (E2), which is nowhere zero (E1). The result is $h$.\qedhere \end{condlist} \end{proof} \begin{remark}[Where the deletions are used]\label{rem:deletions} \RI: the walk lemma deletes at most $\len(e_i)+\len(f_i)\le 6$ elements of $A$; the case CYC of Lemma~\ref{lem:L2} and the end of the proof of Theorem~\ref{thm:master} in the case CYC delete at most $5$. \RII: the preparation in $B$ \textup{(}Lemma~\ref{lem:L1}, and Lemma~\ref{lem:L2} for a step with non-zero $A$-part\textup{)} deletes at most $\len(\eps)-1\le 2$ elements of $B$; the end of the proof uses only that $\SigB$ generates $B$. \RIII{} is used for loops only. \Czero{} is the count $5+2<|A|$ of the case CYC; the target flow needs $|A|\ge 6$. \end{remark} \begin{remark}[Where the counts are used]\label{a3:rem:countuses} The proofs of this section use the order of $A$, the order of $B$ and the numbers $\len(t)$ at the following places, and at no other place. $j_i$ is the number of edges of the path $J$ on the threads of the edges of $N_i$. \begin{condlist}{(C0)} \item[\Czero] Lemma~\ref{lem:L2}, case CYC: at most $5+2$ values of $c$ are excluded. Case SUB, $\quo{\lambda}=0$: the set $\Phi(\eps)$ of one element is not all of $A$. Theorem~\ref{thm:master}: the target needs $|A|\ge 6$. \item[\Cone] The walk lemma: an offset $c_r$ outside a set of at most $2\,|\Lam_f(e_i)|+|\Lam_f(f_i)|\le 2\,\len(e_i)+\len(f_i)$ elements. It is used only through Lemma~\ref{lem:L1}. \item[\Ctwo] Lemma~\ref{lem:L2}, case SUB, $\quo{\lambda}\ne 0$: the reverse choice for lists with $\len(e_i)+\len(f_i)+j_i$ elements at the node $i$, and $j_i\le\max\{\len(e_i),\len(f_i)\}$ by Lemma~\ref{lem:P}. \item[\Cthree] Lemma~\ref{lem:L2}, case SUB, $\quo{\lambda}=0$: the reverse choice for lists with $\len(e_i)+\len(f_i)$ elements and at most two free edges with one excluded value each. \item[\Cfour] Lemma~\ref{a3:lem:prepB}: an element of $B$ outside a set of at most $\len(\eps)$ elements. It is used in Lemma~\ref{a3:lem:step}, hence in Lemma~\ref{lem:L1}, and in Lemma~\ref{lem:L2}, case SUB, $\quo{\lambda}\ne 0$. \end{condlist} The bound $\len(t)\le 3$ is used for: at most $6$ deleted elements in the walk lemma; at most $2$ deleted elements in Lemma~\ref{a3:lem:prepB}; the free pieces with one edge in Lemma~\ref{lem:L2}, case SUB, $\quo{\lambda}=0$; Case~1 of Lemma~\ref{lem:P}. The frame form uses \Czero, \Ctwo, \Cthree, \Cfour{} and, in the place of \Cone, the hypothesis \condLone. \end{remark} \subsection*{Three corollaries} Each corollary is Theorem~\ref{thm:master} together with a check of \Htwo. By Remark~\ref{a3:rem:countuses} this check is a check of the five inequalities \Czero--\Cfour. \begin{corollary}[No condition on the colouring]\label{cor:i} Let $|A|\ge 10$ and $|B|\ge 4$, and let $\Sig$ be a step system with \RI, \RII, \RIII{} and \AVpp. Then the conclusion of Theorem~\textup{\ref{thm:master}} holds. \end{corollary} \begin{proof} \Hone{} is assumed. \Czero: $|A|\ge 10\ge 8$. Let $H$ be a subdivision that \AVpp{} gives. It has a frame (Definition~\ref{def:frame}: every truncated cubic $3$-edge-connected graph has a well colouring, Lemma~\ref{app:wellcolouring}). In every frame $\len(t)\le 3$ for all $t$. So the left sides of \Cone, \Ctwo, \Cthree{} are at most $2\cdot 3+3=9$, $3+3+3=9$ and $3+3+2=8$; all are below $10\le|A|$. The left side of \Cfour{} is at most $3<4\le|B|$. So every frame is counted, and \Htwo{} holds. \end{proof} \begin{corollary}[Quotient of order $9$, condition \condS]\label{cor:ii} Let $|A|\ge 9$ and $|B|\ge 4$, and let $\Sig$ be a step system with \RI, \RII, \RIII{} and with \AVpp{} in this form: the good subgraph can always be chosen to be a cycle with at most $5$ edges, or a subdivision that has a frame with \condS. Then the conclusion of Theorem~\textup{\ref{thm:master}} holds. \end{corollary} \begin{proof} \Hone{} is assumed. \Czero: $|A|\ge 9\ge 8$. Let $H$ be a subdivision with a frame that has \condS. Exchange the names at the nodes so that $\len(e_i)\le\len(f_i)$ for every node $i$; the result is a frame \textup{(iii)}, and \condS{} does not depend on the names. If $\len(e_i)=3$, then $\len(f_i)=3$, and $N_i$ consists of two original edges whose threads have $3$ edges; \condS{} excludes this. So $\len(e_i)\le 2$ and $\len(f_i)\le 3$ for every $i$. The left sides: \Cone{} at most $2\cdot 2+3=7$; \Ctwo{} at most $2+3+3=8$; \Cthree{} at most $2+3+2=7$; all are below $9\le|A|$. \Cfour: at most $3<4\le|B|$. So this frame is counted, and \Htwo{} holds. \end{proof} \begin{corollary}[Subgroup of order $3$, condition \condNR]\label{cor:iii} Let $|A|\ge 10$ and $|B|=3$, and let $\Sig$ be a step system with \RI, \RII, \RIII{} and with \AVpp{} in this form: the good subgraph can always be chosen to be a cycle with at most $5$ edges, or a subdivision that has a frame with \condNR. Then the conclusion of Theorem~\textup{\ref{thm:master}} holds. \end{corollary} \begin{proof} \Hone{} is assumed. \Czero: $|A|\ge 10\ge 8$. Let $H$ be a subdivision with a frame that has \condNR. The left sides of \Cone, \Ctwo, \Cthree{} are at most $9$, $9$, $8$, as in the proof of Corollary~\ref{cor:i}; all are below $10\le|A|$. By \condNR{} every free thread has at most $2$ edges; so the left side of \Cfour{} is at most $2<3=|B|$. So this frame is counted, and \Htwo{} holds. \end{proof} \begin{remark}\label{rem:open} Corollaries~\ref{cor:ii} and~\ref{cor:iii} are conditional: no statement of this paper says that a frame with \condS, or a frame with \condNR, exists for every subdivision that the averaging step gives, so as they stand they give no new modulus and no new group. \end{remark} \subsection*{The integer form} \begin{theorem}[The integer form]\label{thm:Z} Let $k=ab$ with integers $a,b\ge 2$, $\Gam=\Z_k$, $B=a\Z_k$ \textup{(}the multiples of $a$; $b$ elements\textup{)} and $A=\Z_a$ \textup{(}reduction modulo $a$\textup{)}. Let $\Sig$ be a step system all of whose elements are liftable, and let $\Gam$, $B$, $\Sig$ satisfy the hypotheses of Theorem~\textup{\ref{thm:master}} or of Theorem~\textup{\ref{thm:masterframe}}. Then \begin{condlist}{(a)} \item[\textup{(a)}] $\Mk(G)$ is connected for every finite bridgeless multigraph $G$ \textup{(}loops allowed\textup{)}; \item[\textup{(b)}] $\Ik(G)$ is connected for every finite bridgeless loopless multigraph $G$. \end{condlist} \end{theorem} \begin{proof} The quotient $\Z_k/a\Z_k$ is $\Z_a$, and the quotient map is the reduction modulo $a$. \Czero{} gives $a\ge 8$, so $k\ge 16$. (a) $\Mk(G)=\Fl(G,\Z_k)$. It is connected by Theorem~\ref{thm:master} or Theorem~\ref{thm:masterframe}. (b) Let $G$ be loopless, and let $F$, $F'$ be two vertices of $\Ik(G)$. Their reductions $f$, $f'$ modulo $k$ are nowhere-zero $\Z_k$-flows of $G$, because $1\le|F(e)|\le k-1$. By Theorem~\ref{thm:master} or Theorem~\ref{thm:masterframe} there are $\Sig$-moves of $G$ that lead from $f$ through flows $f=f_0,f_1,\dots,f_\nu=f'$. For $r\ge 1$ the step from $f_{r-1}$ to $f_r$ is a move $f_{r-1}\to f_{r-1}+\lambda\chi_C$ of $\Mk(G)$ with a liftable $\lambda$. By Definition~\ref{def:liftable}, all integer lifts of $f_{r-1}$ and all integer lifts of $f_r$ lie in one component of $\Ik(G)$. Every $f_r$ has an integer lift (Lemma~\ref{lem:liftable}~(a)). So this is the same component for all $r$. It contains $F$ and $F'$. For $\nu=0$ use Lemma~\ref{lem:liftable}~(a): all integer lifts of one flow lie in one component. Liftable steps are used only for the $\Sig$-moves of the given loopless graph $G$. Graphs with loops occur only inside the induction in the proof of Theorem~\ref{thm:master}, where all flows have values in $\Gam$. \end{proof} \section{Averaging in a product}\label{a4:sec} This section proves Lemma~\ref{lem:AVG} and, as a first consequence, a theorem on groups \textup{(}Theorem~\ref{thm:G10}\textup{)}. Lemma~\ref{lem:AVG} gives the averaging conditions \AV{} and \AVpp{} of Definition~\ref{def:conditions} for every step system that contains generators of the two factors of a product, if $(|B|-1)/(|\Gam|-1)<1/9$. The subgroup $B$ need not be a direct factor: the proof uses it only through this ratio. The notation is that of Definition~\ref{def:stepsystem}: $\Gam$ is a finite abelian group, $B$ a subgroup, $A=\Gam/B$, $x\mapsto\quo{x}$ the quotient map, and the $A$-coordinate of a $\Gam$-flow $f$ is the $A$-flow $\quo{f}$. A \emph{zero} of the $A$-coordinate of $f$ is an edge $e$ with $\quo{f(e)}=0$. \begin{lemma}[Averaging in a product]\label{lem:AVG}\label{a4:lem:AVG} Let $\Gam=\Gam_1\times\Gam_2$ with both factors non-trivial, and let $B$ be a subgroup of $\Gam$ with \[ \frac{|B|-1}{|\Gam|-1}<\frac19 \qquad\textup{(true for every subgroup of index at least $9$).} \] Let $\Sig$ be a step system that contains a set of generators of $\Gam_1\times 0$ and a set of generators of $0\times\Gam_2$. Then on every multigraph $G$ with $m$ edges, every nowhere-zero $\Gam$-flow is $\Sig$-joined to a nowhere-zero $\Gam$-flow whose $A$-coordinate has at most $m/9$ zeros. In particular \AV{} and \AVpp{} hold. \textup{(}$B$ need not be a direct factor, and it need not be one of the two factors.\textup{)} \end{lemma} \begin{proof}[Proof of Lemma~\ref{lem:AVG}] Let $m$ be the number of edges of $G$ and $f$ a nowhere-zero $\Gam$-flow of $G$. If $m=0$, then $f$ itself has no zero. Let $m\ge 1$, and put $g=|\Gam|$. We write an element $x$ of $\Gam$ as $x=(x_{(1)},x_{(2)})$ with $x_{(1)}$ in $\Gam_1$ and $x_{(2)}$ in $\Gam_2$, and a $\Gam$-flow $\varphi$ as $\varphi=(\varphi_{(1)},\varphi_{(2)})$; here $\varphi_{(i)}$ is a $\Gam_i$-flow. Put $\Sig_1=\Sig\cap(\Gam_1\times 0)$ and $\Sig_2=\Sig\cap(0\times\Gam_2)$. The given set of generators of $\Gam_1\times 0$ is a subset of $\Sig_1$, and $\Sig_1=-\Sig_1$ because $\Sig=-\Sig$. So every element of $\Gam_1\times 0$ is a finite sum of elements of $\Sig_1$ (the empty sum for $0$). In the same way every element of $0\times\Gam_2$ is a finite sum of elements of $\Sig_2$. For a nowhere-zero $\Gam$-flow $\varphi$ of $G$, let $G\langle\varphi_{(2)}\rangle$ be the spanning subgraph of $G$ with the edge set $\supp(\varphi_{(2)})$, and $G\langle\varphi_{(1)}\rangle$ the one with the edge set $\supp(\varphi_{(1)})$. A $\Gam_1$-flow of $G\langle\varphi_{(2)}\rangle$, with the value $0$ on all other edges, is a $\Gam_1$-flow of $G$; and in the same way for $\Gam_2$. \begin{condlist}{(6)} \item[\textup{(1)}] \emph{Rounds.} A \emph{round of type~$1$} replaces a nowhere-zero $\Gam$-flow $\varphi$ by $(\varphi_{(1)}+\xi,\ \varphi_{(2)})$, where $\xi$ is a $\Gam_1$-flow of $G\langle\varphi_{(2)}\rangle$. A \emph{round of type~$2$} replaces $\varphi$ by $(\varphi_{(1)},\ \varphi_{(2)}+\xi)$, where $\xi$ is a $\Gam_2$-flow of $G\langle\varphi_{(1)}\rangle$. In a \emph{random} round, $\xi$ is chosen uniformly at random in the finite group of these flows. \item[\textup{(2)}] \emph{The result of a round is nowhere zero and $\Sig$-joined to $\varphi$.} Take type~$1$; type~$2$ is the same with the coordinates exchanged. Let $e$ be an edge. If $e$ is in $\supp(\varphi_{(2)})$, the second coordinate of the new value is $\varphi_{(2)}(e)\ne 0$. If not, then $\xi(e)=0$ and the value is $\varphi(e)\ne 0$. So the result is nowhere zero. By the cycle decomposition (Lemma~\ref{lem:cycledec}~(a), for the graph $G\langle\varphi_{(2)}\rangle$ and the group $\Gam_1$), \[ \xi=z_1\chi_{C_1}+\dots+z_\nu\chi_{C_\nu} \] with cycles $C_1,\dots,C_\nu$ of $G\langle\varphi_{(2)}\rangle$, each with a direction, and $z_r$ in $\Gam_1$. Every $(z_r,0)$ is a sum of elements of $\Sig_1$. Add these elements one at a time, each one times $\chi_{C_r}$, for $r=1,\dots,\nu$. Every flow in this sequence is $(\varphi_{(1)}+\xi',\ \varphi_{(2)})$, where $\xi'$ is a $\Gam_1$-flow that is $0$ outside $\supp(\varphi_{(2)})$; as above, it is nowhere zero. So every addition is a $\Sig$-move, and the last flow is the result of the round. \item[\textup{(3)}] \emph{The law of one edge after one random round.} Define two matrices $\Pi_1$, $\Pi_2$ with rows and columns indexed by $\Gam\setminus\{0\}$: \begin{align*} \Pi_1(x,y)&=1/|\Gam_1|\ \text{ if } x_{(2)}=y_{(2)}\ne 0, & \Pi_1(x,x)&=1\ \text{ if } x_{(2)}=0, & \Pi_1(x,y)&=0\ \text{ otherwise;}\\ \Pi_2(x,y)&=1/|\Gam_2|\ \text{ if } x_{(1)}=y_{(1)}\ne 0, & \Pi_2(x,x)&=1\ \text{ if } x_{(1)}=0, & \Pi_2(x,y)&=0\ \text{ otherwise.} \end{align*} \emph{Claim:} if $\varphi'$ is the result of a random round of type~$i$ applied to $\varphi$, then for every edge $e$ and every $y$ in $\Gam\setminus\{0\}$ \[ \Pr[\,\varphi'(e)=y\,]=\Pi_i(\varphi(e),y). \] Take type~$1$. If $e$ is not in $\supp(\varphi_{(2)})$, then $\xi(e)=0$, $\varphi'(e)=\varphi(e)$, and $\varphi(e)_{(2)}=0$: this is the second case of the definition of $\Pi_1$. Let $e$ be in $\supp(\varphi_{(2)})$. First, $e$ lies on a cycle of $G\langle\varphi_{(2)}\rangle$. Indeed, suppose it does not. Then $e$ is not a loop, because a loop is a cycle. Let $X$ be the vertex set of the component of $G\langle\varphi_{(2)}\rangle-e$ that contains one end of $e$. The other end of $e$ is not in $X$: otherwise a path of $G\langle\varphi_{(2)}\rangle-e$ between the two ends, together with $e$, is a cycle through $e$. So $e$ is the only edge of $G\langle\varphi_{(2)}\rangle$ with exactly one end in $X$. Add the conservation equations of $\varphi_{(2)}$ at the vertices of $X$: an edge with both ends in $X$ enters $X$ once and leaves it once, so it drops out. The sum is $\pm\varphi_{(2)}(e)$ plus the values of $\varphi_{(2)}$ on the other edges of $G$ with exactly one end in $X$; these edges are not in $\supp(\varphi_{(2)})$. So $\varphi_{(2)}(e)=0$, a contradiction. Hence $e$ lies on a cycle $C$ of $G\langle\varphi_{(2)}\rangle$. For every $x$ in $\Gam_1$ the flow $x\chi_C$ has the value $x$ or $-x$ on $e$. So the map $\xi\mapsto\xi(e)$ is a homomorphism of the group of the $\Gam_1$-flows of $G\langle\varphi_{(2)}\rangle$ onto $\Gam_1$. All its fibres are cosets of its kernel, so they have the same size, and $\xi(e)$ is uniform on $\Gam_1$. So $\varphi'(e)=(\varphi(e)_{(1)} +\xi(e),\ \varphi(e)_{(2)})$ is uniform on the $|\Gam_1|$ elements $y$ with $y_{(2)}=\varphi(e)_{(2)}$; these elements are not $0$. This is the first case of the definition of $\Pi_1$. \item[\textup{(4)}] \emph{The chain.} Put $f^{(0)}=f$. For $r\ge 0$ let $f^{(r+1)}$ be the result of a random round applied to $f^{(r)}$, of type~$1$ if $r$ is even and of type~$2$ if $r$ is odd, with a new independent choice. This is a finite probability space, and every sequence of choices has positive probability. By \textup{(2)}, every $f^{(r)}$ is a nowhere-zero $\Gam$-flow that is $\Sig$-joined to $f$, for every sequence of choices. Fix an edge $e$, and let $\mu_r$ be the law of $f^{(r)}(e)$: a row vector indexed by $\Gam\setminus\{0\}$, with non-negative entries and sum $1$. If $\Pr[f^{(r)}=\varphi]>0$, then under the condition $f^{(r)}=\varphi$ the law of $f^{(r+1)}(e)$ is the row $\Pi_i(\varphi(e),\cdot)$, by \textup{(3)}. Summing over $\varphi$ gives $\mu_{r+1}=\mu_r\Pi_i$, with $i=1$ for even $r$ and $i=2$ for odd $r$. So, with $\Pi=\Pi_1\Pi_2$, \[ \mu_{2h}=\mu_0\Pi^h\qquad\text{for every integer } h\ge 0, \] where $\mu_0$ has the entry $1$ at $f(e)$. Only the law of one edge is used, not the joint law of two edges. \item[\textup{(5)}] \emph{Mixing.} The matrices $\Pi_1$, $\Pi_2$ have non-negative entries and row sums $1$. They are symmetric: for $x\ne y$ the condition $x_{(2)}=y_{(2)}\ne 0$ is symmetric in $x$ and $y$, and the same holds for $\Pi_2$. So their column sums are $1$ too, and the same holds for $\Pi$. Let $\eta$ be the row vector with all entries $1/(g-1)$. Then $\eta\Pi=\eta$. \emph{Claim: every entry of $\Pi^3$ is positive.} Four facts, for $x$, $y$ in $\Gam\setminus\{0\}$: \begin{condlist}{(iii)} \item[\textup{(i)}] $\Pi(x,x)>0$. If $x_{(2)}\ne 0$: $\Pi_1(x,x)=1/|\Gam_1|$, and $\Pi_2(x,x)$ is $1/|\Gam_2|$ or $1$. If $x_{(2)}=0$, then $x_{(1)}\ne 0$: $\Pi_1(x,x)=1$ and $\Pi_2(x,x)=1/|\Gam_2|$. \item[\textup{(ii)}] If $x_{(2)}\ne 0$ and $y_{(1)}\ne 0$, then $\Pi(x,y)>0$: $\Pi_1$ leads from $x$ to $(y_{(1)},x_{(2)})$, and $\Pi_2$ leads from there to $y$, because $y_{(1)}\ne 0$. \item[\textup{(iii)}] If $x_{(2)}=0$, then $\Pi(x,x')>0$ for $x'=(x_{(1)},z)$ and every $z$ in $\Gam_2$: $\Pi_1$ keeps $x$, and $\Pi_2$ leads to $x'$, because $x_{(1)}\ne 0$. \item[\textup{(iv)}] If $y_{(1)}=0$, then $\Pi(x'',y)>0$ for $x''=(z,y_{(2)})$ and every $z$ in $\Gam_1$: $\Pi_1$ leads from $x''$ to $y$, because $y_{(2)}\ne 0$, and $\Pi_2$ keeps $y$, because $y_{(1)}=0$. \end{condlist} Let $x$, $y$ be given. There is $x'$ with $x'_{(2)}\ne 0$ and $\Pi(x,x')>0$: take $x'=x$ by (i) if $x_{(2)}\ne 0$, and otherwise $x'=(x_{(1)},z)$ with $z\ne 0$ by (iii); such a $z$ exists because $\Gam_2$ is not trivial. If $y_{(1)}\ne 0$, then $\Pi(x',y)>0$ by (ii), and $\Pi^3(x,y)\ge\Pi(x,x')\,\Pi(x',y)\,\Pi(y,y)>0$ by (i). If $y_{(1)}=0$, choose $z\ne 0$ in $\Gam_1$ (it exists because $\Gam_1$ is not trivial) and put $x''=(z,y_{(2)})$. Then $\Pi(x',x'')>0$ by (ii), and $\Pi(x'',y)>0$ by (iv); so $\Pi^3(x,y)\ge\Pi(x,x')\,\Pi(x',x'')\,\Pi(x'',y)>0$. \emph{A bound.} Every positive entry of $\Pi_i$ is $1/|\Gam_i|$ or $1$, so it is at least $1/|\Gam_i|$. A positive entry of $\Pi$ is a sum of products $\Pi_1(x,z)\Pi_2(z,y)$, one of which is positive; so it is at least $1/(|\Gam_1|\,|\Gam_2|)=1/g$. In the same way every entry of $\Pi^3$, which is positive, is at least $1/g^3$. Put \[ \delta=(g-1)/g^3,\qquad\text{so } 0<\delta<1, \] and let $\Pi_{\mathrm u}$ be the matrix with all entries $1/(g-1)$. The matrix $\Pi^3-\delta\Pi_{\mathrm u}$ has non-negative entries \textup{(}each is at least $1/g^3-\delta/(g-1)=0$\textup{)} and row sums $1-\delta$. So $\Pi^3=\delta\Pi_{\mathrm u}+(1-\delta)\Pi_{\mathrm r}$ with a matrix $\Pi_{\mathrm r}$ that has non-negative entries and row sums $1$. From $\eta\Pi^3=\eta$ and $\eta\Pi_{\mathrm u}=\eta$ we get $\eta\Pi_{\mathrm r}=\eta$. For a row vector $v$ put $\|v\|=\sum_y|v(y)|$. If $R$ has non-negative entries and row sums $1$, then $\|vR\|\le\|v\|$. Let $\mu$ be a row vector with non-negative entries and sum $1$. Then $\mu\Pi_{\mathrm u}=\eta$, and \[ \mu\Pi^3-\eta=\delta\eta+(1-\delta)\,\mu\Pi_{\mathrm r}-\eta=(1-\delta)\,(\mu-\eta)\,\Pi_{\mathrm r}, \] and so $\|\mu\Pi^3-\eta\|\le(1-\delta)\,\|\mu-\eta\|$. The vector $\mu\Pi^3$ again has non-negative entries and sum $1$. By induction, for every integer $j\ge 0$, \[ \|\mu\Pi^{3j}-\eta\|\le(1-\delta)^j\,\|\mu-\eta\|\le 2\,(1-\delta)^j. \] \item[\textup{(6)}] \emph{Zeros.} Every $f^{(r)}$ is nowhere zero. So an edge $e$ is a zero of the $A$-coordinate of $f^{(r)}$ if and only if $f^{(r)}(e)$ is in $B\setminus\{0\}$. Let $j\ge 0$. After $6j$ rounds, that is $3j$ pairs of rounds, the law of $f^{(6j)}(e)$ is $\mu_{6j}=\mu_0\Pi^{3j}$ by \textup{(4)}. Let $\mathcal{E}_e$ be the event that $e$ is a zero of the $A$-coordinate of $f^{(6j)}$. By \textup{(5)}, \begin{align*} \Pr[\mathcal{E}_e] &=\sum_{y\in B\setminus\{0\}}\mu_{6j}(y) \le\sum_{y\in B\setminus\{0\}}\Bigl(\frac{1}{g-1}+\Bigl|\mu_{6j}(y)-\frac{1}{g-1}\Bigr|\Bigr)\\ &\le\frac{|B|-1}{|\Gam|-1}+\|\mu_{6j}-\eta\| \le\frac{|B|-1}{|\Gam|-1}+2\,(1-\delta)^j. \end{align*} By hypothesis $(|B|-1)/(|\Gam|-1)<1/9$. Fix $j$ so large that $2\,(1-\delta)^j<1/9-(|B|-1)/(|\Gam|-1)$. Then every edge is a zero with probability less than $1/9$. By linearity of expectation, the expected number of zeros of the $A$-coordinate of $f^{(6j)}$ is less than $m/9$. So some sequence of choices gives a flow $f^{(6j)}$ with fewer than $m/9$ zeros. By \textup{(4)} it is a nowhere-zero $\Gam$-flow that is $\Sig$-joined to $f$. This proves the first statement. \end{condlist} \AV{} is the first statement for the loopless $3$-edge-connected multigraphs with at least two vertices. \AVpp{} follows from \AV{} by Lemma~\ref{lem:AVtoAVpp}. The group $B$ is used only in step \textup{(6)}, through its order; it need not be a direct factor. \end{proof} \begin{remark}[The estimate behind Lemma~\ref{lem:AVG}]\label{rem:avground} Steps \textup{(1)}--\textup{(5)} and the first display of step \textup{(6)} of the proof of Lemma~\ref{lem:AVG} do not use the hypothesis $(|B|-1)/(|\Gam|-1)<1/9$. They show: let $\Gam=\Gam_1\times\Gam_2$ with both factors non-trivial, let $B$ be any subgroup of $\Gam$, and let $\Sig$ be as in Lemma~\ref{lem:AVG}. Then for every multigraph $G$, every nowhere-zero $\Gam$-flow $f$ of $G$ and every integer $j\ge 0$, every outcome of the random flow $f^{(6j)}$ of the proof is a nowhere-zero $\Gam$-flow that is $\Sig$-joined to $f$, and every edge of $G$ is a zero of the $A$-coordinate of $f^{(6j)}$ with probability at most \[ \frac{|B|-1}{|\Gam|-1}+2\,(1-\delta)^j,\qquad \delta=\frac{|\Gam|-1}{|\Gam|^3}. \] Part~II uses this estimate also at the ratio $1/9$ \textup{(}Lemma~\ref{lem:average}\textup{)}. \end{remark} \begin{remark}[The ratio]\label{a4:rem:ratio} Put $a=|A|$ and $b=|B|$, so $|\Gam|=ab$. The condition $(b-1)/(ab-1)<1/9$ is the inequality $9b-91$. \end{remark} \begin{remark}[Internal direct sums]\label{a4:rem:internal} Lemma~\ref{lem:AVG} speaks only of the group $\Gam$, its subgroup $B$, the set $\Sig$ and flows; an isomorphism of groups carries all of them along. So the lemma applies to a group $\Gam$ with two non-trivial subgroups $\Gam_1$, $\Gam_2$ such that $\Gam_1\cap\Gam_2=\{0\}$ and $\Gam_1+\Gam_2=\Gam$: the map $(x_1,x_2)\mapsto x_1+x_2$ is an isomorphism of $\Gam_1\times\Gam_2$ onto $\Gam$, and $\Sig$ must contain a set of generators of $\Gam_1$ and a set of generators of $\Gam_2$. For $\Gam=\Z_k$ with $k=cc'$, where $c$, $c'\ge 2$ are coprime, take $\Gam_1=c'\Z_k$ and $\Gam_2=c\Z_k$. Then $\Gam_1$ has order $c$ and is generated by $c'$, and $\Gam_2$ has order $c'$ and is generated by $c$. Their intersection is $\{0\}$, because its order divides $c$ and $c'$; so $|\Gam_1+\Gam_2|=cc'=k$. \end{remark} \subsection*{A consequence for groups} \begin{theorem}[The group theorem for a subgroup of index at least $10$]\label{thm:G10} Let $\Gam$ be a finite abelian group that is not cyclic of prime-power order, and let $|\Gam|$ have a divisor $d$ with $4\le d\le|\Gam|/10$. Then $\Fl(G,\Gam)$ is connected for every finite bridgeless multigraph $G$ \textup{(}loops allowed\textup{)}. The smallest orders are $40$, $44$, $48$, $50$. \textup{(}This is Corollary~\ref{cor:i} with $\Sig=\Gam\setminus\{0\}$ and a subgroup $B$ of order $d$; \AVpp{} by Lemma~\ref{lem:AVG}.\textup{)} \end{theorem} \begin{proof} $|\Gam|\ge 10\,d\ge 40$. The group $\Gam$ is a direct product of cyclic groups of prime-power order. It is not cyclic of prime-power order and not trivial, so there are at least two factors, and $\Gam=\Gam_1\times\Gam_2$ with both factors non-trivial. Let $B$ be a subgroup of $\Gam$ of order $d$; it exists, because $d$ divides $|\Gam|$. Put $A=\Gam/B$. Then $|B|=d\ge 4$ and $|A|=|\Gam|/d\ge 10$. Put $\Sig=\Gam\setminus\{0\}$. This is a step system. Every move of $\Fl(G,\Gam)$ is a $\Sig$-move. Here $\SigB=B\setminus\{0\}$, and $\SigA=A\setminus\{0\}$, because every non-zero element of $A$ is the image of a non-zero element of $\Gam$. So \RI, \RII{} and \RIII{} hold with walks of one step: two different elements of $A\setminus X$, of $B\setminus X$, or of $\Gam\setminus\{0\}$, differ by an element of $\SigA$, of $\SigB$, or of $\Sig$. The set $\Sig$ contains all non-zero elements of $\Gam_1\times 0$ and of $0\times\Gam_2$, and the index of $B$ is $|A|\ge 10$; so \AVpp{} holds by Lemma~\ref{lem:AVG}. Corollary~\ref{cor:i} gives the connectivity of $\Fl(G,\Gam)$. The orders: an order $n$ of such a group has a divisor $d$ with $4\le d\le n/10$, so $n\ge 40$. For $40\le n\le 50$ such a divisor exists exactly for $n=40$, $44$, $48$ \textup{(}$d=4$\textup{)} and $n=50$ \textup{(}$d=5$\textup{)}; for these $n$ the cyclic group $\Z_n$ is not of prime-power order. \end{proof} For the cyclic group $\Z_k$ the theorem is a statement on $\Mk(G)=\Fl(G,\Z_k)$ only. It says nothing about the integer graph $\Ik(G)$: Theorem~\ref{thm:Z} needs a step system of liftable elements, and the lemmas of Section~\ref{a1:sec} do not show that every element of $\Z_k\setminus\{0\}$ is liftable \textup{(}Remark~\ref{a8:rem:master}\textup{)}. \section{The moduli \texorpdfstring{$40$, $48$ and $66$}{40, 48 and 66}; the divisor criterion}\label{a5:sec} This section proves that $\Mk(G)$ and $\Ik(G)$ are connected for every finite bridgeless loopless multigraph $G$ when $k=40$, $k=48$ or $k=66$ (Theorems~\ref{thm:k40} and~\ref{thm:k48}, Corollary~\ref{cor:k66}), and Criterion~\ref{crit:F10} for $k=ab$ with $a\ge 10$ and a composite $b\ge 4$. Every result comes from Theorem~\ref{thm:master}, in the form of Corollary~\ref{cor:i}, and from Theorem~\ref{thm:Z}. Lemma~\ref{a5:lem:route} says which hypotheses must be checked; for each modulus they are checked one by one. Criterion~\ref{crit:F10} uses the theorems of Cauchy and Davenport and of Mader and Watkins, but only in its last sentence; the proofs for $40$, $48$ and $66$ use neither of them. \subsection*{Setting} Throughout this section $k=ab$ with integers $a,b\ge 2$, $\Gam=\Z_k$, $B=a\Z_k$ \textup{(}the multiples of $a$; $b$ elements\textup{)} and $A=\Z_a$; the quotient map is the reduction modulo $a$. This is the setting of Theorem~\ref{thm:Z}. For an integer $z$ we write $z$ also for its residue in $\Z_k$. Because $a$ divides $k$, the image in $A$ of the residue of $z$ is the residue of $z$ modulo $a$. Let $D$ be a set of non-zero elements of $\Z_a$ with $D=-D$. The \emph{graph on $\Z_a$ with the step set $D$} has the vertex set $\Z_a$; two vertices $x$, $y$ are adjacent if $y-x$ is in $D$. It \emph{stays connected after the deletion of at most $6$ elements} if for every set $X$ of at most $6$ elements of $\Z_a$, any two elements of $\Z_a\setminus X$ are joined by a walk in $\Z_a\setminus X$ all of whose steps are in $D$. For $D=\SigA$ this is condition \RI{} of Definition~\ref{def:conditions}. \subsection*{The master theorem for \texorpdfstring{$\Z_k$}{Z\_k} with a quotient of order at least \texorpdfstring{$10$}{10}} \begin{lemma}[The master theorem for $\Z_k$ with a quotient of order at least $10$]\label{a5:lem:route} Let $k=ab$ with integers $a\ge 10$ and $b\ge 4$, $\Gam=\Z_k$, $B=a\Z_k$ and $A=\Z_a$. Let $\Sig$ be a step system with the following three properties. \begin{condlist}{(c)} \item[\textup{(a)}] Every element of $\Sig$ is liftable. \item[\textup{(b)}] $\Sig$ has the properties \RI, \RII{} and \RIII. \item[\textup{(c)}] $\Z_k$ has two non-trivial subgroups $\Gam_1$ and $\Gam_2$ whose orders are coprime and have the product $k$, and $\Sig$ contains a generator of $\Gam_1$ and a generator of $\Gam_2$. \end{condlist} Then $\Gam$, $B$ and $\Sig$ satisfy the hypotheses of Theorem~\ref{thm:master}. Hence $\Mk(G)$ and $\Ik(G)$ are connected for every finite bridgeless loopless multigraph $G$, and $\Mk(G)$ is connected also for every finite bridgeless multigraph $G$ with loops. \end{lemma} \begin{proof} \emph{The averaging condition.} The subgroups of $\Z_k$ are cyclic; so by (c) the set $\Sig$ contains a set of generators of $\Gam_1$ and a set of generators of $\Gam_2$. The order of $\Gam_1\cap\Gam_2$ divides the two coprime orders, so $\Gam_1\cap\Gam_2=\{0\}$, and $|\Gam_1+\Gam_2|=|\Gam_1|\,|\Gam_2|=k$, so $\Gam_1+\Gam_2=\Z_k$. By Remark~\ref{a4:rem:internal}, Lemma~\ref{lem:AVG} applies to these two factors. Its ratio condition holds: \[ \frac{|B|-1}{|\Gam|-1}=\frac{b-1}{ab-1}<\frac1a\le\frac1{10}<\frac19 ; \] the first inequality is $a(b-1)1$. So \AVpp{} holds by Lemma~\ref{lem:AVG}. \emph{\Hone.} \RI, \RII{} and \RIII{} hold by (b), and \AVpp{} holds. \emph{\Htwo.} \Czero: $|A|=a\ge 10\ge 8$. Let $H$ be a good subgraph that \AVpp{} gives. If $H$ is a cycle with at most $5$ edges, nothing is to be checked. If $H$ is a subdivision, it has a frame (Lemma~\ref{app:wellcolouring}; see Definition~\ref{def:frame}). Every frame of $H$ is counted, because $|A|=a\ge 10$ and $|B|=b\ge 4$: the left sides of \Cone, \Ctwo, \Cthree, \Cfour{} are at most $9$, $9$, $8$, $3$ (Remark~\ref{rem:counts}). So \Htwo{} holds. (This is the proof of Corollary~\ref{cor:i}.) \emph{Conclusion.} By (a), every element of $\Sig$ is liftable. So Theorem~\ref{thm:Z} applies: part~(a) gives $\Mk(G)$ for every finite bridgeless multigraph $G$, loops allowed, and part~(b) gives $\Ik(G)$ for every finite bridgeless loopless multigraph $G$. \end{proof} \subsection*{The moduli \texorpdfstring{$40$, $48$ and $66$}{40, 48 and 66}} For these three moduli, every non-zero element of $A$ except two is the image of a step. The following lemma gives \RI{} in this case. \begin{lemma}[At most two non-steps]\label{a5:lem:twononsteps} Let $a\ge 10$, and let $D$ be a set of non-zero elements of $\Z_a$ with $D=-D$ and $|D|\ge a-3$. Then the graph on $\Z_a$ with the step set $D$ stays connected after the deletion of at most $6$ elements. \end{lemma} \begin{proof} Let $D^{\mathrm c}$ be the set of the non-zero elements of $\Z_a$ that are not in $D$. Then $|D^{\mathrm c}|\le 2$ and $D^{\mathrm c}=-D^{\mathrm c}$. Let $X$ be a set of at most $6$ elements of $\Z_a$, and suppose that two elements of $\Z_a\setminus X$ are not joined by a walk in $\Z_a\setminus X$ with steps in $D$. Let $U$ be the set of the elements of $\Z_a\setminus X$ that are joined to the first of them, and $V=\Z_a\setminus(X\cup U)$. Both are not empty, and no step of $D$ leads from $U$ to $V$. So $v-u$ is in $D^{\mathrm c}$ for all $u$ in $U$ and $v$ in $V$. For a fixed $u$ this gives $|V|\le|D^{\mathrm c}|\le 2$, and for a fixed $v$ it gives $|U|\le 2$. But $|U|+|V|=|\Z_a\setminus X|\ge a-6\ge 4$. Hence $a=10$ and $|U|=|V|=|D^{\mathrm c}|=2$. The group $\Z_a$ has at most one non-zero element $x$ with $x=-x$ (namely $a/2$, if $a$ is even). So $D^{\mathrm c}=\{d_0,-d_0\}$ with $d_0\ne -d_0$. Write $U=\{u_1,u_2\}$. Then $V=u_1+D^{\mathrm c}=u_2+D^{\mathrm c}$. The element $u_2+d_0$ is in $u_1+D^{\mathrm c}$ and differs from $u_1+d_0$; so $u_2+d_0=u_1-d_0$. In the same way $u_2-d_0=u_1+d_0$. Subtracting the two equations gives $4d_0=0$ in $\Z_{10}$. So $5$ divides $2d_0$, hence $5$ divides $d_0$, and $2d_0=0$ in $\Z_{10}$, that is $d_0=-d_0$. This is a contradiction. \end{proof} \begin{theorem}[The modulus $40$]\label{thm:k40}\label{a5:thm:k40} $\Mk[40](G)$ and $\Ik[40](G)$ are connected for every finite bridgeless loopless multigraph $G$. \end{theorem} \begin{proof} Here $k=40$, $a=10$, $b=4$, $\Gam=\Z_{40}$, $B=10\Z_{40}=\{0,10,20,30\}$ and $A=\Z_{10}$. We check the hypotheses of Lemma~\ref{a5:lem:route}. \begin{condlist}{(Product)} \item[\textup{(Set)}] $\Sig=\{\pm 1,\pm 2,\pm 4,\pm 5,\pm 8,\pm 10,20\}$, a set of $13$ elements of $\Z_{40}$. They are not $0$, and $\Sig=-\Sig$ because $-20=20$. So $\Sig$ is a step system. \item[\textup{(Lifts)}] $\pm 1$ are liftable by Lemma~\ref{lem:liftable}~(b). The numbers $2$, $4$, $5$, $8$, $10$, $20$ are divisors of $40$ smaller than $40$; so $\pm 2$, $\pm 4$, $\pm 5$, $\pm 8$, $\pm 10$ and $20$ are liftable by Lemma~\ref{lem:liftable}~(d). This is hypothesis~(a) of Lemma~\ref{a5:lem:route}. \item[\textup{(}$\SigA$\textup{)}] The residues modulo $10$ of $1,-1,2,-2,4,-4,5,-5,8,-8$ are $1,9,2,8,4,6,5,5,8,2$; those of $\pm 10$ and $20$ are $0$. So $\SigA=\{1,2,4,5,6,8,9\}$: all non-zero elements of $\Z_{10}$ except $3$ and $7$. \item[\textup{(}$\SigB$\textup{)}] $\SigB=\Sig\cap B=\{10,20,30\}=B\setminus\{0\}$. \item[\RI] $\SigA=-\SigA$ and $|\SigA|=7=a-3$. Lemma~\ref{a5:lem:twononsteps}. With \RII{} and \RIII{} below, this is hypothesis~(b). \item[\RII] Two different elements of $B$ differ by an element of $B\setminus\{0\}=\SigB$. So for every set $X$, any two elements of $B\setminus X$ are joined by a walk of one step. \item[\RIII] $1$ is in $\Sig$. For $1\le x6$. So the deletion of a set $X$ of at most $6$ elements leaves a connected graph on $\Z_a\setminus X$; a path in it is a walk in $\Z_a\setminus X$ with steps in $D$. The proof uses $a$ only through the fact that $\Z_a$ has a set $D$ as in the lemma; such a set exists only if $a-1\ge|D|\ge 9$, that is $a\ge 10$. No larger lower bound on $a$ is used. \end{proof} \begin{remark}[Nine steps for $a\le 12$ without a quoted theorem]\label{a5:rem:ninesmall} For $a=10$, $11$, $12$ Theorem~\ref{q:MW} is not needed in Lemma~\ref{lem:ninesteps}: there $|D|\ge 9\ge a-3$, and Lemma~\ref{a5:lem:twononsteps} gives the same conclusion. \end{remark} \begin{criterion}[The divisor criterion]\label{crit:F10}\label{a5:crit:F10} Let $k$ be not a prime power, and $k=ab$ with $a\ge 10$ and a composite number $b\ge 4$. Let $D$ be the set of the non-zero residues modulo $a$ of $\pm 1$, $\pm 2$ and $\pm q$, for all divisors $q$ of $k$ with $1\le q= 13}: the double step}\label{a6:sec} This section proves Theorem~\ref{a6:thm:4p}: for every prime $p\ge 13$ and $k=4p$, the graphs $\Mk(G)$ and $\Ik(G)$ are connected for every finite bridgeless loopless multigraph $G$ \textup{(}$\Mk(G)$ also with loops\textup{)}. The theorem is Theorem~\ref{thm:Z} together with the frame form of the master theorem, Theorem~\ref{thm:masterframe}. The step system $\Sig$ of this section has only six $A$-parts: $\pm 1$, $\pm 2$, $\pm 4$ modulo $p$. If six elements of $A$ are deleted, one element can be isolated (Remark~\ref{a6:rem:R1}). So \RI{} fails, and the walk lemma and Lemma~\ref{lem:L1} of Section~\ref{a3:sec} cannot be used. The frame form asks instead for \RIm, which holds (Lemma~\ref{a6:lem:V}), and for the conclusion \condLone{} of Lemma~\ref{lem:L1}, which Lemma~\ref{a6:lem:L14p} proves. Its proof walks in the flows of the auxiliary graph that avoid the lists, with ordinary steps and with \emph{double steps} (Lemma~\ref{a6:lem:walk4p}). Lemma~\ref{a6:lem:DS2} makes a double step in the graph. A double step leaves an isolated value through a middle flow that has exactly one zero on the protected threads. It has two routes, and its step $\pm 1$ has two lifts, $\pm 1$ and $\pm(1+2p)$. A parity argument in $B$ shows that one of the four choices works. Two conventions of Section~\ref{a3:sec} hold in this section. If $H$ is a subgraph of a multigraph $G$, the reference orientation of $G$ on $E(H)$ is the direction of the threads. And $u+\pi(Y)$ means: add $\pi(Y)$ on $E(H)$ and nothing outside $E(H)$. \subsection*{The setting} Throughout this section $p$ is a prime with $p\ge 13$, and \[ k=4p,\qquad \Gam=\Z_k,\qquad B=p\Z_k=\{0,p,2p,3p\},\qquad A=\Gam/B=\Z_p, \] where the quotient map $x\mapsto\quo{x}$ is the reduction modulo $p$. So $k=ab$ with $a=p$ and $b=4$, as in Theorem~\ref{thm:Z}. As in Section~\ref{a2:sec}, the $A$-part of $x$ in $\Gam$ is $\quo{x}$. A \emph{protected edge} is an edge of a protected thread. The step system of this section is \[ \Sig=\{\pm 1,\ \pm(1+2p),\ \pm 2,\ \pm 4,\ \pm p,\ 2p\}\subseteq\Gam . \] \begin{condlist}{(S6)} \item[\textup{(S1)}] \emph{The group $B$.} $B$ is cyclic of order $4$, generated by $p$. Its only element of order $2$ is $2p$. So $y+2p\ne y$ for $y$ in $B$, $2p+2p=0$, and $2p\,\xi=2p$ for every odd integer $\xi$. For $x$ in $\Gam$: if $\quo{x}\ne 0$, then $x\ne 0$; if $\quo{x}=0$, then $x$ is in $B$. Adding an element of $B$ does not change the $A$-part. \item[\textup{(S2)}] \emph{Small integers.} An integer $\xi$ with $1\le|\xi|\le 12$ is not $0$ modulo $p$, because $p\ge 13$. \item[\textup{(S3)}] \emph{$\Sig$ is a step system of liftable elements.} $0$ is not in $\Sig$: $1+2p$ is odd, and $1$, $2$, $4$, $p$, $2p$ and their negatives are not multiples of $4p$. $\Sig=-\Sig$, because $-2p=2p$. Every element of $\Sig$ is liftable (Definition~\ref{def:liftable}), by Corollary~\ref{a1:4p}; in detail, by Lemma~\ref{lem:liftable}: $\pm 1$ by~(b); $\pm(1+2p)$ by~(e); $\pm 2$ by~(c), and also by~(d); $\pm 4$, $\pm p$, $2p$ by~(d), because $4$, $p$, $2p$ are divisors of $k$ smaller than $k$, $-p=3p$ and $-2p=2p$. \item[\textup{(S4)}] \emph{$\SigB$, $\SigA$ and the lifts.} The elements $\pm1$, $\pm(1+2p)$, $\pm2$, $\pm4$ are $\pm 1$, $\pm 2$, $\pm 4$ modulo $p$, so they are not in $B$ (S2). Hence $\SigB=\Sig\cap B=\{p,2p,3p\}=B\setminus\{0\}$. The $A$-parts are \[ \SigA=\{1,-1,2,-2,4,-4\}\subseteq\Z_p \] \textup{(}the $A$-part of $1+2p$ is $1$\textup{)}. These six elements are different and non-zero: they and their differences are integers $\xi$ with $1\le|\xi|\le 8$ (S2). The fixed lifts (Definition~\ref{def:stepsystem}) are $\lift(\sigma)=\sigma$, the integer $\sigma$ in $\{\pm1,\pm2,\pm4\}$ read modulo $k$. The elements $\pm(1+2p)$ are used only in the double step (Lemma~\ref{a6:lem:DS2}). \item[\textup{(S5)}] \emph{The two lifts of $\pm 1$.} Let $\mu_1$ be $1$ or $-1$. The elements $\mu_1$ and $\mu_1(1+2p)$ of $\Sig$ both have the $A$-part $\mu_1$, and $\mu_1(1+2p)-\mu_1=2p\,\mu_1=2p$ by (S1). \item[\textup{(S6)}] \emph{Condition \RII.} Let $X$ be a set of at most $2$ elements of $B$. Two elements of $B\setminus X$ differ by an element of $B\setminus\{0\}=\SigB$: they are joined by a walk with one step. So \RII{} holds. Also $|B|=4$. \end{condlist} \subsection*{Six steps modulo \texorpdfstring{$p$}{p}} The proof of the next lemma uses two quoted theorems on sums of sets in $\Z_p$: the theorem of Cauchy and Davenport (Theorem~\ref{q:CD}) and the theorem of Vosper (Theorem~\ref{q:vosper}). For subsets $\mathcal U$, $\mathcal V$ of $\Z_p$ we write $\mathcal U+\mathcal V=\{\,x+x' : x\in\mathcal U,\ x'\in\mathcal V\,\}$. \begin{lemma}[Six steps modulo $p$]\label{lem:V}\label{a6:lem:V} Let $X$ be a set of at most $6$ elements of $A=\Z_p$, and consider the graph on $A\setminus X$ in which $x$ and $x'$ are adjacent if $x'-x$ is in $\SigA=\{\pm1,\pm2,\pm4\}$. \begin{condlist}{(a)} \item[\textup{(a)}] If $|X|\le 5$, this graph is connected. So \RIm{} holds for $\Sig$. \item[\textup{(b)}] If $|X|=6$ and $X\ne c+\SigA$ for every $c$ in $A$, this graph is connected. \item[\textup{(c)}] If $X=c+\SigA$, then $c$ is not in $X$, no other element $c'$ has $X=c'+\SigA$, and the graph has exactly two components: $\{c\}$ and $A\setminus(X\cup\{c\})$. \end{condlist} \end{lemma} \begin{proof} Put $\Sigma^0_A=\SigA\cup\{0\}$; it has $7$ elements (S4). Let $\Omega$ be the vertex set of a component of the graph, and assume $\Omega\ne A\setminus X$. \begin{condlist}{(5)} \item[\textup{(1)}] $\Omega+\Sigma^0_A\subseteq\Omega\cup X$. Reason: for $x$ in $\Omega$ and $\sigma$ in $\Sigma^0_A$, the element $x+\sigma$ is in $X$, or it is a vertex that is equal or adjacent to $x$, hence in $\Omega$. So $|\Omega+\Sigma^0_A|\le|\Omega|+|X|$. Some vertex is not in $\Omega$, so $\Omega\cup X\ne A$ and $|\Omega+\Sigma^0_A|\le p-1$. \item[\textup{(2)}] By the theorem of Cauchy and Davenport, $|\Omega+\Sigma^0_A|\ge\min\{p,\ |\Omega|+6\}$. By (1) this minimum is not $p$; so $|\Omega+\Sigma^0_A|\ge|\Omega|+6$. With (1): $|X|\ge 6$. This proves (a): if $|X|\le 5$, no such $\Omega$ exists. If $|X|=6$, then $|\Omega+\Sigma^0_A|=|\Omega|+6$, and $\Omega+\Sigma^0_A=\Omega\cup X$, because the two sets have the same number of elements \textup{(}$\Omega$ and $X$ are disjoint\textup{)}. From now on $|X|=6$. \item[\textup{(3)}] $\Sigma^0_A$ is not an arithmetic progression. Suppose that $\Sigma^0_A=\{\,b'+r d : r=0,1,\dots,6\,\}$ with $d\ne 0$. The sum of the seven elements is $0$, because $\SigA=-\SigA$; it is also $7b'+21d=7(b'+3d)$. Since $p\ne 7$, $b'=-3d$, and $\Sigma^0_A=\{\,r d : -3\le r\le 3\,\}$. Let $x$ be the inverse of $d$ in $\Z_p$. Then $x\,\Sigma^0_A=\{-3,\dots,3\}$, so $x$, $2x$ and $4x$ are in $\{\pm1,\pm2,\pm3\}$. If $x=\pm1$, then $4x=\pm 4$; if $x=\pm 2$, then $2x=\pm 4$; if $x=\pm 3$, then $2x=\pm 6$. But $\pm4$ and $\pm6$ are not in $\{\pm1,\pm2,\pm3\}$: the differences are integers $\xi$ with $1\le|\xi|\le 9$ (S2). This is a contradiction. \item[\textup{(4)}] By (2), $|\Omega+\Sigma^0_A|=|\Omega|+|\Sigma^0_A|-1$. If $|\Omega|\ge 2$ and $|\Omega|+6\le p-2$, the theorem of Vosper says that $\Sigma^0_A$ is an arithmetic progression, against (3). So $|\Omega|=1$, or $|\Omega|+6\ge p-1$; in the second case $|\Omega|+6=p-1$ by (1). \item[\textup{(5)}] If $|\Omega|=1$, say $\Omega=\{c\}$, then (2) gives $\{c\}\cup(c+\SigA)=\{c\}\cup X$. Here $c$ is not in $X$ (it is a vertex) and not in $c+\SigA$ ($0$ is not in $\SigA$). So $X=c+\SigA$. If $|\Omega|=p-7$, then $|\Omega\cup X|=p-1$, so exactly one element $c$ of $A$ is outside $\Omega\cup X$. It is a vertex. Its component is not $\Omega$, and it lies in $A\setminus(\Omega\cup X)=\{c\}$. So $\{c\}$ is a component, and the first case, for this component, gives $X=c+\SigA$. \end{condlist} (b) By (5), if the graph is not connected, then $X=c+\SigA$ for some $c$. (c) Let $X=c+\SigA$. Then $c$ is not in $X$, because $0$ is not in $\SigA$. If also $X=c'+\SigA$, then the sum of the elements of $X$ is $6c=6c'$, so $c=c'$ by (S2). All neighbours $c+\sigma$ of $c$ are in $X$, so $\{c\}$ is a component. The set $A\setminus(X\cup\{c\})$ has $p-7\ge 6$ elements; let $\Omega$ be a component that meets it. $\Omega\ne A\setminus X$, because $\{c\}$ is another component. By (4) and (5), $|\Omega|=1$ or $|\Omega|=p-7$. If $|\Omega|=1$, then (5) gives $\Omega=\{c'\}$ with $X=c'+\SigA$, so $c'=c$; but $\Omega$ meets $A\setminus(X\cup\{c\})$. So $|\Omega|=p-7$, and $\Omega\subseteq A\setminus(X\cup\{c\})$ is all of this set. Hence the components are $\{c\}$ and $A\setminus(X\cup\{c\})$. \end{proof} \begin{remark}[Condition \RI{} fails]\label{a6:rem:R1} Let $X=\SigA$. Then $0$ is not in $X$, every element $0+\sigma$ with $\sigma$ in $\SigA$ is in $X$, and $p-7\ge 6$ elements other than $0$ are outside $X$. So $0$ is isolated, and \RI{} is false for $\Sig$. Theorem~\ref{thm:master} and Corollary~\ref{cor:i} do not apply. By Lemma~\ref{a6:lem:V}, this is the only way in which six deleted elements disconnect $A$: $X=c+\SigA$, with the one isolated value $c$. \end{remark} \subsection*{Double steps on the auxiliary graph} From here to the end of the proof of Lemma~\ref{a6:lem:L14p}, $H$ is a subdivision as in Definition~\ref{def:good}~(SUB), with the graphs $H_0$, $Q$ of Definition~\ref{def:threads}, and a frame of $H$ is given (Definition~\ref{def:frame}). $C_1,\dots,C_s$ are the official cycles (Definition~\ref{def:official}). For a node $j$, $C_j^{e}=C_j$ and $C_j^{f}$ are the two cycles of the node $j$ (Definition~\ref{a2:twocycles-def}), and $\Delta_j=\chi_{C_j^{f}}-\chi_{C_j^{e}}$. We use three facts of Lemma~\ref{a2:twocycles}: \begin{condlist}{(TC3)} \item[\textup{(TC1)}] $\chi_{C_j^{e}}$ and $\chi_{C_j^{f}}$ are equal on $N_j$ (part (a)). \item[\textup{(TC2)}] If $h>j$, no edge of $N_h$ lies on $C_j^{e}$ or on $C_j^{f}$. If $ij$ is not below $j$.) \item[\textup{(TC3)}] $\Delta_j$ is $0$ on $N_j\cup N_{j+1}\cup\dots\cup N_s$. For every $x$ in $A$, the $A$-flow $x\Delta_j$ is a sum of flows $x'\chi_C$ with $x'$ in $A$ and $C$ a cycle of $Q$ that has no edge in $N_j\cup\dots\cup N_s$. (Part (c): the cycles $C$ lie in $Q_{U'_j}\cup Q_{U''_j}$, whose edges are red edges and edges of pairs $N_h$ with $h$ below $j$, so $hj$, the four flows are equal to $Z$. \item[\textup{(c)}] Let $i12$. \textup{(}This is the count $12j}(g)$, $E_{\ge j}(g)$ in the same way. The valuation of $g(e)$ does not depend on the reference orientation of $e$, because $\operatorname{val}(-x)=\operatorname{val}(x)$. Every edge lies in exactly one set $E_j(g)$, and $E_{\ge j}(g)=E_{>j-1}(g)$ for $j\ge 1$. \begin{lemma}[The flattening process]\label{lem:flatten}\label{a7:flatten} Let $1\le i\le N-1$, and let $\Sig$ be a step system of $\Z_k$ that contains $1$ and $p^j$ for $1\le j\le i-1$. Let $G$ be a finite loopless multigraph and $f$ a nowhere-zero $\Z_k$-flow of $G$. Then $f$ is $\Sig$-joined to a nowhere-zero $\Z_k$-flow $f'$ such that for every $j$ in $\{0,\dots,i-1\}$: \begin{condlist}{(a)} \item[\textup{(a)}] every cycle of $G$ with no edge in $E_j(f')$ has all its edges in $E_{j}(f')|\le n_j-1$, where $n_j$ is the number of components of the multigraph $(V(G),E_{j}(g)$, make the move above with these $j$ and $C$ \textup{(}in either direction of $C$\textup{)}. The sum of the valuations of all values of $g$ is a non-negative integer, and every move lowers it by at least $1$. So the process stops. Let $f'$ be the last flow. It is $\Sig$-joined to $f$. From now on $E_j$, $E_{j}$ are the sets of $f'$. \emph{\textup{(a)}.