Structural Reductions for Monochromatic Matchings and Ramsey Tilings

Extracted source text · arXiv 2606.24863

\documentclass[11pt,reqno]{amsart}
\usepackage{amsmath,amssymb,amsthm,mathtools,calc,verbatim,enumitem,tikz,url,hyperref,mathrsfs,cite,fullpage}
\usepackage{bbm}
\usepackage{pgfplots}
 \pgfplotsset{compat=1.18}
 \usepackage{tikz}
\usepackage{pgfplots}
 \pgfplotsset{compat=1.18}
 \usetikzlibrary{arrows.meta}
\usepackage{todonotes}
\usepackage{stmaryrd}
\usepackage{textcomp}
\usepackage{setspace}
\usepackage{amssymb}
\usepackage{amsthm}
\usepackage{amsmath}
\usepackage{graphicx}
%\usepackage{extarrows}
\usetikzlibrary{calc}
\usepackage{marvosym}
\usepackage{empheq}
\usepackage{latexsym}
%\usepackage{endnotes}
%\usepackage{natbib}
\usepackage{fontenc}
\usepackage{color}
\usepackage{hyperref}
\usepackage{cleveref}
\usepackage{dsfont}
\usepackage{bbm}

\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}

\usepackage{booktabs}
\usepackage{array}

\usepackage{todonotes}
\newcommand{\hong}[1]{\todo[inline,color=green]{Hong: #1}}

\newcommand{\circled}[1]{%
\tikz[baseline=(char.base)]{
  \node[draw, circle, inner sep=1pt] (char) {#1};
}}
\renewcommand{\baselinestretch}{1.1}
\addtolength{\footskip}{\baselineskip/2}

\newcommand\marginal[1]{\marginpar{\raggedright\parindent=0pt\tiny #1}}

\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{conjecture}[theorem]{Conjecture}
\newtheorem{construction}[theorem]{Construction}
\newtheorem{observation}[theorem]{Observation}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{claim}[theorem]{Claim}
\newtheorem*{claim*}{Claim}
\newtheorem{problem}[theorem]{Problem}
	\newtheorem{fact}[theorem]{Fact}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{question}[theorem]{Question}
\newtheorem*{qu*}{Question}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{example}[theorem]{Example}

\newenvironment{poc}{\begin{proof}[Proof of claim]\renewcommand*{\qedsymbol}{$\blacksquare$}}{\end{proof}}
    
\newcommand\N{\mathbb{N}}
\newcommand\R{\mathbb{R}}
\newcommand\Z{\mathbb{Z}}
\newcommand\E{\operatorname{\mathbb{E}}}
\newcommand\cA{\mathcal{A}}
\newcommand\cB{\mathcal{B}}
\newcommand\cC{\mathcal{C}}
\newcommand\cD{\mathcal{D}}
\newcommand\cF{\mathcal{F}}
\newcommand\cG{\mathcal{G}}
\newcommand\cH{\mathcal{H}}
\newcommand\cI{\mathcal{I}}
\newcommand\cJ{\mathcal{J}}
\newcommand\cK{\mathcal{K}}
\newcommand\cL{\mathcal{L}}
\newcommand\cM{\mathcal{M}}
\newcommand\cN{\mathcal{N}}
\newcommand\cP{\mathcal{P}}
\newcommand\cQ{\mathcal{Q}}
\newcommand\cS{\mathcal{S}}
\newcommand\cX{\mathcal{X}}
\newcommand\cT{\mathcal{T}}
\newcommand\rN{\tilde{N}}
\renewcommand\Pr{\operatorname{\mathbb{P}}}
\newcommand\id{\hbox{$1\mkern-6.5mu1$}}
\newcommand\lcm{\operatorname{lcm}}
\newcommand\eps{\varepsilon}
\renewcommand\leq{\leqslant}
\renewcommand\geq{\geqslant}
\renewcommand\le{\leqslant}
\renewcommand\ge{\geqslant}
\renewcommand\to{\rightarrow}
%\renewcommand\prec{\preccurlyeq}
\newcommand\ds{\displaystyle}
\newcommand\bmid{\mathrel{\big|}}