} Let $C$ be a cycle with no edge in $E_j$. The process has stopped, so $C$ has no edge in $E_{>j}$. Every edge has a valuation, so all edges of $C$ are in $E_{j}$; an edge of $E_{>j}$ joins the components of its two ends \textup{(}it is a loop of $G_j$ if they are the same component\textup{)}. We show that $G_j$ has no cycle; then $G_j$ is a forest with $n_j$ vertices, and $|E_{>j}|=|E(G_j)|\le n_j-1$. Suppose that $G_j$ has a cycle $O$, and let $e_1$ be an edge of $O$. Go once round $O$. At every vertex of $O$ \textup{(}a component of $(V(G),E_{j}$. This contradicts \textup{(a)}. \end{proof} \subsection*{Four levels give a good subgraph} \begin{lemma}[Four levels]\label{lem:VF4}\label{a7:VF4} Let $4\le i\le N-1$ \textup{(}so $N\ge 5$\textup{)}, and let $\Sig$ be a step system of $\Z_k$ that contains $1$ and $p^j$ for $1\le j\le i-1$. Let $G$ be a loopless $3$-edge-connected multigraph with $n\ge 2$ vertices, and $f$ a nowhere-zero $\Z_k$-flow of $G$. Then $f$ is $\Sig$-joined to a nowhere-zero $\Z_k$-flow $f'$ such that the multigraph \[ \bigl(V(G),\ \{\,e : f'(e)\notin p^i\Z_k\,\}\bigr) \] contains a good subgraph \textup{(}Definition~\ref{def:good}\textup{)}. So, for $\Gam=\Z_k$, $B=p^i\Z_k$, $A=\Gam/B$ and this $\Sig$, the condition \AVpp{} of Definition~\ref{def:conditions} holds. \end{lemma} \begin{proof} Let $f'$ be the flow of Lemma~\ref{a7:flatten} for this $i$, and let $E_j$, $E_{j}$ be the sets of $f'$. For $0\le j\le i$ let $n_j$ be the number of components of $(V(G),E_{j-1}$, and $j-1$ is in $\{0,\dots,i-1\}$. By Lemma~\ref{a7:flatten}~(b), \[ |E_{j-1}|\ \ge\ \tfrac32\,n-n_{j-1}+1 . \] \emph{Step 2: few edges.} Let a component of $(V(G),E_{0$. \subsection*{The steps \texorpdfstring{$\pm 2^j$}{+-2\^{}j} in \texorpdfstring{$\Z_{2^r}$}{Z\_\{2\^{}r\}}} \begin{lemma}[The powers of $2$]\label{lem:D}\label{a7:D} Let $r\ge 2$, and let $\Delta_r$ be the graph on $\Z_{2^r}$ in which $x$ and $y$ are adjacent if $y-x$ is $2^j$ or $-2^j$ for some $j$ with $0\le j\le r-1$. Then for every set $X$ of at most $2r-2$ elements of $\Z_{2^r}$, the graph $\Delta_r-X$ is connected. \end{lemma} \begin{proof} \emph{The degree.} Let $r'\ge 2$. In $\Z_{2^{r'}}$ the elements $\pm 2^j$, $0\le j\le r'-1$, are $2r'-1$ different non-zero elements: they are not $0$; $2^j\ne 2^l$ for $j0$ of them. \emph{Attached odd survivors.} An odd element $x$ has exactly two even neighbours, $x-1$ and $x+1$ \textup{(}the only odd step is $\pm 1$, and $x-1\ne x+1$\textup{)}. Call an odd element $x$ outside $X$ \emph{attached} if $x-1$ or $x+1$ is outside $X$. An attached element is joined to the connected set of the even survivors. If $x$ is not attached, then $x-1$ and $x+1$ are both in $X$. The even elements form a cycle of length $2^{r-1}\ge 4$ in which $y$ and $y+2$ are consecutive; so $\{x-1,x+1\}$ is an edge of this cycle with both ends in $X$, and different $x$ give different edges. The $w<2^{r-1}$ even elements of $X$ induce a union of paths in this cycle, with at most $\max\{0,w-1\}$ edges. So at most $\max\{0,w-1\}$ odd elements outside $X$ are not attached. \emph{The end.} If $w\le 1$, every odd element outside $X$ is attached, and $\Delta_r-X$ is connected. Let $w\ge 2$. There are at least $2^{r-1}-(|X|-w)\ge 2^{r-1}-2r+2+w$ odd elements outside $X$, and at most $w-1$ of them are not attached; so at least $2^{r-1}-2r+3$ of them are attached. In the odd class, which induces a copy of $\Delta_{r-1}$, every element has degree $2(r-1)-1=2r-3$, so it has exactly $(2^{r-1}-1)-(2r-3)=2^{r-1}-2r+2$ non-neighbours in the odd class other than itself. Let $x$ be an odd element outside $X$ that is not attached. The attached elements are at least $2^{r-1}-2r+3$ odd elements other than $x$, so one of them is a neighbour of $x$. So every element outside $X$ is joined to the even survivors, and $\Delta_r-X$ is connected. \end{proof} \subsection*{The theorem} \begin{theorem}[Prime powers]\label{thm:pp}\label{a7:pp} Let $p$ be a prime, $N\ge 6$ and $k=p^N$. Then \begin{condlist}{(a)} \item[\textup{(a)}] $\Mk(G)=\Fl(G,\Z_k)$ is connected for every finite bridgeless multigraph $G$ \textup{(}loops allowed\textup{)}; \item[\textup{(b)}] $\Ik(G)$ is connected for every finite bridgeless loopless multigraph $G$. \end{condlist} For example $k=64$, $128$, $256$, $729$, $15625$. \end{theorem} \begin{proof} We apply Theorem~\ref{thm:Z} with $\Gam=\Z_k$, $k=ab$, $B=a\Z_k$, $A=\Z_a$, and with the hypotheses of Theorem~\ref{thm:master}. The data of the two cases: \[ \begin{array}{l|l|l} & p=2 & p\text{ odd}\\ \hline a=|A| & 2^{N-2}\ \ (\ge 16) & p^4\ \ (\ge 81)\\ b=|B| & 4 & p^{N-4}\ \ (\ge p^2\ge 9)\\ i\text{ with }a=p^i & N-2 & 4\\ \Sig & \{\,\pm 2^j : 0\le j\le N-1\,\} & \{\pm 1,\pm 2\}\cup\{\,\pm p^j : 1\le j\le N-1\,\} \end{array} \] In both cases $B=a\Z_k=p^i\Z_k$, and $4\le i\le N-1$ \textup{(}for $p=2$ because $N\ge 6$\textup{)}. \begin{condlist}{(Counts)} \item[\textup{($\Sig$)}] \emph{A step system.} No element of $\Sig$ is $0$ in $\Z_k$: the integers $1$, $2$ and $p^j$ with $j\le N-1$ lie between $1$ and $k-1$. And $\Sig=-\Sig$. So $\Sig$ is a step system \textup{(}Definition~\ref{def:stepsystem}\textup{)}. \item[\textup{(Lift)}] \emph{Liftable.} The elements $\pm 1$ are liftable by Lemma~\ref{lem:liftable}~(b), $\pm 2$ by Lemma~\ref{lem:liftable}~(c), and $\pm p^j$ for $0\le j\le N-1$ by Lemma~\ref{lem:liftable}~(d), because $p^j$ is a divisor of $k$ with $1\le p^j2j\,|B|$. Let $G$ be a loopless $3$-edge-connected multigraph on at least two vertices, and for every edge $e$ of $G$ let $\Phi(e)$ be a set of at most $j$ elements of $\Gam$. Then $G$ has a $\Gam$-flow $w$ with $w(e)\notin\Phi(e)$ for every edge $e$. \end{lemma} \begin{proof} Use the frame and the notation above. Define a list assignment: $\Lam(t)=\Phi(t)+B=\{\,z+y : z\in\Phi(t),\ y\in B\,\}$ for every original blue edge $t$, and $\Lam(t)$ empty for every blue expansion edge. Then $|\Lam(t)|\le j\,|B|$, so $|\Lam(e_i)|+|\Lam(f_i)|\le 2j\,|B|<|\Gam|$ for every node $i$. By Lemma~\ref{lem:reverse}~(a) with $Y_0=0$ there is a $\Gam$-flow $Y$ of $Q$ with $Y(t)\notin\Lam(t)$ for every blue edge $t$. So $Y(t)+y\notin\Phi(t)$ for every original blue edge $t$ and every $y$ in $B$. Let $\eps$ be a free edge. The $|B|\ge j+1$ elements $Y(\eps)+y$, $y$ in $B$, are different, and at most $j$ of them are in $\Phi(\eps)$. Choose $y_\eps$ in $B$ with $Y(\eps)+y_\eps\notin\Phi(\eps)$. Put $Y'=Y+\sum_\eps y_\eps\chi_{F_\eps}$, the sum over all free edges. A free edge $\eps$ lies on $F_\eps$ and on no other cycle $F_{\eps'}$, so $Y'(\eps)=Y(\eps)+y_\eps\notin\Phi(\eps)$. On a blue edge $t$ the difference $Y'(t)-Y(t)$ is a sum of elements $\pm y_\eps$, so it is in $B$; hence $Y'(t)\notin\Phi(t)$ if $t$ is original. Every original edge is blue or free. By Lemma~\ref{a2:flows}~(c) the restriction $w$ of $Y'$ to the original edges is a $\Gam$-flow of $G$, and $w(e)\notin\Phi(e)$ for every edge $e$ of $G$. \end{proof} \begin{lemma}[Cyclic list avoidance]\label{lem:Bt}\label{a8:lem:Bt} Let $j\ge 1$ and $N>2j(2j+1)$ be integers. Let $G$ be a loopless $3$-edge-connected multigraph on at least two vertices, and for every edge $e$ of $G$ let $\Phi(e)$ be a set of at most $j$ elements of $\Z_N$. Then $G$ has a $\Z_N$-flow $w$ with $w(e)\notin\Phi(e)$ for every edge $e$. \end{lemma} \begin{proof} Use the frame and the notation above, with $\Gam=\Z_N$, and put $d=j+1$. For every original blue edge $t$ let $\Lam(t)$ be the set of the elements $z-y$ of $\Z_N$ with $z$ in $\Phi(t)$ and an integer $y$ with $|y|\le j$; for a blue expansion edge $\Lam(t)$ is empty. Then $|\Lam(t)|\le j(2j+1)$, so $|\Lam(e_i)|+|\Lam(f_i)|\le 2j(2j+1)6d=2\cdot 3\cdot|B|$. Lemma~\ref{lem:Ct} with $j=3$ gives $w_0$ with $w_0(t)\notin\Phi(t)$ for every edge $t$ of $H_0$. If $N\ge 43$, then $N>42=2\cdot 3\cdot(2\cdot 3+1)$, and Lemma~\ref{lem:Bt} with $j=3$ gives such a $w_0$. The flow of $H$ that has the value $w_0(t)$ on $\thr{t}$ avoids $z$. \end{proof} \subsection*{The theorem of two coprime factors}\label{a8:theorem-sec} From here on $a$, $b$ and $k=ab$ are as in the subsection `Guarded moves'. \begin{theorem}[Two coprime factors]\label{thm:twocoord}\label{a8:thm:twocoord} Let $a\ge 9$ and $b\ge 6$ be coprime integers such that $\Z_a$ and $\Z_b$ have one-value avoidance, and let $k=ab$. \begin{condlist}{(a)} \item[\textup{(a)}] On every bridgeless multigraph $G$ \textup{(}loops allowed\textup{)} any two nowhere-zero $\Z_k$-flows are joined by a sequence of guarded moves. In particular $\Mk(G)$ is connected. \item[\textup{(b)}] $\Ik(G)$ is connected for every finite bridgeless loopless multigraph $G$. \end{condlist} \end{theorem} \begin{proof} A \emph{target} of a multigraph is a $\Z_k$-flow $h$ such that $\rho_a(h)$ and $\rho_b(h)$ are both nowhere zero. A target is nowhere zero. \smallskip \noindent\textsc{Claim.} On every connected bridgeless multigraph $G$ \textup{(}loops allowed\textup{)}, every nowhere-zero $\Z_k$-flow is joined to every target by a sequence of guarded moves. \begin{condlist}{(4)} \item[\textup{(1)}] \emph{The claim gives the theorem.} Let $G$ be bridgeless. By Theorem~\ref{q:flows} \textup{(}Seymour, Tutte\textup{)}, every bridgeless multigraph, loops allowed, has a nowhere-zero flow with values in any abelian group of order at least $6$. So $G$ has a nowhere-zero $\Z_a$-flow $h_a$ and a nowhere-zero $\Z_b$-flow $h_b$, because $a\ge 9$ and $b\ge 6$. Then $h=\iota_a(h_a)+\iota_b(h_b)$ is a target. A cycle of $G$ lies in one component of $G$, and a guarded move of a component is a guarded move of $G$. So, component after component, the claim joins every nowhere-zero $\Z_k$-flow of $G$ to $h$ by guarded moves. The reverse of a guarded move is a guarded move. This gives (a), because guarded moves are moves of $\Mk(G)$. For (b) let $G$ be loopless, and let $F$, $F'$ be vertices of $\Ik(G)$. Their reductions $f$, $f'$ modulo $k$ are nowhere-zero $\Z_k$-flows, because $1\le|F(e)|\le k-1$. By (a) a sequence of guarded moves leads from $f$ to $f'$. By Lemma~\ref{a8:lem:guarded}~(b), all integer lifts of two consecutive flows of the sequence lie in one component of $\Ik(G)$. Every flow of the sequence has an integer lift, and all integer lifts of one flow lie in one component (Lemma~\ref{lem:liftable}~(a); $k\ge 54$). So all of them lie in one component, and it contains $F$ and $F'$. \item[\textup{(2)}] \emph{The induction.} We prove the claim by induction on the number $n$ of vertices of $G$; for each $n$ first for the loopless graphs, then for the graphs with loops. A connected loopless multigraph with one vertex has no edge and only the empty flow. A connected bridgeless loopless multigraph with $n\ge 2$ vertices has a cut with exactly $2$ edges or is $3$-edge-connected: a cut with one edge consists of a bridge. So three cases remain: \textup{(I)}, \textup{(II)}, \textup{(III)}. In all of them $f$ is a nowhere-zero $\Z_k$-flow and $h$ is a target of $G$. \item[\textup{(I)}] \emph{$G$ has a loop.} Let $G^-$ be $G$ without its loops. It is connected, bridgeless and loopless, and it has $n$ vertices; so the claim holds for $G^-$. The restrictions of $f$ and $h$ to $G^-$ are a nowhere-zero flow and a target. A guarded move of $G^-$ is a guarded move of $G$: its cycle has no loop, and the values on the loops do not change. So $f$ is joined by guarded moves to the flow that is equal to $h$ on $E(G^-)$ and to $f$ on the loops. Now let $e$ be a loop, directed, with the values $x=f(e)\ne 0$ and $x'=h(e)$; a loop is a cycle with one edge. If $\rho_b(x)\ne 0$, first add $\iota_a(\rho_a(x')-\rho_a(x))\,\chi_e$: a guarded $a$-move \textup{(}if it is not $0$\textup{)}, because $\rho_b(x)\ne 0$; the new value $x_1$ has $\rho_a(x_1)=\rho_a(x')\ne 0$. Then add $\iota_b(\rho_b(x')-\rho_b(x_1))\,\chi_e$: a guarded $b$-move, because $\rho_a(x_1)\ne 0$. The new value is $x'$. If $\rho_b(x)=0$, then $\rho_a(x)\ne 0$, and the two moves are made in the other order: first the $b$-move, guarded by $\rho_a(x)\ne 0$, then the $a$-move, guarded by $\rho_b(x')\ne 0$. No other edge changes. Do this for every loop. The result is $h$. \item[\textup{(II)}] \emph{$G$ is loopless and has a cut with exactly two edges $e_1$, $e_2$.} Let $G/e_1$ be obtained from $G$ by the identification of the two ends of $e_1$ and the deletion of $e_1$. It is connected and has $n-1$ vertices. It is bridgeless: if an edge $e\ne e_1$ lies on a cycle $O$ of $G$, then the edges of $O$ other than $e_1$ form a closed trail of $G/e_1$ through $e$, and every edge of a closed trail lies on a cycle of the trail. It can have loops. By induction the claim holds for $G/e_1$. The restriction of a flow of $G$ to $E(G/e_1)$ is a flow of $G/e_1$ \textup{(}add the two conservation equations at the ends of $e_1$\textup{)}, and every flow of $G/e_1$ is the restriction of exactly one flow of $G$: conservation at one end of $e_1$ fixes the value on $e_1$. In a flow of $G$, with values in any abelian group, the value on $e_1$ is the value on $e_2$ or its negative \textup{(}conservation over one side of the cut\textup{)}. So a flow of $G$ is nowhere zero, or a target, if and only if its restriction is. Let $g'\to g'+\lambda\chi_{C'}$ be a guarded move of $G/e_1$. The edges of $C'$ form a cycle of $G$, or a path of $G$ between the two ends of $e_1$; in the second case add $e_1$. This gives a cycle $C$ of $G$ with $\chi_C=\chi_{C'}$ on $E(G/e_1)$. In the second case the path crosses the cut and does not use $e_1$, so $C'$ contains $e_2$. Let $g$ be the flow of $G$ with restriction $g'$. Then $g+\lambda\chi_C$ is the flow of $G$ with restriction $g'+\lambda\chi_{C'}$, and $g\to g+\lambda\chi_C$ is a guarded move of the same kind: the residue that guards is not $0$ on the edges of $C'$, and, if $e_1$ is on $C$, also on $e_1$, because its value there is $\pm$ its value on $e_2$. So the claim for $G/e_1$ gives the claim for $G$. \item[\textup{(III)}] \emph{$G$ is loopless and $3$-edge-connected, $n\ge 2$.} Let $m$ be the number of edges. Every vertex has degree at least $3$, so $m\ge 3n/2$. By Lemma~\ref{lem:avgcoord}, $f$ is joined by guarded $a$-moves to a flow $f'$ such that $\rho_a(f')$ has at most $m/a\le m/9$ zeros. So $\supp(\rho_a(f'))$ has at least $(8/9)\,m\ge(4/3)\,n>(4/3)\,n-2$ edges. If $n\ge 3$, Lemma~\ref{lem:extraction}~(a) gives a good subgraph $H$ of the loopless multigraph $(V(G),\supp(\rho_a(f')))$. If $n=2$, then $m\ge 3$, and the support has at least $8m/9>2$ edges; two of them are parallel and form a good subgraph $H$ of the kind (CYC). In both cases $H$ is connected, it has at least two vertices, and $\rho_a(f')$ has no zero on $E(H)$. Let $G^*=G/H$ be obtained from $G$ by the identification of $V(H)$ to one vertex $v_H$ and the deletion of $E(H)$. It has fewer vertices than $G$, and it is connected. It is bridgeless: if an edge $e$ of $G^*$ lies on a cycle $O$ of $G$, the edges of $O$ outside $E(H)$ form a closed trail of $G^*$ through $e$, and every edge of a closed trail lies on a cycle of the trail. It can have loops. By induction the claim holds for $G^*$. The restriction of a flow of $G$ to $E(G^*)$ is a flow of $G^*$ \textup{(}add the conservation equations at the vertices of $H$; the edges of $H$ drop out\textup{)}. So the restriction of $f'$ is a nowhere-zero flow of $G^*$, and the restriction $h^*$ of $h$ is a target of $G^*$. So there are guarded moves of $G^*$ that lead from the restriction of $f'$ through flows $g_0,g_1,\dots,g_\nu=h^*$. \emph{The cycles.} For a cycle $C^*$ of $G^*$ we fix a cycle $C$ of $G$ with $\chi_C=\chi_{C^*}$ on $E(G^*)$ and $E(C)\subseteq E(C^*)\cup E(H)$. If $C^*$ does not pass $v_H$, its edges form a cycle of $G$; take it. Otherwise $C^*$ passes $v_H$ once, and its edges form, in $G$, a cycle or a path $P$ between two different vertices $v_1$, $v_2$ of $H$ whose inner vertices are not in $V(H)$; a loop of $G^*$ at $v_H$ is such a path with one edge, because $G$ is loopless. Add a path of $H$ between $v_2$ and $v_1$; $H$ is connected. Direct $C$ along $C^*$. \emph{The invariant.} We build flows $x_0=f',x_1,\dots,x_\nu$ of $G$, each joined to the one before it by guarded moves, such that the restriction of $x_r$ to $E(G^*)$ is $g_r$ and $\rho_a(x_r)$ has no zero on $E(H)$. Such a flow is nowhere zero. $x_0=f'$ has these properties. Let $x=x_{r-1}$ be built, and let $g_r=g_{r-1}+\lambda\chi_{C^*}$, with the cycle $C$ of $C^*$. \emph{A guarded $b$-move of $G^*$} \textup{(}$\lambda$ in $a\Z_k$, $\rho_a(g_{r-1})$ has no zero on $C^*$\textup{)}. Then $\rho_a(x)$ has no zero on $E(C)$: not on $E(H)$ by the invariant, and not on $E(C^*)$, where $x$ is equal to $g_{r-1}$. So $x\to x+\lambda\chi_C$ is a guarded $b$-move. Put $x_r=x+\lambda\chi_C$. It has the residues of $x$ modulo $a$, and its restriction to $E(G^*)$ is $g_r$. \emph{A guarded $a$-move of $G^*$} \textup{(}$\lambda$ in $b\Z_k$, $\rho_b(g_{r-1})$ has no zero on $C^*$\textup{)}. \begin{condlist}{(iii)} \item[\textup{(i)}] $\Z_b$ has one-value avoidance. So there is a $\Z_b$-flow $w_b$ of $H$ with $w_b(e)\ne-\rho_b(x)(e)$ for every edge $e$ of $H$. Extend it by $0$ to a flow of $G$. By Lemma~\ref{a8:lem:guarded}~(a), with $a$ and $b$ exchanged and $E'=E(H)$ \textup{(}$\rho_a(x)$ has no zero there\textup{)}, $x$ is joined by guarded $b$-moves to $x'=x+\iota_b(w_b)$. Now $\rho_b(x')$ has no zero on $E(H)$; $\rho_a(x')=\rho_a(x)$; and $x'=x$ outside $E(H)$. \item[\textup{(ii)}] $\Z_a$ has one-value avoidance. So there is a $\Z_a$-flow $w_a$ of $H$ with $w_a(e)\ne-\rho_a(x')(e)-\rho_a(\lambda)\chi_C(e)$ for every edge $e$ of $H$. Extend it by $0$. By Lemma~\ref{a8:lem:guarded}~(a) with $E'=E(H)$ \textup{(}$\rho_b(x')$ has no zero there\textup{)}, $x'$ is joined by guarded $a$-moves to $x''=x'+\iota_a(w_a)$. Now $\rho_b(x'')=\rho_b(x')$, and $x''=x$ outside $E(H)$. \textup{(}$\rho_a(x'')$ can have zeros on $E(H)$.\textup{)} \item[\textup{(iii)}] $\rho_b(x'')$ has no zero on $E(C)$: not on $E(H)$ by (i), and not on $E(C^*)$, where $x''$ is equal to $g_{r-1}$. So $x''\to x''+\lambda\chi_C$ is a guarded $a$-move. Put $x_r=x''+\lambda\chi_C$. Its restriction to $E(G^*)$ is $g_{r-1}+\lambda\chi_{C^*}=g_r$. On $E(H)$, $\rho_a(x_r)=\rho_a(x')+w_a+\rho_a(\lambda)\chi_C$, which has no zero by the choice of $w_a$. \end{condlist} \emph{The end.} The flow $x_\nu$ is equal to $h$ outside $E(H)$, and $\rho_a(x_\nu)$ has no zero on $E(H)$. So $\rho_a(x_\nu)$ has no zero at all, because $\rho_a(h)$ has none. By Lemma~\ref{a8:lem:guarded}~(a), with $a$ and $b$ exchanged and $E'=E(G)$, $x_\nu$ is joined by guarded $b$-moves to the flow $x'$ with $\rho_b(x')=\rho_b(h)$ and $\rho_a(x')=\rho_a(x_\nu)$ \textup{(}add $\iota_b$ of the $\Z_b$-flow $\rho_b(h)-\rho_b(x_\nu)$\textup{)}. Now $\rho_b(x')=\rho_b(h)$ has no zero. By Lemma~\ref{a8:lem:guarded}~(a) with $E'=E(G)$, $x'$ is joined by guarded $a$-moves to the flow with the residues $\rho_a(h)$ and $\rho_b(h)$, that is to $h$. Together: $f$ is joined to $f'$, to $x_\nu$, and to $h$ by guarded moves.\qedhere \end{condlist} \end{proof} \begin{corollary}[Two admitted factors]\label{cor:enlarged}\label{a8:cor:enlarged} Call an integer $N\ge 2$ an \emph{admitted factor} if $N=de$ with integers $d\ge 4$ and $e\ge 7$, or if $N\ge 43$. \begin{condlist}{(a)} \item[\textup{(a)}] Let $k=ab$ with coprime admitted factors $a$ and $b$. Then $\Mk(G)$ is connected for every finite bridgeless multigraph $G$ \textup{(}loops allowed\textup{)}, and $\Ik(G)$ is connected for every finite bridgeless loopless multigraph $G$. \item[\textup{(b)}] Every multiple of a product of two coprime admitted factors is again such a product. \item[\textup{(c)}] In particular \textup{(a)} holds for every multiple of $1120=32\cdot 35$, for $k=pq$ with different primes $p,q\ge 43$, and for $k=p^2q$ with different primes $p\ge 7$ and $q\ge 43$. \end{condlist} \end{corollary} \begin{proof} (a) An admitted factor is at least $4\cdot 7=28$. So $a\ge 9$ and $b\ge 6$. By Corollary~\ref{cor:Cprime}, $\Z_a$ and $\Z_b$ have one-value avoidance. Theorem~\ref{thm:twocoord} gives the claim. (b) Let $k=ab$ with coprime admitted factors $a$, $b$, and let $k'$ be a multiple of $k$. For every prime $p$ that divides $k'/k$, let $p^{v}$ be the largest power of $p$ that divides $k'/k$. Put it to $b$ if $p$ divides $b$, and to $a$ otherwise. This gives $k'=a'b'$, where $a'$ is a multiple of $a$ and $b'$ is a multiple of $b$. A prime that divides $a'$ divides $a$ or was put to $a$; in both cases it does not divide $b$, and it was not put to $b$. So $a'$ and $b'$ are coprime. A multiple of an admitted factor is an admitted factor: a multiple of $de$ is $de'$ with a multiple $e'$ of $e$, so $e'\ge 7$; and a multiple of a number $N\ge 43$ is at least $43$. (c) $32=4\cdot 8$ and $35=5\cdot 7$ are coprime admitted factors; use (b). A prime $p\ge 43$ is an admitted factor. For a prime $p\ge 7$, $p^2=p\cdot p$ with $p\ge 4$ and $p\ge 7$ is an admitted factor. \end{proof} \begin{remark}[The admitted factors]\label{a8:rem:admitted} The admitted factors below $43$ are $28$, $32$, $35$, $36$, $40$, $42$. So the admitted factors are the integers $N\ge 28$ other than $29$, $30$, $31$, $33$, $34$, $37$, $38$, $39$, $41$. The smallest product of two coprime admitted factors is $1120$. Indeed, let $ab<1120$ with admitted factors $a\le b$. Then $a^2<1120$ gives $a\le 33$, so $a$ is $28$ or $32$, and $b<1120/28=40$ gives $b$ in $\{28,32,35,36\}$; of these pairs only $32$, $35$ is coprime, and $32\cdot 35$ is not smaller than $1120$. Every multiple of $1120$ is also a multiple of $4$ from $52$ on, so Theorem~\ref{thm:bases} gives it as well. The two lemmas give one-value avoidance only for the admitted factors; other orders are not treated. \end{remark} \begin{remark}[The master theorem for these moduli]\label{a8:rem:master} Let $a\ge 10$ and $b\ge 4$ be coprime and $k=ab$. The connectivity of $\Mk(G)$ in Theorem~\ref{thm:twocoord}~(a) then also follows from Theorem~\ref{thm:master}. Put $\Gam=\Z_k$, $B=a\Z_k$ \textup{(}order $b$\textup{)}, $A=\Z_a$, and $\Sig=\Gam\setminus\{0\}$. Then $\SigA=A\setminus\{0\}$ and $\SigB=B\setminus\{0\}$, so \RI, \RII{} and \RIII{} hold with walks of one step. By the Chinese remainder theorem $\Z_k=\Z_a\times\Z_b$; the element $b$ of $\Sig$ generates the factor $\Z_a\times 0$ and the element $a$ generates $0\times\Z_b$, and $(|B|-1)/(|\Gam|-1)=(b-1)/(ab-1)<1/9$ because $a\ge 10$. So Lemma~\ref{lem:AVG} gives \AV{} and \AVpp. And $|A|=a\ge 10$, $|B|=b\ge 4$: Corollary~\ref{cor:i} applies. This is a special case of Theorem~\ref{thm:G10}, for $\Gam=\Z_k$. Theorem~\ref{thm:Z} does not give the integer graphs $\Ik(G)$ in this way, because it asks that every element of $\Sig=\Gam\setminus\{0\}$ is liftable (Definition~\ref{def:liftable}), and this is not shown. For $k=pq$ a smaller $\Sig$ does not help either. Let $k=pq$ with different primes $p,q\ge 5$. The elements that Lemma~\ref{lem:liftable} shows to be liftable are $\pm 1$, $\pm 2$, $\pm p$, $\pm q$: in part~(e), $c=1$ modulo $p$ and modulo $q$ means $c=1$ modulo $k$. With $B=p\Z_k$, the only ones of them in $B$ are $\pm p$. The map $jp\mapsto j$ identifies $B$ with $\Z_q$ and $\pm p$ with the steps $\pm 1$, and $\Z_q$ without the two elements $1$, $3$ is not connected by these steps. So \RII{} fails; the same holds for $B=q\Z_k$. The guarded moves do not need liftable steps: Lemma~\ref{a8:lem:guarded}~(b) lifts every guarded move, whatever its step. \end{remark} \section{Every multiple of four from 52 on; the moduli by their number of prime factors}\label{a9:sec} This section collects the integer moduli. Theorem~\ref{thm:bases} gives $k=40$, $k=48$ and every multiple of $4$ from $52$ on. Theorem~\ref{thm:omega4} gives every $k$ with at least four prime factors that is not a prime power, except $24$, $36$ and $54$. Proposition~\ref{prop:omega3} and Corollary~\ref{a9:cor:table} say exactly which $k$ these results reach \textup{(}Definition~\ref{a9:def:covered}\textup{)}, by the number of prime factors. Corollary~\ref{cor:density}: the moduli $k$ for which $\Mk(G)$ and $\Ik(G)$ are connected for every finite bridgeless loopless multigraph $G$ have natural density one. $\Omega(k)$ is the number of prime factors of $k$ counted with multiplicity, $\omega(k)$ the number of different prime factors, and $\tau(a)$ the number of positive divisors of $a$. For a set $D$ of non-zero elements of $\Z_a$ with $D=-D$, the \emph{graph on $\Z_a$ with the step set $D$} has the vertex set $\Z_a$; two vertices $x$, $y$ are adjacent if $y-x$ is in $D$. The \emph{sign class} of an element $x$ is $\{x,-x\}$. For a split $k=ab$, $D$ always denotes the step set of Criterion~\ref{crit:F10}: the non-zero residues modulo $a$ of $\pm 1$, $\pm 2$ and $\pm q$, for all divisors $q$ of $k$ with $1\le q15$, then $m\ge 21$ and $2d_0\le 2\sqrt{m}q$: for $pq,r$. Put $a=pq\ge 15$ and $b=pr\ge 15$. The eight residues $\pm 1$, $\pm 2$, $\pm p$, $\pm q$ modulo $pq$ are different. The residues $\pm r$ are two more: $r$ is a unit modulo $pq$, so not $\pm p$, $\pm q$; and $3\le rp$\textup{)}. So $|D|\ge 10$. \emph{Type $uvwz$, $u2$; with $6\le b4$, Chebyshev's inequality \textup{(}Theorem~\ref{q:chebyshev}\textup{)} gives \[ \delta_y=\Pr[X\le 4]\ \le\ \Pr\bigl[\,|X-\mu_y|\ge\mu_y-4\,\bigr]\ \le\ \frac{\mu_y}{(\mu_y-4)^2}. \] The sum of $1/p$ over all primes diverges \textup{(}Euler; Theorem~\ref{q:euler}\textup{)}, so $\mu_y$ tends to infinity, and $\delta_y$ tends to $0$. For every $y$ the upper density of $T$ is at most $\delta_y$. So $T$ has density $0$, and $S^*$ has density $1$. \end{proof} \begin{remark}[Moduli that are not covered]\label{a9:rem:notcovered} The moduli that are not covered include the primes, the prime powers $p^e$ with $e\le 5$, the values $24$, $36$, $54$ and the eleven products of three different primes of Proposition~\ref{prop:omega3}. For $k\ge 6$ in these cases, the connectivity of $\Ik(G)$ for all graphs is open \textup{(}Section~\ref{a12:sec}\textup{)}. For example $k=44=4\cdot 11$ is not covered (Proposition~\ref{prop:omega3}~(i): $4q$ is covered only for $q\ge 13$): it is smaller than $52$ and is not $40$ or $48$; its only split for Criterion~\ref{crit:F10} is $a=11$, $b=4$, with $D=\{\pm 1,\pm 2,\pm 4\}$ modulo $11$; and $44$ is not the product of two coprime admitted factors. \textup{(}The modular graph $\Mk(G)$ with $k=44$ is connected by Theorem~\ref{thm:G10}.\textup{)} \end{remark} \section{One graph: larger moduli, few edges, few vertices, short cycles}\label{a10:sec} Theorem~\ref{thm:bases} gives moduli $k$ for which $\Ik(G)$ and $\Mk(G)$ are connected for \emph{every} finite bridgeless loopless multigraph $G$: $k=40$, $k=48$ and every multiple of $4$ from $52$ on. We call them the \emph{bases}. This section fixes one graph $G$ and gives \emph{every} $k$ from a bound on, prime numbers $k$ included. The bound depends on $G$: \begin{condlist}{(d)} \item[\textup{(a)}] on the length of a longest cycle: every $k\ge\max(48,c(G)+4)$, and every $k\ge 40$ if every cycle has at most $39$ edges (Corollary~\ref{a10:pergraph}); \item[\textup{(b)}] on the number of edges: every $k\ge 40$ if $G$ has at most $189$ edges (Corollary~\ref{a10:C189}); \item[\textup{(c)}] on the number of vertices: every $k\ge 40$ if $G$ has at most $149$ vertices (Corollary~\ref{a10:N149}); \item[\textup{(d)}] on short cycles: every $k\ge 40$ if every edge lies on a cycle with at most $8$ edges, or if $G$ is connected with diameter at most $3$ (Corollaries~\ref{a10:SC1}, \ref{a10:diam}, \ref{a10:SC2}). \end{condlist} All four go up one level at a time, from $k-1$ to $k$, starting at a base. This section does \emph{not} prove that $\Ik(G)$ is connected for every bridgeless $G$ and every large $k$. \subsection*{Setting} In this section $G$ is a finite bridgeless loopless multigraph, and $k\ge 3$ is an integer, unless something else is said. The conventions are those of Section~\ref{sec:flows}. In particular $\Ik(G)$ is the graph on the integer flows $F$ of $G$ with $1\le|F(e)|\le k-1$ for every edge $e$, and $F$, $F'$ are adjacent if $F'-F=\lambda\chi_C$ for a cycle $C$ with a direction and an integer $\lambda\ne 0$ (a move). For a flow $F$ of $\Ik(G)$, $G_F$ is the orientation of $G$ in which $F$ is positive: every edge is directed so that its value is in $[1,k-1]$. We change the reference orientation freely (Section~\ref{sec:flows}). In a statement about one flow $F$ the values are read in $G_F$; an edge of $G$ with this direction is an \emph{arc}, and its \emph{value} $F(e)$ is its size $|F(e)|$. An arc is a \emph{top} if its value is $k-1$, a \emph{$1$-arc} if its value is $1$, and a \emph{middle arc} if its value is in $[2,k-2]$. For a set $X$ of vertices, $\operatorname{in}(X)$ and $\operatorname{out}(X)$ are the numbers of arcs of $G_F$ that enter $X$ and that leave $X$; $\operatorname{in}_v(X)$ and $\operatorname{out}_v(X)$ count only the arcs of value $v$. The sum of the conservation equations over the vertices of $X$ gives \[ \sum_{e\text{ leaves }X}F(e)\;=\;\sum_{e\text{ enters }X}F(e)\qquad\text{(the flow out of $X$ is the flow into $X$).} \] By Lemma~\ref{lem:cycledec}~(b), applied to $F$ with the reference orientation $G_F$, every arc of $G_F$ lies on a \emph{directed cycle} of $G_F$, that is a cycle that passes all its arcs in their direction. \emph{Joined} means: in the same component of $\Ik(G)$, for the $k$ of the context. A flow of $\Ik(G)$ \emph{descends} if it is joined to a flow of $\Ik[k-1](G)$. $c(G)$ is the number of edges of a longest cycle of $G$, and $c(G)=0$ if $G$ has no cycle. For an integer $k_0\ge 2$ put \[ \sigma(k_0)=\sum_{\lambda=1}^{k_0-1}\Bigl\lceil\frac{k_0}{\lambda}\Bigr\rceil . \] One quoted theorem is used: Hoffman's circulation theorem, Theorem~\ref{q:hoffman}. \begin{lemma}[One level]\label{a10:level} Let $G$ be any finite multigraph and $k\ge 3$. \begin{condlist}{(b)} \item[\textup{(a)}] The vertices of $\Ik[k-1](G)$ are the flows of $\Ik(G)$ with all sizes at most $k-2$, and two of them are adjacent in $\Ik[k-1](G)$ if and only if they are adjacent in $\Ik(G)$. \item[\textup{(b)}] If $\Ik[k-1](G)$ is connected and every flow of $\Ik(G)$ descends, then $\Ik(G)$ is connected. \end{condlist} \end{lemma} \begin{proof} (a) holds by the definitions: the adjacency is the same rule in both graphs. (b) If $\Ik[k-1](G)$ has no vertex, then no flow of $\Ik(G)$ descends, so $\Ik(G)$ has no vertex and is connected. Otherwise, by (a), the connected graph $\Ik[k-1](G)$ lies in one component of $\Ik(G)$, and every flow of $\Ik(G)$ is joined to it. \end{proof} \begin{lemma}[From integers to residues]\label{a10:image} Let $G$ be any finite multigraph and $k\ge 2$. If $\Ik(G)$ is connected, then $\Mk(G)$ is connected. So each statement of this section that $\Ik(G)$ is connected gives that $\Mk(G)$ is connected. \end{lemma} \begin{proof} Let $f$, $f'$ be vertices of $\Mk(G)$. By Corollary~\ref{cor:lifting} they have integer lifts $F$, $F'$. Take a path $F=F_0,F_1,\dots,F_\nu=F'$ in $\Ik(G)$, and let $f_r$ be the reduction of $F_r$ modulo $k$. It is a nowhere-zero $\Z_k$-flow, because $1\le|F_r(e)|\le k-1$. If $F_r-F_{r-1}=\lambda\chi_C$, then $f_r-f_{r-1}$ is $\lambda\chi_C$ modulo $k$: either $f_r=f_{r-1}$, or $f_{r-1}$ and $f_r$ are adjacent in $\Mk(G)$. So $f=f_0$ and $f'=f_\nu$ are joined in $\Mk(G)$. \end{proof} \begin{lemma}[Same orientation]\label{a10:sameor} Let $F$ and $T$ be flows of $\Ik(G)$ with $G_T=G_F$. Then $F$ and $T$ are joined, by moves of the form $\chi_O$. \end{lemma} \begin{proof} Use the reference orientation $G_F$. By Lemma~\ref{lem:cycledec}~(b), $T-F=\chi_{O_1}+\dots+\chi_{O_j}$, where every directed cycle $O_r$ passes its edges in the direction of the sign of $T-F$. So on every edge $e$, every partial sum $F+\chi_{O_1}+\dots+\chi_{O_r}$ has a value between $F(e)$ and $T(e)$, that is in $[1,k-1]$. All partial sums are in $\Ik(G)$, and two consecutive ones differ by a move. \end{proof} \subsection*{Larger moduli on one graph} \begin{lemma}[Top-value reduction]\label{lem:top}\label{a10:top} Let $F$ be a flow of $\Ik(G)$ with a top $e$, and let $C$ be a directed cycle of $G_F$ through $e$. Let $\lambda$ be an integer with $1\le\lambda\le k-2$ that is not the value of any arc of $C$. Then $F-\lambda\chi_C$ is a flow of $\Ik(G)$ that is adjacent to $F$ \textup{(}$\chi_C$ is read in $G_F$, so it is $1$ on the arcs of $C$\textup{)}, and its tops are the tops of $F$ that are not on $C$. \end{lemma} \begin{proof} An arc of $C$ has a value $v$ with $1\le v\le k-1$ and $v\ne\lambda$. Its new value $v-\lambda$ is not $0$, and $1-\lambda\le v-\lambda\le k-1-\lambda$; so $|v-\lambda|\le\max\{\lambda-1,k-1-\lambda\}\le k-2$. The arcs outside $C$ keep their values. So $F-\lambda\chi_C$ is in $\Ik(G)$, no arc of $C$ is a top of it, and the tops outside $C$ do not change. It is adjacent to $F$ because $\lambda\ne 0$. \end{proof} \begin{theorem}[Larger moduli on one graph]\label{thm:L}\label{a10:L} Let $k_0\ge 2$ be an integer with $k_0\ge c(G)+1$ such that $\Ik[k_0](G)$ is connected. Then $\Ik(G)$ and $\Mk(G)$ are connected for every $k\ge k_0$. \end{theorem} \begin{proof} Induction on $k$; for $k=k_0$ this is the hypothesis and Lemma~\ref{a10:image}. Let $k>k_0$, so $k\ge 3$, and let $\Ik[k-1](G)$ be connected. A flow of $\Ik(G)$ without a top is in $\Ik[k-1](G)$. Let $F$ be a flow of $\Ik(G)$ with a top $e$. Take a directed cycle $C$ of $G_F$ through $e$. It has at most $c(G)\le k_0-1\le k-2$ arcs, so at most $k-3$ arcs other than $e$. The value of $e$ is $k-1$. So at most $k-3$ of the $k-2$ integers $1,\dots,k-2$ are values of arcs of $C$, and some $\lambda$ in $[1,k-2]$ is not. By Lemma~\ref{a10:top}, $F$ is adjacent to a flow with fewer tops. Repeat this with the new flow and its own orientation. The number of tops goes down at every step; so $F$ is joined to a flow without tops, that is a flow of $\Ik[k-1](G)$. So every flow of $\Ik(G)$ descends, and $\Ik(G)$ is connected by Lemma~\ref{a10:level}~(b). $\Mk(G)$ is connected by Lemma~\ref{a10:image}. \end{proof} \begin{corollary}[One graph: the bound by the longest cycle]\label{cor:pergraph}\label{a10:pergraph} Let $G$ be a finite bridgeless loopless multigraph, and $c=c(G)$. Then $\Ik(G)$ and $\Mk(G)$ are connected for every $k\ge\max(48,\,c+4)$. If $c\le 39$, they are connected for every $k\ge 40$. Since $c\le|V(G)|$, this holds in particular for every $k\ge\max(48,\,|V(G)|+4)$. \end{corollary} \begin{proof} Let $k_0$ be the least multiple of $4$ with $k_0\ge\max(48,c+1)$. It is a base, so $\Ik[k_0](G)$ is connected by Theorem~\ref{thm:bases}, and $k_0\ge c+1$. If $c+1\le 48$, then $k_0=48$; otherwise $k_0\le(c+1)+3=c+4$. So $k_0\le\max(48,c+4)$, and Theorem~\ref{a10:L} gives every $k\ge k_0$. If $c\le 39$, take the base $k_0=40\ge c+1$ in Theorem~\ref{a10:L}. \end{proof} \subsection*{One missing value; few edges} \begin{theorem}[One missing value]\label{thm:V}\label{a10:V} Let $\lambda$ be an integer with $1\le\lambda\le k-2$, and let $F$ be a flow of $\Ik(G)$ with no arc of size $\lambda$. Then $F$ is joined, by moves of the form $-\lambda\chi_O$ with directed cycles $O$ of $G_F$, to a flow with all sizes at most $k-2$. In particular $F$ descends. \end{theorem} \begin{proof} \emph{Claim \textup{($*$)}. Let $F$ be any flow of $\Ik(G)$ and $1\le\lambda\le k-2$. Let $Y$ be an integer flow of $G$ with values in $\{0,1\}$ in the orientation $G_F$, with $Y(e)=1$ on every top and $Y(e)=0$ on every arc of value $\lambda$. Then $F$ is joined, by moves $-\lambda\chi_O$ with directed cycles $O$ of $G_F$, to a flow with all sizes at most $k-2$.} Proof of the claim. Use the reference orientation $G_F$, and write $v(e)=F(e)$, so $1\le v(e)\le k-1$. By Lemma~\ref{lem:cycledec}~(b), $Y=\chi_{O_1}+\dots+\chi_{O_\nu}$ with pairwise edge-disjoint directed cycles $O_1,\dots,O_\nu$ of $G_F$ \textup{(}$Y$ has values in $\{0,1\}$\textup{)}, and every arc with $Y(e)=1$ is on exactly one of them. Put $F_r=F-\lambda\,(\chi_{O_1}+\dots+\chi_{O_r})$ for $r=0,\dots,\nu$. An arc on one of $O_1,\dots,O_r$ has in $F_r$ the value $v(e)-\lambda$. It is not $0$, because $Y(e)=1$ and so $v(e)\ne\lambda$. And $v(e)-\lambda\le k-1-\lambda\le k-2$, $\lambda-v(e)\le \lambda-1\le k-3$. Every other arc keeps its value. So every $F_r$ is in $\Ik(G)$, and $F_r-F_{r-1}= -\lambda\chi_{O_r}$ is a move. In $F_\nu$ an arc on some $O_r$ has a size at most $k-2$, and an arc on no $O_r$ has $Y(e)=0$; so it is not a top of $F$, and its size $v(e)$ is at most $k-2$. This proves the claim. Now let $F$ have no arc of size $\lambda$. In the orientation $G_F$ put $x(e)=F(e)/(k-1)$. Then $0(\text{sum of $u$ over the arcs that enter $X$}), \] that is $\operatorname{out}_{k-1}(X)>\operatorname{in}(X)-\operatorname{in}_\lambda(X)$. All values are in $[1,k-1]$, and the flow out of $X$ is the flow into $X$: \[ (k-1)\operatorname{out}_{k-1}(X)\ \le\ \lambda\operatorname{in}_\lambda(X)+(k-1)\bigl(\operatorname{in}(X) -\operatorname{in}_\lambda(X)\bigr)\ \le\ \lambda\operatorname{in}_\lambda(X)+(k-1)\bigl(\operatorname{out}_{k-1}(X)-1\bigr). \] So $\lambda\operatorname{in}_\lambda(X)\ge k-1$. \end{proof} \begin{corollary}[The count of the edges]\label{cor:V3}\label{a10:V3} Let $F$ be a flow of $\Ik(G)$ that does not descend. Then for every $\lambda$ with $1\le\lambda\le k-2$ the flow $F$ has at least $\lceil(k-1)/\lambda\rceil$ arcs of size $\lambda$, and it has a top. So $|E(G)|\ge 1+\sigma(k-1)$. \end{corollary} \begin{proof} $F$ is not joined to any flow with all sizes at most $k-2$, because such a flow is in $\Ik[k-1](G)$. Corollary~\ref{a10:V2} gives the arcs of size $\lambda$, for every $\lambda$. Arcs of different sizes are different arcs. $F$ has a top, because $F$ is not in $\Ik[k-1](G)$. So $|E(G)|\ge 1+\sum_{\lambda=1}^{k-2}\lceil(k-1)/\lambda\rceil=1+\sigma(k-1)$. \end{proof} \begin{corollary}[One level at a time]\label{cor:Citer}\label{a10:Citer} Let $k_0\ge 3$, let $\Ik[k_0](G)$ be connected, and let $|E(G)|\le\sigma(k_0)$. Then $\Ik(G)$ and $\Mk(G)$ are connected for every $k\ge k_0$. \end{corollary} \begin{proof} First, $\sigma(k+1)\ge\sigma(k)$ for $k\ge 2$: for $\lambda\le k-1$ the term $\lceil(k+1)/\lambda\rceil$ is at least $\lceil k/\lambda\rceil$, and there is one more term. Induction on $k$; $k=k_0$ is the hypothesis. Let $k>k_0$, and let $\Ik[k-1](G)$ be connected. If a flow of $\Ik(G)$ does not descend, Corollary~\ref{a10:V3} gives $|E(G)|\ge 1+\sigma(k-1)\ge 1+\sigma(k_0)$, a contradiction. So every flow descends, and Lemma~\ref{a10:level}~(b) gives the step. $\Mk(G)$: Lemma~\ref{a10:image}. \end{proof} \begin{corollary}[At most $189$ edges]\label{cor:C189}\label{a10:C189} Let $G$ be a finite bridgeless loopless multigraph. \begin{condlist}{(c)} \item[\textup{(a)}] If $G$ has at most $189$ edges, then $\Ik(G)$ and $\Mk(G)$ are connected for every $k\ge 40$. \item[\textup{(b)}] If $G$ has at most $235$ edges, then they are connected for every $k\ge 48$. \item[\textup{(c)}] In general, they are connected for every $k\ge k_0$, for every base $k_0$ with $\sigma(k_0)\ge|E(G)|$. \end{condlist} \end{corollary} \begin{proof} (c) is Corollary~\ref{a10:Citer} with Theorem~\ref{thm:bases}. (a) and (b) are (c) with $\sigma(40)=189$ and $\sigma(48)=235$. The terms $\lceil 40/\lambda\rceil$ and $\lceil 48/\lambda\rceil$ are: \begin{center}\small \setlength{\tabcolsep}{3.6pt} \begin{tabular}{@{}l|ccccccccccc@{}} $\lambda$ & $1$ & $2$ & $3$ & $4$ & $5$ & $6$ & $7$ & $8$, $9$ & $10,\dots,13$ & $14,\dots,19$ & $20,\dots,39$\\ \hline $\lceil 40/\lambda\rceil$ & $40$ & $20$ & $14$ & $10$ & $8$ & $7$ & $6$ & $5$ & $4$ & $3$ & $2$ \end{tabular} \smallskip \begin{tabular}{@{}l|cccccccccccc@{}} $\lambda$ & $1$ & $2$ & $3$ & $4$ & $5$ & $6$ & $7$ & $8$, $9$ & $10$, $11$ & $12,\dots,15$ & $16,\dots,23$ & $24,\dots,47$\\ \hline $\lceil 48/\lambda\rceil$ & $48$ & $24$ & $16$ & $12$ & $10$ & $8$ & $7$ & $6$ & $5$ & $4$ & $3$ & $2$ \end{tabular} \end{center} So $\sigma(40)=40+20+14+10+8+7+6+2\cdot 5+4\cdot 4+6\cdot 3+20\cdot 2=115+16+18+40=189$, and $\sigma(48)=48+24+16+12+10+8+7+2\cdot 6+2\cdot 5+4\cdot 4+8\cdot 3+24\cdot 2=147+16+24+48=235$. \end{proof} \subsection*{Few vertices} In this subsection $k\ge 4$. For a flow $F$ of $\Ik(G)$ let $\Phi(F)=\sum_e|F(e)|$, the sum of the sizes. A \emph{failure} \textup{(}for $G$ and $k$\textup{)} is a component of $\Ik(G)$ that contains no flow of $\Ik[k-1](G)$. Put \[ S(k)=\sum_{\lambda=2}^{k-2}\Bigl\lceil\frac{k-1}{\lambda}\Bigr\rceil=\sigma(k-1)-(k-1). \] \begin{lemma}[A potential]\label{lem:N1}\label{a10:N1} Let $\mathcal{U}$ be a component of $\Ik(G)$, and let $F$ be a flow of $\mathcal{U}$ with the least $\Phi$ in $\mathcal{U}$. Then there is a map $\theta\colon V(G)\to\Z$ such that for every arc $e$ of $G_F$ from $x$ to $y$: \begin{align*} \theta(y)-\theta(x)&=1&&\text{if $e$ is a middle arc;}\\ \theta(y)-\theta(x)&\le 1&&\text{if $e$ is a $1$-arc;}\\ \theta(y)-\theta(x)&\ge 1&&\text{if $e$ is a top.} \end{align*} \end{lemma} \begin{proof} Let $R$ be the directed multigraph on $V(G)$ with the following arcs and costs. Every arc $e$ of $G_F$ from $x$ to $y$ gives an arc $x\to y$ of cost $+1$ if $F(e)\le k-2$, and an arc $y\to x$ of cost $-1$ if $F(e)\ge 2$. A middle arc gives both. \emph{Every directed cycle of $R$ has a cost of at least $0$.} Let $O$ be a directed cycle of $R$ \textup{(}distinct vertices; $R$ has no loop, because $G$ has none\textup{)}. If $O$ uses both arcs that come from one arc $e$ of $G_F$, then after the first of them $O$ is at an end of $e$, and the only arc of $O$ that leaves this vertex is the second one; so $O$ consists of these two arcs, and its cost is $0$. Otherwise $O$ uses pairwise different edges of $G$ and passes pairwise different vertices, so it is a cycle $C$ of $G$ with a direction \textup{(}two parallel edges give a cycle with $2$ edges\textup{)}. Put $F'=F+\chi_C$, with $\chi_C$ read in $G_F$. If $O$ uses the arc $x\to y$ of $e$, then $C$ passes $e$ in its direction and $F'(e)=F(e)+1\le k-1$. If $O$ uses the arc $y\to x$ of $e$, then $C$ passes $e$ against its direction and $F'(e)=F(e)-1\ge 1$. All other values are equal. So $F'$ is in $\Ik(G)$, positive on $G_F$, and adjacent to $F$; so $F'$ is in $\mathcal{U}$, and $\Phi(F')-\Phi(F)$ is the cost of $O$. Since $F$ has the least $\Phi$ in $\mathcal{U}$, the cost is at least $0$. \emph{The potential.} For a vertex $y$ let $\theta(y)$ be the least cost of a directed path of $R$ that ends at $y$ \textup{(}the path with the one vertex $y$ and no arc has cost $0$; there are finitely many paths\textup{)}. Let $x\to y$ be an arc of $R$ with cost $\kappa$, and let $P$ be a path that ends at $x$ with cost $\theta(x)$. If $y$ is not on $P$, then $P$ followed by the arc is a path that ends at $y$, so $\theta(y)\le\theta(x)+\kappa$. If $y$ is on $P$, let $P_1$ be the part of $P$ up to $y$ and $P_2$ the part from $y$ to $x$. Then $P_2$ followed by the arc $x\to y$ is a directed cycle of $R$, so $\operatorname{cost}(P_2)+\kappa\ge 0$, and $\theta(y)\le\operatorname{cost}(P_1)=\theta(x)-\operatorname{cost} (P_2)\le\theta(x)+\kappa$. In both cases $\theta(y)\le\theta(x)+\kappa$. An arc $e$ of $G_F$ from $x$ to $y$ with $F(e)\le k-2$ gives $\theta(y)-\theta(x)\le 1$. If $F(e)\ge 2$, the arc $y\to x$ of cost $-1$ gives $\theta(x)\le\theta(y)-1$. A middle arc has both; a $1$-arc has the first \textup{(}$1\le k-2$\textup{)}; a top has the second \textup{(}$k-1\ge 2$\textup{)}. \end{proof} \begin{lemma}[A forest of middle arcs]\label{lem:N2}\label{a10:N2} Every component $\mathcal{U}$ of $\Ik(G)$ contains a flow $F^*$ that has the least $\Phi$ in $\mathcal{U}$ and whose middle arcs contain no cycle of $G$. \end{lemma} \begin{proof} Among the flows of $\mathcal{U}$ with the least $\Phi$, take a flow $F$ with the least number of middle arcs. Suppose that $C$ is a cycle of $G$ all of whose edges are middle arcs of $F$. Give $C$ a direction, and take $\theta$ from Lemma~\ref{a10:N1}. Every arc of $C$ has $\theta(y)-\theta(x)=1$. The sum of $\theta(y)-\theta(x)$ over the arcs that $C$ passes in their direction, minus the same sum over the arcs that $C$ passes against their direction, is $0$, because it is a sum of differences of $\theta$ once round $C$. So $C$ passes as many arcs in their direction as against it. $C$ has at least $2$ edges, so it passes at least one arc in its direction. Put $F_j=F+j\chi_C$ \textup{(}$\chi_C$ read in $G_F$\textup{)} for $j=0,1,2,\dots$, and let $j_0\ge 1$ be the least $j$ such that $F_j$ has the value $1$ or $k-1$ on some arc of $C$. It exists, because the value of an arc that $C$ passes in its direction grows by $1$ at every step. For $jk_0$, and let $\Ik[k-1](G)$ be connected. Then $S(k)\ge S(k_0+1)\ge n$, so by Theorem~\ref{a10:N} there is no failure for $G$ and $k$: every flow of $\Ik(G)$ descends. Lemma~\ref{a10:level}~(b) gives the step, and Lemma~\ref{a10:image} gives $\Mk(G)$. (a) $S(41)=\sigma(40)-40=149$ and $S(49)=\sigma(48)-48=187$ \textup{(}Corollary~\ref{a10:C189}\textup{)}. Use (b) with the bases $40$ and $48$. (c) $S(k_0+1)=\sum_{\lambda=2}^{k_0-1}\lceil k_0/\lambda\rceil\ge k_0\bigl(\frac12+\dots+\frac1{k_0-1}\bigr)\ge k_0(\ln k_0-1)$, because $1+\frac12+\dots+\frac1{k_0-1}\ge\int_1^{k_0}\frac{dx}{x}=\ln k_0$. (d) Put $\eta=2n/\ln n$, and let $k_0$ be the least multiple of $4$ with $k_0\ge\eta$; so $k_0<\eta+4$. The function $\xi/\ln\xi$ grows for $\xi\ge 3$, and $\ln 150<5.02$; so $\eta\ge 300/5.02>59$, and $k_0\ge 60$ is a base. The function $\xi(\ln\xi-1)$ grows for $\xi\ge 1$, so $k_0(\ln k_0-1)\ge\eta(\ln\eta-1)$. And $\eta(\ln\eta-1)\ge n$ means $2(\ln 2+\ln n-\ln\ln n-1)\ge\ln n$, that is \[ \ln n-2\ln\ln n\ \ge\ 2-2\ln 2 . \] Here $\ln n\ge\ln 150>5$. The function $\xi-2\ln\xi$ grows for $\xi>2$, and at $\xi=5$ it is $5-2\ln 5>1.7$, while $2-2\ln 2<0.62$. So $k_0(\ln k_0-1)\ge n$, and (c) gives every $k\ge k_0$, in particular every $k\ge\eta+4$. \end{proof} \subsection*{Short cycles} In this subsection $F$ is a flow of $\Ik(G)$. A \emph{lift} of $F$ is a flow $\widetilde F$ of $\Ik(G)$ with $\widetilde F=F$ modulo $k$ \textup{(}edge by edge\textup{)}. Let $\Psi(F)$ be the number of arcs of $G_F$ that are $1$-arcs or tops. An edge $e$ is counted in $\Psi(F)$ if and only if $F(e)$ is $1$ or $-1$ modulo $k$; so $\Psi(F)$ depends only on $F$ modulo $k$. \emph{Residues.} Let $C$ be a cycle of $G$ with a direction. For an arc $e$ of $G_F$ on $C$ put $r(e)=F(e)$ if $C$ passes $e$ in its direction, and $r(e)=k-F(e)$ if $C$ passes $e$ against its direction. So $1\le r(e)\le k-1$, and $r(e)$ is the value of $F$ on $e$ measured in the direction of $C$, taken modulo $k$. It is the same for all lifts of $F$. If the direction of $C$ is reversed, $r(e)$ becomes $k-r(e)$. Write $C_{<\lambda}$ and $C_{>\lambda}$ for the sets of the arcs $e$ of $C$ with $r(e)<\lambda$ and with $r(e)>\lambda$. For a set $X$ of vertices and an arc $e$ of the cut of $X$ \textup{(}the edges with one end in $X$\textup{)} put $r_X(e)=F(e)$ if $e$ leaves $X$ and $r_X(e)=k-F(e)$ if $e$ enters $X$. If $C$ passes the arc $e$ of the cut \emph{into} $X$, then $r_X(e)=k-r(e)$; if $C$ passes $e$ \emph{out of} $X$, then $r_X(e)=r(e)$. \textup{(}Check the four cases: $e$ in the direction of $C$ or against it, and $e$ entering or leaving $X$ in $G_F$.\textup{)} \begin{lemma}[Lifts]\label{lem:lifts}\label{a10:lifts} \begin{condlist}{(b)} \item[\textup{(a)}] The lifts of $F$ are exactly the flows $F-kY$, where $Y$ is an integer flow of $G$ with values in $\{0,1\}$ in the orientation $G_F$. All lifts of $F$ are joined to $F$. \item[\textup{(b)}] For every set $X$ of vertices, $k\operatorname{in}(X)$ is the sum of $r_X(e)$ over the arcs $e$ of the cut of $X$. \end{condlist} \end{lemma} \begin{proof} (a) Let $\widetilde F$ be a lift, read in $G_F$. On an arc $e$ the number $\widetilde F(e)$ is $F(e)$ modulo $k$ and lies in $[-(k-1),k-1]\setminus\{0\}$; so it is $F(e)$ or $F(e)-k$. So $Y=(F-\widetilde F)/k$ is an integer flow with values in $\{0,1\}$. Conversely, for such a $Y$ the flow $F-kY$ has on every arc the value $F(e)$ in $[1,k-1]$ or $F(e)-k$ in $[-(k-1),-1]$; so it is a lift of $F$. The lifts of $F$ are the integer lifts of the reduction of $F$ modulo $k$, so they are joined by Lemma~\ref{lem:fibres}. (b) The flow out of $X$ is the flow into $X$. So the sum of $r_X$ is $\sum_{e\text{ leaves }X}F(e)+\sum_{e\text{ enters }X}\bigl(k-F(e)\bigr)=k\operatorname{in}(X)$. \end{proof} \begin{theorem}[A step after a change of lift]\label{thm:R}\label{a10:R} Let $C$ be a cycle of $G$ with a direction, and $\lambda$ an integer with $1\le\lambda\le k-1$ such that $r(e)\ne\lambda$ for every arc $e$ of $C$. The following are equivalent: \begin{condlist}{(ii)} \item[\textup{(i)}] some lift $\widetilde F$ of $F$ has $\widetilde F-\lambda\chi_C$ in $\Ik(G)$; \item[\textup{(ii)}] for every set $X$ of vertices, the number of arcs of $C_{<\lambda}$ that $C$ passes into $X$ is at most $\operatorname{in}(X)$. \end{condlist} If \textup{(i)} holds, then $F$ is joined to $\widetilde F-\lambda\chi_C$, and this flow is $F-\lambda\chi_C$ modulo $k$. \end{theorem} \begin{proof} Measure every value on an arc $e$ of $C$ in the direction of $C$. Measured so, $F$ has a value that is $r(e)$ modulo $k$. A lift $\widetilde F$ has a value that is $r(e)$ modulo $k$ and whose absolute value is in $[1,k-1]$; so it has the value $r(e)$ or $r(e)-k$. After the step it has $r(e)-\lambda$ or $r(e)-k-\lambda$. The first number is not $0$ and lies in $[-(k-2),k-2]$. The second is negative, and it is at least $-(k-1)$ exactly if $r(e)\ge\lambda+1$. The values outside $C$ do not change. So $\widetilde F-\lambda\chi_C$ is in $\Ik(G)$ if and only if $\widetilde F$ has the value $r(e)$, measured in the direction of $C$, on every arc $e$ of $C_{<\lambda}$. Write $\widetilde F=F-kY$ as in Lemma~\ref{a10:lifts}~(a). Let $C^+$ be the set of the arcs of $C_{<\lambda}$ that $C$ passes in their direction, and $C^-$ the set of those that $C$ passes against their direction. On an arc of $C^+$, $F$ has the value $r(e)$ along $C$, and $\widetilde F$ has it if and only if $Y(e)=0$. On an arc of $C^-$, $F$ has the value $-F(e)=r(e)-k$ along $C$, and $\widetilde F$ has the value $r(e)$ if and only if $Y(e)=1$. So \textup{(i)} holds if and only if there is an integer flow $Y$ in $G_F$ with $l\le Y\le u$, where $l=u=0$ on $C^+$, $l=u=1$ on $C^-$, and $l=0$, $u=1$ on all other arcs. By Theorem~\ref{q:hoffman} this holds if and only if for every $X$: \[ (\text{number of arcs of $C^-$ that leave $X$})\ \le\ \operatorname{in}(X)-(\text{number of arcs of $C^+$ that enter $X$}). \] An arc of $C^-$ that leaves $X$ is passed by $C$ into $X$, and so is an arc of $C^+$ that enters $X$; and every arc of $C_{<\lambda}$ that $C$ passes into $X$ is of one of these two kinds. So this is \textup{(ii)}. The last sentence: $F$ is joined to $\widetilde F$ by Lemma~\ref{a10:lifts}~(a), and $\widetilde F\to\widetilde F-\lambda\chi_C$ is a move. \end{proof} \begin{corollary}[A sufficient condition]\label{a10:R1} Let $C$ and $\lambda$ be as in Theorem~\ref{a10:R}. If the sum of $r(e)-1$ over the arcs $e$ of $C_{<\lambda}$ is less than $k$, then \textup{(ii)} of Theorem~\ref{a10:R} holds. In particular \textup{(ii)} holds for $\lambda=1$ \textup{(}$C_{<1}$ is empty\textup{)} and for $\lambda=2$ \textup{(}every arc of $C_{<2}$ has $r(e)=1$\textup{)}. \end{corollary} \begin{proof} Fix $X$. The cycle $C$ passes arcs of the cut of $X$ into $X$ as often as out of $X$; say $j$ times each. Let $j'$ of the $j$ arcs that $C$ passes into $X$ be in $C_{<\lambda}$. By Lemma~\ref{a10:lifts}~(b), and because $r_X\ge 1$ on every arc of the cut, \begin{align*} k\operatorname{in}(X)\ &\ge\sum_{\substack{e\in C_{<\lambda}\\ \text{passed into }X}}\bigl(k-r(e)\bigr)+(j-j')+j \ =\ j'k-\sum_{\substack{e\in C_{<\lambda}\\ \text{passed into }X}}\bigl(r(e)-1\bigr)+2(j-j')\\ &\ge\ j'k-\sum_{e\in C_{<\lambda}}\bigl(r(e)-1\bigr)\ >\ j'k-k . \end{align*} \textup{(}The $j'$ arcs of $C_{<\lambda}$ passed into $X$ have $r_X=k-r(e)$; the other $j-j'$ arcs passed into $X$, the $j$ arcs passed out of $X$, and the arcs of the cut that are not on $C$ have $r_X\ge 1$, so they add at least $(j-j')+j$; and every term $r(e)-1$ is at least $0$.\textup{)} So $\operatorname{in}(X)>j'-1$, that is $\operatorname{in}(X)\ge j'$. \end{proof} \begin{theorem}[The cleaning step]\label{thm:cut}\label{a10:clean} Let $e_0$ be a $1$-arc or a top of $F$, and let $C$ be a cycle through $e_0$, directed so that $r(e_0)=1$ \textup{(}in the direction of a $1$-arc, against the direction of a top\textup{)}. \begin{condlist}{(b)} \item[\textup{(a)}] Let $\lambda$ be an integer with $3\le\lambda\le k-3$ such that no arc of $C$ has $r(e)$ in $\{\lambda-1,\lambda,\lambda+1\}$, and such that the sum of $r(e)-1$ over $C_{<\lambda}$ is less than $k$, or the sum of $k-1-r(e)$ over $C_{>\lambda}$ is less than $k$. Then $F$ is joined to a flow $F'$ with $F'=F-\lambda\chi_C$ modulo $k$ and $\Psi(F')\le\Psi(F)-1$. \item[\textup{(b)}] Let $h\ge 0$ be an integer with $3h^2\lambda}$ is less than $k$. Let $C'$ be $C$ with the other direction; its residues are $k-r(e)$, none of them is $k-\lambda$, and $3\le k-\lambda\le k-3$. The arcs with $k-r(e)\lambda}$, and the sum of $(k-r(e))-1$ over them is less than $k$. Theorem~\ref{a10:R} and Corollary~\ref{a10:R1} for $C'$ and $k-\lambda$ give $F'$ joined to $F$ with $F'=F-(k-\lambda)\chi_{C'}=F+(k-\lambda)\chi_C=F-\lambda\chi_C$ modulo $k$. The count. An edge outside $C$ keeps its value modulo $k$. An arc $e$ of $C$ has in $F'$ the value $r(e)-\lambda$ modulo $k$, measured in the direction of $C$. It is counted in $\Psi(F')$ if and only if $r(e)-\lambda$ is $1$ or $-1$ modulo $k$. Since $r(e)-\lambda$ lies in $[4-k,k-4]$, this means $r(e)=\lambda+1$ or $r(e)=\lambda-1$, which is excluded. So no edge of $C$ is counted in $\Psi(F')$, while $e_0$ is counted in $\Psi(F)$. (b) First case: at most $h$ arcs of $C$ have $r(e)$ in $I_{\mathrm{low}}$. Look at the $3h+1$ integers $\lambda$ in $[3,3h+3]$. An arc with $r(e)$ in $I_{\mathrm{low}}$ has $r(e)$ in $\{\lambda-1,\lambda, \lambda+1\}$ for at most $3$ of them. An arc with $r(e)=1$ or $r(e)\ge 3h+5$ has this for none of them, because $\lambda-1\ge 2$ and $\lambda+1\le 3h+4$. So at most $3h$ values of $\lambda$ are excluded, and some $\lambda$ in $[3,3h+3]$ has no arc of $C$ with $r(e)$ in $\{\lambda-1,\lambda,\lambda+1\}$. For this $\lambda$, an arc of $C_{<\lambda}$ has $r(e)=1$, or $2\le r(e)\le\lambda-2$; the arcs of the second kind have $r(e)$ in $I_{\mathrm{low}}$, so there are at most $h$ of them, each with $r(e)-1\le\lambda-3\le 3h$. So the sum of $r(e)-1$ over $C_{<\lambda}$ is at most $3h^2\lambda}$ has $r(e)=k-1$, or $\lambda+2\le r(e)\le k-2$; the arcs of the second kind have $r(e)$ in $I_{\mathrm{high}}$, so there are at most $h$ of them, each with $k-1-r(e)\le k-3-\lambda\le 3h$. So the sum of $k-1-r(e)$ over $C_{>\lambda}$ is at most $3h^2(k-2)\operatorname{in}(X)$. Then $\operatorname{out}(X)\ge 1$, the flow out of $X$ is positive, so is the flow into $X$, and $\operatorname{in}(X)\ge 1$; hence $\operatorname{out}(X)\ge k-1$. Every arc that leaves $X$ and is not a $1$-arc has a value of at least $2$: \[ (k-1)\operatorname{in}(X)\ \ge\ \text{flow into }X\ =\ \text{flow out of }X\ \ge\ \operatorname{out}_1(X) +2\bigl(\operatorname{out}(X)-\operatorname{out}_1(X)\bigr). \] With $\operatorname{in}(X)\le(\operatorname{out}(X)-1)/(k-2)$ this gives \begin{align*} \operatorname{out}_1(X)\ &\ge\ 2\operatorname{out}(X)-\frac{(k-1)(\operatorname{out}(X)-1)}{k-2} \ =\ \frac{(k-3)\operatorname{out}(X)+(k-1)}{k-2}\\ &\ge\ \frac{(k-3)(k-1)+(k-1)}{k-2}\ =\ k-1 . \end{align*} Also the flow into $X$ is at least $\operatorname{out}(X)>(k-2)\operatorname{in}(X)$, so some arc that enters $X$ has a value greater than $k-2$: it is a top. So the cut of $X$ contains at least $k$ arcs that are $1$-arcs or tops of $F'$. By (a) none of them is short. This contradicts the hypothesis. So $F'$, and hence every flow, descends. (c) Take $h=3$. For every $k\ge 41$: $27\frac{4n}{3}-2 \] edges. Extraction therefore applies, including at $r=1/9$. If $n=2$, then $m\ge3$ and the support has more than two edges, so it contains a two-edge cycle. In both cases we obtain a good subgraph entirely in nonzero quotient support. \section{A binary list argument} In this section $A$ is a finite elementary abelian $2$-group. Cycle vectors with coefficients in $A$ can be written as unsigned indicators; this convention is not applied to $\Gamma$-valued moves. \begin{lemma}[Repeated values]\label{lem:duplicates} Let $W$ be a finite set with $|W|<|A|$, let $v:W\to A$, and put $V=v(W)$. Suppose $R_1,\ldots,R_r\subseteq W$, $r\ge2$, have empty common intersection. For every $\sigma\in A$, some $j$ satisfies \[ V\cup(\sigma+v(R_j))\ne A. \] \end{lemma} \begin{proof} Put $M=A\setminus V$. If all the displayed unions were $A$, every value of $\sigma+M$ would lie in every $v(R_j)$. A value with a unique preimage in $W$ cannot do so, because that preimage would belong to all $R_j$. Thus each of the $|M|$ values in $\sigma+M$ has at least two preimages. The number of repeated values is at most $|W|-|V|$, whereas $|M|=|A|-|V|>|W|-|V|$, a contradiction. \end{proof} An edge set $D$ of $Q$ is \emph{even} if every vertex has even degree in $D$. For $\sigma\ne0$, the map $\sigma\ind_D$ is an $A$-flow precisely when $D$ is even. Every nonempty even set decomposes into edge-disjoint cycles. Call $\sigma\ind_D$ a \emph{general step}. It is not required to be supported on just one or two official cycles. \begin{lemma}[Binary list walk]\label{lem:binwalk} For a frame and a list assignment $\Lambda$, suppose \[ |\Lambda(e_i)|+|\Lambda(f_i)|<|A|\quad\text{at every node }i. \] Any two flows in $\PL{\Lambda}$ are joined inside $\PL{\Lambda}$ by a sequence of general steps. \end{lemma} \begin{proof} Let $\Lambda_i$ keep only the lists at nodes $i,i+1,\ldots,s$. Prove the assertion for $\Lambda_i$ by downward induction on $i$. With no lists, the official cycle expansion gives the assertion. Suppose it holds for $i+1$, put $\chi=\ind_{C_i}$, and set \[ X_e(Z)=Z(e_i)+\Lambda(e_i),\quad X_f(Z)=Z(f_i)+\Lambda(f_i),\quad X(Z)=X_e(Z)\cup X_f(Z). \] The official-cycle properties imply that for $Z\in\PL{\Lambda_{i+1}}$, adding $c\chi$ leaves all later lists satisfied, and \begin{equation}\label{eq:compensator} Z+c\chi\in\PL{\Lambda_i}\quad\Longleftrightarrow\quad c\notin X(Z). \end{equation} Take a walk $Y=Y_0,\ldots,Y_n=Y'$ in $\PL{\Lambda_{i+1}}$, each step $Y_r-Y_{r-1}=\sigma\ind_D$, supplied by induction. We follow it with compensated states $Y_r+c_r\chi$ satisfying~\eqref{eq:compensator}; start with $c_0=0$. For one step write $X_e=X_e(Y_{r-1})$, $X_f=X_f(Y_{r-1})$, and let $k_e,k_f\in\{0,1\}$ indicate whether $e_i,f_i$ lie in $D$. Then \[ X(Y_r)=(X_e+k_e\sigma)\cup(X_f+k_f\sigma). \] We seek $c\notin X_e\cup X_f$ such that either \[ \text{(I) }c\notin X(Y_r),\qquad\text{or}\qquad \text{(II) }c+\sigma\notin X(Y_r). \] For $(k_e,k_f)=(0,0)$ the current $c_{r-1}$ satisfies (I); for $(1,1)$ it satisfies (II). For $(1,0)$ we need $c$ outside one of \[ (X_e\cup X_f)\cup(\sigma+X_e),\qquad (X_e\cup X_f)\cup(\sigma+X_f). \] Apply Lemma~\ref{lem:duplicates} to disjoint tagged copies of $X_e$ and $X_f$, with the value map forgetting the tag. There are fewer than $|A|$ positions, and the two candidate subsets are disjoint. This gives $c$. The case $(0,1)$ exchanges the two labels. First change $c_{r-1}$ to $c$ along $C_i$ if necessary; both compensated endpoints are legal. In case (I) follow the original general step and set $c_r=c$. In case (II) set $c_r=c+\sigma$ and use \[ \sigma\ind_D+\sigma\ind_{C_i} =\sigma\ind_{D\mathbin\triangle E(C_i)}. \] The symmetric difference is even; if empty, omit the step. Thus this too is an allowed general step. At the end, change $c_n$ to zero along $C_i$; the final state $Y'$ satisfies its lists. This proves the induction. Allowing arbitrary even supports is essential to closure under the displayed symmetric difference. \end{proof} \section{Signed realization inside a good subgraph}\label{sec:local} Throughout this and the next section, $B$ and $A$ are binary, $|B|\ge4$, and $|A|\ge8$, as in Theorem~\ref{thm:main}. We use any frame; there are no additional counting restrictions on it. \begin{lemma}[Free-thread preparation]\label{lem:prep} Let $f$ be $H$-good on a subdivided good subgraph and put $u=\quo f$. For selected free threads, let $T_\varepsilon\subsetneq B$ contain \[ W_\varepsilon=\{-f(a):a\in E(\thr{\varepsilon}),\ u(a)=0\}. \] One can translate each selected free thread by any chosen $y_\varepsilon\in B\setminus T_\varepsilon$, using a legal cycle move on $\Deps$ (or no move for $y_\varepsilon=0$). The quotient is unchanged, other free threads are unchanged, and the resulting flow is $H$-good. \end{lemma} \begin{proof} On its own free thread $\chi_{\Deps}=1$. Avoiding $W_\varepsilon$ protects all old quotient-zero edges of that thread. Every other edge of the cycle is protected by its nonzero quotient. Distinct preparations affect no other free thread, so perform them successively. Each nonzero coefficient is an allowed move. \end{proof} \begin{lemma}[One auxiliary cycle]\label{lem:onecycle} Let $f$ be $H$-good and $K$ a cycle of $Q$. Suppose $0\ne\sigma\in A$ and $u+\pi(\sigma\chi_K)$ is nonzero on every protected edge. Then $f$ is joined by moves supported in $H$ to an $H$-good flow with that quotient. \end{lemma} \begin{proof} By Proposition~\ref{prop:frame}(i), $\pi(\chi_K)$ is a sum of signed vectors of edge-disjoint cycles $O_1,\ldots,O_j$ of $H$. Choose a lift $\gamma$ of $\sigma$. On a free thread each $O_l$ either contains the entire thread or misses it, and no two $O_l$ use the same thread. For the prospective signed addition $\gamma\pi(\chi_K)$, on each free thread forbid $-f(a)$ when $u(a)=0$, and forbid $-f(a)-\gamma\pi(\chi_K)(a)$ when the new quotient is zero and that edge changes. These forbidden values lie in $B$. A changed edge cannot have quotient zero both before and after, since $\sigma\ne0$. Hence at most one value is forbidden per thread edge, giving at most three forbidden values. Use Lemma~\ref{lem:prep}, then add $\gamma\chi_{O_l}$ successively. Protected edges and nonzero final quotients are safe. The preparations protect every newly zero quotient on a free thread, as well as every unchanged zero quotient. Because the cycles are edge-disjoint, during these cycle additions every edge has its prepared value or its final value. This proves the assertion. The coefficient $\gamma$ need not have order two; all additions in this proof are signed. \end{proof} This is the signed one-cycle mechanism of Lemma~\ref{a3:lem:step} in Part~I, reproduced here for the all-coefficients setting. \begin{lemma}[Internal quotient adjustment]\label{lem:internal} Let $f$ be $H$-good on a subdivided good subgraph. For every $Y\in\PL{\Lambda_f}$ there is an $H$-good flow joined to $f$ by moves on $H$, equal to $f$ outside $H$, with quotient $u+\pi(Y)$. \end{lemma} \begin{proof} At a node the two lists have total size at most six, which is less than $|A|$. Lemma~\ref{lem:binwalk} joins $0$ to $Y$ by general steps in $\PL{\Lambda_f}$. Decompose each even support into edge-disjoint cycles of $Q$. For each partial sum, every auxiliary edge has its value from the old or the final state of this general step, so the lists remain satisfied. Apply Lemma~\ref{lem:onecycle} to each auxiliary cycle, starting each time from the currently reached full flow. That lemma requires only current $H$-goodness and nonzero final protected quotients; it does not require the original full values or the original lists. Thus the successive realizations give the stated quotient and leave the exterior unchanged. \end{proof} \section{Lifting a contracted move}\label{sec:contracted} Contract all vertices of a connected good subgraph $H$ to a single vertex, delete its edges, and retain every other edge, including new loops. Denote the resulting graph by $G/H$. \begin{lemma}[Local lifting]\label{lem:lift} Let $f$ be an $H$-good $\Gamma$-flow of $G$ and let $f^*$ be its restriction to $G/H$. Every legal move \[ g^*=f^*+\lambda\chi_{C^*} \] can be lifted to a path from $f$ to an $H$-good flow $g$ whose restriction is $g^*$. \end{lemma} \begin{proof} If the edges of $C^*$ already form a cycle in $G$, perform that move directly. Otherwise they form a path with distinct ends $p,p'\in V(H)$, all other vertices outside $H$. It suffices to find a path $J\subseteq H$ between them and prepare inside $H$ before adding $\lambda\chi_C$, where $C=C^*\cup J$ is directed to agree with $C^*$ outside $H$. Write $\sigma=\quo\lambda$ and $u=\quo f$. Call the move a \emph{quotient step} if $\sigma\ne0$, and a \emph{kernel step} if $\sigma=0$, that is, if $\lambda\in B$. \subsection*{A short-cycle good subgraph} Choose an arc $J$ of $H$ with at most two edges. If $\sigma=0$, the quotient protects all edges of $H$ and no preparation is needed. Otherwise, a quotient addition $c\chi_H$ has at most five forbidden values of $c$ for maintaining a nonzero quotient on $H$, and the further addition $\sigma\chi_C$ on $J$ forbids at most two more. As $|A|\ge8$, choose $c$ avoiding all seven. If $c\ne0$, add any lift of $c$ on the signed cycle $H$; the final quotient is nonzero throughout $H$, so this is legal. Now perform $\lambda\chi_C$. The flow is still $H$-good and has the prescribed restriction. \subsection*{A subdivision and a nonzero quotient step} Suppose $\sigma\ne0$. We construct $Y\in\PL{\Lambda_f}$ and a path $J$ such that \begin{equation}\label{eq:pathprotection} \begin{gathered} u_1=u+\pi(Y)\text{ is nonzero on protected edges, and}\\ u_1(a)+\sigma\ne0\text{ on protected edges of }J. \end{gathered} \end{equation} Let $W_i$ be the edge positions on the original threads of $e_i,f_i$; thus $|W_i|\le6$. Choose official coefficients $x_i\in A$ in decreasing order of $i$. At stage $i$ put \[ Z_i=\sum_{l>i}x_l\ind_{C_l},\qquad v_i(a)=Z_i(t)+u(a)\quad(a\in E(\thr{t}),\ t\in N_i),\qquad V_i=v_i(W_i). \] For $Y=\sum_i x_i\ind_{C_i}$ the triangular official-cycle properties give \begin{equation}\label{eq:triangular} (u+\pi(Y))(a)=v_i(a)+x_i\qquad(a\in W_i). \end{equation} At the same time choose a permitted position set $R_i\subseteq W_i$ and ensure \begin{equation}\label{eq:permission} x_i\notin V_i\cup(\sigma+v_i(R_i)). \end{equation} The later coefficients already determine $v_i$, so all these choices are well-defined in this order. If both $p,p'$ are interior to the same thread, it has length three and they are joined by its middle edge. Take that edge as $J$. At its node, if protected, let $R_i$ be this single edge; take all other $R_l$ empty. At most $6+1<8$ values are excluded at any node. If the thread is free all $R_l$ are empty; its middle edge is handled by free-thread preparation below. Otherwise classify each node by the interiors containing the endpoints. There are three possibilities. \begin{enumerate} \item \emph{Neither endpoint is inside its pair.} The two candidate sets are the entire thread edge sets of $e_i$ and $f_i$, empty for an expansion edge. They are disjoint. Lemma~\ref{lem:duplicates} chooses one, $R_i=E(\thr{g_i})$, and a coefficient satisfying~\eqref{eq:permission}. \item \emph{Exactly one endpoint, say $p_i$, is inside its pair.} Let its thread be $\thr{t_i}$, with ends $z,z'$, and write $S_z,S_{z'}$ for the edge sets of the two pieces from $p_i$ to these ends. Let $T'$ be the other thread edge set in the pair. The three candidate sets \[ S_z\cup T',\qquad S_{z'}\cup T',\qquad S_z\cup S_{z'} \] have empty common intersection. Apply Lemma~\ref{lem:duplicates}. For the first or second choice select the other edge of the pair as $g_i$ and attach the corresponding piece from $p_i$ to $z$ or $z'$. For the third choose $g_i=t_i$ and attach no extra piece; the endpoint then lies in $H_g$ itself. \item \emph{The two endpoints are inside different threads of this pair.} Each thread has one end on each operand side of node $i$. The two candidate sets consist of the two endpoint pieces going to the first side, or the two pieces going to the second side. They are disjoint. Lemma~\ref{lem:duplicates} selects one side and a coefficient satisfying~\eqref{eq:permission}. Choose any $g_i$ to complete the tree selection; only the selected endpoint pieces will be used at this node. \end{enumerate} There is either one node of the third type and all others have the first type, or at most two nodes of the second type and all others have the first type. The types depend only on the fixed endpoints and frame. All applications of Lemma~\ref{lem:duplicates} use $|W_i|\le6<|A|$. We now build the path from these independent permissions. If there is no node of the third type, each endpoint lies in $H_g$ or has an attached piece ending in $H_g$. Join the two resulting ends by a path in connected $H_g$ and attach those pieces. The interiors of attached pieces lie on unselected threads, outside $H_g$, and two attached pieces are on different threads. Hence the union is a path. The selected thread and attached piece rules ensure that it uses only $R_i$ at each pair. If node $i$ has the third type, use the spanning tree of its chosen operand stage to join the corresponding ends of its two pair edges. Neither pair edge is in that operand stage. Contract expansion edges and restore original threads. The image is a connected subgraph $L$ containing both required thread ends, using neither thread of $N_i$, and using only selected threads at other nodes. Adjoin the chosen endpoint pieces and choose a simple $p$--$p'$ path in this connected union. Again it uses only $R_i$ at each node. We do not require the image of the auxiliary tree path itself to be a simple path after contraction. Equations~\eqref{eq:triangular}--\eqref{eq:permission} prove~\eqref{eq:pathprotection}. Use Lemma~\ref{lem:internal} to reach a full flow $f_1$ with quotient $u_1$. For every free thread meeting $J$, forbid \begin{align*} W_\varepsilon&=\{-f_1(a):a\in E(\thr{\varepsilon}),\ u_1(a)=0\},\\ T_\varepsilon&=W_\varepsilon\cup \bigl\{-f_1(a)-\lambda\chi_C(a): a\in E(\thr{\varepsilon})\cap E(J),\\ &\hspace{48mm}u_1(a)+\sigma\chi_C(a)=0\bigr\}. \end{align*} These are elements of $B$. Because $\sigma\ne0$, each thread edge contributes at most one forbidden value; thus $|T_\varepsilon|\le3<|B|$. Lemma~\ref{lem:prep} gives a flow $f_2$ with all these obstructions avoided. Now $f_2+\lambda\chi_C$ is legal and $H$-good. Protected edges of $J$ are safe by~\eqref{eq:pathprotection}, new quotient zeros on free threads are safe by the second forbidden set, and the first forbidden set preserves old quotient zeros throughout each prepared thread. The signs in $\lambda\chi_C$ and in the forbidden set are retained even though $\sigma=-\sigma$. \subsection*{A subdivision and a kernel step} Suppose $\sigma=0$. Then $\lambda\in B$ and $\lambda=-\lambda$. In this part, and only for these $B$-coefficients, signs disappear. On a free thread put \[ Z_\varepsilon=\{a\in E(\thr{\varepsilon}):u(a)=0\},\qquad F_\varepsilon=\{f(a):a\in Z_\varepsilon\}\subseteq B\setminus\{0\}. \] For any thread edge set $S$, define \[ T_\varepsilon(S)=F_\varepsilon\cup \{f(a)+\lambda:a\in Z_\varepsilon\cap S\}. \] Choosing $y\in B\setminus T_\varepsilon(S)$ allows preparation on $\Deps$, protects the entire thread before the move, and protects the edges of $S$ after the further $\lambda$ addition. For two disjoint sets $S_1,S_2$, at least one $T_\varepsilon(S_j)$ is proper in $B$. Otherwise each contains zero, forcing two distinct positions of value $\lambda$, one in each piece. Since $|Z_\varepsilon|\le3$, this gives $|F_\varepsilon|\le2$. Each equality $T_\varepsilon(S_j)=B$, with $|B|\ge4$, then requires at least two positions in $Z_\varepsilon\cap S_j$. The two disjoint pieces would require at least four positions, a contradiction. If $p,p'$ lie in the same free thread, choose between its middle edge and its two outer edges, which are disjoint candidate sets. For the second choice join the distinct thread ends by a path in $H_P$. A preparation outside the corresponding proper forbidden set makes the chosen path safe. Otherwise, for each endpoint interior to a free thread choose one of its two disjoint pieces to a thread end with a proper forbidden set, and prepare that thread. An endpoint not interior to a free thread is already in $H_P$. Join the resulting endpoints by a path in $H_P$ and attach the selected pieces. The resulting union is a path: $H_P$ contains no interior vertex of a free thread, and any two pieces are on different threads. Preparations on different free threads leave each other's full values unchanged. Finally add $\lambda\chi_C$. On protected edges the unchanged quotient is nonzero; on the free pieces the selected preparation prevents zero; elsewhere in $H$ nothing changes. Outside $H$ the resulting flow is $g^*$. This completes all cases. \end{proof} \section{Universal connectivity} \begin{proof}[Proof of Theorem~\ref{thm:main}] Every $x\in\Gamma$ satisfies $2x\in B$, hence $4x=0$. Since $|\Gamma|\ge32$, the finite abelian structure theorem shows that $\Gamma$ is a product of two nontrivial groups. This product is used for averaging only; it need not split $B$ or the quotient map onto $A$. Lemma~\ref{lem:average} and the assumed ratio supply a good subgraph in nonzero quotient support in every loopless $3$-edge-connected graph on at least two vertices. Call a $\Gamma$-flow a \emph{target} if its $A$-coordinate is nowhere zero. Every bridgeless graph has a target: use the order-at-least-six existence theorem for $A$, then Lemma~\ref{lem:elementary} to lift it. The lift is automatically nowhere zero. We prove that every nowhere-zero flow is joined to every target on a connected bridgeless graph. Induct on the number of vertices, first proving the loopless case at each order and then allowing loops. A connected loopless graph on one vertex has only its empty flow. If loops are present, apply the loopless result to their deletion, then adjust each loop separately to its target value with a single nonzero cycle coefficient, omitting changes already achieved. Suppose the loopless graph has a cut of size two, with edges $a,b$. Contract $a$. The resulting graph is bridgeless and has fewer vertices, though it may have loops. Every flow on it lifts uniquely to a flow on the original graph by conservation at the two ends of $a$. Conservation across the two-edge cut forces the lifted value of $a$ to be the value of $b$ or its negative. Thus restriction preserves, in both directions, being nowhere zero and being a target. A cycle of the contracted graph lifts to a cycle either directly or by adjoining $a$ to the path between its ends. The unique flow lift therefore takes every legal contracted move to a legal original move. Induction applies. It remains to treat a loopless $3$-edge-connected graph. By averaging and extraction join the initial flow $f$ to an $H$-good flow $f'$ for some good $H$ with at least two vertices. The graph $G/H$ is bridgeless and has fewer vertices: a cycle of $G$ through a retained edge projects to a closed trail containing that edge, hence to a cycle through it. By induction the restriction of $f'$ is joined to the restriction $h^*$ of the chosen target $h$. Lift this path successively by Lemma~\ref{lem:lift}. We reach an $H$-good flow $f_0$ equal to $h$ outside $H$. If $H$ is a short cycle, the quotient discrepancy $\quo h-\quo{f_0}$ is $c\chi_H$. If $c\ne0$, add a lift of $c$ on $H$; the final quotient is the nowhere-zero target quotient, so the move is legal. The remaining discrepancy is $y\chi_H$ with $y\in B$ and is corrected in one move if nonzero, protected by the quotient. If $H$ is a subdivision, the quotient discrepancy is a flow supported on $H$. By Proposition~\ref{prop:frame}(i) it is $\pi(Y)$ for an $A$-flow $Y$ on $Q$. Since the target quotient is nonzero everywhere, $Y\in\PL{\Lambda_{f_0}}$. Lemma~\ref{lem:internal} aligns the quotient on $H$. The remaining full discrepancy is a $B$-valued flow supported in $H$. Decompose it into signed cycles of $H$ and add them successively. Every changed edge keeps its nonzero target quotient, so all moves are legal. We reach $h$. This completes the induction. For a disconnected graph apply it component by component. Existence of a target proves nonemptiness as well as connectivity, including the graph with no edges and its single empty flow. \end{proof} \begin{proof}[Proof of Corollary~\ref{cor:exp4}] Write $\Gamma=\Z_4^a\times\F_2^b$ with $2a+b\ge5$. The binary subgroups $2\Gamma$ and $\Gamma[2]$ have dimensions $a$ and $a+b$, respectively. Set $t=\max(a,2)$ and choose \[ 2\Gamma\subseteq B\subseteq\Gamma[2],\qquad \dim_{\F_2}B=t. \] Such a subgroup exists since $a+b\ge t$. The quotient is binary and has dimension $d=2a+b-t\ge3$. If $q=|A|\ge16$, the required ratio holds, since it is equivalent to \[ (9-q)|B|\le8. \] If $d=3$, the possibilities for $(a,b,t)$ are \[ (0,5,2),\quad(1,3,2),\quad(2,1,2),\quad(3,0,3). \] The first three give ratio $3/31<1/9$. The last, $\Gamma=\Z_4^3$ with $B=2\Gamma$, gives $7/63=1/9$. Apply Theorem~\ref{thm:main} in every case. For the binary specialization one can simply take a subgroup of order four in $\F_2^n$; the quotient has order at least eight and the ratio is $3/(2^n-1)\le3/31$. \end{proof} \begin{remark}[Where the hypotheses enter] The quotient's exponent two is used in the symmetric-difference compensation of Lemma~\ref{lem:binwalk} and in the path permissions. The bounds $6<8$ and $7<8$ control paired lists and the same-thread or short-cycle cases. The kernel's exponent two is used in the disjoint-piece argument; $3<4$ permits preparation. The ratio is used only for averaging and extraction. A quotient of order four, a kernel of order two, or a cyclic kernel of order four is not covered by these local proofs. \end{remark} \section{Four-edge-connected binary flows}\label{sec:four} This section proves Theorem~\ref{thm:four}. Put $\Gamma=\F_2^m$, $m\ge3$. All coefficients in this section are binary, so a cycle move adds $\lambda\ind_C$ for $\lambda\ne0$ and is legal exactly when no edge of $C$ currently has value $\lambda$. Applying the same move again reverses it. Cycles are simple, with a parallel pair allowed as a two-edge cycle. We first work without loops and restore them at the end. For an edge map on a graph $H$, its \emph{boundary} at a vertex is the sum of incident values. A prescribed boundary is a map $\beta:V(H)\to\Gamma$ with total sum zero. In the lemmas involving $\beta$, all edge maps are nowhere zero and have that same boundary. Their differences have boundary zero, and cycle additions preserve the boundary. A zero-boundary flow is called \emph{main} if it is joined to a flow omitting a nonzero value. The definition alone does not imply that all main flows on an arbitrary graph belong to one component. We use the Nash-Williams--Tutte tree-packing criterion~\cite{NashWilliamsTrees,TutteTrees}: a finite multigraph contains $k$ edge-disjoint spanning trees if and only if every partition into $q$ nonempty vertex sets has at least $k(q-1)$ edges between parts. In particular a $4$-edge-connected graph has two edge-disjoint spanning trees, since summing its part cuts gives at least $2q$ crossing edges when $q\ge2$. \subsection{Linear changes and two spanning trees} \begin{lemma}[Legal linear changes]\label{lem:linear} Let $j$ have boundary zero, let $d\in\Gamma\setminus\{0\}$, and let $\ell:\Gamma\to\F_2$ be linear with $\ell(d)=0$. Applying $x\mapsto x+\ell(x)d$ to every value of $j$ gives a flow joined to $j$. Consequently a flow joined to one omitting a value $v$ is joined to one omitting any specified nonzero value $b$. \end{lemma} \begin{proof} The edges with $\ell(j(e))=1$ form an even set, and none has value $d$. Decompose this set into edge-disjoint cycles and add $d$ on them successively. These legal moves implement the stated map. Such transvections generate $\operatorname{GL}(m,2)$: elementary row additions generate the invertible binary matrices, with row interchanges obtained by three row additions. An invertible linear map can send $v$ to $b$, proving the final assertion. \end{proof} \begin{lemma}[Two-tree boundary adjustment]\label{lem:twotrees} Let $T$ be a spanning tree of $H$ such that the spanning graph with edge set $R=E(H)\setminus E(T)$ is connected. Fix $b\ne0$ and a boundary $\beta$. \begin{enumerate} \item Any two boundary-$\beta$ edge maps having no value $b$ on $T$ are joined by cycle moves in $H$, through maps with that property. \item For every basis $x_1,x_2,x_3=b,x_4,\ldots,x_m$ and every spanning tree $T_2$ in $R$, there is such a map $g$ omitting $b$ on all of $H$, with only the values $x_1,x_1+x_2$ on $T_2$. \end{enumerate} \end{lemma} \begin{proof} Write $u$ for the quotient of the current map modulo $\langle b\rangle$. Its values on $T$ are nonzero. We use two operations. First, for an edge $e\in R$ with $u(e)\ne0$, add $b$ on its fundamental cycle in $T+e$. This changes no other $R$-edge and preserves every quotient value. It is legal because $e$ and all tree edges on that cycle have nonzero quotient. Call this a \emph{bit correction}. Second, let $t\in T$ and choose $\lambda$ whose quotient $\tau$ is neither zero nor $u(t)$. A simple $R$-path $P$ joins the ends of $t$. At each edge of $P$ currently having full value $\lambda$, use a bit correction to change it to $\lambda+b$. Distinct corrections leave each other's $R$-edges unchanged and preserve all tree quotients. Now add $\lambda$ on $t+P$. This is legal; among tree quotients only $u(t)$ changes, becoming the nonzero value $u(t)+\tau$. Given maps $j,j'$ in (i), perform the second operation on each tree edge where their quotient values differ, using $\tau=u(t)+u'(t)$. It is nonzero and different from $u(t)$. After these operations the quotient discrepancy is a zero-boundary flow supported in $R$. Decompose it into quotient cycle flows within $R$. Lift and perform these cycle additions one at a time, before each addition correcting its currently dangerous edges by bit corrections. The tree quotients remain fixed throughout, although their full values can change by $b$. Edges of $R$ may acquire the full value $b$, which is permitted. At the end the quotient maps agree everywhere. The remaining discrepancy is $b\ind_D$ for an even set $D$. Any $R$-edge of $D$ has nonzero quotient: where the common quotient is zero, both full values are $b$. Correct each such $R$-edge using its bit correction. Their symmetric difference agrees with $D$ off $T$; it therefore equals $D$, because an even edge set supported on a tree is empty. This reaches $j'$ and proves (i). For (ii), set $T_1=T$ and write $\beta_i$ for the coordinates of $\beta$ in the chosen basis. There is a unique binary edge set $D_1$ with boundary $\beta_1$ containing $E(H)\setminus E(T_1)$: solve the remaining boundary on $T_1$ by successively removing leaves. Define $D_2$ similarly using $T_2,\beta_2$. For $i\ge3$, let $D_i$ have boundary $\beta_i$ and be supported on $T_1$. Put \[ g=\sum_{i=1}^m x_i\ind_{D_i}. \] Every edge is outside $T_1$ or outside $T_2$, so it belongs to $D_1$ or $D_2$. Hence $g$ is nowhere zero and never equals $b=x_3$. An edge of $T_2$ belongs to $D_1$ and to no $D_i$ with $i\ge3$; its value is $x_1$ or $x_1+x_2$. The construction has boundary $\beta$, as required. \end{proof} \subsection{Splitting and contraction} \begin{lemma}[Balanced degree-four splitting]\label{lem:splitfour} Suppose $s$ has degree four with edges $e_i=su_i$ to distinct vertices, and $j(e_1)=j(e_2)$, hence $j(e_3)=j(e_4)$. Delete $s$ and replace its two pairs by edges $u_1u_2,u_3u_4$ with their common values, obtaining $U'$ and $j'$. If $j'$ is main, then $j$ is main. \end{lemma} \begin{proof} Lift a cycle of $U'$ by replacing each new edge by its two-edge path through $s$. The lift is one cycle or decomposes into two edge-disjoint cycles meeting only at $s$. A legal coefficient downstairs is absent on both lifted cycles. Add it successively on them. The second is unchanged by the first, and after both moves the values in each pair are equal again. Pair equality is not needed at the intermediate state. This lifts every downstairs step. At the final endpoint, omitting a value downstairs omits it on the two restored paths as well. \end{proof} For a flow $j$ on $U$ and $b\ne0$, call a vertex set $X$, $|X|\ge2$, \emph{good for $(j,b)$} if $U[X]$ has a spanning tree $T_X$ without value $b$ whose edge complement in $U[X]$ is connected and spanning. In $U/X$ identify the vertices of $X$ and delete all edges of $U[X]$, retaining every other edge. \begin{lemma}[Contraction of a good set]\label{lem:goodcontract} If $X$ is good for $(j,b)$ and the restriction of $j$ to $U/X$ is main, then $j$ is main. \end{lemma} \begin{proof} By Lemma~\ref{lem:linear}, a downstairs path reaches a flow omitting $b$. Lift it while keeping $T_X$ free of $b$. A downstairs cycle already forming a cycle of $U$ is handled directly. Otherwise it is an exterior path between distinct vertices $p,p'\in X$. If the step coefficient is $b$, close this path along $T_X$. Every edge on the resulting cycle avoids $b$, and each changed tree value remains nonzero and different from $b$. If the step coefficient is $k\ne b$, use the current boundary of the restriction to $U[X]$. By Lemma~\ref{lem:twotrees}, move inside $U[X]$ to its canonical map for a basis with $x_2=k,x_3=b$ and a spanning tree $T_2$ of $U[X]\setminus E(T_X)$. Its $T_2$-values are $x_1,x_1+k$, both different from $k$. Close the exterior path through $T_2$ and add $k$. This final move leaves $T_X$ unchanged. Preparations keep the exterior fixed and preserve the required tree condition. Once the exterior omits $b$, apply Lemma~\ref{lem:twotrees} once more inside $X$ to reach a map omitting $b$ there too. This is a full flow omitting $b$. \end{proof} \subsection{Proof of the four-edge-connected theorem} \begin{proof}[Proof of Theorem~\ref{thm:four}] First suppose $G$ is loopless. A graph on one vertex has only its empty flow. On a graph with at least two vertices, the tree-packing criterion gives two edge-disjoint spanning trees. Lemma~\ref{lem:twotrees}(ii), with boundary zero, constructs a nowhere-zero flow, proving nonemptiness. For a fixed $b\ne0$, Lemma~\ref{lem:twotrees}(i) connects all flows omitting $b$. Every main flow is joined to one omitting $b$ by Lemma~\ref{lem:linear}. Thus it remains to prove that every flow is main. Fix $m\ge3$ and suppose a counterexample $(U,j)$ exists with the least number of vertices. No flow joined to $j$ is main. Contracting any vertex set preserves the cut condition, since every cut of the contraction is a cut of $U$. \subsection*{Parallel edges are impossible} If $x,y$ have two parallel edges $e,e'$, then $X=\{x,y\}$ is good for $(j,b)$ for any nonzero $b\ne j(e)$: take $T_X=\{e\}$ and use $e'$ in its connected complement. The contraction has fewer vertices, so its restricted flow is main. Lemma~\ref{lem:goodcontract} gives a contradiction. Hence $U$ is simple. \subsection*{At least two pairings at a degree-four vertex are admissible} Suppose $s$ has incident edges $e_i=su_i$, $1\le i\le4$, with distinct neighbors, and put $p_i=j(e_i)$. Their sum is zero. A pairing $\rho$ of the four edges gives a split graph $U_\rho$ by removing $s$ and joining the two endpoints in each pair. For $\varnothing\ne X\subsetneq V(U)\setminus\{s\}$, let $t$ count the neighbors $u_i$ in $X$ and let $\sigma$ count the pairs of $\rho$ with exactly one endpoint in $X$. Then \[ d_{U_\rho}(X)=d_U(X)-t+\sigma,\qquad d_U(X)\ge4,\qquad d_U(X)+4-2t=d_U(X\cup\{s\})\ge4. \] For $t=0,1,3,4$ these inequalities give a split cut of size at least four. For $t=2$ they do so unless $\sigma=0$ and $d_U(X)\le5$. Call such a set a \emph{blocker} of the pairing: it contains exactly the two neighbors in one pair. Its complement in $V(U)\setminus\{s\}$ is also a blocker of the same cut size. Two different pairings cannot both have blockers. After relabeling and choosing complementary blockers if necessary, suppose $X_1$ contains $u_1,u_2$ only, $X_2$ contains $u_1,u_3$ only, and both cuts have size at most five. Write $e_U(P,Q)$ for the number of edges joining disjoint vertex sets $P,Q$. In $U$, \begin{align*} d_U(X_1)+d_U(X_2) ={}&d_U(X_1\setminus X_2)+d_U(X_2\setminus X_1)\\ &+2e_U\bigl(X_1\cap X_2,V(U)\setminus(X_1\cup X_2)\bigr). \end{align*} The first two terms on the right are at least four, and the final edge count includes $su_1$. All bounds must therefore be equalities: both left cuts have size five, and each difference cut has size four. Writing $I=X_1\cap X_2$, the identity \[ d_U(X_1)=d_U(X_1\setminus X_2)+d_U(I) -2e_U(X_1\setminus X_2,I) \] shows that $d_U(I)$ is odd, hence at least five. Cut submodularity gives $d_U(X_1\cup X_2)\le5$. But this union contains exactly three neighbors of $s$, so \[ d_U(X_1\cup X_2\cup\{s\}) =d_U(X_1\cup X_2)+4-6\le3, \] contradicting the cut condition, with $u_4$ still outside. Thus at least two of the three pairings preserve $4$-edge-connectivity. Call them $12|34$ and $13|24$. Their split graphs have fewer vertices, so all their flows are main. \subsection*{The flow values rule out degree four} For pairing $12|34$, equality $p_1=p_2$ would allow Lemma~\ref{lem:splitfour}, contradicting non-mainness. Thus \[ \lambda_1=p_1+p_2=p_3+p_4\ne0. \] None of the four incident values equals $\lambda_1$, since its partner would then be zero. If a cycle through $s$ using one edge from each pair avoided $\lambda_1$, adding $\lambda_1$ would make the values in both pairs equal. Lemma~\ref{lem:splitfour} would again make the reached flow main. Consequently, in $U-s$ no path avoiding $\lambda_1$ joins $\{u_1,u_2\}$ to $\{u_3,u_4\}$. Grouping components after deletion of the $\lambda_1$-edges gives a partition $A_1,B_1$ of $V(U)\setminus\{s\}$, with $u_1,u_2\in A_1$, $u_3,u_4\in B_1$, and only $\lambda_1$-edges crossing it. Pairing $13|24$ similarly gives $\lambda_2=p_1+p_3$ and a partition $A_2,B_2$ separating $\{u_1,u_3\}$ from $\{u_2,u_4\}$. The set $Q=A_1\cap A_2$ contains $u_1$ and no other neighbor of $s$. Its cut consists of $e_1$ and edges with values among $\lambda_1,\lambda_2$. Conservation across this cut forces \[ p_1\in\{0,\lambda_1,\lambda_2,\lambda_1+\lambda_2\}. \] These alternatives give, respectively, $p_1=0$, $p_2=0$, $p_3=0$, or $p_4=0$, a