\pagestyle{plain}

	\def\eps{\varepsilon}
	\def\HH{\mathcal{H}}
	\def\TT{\mathcal{T}}
	\def\EE{\mathbb{E}}
	\def\C{\mathbb{C}}
	\def\R{\mathbb{R}}
	\def\Z{\mathbb{Z}}
	\def\PP{\mathcal{P}}
	\def\Prob{\mathbb{P}}	
	 
    
	\def\s{\sigma}
	\def\t{\theta}
	\def\a{\alpha}
	\def\Xt{\widetilde{X}}
	\def\Pt{\widetilde{P}}
	\def\Var{\mathrm{Var}}
	\def\PV{\mathrm{PV}}
	\def\NV{\mathrm{NV}}
	\def\sp{\mathrm{span}}
	\def\N{\mathbb{N}}
	\def\CC{\mathcal{C}}
	\def\P{\mathcal{P}}
	\def\AA{\mathcal{A}}
	\def\cA{\mathcal{A}}
	\def\QQ{\mathcal{Q}}
	\def\cQ{\mathcal{Q}}	
	\def\QC{\mathcal{C}}
	\def\cF{\mathcal{F}}
	\def\tm{\tilde{\mu}}
	\def\ts{\tilde{\sigma}}
	
	\def\g{\gamma}
	\def\G{\Gamma}
    \def\diff{\triangle}
	\def\Q{\mathcal{Q}}		
	
	\def\p{\partial }	
	\def\la{\langle }
	\def\ra{\rangle }
	\def\<{\langle }
	\def\>{\rangle }


% Part dividers for the main body.
% These do not reset section counters or theorem counters.
\newcounter{paperpart}
\renewcommand{\thepaperpart}{\Roman{paperpart}}

\newcommand{\paperpart}[1]{%
  \par\bigskip\bigskip
  \refstepcounter{paperpart}%
  \phantomsection
  \addcontentsline{toc}{part}{Part~\thepaperpart. #1}%
  \begin{center}
    {\large\bfseries Part~\thepaperpart. #1}
  \end{center}
}

%the following is to see changes
\newcommand{\old}[1]{{\color{red}#1}}
\newcommand{\new}[1]{{\color{blue}#1}}

%for the final-clean mode, redefine them as:
%\renewcommand{\old}[1]{}
%\renewcommand{\new}[1]{#1}

\begin{document}

\title{Structural Reductions for Monochromatic Matchings and Ramsey Tilings}
\author{Hong Liu \and Maksim Turevskii \and Lanchao Wang \and Zhifei Yan}

\address{ECOPRO, Institute for Basic Science, 55 Expo-ro, Yuseong-gu, Daejeon, 34126, Korea}\email{\{hongliu,zhifeiyan\}@ibs.re.kr}  
\address{Faculty of Mathematics and Computer Science, Saint Petersburg State University, Saint Petersburg, 199178, Russia}\email{turmax20052005@gmail.com} 

\address{School of Mathematics, Nanjing University, Nanjing, China, and ECOPRO, Institute for Basic Science, 55 Expo-ro, Yuseong-gu, Daejeon, 34126, Korea}\email{lanchaowang@foxmail.com}


% Supported by the Institute for Basic Science (IBS-R029-C4)

\thanks{}
	

\begin{abstract}
The Alon--Frankl--Lov\'asz theorem determines the chromatic number of Kneser
hypergraphs; equivalently, it gives the sharp minimum size of a monochromatic
matching in every \(r\)-edge-colouring of the complete \(t\)-uniform
hypergraph. The known proofs of the exact theorem are topological. We develop
a topology-free structural framework for its asymptotic form and for related
sparse and tiling problems.

Our main theorem shows that every \(r\)-colouring of a sufficiently
pseudo-random \(t\)-graph can be reduced, with only \(o(n)\) loss in the
largest monochromatic matching, to a colouring of \(K_n^{(t)}\) whose vertex
set is partitioned into at most \(r\) parts and whose edge colours depend only
on intersection profiles. The proof combines hypergraph regularity, LP
duality, and convex-geometric compression. As consequences, we obtain a topology-free proof of the asymptotic
AFL theorem, a sparse random transference theorem, and near-exact bounds in a
linear-uniformity regime of Meunier's stable Kneser conjecture.

For a graph \(H\), let \(Rt_r(H;K_n)\) be the minimum, over all
\(r\)-edge-colourings of \(K_n\), of the largest monochromatic \(H\)-tiling.
We prove
\[
Rt_r(H;K_n)=(\beta_{r,H}+o(1))n,
\]
where \(\beta_{r,H}\) is effectively computable from finitely many linear
programs depending only on \(H\) and \(r\). An additional multipartite Ramsey
extraction is the key ingredient needed to reconstruct consistent graph
templates. This gives an effective asymptotic solution to the complete-host
multicolour Ramsey-tiling problem, extending the classical two-colour theorem
of Burr, Erd\H{o}s and Spencer. We also determine explicit constants for several natural families, including
connected non-bipartite graphs, balanced Hall-type bipartite graphs, complete
bipartite graphs with three, four, and five colours, and a non-Hall bipartite example.
\end{abstract}	

    
	\maketitle

%\newpage
%\tableofcontents



\section{Introduction}

A central theme in extremal combinatorics and Ramsey theory is to understand how much
monochromatic structure is forced by an arbitrary colouring. In this paper we
study this question for matchings in uniform hypergraphs and, more generally,
for tilings in graphs. The starting point is the following theorem of Alon, Frankl and
Lov\'asz~\cite{AFL}, which resolved a conjecture of Erd\H{o}s~\cite{Erdos}
and generalized Kneser's original conjecture~\cite{Kneser,Lovasz}. We state it
in the equivalent language of edge-coloured complete hypergraphs.

\begin{theorem}[Alon, Frankl and Lov\'asz]\label{thm:AFL}
Let \(r,t\geq 2\) and \(n\in\mathbb N\), and let \(\chi\) be an arbitrary
\(r\)-edge-colouring of \(K_n^{(t)}\). Then \(\chi\) contains a monochromatic
matching consisting of
\[
\left\lfloor\frac{n+r-1}{r+t-1} \right\rfloor
\]
vertex-disjoint \(t\)-edges. Moreover, this bound is tight.
\end{theorem}

Theorem~\ref{thm:AFL} is one of the fundamental results connecting
Kneser-type colouring problems, Ramsey theory and topological combinatorics.
Known proofs of the exact theorem and its refinements use equivariant
topology, through the Borsuk--Ulam theorem, Tucker-type lemmas, or related
topological obstructions; see, for example,
\cite{AFL,Kriz,Matousek,Ziegler2002,Haviv}. These methods are powerful, but
they do not explain the extremal colourings in elementary combinatorial terms.
In particular, they do not provide a structural reduction from an arbitrary
colouring to a small family of explicit model colourings.

Our first goal is to give such a reduction asymptotically. We show that every
colouring of a sufficiently pseudo-random host hypergraph can be compressed,
for the purpose of finding large monochromatic matchings, to a highly
structured colouring of a complete hypergraph. The structured colourings are
the following. An \emph{\(r\)-strip colouring} is an edge-colouring for which the vertex set is
partitioned into at most \(r\) parts and the colour of an edge depends only on
its intersection profile with these parts. Thus, once the part sizes are fixed,
the colouring is described by finitely many profile colours. Our main
hypergraph theorem says that arbitrary colourings can be reduced to such
strip colourings with only \(o(n)\) loss in the maximum monochromatic matching
size. If \(\chi\) is an
edge-colouring, we write \(\nu(\chi)\) for the maximum size of a monochromatic
matching in \(\chi\).