contradiction. This also covers coincident $\lambda_1,\lambda_2$. No vertex of degree four exists in $U$. \subsection*{The density contradiction} Every vertex now has degree at least five. Write $n=|V(U)|$, $e=|E(U)|$, so $e\ge5n/2$. Choose a least-used nonzero value $b$ and let $N$ be the edges not carrying it. Since $2^m-1\ge7$, \[ |N|\ge\frac67e\ge\frac{15}{7}n\ge2(n-1). \] Choose a vertex set $S$ of least size among those with $|S|\ge2$ and $e_N(S)\ge2|S|-2$, where $e_N(S)$ counts internal edges of $N$. Such a set exists because the full vertex set qualifies. Any proper part of $S$ with at least two vertices has fewer than twice its order minus two internal $N$-edges; a singleton has zero. For every partition of $S$ into $q\ge2$ parts, the number of crossing $N$-edges is therefore at least \[ (2|S|-2)-(2|S|-2q)=2(q-1). \] The tree-packing criterion supplies two edge-disjoint spanning trees in $N[S]$. One is a $b$-free tree, and the other is in its connected spanning complement in $U[S]$. Thus $S$ is good for $(j,b)$. Contracting $S$ reduces the vertex count and preserves the cut condition. Minimality and Lemma~\ref{lem:goodcontract} force $j$ to be main, the final contradiction. This proves the loopless result. With loops, first apply it after deleting the loops. Once the nonloop flow omits a value $b$, change any loop of value $b$ to a different nonzero value by a move on that loop. Loops contribute zero boundary and can also be adjusted independently when joining two flows. Thus the theorem, including nonemptiness and reachability of a value-omitting state, holds with loops. If the graph with no vertices is included, its unique empty flow supplies the additional vacuous case. \end{proof} \subsection{Consequences and the rank-two boundary} \begin{corollary}[Partition density]\label{cor:partition} For $\Gamma=\F_2^m$, $m\ge3$, suppose every partition of $V(U)$ into $q$ nonempty parts has at least \[ \frac{2(2^m-1)}{2^m-2}(q-1) \] crossing edges. Then $\Fl(U,\Gamma)$ is nonempty and connected, and every flow is main. At order eight the coefficient is $7/3$. \end{corollary} \begin{proof} Loops do not affect the partition condition. Delete them first and apply the following argument to the resulting loopless graph; afterward restore their values independently as in the proof of Theorem~\ref{thm:four}. The hypothesis is preserved by contraction and implies the two-tree criterion. Its singleton partition gives enough total edges that deleting a least-used value leaves at least $2(n-1)$ edges. The minimal dense-set argument and good-set contraction in the preceding proof then suffice without the degree-four step. Nonemptiness and connection of main flows follow from the same two-tree lemmas. \end{proof} \begin{corollary}[All qualifying trees and values, at fixed boundary]\label{cor:alltrees} Let $H$ have two edge-disjoint spanning trees, and fix a boundary $\beta$ in $\F_2^m$, $m\ge3$. All nowhere-zero boundary-$\beta$ maps that omit some nonzero value on some spanning tree with connected spanning complement lie in one reconfiguration component. \end{corollary} \begin{proof} For a fixed value and tree, use Lemma~\ref{lem:twotrees}(i). A map omitting $b$ everywhere belongs to the family for every qualifying tree, so these families connect for fixed $b$. The canonical map in part (ii) omits both $b=x_3$ and $x_1+b$: on $T_1$ its second coordinate is one, and off $T_1$ its third coordinate is zero. Given distinct $b,b'$, choose $x_1=b+b'$ and complete $x_1,b$ to the specified basis. The canonical map omits both values and joins their families. This uses no global linear change at nonzero boundary. \end{proof} \begin{corollary}[Rounded partition density at order eight]\label{cor:roundedpartition} Let $U$ be a finite multigraph with at least one vertex. Suppose every partition of $V(U)$ into $r$ nonempty parts, with $c$ crossing edges, satisfies \[ 6c\ge14r-20. \] Then $\Fl(U,\F_2^3)$ is nonempty and connected, and every flow is main. \end{corollary} \begin{proof} First delete all loops, which affect neither the hypothesis nor conservation on the other edges. The loopless one-vertex graph has its unique empty flow, which omits every nonzero value. Suppose $n=|V(U)|\ge2$. For every partition with $r\ge2$, \[ 14r-20>6(2r-3). \] Since the crossing count is integral, it is at least $2r-2$. Thus $U$ has two edge-disjoint spanning trees. Lemma~\ref{lem:twotrees} gives a nowhere-zero flow, and Corollary~\ref{cor:alltrees} joins all value-omitting flows at boundary zero. It remains to prove that every flow reaches one. Suppose the contrary. Among all loopless multigraphs that satisfy the partition hypothesis and carry a nowhere-zero flow that is not main, choose one with the least number of vertices; call it again $U$, with $n$ vertices, and let $j$ be a flow on it that is not main. Then $n\ge2$ by the one-vertex case. The partition hypothesis is preserved by contraction: a partition of the contracted graph lifts to one with the same number of parts and crossing edges. Write $e=|E(U)|$. Each edge belongs to the complements of exactly six of the seven nonzero value classes of $j$. If all seven complements had at most $2n-3$ edges, then \[ 6e\le7(2n-3)=14n-21, \] contrary to the hypothesis on the partition into single vertices. Some nonzero value $b$ therefore has a complement $N$ with at least $2n-2$ edges. Any such value serves; it need not be a least-used one. Choose $S\subseteq V(U)$ of least size with $|S|\ge2$ and $e_N(S)\ge2|S|-2$. For a partition of $S$ into $q\ge2$ parts, minimality bounds the sum of the internal $N$-edge counts in those parts by $2|S|-2q$. Hence at least $2(q-1)$ edges of $N[S]$ cross that partition. The tree-packing criterion gives two edge-disjoint spanning trees in $N[S]$. Both are $b$-free, as all of $N$ is. Take one of them as the tree; the other lies in its edge complement in $U[S]$, which is therefore connected and spanning. So $S$ is good for $(j,b)$. The graph $U/S$ is loopless, has fewer vertices than $U$ and satisfies the partition hypothesis, so the restriction of $j$ to it is main by the choice of $U$. Lemma~\ref{lem:goodcontract} then makes $j$ main, a contradiction. Hence every flow is main. Together with the common component of value-omitting flows, this proves connectedness. Restore the loops independently. Once the nonloop flow omits a value $b$, any loop carrying $b$ can be changed to another nonzero value by a move on that loop. Loop values can also be joined independently, so nonemptiness, connectedness and mainness all persist. This includes a single vertex with loops. \end{proof} \begin{corollary}[A matching criterion at order sixteen]\label{cor:matching} Let $G$ be cubic and $3$-edge-connected, with every three-edge cut the cut at a single vertex. Let $f$ be a nowhere-zero $\F_2^4$-flow. If some perfect matching omits some nonzero value of $f$, then $f$ is joined to a flow omitting a nonzero value on all of $G$. \end{corollary} \begin{proof} Let the matching be $M$, let its missing value be $b$, and choose a hyperplane $K$ not containing $b$. Put $W=\{e:f(e)\notin K\}$ and $W'=E(G)\setminus M$. Both are even sets and both contain every $b$-edge. Hence $W\mathbin\triangle W'$ is even and contains no $b$-edge. Adding $b$ on its edge-disjoint cycles gives a legal path to a flow with quotient support $W'$ modulo $K$. The set $W'$ is a $2$-factor. Contract each of its cycles, retaining loops. A nontrivial cut in the quotient lifts to a cut in $G$ whose two sides are unions of whole $W'$-cycles. Neither side is a single vertex, so its size cannot be three; it is at least four. The restriction to the matching is a nowhere-zero $K$-flow on this quotient. By Theorem~\ref{thm:four}, it is joined to one missing some $v\in K\setminus\{0\}$. Lift each quotient cycle by paths in its visited $W'$-cycles. The resulting $K$-move is safe on those paths because their full values lie outside $K$, and safe on the matching by its prescribed downstairs endpoints. A quotient loop lifts to a chord and an arc of its $W'$-cycle. Thus the path lifts while all $W'$-values stay outside $K$. Its endpoint omits $v$ both on the matching and on $W'$, proving the claim. In particular, in any component containing no main flow, every perfect matching carries all fifteen values at every state of that component. \end{proof} \begin{remark}[Mainness at rank two]\label{rem:ranktwo} The omitted-value conclusion fails for $\F_2^2$. Let its nonzero values be $a,b,c=a+b$. On $K_6$, alternate $b,c$ around a Hamiltonian six-cycle and put $a$ on its other nine edges. Each vertex has sum $3a+b+c=0$, giving a nowhere-zero flow. Every nontrivial cut has size $s(6-s)\ge5$. But omitting a value would leave at most two independent nonzero values, each occurring an even number of times at every vertex by conservation, contrary to degree five. Thus no flow on $K_6$ omits a nonzero value. This example concerns only the omitted-value conclusion, not the connectivity of $\Fl(K_6,\F_2^2)$. The proof above needs three coordinates in the canonical two-tree construction and at least seven nonzero values in its density count. \end{remark} \section{At most five exceptional star cuts}\label{sec:fivestars} Degrees in this section are measured after deleting loops. Let $\mathcal C_5$ consist of finite multigraphs with at least one vertex satisfying the following condition: either the graph is a singleton, or it has a designated set $T$ of at most five degree-three vertices such that every edge cut determined by a nonempty proper vertex set has size at least four, except cuts of the form $\delta(\{t\})$ for $t\in T$. Complementary shores determine the same edge cut. In particular, every nonsingleton graph in $\mathcal C_5$ is connected and has minimum cut at least three. \begin{theorem}[Five exceptional stars]\label{thm:fivestars} For $U\in\mathcal C_5$ and $m\ge3$, the graph $\Fl(U,\F_2^m)$ is nonempty and connected. Every flow is joined to a flow whose values lie among the three nonzero elements of a two-dimensional subspace. \end{theorem} The designation concerns cuts, not a literal count of all degree-three vertices: the three-edge dipole belongs to $\mathcal C_5$ with just one designated vertex. For a connected graph on at least three vertices, different degree-three vertices have different star cuts, since equality of two such edge sets would make them an entire dipole component. Consequently every degree-three vertex is then designated. \subsection{Contraction, two trees, and a short cycle} The class $\mathcal C_5$ is closed under contraction of a vertex set, with internal edges deleted. A surviving three-edge cut lifts to an original designated star cut. Its designated singleton is an unswallowed vertex, or no such cut survives the contraction; a singleton contraction is the identity. The surviving designated vertices retain degree three and witness the quotient condition. Their number does not increase. Every member has two edge-disjoint spanning trees. For a partition into $q\ge3$ parts, only a designated singleton part can have a cut of size three: the complement of a singleton cannot be a part of a partition with at least three parts. The sum of part-cut sizes is at least $4q-5$; being even, it is at least $4q-4$. Thus there are at least $2(q-1)$ crossing edges. For $q=2$ there are at least three, and the tree-packing criterion applies. These trees also give a three-value anchor. For disjoint spanning trees $T_1,T_2$, let $D_i$ be the unique even set containing $E(U)\setminus E(T_i)$, obtained by solving the remaining binary boundary on $T_i$. If $x_1,x_2$ are independent, then \[ g=x_1\ind_{D_1}+x_2\ind_{D_2} \] is nowhere zero: every edge is outside at least one tree. Its values lie in $\{x_1,x_2,x_1+x_2\}$. By Lemmas~\ref{lem:linear} and~\ref{lem:twotrees}, all main flows on a member of $\mathcal C_5$ lie in one component. It is therefore enough to prove that every flow is main. We will also use the following binary short-cycle lifting fact. It is the needed direction of the contraction criterion of~\cite[Theorem 5.9, Theorem 5.11 and Corollary 5.14]{CLSWX}; we include its elementary proof. \begin{lemma}[Lifting across a short induced cycle]\label{lem:shortbinary} Let $A$ be elementary abelian of exponent two, and let $C$ be an induced cycle of length $r<|A|$ in a loopless graph $H$. If $\Fl(H/C,A)$ is nonempty and connected, then so is $\Fl(H,A)$. \end{lemma} \begin{proof} For any quotient flow the prescribed boundary on $C$ has total zero, so it has a solution on $C$. Adding a constant around $C$ changes that solution; at most $r$ constants produce a zero edge. Hence every quotient flow has a nowhere-zero extension. Two such extensions of one quotient flow differ by a constant on $C$, and are adjacent or equal. Consider a legal quotient cycle move of coefficient $\lambda$. If it is already a cycle upstairs, perform it there. Otherwise its lift off $C$ is a path $P$ between distinct vertices of $C$, avoiding $\lambda$. Let $S,S'$ be the two complementary arcs of $C$, and let $V$ be the set of current values on $C$. There exists $c\notin V$ for which at least one of \[ c\notin\lambda+f(S),\qquad c\notin\lambda+f(S') \] holds. Indeed, if both tests failed for every $c\in M=A\setminus V$, each value in $\lambda+M$ would occur on both disjoint arcs. The number of distinct repeated values among the $r$ edges is at most $r-|V|$, whereas $|M|=|A|-|V|>r-|V|$, a contradiction. Add $c$ around $C$, doing nothing if $c=0$. The cycle stays nowhere zero. Now add $\lambda$ on $P$ together with the permitted arc. This is legal and realizes the quotient move. The connected fibres adjust the final extension, proving the assertion. \end{proof} \subsection{Proof of the five-star theorem} \begin{proof}[Proof of Theorem~\ref{thm:fivestars}] Remove loops first; they can later be changed independently to anchor values. Fix $m\ge3$ and choose a nonmain flow $j$ in this class on a graph $U$ with the least possible number $n$ of vertices. The singleton is immediate. A pair of vertices with parallel edges is a good set for $(j,b)$ for some $b$ different from the value on one of these edges. Contraction, minimality and Lemma~\ref{lem:goodcontract} give a contradiction. This removes the dipole case before counting designated vertices. Thus $U$ is simple. A shortest cycle of length at most seven is induced and shorter than $2^m$; contraction stays in $\mathcal C_5$, and Lemma~\ref{lem:shortbinary} and minimality give a contradiction. Hence $U$ has minimum degree three and girth at least eight. \subsection*{Degree-four vertices must meet degree-three vertices} Suppose $s$ has degree four and none of its distinct neighbors $u_1,\ldots,u_4$ has degree three. None of these neighbors is designated. For a pairing $\rho$ of the incident edges and a nontrivial shore $X\subset V(U)\setminus\{s\}$, let $t$ be the number of neighbors in $X$ and $\sigma$ the number of paired edges crossing $X$. Then \[ d_{U_\rho}(X)=d_U(X)-t+\sigma, \qquad d_U(X\cup\{s\})=d_U(X)+4-2t. \] For $t=0$, a cut retains its size and any exceptional singleton remains allowed. For $t=1,2$, the shore contains a nondesignated neighbor and its complement contains several vertices, so $d_U(X)\ge4$. For $t=3$, the complement of $X\cup\{s\}$ contains a nondesignated neighbor, so $d_U(X)-2\ge4$. For $t=4$, the split cut is $d_U(X\cup\{s\})$, at least four unless its complement is a designated singleton, whose star remains an allowed exception. The only possible new forbidden cuts therefore have $t=2$, $\sigma=0$, and $d_U(X)\le5$. Two different pairings cannot both have such blockers. Choose shores $X_1,X_2$ containing respectively just $u_1,u_2$ and just $u_1,u_3$. The difference shores contain nondesignated neighbors $u_2,u_3$, and their cuts have size at least four. The uncrossing identity used in the proof of Theorem~\ref{thm:four}, with $su_1$ joining the intersection to the outside union, forces \[ d_U(X_1)=d_U(X_2)=5, \qquad d_U(X_1\setminus X_2)=d_U(X_2\setminus X_1)=4. \] The intersection cut has odd size and is at least four, hence at least five. Submodularity gives $d_U(X_1\cup X_2)\le5$, and adding $s$ lowers this by two. The resulting cut has size at most three, with several vertices on one shore and nondesignated $u_4$ on the other. This violates the class condition. At least two pairings therefore preserve $\mathcal C_5$. Their split graphs are smaller. A balanced preserved pairing would make $j$ main by Lemma~\ref{lem:splitfour}. For an unbalanced pairing $12|34$, put $p_i=j(su_i)$ and $\lambda_1=p_1+p_2=p_3+p_4\ne0$. A crossing cycle through $s$ avoiding $\lambda_1$ would balance this pairing and again give mainness. Thus, in $U-s$, the pairs are separated by a cut of $\lambda_1$-edges. The other preserved pairing $13|24$ gives a cut of $\lambda_2=p_1+p_3$-edges. The intersection of the shores containing $u_1$ has precisely that neighbor of $s$, so conservation gives \[ p_1\in\{0,\lambda_1,\lambda_2,\lambda_1+\lambda_2\}. \] Every alternative makes one of $p_1,p_2,p_3,p_4$ zero, exactly as in the degree-four step of Theorem~\ref{thm:four}. This contradiction proves the claim about neighbors. Let $k$ be the actual number of degree-three vertices, $a$ the number of degree-four vertices, and $h=n-k-a$ the remaining number. We have $k\le5$ and $a\le3k$. Writing $e=|E(U)|$, \begin{equation}\label{eq:five-density-lower} 2e\ge3k+4a+5h=5n-2k-a\ge5n-5k. \end{equation} \subsection*{An integer density bound} For any nonzero $b$, let $N_b$ be the set of edges not carrying $b$. If $|N_b|\ge2(n-1)$, the minimal dense-set argument in Theorem~\ref{thm:four} supplies a vertex set with two edge-disjoint spanning trees in $N_b$. It is good for $(j,b)$, and its contraction contradicts minimality via Lemma~\ref{lem:goodcontract}. Hence $|N_b|\le2n-3$ for every $b$. For a least-used value, $|N_b|\ge6e/7$, so \[ 5n-5k\le2e\le\frac73(2n-3), \qquad n\le15k-21\le54. \] \subsection*{The girth count} Root a search at the two ends $u,v$ of an edge, without traversing that edge again, and grow three levels on each side. Girth at least eight makes all counted vertices distinct: a collision within a side creates a cycle of length at most six, and a collision across sides one of length at most seven. Every intermediate vertex has at least two forward neighbors. Thus \begin{equation}\label{eq:edge-rooted} n\ge2+(d(u)+d(v)-2)(1+2+4)=7(d(u)+d(v))-12. \end{equation} In particular $n\ge30$. Two adjacent vertices of degree at least five would give $n\ge58$, contradicting $n\le54$. The $h$ vertices of degree at least five are consequently independent. Counting their incident edges gives \[ 5h\le3k+4a\le15k, \qquad n=k+a+h\le7k\le35. \] If $h>0$, an edge from a vertex of degree at least five has degree sum at least eight and gives $n\ge44$, again impossible. Hence $h=0$, and $n=k+a\le4k\le20<30$. This final contradiction shows every flow is main. The two-tree argument connects all flows to the three-value anchor. Restoring loops, and independently assigning them any of the anchor values, proves the theorem as stated. \end{proof} \begin{corollary}[A sufficient support condition at order sixteen]\label{cor:fivesupport} Let $G$ be cubic and $3$-edge-connected, with every three-edge cut a vertex star, and let $f$ be a nowhere-zero $\F_2^4$-flow. If an even edge set $W'$ contains all edges of some nonzero value $b$ and leaves at most five vertices uncovered, then $f$ is main. Equivalently, it suffices that $G-f^{-1}(b)$ have a spanning subgraph with every degree odd and at most five vertices of degree three. \end{corollary} \begin{proof} Choose a hyperplane $K$ avoiding $b$, and let $W=\{e:f(e)\notin K\}$. The even set $W\mathbin\triangle W'$ has no $b$-edge. Adding $b$ on its cycle components gives a flow whose outside-$K$ support is $W'$. Since $G$ is cubic, $W'$ is a union of disjoint cycles. Contract them, retaining loops. The remaining edges form a nowhere-zero $K$-flow. Every three-edge cut of the quotient lifts to a vertex star of $G$; its singleton shore must be a vertex outside $W'$. These at most five vertices designate the exceptional cuts. In particular their degrees, after deleting loops, remain three. The quotient belongs to $\mathcal C_5$. Theorem~\ref{thm:fivestars} joins this quotient flow to one omitting a nonzero $K$-value. Lift each quotient cycle with arcs in the contracted cycles of $W'$, as in Corollary~\ref{cor:matching}. Those arcs stay outside $K$, so the lifted $K$-moves are safe. A quotient loop lifts to a chord and an arc. The endpoint omits the chosen value everywhere. Finally, the complement of $W'$ has degree one or three at every vertex, with degree three exactly at uncovered vertices, and contains no $b$-edge. Taking complements proves the equivalent formulation. We do not show that this spanning subgraph exists. \end{proof} \section{A conditional four-edge-cut reduction at order sixteen}\label{sec:fourcut} In this section $\Gamma=\F_2^4$, and all graphs are finite and loopless; parallel pairs count as cycles. A cut is always determined by a nonempty proper vertex set. All statements of joining refer to the unrestricted flow graph: a constructed path may leave the boundary types being compared. \begin{theorem}[Conditional four-cut reduction]\label{thm:fourcut} Suppose that $G$ satisfies the following hypotheses. \begin{enumerate}[label=\textup{(H\arabic*)}] \item $G$ is cubic and $3$-edge-connected, and every three-edge cut is a vertex star. \item For every bridgeless graph $G'$ with fewer odd-degree vertices than $G$, any two nowhere-zero $\Gamma$-flows of $G'$ are joined. \item Some vertex set $X$ has $d_G(X)=4$, $|X|\ge4$, and $|V(G)\setminus X|\ge4$. \end{enumerate} Then $\Fl(G,\Gamma)$ is nonempty and connected. \end{theorem} Only six smaller graphs will use (H2): the three wire closures of each side of the cut. They are cubic and bridgeless, with at most $|V(G)|-4$ vertices, hence fewer odd-degree vertices. They need not inherit (H1). For existence we use Jaeger's theorem that every bridgeless multigraph has a nowhere-zero binary rank-three flow~\cite{Jaeger,JaegerSurvey}; an isomorphism gives a flow in any specified order-eight subgroup of $\Gamma$. \subsection{Sides, linear transport, and classes} In a cubic $3$-edge-connected graph, $G-v$ is connected: different components would each require at least three edges to $v$, exceeding its degree. Consequently a cut equal to the star of $v$ has one shore $\{v\}$. This explains the singleton-shore interpretation of (H1). Write the four cut edges as $e_1,\ldots,e_4$, and the two sides as $Z_1,Z_2$. For either side $Z$, put $D=G[Z]$ and let $z_i$ be the endpoint of $e_i$ in $Z$. The terminals are distinct. A vertex incident with three cut edges would leave a one-edge cut on deleting it from its shore. A vertex incident with two would leave a three-edge cut with at least three and five vertices on its shores, violating (H1). The graph $D$ is connected, since two components would need at least six cut edges. It is also bridgeless: an internal bridge separating $t$ terminals from $4-t$ gives cuts in $G$ of sizes $t+1$ and $5-t$. These force $t=2$, and then give a three-edge cut with neither shore a singleton, a contradiction. A \emph{side flow} assigns nonzero values to $E(D)$ and the four cut edges, satisfying conservation at each vertex of $Z$. Its ordered boundary $b=(b_1,\ldots,b_4)$ has sum zero. A full flow is a pair of side flows with identical boundaries. An \emph{inner move} is a cycle move entirely in $D$. Write $f\equiv g$ when side flows with the same boundary are joined by inner moves. Reversibility makes this an equivalence relation. There are five boundary types: $Q$, all four entries equal; $P_\pi$, two distinct equal pairs for one of the three pairings $\pi$; and $D_4$, all four entries distinct. In type $D_4$ any three entries are independent, since a dependence would make the fourth entry zero. We repeatedly use the following even-set observation. If nowhere-zero endpoint maps of the same boundary differ by $d\ind_E$ on an internal even set, they are inner-equivalent: every edge of $E$ avoids $d$ at the first endpoint, since otherwise the second endpoint would be zero. Its edge-disjoint cycle decomposition gives legal successive moves. This also holds for full flows. \begin{lemma}[Internal stabilizers and linear transport]\label{lem:cutlinear} Let $f$ be a side flow with boundary $b$. \begin{enumerate} \item An invertible linear map fixing every $b_i$ sends $f$ to an inner-equivalent flow. \item Let $I\subseteq\{1,2,3,4\}$ have size two or four. Suppose $R\subseteq E(D)$ has odd degree exactly at $z_i$ for $i\in I$, avoids $\lambda\ne0$, and $b_i\ne\lambda$ for $i\in I$. Let $f'$ add $\lambda$ on $R$ and those cut edges. If a functional $\ell$ satisfies \[ \ell(b_i)=\ind_I(i),\qquad \ell(\lambda)=0, \] then $f'\equiv t(f)$ for $t(x)=x+\ell(x)\lambda$. \end{enumerate} \end{lemma} \begin{proof} Put $U=\langle b_1,\ldots,b_4\rangle$. In a basis extending one of $U$, its pointwise stabilizer has matrices \[ \begin{pmatrix}I&A\\0&B\end{pmatrix},\qquad B\text{ invertible}. \] Row additions on the complementary coordinates generate $B$; the remaining maps add vectors of $U$ to complementary basis vectors. Thus this stabilizer is generated by transvections $x\mapsto x+\ell(x)d$ with $\ell|_U=0$ and $\ell(d)=0$. For a side flow, the support of $\ell(f)$ is internal and even, and avoids $d$. Each generator therefore acts by inner moves, proving (i). For (ii), the internal set $S=\{e:\ell(f(e))=1\}$ has the same odd terminals as $R$. The states $t(f)$ and $f'$ have identical boundaries and differ by $\lambda$ on the internal even set $R\mathbin\triangle S$. The even-set observation proves the claim. \end{proof} For side flows of the same type, define $f\sim g$ if $L(f)\equiv g$ for an invertible linear map $L$ taking the boundary of $f$ to that of $g$ in order. Linear maps preserve inner paths, so this is an equivalence relation; call its equivalence classes \emph{side classes}. Every side class has a representative over every boundary of its type, by transitivity on one, two, or three independent vectors. By Lemma~\ref{lem:cutlinear}(i), two representatives of one class at the same boundary are inner-equivalent. In particular, full flows of the same boundary type with the same two side classes are joined. First apply a global invertible linear map to align their boundaries, using Lemma~\ref{lem:linear}; then use inner paths on the two sides. The global path may pass through other boundary types. Lemma~\ref{lem:cutlinear}(ii) also shows that any step meeting its functional conditions preserves the side class. \subsection{Two walls and a direct move to equal pairs} For $\lambda\ne0$, call $Y\subseteq Z$ a \emph{$\lambda$-wall} if every edge of $D$ leaving $Y$ has value $\lambda$. Summing the side-flow equations on the component reachable from $z_i$ in the graph without $\lambda$-edges shows that its terminal-value sum is $0$ or $\lambda$. \begin{lemma}[Wall intersection]\label{lem:walls} If $Y$ is a $\lambda$-wall and $Y'$ a $\lambda'$-wall, then \[ \sum_{i:z_i\in Y\cap Y'}b_i\in\langle\lambda,\lambda'\rangle. \] \end{lemma} \begin{proof} Sum conservation over $Y\cap Y'$. Internal edges cancel. Every remaining internal edge leaves $Y$ or $Y'$, so has value $\lambda$ or $\lambda'$. The cut edges contribute precisely the displayed sum. The empty intersection gives zero as well. \end{proof} For a type-$D_4$ boundary and a pairing $\pi=\{i,k\}\mid\{j,l\}$, put $d_\pi=b_i+b_k=b_j+b_l$. It is nonzero and equals no boundary entry. A \emph{$\pi$-wall} is a $d_\pi$-wall separating the two terminal pairs. Either all terminals are joined by $d_\pi$-free paths, or a $\pi$-wall exists: the reachable-terminal count excludes sizes one and three, and a reachable pair containing $i$ must be $\{i,k\}$. A side flow of type $D_4$ has a wall for at most one pairing. Otherwise complement two walls so both contain terminal $1$, with other terminals $k,k'$, where $k\ne k'$. Their terminal intersection is just $1$. Lemma~\ref{lem:walls} forces \[ b_1\in\{0,b_1+b_k,b_1+b_{k'},b_k+b_{k'}\}. \] The first three possibilities make a boundary entry zero; the last makes the remaining fourth entry zero. All are impossible. Across the two sides at most two of the three pairings have a wall. Choose a pairing free on both sides. Their $d_\pi$-free paths from $z_i$ to $z_j$, together with $e_i,e_j$, form one legal cycle of $G$. Adding $d_\pi$ changes $b_i$ to $b_k$ and $b_j$ to $b_l$. Thus every full flow of type $D_4$ reaches type $P_\pi$ in one move. No smaller-graph hypothesis is used here. \subsection{Wire closures and their class graphs} For a side $Z$ and pairing $\pi=\{i,k\}\mid\{j,l\}$, form $H(Z,\pi)$ by adding to $D$ the two \emph{wire} edges $w=z_iz_k$ and $w'=z_jz_l$. This is a loopless cubic bridgeless graph on $|Z|\le |V(G)|-4$ vertices: $D$ is connected and bridgeless, and a wire can be deleted without disconnecting the graph. Parallel edges are permitted. Hypothesis (H2) makes its flow graph connected. Its flows are exactly the side flows of types $Q$ and $P_\pi$, identifying each wire value with the two corresponding boundary entries. Both types occur. Take a nowhere-zero flow of $H(Z,\pi)$ in an order-eight subgroup $K$. If the wire values differ, this already gives type $P_\pi$. If they agree, add a value outside $K$ on the cycle consisting of $w$ and a path in $D$ between its ends. This avoids $w'$, and all changed values lie outside $K$, so the two wire values are now different. For type $Q$ with any prescribed $\beta\ne0$, choose an order-eight subgroup $K$ avoiding $\beta$ and a nowhere-zero $K$-flow $h$ on $D$. Choose an edge set $J$ of $D$ with odd degree exactly at its four terminals, for example the symmetric difference of two paths pairing them. Put \[ p(e)=h(e)+\beta\ind_J(e)\quad(e\in E(D)), \qquad p(e_i)=\beta\quad(1\le i\le4). \] Off $J$ the internal values are nonzero in $K$; on $J$ they lie in the nonzero coset $\beta+K$. The extra boundary at each terminal is balanced by its cut edge. Hence $p$ is a nowhere-zero side flow of type $Q$. Let $B(Z,\pi)$ be the bipartite graph whose vertices are side classes of types $Q$ and $P_\pi$, with an edge when representatives are related by a cycle move in $H(Z,\pi)$ using exactly one wire and changing the type. Edges are undirected because every move reverses. Every cycle move of $H(Z,\pi)$ either keeps its side class or follows an edge of $B(Z,\pi)$. With no wire it is inner. With both wires it preserves the type: for coefficient $\lambda$, a functional equal to one on both wire values and zero on $\lambda$ exists. For equal wire values this uses $\lambda\ne\beta$; for distinct values $a,c$, the only elements of their span where this functional is one are $a,c$, both excluded by legality. Extend the functional if $\lambda$ is outside that span, and apply Lemma~\ref{lem:cutlinear}(ii). With just one wire, a $Q$ state goes to $P_\pi$. From a $P_\pi$ state with wire values $a,c$, the coefficient $a+c$ gives type $Q$. Every other legal coefficient preserves $P_\pi$: choose $\ell(a)=1$, $\ell(c)=0$, $\ell(\lambda)=0$. The only obstructions in the span are $a$ and $a+c$, already excluded. The same transport lemma applies. Since $\Fl(H(Z,\pi),\Gamma)$ is connected and both types occur, $B(Z,\pi)$ is connected, has both parts nonempty, and has no isolated vertex. \subsection{Synchronizing the two sides} Two facts at a type-$Q$ boundary $(\beta,\beta,\beta,\beta)$ control which wire an exit uses. Fix $\lambda\ne0,\beta$. Every component after deleting the $\lambda$-edges contains an even number of terminals, by the terminal-sum count. Consequently, if one pair of $\pi$ is joined by a $\lambda$-free path, so is its complementary pair. Moreover the two choices lead to the same side class. Let $P$ and $P'$ be $\lambda$-free paths for the two complementary pairs, and let $g,g'$ be the two resulting side flows after changing the chosen path and its two cut edges. Choose $\ell(\beta)=1$, $\ell(\lambda)=0$, and $t(x)=x+\ell(x)\lambda$. This exchanges $\beta$ and $\beta+\lambda$. If $S=\{e\in E(D):\ell(f(e))=1\}$ for the starting flow, then on internal edges \[ t(g)=f+\lambda\ind_{P\mathbin\triangle S}, \qquad g'=f+\lambda\ind_{P'}. \] On the first pair of cut edges $t(g)$ has value $\beta$, and on the complementary pair it has value $\beta+\lambda$, exactly as $g'$. Their internal difference is $\lambda$ on the even set $P\mathbin\triangle S\mathbin\triangle P'$, so $t(g)\equiv g'$. Thus $g\sim g'$. Two paths for the same pair similarly give inner-equivalent endpoints by their even symmetric difference. \begin{lemma}[Paired class transitions]\label{lem:syncclasses} Fix $\pi$. On each side let a class $q_s$ of type $Q$ be adjacent to a class $p_s$ of type $P_\pi$ in $B(Z_s,\pi)$. Any full flow with side classes $(q_1,q_2)$ is joined to any full flow with side classes $(p_1,p_2)$. \end{lemma} \begin{proof} Let the starting common boundary be $\beta$. Transport each edge witness linearly to this boundary. The fixed-boundary class property gives inner paths from the starting sides to these witnesses. Let their exit coefficients be $\lambda_1,\lambda_2$, each nonzero and different from $\beta$. An invertible linear map fixing $\beta$ and taking $\lambda_2$ to $\lambda_1$ exists by extending independent vectors to bases. Lemma~\ref{lem:cutlinear}(i) implements it internally on side two, without changing either prescribed class. If its witness uses the complementary wire, the preceding two facts supply a path for the same wire used by side one, still reaching $p_2$. We now have two simple internal paths with the same terminal pair and coefficient. Together with the two corresponding cut edges they form a legal full cycle, reaching the prescribed pair of classes. Global boundary alignment followed by inner paths joins this endpoint to any requested representative. Reversibility gives the assertion in the opposite direction as well. \end{proof} For fixed $\pi$, consider pairs of vertices from $B(Z_1,\pi)$ and $B(Z_2,\pi)$ with matching types. Join two pairs when both coordinates are adjacent. This graph is connected. Between two pairs of type $Q$, the two coordinate walks have even lengths; pad the shorter by going to a neighbor and back until their lengths agree. The simultaneous walk has matching types at every step. Every pair of type $P_\pi$ has a neighbor of type $Q$, since neither factor has an isolated vertex. Every such pair has a full-flow representative: the two classes admit a common ordered boundary. Full flows representing one pair are joined by boundary alignment and inner moves, and adjacent pairs lift by Lemma~\ref{lem:syncclasses}. Hence all full flows of types $Q$ and $P_\pi$ lie in one component of the unrestricted flow graph. The argument uses paths in the whole flow graph, not only in the subgraph induced by those boundary types. \begin{proof}[Completion of Theorem~\ref{thm:fourcut}] The type-$Q$ constructions on the two sides with one common $\beta$ give a full nowhere-zero flow. For each of the three pairings the preceding argument puts all $Q$ and $P_\pi$ flows into one component. They share the nonempty type-$Q$ family, so these components agree. The direct wall argument brings every type-$D_4$ flow into this component. The boundary types exhaust all possibilities, proving nonemptiness and connectedness. \end{proof} \begin{corollary}[Conditional least-core consequence]\label{cor:fourcore} Let $G$ be a bridgeless counterexample to $\F_2^4$-flow connectivity with the minimum number of odd-degree vertices among all bridgeless counterexamples. If $G$ satisfies \textup{(H1)}, then every four-edge cut is the cut around two vertices joined by one edge. Every cut with a cycle on both sides has size at least five. \end{corollary} \begin{proof} Minimality gives (H2), so (H3) is impossible. In a cubic graph a shore of a four-cut has even order, since $3|X|-2|E(G[X])|=4$. If its order is less than four, it is two and has exactly one internal edge, hence no cycle. Hypothesis (H1) excludes smaller cuts with cycles on both sides. These observations prove both assertions. \end{proof} The normal-form interface is separate from this four-cut proof. Lemma~\ref{lem:widecore} in Section~\ref{sec:oddbound} supplies a connected simple cubic representative at the same minimum odd count, satisfying (H1) and the wide hypothesis (H2). Section~\ref{sec:oddbound} justifies the general binary interface and gives order-sixteen girth at least seventeen and core size at least $766$, independently of the four-cut theorem. Section~\ref{sec:c17} then uses Corollary~\ref{cor:fourcore} to exclude seventeen-cycles, giving girth at least eighteen and at least $1022$ vertices. Section~\ref{sec:smallshores} later strengthens the size bound to $3038$, Section~\ref{sec:b3996} to $3998$, and Section~\ref{sec:finalquotient} to $19024$. The resulting representative is cyclically five-edge-connected. These statements do not show that every minimum-odd-count representative is cubic, and they do not settle connectivity across a five-edge cut. \section{A cycle-order improvement of the odd-vertex bound}\label{sec:oddbound} Write $A=\F_2^m$ and $q=|A|=2^m$, with $m\ge3$. In particular, $q$ denotes the group \emph{order}, not its rank. Graphs are finite multigraphs; loops contribute two to degree and zero to binary conservation. As throughout, a cycle is a connected $2$-regular subgraph, including a loop or a digon, and a legal move adds a nonzero coefficient absent from that cycle. In this section and in Section~\ref{sec:c17}, $\Sigma$ denotes a set of group values, not a step system as in Definition~\ref{def:stepsystem}; in Section~\ref{sec:finalquotient} it denotes a sum of degrees. \begin{theorem}[Improved odd-vertex bound]\label{thm:oddbound} Let $A=\F_2^m$ have order $q\ge8$. If each connected component of a finite bridgeless multigraph $G$ has at most \[ B_*(q)=3\cdot2^{q/2}-4 \] odd-degree vertices, then $\Fl(G,A)$ is nonempty and connected. \end{theorem} \begin{corollary}\label{cor:44-764} The conclusion holds for $A=\F_2^3$ when each component has at most $44$ odd-degree vertices, and for $A=\F_2^4$ when each has at most $764$ odd-degree vertices. \end{corollary} The proof uses \cite[Theorem~1.2(c)]{Eulerian}, an unpublished manuscript, through Lemma~\ref{lem:widecore}. The cycle-index argument below requires $q\ge8$, so the theorem is stated only for $q\ge 8$. For $q\ge32$, Corollary~\ref{cor:exp4} already gives the stronger universal result. The proof of Theorem~\ref{thm:oddbound} uses no four-cut reduction. Its new step handles a cycle of length \emph{equal} to the group order, rather than only shorter cycles. The order-sixteen value $764$ remains a consequence of this general formula; Theorem~\ref{thm:1020} later improves it using the four-cut theorem. \subsection{The general least-odd-count interface} We first state the local inputs precisely. The two-odd-vertex case is~\cite[Theorem 1.2(c)]{Eulerian}: a bridgeless graph in which each component has at most two odd-degree vertices has a connected flow graph over every finite abelian group of order at least four. The other local inputs, balanced splitting of two equal-valued nonloop incident edges, the degree-four local entry and family joining, and triangle extension and projection, are written out and proved below; they follow the corresponding lemmas of~\cite[Part 2]{Eulerian} (Lemmas 17.1, 17.2, 18.1 and 18.2, Proposition 18.3 and Lemma 21.1; for exponent two, also Lemma 69.2 of its Part 6), which we cite for credit, and are local statements, not an invocation of a bounded theorem beyond its range. The small-cut gluing is proved below as Lemma~\ref{lem:smallcutbinary}. We explain the use of these inputs, including the rank requirements, to preserve the quantifiers of minimality. Balanced splitting replaces edges $uv,vw$ of equal outward value by an edge $uw$, retaining $v$; $u=w$ is permitted and gives a loop. It decreases edge count, preserves degree parity and does not increase maximum degree. A cycle move downstairs lifts through the two-edge path, possibly decomposing into more than one cycle upstairs; the common coefficient is safe on every lifted edge. Thus connectedness downstairs joins the corresponding balanced family in the full flow graph. At a loopless degree-four vertex, every binary flow reaches a balanced pair in at most one move, and if every nonempty balanced-pair family is joined then all flows are joined. This is~\cite[Lemma 69.2]{Eulerian}; we include its short proof, which needs no rank-four assumption. With four distinct incident values $a,b,c,d$, their sum is zero and they form a coset of $K=\langle a+b,a+c\rangle$, disjoint from $K$. The nonzero quotient support modulo $K$ is bridgeless: a bridge in that support would violate conservation across its cut. A cycle through the vertex in this support permits a safe $K$-increment that matches its two used values to the two unused values. If all four values can become equal, the three pairing families intersect. Otherwise a balanced state has values $a,a,b,b$ with $a\ne b$; the support modulo $\langle a+b\rangle$ has a cycle through the vertex, and its safe $(a+b)$-move links this family to a different one. The graph of nonempty pairing families has at most three vertices and no isolated vertex, so is connected. No rank-four assumption or bound on the number of odd vertices occurs in this local reasoning. For the triangle step, remove loops and put \[ \Psi(G)=\sum_{v\in V(G)}\max\{\deg(v)-4,0\}. \] Replace an even-degree vertex of degree $d\ge6$ by a triangle with external groups $2,2,d-4$, giving degrees $4,4,d-2$. For odd $d\ge5$, use groups $1,2,d-3$, giving degrees $3,4,d-1$. The odd count is preserved and $\Psi$ decreases. Every prescribed exterior flow extends over the triangle: its three edge values have the form of one parameter plus fixed offsets, so at most three parameters are forbidden, fewer than $q$. An upstairs cycle projects to an even support downstairs; decompose it into edge-disjoint cycles and retain its coefficient. Each retained edge remains old or final during this path, so no zero is created. A projected step may be stationary. These are the extension and projection properties of~\cite[Lemma 21.1]{Eulerian}. \begin{lemma}[Pinning and gluing at two or three edges]\label{lem:smallcutbinary} For $A=\F_2^m$, $m\ge3$, two joined flows agreeing on the star of a loopless vertex of degree two or three are joined by moves avoiding that vertex. Consequently a bridgeless graph with a two- or three-edge cut and connected shores has connected flow graph whenever both capped shores have connected flow graphs. \end{lemma} \begin{proof} Normalize a path between the two given flows, one move at a time. At a degree-two vertex the star values are $a,a$; a through-step of coefficient $t$ excludes $0,a$, so choose $\ell(a)=1$, $\ell(t)=0$. At degree three the used star values are $a,b$ and the unused one is $a+b$. Legality excludes $t=a,b$, so $\ell(a)=\ell(b)=1$, $\ell(t)=0$ is consistent: on their plane the remaining allowed nonzero value is $a+b$, where $\ell$ is zero, and outside that plane the functional extends. The transvection $T(x)=x+\ell(x)t$ restores the changed star. If the move was on $C$, then \[ T(f+t\ind_C)=f+t(\ell(f)+\ind_C). \] The difference support is even and avoids the star. Its two endpoints are nowhere zero, so the even-set argument realizes it by moves avoiding the vertex. Along the path retain the accumulated invertible linear map. At the last endpoint it fixes the span of the original star values pointwise. In a basis starting with that span, the pointwise stabilizer has block form $\left(\begin{smallmatrix}I&B\\0&C\end{smallmatrix}\right)$. Row additions generate $C$, and shears into the fixed span generate $B$. Each generator has its functional zero on the star span and acts by safe cycles avoiding the star. Undoing the accumulated map completes the pinned path. This factorization works in rank three, and indeed already in rank two. Now cap each shore of the small cut by a new vertex. Project two full flows to the capped sides and join the first pair of side projections. A step avoiding its cap lifts directly. For a step through two cap edges, the functional just constructed on the other side has support of degree two at its cap, on exactly the corresponding edges, and avoids the coefficient. A cycle decomposition supplies a cycle through these edges. Splicing the two cap-deleted paths gives a legal full-graph cycle; a zero-length side path can give a permitted digon. Thus the first side path lifts while the second stays in its ordinary component. At its end the second-side star agrees with the target, so the pinned assertion finishes on that side with the first side fixed. \end{proof} \begin{lemma}[Wide minimal core]\label{lem:widecore} If a bridgeless graph has disconnected $A$-flow graph, let $p$ be the minimum total number of odd-degree vertices among all such graphs. There is a counterexample with the \emph{same} $p$ which is connected, simple, cubic, $3$-edge-connected, has only vertex-star three-cuts, and has girth at least $q$. Every bridgeless graph with fewer than $p$ odd vertices has connected flow graph, with no restriction on its degrees. \end{lemma} This lemma, and every result below that uses it, depends on Theorem~1.2(c) of the unpublished manuscript~\cite{Eulerian}. \begin{proof} The final statement is the original minimization, made before imposing any structural restrictions. Existence for all bridgeless graphs follows by embedding an order-eight flow, including independent loop values. The two-odd-vertex theorem~\cite[Theorem 1.2(c)]{Eulerian} gives $p\ge4$. If all subquartic graphs at this odd count were connected, induction on $\Psi$ using the triangle extension and projection would give connectivity for arbitrary degrees at the same count. Hence a subquartic counterexample at odd count $p$ exists. Choose one with the fewest edges. Independent loops, isolated vertices, component products and suppression of degree-two vertices reduce to a connected loopless graph of degrees three or four. A bridge is impossible in a graph with a nowhere-zero flow. A nonempty balanced family at a degree-four vertex splits to fewer edges, preserving the odd count and the subquartic property. The split graph has a nowhere-zero flow from that family, so it is bridgeless and edge minimality applies. Its smaller flow space is connected, and the balanced lift joins that family. The local family-joining result then excludes degree four. Thus the graph is cubic on exactly $p$ vertices. A parallel pair is an induced digon unless it forms the whole triple-edge dipole; the latter has only two odd vertices and is already excluded. Contracting the digon reduces odd count by two. More generally a shortest cycle of length $\ell|P|-2. \] An adjacency supplies a flow common to $\mathcal E(a,b,p)$ and $\mathcal E(a,b,p')$. Thus all triple families containing $a,b$ lie in one component, denoted $S(a,b)$. Given any two admissible pairs $\{a,b\},\{a',b'\}$, choose $d$ avoiding the closed neighbourhoods of their four edges, then $d'$ avoiding those and $d$. At most twelve and fifteen of seventeen positions are excluded, respectively. The matching triples \[ \{a,b,d\},\quad\{a,d,d'\},\quad \{d,d',a'\},\quad\{d',a',b'\} \] have nonempty families and link $S(a,b)$ to $S(a',b')$. This works also if the pairs share an edge. Hence one component $S_T$ contains every admissible triple family, and therefore every flow with a value occurring at least three times on $C$. We need one more use of the four-cut classification. If $h_1,h_2$ are the two cycle edges at $v$, and $e_1,e_2$ are two other nonincident cycle edges, then \begin{equation}\label{eq:c17star} \{h_1,h_2,e_1,e_2,f\}\text{ is not a cut for any }f\notin E(C). \end{equation} Indeed, take symmetric difference with the star $\{h_1,h_2,p_v\}$. If $f=p_v$ it is a two-cut. Otherwise the four-cut $\{e_1,e_2,f,p_v\}$ bounds adjacent $u,w$. If only one is on $C$, $e_1,e_2$ meet, impossible. If neither is on $C$, no cycle edge leaves. If both are on $C$, their joining edge is on $C$ and their boundary spokes include $p_v$, so one is $v$. Then one of $h_1,h_2$ must leave $\{u,w\}$ and be $e_1$ or $e_2$, again impossible. \begin{lemma}[Noncrossing pairs]\label{lem:c17noncross} Every flow with two noncrossing equal-value pairs on $C$ lies in $S_T$. \end{lemma} \begin{proof} Choose the pairs in cyclic order $a,b,c,d$, with $\phi(a)=\phi(b)$ and $\phi(c)=\phi(d)$. Their shape-$(2,2)$ family is nonempty, so Lemma~\ref{lem:c17fibres} joins the whole family. It suffices to find one member in $S_T$. Within-pair meetings are impossible. The only possible cross-pair meetings are $b,c$ and $d,a$. If there is no meeting, the quadruple is admissible and unblocked: a cut consisting of these four edges and $f$ would have nonzero sum $\phi(f)$. A four-equal flow exists and joins the family to $S_T$. Here is a common tool for the other cases. Take an edge $e$, nonincident with all four anchors, on the open $a$--$b$ arc or the open $d$--$a$ arc. The patterns $(a,e,b;c,d)$ and $(e,a,b;c,d)$ are of shape $(3,2)$. A flow in either is a triple flow in the original two-pair family. By Lemma~\ref{lem:c17bridge}, the only bridge obstructions are \[ \{a,b,c,d,f\},\quad\{b,e,c,d,f\},\quad\{a,e,c,d,f\}. \] The first is excluded by the given flow. If $b,c$ meet, the second is excluded by~\eqref{eq:c17star}; if $d,a$ meet, the third is excluded too. If both meetings occur, the two inner arcs contain thirteen edges in total. Rotate $(a,b,c,d)$ to $(c,d,a,b)$ if necessary so the $a$--$b$ arc has at least seven edges. An edge other than its first or last meets none of the anchors. Both remaining obstructions are excluded, so its test graph is bridgeless and supplies the required triple flow. If exactly one meeting occurs, rotate the names so it is $b,c$. Let the open $a$--$b$, $c$--$d$, $d$--$a$ arcs have lengths $p',q',r'\ge1$, with $p'+q'+r'=13$. Put \[ A=\max(p'-2,0)+\max(r'-2,0),\qquad B=\max(q'-2,0)+\max(r'-2,0). \] At least one is at least four. If $r'\le2$, one of $p',q'$ is at least six. If $r'\ge3$, then $A+B\ge r'+5\ge8$. Reversing the names to $(d,c,b,a)$ exchanges $A,B$ and keeps the same meeting pattern, so assume $A\ge4$. These $A$ choices are exactly the eligible $e$ on the two arcs of the common tool. The meeting at $b,c$ excludes the second bridge obstruction. For the third, $\{a,c,d\}$ is one fixed admissible triple across \emph{both} candidate arcs. Lemma~\ref{lem:c17window} blocks at most three choices. Some choice therefore has a bridgeless test graph and a triple flow in the original family. This finishes all cases. \end{proof} \subsection{Crossing configurations, a switch, and ears} Call a flow of \emph{type $T$} if some value occurs at least three times on $C$, and of \emph{type $N$} if it is not of type $T$ and has two noncrossing pairs. All other flows have \emph{type $X$}: every value occurs at most twice and all repeated pairs cross pairwise. If $k$ values repeat, there are $2k$ \emph{anchors}, $17-2k$ singleton edges and $17-k$ distinct used values. Thus \[ 2\le k\le8,\qquad\text{the number of missing nonzero values is }k-2. \] Write $X_k$ for this type. Call an $X_k$ flow \emph{shut} if no single legal move takes it to type $T$ or $N$. The arguments below analyze this condition only. For an $X_k$ flow, the anchors have cyclic values \begin{equation}\label{eq:c17order} r_1,\ldots,r_k,r_1,\ldots,r_k. \end{equation} Indeed, each side of one repeated pair contains exactly one anchor of each other pair, so the partner of the $j$th anchor is the $(j+k)$th. The open intervals between consecutive anchors, possibly empty, are \emph{anchor arcs}; each borders two repeated labels. \begin{lemma}[Two-pair switch]\label{lem:c17switch} In an $X_k$ flow, let $x,y$ be distinct repeated values and $s=x+y$. If a cycle avoiding $s$ contains a nonempty proper subset of their four anchors, adding $s$ on it gives type $T$ or $N$. \end{lemma} \begin{proof} The four anchors occur with values $x,y,x,y$. The move swaps $x,y$ on its selected anchors. Selecting one or three produces a triple. Selecting two of the same old value makes all four equal. Selecting one of each changes the crossing pairing to one of the two noncrossing pairings. Other changed edges cannot erase these selected equalities. In particular, for a shut flow such a cycle contains all four anchors or none. \end{proof} For a nonzero linear functional $\ell:\Gamma\to\F_2$, the edges whose values have $\ell$-value one form a vertex-disjoint union of cycles: conservation and cubicity give degree zero or two. Denote this support by $S_\ell$. Let $L_1,L_0$ be the repeated values with functional value one and zero. A nonzero kernel value outside $L_0$ occurs on $C$ at most once. A \emph{zero gap} is a maximal nonempty run of $\ell$-zero cycle edges. At each gap endpoint the spoke has $\ell$-value one. \begin{lemma}[Ear construction]\label{lem:c17ears} Let $\phi$ be shut, and suppose $|L_1|\ge2$. Form a graph on $L_1$ joining $w,w'$ when $w+w'\notin L_0$, and suppose it is connected. Then: \begin{enumerate} \item One cycle $D$ of $S_\ell$ contains every $\ell$-one edge of $C$. \item Every maximal path $Q$ of $D$ outside $C$ is an \emph{ear}: its endpoints are distinct vertices of $C$, its interior avoids $C$, and all its values have $\ell$-value one. Of the two endpoint paths on $C$, one, $J'$, contains all $L_1$ anchors; the other, $J$, contains none. \item If a singleton $\ell$-one value $u$ admits $w\in L_1$ such that $u+w$ is absent from $J$ or absent from $J'$, its edge lies on $J'$. In particular it suffices that $u+w\notin L_0$. \item If the preceding condition holds for every ear and every singleton $\ell$-one value, every ear closes a whole zero gap and every zero gap $J$ has its closing ear $Q_J$. The union $Z_J=J\cup Q_J$ is a cycle. Every nonzero $\lambda\in\ker\ell$ absent from $J$ gives a legal move on $Z_J$, changing precisely the edges of $J$ on $C$. \end{enumerate} \end{lemma} \begin{proof} Choose a joined pair $x,y\in L_1$ and the cycle $D$ of $S_\ell$ through an $x$-anchor. It avoids $s=x+y$, so Lemma~\ref{lem:c17switch} puts both $x$- and both $y$-anchors on $D$. For any other $\ell$-one value $u\ne x$ occurring on $C$, the coefficient $u+x$ is a nonzero kernel value. Its legal move on $D$ sends the two $x$-anchors to $u$. An original $u$-edge outside $D$ would create a triple. Hence every such occurrence is on $D$, proving (i). If $C$ has no $\ell$-zero edge, $D=C$ and there are no ears or gaps. Otherwise, at an endpoint of a zero gap, $D$ leaves $C$ along the spoke. An internal ear vertex cannot lie on $C$: it would require two edges outside $C$ at a cubic cycle vertex. The endpoints are distinct because the ear is a proper path of $D$. This remains valid for a chord, which is a one-edge ear. The value $s$ is in the kernel and outside $L_0$, so it occurs on $C$ at most once. One endpoint path $P$ of an ear therefore avoids it, as does the ear itself. The switch on $Q\cup P$ puts all four base anchors on one endpoint path $J'$ and none on the other $J$. Let $W\subseteq L_1$ be the labels with both anchors on $J'$. It contains $x,y$. If $w\in W$ is joined to $u\in L_1$, the coefficient $u+w$ again has a free endpoint path. If that path is $J'$, the two $w$-anchors become $u$ and any $u$-anchor on $J$ would give a triple. If it is $J$, that anchor becomes $w$ while the original two $w$-anchors stay fixed. Thus $u\in W$. Connectedness propagates this to all of $L_1$, proving (ii). The same two tests apply to a singleton $u$ under the free-path condition in (iii): either its edge on $J$ stays $u$ while two $w$-anchors become $u$, or it becomes $w$ while those two anchors remain. Hence it must be on $J'$. Under (iv), $J$ contains no $\ell$-one edge. At both endpoints the continuation on $J'$ is $\ell$-one, so $J$ is a maximal zero gap. Conversely, $D$ leaves $C$ at any endpoint of a gap, starting an ear. That ear closes a gap at the same endpoint, which can only be the original gap. Finally its ear values are outside the kernel and its gap avoids $\lambda$, proving legality and the exact changed portion of $C$. \end{proof} \subsection{Exactly two crossing pairs} \begin{lemma}\label{lem:c17x2} Every $X_2$ flow has a one-move exit to type $T$ or $N$. \end{lemma} \begin{proof} Suppose it is shut. Its four anchors have cyclic values $x,y,x,y$, and all thirteen other nonzero values occur once. Put $s=x+y$, $W=\langle x,y\rangle$ and $g=\eta(s)$, where $\eta(v)$ denotes the unique edge of a singleton value. The four anchor arcs are $A_1,A_2,A_3,A_4$ in order. A pair of other edges crosses both original pairs exactly when its edges lie in opposite arcs: $A_1,A_3$ or $A_2,A_4$. Take any of the four functionals with $\ell(x)=\ell(y)=1$. Here $L_1=\{x,y\}$ and $L_0=\varnothing$, so all hypotheses of Lemma~\ref{lem:c17ears} hold. The seven nonzero kernel values occur once each on the zero gaps. Let $X_J$ be the values of a gap $J$. If $\lambda\in\ker\ell\setminus\{0\}$ is absent from $J$ and $v\in X_J$ has $w=v+\lambda\notin X_J$, the gap move creates the pair $\{\eta(v),\eta(w)\}$. Both original pairs remain unchanged, so the new pair must cross both. Its edges lie in opposite anchor arcs. For two different gaps choose $v,w$ on them and $\lambda=v+w$. Their singleton value sets are disjoint, so $\lambda$ is absent from at least one of the two gaps. Apply the preceding move there. Thus any two gaps lie in opposite arcs, and there cannot be three gaps. If $A_g$ is the arc containing $g$, all zero edges lie in $A_g$ and its opposite $A_g'$, in one interval in each; the interval in $A_g$ contains $g$. For $w\notin W$, let $Z(w)$ be the set of these four functionals vanishing at $w$. It has size two. To see both its size and its multiplicities, write the four functionals as $\ell_0+m$, with $m$ in the two-dimensional annihilator of $W$. Evaluation at $w\notin W$ takes each binary value twice. Equality $Z(w)=Z(w')$ fixes both the evaluation functional on this annihilator and $\ell_0(w)$; hence $w'\in w+W$ and $\ell_0(w')=\ell_0(w)$. This leaves exactly $w'=w$ or $w+s$. Each zero set therefore belongs to just two values. Every one of the twelve outside-$W$ singleton edges lies in $A_g$ or $A_g'$, because it is zero for some functional. Together with $g$ these are all thirteen singleton edges. On either side of $g$ in $A_g$, the sets $Z(w)$ are nested as one moves away from $g$: a zero interval containing a farther edge and $g$ contains all intermediate edges. Since the sets all have size two, they are equal along that side. At most two edges occur on each side, so $|A_g|\le5$. In $A_g'$, group consecutive equal zero sets into blocks. Each block has at most two edges. Each functional occurs in a single interval, so every change of block introduces a functional not used in any previous block: a vanished one cannot reappear. Starting with two of four functionals allows at most two changes, hence three blocks and $|A_g'|\le6$. The other two arcs are empty. The cycle has at most $4+5+6=15$ edges, contradicting seventeen. \end{proof} \subsection{One exceptional repeated value} \begin{lemma}[One exception]\label{lem:c17one} Let a shut $X_k$ flow, $k\ge3$, admit a functional zero on one repeated value $z$ and one on all the others. For $k=3$ assume additionally that $z$ is not the sum of the other two. There are exactly two zero gaps, each containing one $z$-anchor and no other anchor. Every singleton edge whose value is mapped to zero by the functional therefore lies in an anchor arc bordering $z$. \end{lemma} \begin{proof} Here $L_0=\{z\}$. For $k=3$ the safe-sum graph on the two one-labels is connected by hypothesis. For $k\ge4$, any missing adjacency $w+w'=z$ is replaced by a path through a third label $w''$, since neither $w+w''$ nor $w'+w''$ equals $z$. For any singleton one-value $u$, its sums with two distinct one-labels differ, so one is not $z$. Lemma~\ref{lem:c17ears} applies in full. The two $z$-anchors are separated on both sides by one-support anchors, so belong to different gaps $J_a,J_b$. A third gap $J$ would contain only singleton values. Choose $v$ there. If $v+z$ is absent from $J$, its gap move creates a third $z$-edge. Otherwise $v+z$ is a singleton on $J$ and is absent from both $J_a,J_b$. Use it separately on these two gaps. Each test makes a new $v$-pair with the same untouched edge $\eta(v)$ and preserves every one-support pair. Fix one such pair. The two $z$-anchors lie on opposite sides of it, so $\eta(v)$ cannot be opposite both. One test produces noncrossing pairs, a contradiction. Hence only the two anchor gaps exist, proving the last assertion too. \end{proof} \begin{lemma}[Three crossing pairs]\label{lem:c17x3} An $X_3$ flow reaches type $T$ or $N$ in at most two moves, and in one move if its three repeated values are independent. \end{lemma} \begin{proof} Let the repeated values be $x,y,z$. First assume independence and shutness. For each exceptional label $z_0$ among them, there are two functionals taking zero on $z_0$ and one on the others $w_1,w_2$. Both are one at precisely \[ w_1,\quad w_2,\quad w_1+z_0,\quad w_2+z_0. \] Indeed their difference has kernel $\langle x,y,z\rangle$, and these are its four elements where either prescribed functional is one. Lemma~\ref{lem:c17one} applies. A singleton whose value is neither $w_1+z_0$ nor $w_2+z_0$ must lie in an anchor arc bordering $z_0$. A singleton outside $\{x+y,x+z,y+z\}$ would therefore have to border all three labels, impossible. Eleven distinct singleton values cannot lie in a set of size three. If the labels are dependent, their distinctness and nonzeroness give $x+y+z=0$. There is exactly one missing nonzero value $c$. The legal whole-cycle $c$-move preserves every equality class; its new repeated values are distinct, nonzero and sum to $c\ne0$, hence independent. The first case supplies one further move. \end{proof} \begin{lemma}[Four labels with nonzero sum]\label{lem:c17x4a} An $X_4$ flow whose four repeated values have nonzero sum has a one-move exit to type $T$ or $N$. \end{lemma} \begin{proof} Suppose it is shut. Four distinct nonzero binary values have rank four, or rank three with a unique relation of size three or four. The nonzero-sum hypothesis excludes the last case. In rank four, index the basis $r_1,\ldots,r_4$ by its repeated cyclic order. Let $\ell_i(r_i)=0$ and $\ell_i(r_j)=1$ for $j\ne i$. A singleton in the arc bordering $r_i,r_{i+1}$ can have zero under $\ell_j$ only for $j=i,i+1$, by Lemma~\ref{lem:c17one}. Write its value as the sum over a subset $U$ of the basis. Weight one is a repeated value, and weight zero is prohibited. At weight two the zero indices are exactly the two complementary indices, which must be cyclically adjacent; four values qualify. At weight three there are three zero indices, impossible. At weight four one value qualifies. Thus at most five distinct singleton values are possible, against nine. In rank three, write $R=\{y_1,y_2,y_3,t\}$ with $y_1+y_2+y_3=0$ and $t$ outside their plane, and let $H=\langle R\rangle$. For each exceptional $z_0\in\{y_1,y_2,y_3\}$, prescribing zero there and one on the other three is consistent on $H$. Its two extensions to $\Gamma$ differ by the functional with kernel $H$. Every value outside $H$ is zero for one extension for each of the three choices of $z_0$. Its singleton edge would have to border three labels, impossible. But there are eight outside-$H$ values and only two missing values, so at least six such singleton edges occur. This is a contradiction. \end{proof} \begin{lemma}[Five labels with a five-term relation]\label{lem:c17x5a} An $X_5$ flow with rank-four repeated values whose unique relation is the sum of all five has a one-move exit to type $T$ or $N$. \end{lemma} \begin{proof} Suppose it is shut, with labels $r_1,\ldots,r_5$. For each $i$, the functional zero at $r_i$ and one at the four others exists: its assigned values respect the sole relation. The ten sums $r_a+r_b$, $a3+3+3+5$ vertices, so some $X_i\cap Y_j$ contains a cycle; two of them would give $|Y_j|\ge2022>1969$. So there is exactly one such $i=\sigma(j)$, and $\sigma$ is a bijection, since two indices $j$ with one $i$ would give $|X_i|\ge2022$. We renumber so that $\sigma$ is the identity. So for each $i$ the set $X_i\cap Y_i$ is a shore with boundary five, and $|X_i\cap Y_j|\le3$ for all $j\ne i$. \emph{Step 2: the leftovers.} Put $P_i=X_i\setminus Y_i$, $Q_i=Y_i\setminus X_i$, $m_i=|Z'\cap X_i|$ and $m'_i=|Z\cap Y_i|$. The set $P_i$ lies in $(X_i\cap Y_j)\cup(X_i\cap Y_k)\cup(Z'\cap X_i)$, so $|P_i|\le3+3+5=11<18$, and $G[P_i]$ is a forest by the girth. As $X_i$ is the disjoint union of $X_i\cap Y_i$ and $P_i$, we get $5=d(X_i)=d(X_i\cap Y_i)+d(P_i)-2e(P_i,X_i\cap Y_i)=5+d(P_i)-2e(P_i,X_i\cap Y_i)$, so $d(P_i)=2e(P_i,X_i\cap Y_i)$. An edge from $P_i$ into $X_i\cap Y_i\subseteq Y_i$ starts at a vertex of $Z'$: a vertex of $P_i$ outside $Z'$ lies in $Y_j$ or $Y_k$, and no edge joins two different $Y$'s. Every vertex of $Z'$ has exactly one neighbour in $Y_i$ (Lemma~\ref{lem:rig}(a) for $\mathcal D_2$). So $e(P_i,X_i\cap Y_i)\le m_i$; note that $Z'\cap X_i\subseteq P_i$, as $Z'$ misses $Y_i$. So $|P_i|+2\,\mathrm{comp}(P_i)=d(P_i)\le2m_i$, and with $|P_i|=m_i+q_i$, where $q_i$ counts the vertices of $P_i$ outside $Z'$, we get $q_i+2\,\mathrm{comp}(P_i)\le m_i$. If $m_i\ge1$, then $q_i\ge1$: a vertex $y\in Z'\cap X_i$ has at most one neighbour in $Z$ (Lemma~\ref{lem:rig}(d) for $\mathcal D_1$), so at least two neighbours in $X_i$; its three neighbours lie one in each of $Y_1,Y_2,Y_3$ (Lemma~\ref{lem:rig}(a) for $\mathcal D_2$); so a neighbour of $y$ in $X_i$ lies in $Y_j$ or $Y_k$, hence in $P_i$, and it is not in $Z'$, which is independent (Lemma~\ref{lem:rig}(c) for $\mathcal D_2$). So if $P_i$ is not empty, then $\mathrm{comp}(P_i)\ge1$, $m_i=0$ is impossible, and $m_i\ge q_i+2\ge3$. The same holds with the roles of $\mathcal D_1$ and $\mathcal D_2$ exchanged: if $Q_i$ is not empty, then $m'_i\ge3$. (Here $d(Q_i)=2e(Q_i,X_i\cap Y_i)$; an edge from $Q_i$ into $X_i\cap Y_i$ starts in $Z$; every vertex of $Z$ has one neighbour in $X_i$ (Lemma~\ref{lem:rig}(a) for $\mathcal D_1$); a vertex of $Z\cap Y_i$ has at most one neighbour in $Z'$ (Lemma~\ref{lem:rig}(d) for $\mathcal D_2$), so at least two neighbours in $Y_i$, at most one of them in $X_i$ (Lemma~\ref{lem:rig}(a) for $\mathcal D_1$), so one in $X_j$ or $X_k$, hence in $Q_i$, and not in $Z$ (Lemma~\ref{lem:rig}(c) for $\mathcal D_1$).) \emph{Step 3.} Suppose $Z'\ne Z$. Some vertex of $Z'$ is not in $Z$, so it lies in some $X_i$; then $m_i\ge1$, so $P_i$ is not empty and $m_i\ge3$. As $m_1+m_2+m_3\le5$, this holds for one index $i_0$ only; for the other two indices $m_i=0$, so $P_i$ is empty, that is, $X_i\subseteq Y_i$. Some vertex of $Z$ is not in $Z'$, so it lies in some $Y_i$; then $Q_i$ is not empty and $m'_i\ge3$, for one index $i_1$ only; for the other two indices $Q_i$ is empty, that is, $Y_i\subseteq X_i$. Take an index $i$ that is neither $i_0$ nor $i_1$; then $X_i=Y_i$. Every $z\in Z$ has a neighbour in $X_i=Y_i$ and is not in $Y_i$, as it is not in $X_i$; since $Y_i$ is a component of $G-Z'$ (Lemma~\ref{lem:rig}(c) for $\mathcal D_2$), $z$ is in $Z'$. So $Z\subseteq Z'$, and $Z=Z'$, as both have five vertices: a contradiction. \end{proof} \begin{theorem}[The bound 3996]\label{thm:b3996} If each connected component of a finite bridgeless multigraph $H$ has at most $3996$ odd-degree vertices, then $\Fl(H,\F_2^4)$ is nonempty and connected. A least order-sixteen core has at least $3998$ vertices. \end{theorem} The proof uses \cite[Theorem~1.2(c)]{Eulerian}, an unpublished manuscript, through Lemma~\ref{lem:widecore}. \begin{proof} As in the proof of Theorem~\ref{thm:3036}, take the least core $G$ on $p=n$ vertices, suppose $n\le3996$, and take a flow $f$ that is not main, that is, not joined to a flow that omits a value; it exists, since all value-omitting flows lie in one component by Lemma~\ref{lem:clawmain} and the flow graph is not connected. Let $b$ be any value that occurs in $f$. Every component of $G-f^{-1}(b)$ is bridgeless, of even order, with all degrees two or three (the proof of Theorem~\ref{thm:cyclic11} shows this for any value that occurs), so it has an edge-and-claw factor by Theorem~\ref{thm:clawfactor}. For any such factor $J$ of $G-f^{-1}(b)$, the quotient $G/W$, with $W=E(G)\setminus E(J)$, fails Corollary~\ref{cor:roundedpartition}: otherwise $f$ is main, exactly as in the proof of Theorem~\ref{thm:3036}, whose argument uses of $J$ only that $W$ is even and contains $f^{-1}(b)$ and that the claw centres are independent, which holds for every edge-and-claw factor. So Lemmas~\ref{lem:roundedshores} and~\ref{lem:s6} apply to a failing partition. The case $k\ge4$ gives $n\ge3998$ by Lemma~\ref{lem:s6}, which is impossible. So $k\le3$, and the second outcome holds: an exceptional decomposition whose five connectors are claw centres of $J$. Take one such decomposition $\mathcal D=(X_1,X_2,X_3;Z)$, for some value $b_0$ and factor $J_0$. By Lemma~\ref{lem:uniq} it is the only exceptional decomposition of $G$, up to the order of the three shores. Now let $z_1\in Z$, let $x_1$ be its neighbour in $X_1$, and let $b=f(z_1x_1)$; this value is not $0$ and occurs in $f$. Take any factor $J$ of $G-f^{-1}(b)$. The edge $z_1x_1$ has value $b$, so it lies in $W=E(G)\setminus E(J)$, and $z_1$ lies on $W$: it is not a claw centre of $J$. The second outcome for $(b,J)$ gives an exceptional decomposition whose connectors are claw centres of $J$, so its set of connectors is not $Z$, as $z_1\in Z$. This is a second exceptional decomposition with another set of connectors, against Lemma~\ref{lem:uniq}. This contradiction shows that no such $G$ exists with $n\le3996$. So $n\ge3997$, and $n$ is even as $G$ is cubic, so $n\ge3998$. Hence no finite bridgeless multigraph each of whose components has at most $3996$ odd-degree vertices has a disconnected flow graph; the passage to components, loops and nonemptiness is the last paragraph of the proof of Theorem~\ref{thm:3036}. \end{proof} The value $3998$ is what this argument gives; we do not know whether it is best possible. \section{Chains of good sets and short cycles}\label{sec:cqplus} In this section $K$ is a subgroup of order eight of $\F_2^4$, with a basis $x_1,x_2,x_3$; so $K\cong\F_2^3$. A flow is a nowhere-zero flow, and $f\sim g$ means that $f$ and $g$ are joined by moves. For a multigraph $U$ and a set $X$ of at least two vertices, $U/X$ identifies the vertices of $X$ and deletes all edges of $U[X]$. \subsection{Three spanning trees through a linear forest} We say that a graph $G$ has \emph{property \textup{(H)}} if $G$ is cubic, simple and connected, no cut of $G$ has fewer than three edges, every cut of three edges is the set of edges at one vertex, and every cut of four edges is the set of the four edges that leave two adjacent vertices. A least order-sixteen core, in the sense of Lemma~\ref{lem:widecore} for the group $\F_2^4$, has property (H) by Lemma~\ref{lem:widecore} and Corollary~\ref{cor:fourcore}. \begin{lemma}\label{lem:cqT} Let $G$ have property \textup{(H)}, and let $F$ be a set of edges of $G$ with no cycle and with at most two edges at each vertex. Then $G$ has spanning trees $T^1,T^2,T^3$ such that $F\subseteq T^3$ and no edge lies in all three trees. \end{lemma} \begin{proof} \emph{The matroid form.} Put $E_0=E(G)\setminus F$ and $n=|V(G)|$. Take two copies of each edge of $E_0$ and one copy of each edge of $F$; this is the ground set $\Omega$ of a multigraph $G'$. Let $M_1=M_2$ be the cycle matroid of $G'$. In $M_3$ a set $I$ of copies of edges of $E_0$ is independent if $I\cup F$ has no cycle, and the edges of $F$ are loops; its rank is $n-1-|F|$. If $M_1,M_2,M_3$ have pairwise disjoint bases $B_1,B_2,B_3$, then $T^1=B_1$, $T^2=B_2$ and $T^3=F\cup B_3$, with the copies forgotten, are spanning trees. An edge of $E_0$ has two copies, so it is in at most two of the trees; an edge of $F$ is one element, so it is in at most one of $T^1,T^2$. So no edge is in all three. By the matroid union theorem~\cite{MatroidUnion}, such bases exist if and only if, for every $A\subseteq\Omega$, \[ |\Omega\setminus A|+r_1(A)+r_2(A)+r_3(A)\ge3(n-1)-|F|. \] \emph{Reduction to partitions.} Adding the second copy of an edge to $A$ changes no rank and does not raise $|\Omega\setminus A|$. Adding to $A$ an edge whose two ends are in one component of the graph of $A$ changes no rank, also not $r_3$, since $r_3(A)=r_G(A\cup F)-|F|$, and does not raise $|\Omega\setminus A|$. So it suffices to take $A$ to be all copies of all edges inside the parts of a partition of $V(G)$ into $r$ parts, each part connected. Let $m$ be the number of edges of $G$ between different parts and $m_F$ the number of them in $F$. Let $\Phi$ be the multigraph on the parts whose edges are these $m_F$ edges, and $\nu$ its cyclomatic number (edges minus nodes plus components). Then $|\Omega\setminus A|=2(m-m_F)+m_F$, $r_1(A)=r_2(A)=n-r$ and $r_3(A)=\bigl(n-(r-(m_F-\nu))\bigr)-|F|$, and the inequality becomes \begin{equation}\label{eq:cqT} \nu\le2m-3r+3. \end{equation} \emph{Proof of~\eqref{eq:cqT}.} For $r=1$ we have $m=0$, $\nu=0$ and the right side is $0$. Let $r\ge2$. For a part $k$ let $d_k\ge3$ be the number of edges that leave it and $\delta_k$ the number of them in $F$; then $\sum_kd_k=2m$. \emph{Case A: some part $X$ has $d(X)\le4$ and is neither one vertex nor two adjacent vertices.} By (H), the cut of $X$ is the cut of one vertex or of two adjacent vertices; in a connected graph two vertex sets with the same cut are equal or complementary, so $V(G)\setminus X$ is one vertex or two adjacent vertices, and $r\le3$. If $r=2$ and the rest is one vertex, then $m=3$, the right side is $3$, $m_F\le2$ and $\nu\le1$. If $r=2$ and the rest is two vertices in one part, then $m=4$, the right side is $5$, $m_F\le4$ and $\nu\le3$. If $r=3$, the two vertices being two parts, then $m=5$, the right side is $4$, $m_F\le4$ and $\nu\le3$. Here we use that $\nu=0$ if $m_F=0$, and that $\nu\le m_F-1$ if $m_F\ge1$, because $\Phi$ then has an edge. So~\eqref{eq:cqT} holds. \emph{Case B: every part with $d_k\le4$ is one vertex \textup{(}$d_k=3$\textup{)} or two adjacent vertices \textup{(}$d_k=4$\textup{)}.} For a part with $\delta_k\ge1$ put $t_k=d_k-2-\delta_k/2$. For one vertex, $\delta_k\le2$ by the hypothesis on $F$, so $t_k\ge0$, with equality only if $\delta_k=2$. For two adjacent vertices, $\delta_k\le4$, so $t_k\ge0$, with equality only if all four leaving edges are in $F$. For $d_k\ge5$, $t_k\ge d_k/2-2\ge\frac12$. Let $\Phi_i$ be a component of $\Phi$ with an edge, $N_i$ its set of parts and $\nu_i$ its cyclomatic number. Then $\nu_i=1+\sum_{k\in N_i}(\delta_k/2-1)$, so \[ \sum_{k\in N_i}t_k=\sum_{k\in N_i}(d_k-3)-\nu_i+1, \] an integer that is at least $0$. If it is $0$, every part of $N_i$ is one vertex with two leaving edges in $F$ or two adjacent vertices with all four leaving edges in $F$. Then every vertex of these parts is on exactly two edges of $\Phi_i$, and the other ends are again in parts of $N_i$. So the edges of $\Phi_i$ form a graph in which every vertex has degree two; it has a cycle, which is a cycle of $G$ inside $F$. This is impossible. So the sum is at least $1$, that is, $\nu_i\le\sum_{k\in N_i}(d_k-3)$. The sets $N_i$ are disjoint and $d_k-3\ge0$ for all parts, so $\nu\le\sum_k(d_k-3)=2m-3r$, and~\eqref{eq:cqT} holds. \end{proof} \subsection{The flow of three trees} \begin{lemma}\label{lem:cqJ} Let $G$ be connected and let $T^1,T^2,T^3$ be spanning trees with no edge in all three. Let $D_i$ be the even edge set that contains $E(G)\setminus T^i$; it exists and is unique, being the sum of the fundamental cycles of the edges outside $T^i$. Put $h=x_1\ind_{D_1}+x_2\ind_{D_2}+x_3\ind_{D_3}$. Then $h$ is a nowhere-zero flow with values in $K$, and $h(e)\ne x_3$ for every edge $e$ of $T^3$. \end{lemma} \begin{proof} $h$ is a sum of flows. Every edge $e$ is outside some $T^i$, so $e\in D_i$ and $h(e)\ne0$. Let $e\in T^3$. Then $e$ is outside $T^1$ or outside $T^2$, so $e$ is in $D_1$ or in $D_2$; hence $h(e)$ has a coordinate $x_1$ or $x_2$, and $h(e)\ne x_3$. \end{proof} \subsection{A good set, for two flows} \begin{lemma}\label{lem:cqGC} Let $U$ be a finite multigraph without loops, $X$ a set of at least two vertices, and $T$ a spanning tree of $U[X]$ whose edge complement in $U[X]$ is connected and spanning. Let $j$ and $y$ be flows of $U$ with values in $K$, and let $v,v'$ be nonzero elements of $K$ such that $j$ does not take the value $v$ on $T$ and $y$ does not take the value $v'$ on $T$. If the restrictions of $j$ and $y$ to $U/X$ are joined in $\Fl(U/X,K)$, then $j\sim y$ in $\Fl(U,K)$. \end{lemma} \begin{proof} First lift the walk of $U/X$ step by step, as in the proof of Lemma~\ref{lem:goodcontract}. A step on a cycle of $U/X$ that is a cycle of $U$ is done as it is. A step on an exterior path between two different vertices of $X$ is closed through $T$ if its coefficient is $v$, and through a spanning tree $T_2$ of the complement, after the preparation inside $U[X]$ by Lemma~\ref{lem:twotrees}, if its coefficient is another value. That proof lifts every step of any walk and keeps the property ``no value $v$ on $T$''; it does not use where the walk ends. The end is a flow $j''\sim j$ with $j''=y$ on all edges outside $U[X]$ and with no value $v$ on $T$. Now $j''$ and $y$ are nowhere zero on $U[X]$ and have the same boundary there. The graph $U[X]$ has two edge-disjoint spanning trees, $T$ and a tree of the complement. The map $j''$ omits $v$ on $T$ and $y$ omits $v'$ on $T$. By Corollary~\ref{cor:alltrees} the two maps are joined by moves inside $U[X]$, and these are moves of $U$. \end{proof} \subsection{Chains} \emph{The setting.} Let $G$ have property (H), let $f$ be an $\F_2^4$-flow of $G$, $b$ a value of $f$, and $W'$ an even edge set that contains $f^{-1}(b)$. Put $J'=E(G)\setminus W'$. The graph $G/W'$ contracts every cycle of $W'$ to one node and keeps the vertices off $W'$. An edge of $J'$ with both ends on one cycle of $W'$, a \emph{chord}, would be a loop; throughout, $U$ denotes $G/W'$ with these loops deleted. So $E(U)$ is $J'$ without the chords. Let $K$ be a subgroup of order eight that does not contain $b$. Then $f\sim f_1$ for a flow $f_1$ with all values on $W'$ outside $K$ and all values on $J'$ in $K$: the set of edges with a value of $f$ outside $K$ and the set $W'$ are two even sets that both contain $f^{-1}(b)$; their symmetric difference is a set of disjoint cycles with no edge of value $b$; add $b$ on these cycles. Let $j$ be $f_1$ on the edges of $U$; it is a flow of $U$ with values in $K$. \emph{Chains.} A \emph{chain} for $j$ is a sequence $U_0=U,U_1,\ldots,U_k$ in which each $U_i$ arises from $U_{i-1}$ by one of two steps. \begin{enumerate}[label=\textup{(\Alph*)},start=7] \item\label{it:stepG} A \emph{good-set step}: a set $S_i$ of at least two vertices of $U_{i-1}$, a spanning tree $T_i$ of $U_{i-1}[S_i]$ whose edge complement in $U_{i-1}[S_i]$ is connected and spanning, and a nonzero element $v_i$ of $K$ that $j$ does not take on $T_i$; and $U_i=U_{i-1}/S_i$. So $S_i$ is good for $(j,v_i)$ in $U_{i-1}$, in the sense defined before Lemma~\ref{lem:goodcontract}. \end{enumerate} \begin{enumerate}[label=\textup{(\Alph*)},start=3] \item\label{it:stepC} A \emph{short-cycle step}: an induced cycle $C$ of $U_{i-1}$ with at most seven edges; and $U_i=U_{i-1}/V(C)$. A cycle with two edges is induced if its two vertices are joined by exactly two edges. \end{enumerate} No graph $U_i$ has a loop: a good-set step deletes all edges inside the set, and an induced cycle has no chord. \begin{theorem}\label{thm:cqplus} If for some chain the flow graph $\Fl(U_k,K)$ is connected, then $f$ is joined to a flow of $G$ that omits the value $b$. \end{theorem} \begin{proof} Let $F$ be the union of the edge sets of the trees $T_i$ of the good-set steps. These are edges of $J'$, so edges of $G$. \emph{\textup{(a)} $F$ has at most two edges at each vertex of $G$.} A vertex on $W'$ has one edge of $J'$. Let $c$ be a vertex off $W'$; it is a vertex of $U$ of degree three. If $c$ is in no merged set, no edge at $c$ is in $F$. Otherwise let $i$ be the first step whose set contains $c$. A tree of an earlier step has no edge at $c$, since its edges have both ends in its own set. If step $i$ is a good-set step, the complement of $T_i$ in $U_{i-1}[S_i]$ is connected and spanning and $S_i$ has at least two vertices, so it has an edge at $c$; this edge is not in $T_i$, and it is deleted in $U_i$, so it is in no later tree. If step $i$ is a short-cycle step, the two edges of $C$ at $c$ are deleted and are in no tree. In both cases at most two edges at $c$ are in $F$. \emph{\textup{(b)} $F$ has no cycle in $G$.} First, the union of the trees of the steps $i,\ldots,k$ has no cycle in $U_{i-1}$, by induction down from $i=k$. Let $Z$ be a cycle of $U_{i-1}$ inside this union. If step $i$ is a good-set step, $Z$ is not inside the tree $T_i$; its edges outside $T_i$ are edges of $U_i$, because they lie in later trees, and after the contraction of $S_i$ they form a nonempty edge set with even degree at each vertex of $U_i$, which contains a cycle inside the later trees. If step $i$ is a short-cycle step, $Z$ has no edge of $C$, since the edges of $C$ are not edges of $U_i$, and no chord of $C$; so its image in $U_i$ is not empty, has even degrees and no loop, and it contains a cycle of $U_i$ inside the later trees. Both are impossible. So $F$ has no cycle in $U$. A cycle of $G$ inside $F$ has two edges of $J'$ at each of its vertices, so all its vertices are off $W'$, and it is a cycle of $U$. This is impossible. \emph{\textup{(c)} The target.} By (a), (b) and Lemma~\ref{lem:cqT} there are spanning trees $T^1,T^2,T^3$ of $G$ with $F\subseteq T^3$ and no common edge. By Lemma~\ref{lem:cqJ} there is a flow $h$ of $G$ with values in $K$, nowhere zero, with $h(e)\ne x_3$ for every edge $e$ of $F$. Let $y$ be $h$ on the edges of $U$; it is a flow of $U$ with values in $K$, and its restriction is a flow of each $U_i$. \emph{\textup{(d)} Down the chain.} We claim that $j\sim y$ on $U_i$ for $i=k,k-1,\ldots,0$, where both maps are restricted to $U_i$. The same two flows $j$ and $y$ are used at every level; only the relation $\sim$ passes down the chain, not a walk. For $i=k$ this holds because $\Fl(U_k,K)$ is connected. From $i$ to $i-1$ at a good-set step, apply Lemma~\ref{lem:cqGC} in $U_{i-1}$ with $X=S_i$, $T=T_i$, $v=v_i$ and $v'=x_3$; the map $y$ does not take the value $x_3$ on $T_i$ because $T_i\subseteq F$. From $i$ to $i-1$ at a short-cycle step, the proof of Lemma~\ref{lem:shortbinary} lifts every move of the quotient across the cycle, which has fewer than eight edges, and shows that two extensions of one quotient flow are equal or adjacent; so $j\sim y$ on $U_{i-1}/V(C)$ gives $j\sim y$ on $U_{i-1}$, with no condition on $y$. For $i=0$ we get $j\sim y$ in $\Fl(U,K)$. \emph{\textup{(e)} The end.} Lift the moves of $U$ through arcs of the cycles of $W'$; a coefficient in $K$ is safe there, because the values on $W'$ stay outside $K$. Then set every chord to its value under $h$ by one move on the chord and an arc of its cycle; the coefficient is in $K$ and is not the value of the chord. This gives $f_1\sim f_2$ with $f_2=h$ on $J'$ and all values on $W'$ outside $K$. Put $g=h$ on $J'$ and $g=h+b$ on $W'$. Then $g$ is a flow, because $W'$ is even; it is nowhere zero; and it omits $b$, because $h$ is nowhere zero on $W'$ and the values on $J'$ are in $K$. The difference $f_2-g$ is $0$ on $J'$ and has values in $K$ on $W'$, so it is a constant of $K$ on each cycle of $W'$; adding this constant on the cycle is a move. So $f\sim g$. \end{proof} \subsection{The consequence for a least core} \begin{corollary}\label{cor:cqleast} Let $G$ be a least order-sixteen core and $f$ a flow of $G$ that is not joined to a flow omitting a value. Then for every $b$, $W'$, $K$ as above and every chain, the flow graph $\Fl(U_k,K)$ is not connected, and $U_k$ has at least $46$ vertices of odd degree. Moreover, $U_k$ has a cut of at most three edges, and every such cut is the set of the three edges at a vertex off $W'$ that lies in no merged set; so in every chain some vertex off $W'$ is left out. \end{corollary} The proof uses \cite[Theorem~1.2(c)]{Eulerian}, an unpublished manuscript, through Lemma~\ref{lem:widecore}. \begin{proof} The first statement is Theorem~\ref{thm:cqplus}. The graph $U_k$ is a finite bridgeless multigraph, because $G$ has no bridge and neither a contraction nor the deletion of a loop makes one; so Corollary~\ref{cor:44-764} gives more than $44$ vertices of odd degree, and their number is even. By Theorem~\ref{thm:four}, $U_k$ is not $4$-edge-connected; it has at least two vertices, so it has a cut of at most three edges. A cut of $U_k$ is a cut of $G$ all of whose edges are in $J'$. By (H) it is the set of the three edges at one vertex $c$; so $c$ is off $W'$, $c$ is in no merged set, and $c$ is a vertex of degree three of $U_k$. \end{proof} A chain is \emph{complete} if no further step is possible. \begin{lemma}\label{lem:cqcomplete} The last graph $U_k$ of a complete chain is simple, has girth at least eight, and has no good set: no set of at least two vertices of $U_k$ is good for $(j,v)$ for a nonzero element $v$ of $K$. \end{lemma} \begin{proof} A set that is good for $(j,v)$, with $v$ a nonzero element of $K$, would allow a further good-set step. Two vertices joined by two or more parallel edges form such a set, with one of these edges as the tree and any nonzero element of $K$ other than its value; so $U_k$ is simple. A cycle with the fewest edges of a simple graph is induced; so a cycle with at most seven edges would allow a further short-cycle step. \end{proof} \section{The final quotient of a complete chain}\label{sec:finalquotient} Throughout this section $G$ is a least order-sixteen core, as in the proof of Theorem~\ref{thm:3036}: a connected simple cubic graph on $p$ vertices whose $\F_2^4$-flow graph is not connected, $3$-edge-connected with only vertex-star three-cuts (Lemma~\ref{lem:widecore}), of girth at least eighteen (Lemmas~\ref{lem:cycleorder} and~\ref{lem:c17}), and with every cut that has a cycle on both sides of size at least five (Corollary~\ref{cor:fourcore}). We write $n=|V(G)|=p$; $n$ is even. All value-omitting flows of $G$ lie in one component (Lemma~\ref{lem:clawmain}), so some flow $f$ of $G$ is not joined to a flow omitting a value. \subsection{The setting}\label{ssec:fqsetting} Fix a value $b$ of $f$. As in the proof of Theorem~\ref{thm:3036}, the graph $G-f^{-1}(b)$ has an edge-and-claw factor $J$ (Theorem~\ref{thm:clawfactor} and the proof of Theorem~\ref{thm:cyclic11}), and $W=E(G)\setminus E(J)$ is a nonempty even edge set that contains $f^{-1}(b)$. The vertices not on $W$ are the claw centres; no two of them are adjacent, since an edge between two claw centres would lie in one claw, and a claw has one centre. Fix a subgroup $K$ of order eight with $b\notin K$. Let $f_1$, $U$ and $j$ be as in the setting of Theorem~\ref{thm:cqplus} with $W'=W$: we have $f\sim f_1$ by adding $b$ on the cycles of the symmetric difference of $W$ and the set of edges with an $f$-value outside $K$, so the values of $f_1$ outside $K$ are exactly those on $W$; $U=G/W$ with its loops (the chords of the cycles of $W$) deleted; and $j$ is the restriction of $f_1$ to $E(U)$, a nowhere-zero $K$-flow of $U$. Take a complete chain for $j$ (Section~\ref{sec:cqplus}) and let $U^*$ be its last graph, with $n^*$ vertices and $m^*$ edges. We call the vertices of $U^*$ \emph{nodes}. \begin{lemma}\label{lem:fqfacts} \begin{enumerate}[label=\textup{(Q\arabic*)}] \item $U^*$ is simple, has girth at least eight, and has no good set for $j$. \item $\Fl(U^*,K)$ is not connected, and $U^*$ has at least $46$ vertices of odd degree; in particular $n^*\ge46$. \item $6m^*\le14n^*-21$. \end{enumerate} \end{lemma} \begin{proof} (Q1) is Lemma~\ref{lem:cqcomplete}. (Q2) is Corollary~\ref{cor:cqleast} for the complete chain; it applies because $U^*$ is connected (as $G$ is), loopless (the loops are deleted at the start, and every later step deletes the edges inside the merged set), and bridgeless (a cut of $U^*$ is a cut of $G$ with the same edges, and $G$ is $3$-edge-connected). (Q3) follows the argument of Corollary~\ref{cor:roundedpartition}. Fix a nonzero value $v$ of $K$ and let $N$ be the set of edges of $U^*$ whose $j$-value is not $v$. Suppose $|N|\ge2n^*-2$. As $n^*\ge2$ by (Q2), the whole node set has at least two nodes and at least $2n^*-2$ edges of $N$ inside; take a set $S$ of least size with $|S|\ge2$ and at least $2|S|-2$ edges of $N$ inside. For a partition of $S$ into $q\ge2$ parts, a part $P$ with at least two nodes has at most $2|P|-3$ edges of $N$ inside, since it is smaller than $S$, and a part of one node has none, since $U^*$ has no loop. So at least $(2|S|-2)-(2|S|-2q)=2(q-1)$ edges of $N[S]$ cross the partition. By the tree-packing theorem of Nash-Williams and Tutte~\cite{NashWilliamsTrees,TutteTrees}, $N[S]$ has two edge-disjoint spanning trees. One of them has no value $v$; the other lies in its complement in $U^*[S]$, which is therefore connected and spanning. So $S$ is a good set for $(j,v)$, against (Q1). Hence $|N|\le2n^*-3$ for each of the seven values $v$. Every edge lies in $N$ for six of the seven values (all but its own value), so $6m^*\le7(2n^*-3)=14n^*-21$. \end{proof} \emph{Nodes.} A \emph{single centre} is a node of $U^*$ that is one claw centre; a \emph{lump} is any other node. Every lump contains a cycle of $W$: each step of the chain merges a connected set (a good set has a spanning tree, and a cycle is connected), and claw centres are pairwise not adjacent in $U$, so every merged set contains a node that contains a cycle of $W$. A single centre has degree three in $U^*$ and three different neighbours, by (Q1), and all of them are lumps. We write $s$ for the number of single centres, $t$ for the number of lumps, and $\varepsilon$ for the number of edges between two lumps. For a lump $L$, we also write $L$ for the set of vertices of $G$ that it contains; then $G[L]$ is connected, and the degree $d(L)$ of $L$ in $U^*$ is the number of edges of $G$ with exactly one end in $L$. \subsection{Small shores of boundary at most fourteen} \begin{lemma}\label{lem:S14} Let $G$ be a finite simple cubic graph of girth at least eighteen, and let $X$ be a set of vertices such that $G[X]$ is connected and contains a cycle. If $d(X)\le14$, then $|X|\ge988$. \end{lemma} \begin{proof} Prune vertices of degree one as in the proof of Lemma~\ref{lem:smallshore}: this gives a connected core $F=G[X']$ of minimum degree two, with $D=d(X')\le d(X)\le14$ vertices of degree two and $N=|X'|$ vertices. $F$ is not a cycle, since a cycle would have $D=N\ge18$. \emph{First bound: $N\ge44$.} (The first bound in the proof of Lemma~\ref{lem:smallshore} gives only $N\ge32$ when $D\le14$, which lies below the smaller root $38.7$ of the quadratic below; so a finer count is needed.) Root the distance levels of $F$ at a vertex of degree three. Girth at least eighteen gives unique parents through level eight and no edge inside levels one to seven. Write $n_i$ for the size of level $i$ and $c_i$ for the number of vertices of degree three in level $i$; then $n_1=3$ and $n_{i+1}=n_i+c_i$. At most fourteen vertices of levels one to seven have degree two, so \[ 21+5c_1+4c_2+3c_3+2c_4+c_5-c_7\le14,\quad\text{that is}\quad c_7\ge7+5c_1+4c_2+3c_3+2c_4+c_5 . \] With $c_7\le n_7=3+c_1+\cdots+c_6$ this gives $c_6\ge4+4c_1+3c_2+2c_3+c_4$, and with $c_6\le n_6$ it gives $c_5\ge1+3c_1+2c_2+c_3$. So $c_5\ge1$, $c_6\ge4$, $c_7\ge8$, and the root with levels one to eight has $25+(7c_1+6c_2+\cdots+1c_7)\ge25+3+8+8=44$ vertices. \emph{The averaging.} The averaging in the proof of Lemma~\ref{lem:smallshore} (equation~\eqref{eq:shoredeficit} and the bound $3586$ for each vertex of degree two) is used here for $D\le14$. That proof uses $D\le10$ in three places only, and none of them blocks $D\le14$: ``$F$ is not a cycle'' needs $D<18$; its first bound $N\ge36$ is replaced by $N\ge44$ above; and its last step puts the bound on $D$ into $D/(3N-D)$, which grows with $D$. It gives a root edge with loss at most \[ \frac{7172D}{3N-D}\le\frac{100408}{3N-14}, \] so $N\ge1022-100408/(3N-14)$, that is, \[ 3N^2-3080N+114716\ge0. \] This convex quadratic has the values $2008$, $-841$, $-2737$ and $108$ at $N=38$, $39$, $987$ and $988$, so it is negative exactly for $39\le N\le987$. With $N\ge44$ we get $N\ge988$, and $|X|\ge N$. \end{proof} \subsection{Linear and triangle-free claw triples} \begin{lemma}\label{lem:fqlin} \begin{enumerate} \item Two single centres have at most one common neighbour. Hence $s\le t(t-1)/6$. \item Let $L$ be a lump with $x$ single-centre neighbours $z_1,\ldots,z_x$, and let $N_i$ be the set of the two neighbours of $z_i$ other than $L$. A single centre that is not adjacent to $L$ has at most one neighbour in $N=N_1\cup\cdots\cup N_x$. Hence, for the lump $L$ with the most single-centre neighbours, \begin{equation}\label{eq:claimA} s\le x+\binom{t-1-2x}{2},\qquad \frac{3s}{t}\le x\le\frac{t-1}{2}, \end{equation} where $\binom m2=m(m-1)/2$, and $\binom m2=0$ for $m\le1$. \end{enumerate} \end{lemma} \begin{proof} (i) Two common neighbours would give a cycle of four edges, against girth eight. So the $s$ neighbour triples of the single centres cover $3s$ different pairs of lumps, and $3s\le\binom t2$. (ii) By (i) the sets $N_i$ are pairwise disjoint, so $N$ has $2x$ lumps, and $L\notin N$. Let $A,B$ be two neighbours in $N$ of a single centre $z$ not adjacent to $L$. If both lie in one $N_i$, then $z$ and $z_i$ share $A$ and $B$, against (i). If $A\in N_i$ and $B\in N_j$ with $i\ne j$, then $z_i\,L\,z_j\,B\,z\,A\,z_i$ is a cycle of six edges, against girth eight. Its six nodes are different: the centres $z_i,z_j$ are different, and $z$ is neither of them, as $z$ is not adjacent to $L$; the lumps $A,B$ are different, as $N_i$ and $N_j$ are disjoint, and $L$ is neither of them, as $L\notin N$; and a centre is never a lump. So each of the $s-x$ single centres not adjacent to $L$ has at least two neighbours in the set $R$ of the lumps other than $L$ and those in $N$, and $|R|=t-1-2x$. Choose one such pair in $R$ for each of them; by (i) the pairs are different. This gives the first inequality of~\eqref{eq:claimA}. For the lump with the most single-centre neighbours, $x\ge3s/t$ (the average), and $2x\le t-1$, since $L$ and $N$ are $2x+1$ different lumps. \end{proof} For $t\ge2$ let $S(t)$ be the largest whole $s\le t(t-1)/6$ for which some whole $x$ with $3s/t\le x\le(t-1)/2$ satisfies~\eqref{eq:claimA}. For $t=15,\ldots,30$, \[ S(t)=19,\ 21,\ 22,\ 26,\ 31,\ 33,\ 35,\ 42,\ 46,\ 48,\ 52,\ 60,\ 63,\ 65,\ 74,\ 80 . \] For example, at $t=19$ the value $s=32$ forces $x\ge6$, and $6+\binom62=21$, $7+\binom42=13$, $8+\binom22=9$ and $9+0=9$ are all below $32$; the value $s=31$ has $x=5$ with $5+\binom82=33$. \subsection{A ball around the best edge} We use the setting and the notation of Section~\ref{ssec:fqsetting}, Lemma~\ref{lem:fqfacts} and Lemma~\ref{lem:fqlin}, and the count of the lumps: with $\Sigma=3s+2\varepsilon$ the sum of the degrees of the lumps and $a$ the number of lumps $L$ with $d(L)\le14$, \begin{equation}\label{eq:fqcountB} 2s+3\varepsilon\le7t-11,\qquad \Sigma\ge5a+15(t-a),\qquad n\ge988a+18(t-a)+s, \end{equation} by (Q3) (as $m^*=3s+\varepsilon$ and $n^*=s+t$, and both sides are whole numbers), by $d(L)\ge5$ for every lump (since $t\ge2$, a lump and its complement both contain a cycle), and by Lemma~\ref{lem:S14} for the lumps with $d(L)\le14$ and girth eighteen for the others, as the lumps are pairwise disjoint vertex sets of $G$ and each single centre is one more vertex. Here $t\ge2$: $t\ge1$ as $W$ is not empty, and $t=1$ would force $s=0$, as $U^*$ is simple, and $n^*=1$, against (Q2). Moreover $s\ge1$: from~\eqref{eq:fqcountB}, $5t\le\Sigma\le(5s+14t-22)/3$, that is, $5s\ge t+22$. \begin{lemma}\label{lem:ball} Let $L$ be a lump with the most single-centre neighbours, $x$ of them, and $d=d(L)$. Then $s+t\ge28+11(d-1)$, where $d\ge5$ and $d\ge x\ge3s/t$. Hence \begin{equation}\label{eq:ball} s+t\ge\max\Bigl(72,\ 17+\frac{33s}{t}\Bigr). \end{equation} \end{lemma} \begin{proof} The $3s$ edges at the single centres end at lumps, so the average number of single-centre neighbours of a lump is $3s/t$, and $x\ge3s/t>0$ as $s\ge1$. Let $c$ be a single centre adjacent to $L$, and for a node $v$ let $r(v)$ be its distance to the set $\{c,L\}$. If a node $v$ with $r(v)\le3$ had two different shortest paths from $\{c,L\}$, these paths, with the edge $cL$ if they start at different ends, would contain a cycle of at most $2r(v)+1\le7$ edges; and if a node $v$ with $r(v)\le2$ had a neighbour $w$ with $r(w)\le r(v)$ other than its parent (the parent of $c$ is $L$, and the parent of $L$ is $c$), the shortest paths to $v$ and to $w$ with the edge $vw$, and with $cL$ if needed, would contain a cycle of at most $2+2+1+1=6$ edges. Both are excluded by girth eight (Q1). So the nodes with $r\le3$ are reached along a tree that contains the edge $cL$: every node $v$ with $r(v)\le2$ has all its neighbours other than its parent at distance $r(v)+1$, and two different nodes with $r\le2$ have no such neighbour in common. On the side of $c$: $c$ has two other neighbours, which are lumps; each has at least four further neighbours, as its degree is at least five, which gives at least eight nodes at distance two; each of these has at least two further neighbours, which gives at least sixteen nodes at distance three. Together at least $26$. On the side of $L$: $L$ has $d-1$ other neighbours. If such a neighbour is a single centre, it has two further neighbours, which are lumps with at least four further neighbours each: $1+2+8=11$ nodes. If it is a lump, it has at least four further neighbours with at least two further neighbours each: $1+4+8=13$ nodes. So each of the $d-1$ branches has at least eleven nodes, and with $c$ and $L$ we get $s+t=n^*\ge2+26+11(d-1)$. Finally $d\ge5$ gives $72$, and $d\ge x\ge3s/t$ gives $28+11(3s/t-1)=17+33s/t$. \end{proof} For $t\le32$, \eqref{eq:ball} says $s(33-t)\le t(t-17)$; for $t\ge33$ its second part holds for every $s$. \begin{lemma}\label{lem:claimAclosed} Let $t\ge3$ and $s>(t-1)/2$. If~\eqref{eq:claimA} holds for some whole $x$ with $3s/t\le x\le(t-1)/2$, then \[ Q_t(s)=36s^2-2t(7t-12)s+t^2(t-1)(t-2)\ge0\qquad\text{and}\qquad s\le\frac{t(t-3)}6\le\frac{t(7t-12)}{36}, \] the last number being the vertex of $Q_t$. \end{lemma} \begin{proof} Put $m=t-1-2x$. If $m\le1$, then $\binom m2=0$ and $s\le x\le(t-1)/2$, which is excluded; so $m\ge2$. The function $g(y)=y+(t-1-2y)(t-2-2y)/2$ has derivative $2-2(t-1-2y)$, which is negative while $t-1-2y>1$, so on the whole interval $[3s/t,x]$, where $t-1-2y\ge m\ge2$. Hence $s\le x+\binom m2=g(x)\le g(3s/t)$. With $u=6s/t$, so that $3s/t=u/2$ and $s=ut/6$, this reads $ut/6\le u/2+(t-1-u)(t-2-u)/2$, that is, $u^2-\frac{7t-12}3u+(t-1)(t-2)\ge0$; multiplied by $t^2$ this is $Q_t(s)\ge0$. And $u\le2x\le t-3$, as $m\ge2$, so $s\le t(t-3)/6$, which is at most $t(7t-12)/36$ since $6t-18\le7t-12$. \end{proof} As $Q_t$ is convex and $s$ lies at or below its vertex, $s$ lies at or below the smaller root of $Q_t$. For $t=31,\ldots,38$ the largest such whole $s$ is \[ S'(t)=87,\ 93,\ 99,\ 105,\ 111,\ 118,\ 125,\ 132, \] as $Q_t(S'(t))\ge0>Q_t(S'(t)+1)$, both points below the vertex: the pairs of values are $(2784,-3626)$, $(1860,-4976)$, $(2178,-5112)$, $(3996,-3776)$, $(7596,-686)$, $(4464,-4284)$, $(2690,-6552)$, $(2544,-7220)$. \subsection{The count and the bound} \begin{theorem}\label{thm:order16bound} If each connected component of a finite bridgeless multigraph $H$ has at most $19022$ odd-degree vertices, then $\Fl(H,\F_2^4)$ is nonempty and connected. A least order-sixteen core has at least $19024$ vertices. \end{theorem} The proof uses \cite[Theorem~1.2(c)]{Eulerian}, an unpublished manuscript, through Lemma~\ref{lem:widecore}. \begin{proof} We show $n\ge19023$; as $n$ is even, then $n\ge19024$. \emph{\textup{(1)} No case with $t\le26$.} By~\eqref{eq:ball}, $s\ge72-t\ge46$ and $s(33-t)\le t(t-17)$. For $t\le17$ the right side is at most $0$, against $s\ge1$ (and against $s\ge46$). For $18\le t\le26$, $t(t-17)/(33-t)$ grows with $t$ and is $234/7<34$ at $t=26$, against $s\ge46$. \emph{\textup{(2)} The rule for one row.} From~\eqref{eq:fqcountB}, $\Sigma\le(5s+14t-22)/3$ and $\Sigma\ge15t-10a$, so $a\ge(31t-5s+22)/30$, and $n\ge970a+18t+s$. Let $\mathrm{lo}=72-t$, by~\eqref{eq:ball}, and let $\mathrm{hi}$ be the largest allowed $s$: $S(t)$ for $t=29,30$; $S'(t)$ for $t=31,\ldots,36$ (Lemma~\ref{lem:claimAclosed} applies, as $s\ge\mathrm{lo}\ge36>(t-1)/2$); \eqref{eq:ball} for $t=27,28$; and $s\le(7t-11)/2$ for $t=37,38$. Let $a_0$ be the least whole number at least $(31t-5\,\mathrm{hi}+22)/30$; then $a\ge a_0$. If $a\ge a_0+1$, then $n\ge970(a_0+1)+18t+\mathrm{lo}$. If $a=a_0$, then $\Sigma\ge15t-10a_0$; but $\Sigma\le M(s)=3s+2\lfloor(7t-11-2s)/3\rfloor$, and $M$ grows with $s$ (by one or three at each step). Let $s_0$ be the least $s\ge\mathrm{lo}$ with $M(s)\ge15t-10a_0$; then $s\ge s_0$ and $n\ge970a_0+18t+s_0$. \emph{\textup{(3)} The table.} \[ \begin{array}{c|c|c|c|c|c} t & \mathrm{lo}..\mathrm{hi} & a_0 & 15t-10a_0 & s_0\ (M(s_0),\,M(s_0-1)) & \text{least }n\\\hline 27 & 45..45 & 22 & 185 & 45\ (193;\ s_0=\mathrm{lo}) & 21871\\ 28 & 44..61 & 20 & 220 & 58\ (220,\,217) & 19962\\ 29 & 43..74 & 19 & 245 & 71\ (245,\,244) & 19023\\ 30 & 42..80 & 19 & 260 & 77\ (261,\,258) & 19047\\ 31 & 41..87 & 19 & 275 & 83\ (275,\,274) & 19071\\ 32 & 40..93 & 19 & 290 & 90\ (292,\,289) & 19096\\ 33 & 39..99 & 19 & 305 & 95\ (305,\,302) & 19119\\ 34 & 38..105 & 19 & 320 & 102\ (320,\,319) & 19144\\ 35 & 37..111 & 19 & 335 & 108\ (336,\,333) & 19168\\ 36 & 36..118 & 19 & 350 & 114\ (350,\,349) & 19192\\ 37 & 35..124 & 19 & 365 & 121\ (367,\,364) & 19217\\ 38 & 34..127 & 19 & 380 & 126\ (380,\,377) & 19240 \end{array} \] In each row the last entry is the smaller of $970a_0+18t+s_0$ and $970(a_0+1)+18t+\mathrm{lo}$; the second is always the larger one (it is $22841$, $20918$, $19965$, $19982$, $19999$, $20016$, $20033$, $20050$, $20067$, $20084$, $20101$, $20118$). \emph{\textup{(4)} $t\ge39$.} $s\le(7t-11)/2$ gives $\Sigma\le10.5t-16.5$, so $a\ge(4.5t+16.5)/10>19$, $a\ge20$, and $n\ge970a+18t\ge97(4.5t+16.5)+18t=454.5t+1600.5\ge19326$. So $n\ge19023$ in every case, the least row being $t=29$, $s=71$, $a=19$. The passage to components, loops and nonemptiness is the last paragraph of the proof of Theorem~\ref{thm:3036}. \end{proof} The proof does not show that $19022$ is best possible. \section{A boundary theorem with at most seven terminals}\label{sec:gen} This section concerns existence only; it uses no moves. Throughout it $K=\F_2^3$. Its two external inputs are Jaeger's theorem~\cite{Jaeger}, that every bridgeless multigraph has a nowhere-zero $K$-flow, and Edmonds' matroid intersection theorem~\cite{EdmondsMI}: two matroids $M_A$, $M_B$ on one ground set $E$ have a common independent set with $t$ elements if and only if $r_A(S)+r_B(E\setminus S)\ge t$ for every $S\subseteq E$. \subsection{Instances and the statement} \begin{definition}\label{def:instance} An \emph{instance} is a pair $(H,\beta)$, where $H$ is a finite connected bridgeless multigraph with at least one vertex (loops and parallel edges are allowed) and $\beta:V(H)\to K$ satisfies $\sum_{v}\beta(v)=0$. The vertices with $\beta(v)\ne0$ are the \emph{terminals}. A \emph{solution} is a map $f:E(H)\to K\setminus\{0\}$ such that, for every vertex $v$, the sum of $f$ over the edges at $v$ equals $\beta(v)$; a loop at $v$ is counted twice in this sum, so it contributes $0$. \end{definition} Every edge is counted twice in the sum of all the vertex sums; this is why the total of $\beta$ must be $0$. For a vertex set $X$ we write $d_H(X)$ for the number of edges with exactly one end in $X$, and $H/X$ for the graph in which $X$ becomes one vertex and the edges inside $X$ are deleted. \begin{theorem}[Seven terminals]\label{thm:gen} Every instance with at most seven terminals has a solution. \end{theorem} The number seven cannot be raised. Let $H$ be a cycle $v_1\cdots v_8$ whose eight vertices are terminals and whose partial sums $s_i=\beta(v_1)+\cdots+\beta(v_i)$ are the eight elements of $K$. A map with boundary $\beta$ has the form $p+s_i$ on the edge $v_iv_{i+1}$, and it takes the value $0$ on some edge. The hypothesis that $H$ is bridgeless cannot be dropped either: the value on a bridge is the sum of $\beta$ over one of its sides, which can be $0$. For a cubic graph the theorem takes the following form, in which the values on the edges leaving a vertex set are prescribed. \begin{corollary}\label{cor:genpole} Let $G$ be a cubic graph and $L\subseteq V(G)$ a set such that $G[L]$ is connected and bridgeless, with at most seven edges leaving $L$, all at different vertices of $L$. Then every assignment of nonzero values of $K$ with sum $0$ to the leaving edges extends to a map $E(G[L])\to K\setminus\{0\}$ whose sum at every vertex of $L$, the leaving edges included, is $0$. \end{corollary} \begin{proof} Put $H=G[L]$ and let $\beta(v)$ be the prescribed value of the leaving edge at $v$, and $\beta(v)=0$ if no edge leaves $L$ at $v$. Then $(H,\beta)$ is an instance with at most seven terminals, and a solution is the required map. \end{proof} \subsection{Cutting off a set} \begin{lemma}\label{lem:genR} Let $(H,\beta)$ be an instance with at most seven terminals, and suppose Theorem~\ref{thm:gen} holds for all instances with fewer edges than $H$. Let $X\subsetneq V(H)$ have at least two vertices, with $H[X]$ connected and bridgeless, and suppose that the number of terminals in $X$ plus $d_H(X)$ is at most seven. Then $(H,\beta)$ has a solution. \end{lemma} \begin{proof} Let $H_1=H/X$, with $x$ the vertex that replaces $X$, and put $\beta_1(x)=\sum_{v\in X}\beta(v)$ and $\beta_1=\beta$ elsewhere. The graph $H_1$ is connected, is bridgeless because a cut of $H_1$ is a cut of $H$, and has fewer edges because $H[X]$ has an edge. It has no more terminals than $H$, since $x$ is a terminal only if $X$ contains one. Take a solution $f_1$. On $H[X]$ let $\beta_2(v)$ be $\beta(v)$ plus the sum of $f_1$ over the edges that leave $X$ at $v$. Its total is $\beta(X)+\beta_1(x)=0$, and it has at most (terminals in $X$)${}+d_H(X)\le7$ terminals. The graph $H[X]$ is connected and bridgeless, and it has fewer edges than $H$ because an edge leaves $X$. Take a solution $f_2$ of $(H[X],\beta_2)$. Then $f_1$ outside $H[X]$ and $f_2$ inside form a solution of $(H,\beta)$. \end{proof} \begin{lemma}\label{lem:genRprime} Let $(H,\beta)$ be an instance with at most seven terminals, and suppose Theorem~\ref{thm:gen} holds for all instances with fewer edges than $H$. Let $X$ have at least two vertices, with $H[X]$ connected, $d_H(X)=2$, and at most one terminal in $X$. Then $(H,\beta)$ has a solution. Here $H[X]$ may have bridges. \end{lemma} \begin{proof} First solve $H/X$, which has fewer edges and no more terminals. The vertex $x$ has two edges, with values $p$ and $p+s$, where $s=\beta(X)$; both are nonzero. Let $u_a,u_b$ be the ends in $X$ of the two leaving edges and $y$ the terminal of $X$, if there is one; some of these vertices may coincide. The new boundary on $H[X]$ is nonzero at most at these three vertices. Its total is $0$, so it is nonzero at exactly zero, two or three vertices: at two vertices the two values are equal, and at three vertices they are $c,d,c+d$ with $c,d$ different and nonzero. If $u_a\ne u_b$, the new boundary is nonzero at both. Indeed, at $u_a$ it is the value of $f_1$ on the leaving edge at $u_a$ if $u_a$ is not the terminal; if $u_a$ is the terminal, then $s=\beta(y)$ and the new boundary at $u_a$ is this value plus $\beta(y)$, which is the value of $f_1$ on the other leaving edge. The same holds at $u_b$. So the case of zero vertices occurs only with $u_a=u_b$, and in the other two cases with $u_a\ne u_b$ both $u_a$ and $u_b$ are among the vertices with nonzero new boundary. Close $H[X]$: with nothing in the case of zero vertices, with one new edge between the two vertices, or with one new vertex joined to the three vertices. The closed graph is bridgeless. A bridge of $H[X]$ separates $u_a$ from $u_b$, because it is not a bridge of $H$, and the new edge or the new vertex joins $u_a$ to $u_b$; if $u_a=u_b$, then $H[X]$ has no bridge at all; and a new edge is not a bridge because $H[X]$ is connected. By Jaeger's theorem the closed graph has a nowhere-zero $K$-flow. A linear automorphism of $K$ carries it to a flow with the wanted values on the new edges: for one new edge, any nonzero value goes to any nonzero value; for a new vertex of degree three, any two different nonzero values go to any two different nonzero values, and the third value is their sum. Its restriction to $H[X]$ is a solution for the new boundary, and together with $f_1$ it solves $(H,\beta)$. \end{proof} \subsection{A least counterexample} Let $(H,\beta)$ be a counterexample to Theorem~\ref{thm:gen} with the fewest edges, and let $k\le7$ be the number of its terminals. \begin{lemma}\label{lem:genleast} The following hold. \begin{enumerate}[label=\textup{(M\arabic*)},start=0] \item\label{it:M0} $H$ has a vertex of degree at least three. \item\label{it:M1} $H$ has no loop, no vertex of degree two with $\beta=0$, and no vertex of degree two both of whose edges go to one vertex. \item\label{it:M2} If $X\subsetneq V(H)$ has at least two vertices and $H[X]$ is connected and bridgeless, then the number of terminals in $X$ plus $d_H(X)$ is at least eight. \end{enumerate} Let $G'$ be the graph obtained from $H$ by suppressing all vertices of degree two; its vertices are the vertices of $H$ of degree at least three. Then: \begin{enumerate}[label=\textup{(M\arabic*)},start=3] \item\label{it:M3} $G'$ has no loop and at least two vertices, and every cut of $G'$ has at least three edges. \item\label{it:M4} Every edge of $G'$ carries at most one vertex of degree two of $H$, and such a vertex is a terminal. \end{enumerate} \end{lemma} \begin{proof} \ref{it:M0}. Otherwise $H$ is connected and bridgeless with all degrees at most two. If it is one vertex without an edge, then $\beta=0$ and the empty map is a solution. Otherwise it is a cycle $v_1\cdots v_n$ (a loop is a cycle with one edge). Put $f=p+s_i$ on the edge $v_iv_{i+1}$, where $s_i=\beta(v_1)+\cdots+\beta(v_i)$. The $s_i$ take at most $\max(1,k)\le7$ values, so some $p$ is none of them, and $f$ is a solution. \ref{it:M1}. Suppose a vertex of degree two has $\beta=0$. Its two edges are different: a vertex whose only edge is a loop would be all of $H$, since $H$ is connected, against~\ref{it:M0}. Suppress the vertex, so that its two edges become one edge. The new graph is connected and bridgeless, has one edge fewer and the same terminals; a solution of it gives both old edges the value of the new edge. Suppose next that $H$ has a loop. Delete it; the graph stays connected and bridgeless. Solve it and give the loop any nonzero value; the loop adds $0$ at its vertex. Finally suppose a vertex $y$ of degree two has both its edges to one vertex $v$. Then $H-y$, with $\beta(v)+\beta(y)$ at $v$, is an instance with fewer edges and no more terminals: it is connected, and a bridge of $H-y$ would be a bridge of $H$. Take a solution, and give the two edges at $y$ the values $p$ and $p+\beta(y)$ with $p\notin\{0,\beta(y)\}$. \ref{it:M2} is Lemma~\ref{lem:genR}. \ref{it:M3}, first part. By~\ref{it:M0} the graph $G'$ is defined. A loop of $G'$ is a cycle of $H$ with exactly one vertex $v$ of degree at least three and $r$ vertices of degree two. The case $r=0$ is a loop of $H$ and the case $r=1$ is a vertex of degree two with both edges to $v$; both are excluded by~\ref{it:M1}. For $r\ge2$ let $X$ be the set of all other vertices. Then $H[X]$ is connected, a bridge of it would be a bridge of $H$, and $X$ has at least two vertices, because $v$ has a third edge, which by~\ref{it:M1} is not a loop and does not end on the cycle. By~\ref{it:M1} the $r$ vertices of degree two are terminals, so the number of terminals in $X$ plus $d_H(X)$ is at most $(k-r)+2\le7$, against~\ref{it:M2}. So $G'$ has no loop. A vertex of degree at least three without a loop has a neighbour in $G'$, so $G'$ has at least two vertices. \ref{it:M3}, the cuts. $G'$ has no cut of zero or one edges, because $H$ is connected and bridgeless. Suppose $G'$ has a cut of two edges and take an inclusion-least side $X'$ of such a cut. Then $G'[X']$ is connected, since a component of it is a side of a cut of at most two edges and so is all of $X'$. It is bridgeless, since a bridge would give a smaller side of a cut of two edges. And $X'$ has at least two vertices, since one vertex of $G'$ has degree at least three and no loop. Let $X$ be $X'$ together with the vertices of degree two on the edges of $G'$ inside $X'$. Then $H[X]$ is connected and bridgeless, $X$ has at least two vertices and $d_H(X)=2$; by~\ref{it:M2}, $X$ contains at least six terminals. Let $Z=V(H)\setminus X$; it contains the vertices of degree two on the two cut edges, and it contains at most one terminal. The graph $H[Z]$ is connected: the other side of the cut is connected in $G'$, because a component of a side has at least two leaving edges ($G'$ has no cut of zero or one edges) and only two edges leave the side; and the vertices on the cut edges hang on it as end paths. The set $Z$ has at least two vertices, since the other side is not a single vertex of $G'$ (such a vertex has degree at least three and no loop). And $d_H(Z)=2$. Lemma~\ref{lem:genRprime} for $Z$, which does not ask that $H[Z]$ be bridgeless, gives a solution: a contradiction. \ref{it:M4}. A vertex of degree two is a terminal by~\ref{it:M1}. Let $r\ge2$ be the number of vertices of degree two on an edge $e$ of $G'$; $e$ is not a loop. Let $X$ be the set of all other vertices of $H$. Then $H[X]$ is connected, because $G'-e$ is connected by~\ref{it:M3}. It is bridgeless, because a bridge $g$ of $G'-e$ would give the cut $\{e,g\}$ of $G'$, against~\ref{it:M3}. The set $X$ has at least two vertices, the ends of $e$. And the number of terminals in $X$ plus $d_H(X)$ is $(k-r)+2\le7$, against~\ref{it:M2}. \end{proof} Thus the vertices of $H$ are the vertices of $G'$ together with at most one vertex of degree two on each edge of $G'$, each of them a terminal. At most seven edges of $G'$ carry a terminal, and the other terminals are vertices of $G'$. \subsection{Three trees and a flag} In this subsection and the next, $G'$ is any multigraph without loops, with at least two vertices, in which every cut has at least three edges. Different \emph{marked} edges $e_1,\ldots,e_j$ carry values $z_1,\ldots,z_j\in K\setminus\{0\}$, and $b:V(G')\to K$ has total $0$. We want a map $a:E(G')\to K$ with boundary $b$, that is, with sum $b(v)$ over the edges at each vertex $v$, such that $a(e)\ne0$ for every edge and $a(e_i)\ne z_i$ for every marked edge. An instance as at the end of the last subsection gives this setting. The marked edges are the edges of $G'$ that carry a terminal $y_i$ of degree two, and $z_i=\beta(y_i)$. Choose one end $v_i$ of $e_i$ and let $b(v)$ be $\beta(v)$ plus the sum of the $z_i$ with $v_i=v$; its total is the total of $\beta$, which is $0$. A wanted map $a$ gives a solution $f$: put $f=a(e)$ on an edge without a terminal; on $e_i$ put $a(e_i)$ on the half away from $v_i$ and $a(e_i)+z_i$ on the half at $v_i$. A \emph{flag} is a line $\ell=\{0,x\}$ inside a plane $P$ of $K$. The \emph{colour} of a nonzero value $z$ is $1$ if $z\in\ell$, is $2$ if $z\in P\setminus\ell$, and is $3$ if $z\notin P$. Take a basis $\varphi_1,\varphi_2,\varphi_3$ of the linear forms with $P=\ker\varphi_3$ and $\ell=\ker\varphi_2\cap\ker\varphi_3$. Write $c_m=\varphi_m\circ a$ and $q(z)=(\varphi_2(z),\varphi_3(z))$. Then colour $1$ means $q(z)=(0,0)$ and $\varphi_1(z)=1$; colour $2$ means $q(z)=(1,0)$; and colour $3$ means $\varphi_3(z)=1$. Three spanning trees $T_1,T_2,T_3$ of $G'$ satisfy \emph{the rule} if \begin{enumerate} \item no edge lies in all three trees, and \item a marked edge of colour $m$ does not lie in both trees other than $T_m$. \end{enumerate} \begin{lemma}\label{lem:genTR} If three spanning trees satisfying the rule exist for some flag, then the wanted map $a$ exists, for every $b$. \end{lemma} \begin{proof} For each $m$: once $c_m$ is given on the edges outside $T_m$, exactly one choice on the edges of $T_m$ gives $c_m$ the boundary $\varphi_m\circ b$, because $T_m$ is a spanning tree and the total of $\varphi_m\circ b$ is $0$. We choose the coordinates in three stages. \emph{Stage 3.} Put $c_3(e)=1$ on every edge outside $T_3$. Now $c_3$ is known on all edges. \emph{Stage 2.} On an edge $e$ outside $T_2$: if $e$ is marked with value $z$, $c_3(e)=1$ and $\varphi_3(z)=1$, put $c_2(e)=\varphi_2(z)+1$; in all other cases put $c_2(e)=1$. Now $c_2$ is known on all edges; write $q(e)=(c_2(e),c_3(e))$. \emph{Stage 1.} On every edge $e$ outside $T_1$ choose $c_1(e)$ so that $a(e)\ne0$ and, if $e$ is marked with value $z$, $a(e)\ne z$. We show that stage 1 is possible and that the edges of $T_1$ are also good. There are three kinds of edges: (A) $e\notin T_3$; (B) $e\in T_3$ and $e\notin T_2$; (C) $e\in T_2\cap T_3$, and then $e\notin T_1$ by rule~(i). \emph{Fact 1: in cases (A) and (B), $q(e)\ne(0,0)$.} In (A), $c_3(e)=1$. In (B), if $c_3(e)=0$, then stage 2 gave $c_2(e)=1$. \emph{Fact 2: in cases (A) and (B), if $e$ is marked with value $z$ and $q(e)=q(z)$, then $e\notin T_1$.} In (A), $c_3(e)=1$, so $\varphi_3(z)=1$ and $z$ has colour $3$. If $e$ were outside $T_2$, stage 2 would give $c_2(e)=\varphi_2(z)+1$ and $q(e)\ne q(z)$. So $e\in T_2$, and by rule~(ii) for colour $3$, $e\notin T_1$. In (B), if $c_3(e)=1$, then $\varphi_3(z)=1$ and stage 2 gave $c_2(e)=\varphi_2(z)+1$, which is impossible. So $c_3(e)=0=\varphi_3(z)$ and $c_2(e)=1=\varphi_2(z)$: colour $2$. As $e\in T_3$, rule~(ii) for colour $2$ gives $e\notin T_1$. \emph{The edges in $T_1$} are of kinds (A) and (B). By Fact 1, $q(e)\ne(0,0)$, so $a(e)\ne0$; by Fact 2, $q(e)\ne q(z)$ if $e$ is marked, so $a(e)\ne z$; whatever $c_1(e)$ is. \emph{The edges outside $T_1$.} The value $c_1(e)=0$ is forbidden only if $q(e)=(0,0)$, and the value $c_1(e)=\varphi_1(z)$ is forbidden only if $q(e)=q(z)$. In (A) and (B), $q(e)\ne(0,0)$, so at most one value is forbidden. In (C), both values are forbidden only if $q(z)=(0,0)$, that is, colour $1$; but by rule~(ii) an edge of colour $1$ is not in $T_2$ and $T_3$ together. So stage 1 always has a value. \end{proof} \subsection{The trees exist} \begin{lemma}\label{lem:genPK} Let $G'$ be as above, with at most eight marked edges, each with a colour $1$, $2$ or $3$ \textup{(}any colours\textup{)}. Then three spanning trees satisfying the rule exist. The same holds with at most eleven marked edges, at most eight of each colour. \end{lemma} \begin{proof} The ground set consists of the pairs $(e,m)$ with $e\in E(G')$ and $m\in\{1,2,3\}$, read as ``tree $m$ uses $e$''. In the matroid $M_A$ a set $I$ is independent if, for each $m$, the edges $e$ with $(e,m)\in I$ form a forest. In the matroid $M_B$ a set $I$ is independent if for every edge $e$ the set $I_e=\{m:(e,m)\in I\}$ is allowed: for an unmarked edge, at most two elements; for a marked edge of colour $m$, not both elements other than $m$. Thus $M_B$ is a direct sum of matroids of rank two on the three elements $(e,1),(e,2),(e,3)$: the uniform matroid for an unmarked edge; for colour $m$, the two elements other than $m$ are parallel and $m$ is free. The rank $r_e(T)$ of $T\subseteq\{1,2,3\}$ is $\min(|T|,2)$ for an unmarked edge, and $[m\in T]+[T\not\subseteq\{m\}]$ for a marked edge of colour $m$. Let $n=|V(G')|$. The graph $G'$ is connected, since every cut has at least three edges, so a forest with $n-1$ edges is a spanning tree. A common independent set with $3(n-1)$ elements is the same as three spanning trees satisfying the rule, because each of the three forests has at most $n-1$ edges and hence exactly $n-1$. By Edmonds' theorem such a set exists if and only if $r_A(S)+r_B(S^c)\ge3(n-1)$ for every subset $S$ of the ground set. For such an $S$ let $R_m$ be the set of edges $e$ with $(e,m)\notin S$, let $p_m$ be the number of components of $G'-R_m$, and let $T_e=\{m:e\in R_m\}$. Then $r_A(S)=(n-p_1)+(n-p_2)+(n-p_3)$, and the condition becomes \begin{equation}\label{eq:genPK} (p_1-1)+(p_2-1)+(p_3-1)\le\sum_{e\in E(G')}r_e(T_e). \end{equation} To prove~\eqref{eq:genPK}, let $A=\{m:p_m\ge2\}$. Making $R_m$ empty for every $m\notin A$ does not change the left side, where $p_m=1$ already, and does not increase the right side, which is a rank. So assume $R_m=\emptyset$ for $m\notin A$. For $m\in A$, every component of $G'-R_m$ has at least three leaving edges, all of them in $R_m$, and an edge is counted at most twice; so $|R_m|\ge3p_m/2$ and $p_m-1\le\frac23|R_m|-1$. With $N=|R_1|+|R_2|+|R_3|=\sum_e|T_e|$, the left side of~\eqref{eq:genPK} is at most $\frac23N-|A|$. For every edge, $r_e(T)\ge\frac23|T|$, with one exception: a marked edge of colour $m$ with $T$ equal to the two elements other than $m$, where $r_e(T)=1=\frac23|T|-\frac13$. Indeed, for an unmarked edge, $|T|=1,2,3$ gives $r_e=1,2,2$ against $\frac23,\frac43,2$; for colour $m$, the set $\{m\}$ and a single other element have rank $1$, the sets $\{m,m'\}$ and $\{1,2,3\}$ have rank $2$, and the two other elements together have rank $1$ against $\frac43$. Let $D$ be the number of exceptional edges. The right side of~\eqref{eq:genPK} is at least $\frac23N-\frac13D$, so the right side minus the left side is at least $|A|-\frac13D$. If $|A|\le1$, every $T_e$ has at most one element and $D=0$. If $|A|=2$, every exceptional edge has the colour outside $A$, so $D\le8$ and $|A|-\frac13D\ge-\frac23$. If $|A|=3$, then $D\le11$ and $|A|-\frac13D\ge-\frac23$. Both sides of~\eqref{eq:genPK} are integers, so the inequality holds. \end{proof} \subsection{Proofs of the theorems} \begin{proof}[Proof of Theorem~\ref{thm:gen}] Take a counterexample with the fewest edges. By Lemma~\ref{lem:genleast}, $G'$ has no loop and at least two vertices, every cut of $G'$ has at least three edges, and every edge of $G'$ carries at most one vertex of degree two of $H$, which is a terminal. So we are in the setting of the last two subsections with $j\le7$ marked edges: each marked edge holds its own terminal of degree two, and there are at most seven terminals. Take any flag. Lemma~\ref{lem:genPK} gives three spanning trees satisfying the rule, Lemma~\ref{lem:genTR} gives the wanted map $a$, and $a$ gives a solution of $(H,\beta)$: a contradiction. \end{proof} The two lemmas give more when no induction is needed. \begin{theorem}\label{thm:genMK} Let $G'$ be a multigraph without loops, with at least two vertices, in which every cut has at least three edges. Let at most eight different edges $e_i$ be marked, with any nonzero values $z_i\in K$, and let $b:V(G')\to K$ be any map with total $0$. Then there is a map $a:E(G')\to K$ with boundary $b$ such that $a(e)\ne0$ for every edge and $a(e_i)\ne z_i$ for every marked edge. The same holds with at most eleven marked edges if some flag gives no colour to more than eight of them. \end{theorem} \begin{proof} Lemma~\ref{lem:genPK} and Lemma~\ref{lem:genTR}. \end{proof} \section{A component correspondence for free edge sets}\label{sec:free} This section allows any finite abelian $\Gamma$, any subgroup $B$ with $|B|\ge2$, and $A=\Gamma/B$. Neither group is assumed binary and the quotient need not split. All moves are signed. Subgroup-valued corrections on fundamental cycles already appear in the proof of~\cite[Theorem 6.3]{EHLMNS}, where tree edges are protected modulo a subgroup while cotree edges are corrected. A complementary-subgroup preparation appears in~\cite[Lemma 6.12]{CLSWX}, under different hypotheses and with a different conclusion. The statement below concerns exact fibres and connected components of a specified induced family; kernel cycle correction itself is not new. Let $G$ be connected and let $R\subseteq E(G)$ satisfy that $G-R$ is connected as a spanning graph. Contract each component of $(V(G),R)$ and delete the edges of $R$, retaining all other edges, including loops; call the result $G/R$. Let $\mathcal K_R$ be the induced subgraph of $\Fl(G,\Gamma)$ on states satisfying \[ \quo{f(e)}\ne0\qquad(e\notin R). \] Conservation summed over each $R$-component shows that restriction defines a map \[ p:V(\mathcal K_R)\longrightarrow V(\Fl(G/R,A)),\qquad p(f)=\quo f\big|_{E(G)\setminus R}. \] \begin{theorem}[Free-set component correspondence]\label{thm:free} The map $p$ is surjective, every exact fibre is connected in $\mathcal K_R$, and $p$ induces a bijection between the connected components of $\mathcal K_R$ and those of $\Fl(G/R,A)$. \end{theorem} This is a correspondence for the \emph{restricted induced graph} $\mathcal K_R$. An edge upstairs may project to a path of several moves downstairs; $p$ need not be a graph homomorphism. The two vertex sets are empty together when no downstairs flow exists. \begin{proof} For every $e\in R$, connectedness of $G-R$ gives a cycle $D_e$ containing $e$ and no other edge of $R$: use a simple path in $G-R$ between its ends, or take the loop itself. Direct it so that $\chi_{D_e}(e)=1$. A correction on $D_e$ changes no other $R$-edge. If its coefficient lies in $B$, it preserves all outside-$R$ quotients. \subsection*{Surjectivity on states} Let $h$ be a nowhere-zero $A$-flow on $G/R$. Extend it to an $A$-flow $u$ on $G$. Inside each $R$-component the prescribed vertex imbalances sum to zero by conservation at the contracted vertex, so solve them on a spanning tree of that $R$-component; assign zero to its other $R$-edges, keeping every prescribed outside-$R$ value unchanged. Lemma~\ref{lem:elementary} lifts $u$ to a $\Gamma$-flow $F$, which may have full zeros only on quotient-zero edges of $R$. For every edge $e$ of $R$ whose quotient is zero, choose $z_e\in B\setminus\{0\}$ and set $\beta_e=z_e-F(e)\in B$. Then \[ f=F+\sum_{\substack{e\in R\\u(e)=0}}\beta_e\chi_{D_e} \] has value $z_e$ at each such edge, changes no other $R$-edge, and preserves the nonzero quotient on every outside edge. Thus $f$ is a nowhere-zero state with $p(f)=h$. This constructs a state, not a legal path starting from the possibly zero flow $F$. \subsection*{Realizing a quotient cycle change} Let $f\in V(\mathcal K_R)$ have quotient $u$. Suppose $C$ is a cycle of $G$, $0\ne\sigma\in A$, and $u'=u+\sigma\chi_C$ is nonzero outside $R$. Choose a lift $\gamma\in\Gamma$ of $\sigma$. Every newly zero quotient edge lies in $C\cap R$ and currently has nonzero quotient. For each such edge $e$, choose $\beta_e\in B$ avoiding the single value \[ -f(e)-\gamma\chi_C(e). \] There is a choice since $|B|\ge2$. If $\beta_e\ne0$, add $\beta_e\chi_{D_e}$. This is legal: $e$ has nonzero quotient before and during preparation, and all other changed edges lie outside $R$ with nonzero quotient. Preparations leave other $R$-edge data unchanged. Now add $\gamma\chi_C$. Newly zero quotients have nonzero full values by construction. An edge with quotient zero both before and after cannot lie on $C$, since $\sigma\ne0$; it belongs to $R$ and was untouched by the preparations. All other edges are protected by their final quotients. The result is a state with quotient $u'$, reached within $\mathcal K_R$. \subsection*{Connected exact fibres} Suppose $p(f)=p(g)$. The quotient difference $\quo g-\quo f$ is supported on $R$. Decompose it into signed cycle flows within $R$, using Lemma~\ref{lem:elementary}, and realize the nonzero terms one at a time by the preceding construction. Outside-$R$ quotients never change, so the path stays in the exact fibre. Now the quotients agree everywhere. At each quotient-zero edge $e$, both the current full value and $g(e)$ are nonzero elements of $B$. Add their difference on $D_e$, skipping a zero coefficient. This changes $e$ directly to the nonzero target value, leaves every other $R$-edge unchanged, and changes only quotient-protected edges outside $R$. After these legal alignments the remaining difference is a $B$-valued flow supported on the subgraph of nonzero quotient edges. Decompose it into signed cycles \emph{of that subgraph} and add them successively. Every intermediate value is protected by its unchanged quotient. This reaches $g$ in the exact fibre. The explicit full-value alignment at quotient-zero edges is necessary when $B$ has more than two elements. \subsection*{Lifting paths from the contracted graph} A directed cycle of $G/R$ lifts to a directed cycle of $G$ by joining its entry and exit vertices with a simple $R$-path inside each visited contracted component. Distinct visited components are disjoint. A quotient loop uses its single outside edge and an internal $R$-path, or that edge alone if already a loop. The lifted cycle's signed restriction is the prescribed downstairs vector. The quotient-cycle construction therefore realizes any downstairs step starting from any lift. Connected exact fibres then allow any prescribed terminal lift. Consequently downstairs-connected states have all their lifts in one component of $\mathcal K_R$. \subsection*{Projecting paths to the contracted graph} Consider a legal step $f\to f+\gamma\chi_C$ within $\mathcal K_R$. If $\quo\gamma=0$, its projection is unchanged. Otherwise set $\sigma=\quo\gamma$. Restrict the \emph{integer} signed vector $\chi_C$ to outside-$R$ edges after contraction. It is an integer circulation $z$ with coefficients in $\{0,1,-1\}$, each used edge occurring once. Orient its used edges according to these signs. The resulting directed graph is balanced at each vertex, and so decomposes into edge-disjoint directed cycles, including loops. There is no cancellation and every edge keeps its sign in $z$. Add $\sigma$ on these directed quotient cycles successively. Each edge has either its old or its final value, both nonzero because the two full endpoint states lie in $\mathcal K_R$. All intermediate quotient flows are therefore nowhere zero. If the projected support is empty, no move is needed. This is an integer signed decomposition, not an even-support decomposition modulo two. Thus an upstairs path projects to a downstairs path. Together with state surjectivity and the converse path implication, this proves the bijection on components. \end{proof} \begin{corollary}[A common ambient component for fixed $B$]\label{cor:anchor} Suppose $G$ has a nowhere-zero $A$-flow. For each edge set $R$ with $G-R$ connected and $\Fl(G/R,A)$ connected, the family $\mathcal K_R$ is nonempty and connected. All such families lie in one component of $\Fl(G,\Gamma)$. \end{corollary} \begin{proof} Lift a nowhere-zero $A$-flow on $G$ by Lemma~\ref{lem:elementary}. The lift has nonzero quotient on every edge, hence belongs to every such family. Theorem~\ref{thm:free} makes each family connected. \end{proof} The assumption $|B|\ge2$ cannot simply be dropped. Take three parallel edges between two vertices, let $R$ be one edge, and let $\Gamma=A=\Z_2$ with trivial $B$. The graph $G-R$ is connected and $G/R$ consists of two loops, which has a nowhere-zero $A$-flow. The original cubic graph has none: three values equal to one do not conserve at a vertex. Thus state surjectivity already fails. Distinct components of $\mathcal K_R$ can in principle merge along paths leaving the family. A disconnected downstairs graph therefore does not furnish a counterexample to unrestricted $\Gamma$-flow connectivity. Corollary~\ref{cor:anchor} also fixes $B$; it makes no identification across different subgroup choices. \section{Outlook and limitations}\label{sec:outlook} Universal connectivity for $\F_2^3$ or $\F_2^4$ on all bridgeless graphs remains open. Theorems~\ref{thm:oddbound} and~\ref{thm:order16bound} give the respective bounds $44$ and $19022$ on odd-degree vertices per component. A least core has at least $46$ vertices at order eight and $19024$ at order sixteen. The general formula and its intermediate order-sixteen bound $764$ remain valid without the four-cut theorem. The order-sixteen bounds $1020$, $3036$, $3996$ and $19022$ use the four-cut reduction and seventeen-cycle exclusion; the last three also use rounded partition density, the small-shore count and the edge-and-claw factor of Section~\ref{sec:claws}. Lemma~\ref{lem:widecore} uses Theorem~1.2(c) of the unpublished manuscript~\cite{Eulerian}, and so, through it, do Theorems~\ref{thm:oddbound}, \ref{thm:1020}, \ref{thm:3036}, \ref{thm:b3996} and~\ref{thm:order16bound} and Corollaries~\ref{cor:44-764}, \ref{cor:cyclic5to10} and~\ref{cor:cqleast}; the order-$32$ results, Theorem~\ref{thm:main} and Corollary~\ref{cor:exp4}, do not use it. The bounds $3036$, $3996$ and $19022$ also use Lemma~\ref{lem:clawmain}, whose proof takes a frame as in Proposition~\ref{prop:frame}; Theorem~\ref{thm:cyclic11} uses Lemma~\ref{lem:clawmain} but not Lemma~\ref{lem:widecore}. Whether cycles of length eighteen can be excluded, and whether these bounds are optimal, is open. Theorem~\ref{thm:fivestars} permits at most five designated exceptional star cuts; it does not cover arbitrary collections of five small cuts. In the cubic setting, Corollaries~\ref{cor:matching} and~\ref{cor:fivesupport} give sufficient matching and support tests, but no construction meeting either test for every flow is supplied. Theorem~\ref{thm:fourcut} removes a four-cut only under its explicit structural and smaller-graph hypotheses; it does not remove a five-cut. Another possible route would be to find, after legal reconfiguration, an edge set $R$ with connected complement for which the appropriate quotient flow graph of $G/R$ is connected. Theorem~\ref{thm:free} would then control that restricted family. To prove a universal conclusion this route still needs an accessibility theorem and a way to place all starting states in a common ambient component. The component correspondence alone does not give these. The bounds for threads of up to three edges do not permit the same proof with order $32$ replaced by order $16$: a paired list can have six entries, exceeding a quotient of order four, and a free thread can forbid three kernel values, exceeding a kernel of order two. Lemma~\ref{lem:clawmain} uses the separate one-edge-thread case. The short-cycle contraction families of~\cite[Theorem 5.9 and Corollary 5.14]{CLSWX} are not new here. Theorem~\ref{thm:cyclic11} supplies a different sufficient support construction for cubic graphs with all cyclic cuts of size at least eleven. Corollary~\ref{cor:cyclic5to10} shows that a remaining least core has a cyclic cut of size five through ten. The factor theorem places no upper bound on the number of uncovered vertices or support cycles, so it does not by itself settle these smaller cyclic cuts. Sections~\ref{sec:b3996} and~\ref{sec:finalquotient} apply the same factor to every least core; they bound its size from below and do not exclude it. Theorem~\ref{thm:gen} is an existence statement and enters no connectivity proof here. We do not know whether the factor theorem is new. \subsection*{Dependencies and status} The all-bridgeless binary-extension proof is an extension of Part~I, with extraction and frame existence inherited from Section~\ref{a2:sec} and Appendix~\ref{app:sec}, and the finite-round marginal estimate inherited from the proof of Lemma~\ref{lem:AVG}. Their statements and uses are explicit in Section~\ref{sec:inherited}. The local arguments are written here in the all-nonzero-coefficients setting; the signed global induction is included in full. The separate four-edge-connected proof in Section~\ref{sec:four} uses the classical tree-packing criterion, elementary boundary solving on a tree, and binary cycle decomposition; its other lemmas are proved there. Section~\ref{sec:fivestars} builds on those proved lemmas and includes the binary short-cycle lift used in its argument. Section~\ref{sec:fourcut} proves the conditional four-cut theorem, with Jaeger's existence theorem as an external input. Section~\ref{sec:oddbound} proves the wide least-odd-count interface, with the existence input of Section~\ref{sec:inherited} and the two-odd-vertex theorem~\cite[Theorem 1.2(c)]{Eulerian} as its outside inputs, and supplies the complete cycle-order lifting and family-intersection proof. It is independent of the four-cut theorem and does not use the bound $2^{q/2+1}-4$ of~\cite[Theorem 1.4]{Eulerian} beyond its range. The independent free-set correspondence uses only elementary cycle decomposition and connectedness of $G-R$. Section~\ref{sec:c17} gives the full special order-sixteen argument. Its core input is Lemma~\ref{lem:widecore}; its family joining uses Corollary~\ref{cor:fourcore}; its local crossing exits use only cubicity and the fixed cycle length and group order. Lemma~\ref{lem:cycleorder} supplies the exclusion of sixteen-cycles, which the proof of Theorem~\ref{thm:1020} combines with Lemma~\ref{lem:c17} to obtain girth at least eighteen. The component-product argument is explicit in the proof of Theorem~\ref{thm:1020}. Section~\ref{sec:claws} proves the factor theorem in full; the proof uses no enumeration. The partition count and support transfer are written out there. Lemma~\ref{lem:clawmain} specializes the earlier frame proofs to single-edge threads, explicitly checking the smaller kernel bound; Theorem~\ref{thm:cyclic11} also uses the nowhere-zero existence input from Section~\ref{sec:inherited}. Within that section, only the least-core corollary uses the four-cut reduction. Corollary~\ref{cor:roundedpartition} gives the stronger integer criterion used in Section~\ref{sec:smallshores}. The proofs of Section~\ref{sec:smallshores} use no graph enumeration. Section~\ref{sec:b3996} uses the properties of a least core (Lemmas~\ref{lem:widecore}, \ref{lem:cycleorder} and~\ref{lem:c17}, Corollary~\ref{cor:fourcore}), Sections~\ref{sec:claws} and~\ref{sec:smallshores} and Corollary~\ref{cor:roundedpartition}. Section~\ref{sec:cqplus} uses Lemmas~\ref{lem:twotrees}, \ref{lem:goodcontract} and~\ref{lem:shortbinary}, Corollary~\ref{cor:alltrees}, Theorem~\ref{thm:four} and Corollary~\ref{cor:44-764}, and, for a least core, Lemma~\ref{lem:widecore} and Corollary~\ref{cor:fourcore}, with the matroid union theorem~\cite{MatroidUnion} as an external input. Section~\ref{sec:finalquotient} uses Section~\ref{sec:cqplus}, the properties of a least core, Section~\ref{sec:claws}, the argument of Corollary~\ref{cor:roundedpartition}, the small-shore count of Lemma~\ref{lem:smallshore}, the last paragraph of the proof of Theorem~\ref{thm:3036} and the tree-packing theorem~\cite{NashWilliamsTrees,TutteTrees}; its finite arithmetic is a table of twelve rows and the values of two functions at twenty-four arguments, all checkable by hand; no graph enumeration is used. Section~\ref{sec:gen} is independent of all other sections; its external inputs are Jaeger's theorem~\cite{Jaeger} and Edmonds' matroid intersection theorem~\cite{EdmondsMI}. \part{Recolouring 2-degenerate graphs with four colours} \setlist[enumerate]{} \section{Recolouring: introduction}\label{sec:introIII} In Part~III all graphs are finite and simple. Let $G$ be a graph and $k$ a positive integer. The \emph{$k$-recolouring graph} $\Rk{k}(G)$ has as vertices the proper $k$-colourings of $G$; two colourings are adjacent if they differ at exactly one vertex. A graph is \emph{$d$-degenerate} if every subgraph has a vertex of degree at most $d$. A $d$-degenerate graph has a proper $(d+1)$-colouring, and for $k \ge d+2$ the graph $\Rk{k}(G)$ is connected (Dyer, Flaxman, Frieze and Vigoda \cite{DFFV06}; see also \cite[Thm.~2.7]{Cer07}). For $k = d+1$ it can be disconnected ($G = K_{d+1}$). For $d = 2$ every component of $\Rk{3}(G)$ has diameter $O(n^2)$ \cite{CvdHJ11}, but for every $d \ge 3$ there are $d$-degenerate graphs $G$ for which $\Rk{d+1}(G)$ has a component of superpolynomial diameter, at least $2^{c\sqrt{n}}$ for some $c > 0$ \cite[Thm.~1.4]{BC09} (these graphs have degeneracy $d$; see \cite[p.~109]{Cer07}). Cereceda asked how large the diameter is for $k \ge d+2$. \begin{conjecture}[Cereceda \cite{Cer07}]\label{conj:cereceda} For every $d \ge 1$ and every $d$-degenerate graph $G$ on $n$ vertices, $\diam \Rk{d+2}(G) = O(n^2)$. \end{conjecture} The conjecture holds for $d = 1$ (forests): Cereceda proved $\diam \Rk{k}(G) = O(n^2)$ for $k \ge 2d+1$ \cite[Thm.~5.22]{Cer07}, and $2d+1 = d+2$ when $d = 1$; see also \cite[Thms.~2 and~6]{BJLPP14}. Bousquet and Heinrich \cite{BH22} proved $\diam \Rk{k}(G) = O(n^{d+1})$ for every $d$-degenerate $G$ and $k \ge d+2$, the first polynomial bound for $k = d+2$; for $d = 2$ this is $O(n^3)$. They also proved the quadratic bound once $k \ge \tfrac32 (d+1)$, which for $d = 2$ means $k \ge 5$. Before their work no polynomial bound was known for $\Rk{4}(G)$ over all $2$-degenerate graphs $G$. For $d = 2$ the following partial results are known ($k = 4$ unless stated otherwise): \begin{itemize} \item sparse graphs, measured by the maximum average degree $\mad(G)$: Feghali \cite{Feg21}, improving the polynomial bound of Bousquet and Perarnau \cite[Thm.~2]{BP16}, proved $\diam \Rk{k}(G) = O(n (\log n)^{d-1})$ when $\mad(G) \le d - \varepsilon$ and $k \ge d+1$ (for $k = 4$: $\mad(G) \le 3 - \varepsilon$ for a fixed $\varepsilon > 0$, and the bound $O(n \log^2 n)$), and Zheng and Lu \cite{ZL26} sharpened the exponent of the logarithm; with lists of size $4$, Cranston \cite[Thm.~3]{Cra22} proved that each vertex need be recoloured at most $14$ times when $\mad(G) < 22/9$. Pan, Wang and Liu \cite[Thm.~1.4]{PWL25} extended the density range to $\mad(G) < 5/2$, with at most $18$ recolourings per vertex, including planar graphs of girth at least $10$; \item $2$-degenerate graphs of maximum degree at most $3$: $\diam \Rk{4}(G) = O(n^2)$ by Feghali, Johnson and Paulusma \cite[Thm.~5]{FJP16} (for $\Delta = 3$, by \cite[Thm.~2]{FJP16}: for every connected $G$ with $\Delta \ge 3$, $\Rk{\Delta+1}(G)$ has at most one component that is not a single vertex, and its diameter is $O(n^2)$; for $2$-degenerate $G$ the graph $\Rk{4}(G)$ is connected); \item chordal bipartite graphs (not all of them $2$-degenerate): at most $2n^2$ moves for every $k \ge 3$ \cite[Thms.~2, 8, 10]{BJLPP14}; \item planar graphs of large girth (see Table~1 of \cite{BBFHMP21}): for $4$ colours, $n$ times a power of $\log n$ at girth $7$ and more \cite{Feg21}, and $O(n^3)$ at girth $6$ \cite{BH22}; at girth $5$ (planar graphs of girth $5$ need not be $2$-degenerate) the graph $\Rk{4}$ is connected \cite{BBFHMP21}; \item graphs of treewidth at most $2$: $2n(n+1)$ for $k \ge 4$ by Bonamy and Bousquet \cite[Thm.~2]{BB18} (who prove $2(n^2+n)$ for $k \ge \mathrm{tw}+2$), linear for $k \ge 5$ \cite{BBH20}, while $k = 4$ needs order $n^2$ moves \cite[Thm.~12, $k = 3$]{BJLPP14}; and $d$-degenerate chordal graphs, linear for $k \ge 2d+1$ \cite{Chordal25}. \end{itemize} A $2$-degenerate graph can have maximum average degree arbitrarily close to $4$, for example $K_{2,m}$. Known linear subfamilies already have unbounded treewidth and maximum average degree at least $3$: graphs obtained by subdividing every edge once have Grundy number at most $3$, so Bonamy and Bousquet's theorem \cite[Thm.~1]{BB18} gives four-colour diameter at most $8n$. Here the Grundy number is the largest number of colours used by a greedy colouring over all vertex orders (in this paragraph greedy colours are numbered $1, 2, \dots$). Indeed, a subdivision vertex has degree $2$; if an original vertex first received greedy colour at least $4$, its earlier colour-$3$ neighbour would be a subdivision vertex with at most one earlier neighbour, a contradiction. Once-subdivided $K_m$ has treewidth $m-1$ and average degree $4(m-1)/(m+1)$, which is at least $3$ for $m \ge 7$. Our result covers the entire class of bipartite $2$-degenerate graphs. \subsection*{Our results} \begin{theorem}\label{thm:bip} For every bipartite graph $G$ on $n$ vertices of degeneracy at most $2$, $\diam \Rk{4}(G) \le 2n(n+1)$. \end{theorem} The class contains all trees, even cycles, square grids, every graph obtained by subdividing each edge of a graph once, and graphs of every treewidth. \begin{proposition}\label{prop:lb} The order $n^2$ in Theorem~\ref{thm:bip} cannot be improved: there are bipartite planar graphs $H_n$ of degeneracy $2$ on $N = n+2$ vertices, for $n \ge 5$, with $\diam \Rk{4}(H_n) \ge (N-7)^2/8$, and bipartite graphs $J_n$ of degeneracy $2$ and girth $6$ on $N = n+4$ vertices, for $n \ge 29$, with $\diam \Rk{4}(J_n) \ge (N-33)^2/128$. ($J_n$ has girth $6$, so no bound of lower order than $n^2$ holds for $2$-degenerate graphs of girth at least $5$ or $6$.) \end{proposition} Quadratic lower bounds for four colours were known before, already for graphs of treewidth $2$: the fan (a path together with a vertex joined to all its vertices) has $\diam \Rk{4} = \Theta(n^2)$ \cite[Thm.~12, $k = 3$]{BJLPP14} (see also \cite{BBH20}); the point of $H_n$ and $J_n$ is that they lie in the class of Theorem~\ref{thm:bip}: they are bipartite, and $J_n$ even has girth $6$. \begin{theorem}[computer-assisted]\label{thm:conewidth} For every graph $G$ on $n$ vertices of cone width at most $2$ (Section~\ref{sec:cw}), $\diam \Rk{4}(G) \le 4n(n+1)$. The class contains every graph of treewidth at most $2$ and graphs of treewidth $3$ with triangles, and it is closed under subgraphs (Proposition~\ref{prop:cwinside}). \end{theorem} \begin{theorem}\label{thm:terminal} Let $G$ be a graph on $n$ vertices with an independent set $Z$ of vertices of degree at most $2$ such that $G - Z$ is bipartite and $2$-degenerate. Then $\diam \Rk{4}(G) \le 5n(n+1)$. (A more general form is Theorem~\ref{thm:terminalclass}.) \end{theorem} All upper bounds go through one reduction (Lemma~\ref{lem:sweep}, the \emph{sweep}): a quadratic bound for $G$ follows from a \emph{linear} bound for the instances on the cones, in which the colour of one vertex (the sink) is prescribed and every other vertex is greedy. In every instance we use, each vertex is recoloured a bounded number of times (two for bipartite instances). \section{Setting: orders, cones, instances, and the sweep}\label{sec:setting} Colours are $0 < 1 < 2 < 3$. Let $G$ have an order $v_1, \dots, v_n$ in which every vertex has at most two \emph{later} neighbours, its \emph{out-neighbours}; such an order exists exactly when $G$ is $2$-degenerate (take a degeneracy order). An arc goes from a vertex to each of its out-neighbours. \begin{definition}[cones] For $1 \le t \le n$, the \emph{cone} $C_t$ is the set of vertices $v_j$, $j \le t$, from which $v_t$ can be reached along arcs inside $\{v_1, \dots, v_t\}$. Note $v_t \in C_t$ and $\sum_t |C_t| \le 1 + 2 + \dots + n$. \end{definition} \begin{definition}[instances] An \emph{instance} has vertices in a \emph{creation order}, the \emph{sink} $s$ first; every vertex $v$ has a set $N^+(v)$ of at most two \emph{out-neighbours} among the vertices before it, the sink has none, and the edges are the pairs $\{v,u\}$, $u \in N^+(v)$. A vertex other than the sink without out-neighbours is a \emph{root vertex}. Every vertex has a list $L(v) \subseteq \{0,1,2,3\}$ with $|L(v)| \ge |N^+(v)| + 2$; the sink has a list $\{g, x\}$ with $0 \le g < x \le 3$ (the \emph{main case} is $g = 0$). A list is \emph{tight} if $|L(v)| = |N^+(v)| + 2$ (the sink's list always is). The colourings $\alpha$ and $\beta$ are \emph{greedy}: $\alpha(s) = x$, $\beta(s) = g$, and every other vertex, in creation order, takes the least colour of its list that none of its out-neighbours has (so $\alpha$ and $\beta$ depend only on the out-neighbours and the lists, not on the creation order). A \emph{sequence} recolours one vertex at a time, to a colour of its list that none of its neighbours has at that time. Whether an instance lies in one of the classes below depends only on its graph and its arcs, not on the lists or the colours. \end{definition} \begin{lemma}\label{lem:caseA} In an instance whose sink has the list $\{g, x\}$ with $0 < g < x$ ($\alpha$ has the sink on $x$, $\beta$ on $g$, all other vertices greedy) there is a sequence from $\alpha$ to $\beta$ of at most $2t$ moves, $t$ the number of vertices. \end{lemma} We use heights on trees. The \emph{residue} of the colours $1, 2, 3$ is $1, 2, 0$. Let $T$ be a tree and $c$ a proper colouring of $T$ with colours in $\{1,2,3\}$. A \emph{height} of $c$ is an integer function $h$ on $V(T)$ with $h(v) \equiv \mathrm{res}\,c(v) \pmod 3$ and $|h(u) - h(v)| = 1$ on every edge; it exists (adjacent residues differ by $\pm 1 \bmod 3$) and is unique up to an additive constant in $3\mathbb{Z}$. \begin{lemma}[one tree]\label{lem:T} Let $T$ be a tree with root $r$; every vertex other than $r$ may use the colours $1,2,3$, and $r$ may use a set $P(r) \subseteq \{1,2,3\}$. Let $c, c'$ be proper colourings of $T$ within these sets, $h$ a height of $c$, and $E$ a function to the even integers such that $h + E$ is a height of $c'$ and $|E(r)| \le 2$ if $|P(r)| \le 2$. Then $c$ can be changed into $c'$ inside $T$ by exactly $\frac12 \sum_v |E(v)|$ moves, every intermediate colouring proper and within the allowed colours. \end{lemma} \begin{proof} Induction on $\sum |E|$. If $E = 0$ then $c' = c$. Suppose some $E(v) > 0$, and choose such $v$ with $h(v)$ least. If a neighbour $u$ had $h(u) = h(v) - 1$, then $(h+E)(u) \ge (h+E)(v) - 1 \ge h(v) + 1 = h(u) + 2$, so $E(u) \ge 2$ and $h(u) < h(v)$, against the choice of $v$. So every neighbour of $v$ has height $h(v)+1$ and hence the colour of residue $h(v)+1$. Recolour $v$ with the colour $y$ of residue $h(v) + 2$: the colouring stays proper. $y$ is allowed at $v$: clear unless $v = r$ and $|P(r)| \le 2$; then $E(r) = 2$, so $\mathrm{res}\,c'(r) = h(r) + 2$ and $y = c'(r)$. The new colouring has the height $h + 2[v]$, where $[v]$ is the indicator function of $v$, and $E - 2[v]$ satisfies the hypotheses with $\sum |E|$ smaller by $2$. If no $E(v) > 0$, argue symmetrically with $E(v) < 0$ and $h(v)$ largest. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:caseA}] The sink is never on $0$, so by induction along the creation order a vertex $v \ne s$ has $\alpha(v) = 0$ if and only if $0 \in L(v)$ and no out-neighbour has the colour $0$, and the same for $\beta$; hence $K = \alpha^{-1}(0) = \beta^{-1}(0)$. Let $F$ be the set of the other vertices. (F1) Every $v \in F$ has at most one out-neighbour in $F$: if $0 \in L(v)$, some out-neighbour is on $0$; if not, $|L(v)| \le 3$ and $v$ has at most one out-neighbour. So $F$ induces a forest in which each vertex has at most one \emph{parent} (an out-neighbour in $F$); all neighbours of a vertex of $F$ other than its parent and its children lie in $K$ (so every neighbour of a vertex of a component $T$ of $F$ that lies outside $T$ is in $K$). (F2) A vertex of $F$ with a parent has $\{1,2,3\} \subseteq L(v)$ (if $0 \in L(v)$ it has two out-neighbours and $|L(v)| = 4$; else $L(v) = \{1,2,3\}$). A root $r \ne s$ of $F$ has all its out-neighbours on $0$, so $\alpha(r) = \beta(r) = \min (L(r) \setminus \{0\})$; a vertex of $F$ with a parent $p$ takes the least colour of $\{1,2,3\}$ other than the colour of $p$, in $\alpha$ and in $\beta$. Hence $\alpha = \beta$ outside the tree $T$ of $F$ that contains $s$, and in a greedy colouring $c$ the child of a vertex on $1$ is on $2$ and the child of a vertex on $2$ or $3$ is on $1$. For a height $h$ of $c$ on $T$ this gives, for $v \in T \setminus \{s\}$: $h(v) - h(s) \in \{0,1\}$ if $c(s) = 1$, $\in \{-1, 0\}$ if $c(s) = 2$, and $\in \{1,2\}$ if $c(s) = 3$. Let $h_\alpha, h_\beta$ be heights of $\alpha|_T, \beta|_T$ with $E = h_\beta - h_\alpha$ and $E(s) \in \{-2, 2\}$ (possible since $\alpha(s) \ne \beta(s)$). Then $E(v) - E(s) = (h_\beta(v) - h_\beta(s)) - (h_\alpha(v) - h_\alpha(s)) \in [-3, 3]$, and $E(v)$ is even (every height has $h(v) - h(s) \equiv \mathrm{dist}_T(v,s) \pmod 2$), so $|E(v)| \le 4$. Apply Lemma~\ref{lem:T} on $T$ with root $s$ and $P(s) = L(s) = \{g, x\}$; every other vertex of $T$ has a parent in $T$, so by (F2) the colours $1, 2, 3$ are in its list. While $K$ stays on $0$ (no vertex of $F$ uses $0$), this turns $\alpha$ into $\beta$ in $\frac12 \sum_{v \in T} |E(v)| \le 2|T| \le 2t$ moves. \end{proof} \begin{lemma}[the sweep]\label{lem:sweep} Let $\mathcal{C}$ be a class of instances (closed under changing the lists) and $C \ge 2$ such that every instance $I$ of $\mathcal{C}$ in the main case has a sequence from $\alpha$ to $\beta$ of at most $C|V(I)|$ moves. Let $G$ have an order $v_1, \dots, v_n$ in which every vertex has at most $2$ later neighbours and in which, for every $t$, the instance on $C_t$ with sink $v_t$ defined below lies in $\mathcal{C}$. Then \[ \diam \Rk{4}(G) \le 2C\,(|C_1| + \dots + |C_n|) \le C n(n+1). \] \end{lemma} \begin{proof} A vertex $v$ is \emph{greedy} in a colouring $c$ of $G$ if $c(v)$ is the least colour that no out-neighbour of $v$ has. There is exactly one colouring $\gamma$ in which all vertices are greedy (colour $v_n, v_{n-1}, \dots, v_1$ in turn). We show that from any colouring $\gamma$ is reached in at most $C \sum_t |C_t|$ moves; two colourings are then joined through $\gamma$. The sweep has steps $t = 1, \dots, n$; after step $t$ the vertices $v_1, \dots, v_t$ are greedy. At step $t$ let $c$ be the current colouring, $x = c(v_t)$, and $g$ the least colour that no out-neighbour of $v_t$ has. If $x = g$ nothing is done. Otherwise take the instance on $C_t$ with sink $v_t$, out-neighbours of each $v \in C_t$ its out-neighbours in $G$ that lie in $C_t$, creation order $v_t$ followed by the other vertices of $C_t$ by decreasing index (so every vertex comes after its out-neighbours), $L(v_t) = \{g, x\}$ (we show below that $g < x$), and for every other $v \in C_t$ the list $L(v) = \{0,1,2,3\}$ minus the colours of the out-neighbours of $v$ outside $C_t$. Then $|L(v)| \ge 4 - |N^+(v) \setminus C_t| \ge |N^+(v) \cap C_t| + 2$. Every neighbour of a vertex $v \in C_t$ that lies outside $C_t$ is an out-neighbour of $v$: an in-neighbour of $v$ is earlier than $v$, hence among $v_1, \dots, v_t$, and it reaches $v_t$ through $v$, so it lies in $C_t$. Hence a sequence of the instance is a sequence in $G$: the neighbours outside $C_t$ do not move and their colours are not in the lists. The colouring $c$ restricted to $C_t$ is $\alpha$: the vertices of $C_t$ other than $v_t$ are greedy before the step, and the greedy colour of a vertex in $G$ is the least colour of its list that no out-neighbour inside $C_t$ has (by induction along the creation order). Likewise $\beta$ makes all of $C_t$ greedy. A vertex $v_j$, $j < t$, outside $C_t$ has no out-neighbour in $C_t$ (else it would lie in $C_t$), so it stays greedy. If $g = 0$ the instance lies in $\mathcal{C}$ and has the list $\{0, x\}$ at the sink: at most $C|C_t|$ moves. If $g \ne 0$, then $x$ is also a colour no out-neighbour of $v_t$ has, so $0 < g < x$, and Lemma~\ref{lem:caseA} gives at most $2|C_t| \le C|C_t|$ moves. \end{proof} \begin{remark}\label{rem:noroots} In a cone every vertex other than the sink has an out-neighbour inside the cone, so the instances of the sweep have no root vertices. Membership of an instance in the classes below does not depend on the lists or on the colouring. \end{remark} \section{Bipartite graphs: proof of Theorem~\ref{thm:bip}}\label{sec:bip} \begin{theorem}\label{thm:BIPinst} In every instance without root vertices whose graph is bipartite there is a sequence from $\alpha$ to $\beta$ in which the sink moves exactly once and every other vertex at most twice. In particular $\mathrm{dist}(\alpha,\beta) \le 2t - 1$. \end{theorem} Throughout this section the instance has no root vertices, so every vertex other than the sink has one or two out-neighbours, and a vertex with two out-neighbours has the list $\{0,1,2,3\}$. If the sink has the list $\{g, x\}$ with $g \ne 0$, the sequence of Lemma~\ref{lem:caseA} already has the required form: in its proof, Lemma~\ref{lem:T} moves every vertex $v$ of the tree $T$ exactly $|E(v)|/2 \le 2$ times and the sink exactly once ($|E(s)| = 2$), and no other vertex moves. So from now on $g = 0$ (the main case). \begin{lemma}\label{lem:s1} For every vertex $v \ne s$, $\alpha(v), \beta(v) \in \{0,1,2\}$. \end{lemma} \begin{proof} At most one colour of $L(v)$ is $3$, so at least $|N^+(v)| + 1$ colours of $L(v)$ lie in $\{0,1,2\}$; the out-neighbours occupy at most $|N^+(v)|$ of them, and the greedy colour is the least free colour of the list. \end{proof} The \emph{pattern} of $v$ is the pair $(\alpha(v), \beta(v))$; the pattern of the sink is $(x, 0)$. A vertex is \emph{fixed} if $\alpha(v) = \beta(v)$, a \emph{low swapper} if its pattern is $(0,1)$ or $(1,0)$, and a \emph{high mover} if it is not fixed and has the colour $2$ at one end. By Lemma~\ref{lem:s1} every vertex other than the sink is of exactly one of these kinds; the sink is a low swapper if $x = 1$, a high mover if $x = 2$, and none of the three if $x = 3$. A \emph{leaver} has a pattern $(0, b)$ with $b \ne 0$, a \emph{joiner} a pattern $(a, 0)$ with $a \ne 0$; the sink is a joiner. \begin{lemma}\label{lem:s3} Every leaver has an out-neighbour that is a joiner, and every joiner other than $s$ has an out-neighbour that is a leaver. \end{lemma} \begin{proof} A leaver $v$ has $0 \in L(v)$ and is not on $0$ in $\beta$, so some out-neighbour $u$ has $\beta(u) = 0$; and $\alpha(u) \ne 0$ because $\alpha(v) = 0$. A joiner $v \ne s$ has $0 \in L(v)$ and is not on $0$ in $\alpha$, so some out-neighbour $u$ has $\alpha(u) = 0$, and $\beta(u) \ne 0$; $u \ne s$ because $\alpha(s) = x \ne 0$. \end{proof} \begin{lemma}[upper middle]\label{lem:um} Let $\mu$ be a proper colouring of an instance with $\mu(v) \in L(v)$, $\mu(v) \ge \alpha(v)$ and $\mu(v) \ge \beta(v)$ for every $v$, and $\mu(s) = x$. Recolour first every vertex from $\alpha(v)$ to $\mu(v)$, the vertices in falling order of $\mu$, and then every vertex from $\mu(v)$ to $\beta(v)$, in rising order of $\mu$ (ties in any order; no move when the colours agree). This is a sequence from $\alpha$ to $\beta$ in which every vertex moves at most twice and the sink once. \end{lemma} \begin{proof} Let $v, w$ be adjacent, so $\mu(v) \ne \mu(w)$; say $\mu(w) > \mu(v)$. In the first part $w$ moves first, while $v$ is on $\alpha(v) \le \mu(v) < \mu(w)$; when $v$ moves, $w$ is on $\mu(w) \ne \mu(v)$. In the second part $v$ moves first, to $\beta(v)$, while $w$ is on $\mu(w) > \mu(v) \ge \beta(v)$; when $w$ moves to $\beta(w)$, $v$ is on $\beta(v) \ne \beta(w)$. All colours used lie in the lists. The sink starts on $x = \mu(s)$ and moves only in the second part. \end{proof} So Theorem~\ref{thm:BIPinst} follows once we find such a $\mu$. For the rest of the section the instance is bipartite; it is connected (every vertex reaches $s$), $S_0$ denotes the side of $s$ and $S_1$ the other side, and the out-neighbours of a vertex lie on the other side. \begin{lemma}[sides]\label{lem:sides} In a bipartite instance without root vertices, a vertex of $S_0 \setminus \{s\}$ has one of the patterns $(0,0)$, $(1,1)$, $(2,2)$, $(1,0)$, $(2,0)$, $(1,2)$, and a vertex of $S_1$ one of $(0,0)$, $(1,1)$, $(2,2)$, $(0,1)$, $(0,2)$, $(2,1)$. Equivalently: every joiner lies on $S_0$ and every leaver on $S_1$; and a vertex that is not fixed and avoids $0$ at both ends has the pattern $(1,2)$ on $S_0$ and $(2,1)$ on $S_1$. \end{lemma} \begin{proof} Induction along the creation order; the statement is used only for out-neighbours of the new vertex $v$, which are earlier. If $v$ is a leaver, it has a joiner out-neighbour (Lemma~\ref{lem:s3}), which lies on $S_0$ (by induction, or it is $s$); so $v \in S_1$. If $v \ne s$ is a joiner, it has a leaver out-neighbour, which lies on $S_1$; so $v \in S_0$. Suppose $v \in S_0$ has the pattern $(2,1)$. Then $1 \in L(v)$, and $1$ is not free in $\alpha$: some out-neighbour $p$ has $\alpha(p) = 1$, and $\beta(p) \ne 1$ since $\beta$ is proper. But $p \in S_1$, and the only pattern of $S_1$ with first entry $1$ is $(1,1)$; a contradiction. Suppose $v \in S_1$ has the pattern $(1,2)$. Then $1 \in L(v)$ and some out-neighbour $p$ has $\beta(p) = 1 \ne \alpha(p)$. But $p \in S_0$, and the only pattern there with second entry $1$ is $(1,1)$ (the sink has second entry $0$); a contradiction. The remaining patterns are fixed. \end{proof} \begin{definition}[roles] Each vertex gets a role $D$ (direct), $P_2$ (parked on $2$) or $P_3$ (parked on $3$), in creation order. A colour $r \in \{2,3\}$ is \emph{free for parking} at $v$ if $r \in L(v)$, no out-neighbour of $v$ is parked on $r$, and, if $r = 2$, no out-neighbour of $v$ is fixed on $2$. The sink gets $D$ if $x = 1$, $P_2$ if $x = 2$, $P_3$ if $x = 3$. Every other vertex $v$ gets: (a) $D$ if $v$ is fixed; (b) if $v$ is a low swapper: $D$ if no out-neighbour of $v$ is a low swapper with role $D$, and otherwise $P_r$ for a colour $r$ free for parking; (c) if $v$ is a high mover: $P_r$ for a colour $r$ free for parking. \end{definition} \begin{proposition}\label{prop:roles} In a bipartite instance without root vertices a colour free for parking exists whenever the rule asks for one, and the roles satisfy: (p1) a vertex in $P_r$ has $r$ in its list, no two adjacent vertices share a role $P_2$ or a role $P_3$, and the sink is in $D$, $P_2$, $P_3$ for $x = 1, 2, 3$ (and $x \in L(s)$); (p2) every vertex of $D$ is fixed or a low swapper; (p3) no vertex of $D$ fixed on the colour $2$ has a neighbour in $P_2$; (p4) no two low swappers in $D$ are adjacent. \end{proposition} \begin{proof} Every edge joins a vertex to an out-neighbour, so (p1), (p3), (p4) can be checked when the later end of an edge gets its role; (p2) holds by the rule. For (p3) with the vertex of $P_2$ as the later end, the clause ``$r = 2$'' of the rule applies; the case where the vertex fixed on $2$ is the later end is case (a) below. Let $v$ be the new vertex. (a) $v$ fixed. Only (p3) can fail, if $v$ is fixed on $2$ and has an out-neighbour in $P_2$. We show that every out-neighbour $p$ of a vertex $v$ fixed on $2$ is fixed (hence in $D$). First, $p$ never has the colour $2$, so $p$ is not the sink with $x = 2$. Nor is $p$ the sink with $x = 3$: then $s$ is on $3$ in $\alpha$, and one of $0, 1$ lies in $L(v)$ and is free in $\alpha$ (with two out-neighbours $L(v) = \{0,1,2,3\}$ and the other one takes at most one of $0,1$; with one out-neighbour $L(v)$ has $3$ or $4$ colours, so it contains $0$ or $1$), so $\alpha(v) \in \{0,1\}$. Hence if $p$ is not fixed, it is a low swapper. If $p = (1,0)$ (so $p \in S_0$, $v \in S_1$): $\alpha(v) = 2$ forces $0 \notin L(v)$ or a second out-neighbour $p'$ with $\alpha(p') = 0$, and $\beta(v) = 2$ forces $1 \notin L(v)$ or $\beta(p') = 1$. With one out-neighbour $L(v)$ would miss $0$ and $1$, too small; with two, $L(v) = \{0,1,2,3\}$, so $p'$ has the pattern $(0,1)$ and lies on $S_0$, against Lemma~\ref{lem:sides}. If $p = (0,1)$ (so $p \in S_1$, $v \in S_0$) the same argument gives a second out-neighbour with pattern $(1,0)$ on $S_1$, again impossible. (b) $v$ a low swapper with an out-neighbour $p$ that is a low swapper in $D$. Then $0, 1 \in L(v)$. If $p$ is the only out-neighbour, $L(v)$ has a third colour, $2$ or $3$, and $p$ blocks neither. If there is a second out-neighbour $p'$, then $L(v) = \{0,1,2,3\}$ and $p'$ blocks at most one of $2, 3$ (it is parked on one colour, or fixed on $2$, or neither). (c) $v$ a high mover. Then $2, 3 \in L(v)$: with two out-neighbours $L(v) = \{0,1,2,3\}$; with one, the greedy colour is the least or second least colour of the list and $v$ has the colour $2$ at one end, so $L(v)$ is $\{0,2,3\}$ or $\{1,2,3\}$. No out-neighbour is fixed on $2$, since $v$ has the colour $2$ at one end. So the rule can fail only if $v$ has two out-neighbours parked on different colours. But a high mover with two out-neighbours has a fixed out-neighbour: at the end where $v$ has the colour $2$, its out-neighbours have the colours $0$ and $1$, and by Lemma~\ref{lem:sides}: for $v = (0,2) \in S_1$, the out-neighbour on $1$ in $\beta$ has the pattern $(1,1)$; for $v = (2,1) \in S_1$, the out-neighbour on $0$ in $\alpha$ has $(0,0)$; for $v = (2,0) \in S_0$, the out-neighbour on $1$ in $\alpha$ has $(1,1)$; for $v = (1,2) \in S_0$, the out-neighbour on $0$ in $\beta$ has $(0,0)$. So at most one out-neighbour is parked and the other parking colour is free. The choice of $r$ gives (p1), and (p3) for an out-neighbour fixed on $2$; (p4) holds by (b). \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:BIPinst}] Take the roles of Proposition~\ref{prop:roles} and put $\mu(v) = 3$ on $P_3$, $\mu(v) = 2$ on $P_2$, $\mu(v) = \alpha(v) = \beta(v)$ on a fixed vertex of $D$, and $\mu(v) = 1$ on a low swapper of $D$. By (p2) every vertex gets a value. $\mu(v) \in L(v)$ by (p1), and because $1$ is an end colour of a low swapper. $\mu \ge \alpha, \beta$ by Lemma~\ref{lem:s1} (for the sink: $\mu(s) = x$ by the rule, and $x \ge \alpha(s), \beta(s)$). $\mu$ is proper: two parked neighbours have $3$ and $2$ by (p1); a vertex of $D$ has a value in $\{0,1,2\}$, so it differs from a neighbour in $P_3$, and from a neighbour in $P_2$ by (p3) (a low swapper has $1$); two low swappers of $D$ are not adjacent (p4); a vertex fixed on $1$ is not adjacent to a low swapper (the swapper is on $1$ in $\alpha$ or in $\beta$); two fixed neighbours differ because $\alpha$ is proper. Lemma~\ref{lem:um} finishes the proof. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:bip}] Apply Lemma~\ref{lem:sweep} with $\mathcal{C}$ = the bipartite instances without root vertices and $C = 2$. A cone of a bipartite graph induces a bipartite graph, and every vertex of a cone other than its sink has an out-neighbour inside the cone; Theorem \ref{thm:BIPinst} holds for all lists of the sizes required, not only tight ones. \end{proof} \begin{corollary}[every cone bipartite]\label{cor:par} If $G$ has an order with at most $2$ later neighbours per vertex in which every cone induces a bipartite graph, then $\diam \Rk{4}(G) \le 4 \sum_t |C_t| \le 2n(n+1)$. Such graphs need not be bipartite (the $5$-cycle with a suitable order), but they are triangle-free. \end{corollary} \begin{proof} The proof of Theorem~\ref{thm:bip} uses bipartiteness only for the cones. \end{proof} \begin{remark}\label{rem:eight} Two moves per vertex do not suffice for every instance with odd cycles. Take the sink $v_0$ with the list $\{0,1\}$ and $x = 1$, and $v_1, \dots, v_7$ in this creation order, with out-neighbours and lists $v_1, v_2$: $\{v_0\}$, $\{0,1,2\}$; $v_3$: $\{v_1\}$, $\{0,2,3\}$; $v_4$: $\{v_2, v_3\}$, $\{0,1,2,3\}$; $v_5$: $\{v_4\}$, $\{0,1,2\}$; $v_6$: $\{v_4\}$, $\{1,2,3\}$; $v_7$: $\{v_3, v_6\}$, $\{0,1,2,3\}$. Its graph is a $5$-cycle $v_0 v_1 v_3 v_4 v_2$ and a $4$-cycle $v_3 v_4 v_6 v_7$ sharing the edge $v_3 v_4$, plus the pendant vertex $v_5$. On $(v_0, \dots, v_7)$, $\alpha = (1,0,0,2,1,0,2,0)$ and $\beta = (0,1,1,0,2,0,1,2)$. The sequence $v_2{\to}2$, $v_3{\to}3$, $v_1{\to}2$, $v_0{\to}0$, $v_1{\to}1$, $v_3{\to}2$, $v_4{\to}3$, $v_2{\to}1$, $v_6{\to}1$, $v_7{\to}3$, $v_3{\to}0$, $v_4{\to}2$, $v_7{\to}2$ moves $v_3$ three times and every other vertex at most twice. No sequence moves every vertex at most twice. Indeed, when $v_0$ moves, $v_1$ and $v_2$ are on $2$; so $v_2$ goes $0, 2, 1$, and $v_3$ goes $2, 3, 0$, reaching $3$ before $v_1$ reaches $2$. The vertex $v_4$ must leave $1$ before $v_2$ returns to $1$ and can enter $2$ only after $v_2$ has left $2$, so it does not go to $2$ directly; through $0$ it would cost $v_5$ three moves; so $v_4$ goes $1, 3, 2$, entering $3$ after $v_3$ has left $3$ (if $v_4$ were on $3$ while $v_3$ is still on $2$, neither could move on). Hence $v_3$ enters $0$ while $v_4$ is on $1$, and $v_7$ must leave $0$ before that. The vertex $v_6$ cannot use $3$ (it would have to leave $3$ before $v_4$ enters $3$, for $1$, which $v_4$ holds until then, or enter $3$ after $v_4$ has left it, but $v_4$ enters $2$ only after $v_6$ has left $2$), so $v_6$ stays on $2$ until it moves to $1$; it cannot be on $1$ while $v_4$ is on $1$, so it is still on $2$ when $v_7$ leaves $0$, and $v_7$ cannot go to $2$ directly. The vertex $v_7$ cannot use $3$ either: before $v_3$ enters $3$ it would block $v_3$, which can leave $2$ only for $3$, while $v_7$ could leave $3$ only for its end colour $2$; afterwards $v_3$ holds $3$ until it enters $0$. So $v_7$ goes to $1$, and then needs $v_6$ off $2$, while $v_6$ needs $v_7$ off $1$. An exhaustive computer search over the instances without root vertices and with $g = 0$ shows that every triangle-free one with at most $7$ vertices has a sequence with at most two moves per vertex, and that among all $800{,}613{,}300$ triangle-free ones with $8$ vertices and an odd cycle exactly $21$ (counted with their creation orders and lists) have none, all on this graph. (For $g \ne 0$, the sequence built in the proof of Lemma~\ref{lem:caseA} moves every vertex at most twice, see the start of Section~\ref{sec:bip}; and an instance with root vertices reduces to a smaller one without them: delete the vertices that do not reach the sink, which have $\alpha = \beta$, and remove their colours from the lists of their in-neighbours.) So $8$ is the least number of vertices of a triangle-free instance with this property. \end{remark} \section{Lower bounds: proof of Proposition~\ref{prop:lb}}\label{sec:lb} For $n \ge 2$ let $H_n$ be the path $v_1 v_2 \cdots v_n$ together with a vertex $a$ joined to every $v_i$ with $i$ odd and a vertex $b$ joined to every $v_i$ with $i$ even. It has $N = n+2$ vertices and $2n-1$ edges; it is bipartite (sides: odd path vertices and $b$; even path vertices and $a$), planar, and $2$-degenerate (in the order $v_1, \dots, v_n, a, b$ every vertex has at most two later neighbours). \begin{proposition}\label{prop:LBH} For $n \ge 5$ there are proper $4$-colourings $\sigma^+, \sigma^-$ of $H_n$ at distance at least $(n-5)^2/8$ in $\Rk{4}(H_n)$. Hence $\diam \Rk{4}(H_n) \ge (N-7)^2/8$. \end{proposition} Let $\sigma^\pm$ give $a$ and $b$ the colour $3$ and $v_i$ the colour $\pm(i-1) \bmod 3$. Call a colouring \emph{plain} if $a$ and $b$ have the colour $3$. In a plain colouring the path uses only $0,1,2$, and the \emph{step} $s_i = \mathrm{col}(v_{i+1}) - \mathrm{col}(v_i) \in \{+1,-1\} \pmod 3$ is defined for $1 \le i < n$. A \emph{height function} is $h \colon \{1,\dots,n\} \to \mathbb{Z}$ with $h(i+1) - h(i) = s_i$; the \emph{winding} is $W = h(n) - h(1)$. For $\sigma^+$ all steps are $+1$; for $\sigma^-$ all are $-1$. \begin{lemma}\label{lem:h1} A move of a path vertex between two plain colourings changes one value of a height function by exactly $2$ (and keeps the others). \end{lemma} \begin{proof} Let $v_i$ move, $1 < i < n$. Its new colour is not $3$ (it is joined to $a$ or $b$), not its old colour, and not the colours of $v_{i-1}, v_{i+1}$, which also differ from its old colour. As only $0,1,2$ are available, $v_{i-1}$ and $v_{i+1}$ have the same colour, so $s_{i-1} = -s_i$, and after the move both steps change sign: changing $h(i)$ by $2$ gives a height function of the new colouring. For $i = 1$ (or $n$) only $s_1$ (or $s_{n-1}$) changes sign, and $h(1)$ (or $h(n)$) changes by $2$. \end{proof} \begin{lemma}[flat]\label{lem:flat} If a plain colouring allows a move of $a$ or of $b$, then $|W| \le 4$. \end{lemma} \begin{proof} Let $a$ move to $c \in \{0,1,2\}$; then no odd path vertex has the colour $c$, and every odd path vertex has one of the other two colours $c_1, c_2$. For consecutive odd vertices $u = v_{2k-1}$, $u' = v_{2k+1}$ with middle vertex $e = v_{2k}$: if $u, u'$ have the same colour, $h(2k+1) = h(2k-1)$. Otherwise $e$ has neither $c_1$ nor $c_2$ nor $3$, so it has $c$; then $u, e, u'$ have three different colours, both steps have the same sign, and $h(2k+1) - h(2k-1) = \pm 2$, where the sign is $+$ exactly when $c = \mathrm{col}(u) + 1$. Since $c = c_1 + 1$ if and only if $c = c_2 - 1$, every change from $c_1$ to $c_2$ along the odd vertices moves the height by $2\eta$ and every change from $c_2$ to $c_1$ by $-2\eta$, for one sign $\eta$; these changes alternate. So the heights of the odd vertices take at most two values, $2$ apart; the first odd vertex is $v_1$ and the last is $v_n$ or $v_{n-1}$, so $|h(n) - h(1)| \le 2 + 1 = 3$. For $b$ the same argument works on the even vertices (when two consecutive even vertices differ, the odd vertex between them has $c$, because it is joined to $a$, which is on $3$); the first even vertex is $v_2$ and the last is $v_n$ or $v_{n-1}$, so $|h(n) - h(1)| \le 1 + 2 + 1 = 4$. \end{proof} \begin{lemma}[cost]\label{lem:cost} Let $h, g$ be functions on $\{1,\dots,n\}$ with all steps $\pm 1$, all steps of $h$ equal to $+1$, and let $q \ge 0$ with $g(n) - g(1) \le (n-1) - q$ and $q/4 + 1 \le n/2$. Then $\sum_i |g(i) - h(i)| \ge q^2/8$. The same holds with all signs reversed (all steps of $h$ equal to $-1$ and $g(n) - g(1) \ge -(n-1) + q$). \end{lemma} \begin{proof} Put $f = g - h$. The steps of $f$ are $0$ or $-2$, so $f$ does not increase, and $f(1) - f(n) \ge q$. For $i \le n/2$ and $i' = n+1-i$, $f$ falls by at most $2(i-1)$ on $[1,i]$ and on $[i',n]$, so $|f(i)| + |f(i')| \ge f(i) - f(i') \ge q - 4(i-1)$. With $m = \lfloor q/4 \rfloor$, the indices $i = 1, \dots, m+1$ are at most $n/2$, the pairs $\{i, i'\}$ are disjoint, and $\sum_i |f(i)| \ge \sum_{j=0}^{m} (q - 4j) = (m+1)(q - 2m) \ge (q/4)(q/2) = q^2/8$. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:LBH}] Consider a sequence from $\sigma^+$ to $\sigma^-$. If $a$ and $b$ never move, all colourings are plain and Lemma~\ref{lem:h1} carries one height function $h_t$ along the sequence, from $h_0(i) = i - 1$ to a function $g$ with all steps $-1$. Then $f = g - h_0$ satisfies $|f(i)| + |f(n+1-i)| \ge 2(n+1-2i)$, so $\sum |f| \ge 2\lfloor n/2\rfloor (n - \lfloor n/2\rfloor) \ge (n^2-1)/2$; each move changes $\sum_i |h_t(i) - h_0(i)|$ by at most $2$, so there are at least $(n^2-1)/4 > (n-5)^2/8$ moves. Otherwise let $T_1 < T_2$ be the first and the last move of $a$ or $b$ (whichever leaves $3$ must return). Before $T_1$ all colourings are plain, and the colouring just before $T_1$ allows a move of $a$ or $b$, so by Lemma~\ref{lem:flat} its winding is at most $4 = (n-1) - q$ with $q = n - 5$, and by Lemma~\ref{lem:cost} its height function is at $\ell^1$-distance at least $(n-5)^2/8$ from $h_0$: at least $(n-5)^2/16$ moves precede $T_1$. After $T_2$ neither $a$ nor $b$ moves and $\sigma^-$ is plain, so all colourings after $T_2$ are plain, and the colouring just after $T_2$ allows the move $T_2$ to be undone; reading the tail backwards from $\sigma^-$ and using the mirror form of Lemma~\ref{lem:cost}, at least $(n-5)^2/16$ moves follow $T_2$. \end{proof} \begin{remark} The exact distance from $\sigma^+$ to $\sigma^-$ for $n = 10, \dots, 13$ (breadth-first search) is $24$, $28$, $33$, $39$, about $n^2/4$. We have not found $H_n$ in the literature. It is the bipartite analogue of the fan of \cite[Thm.~12]{BJLPP14}, obtained by splitting its apex into two vertices. Height functions of $3$-colourings are used for recolouring in Cereceda's thesis \cite[Ch.~3 and Thm.~4.13]{Cer07}, following Goldberg, Martin and Paterson \cite{GMP04}; see also the weighted sum in \cite[Thm.~11]{BJLPP14}. Since $H_n$ is planar of girth $4$ and $\Rk{4}(H_n)$ is connected (Theorem~\ref{thm:bip}), it gives a quadratic lower bound for four colours on planar graphs of girth $4$ (compare the remark after Table~1 of \cite{BBFHMP21} on lower bounds for planar graphs of given girth). \end{remark} \subsection*{Girth six} For $n \ge 5$ let $J_n$ be the path $v_1 \cdots v_n$ together with four vertices $a_0, \dots, a_3$, where $a_j$ is joined to every $v_i$ with $i \equiv j \pmod 4$. It has $N = n + 4$ vertices; it is $2$-degenerate (order $v_1, \dots, v_n, a_0, \dots, a_3$), bipartite (odd path vertices with $a_0, a_2$ against even path vertices with $a_1, a_3$), and has girth $6$. Indeed the path is acyclic and the $a_j$ are pairwise non-adjacent; a cycle through exactly one $a_j$ is $a_j v_i \cdots v_l a_j$ with $l - i \ge 4$, of length at least $6$; on a cycle through two or more $a_j$, between two consecutive ones lies a path segment of at least two vertices (each path vertex is joined to exactly one $a_j$), and there are at least two such segments, so the length is at least $2+2+2 = 6$. Finally $a_1 v_1 v_2 v_3 v_4 v_5$ is a $6$-cycle. So $J_n$ is a $2$-degenerate graph of girth $6$. \begin{proposition}\label{prop:LB6} For $n \ge 29$, $\sigma^+$ and $\sigma^-$ (all $a_j$ on $3$, $v_i$ on $\pm(i-1) \bmod 3$) are at distance at least $(n-29)^2/128$ in $\Rk{4}(J_n)$. Hence $\diam \Rk{4}(J_n) \ge (N-33)^2/128$. \end{proposition} Plain colourings (all $a_j$ on $3$), steps, heights and windings are as before, and Lemma~\ref{lem:h1} holds with the same proof. \begin{lemma}\label{lem:flat4} If a plain colouring of $J_n$ allows a move of some $a_j$, then $|W| \le 3(n-1)/4 + 7$. \end{lemma} \begin{proof} We may assume $n \ge 4$. The move goes to a colour $c \in \{0,1,2\}$ that no neighbour of $a_j$ has; write the other two colours as $c_1$ and $c_2 = c_1 + 1 \pmod 3$. Let $p_0 < \dots < p_r$ be the indices of the path neighbours of $a_j$ ($p_{k+1} = p_k + 4$, $p_0 \le 4$, $p_r \ge n-3$), and $D_k = h(p_{k+1}) - h(p_k)$, a sum of four steps, so $D_k \in \{-4,-2,0,2,4\}$ and $D_k \equiv \mathrm{col}(v_{p_{k+1}}) - \mathrm{col}(v_{p_k}) \pmod 3$. If the two colours agree, $D_k = 0$; a change from $c_1$ to $c_2$ gives $D_k \in \{4, -2\}$, and from $c_2$ to $c_1$ gives $D_k \in \{2, -4\}$. The two kinds of change alternate, so at most $\lceil r/2 \rceil$ of the $D_k$ equal $4$ and the others are at most $2$: $\sum_k D_k \le 3r + 1$, and symmetrically $\ge -(3r+1)$. At most three steps lie before $p_0$ and after $p_r$ each, and $r \le (n-1)/4$, so $|W| \le (3r + 1) + 6 \le 3(n-1)/4 + 7$. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:LB6}] Put $q = (n-29)/4 \ge 0$, so $3(n-1)/4 + 7 = (n-1) - q$ and $q/4 + 1 \le n/2$. If no $a_j$ moves, at least $(n^2-1)/4$ moves are needed as for $H_n$. Otherwise let $T_1 < T_2$ be the first and last moves among ALL apex vertices $a_j$ (possibly different apices; any apex that leaves $3$ must return). Before $T_1$ and after $T_2$ all colourings are plain and only path vertices move; Lemmas~\ref{lem:flat4} and~\ref{lem:cost} give at least $q^2/16$ moves before $T_1$ and, reading backwards from $\sigma^-$, at least $q^2/16$ after $T_2$; these are different moves, so in total at least $q^2/8 = (n-29)^2/128$. \end{proof} \begin{remark} With $k \ge 4$ apex vertices and $n \ge k+2$ path vertices, the analogous graph has girth $6$ ($a_j v_i v_{i+1} a_{j+1} v_{i+k+1} v_{i+k}$), so this family gives nothing for girth $7$ or more. Whether the diameter can be of order $n^2$ for $2$-degenerate graphs of girth $\ge 7$ is open; for planar graphs of girth $\ge 7$ it cannot (their maximum average degree is below $14/5$, and \cite{Feg21} gives $n$ times a power of $\log n$). The exact distances from $\sigma^+$ to $\sigma^-$ in $\Rk{4}(J_n)$ for $n = 5, \dots, 18$ (breadth-first search) are $6, 8, 8, 9, 11, 14, 15, 17, 20, 24, 26, 29, 31, 33$, all below $2n$: the square appears late. \end{remark} \section{Cone width two}\label{sec:cw} \subsection*{Cones and their width} An \emph{orientation} of $G$, in this section, is an acyclic orientation in which every vertex has at most $2$ out-neighbours. An order as in Section~\ref{sec:setting} gives one (every edge points to its later end), and every orientation comes from such an order (any order that puts every vertex before its out-neighbours). The \emph{cone} of a vertex $x$ is $x$ together with all vertices from which $x$ can be reached along arcs; $x$ is its \emph{top}. A vertex that reaches $v_t$ along arcs is earlier than $v_t$, and so is every vertex on the way; hence the set $C_t$ of Section~\ref{sec:setting} is the cone of $v_t$, and it depends only on the orientation. The instance on $C_t$ (Lemma~\ref{lem:sweep}) gives each $v \in C_t$ its out-neighbours inside $C_t$; so its graph is the graph that $G$ induces on the cone (an edge of $G$ between two vertices of the cone is an arc from the earlier to the later one), its sink is the top, and it has no root vertex. An \emph{elimination order} of a graph $H$ is an order of $V(H)$ in which the vertices are removed. The \emph{neighbours left} of $v$ are the vertices $u$ after $v$ that are adjacent to $v$ or joined to $v$ by a path all of whose inner vertices are before $v$. The \emph{width} of the order is the largest number of neighbours left. An elimination order of a graph with arcs (a cone, or the graph of an instance) is \emph{topological} if every vertex comes before its out-neighbours (so the top of a cone comes last). The \emph{width of a cone} is the least width of a topological elimination order of the graph it induces, and the \emph{cone width} $\mathrm{cw}(G)$ is the least, over all orientations of $G$, of the largest width of a cone. A graph with no orientation (one that is not $2$-degenerate) has no cone width and lies in no class $\{\mathrm{cw} \le w\}$. \begin{lemma}[two forms]\label{lem:forms} Remove the vertices of $H$ in an elimination order, and when a vertex is removed join its neighbours of that time to each other. Then the neighbours of $v$ at the time of its removal are exactly its neighbours left. \end{lemma} \begin{proof} We show by induction on the number of removed vertices that two vertices $u, w$ that are left are adjacent at that time exactly when they are adjacent in $H$ or joined in $H$ by a path whose inner vertices are all removed. Let $z$ be removed next. A new edge $uw$ appears exactly when $u$ and $w$ are neighbours of $z$; by induction each is joined to $z$ through removed vertices, and the two together give a walk from $u$ to $w$ through removed vertices and $z$, which contains such a path. Conversely, on such a path from $u$ to $w$ let $y$ be the inner vertex removed last; just before $y$ is removed, the two parts of the path make $u$ and $w$ neighbours of $y$ (induction), so $uw$ is an edge after $y$ goes, and edges stay. (See also~\cite{RTL76}.) \end{proof} \begin{theorem}[Theorem~\ref{thm:conewidth}; computer-assisted]\label{thm:cwmain} If $\mathrm{cw}(G) \le 2$, then $\diam \Rk{4}(G) \le 8 \sum_t |C_t| \le 4n(n+1)$, where $C_t$ are the cones of any orientation in which every cone has width at most $2$. \end{theorem} \begin{proposition}\label{prop:cwinside} (a) Every graph of treewidth at most $2$ has cone width at most $2$. More generally, a graph with an elimination order of width at most $w$ in which every vertex has at most $2$ later neighbours has cone width at most $w$. (b) If $\mathrm{cw}(G) \le w$, then $\mathrm{cw}(G') \le w$ for every subgraph $G'$ of $G$. (c) The class $\{\mathrm{cw} \le 2\}$ contains graphs of treewidth $3$ with triangles, for instance the graph $W$ below. \end{proposition} \begin{proof} (a) Take an elimination order of width at most $w$ and orient every edge to its later end. The out-neighbours of a vertex are its later neighbours: at most $2$ by hypothesis (for $w = 2$ this follows, since they are among its at most $2$ neighbours left). The order cut down to a cone is topological, and cutting an elimination order down to an induced subgraph does not raise its width (a path through removed vertices of the subgraph is such a path of $G$). A graph of treewidth at most $2$ has an elimination order of width at most $2$ (see~\cite{Bod98}). (b) Keep the orientation. The new cone of $x$ lies inside the old one; cut a topological elimination order of the old cone down to the new cone. Every vertex still comes before its out-neighbours, and its neighbours left are among the old ones (a path of $G'$ through removed vertices of the new cone is such a path of the old cone). (c) Let $W$ have the vertices $a, b, c, d$, a vertex $xy$ for each pair $\{x, y\} \subseteq \{a,b,c,d\}$ joined to $x$ and $y$ ($K_4$ with every edge subdivided once), a triangle $t_0 t_1 t_2$ with three further vertices $s_{01}, s_{12}, s_{20}$, where $s_{ij}$ is joined to $t_i$ and $t_j$, and one more edge $a t_0$ ($16$ vertices, $22$ edges). Its treewidth is at least $3$ (it contains a subdivided $K_4$) and at most $3$ (the order below, as an elimination order of all of $W$, has width $3$). Take the order $ab, ac, ad, bc, bd, cd, b, c, d, a, s_{01}, s_{12}, s_{20}, t_0, t_1, t_2$ and orient every edge to its later end; every vertex has at most $2$ later neighbours. The cones of $ab, \dots, cd, s_{01}, s_{12}, s_{20}$ are single vertices; the cones of $b$, $c$, $d$ and $a$ are stars (width $1$). The cone of $t_2$ consists of $ab, ac, ad, a, s_{01}, s_{12}, s_{20}, t_0, t_1, t_2$; in this order it is topological, and the neighbours left are $\{a\}$ for $ab, ac, ad$; $\{t_0\}$ for $a$; $\{t_0, t_1\}$, $\{t_1, t_2\}$, $\{t_2, t_0\}$ for $s_{01}, s_{12}, s_{20}$; $\{t_1, t_2\}$ for $t_0$; $\{t_2\}$ for $t_1$: width $2$. The cones of $t_0$ and $t_1$ lie inside it; cut this order down to them, as in (a). \end{proof} \begin{remark} Treewidth is not bounded by cone width: for $m \ge 2$, $K_m$ with every edge subdivided once has cone width $1$ (orient each subdividing vertex to its two ends; every cone is a star) and treewidth $m - 1$. Among the graphs on at most $8$ vertices that have an orientation, a census by two independent programs found $21$ minimal graphs outside $\{\mathrm{cw} \le 2\}$ ($1, 1, 3, 16$ with $5, 6, 7, 8$ vertices), all of cone width $3$ and treewidth $3$; the smallest is $K_4$ with one edge subdivided ($5$ vertices). For treewidth at most $2$ the bound $2n(n+1)$ with four colours is known~\cite[Thm.~2]{BB18}; Theorem~\ref{thm:cwmain} has the weaker constant $4$ but covers the larger class. Neither of the classes of Theorems~\ref{thm:bip} and~\ref{thm:conewidth} contains the other: the triangle has cone width $2$ and is not bipartite, and the bipartite $2$-degenerate graph on $\{0, \dots, 7\}$ with the edges $04, 05, 07, 14, 16, 25, 26, 27, 35, 36, 37$ has cone width $3$ (by two independent programs: none of its $130$ orientations has all cones of width at most $2$). \end{remark} \subsection*{Instances of bag width two} Let $I$ be an instance (Section~\ref{sec:setting}). Sets of at most two vertices become \emph{available} as vertices are created: at the start $\{s\}$ is available; every vertex $v \ne s$ is \emph{created on} an available set $B(v) \supseteq N^+(v)$; if $B(v) = \{s\}$ the set $\{s, v\}$ becomes available, and if $B(v) = \{a, b\}$ the sets $\{a, v\}$ and $\{b, v\}$ become available; available sets stay available. The instance is in $\mathrm{BAG}(2)$ if some creation order (the sink first, every vertex after its out-neighbours) admits such sets $B(v)$. Membership does not depend on the lists. \begin{lemma}\label{lem:C} Let $s = u_0, u_1, \dots, u_{m-1}$ be an order of the vertices of an instance in which every vertex comes after its out-neighbours. Sets $B(v)$ as above exist for this order if and only if the reversed order $u_{m-1}, \dots, u_1, s$ is an elimination order of width at most $2$ of the graph of the instance. Hence an instance lies in $\mathrm{BAG}(2)$ if and only if its graph has a topological elimination order of width at most $2$ that ends with the sink; and the instance on $C_t$ lies in $\mathrm{BAG}(2)$ if and only if the cone $C_t$ has width at most $2$. \end{lemma} \begin{proof} ``Before'' and ``after'' refer to the creation order; the elimination removes $u_{m-1}$ first, so the neighbours left of $v$ are the vertices $u$ before $v$ that are adjacent to $v$ or joined to $v$ by a path whose inner vertices are all after $v$. Sets to width. Let $T$ be the graph whose edges are the available sets of size $2$. Such a set is made when its later vertex $v$ is created, with its other vertex in $B(v)$, and every $\{c, v\}$, $c \in B(v)$, is made. So the neighbours of $v$ in $T$ before $v$ are exactly the vertices of $B(v)$: at most two, and adjacent in $T$ if there are two. Removing the vertices of $T$ last-created first, the graph at the time $v$ is removed is $T$ on $v$ and the vertices before it (induction: the neighbours of $v$ at that time form $B(v)$, which is already a clique, so no edge is added). By Lemma~\ref{lem:forms} every vertex has at most $2$ neighbours left in $T$. The graph of the instance is a subgraph of $T$ on the same vertices ($N^+(v) \subseteq B(v)$), and a path in it is a path in $T$. Width to sets. Let $H(v)$ be the set of neighbours left of $v$; $|H(v)| \le 2$ and $N^+(v) \subseteq H(v)$. If $a, b \in H(v)$ with $a$ before $b$, then $a \in H(b)$ (go from $b$ to $v$ and on to $a$: a walk whose inner vertices are $v$ or after $v$, hence after $b$; a path from $b$ to $a$ lies inside it). Create the vertices in order, maintaining: (i) after $v$ is created, $\{a, v\}$ is available for every $a \in H(v)$. If $H(v) = \{a, b\}$ with $a$ before $b$: $a \in H(b)$, so $\{a, b\}$ is available by (i) for $b$; take $B(v) = \{a,b\}$. If $H(v) = \{a\}$: take $B(v) = \{s\}$ if $a = s$, and otherwise any available set $\{c, a\}$ of size $2$ (one was made when $a$ was created). If $H(v) = \emptyset$, then $v$ is a root vertex; take $B(v) = \{s\}$. In each case $N^+(v) \subseteq H(v) \subseteq B(v)$ and (i) holds for $v$. The last sentence: as said at the start of this section, the instance on $C_t$ has the graph that $G$ induces on the cone, with the top as sink; a topological elimination order of the cone, read backwards, has the sink first and every vertex after its out-neighbours, so it is a creation order, and conversely. \end{proof} \subsection*{Four moves per vertex} Fix $C \ge 1$. A \emph{word} is a sequence over $\{0,1,2,3\}$ with between $1$ and $C + 1$ letters and consecutive letters different (the colours a vertex takes in a sequence; at most $C$ moves). A \emph{state} of two vertices (places $0$ and $1$) consists of their two words and an \emph{order}: a word over $\{0, 1\}$, with as many digits $p$ as place $p$ has moves, saying which place makes the first, second, \dots{} of their moves (the state with the places exchanged is the same state). A \emph{new vertex} for a state is a set $P \subseteq \{0,1\}$ of places (its out-neighbours) with a tight list $L$: $|L| = 2$ if $P = \emptyset$, $|L| = 3$ if $|P| = 1$, and $L = \{0,1,2,3\}$ if $P = \{0,1\}$ ($6 + 4 + 4 + 1 = 15$ new vertices; for the sink alone $P \subseteq \{0\}$, $10$ new vertices). An \emph{answer} to a new vertex $(P, L)$ is a word $w = w_0 \dots w_m$ and a word $\pi$ over $\{0,1,2\}$ with (A1) $m \le C$, all $w_i \in L$, $w_i \ne w_{i+1}$; (A2) $w_0$ is the least colour of $L$ that no place of $P$ has as its first colour, and $w_m$ the least colour of $L$ that no place of $P$ has as its last colour; (A3) $\pi$ has exactly $m$ digits $2$, and deleting them leaves the order of the state; (A4) replaying $\pi$ from the first colours, at a digit $2$ the new vertex takes its next colour, which no place of $P$ has at that time, and at a digit $p \in P$ the place $p$ takes its next colour, which the new vertex does not have at that time. The two \emph{states after} the answer are (place $1$, new vertex) and (place $0$, new vertex), with $\pi$ restricted to the corresponding digits (renamed $0, 1$ in the order old place, new vertex). For the sink alone (place $0$, its order $0\dots0$), an answer is defined in the same way with $\pi$ over $\{0, 2\}$, and there is one state after it, (sink, new vertex). A \emph{closed core} for $C$ is a set $K$ of states and, for each $x \in \{1,2,3\}$, a word of the sink from $x$ to $0$ inside $\{0, x\}$ with at most $C$ moves, such that (K1) every state of $K$ has, for each of its $15$ new vertices, an answer whose two states after it lie in $K$, and (K2) the word of the sink has, for each of its $10$ new vertices, an answer whose state after it lies in $K$. \begin{proposition}\label{prop:core} If a closed core for $C$ exists, then every instance of $\mathrm{BAG}(2)$ in the main case with tight lists has a sequence from $\alpha$ to $\beta$ in which every vertex moves at most $C$ times. \end{proposition} \begin{proof} Go through the instance in a creation order with sets $B(v)$, keeping one list $M$ of moves with: (I1) the moves of $M$, applied from $\alpha$ on the vertices created so far, form a sequence that ends in $\beta$ there, with at most $C$ moves per vertex; (I2) for every available set $\{a, b\}$ the words of $a$ and $b$ in $M$ and the order of their moves in $M$ form a state of $K$, and the word of $s$ in $M$ is the core's word for $x$. Start with the moves of the sink's word. When $v$ is created on $B(v) = \{a, b\}$, $v$ is a new vertex of the state of $B(v)$ ($P$ the places of its out-neighbours, $L$ its tight list); take the answer given by (K1), give $v$ the word $w$, and insert the moves of $v$ into $M$ as $\pi$ prescribes (between the same two moves of $a$ and $b$; the order of the old moves is unchanged). By (A2) and greediness, $w$ runs from $\alpha(v)$ to $\beta(v)$; by (A1) it stays in the list with at most $C$ moves. Neighbours are a vertex and one of its out-neighbours; pairs of old vertices are unaffected, and for $u \in N^+(v) \subseteq B(v)$ condition (A4) is exactly properness on the edge $uv$ (initially $\alpha(v) \ne \alpha(u)$ by greediness), because the colours of $u$ and $v$ at any time depend only on the moves of $a, b, v$, whose order in $M$ is $\pi$. The old available sets keep their words and orders (the order of the old moves is unchanged), and the new ones $\{a, v\}, \{b, v\}$ carry the two states after the answer, which lie in $K$. Creation on $\{s\}$ is the same with (K2). At the end $M$ is the required sequence. \end{proof} \begin{lemma}[tight lists suffice]\label{lem:tight} Let a class of instances be such that membership does not depend on the lists. If every instance of the class in the main case with tight lists has a sequence with at most $C$ moves per vertex, so does every instance of the class in the main case. \end{lemma} \begin{proof} Replace each list $L(v)$ by a tight list $L'(v) \subseteq L(v)$ containing $\alpha(v)$ and $\beta(v)$ (the sink's list $\{0, x\}$ is unchanged; for a vertex with two out-neighbours $L(v)$ is already tight; a root vertex has $\alpha(v) = \beta(v)$). By induction along the creation order, greedy colouring with the lists $L'$ gives the same $\alpha$ and $\beta$: a smaller colour of $L'(v)$ free of the out-neighbours' colours would be a smaller such colour of $L(v)$. A sequence for the lists $L'$ stays inside the lists $L$. \end{proof} \begin{theorem}[computer-assisted]\label{thm:MV4} A closed core for $C = 4$ exists. Hence, by Proposition~\ref{prop:core} and Lemma~\ref{lem:tight}, every instance of $\mathrm{BAG}(2)$ in the main case has a sequence from $\alpha$ to $\beta$ in which every vertex moves at most $4$ times; in particular $\mathrm{dist}(\alpha,\beta) \le 4t$. The $4$ cannot be lowered: there are instances of $\mathrm{BAG}(2)$ without root vertices, with $27$, $30$ and $33$ vertices ($x = 1, 2, 3$), with no sequence of at most $3$ moves per vertex; so no closed core exists for $C \le 3$ (a closed core for $C \le 3$ is one for $C = 3$, and it would give such a sequence). \end{theorem} \begin{proof}[About the computation] For $C = 4$ there are $484$ words and $5{,}261{,}242$ states. Starting from all states and deleting, until nothing changes, every state that has a new vertex without an answer whose states after it remain, leaves $3{,}258{,}618$ states; the sink word $x0$ satisfies (K2) for $x = 1, 2, 3$. From it a closed core of $17{,}485$ states with $262{,}275$ answers for (K1) ($15$ per state), together with the three sink words and their $30$ answers for (K2) (ten for each), was extracted and written to a file. The file is verified line by line, without search, against (A1)--(A4), (K1), (K2) by three checkers: one written with the search (separate code), and two written independently from the definitions alone. The two independent checkers reject damaged copies and reject the file when the cap is $3$; the first rejects a copy with one changed answer. One independent run also built, for every state of the core, a real instance using all $15$ new vertices and replayed it, so every answer line was used. An end-to-end test built the sequences of Proposition~\ref{prop:core} from the file for $1{,}800$ random instances of $\mathrm{BAG}(2)$ (up to $600$ vertices) and replayed them move by move: all were legal, ended in $\beta$, and moved every vertex at most $4$ times. For $C \le 3$, after the same deletion no word of the sink satisfies (K2), for $x = 1, 2, 3$; the three explicit instances (which prove this by themselves, as stated) were confirmed by several SAT models from two independent implementations and by a search that uses no solver. The core file \texttt{core\char`\_C4.txt} ($17{,}485$ states, $262{,}275$ answers and $30$ answers at the sink; SHA-256 beginning \texttt{24881187}), the three checkers, the three instances with $27$, $30$ and $33$ vertices, the search without a solver, the end-to-end test and a README are ancillary files of this paper. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:cwmain}] Take an order inducing an orientation in which every cone has width at most $2$. By Lemma~\ref{lem:C} every instance of the sweep lies in $\mathrm{BAG}(2)$, a class that does not depend on lists or colourings. By Theorem~\ref{thm:MV4}, every instance of $\mathrm{BAG}(2)$ in the main case has a sequence of at most $4t$ moves. Lemma~\ref{lem:sweep} with $C = 4$ gives the bound. \end{proof} \section{The terminal extension}\label{sec:terminal} In an instance, an \emph{in-neighbour} of $v$ is a vertex that has $v$ as an out-neighbour. \begin{lemma}[terminal vertices]\label{lem:terminal} Let $I$ be an instance without root vertices on $t$ vertices, and $Z$ a set of vertices other than the sink without in-neighbours such that $I - Z$ is bipartite. Then there is a sequence from $\alpha$ to $\beta$ of at most $2(t - |Z|) - 1 + \sum_{z \in Z} (2\deg_I(z) + 1) \le 5t - 1$ moves, where $\deg_I$ is the degree in $I$. \end{lemma} \begin{proof} $Z$ is independent (on an edge between two members, the later-created end is an in-neighbour of the other), and every neighbour of $z \in Z$ is an out-neighbour, so $\deg_I(z) \in \{1,2\}$ and $|L(z)| \ge \deg_I(z) + 2$. A vertex outside $Z$ has no out-neighbour in $Z$, so $I - Z$ is an instance without root vertices, with the same lists and sink, whose greedy colourings are the restrictions of $\alpha, \beta$. By Theorem~\ref{thm:BIPinst} it has a sequence in which the sink moves once and every other vertex at most twice. Follow this sequence; immediately before a vertex $p \notin Z$ moves from a colour $c_{\rm old}$ to a colour $c_{\rm new}$, every $z \in Z$ adjacent to $p$ and on $c_{\rm new}$ first moves to a colour of $L(z)$ other than $c_{\rm old}$, $c_{\rm new}$ and the colours of its other neighbours; at most $\deg_I(z) + 1$ colours are excluded, so such a colour exists, and the move stays proper after $p$ moves. Members of $Z$ are pairwise non-adjacent, so these moves do not interfere. At the end every $z$ moves to $\beta(z)$; its neighbours are then on their $\beta$ colours, so this is proper. A vertex $z$ moves at most once per move of a neighbour and once at the end: at most $2\deg_I(z) + 1 \le 5$ times. \end{proof} \begin{theorem}\label{thm:terminalclass} Let $G$ have an order with at most $2$ later neighbours per vertex such that, for every $t$, the cone $C_t$ becomes bipartite when the vertices of $C_t$ other than $v_t$ without an earlier neighbour (in the order of $G$) in $C_t$ are removed. Then $\diam \Rk{4}(G) \le 10 \sum_t |C_t| \le 5n(n+1)$. \end{theorem} \begin{proof} The creation order of the instance on $C_t$ is the reverse of the order of $G$, so the in-neighbours of a vertex in the instance are its earlier neighbours in $C_t$ in the order of $G$, and it has no root vertices (Remark~\ref{rem:noroots}); apply Lemma~\ref{lem:sweep} with $C = 5$ and Lemma~\ref{lem:terminal}. \end{proof} \begin{corollary}\label{cor:nearbip} Let $G$ be a graph on $n$ vertices with an independent set $Z$ of vertices of degree at most $2$ such that $G - Z$ is bipartite and $2$-degenerate. Then $\diam \Rk{4}(G) \le 5n(n+1)$. \end{corollary} \begin{proof} Order $Z$ first, then $G - Z$ in an order with at most $2$ later neighbours. Every $z \in Z$ has at most two later neighbours (all its neighbours), and no vertex of $G - Z$ gains a later neighbour. A cone $C_t$ with $v_t \in Z$ is $\{v_t\}$ ($Z$ is independent and comes first). For $v_t \notin Z$, the members of $Z$ in $C_t$ have no earlier neighbour, and $C_t \setminus Z$ induces a subgraph of $G - Z$, which is bipartite; removing the vertices of $C_t$ other than $v_t$ without an earlier neighbour in $C_t$ removes all of $Z \cap C_t$ (as $v_t \notin Z$) and leaves a subset of $C_t \setminus Z$. Apply Theorem~\ref{thm:terminalclass}. (A vertex of $Z$ that is isolated in $G$ is harmless: its cone is itself.) \end{proof} \begin{remark} The class of Corollary~\ref{cor:nearbip} contains graphs with triangles (a triangle: $Z$ = one vertex), so it is not contained in the class of Corollary~\ref{cor:par}, whose graphs are triangle-free; both contain the bipartite $2$-degenerate graphs ($Z = \emptyset$). For example, add to a square grid any number of new vertices of degree $2$, each joined to two grid vertices at odd distance (for instance the two ends of an edge); every new vertex closes odd cycles (a triangle in that case), and the graph satisfies the hypothesis of Theorem~\ref{thm:terminal} with $Z$ the set of new vertices. \end{remark} \section{Open problems}\label{sec:open} \begin{enumerate} \item Cereceda's conjecture for $d = 2$ remains open. By Lemma~\ref{lem:sweep} it would follow from a bound $O(t)$ on the length of a shortest sequence for every instance in the main case, in particular from a constant bound on the number of moves per vertex. \item Girth. The graphs $J_n$ show that $\diam \Rk{4}$ can be of order $n^2$ for $2$-degenerate graphs of girth $6$, and Remark~\ref{rem:eight} shows that two moves per vertex no longer suffice once odd cycles are allowed (its instance has girth $4$). Is $\diam \Rk{4}(G) = O(n^2)$ for every $2$-degenerate graph $G$ of girth at least $5$? Is there a constant $C$ such that every instance of girth at least $5$ has a sequence moving every vertex at most $C$ times? \item The constant for bipartite graphs: Theorem~\ref{thm:bip} and Proposition~\ref{prop:lb} leave a factor of $16$ between the leading terms $2n^2$ and $n^2/8$. \item Cone width: does Theorem~\ref{thm:conewidth} extend to cone width $3$, for instance by a finite certificate as in Theorem~\ref{thm:MV4}? \item Larger $d$: is $\diam \Rk{d+2}(G) = O(n^2)$ for bipartite $d$-degenerate graphs? (Bousquet and Heinrich \cite{BH22} proved this for planar bipartite graphs, which are $3$-degenerate, with $5$ colours; Cranston and Mahmoud \cite{CM24} extended it to planar graphs with no cycles of length $3$ or $5$, with at most $4n^2$ moves.) The per-vertex form of Theorem~\ref{thm:BIPinst} does not extend to $d \ge 3$ (lists in $\{0, \dots, d+1\}$, at most $d$ out-neighbours): the instance with creation order $s, v_1, \dots, v_5$, out-neighbours $N^+(v_1) = \{s\}$, $N^+(v_2) = N^+(v_3) = \{v_1\}$, $N^+(v_4) = \{v_2, v_3\}$, $N^+(v_5) = \{v_1, v_4\}$ and lists $L(s) = \{0,1\}$, $L(v_1) = \{0,1,2\}$, $L(v_2) = \{0,2,3\}$, $L(v_3) = \{1,2,4\}$, $L(v_4) = \{1,2,3,4\}$, $L(v_5) = \{0,1,2,3\}$ has no root vertex and a bipartite graph ($K_{2,3}$ with the sink as a pendant vertex), but no sequence from $\alpha$ to $\beta$ in which every vertex moves at most twice; its shortest sequences still have $11 = 2t - 1$ moves, and one of them moves every vertex at most three times. \end{enumerate} \noindent Theorem numbers for \cite{BB18}, \cite{BJLPP14}, \cite{BC09}, \cite{BP16}, \cite{Cra22} and \cite{FJP16} refer to the open versions named in the entries. \pdfbookmark[0]{Appendix, acknowledgments and references}{backmatter} \appendix \providecommand{\TC}{\ensuremath{\mathrm{TC3}}} \providecommand{\twoC}{\ensuremath{\mathrm{2C}}} \section{The closure lemma of Jaeger, Linial, Payan and Tarsi}\label{app:sec} This appendix proves the two facts on truncated cubic $3$-edge-connected graphs that Section~\ref{a2:sec} uses: every such graph has a well colouring \textup{(}Lemma~\ref{app:wellcolouring}\textup{)}, and a well colouring gives a closure sequence of cycles \textup{(}Lemmas~\ref{app:closureseq} and~\ref{app:closure}\textup{)}. Both are results of Jaeger, Linial, Payan and Tarsi~\cite{JLPT}: the generation and the colouring are \cite[Lemmas~3.2 and~3.3, pp.~175 and~177]{JLPT}, and the closure sequence is the observation at the start of the proof of their Theorem~3.2 \cite[p.~179]{JLPT}. The proofs are written out in full. This appendix contains no new result. \subsection*{Definitions and counts}\label{app:defs} \TC{} is the class of the truncated cubic $3$-edge-connected graphs of Definition~\ref{a2:tc}: the graphs $Q=Q^+-v$, where $Q^+$ is a cubic $3$-edge-connected graph on at least four vertices and $v$ is a vertex of $Q^+$. The words $2$-join, $2$-constructible, $2$-construction, stage, node and operand are those of Definition~\ref{a2:twoC}, and \twoC{} is the class of the $2$-constructible graphs. Graphs can have parallel edges \textup{(}a $2$-constructible graph often has them\textup{)}. \begin{definition}[Well colouring]\label{app:wellcoloured} Let $Q$ be a graph in which every edge is blue or red. A \emph{red component} of $Q$ is a component of the graph with the vertex set $V(Q)$ and the red edges. $Q/\red$ is the graph whose vertices are the red components of $Q$ and whose edges are the blue edges of $Q$; the ends of a blue edge in $Q/\red$ are the red components of its two ends. The colouring is a \emph{well colouring} if \begin{condlist}{(ii)} \item[\textup{(i)}] the blue edges form a spanning tree of $Q$, and \item[\textup{(ii)}] $Q/\red$ is $2$-constructible. \end{condlist} This is the notion of Definition~\ref{def:frame}~\textup{(F1)}. \end{definition} \begin{lemma}[Counts]\label{app:count} \begin{condlist}{(a)} \item[\textup{(a)}] Let $Q=Q^+-v$ be in \TC, with $n$ vertices. Then $Q^+$ and $Q$ are simple graphs, $Q$ is connected, $n$ is odd, $Q$ has $3(n-1)/2$ edges, exactly three vertices of $Q$ have degree $2$ \textup{(}the neighbours of $v$\textup{)}, and all other vertices of $Q$ have degree $3$. The smallest graph of \TC{} is the triangle \textup{(}$Q^+=K_4$\textup{)}. \item[\textup{(b)}] Let $U_0$ be a graph on $j$ vertices with a $2$-construction. Then $U_0$ is connected, has no loop, and has $2(j-1)$ edges. Two stages are disjoint, or one is a subgraph of the other. Every edge of $U_0$ is an edge of exactly one node. Every stage $U$ is an induced subgraph of $U_0$, and the edges of $U$ are the edges of the nodes that are subgraphs of $U$. \item[\textup{(c)}] In a well coloured graph no blue edge has both ends in one red component. If a graph of \TC{} with $n$ vertices is well coloured and has $j$ red components, then $2(j-1)=n-1$, and the red edges form a forest. \end{condlist} \end{lemma} \begin{proof} (a) A loop at a vertex $x$ of the cubic graph $Q^+$ makes the third edge at $x$ a bridge. Two parallel edges with the ends $x$, $y$ leave at most $2$ edges between $\{x,y\}$ and the other vertices, and there are other vertices because $Q^+$ has at least four. Both are impossible in a $3$-edge-connected graph. So $Q^+$ is simple, and $Q$ is simple. A cubic graph has an even number of vertices, so $n=|V(Q^+)|-1$ is odd, and $Q$ has $3(n+1)/2-3=3(n-1)/2$ edges. The three neighbours of $v$ are different; they lose one edge, and no other vertex does. If $Q$ had more than one component, one of them would get at most one of the three edges at $v$; this edge alone, or no edge, would separate it from the rest of $Q^+$. So $Q$ is connected. (b) Induction on the $2$-construction. For one vertex all is clear. Let $U_0$ be the $2$-join of $U'_0$ and $U''_0$, with $j'$ and $j''$ vertices. The two graphs are connected and disjoint, and a new edge joins them; so $U_0$ is connected. A new edge has its ends in two different graphs, so it is no loop. The number of edges is $2(j'-1)+2(j''-1)+2=2(j-1)$. The stages other than $U_0$ are the stages of $U'_0$ and the stages of $U''_0$; the first are subgraphs of $U'_0$ and the second are subgraphs of $U''_0$. This gives the statement on two stages, and the statements on the edges. Let $U\ne U_0$ be a stage, say a stage of $U'_0$, and let $t$ be an edge of $U_0$ with both ends in $V(U)$. Then $t$ is not a new edge and not an edge of $U''_0$. So $t$ is an edge of $U'_0$, and by induction it is an edge of $U$. (c) A blue edge with both ends in one red component is a loop of $Q/\red$; by (b) there is none. Let $Q$ be in \TC. By (b), $Q/\red$ has $2(j-1)$ edges. They are the blue edges, and the blue spanning tree has $n-1$ edges. So $2(j-1)=n-1$. By (a) there are $3(n-1)/2-(n-1)=(n-1)/2=n-j$ red edges. A graph with $n$ vertices, $j$ components and $n-j$ edges is a forest. \end{proof} \smallskip \noindent\textbf{The graphs of the stages.} Let $Q$ be a graph with a well colouring and with a $2$-construction of $Q/\red$. For a stage $U$ let $V_U$ be the union of the red components that are the vertices of $U$, and let $Q_U$ be the subgraph of $Q$ induced on $V_U$. We call $Q_U$ the \emph{graph of $U$}. For $U=Q/\red$ it is $Q$. \begin{lemma}[The graphs of the stages]\label{app:operands} Let $Q$ be a graph with a well colouring and with a $2$-construction of $Q/\red$, and let $U$ be a stage. \begin{condlist}{(a)} \item[\textup{(a)}] The edges of $Q_U$ are the red edges with both ends in $V_U$ and the edges of the nodes that are subgraphs of $U$. \item[\textup{(b)}] $Q_U$ is connected. \item[\textup{(c)}] If $U$ is a node with the operands $U'$, $U''$ and the two edges $t_1$, $t_2$, then $V_U$ is the disjoint union of $V_{U'}$ and $V_{U''}$, each of $t_1$, $t_2$ has one end in $V_{U'}$ and one end in $V_{U''}$, and $E(Q_U)=E(Q_{U'})\cup E(Q_{U''})\cup\{t_1,t_2\}$. \end{condlist} \end{lemma} \begin{proof} (a) A red edge has both ends in one red component. A blue edge with both ends in $V_U$ is an edge of $Q/\red$ whose two ends are vertices of $U$. By Lemma~\ref{app:count}~(b), $U$ is an induced subgraph of $Q/\red$, so this edge is an edge of $U$, and the edges of $U$ are the edges of the nodes that are subgraphs of $U$. (c) follows from (a) and Lemma~\ref{app:count}~(b). (b) Induction on the $2$-construction. If $U$ has one vertex, $Q_U$ is one red component with its red edges. If $U$ is a node, then $Q_{U'}$ and $Q_{U''}$ are connected, and $t_1$ joins them. \end{proof} \smallskip \noindent\textbf{Two operations.} Let $Q'$ and $Q''$ be disjoint graphs, and let $d$ be $2$ or $3$. \begin{itemize} \item[--] A \emph{$d$-sticking} of $Q'$ and $Q''$: $z$ is a vertex of $Q'$ of degree $d$ with the edges $t_1,\dots,t_d$, and $z_1,\dots,z_d$ are $d$ different vertices of degree $2$ of $Q''$. The result is the graph $(Q'-z)\cup Q''$ in which every $t_i$ keeps its end in $Q'-z$ and gets $z_i$ as its other end. \item[--] A \emph{growing} of $Q'$ at an edge $y_1y_2$ and a vertex $x'$ of degree $2$: put a new vertex $y$ on the edge $y_1y_2$ \textup{(}the edge becomes the path $y_1,y,y_2$\textup{)}, and add a new vertex $x$ with two edges, to $x'$ and to $y$. \end{itemize} \subsection*{Generation}\label{app:generation-sec} \begin{lemma}[Generation]\label{lem:A2}\label{app:generation} Every graph $Q$ of \TC{} other than the triangle is a $3$-sticking or a $2$-sticking of two graphs of \TC{} with fewer vertices, or a growing of a graph of \TC{} with fewer vertices. \end{lemma} \begin{proof} Let $Q=Q^+-v$, and let $Q$ be not the triangle. Then $|V(Q^+)|\ge 6$. \smallskip \textsc{Case 1}: $V(Q^+)$ is the disjoint union of two sets $V_1$, $V_2$ with $|V_1|,|V_2|\ge 2$ and with exactly three edges between $V_1$ and $V_2$. \begin{itemize} \item[(1a)] The three edges are pairwise disjoint. Suppose that a vertex $x$ of $V_1$ meets two or three of them \textup{(}the same for $V_2$\textup{)}. Move $x$ to the other side. Then $2$ or $0$ edges join the two sides, and both sides are non-empty because $|V_1|\ge 2$. This is impossible. \item[(1b)] $|V_i|$ is odd: three times $|V_i|$ is twice the number of edges inside $V_i$, plus $3$. So $|V_i|\ge 3$. \item[(1c)] Let $Q^+/V_2$ be the graph obtained from $Q^+$ by the identification of all vertices of $V_2$ to one new vertex and the deletion of the edges inside $V_2$. The new vertex has degree $3$ \textup{(}the three edges between $V_1$ and $V_2$\textup{)}, and every other vertex keeps its degree $3$. Put $Q_1^+=Q^+/V_2$ and $Q_2^+=Q^+/V_1$. Both graphs are cubic. They are $3$-edge-connected: a partition of the vertex set of $Q_1^+$ into two non-empty sets gives such a partition of $V(Q^+)$ with the same edges between the two sets \textup{(}put $V_2$ on the side of the new vertex\textup{)}; the same for $Q_2^+$. They have $|V_i|+1\ge 4$ vertices. \end{itemize} By symmetry let $v$ be in $V_1$, and let $z$ be the vertex of $Q_1^+$ that comes from $V_2$. Put $Q'=Q_1^+-v$, and let $Q''$ be the subgraph of $Q^+$ induced on $V_2$; this is $Q_2^+$ minus the vertex that comes from $V_1$. Both are in \TC. The vertices of degree $2$ of $Q''$ are the three ends in $V_2$ of the edges between $V_1$ and $V_2$. By (1a), at most one of these three edges is at $v$. \begin{itemize} \item[--] No one is at $v$: then $z$ has degree $3$ in $Q'$, and $Q$ is the $3$-sticking of $Q'$ and $Q''$ at $z$. \item[--] One of them is at $v$; let $y'$ be its end in $V_2$: then $z$ has degree $2$ in $Q'$, and $Q$ is the $2$-sticking of $Q'$ and $Q''$ at $z$. The two edges of $z$ go to two of the three vertices of degree $2$ of $Q''$; the third one is $y'$. \end{itemize} $Q'$ has $|V_1|$ vertices and $Q''$ has $|V_2|$ vertices. $Q$ has $|V_1|+|V_2|-1$ vertices; by (1b) this is more than each of the two numbers. \smallskip \textsc{Case 2}: there is no such pair $V_1$, $V_2$. Then $Q^+$ has no triangle: exactly three edges leave a triangle of a simple cubic graph, and $|V(Q^+)|\ge 6$, so a triangle would give a pair as in Case~1. Let $x$ be a neighbour of $v$, and $xy$ an edge of $Q$. In $Q^+$ the neighbours of $x$ are $v$, $y$, $x'$, and the neighbours of $y$ are $x$, $y_1$, $y_2$. Because $Q^+$ is simple and has no triangle, $v$, $x'$, $y_1$, $y_2$ are four different vertices, and they are different from $x$ and $y$. Let $\widehat{Q}^{+}$ be the graph obtained from $Q^+$ by the deletion of the edge $xy$ and the suppression of $x$ and $y$: the path $v,x,x'$ becomes a new edge $vx'$, and the path $y_1,y,y_2$ becomes a new edge $y_1y_2$. \textsc{Claim}: $\widehat{Q}^{+}$ is cubic and $3$-edge-connected, and it has $|V(Q^+)|-2\ge 4$ vertices. \emph{Proof of the claim.} $\widehat{Q}^{+}$ is cubic by its construction. Suppose that $V(\widehat{Q}^{+})$ is the disjoint union of two non-empty sets $V'$, $V''$ such that the set $E^*$ of the edges between them has at most $2$ elements \textup{(}$E^*$ can be empty\textup{)}. Put $x$ back: on the side of $v$ and $x'$ if they are on the same side; on any side if the edge $vx'$ is in $E^*$ \textup{(}then exactly one of the edges $vx$, $xx'$ joins the two sides, in the place of $vx'$\textup{)}. Do the same for $y$ with $y_1$, $y_2$. This is a partition of $V(Q^+)$ into two non-empty sets, and the number of edges of $Q^+$ between them is $|E^*|$ if $x$ and $y$ are on the same side, and $|E^*|+1$ if they are not. If $vx'$ or $y_1y_2$ is in $E^*$, we can choose the sides such that $x$ and $y$ are together. Then at most $2$ edges of $Q^+$ join the two sides: impossible. In the other case $x$ and $y$ have no choice. If they are on the same side, again at most $2$ edges join the sides: impossible. If they are on different sides and $|E^*|\le 1$, at most $2$ edges join the sides: impossible. If they are on different sides and $|E^*|=2$, exactly three edges join the sides, and the sides contain $\{x,v,x'\}$ and $\{y,y_1,y_2\}$: a pair as in Case~1, which does not exist. So $\widehat{Q}^{+}$ is connected \textup{(}$E^*$ empty\textup{)} and $3$-edge-connected. This proves the claim. \smallskip The vertex $v$ is a vertex of $\widehat{Q}^{+}$, so $\widehat{Q}=\widehat{Q}^{+}-v$ is in \TC. The vertex $x'$ is a neighbour of $v$ in $\widehat{Q}^{+}$, so it has degree $2$ in $\widehat{Q}$. And $y_1y_2$ is an edge of $\widehat{Q}$, because $y_1$ and $y_2$ are not $v$. The graph $Q$ is the growing of $\widehat{Q}$ at the edge $y_1y_2$ and the vertex $x'$, with the new vertices $y$ and $x$. $\widehat{Q}$ has two vertices fewer than $Q$. \end{proof} \subsection*{Well colouring}\label{app:wellcolouring-sec} \begin{lemma}[Well colouring]\label{lem:A3}\label{app:wellcolouring} Every graph of \TC{} has a well colouring. \end{lemma} \begin{proof} Induction on the number of vertices. By Lemma~\ref{app:generation} it is enough to show three things: the triangle has a well colouring; a graph of \TC{} that is a sticking of two well coloured graphs of \TC{} has a well colouring; a graph of \TC{} that is a growing of a well coloured graph of \TC{} has a well colouring. \smallskip \textsc{Triangle}. Colour two edges blue and one edge red. The blue edges form a spanning tree. There are two red components, and the two blue edges join them: $Q/\red$ is the $2$-join of two single vertices. \smallskip \textsc{Sticking}. Let $Q'$ and $Q''$ in \TC{} be well coloured, with the blue trees $T'$ and $T''$. Let $Q$ be the $d$-sticking of $Q'$ and $Q''$ at $z$, with the edges $t_1,\dots,t_d$ of $z$ and their new ends $z_1,\dots,z_d$. Keep all colours. (i) The blue edges of $Q$ are the edges of $T'$ and of $T''$; the edges of $T'$ at $z$ have a new end. Their number is $(|V(Q')|-1)+(|V(Q'')|-1)=|V(Q)|-1$. The tree $T''$ connects $V(Q'')$. Every component of $T'-z$ contains a neighbour of $z$ in $T'$, and the edge from this neighbour to $z$ ends now in $V(Q'')$. So the blue edges form a connected spanning subgraph of $Q$ with $|V(Q)|-1$ edges: a spanning tree. (ii) Let $\Omega$ be the red component of $z$ in $Q'$; it is a vertex of $Q'/\red$. By Lemma~\ref{app:count}~(c) the red edges of $Q'$ form a forest. So $\Omega$ without $z$ is the disjoint union of trees $\Omega_i$, one for every \emph{red} edge $t_i$ at $z$; the end of $t_i$ other than $z$ is in $\Omega_i$. \textup{(}If no edge at $z$ is red, then $\Omega=\{z\}$ and there is no $\Omega_i$.\textup{)} In $Q$, a red edge $t_i$ joins $\Omega_i$ to the red component of $z_i$ in $Q''$, and a blue edge $t_i$ is a blue edge at the red component of $z_i$. All other red components of $Q'$ and of $Q''$ are not changed. So the red components of $Q$ are: the red components of $Q'$ other than $\Omega$; and the red components of $Q''$, some of them enlarged by trees $\Omega_i$. An edge of $Q'/\red$ at $\Omega$ is a blue edge of $Q'$ with one end in $\Omega$. This end is $z$ or a vertex of some $\Omega_i$. In both cases the edge ends now in a red component that comes from $Q''$. Its other end is a vertex of $Q'/\red$ other than $\Omega$, because $Q'/\red$ has no loop. So $Q/\red$ is obtained in this way: take $Q'/\red$ without the vertex $\Omega$, and $Q''/\red$, and every edge of $Q'/\red$ at $\Omega$, with its end $\Omega$ moved to some vertex of $Q''/\red$. It remains to show that such a graph is $2$-constructible. \textsc{Claim}: let $U_a$ and $U_b$ be disjoint $2$-constructible graphs and $\omega$ a vertex of $U_a$; let $U$ be obtained from $U_a$ without $\omega$, from $U_b$, and from every edge of $U_a$ at $\omega$, with its end $\omega$ moved to some vertex of $U_b$. Then $U$ is $2$-constructible. Proof by induction on a $2$-construction of $U_a$. If $U_a$ has only the vertex $\omega$, then $U=U_b$. If $U_a$ is the $2$-join of $U'_a$ and $U''_a$ and $\omega$ is in $U'_a$, make the same change in $U'_a$; by induction the result $U'$ is $2$-constructible. Each of the two new edges of the $2$-join has one end in $U''_a$ and one end in $U'_a$; if this end was $\omega$, it is now a vertex of $U_b$, which is in $U'$. So $U$ is the $2$-join of $U'$ and $U''_a$. This proves the claim. So $Q$ is well coloured. \smallskip \textsc{Growing}. Let $Q'$ in \TC{} be well coloured with the blue tree $T'$, and let $Q$ be the growing of $Q'$ at the edge $y_1y_2$ and the vertex $x'$, with the new vertices $y$ and $x$. Colour the edges $xx'$ and $xy$ blue. Keep the colours of the old edges other than $y_1y_2$. \emph{The edge $y_1y_2$ is red.} Colour $y_1y$ and $yy_2$ red. The blue edges are those of $T'$, and $xx'$, and $xy$: a spanning tree of $Q$ \textup{(}$x$ hangs on $x'$, and $y$ hangs on $x$\textup{)}. The vertex $y$ is now in the red component of $y_1$ and $y_2$, and $x$ is a new red component with one vertex. So $Q/\red$ is $Q'/\red$ and one new vertex $x$ with two edges, to the red components of $x'$ and of $y$: a $2$-join of $Q'/\red$ and a single vertex. \emph{The edge $y_1y_2$ is blue.} $T'$ without the edge $y_1y_2$ has two components, $T'_1$ with $y_1$ and $T'_2$ with $y_2$. If $x'$ is in $T'_1$, colour $yy_2$ blue and $y_1y$ red. If $x'$ is in $T'_2$, colour $y_1y$ blue and $yy_2$ red. Look at the first case; the second one is the same with $y_1$ and $y_2$ exchanged. The number of blue edges is $(|V(Q')|-1)-1+3=|V(Q)|-1$, and the blue edges connect $V(Q)$: $T'_1$, then the path $x',x,y,y_2$, then $T'_2$. So they form a spanning tree. The vertex $y$ is in the red component of its red neighbour $y_1$. The vertices $y_1$ and $y_2$ are in different red components of $Q'$, by Lemma~\ref{app:count}~(c). So the blue edge $yy_2$ joins the same two red components that $y_1y_2$ joined. And $x$ is a new red component with one vertex and with two blue edges. Again $Q/\red$ is a $2$-join of $Q'/\red$ and a single vertex. In all cases $Q$ is well coloured. \end{proof} \subsection*{From a 2-construction to a closure sequence}\label{app:closure-sec} \begin{lemma}[Closure sequence]\label{lem:A4}\label{app:closureseq} Let $Q$ be a graph with a well colouring and with a $2$-construction of $Q/\red$. Number the nodes $1,\dots,s$ such that every node comes after all nodes that are subgraphs of its two operands. Then there are cycles $C_1,\dots,C_s$ of $Q$ such that, for every $i$, the cycle $C_i$ contains the two edges of the node $i$, and all other edges of $C_i$ are red edges or edges of nodes with a smaller number. So, with the set $R$ of the red edges, \[ E(Q)=R\cup E(C_1)\cup\dots\cup E(C_s), \] and each $C_i$ has at most two edges outside $R\cup E(C_1)\cup\dots\cup E(C_{i-1})$. \textup{(}Here it is not used that the red edges form a forest.\textup{)} \end{lemma} \begin{proof} Let $i$ be a node with the operands $U'$, $U''$ and the edges $t_1$, $t_2$. By Lemma~\ref{app:operands}~(b),~(c) the graphs $Q_{U'}$ and $Q_{U''}$ are connected and disjoint, and each of $t_1$, $t_2$ has one end in each of them. Take a path of $Q_{U'}$ between the ends of $t_1$ and $t_2$ in $V_{U'}$, and a path of $Q_{U''}$ between their ends in $V_{U''}$; a path can have length $0$. The two paths with $t_1$ and $t_2$ form a cycle $C_i$. \textup{(}If both paths have length $0$, it is the cycle of length $2$ with the edges $t_1$, $t_2$.\textup{)} By Lemma~\ref{app:operands}~(a), an edge of $Q_{U'}$ or of $Q_{U''}$ is red, or it is an edge of a node that is a subgraph of $U'$ or of $U''$; such a node has a number $h