\begin{theorem}[Structural reduction via strip colourings, informal]\label{thm:informal-hypergraph-structure}
Let \(r,t\ge2\) be fixed, and let \(\cG\) be a sufficiently large
pseudo-random \(t\)-graph on \(n\) vertices. For every \(r\)-edge-colouring
\(\chi\) of \(\cG\), there exists an \(r\)-strip \(r\)-edge-colouring \(\psi\)
of \(K_n^{(t)}\) such that
\[
\nu(\chi)\ge \nu(\psi)-o(n).
\]
\end{theorem}

The formal statement is Theorem~\ref{thm:hypergraphstructure}. We view it as
a compression theorem. Weak hypergraph regularity reduces the original
colouring to a bounded object on \(M=O(1)\) clusters, but this reduced object
still consists of \(O(M^t)\) coloured \(t\)-tuples and has no apparent strip
structure. The main task is to compress the \(M\) cluster types to at most
\(r\) strip types while preserving the information that controls
monochromatic matchings.

This information comes from LP duality. For each colour, the fractional
matching problem in the reduced hypergraph has a dual fractional-cover
problem, assigning a weight to every cluster. Grouping the dual weights over
all \(r\) colours gives vectors
$S_i=(s_i^{(1)},\ldots,s_i^{(r)})\in\mathbb R^r .$
The vector sum \(\sum_i S_i\) records the dual objectives, and its
\(\ell_\infty\)-norm controls the matching bound. The dual constraints say that almost every reduced \(t\)-set is covered in some colour. 

There are two obstacles. First, this covering condition holds only almost
everywhere, so the exceptional tuples must be repaired without substantially
increasing the total weight. Second, even after repair, there may still be
many vector types. We handle the first issue by a rounding-and-frequency
repair, and the second by a Carath\'eodory-type compression which replaces the
repaired vectors by a new multiset with at most \(r\) distinct values, without
increasing the relevant \(\ell_\infty\)-norm by more than \(o(M)\). These values
define the parts of the strip colouring.


Theorem~\ref{thm:informal-hypergraph-structure} reduces the asymptotic
Alon--Frankl--Lov\'asz theorem to a finite profile-packing statement for strip
colourings. This gives the following topology-free asymptotic form.

\begin{corollary}[Asymptotic AFL]\label{cor:intro-asyAFL}
For every fixed \(r,t\ge2\), every \(r\)-edge-colouring \(\chi\) of
\(K_n^{(t)}\) satisfies
\[
\nu(\chi)\ge \frac{n}{r+t-1}-o(n).
\]
\end{corollary}

The final profile-packing step is elementary but not vacuous: it is the
only point where the constant \((r+t-1)^{-1}\) enters. Thus the topological part of the classical proof is replaced, asymptotically,
by two finite ingredients: an LP-duality structural reduction and an elementary
profile-packing lemma for strip colourings.

\subsection{Hypergraph applications}

The structural theorem is stated for pseudo-random hosts, and therefore
applies beyond complete hypergraphs. Our first application is a sparse
transference version of AFL. In recent years, a growing body of work in combinatorics has demonstrated that
many classical results for dense graphs and hypergraphs admit natural
analogues in sparse random settings. Such results are commonly referred to as
\emph{transference theorems}. Notable examples include the work of R\"odl and
Ruci\'nski~\cite{RR} on the transference of Ramsey's theorem, as well as the
transference of Tur\'an's theorem due to Conlon and Gowers~\cite{CG}, Balogh,
Morris, and Samotij~\cite{BMS}, Saxton and Thomason~\cite{ST}, and
Schacht~\cite{Schacht}. More recently, Gishboliner, Glock, Michaeli and
Sgueglia~\cite{GGMS} proved such a result for random hypergraphs (see Theorem~\ref{thm:randomAFL}). Their proof
uses the dense AFL theorem as a black box, through a defect version of AFL and
additional extremal set-theoretic input. Our method gives an independent
structural route.

\begin{theorem}\label{thm:randomaaa0}
Let \(t,r\ge 2\) and \(\varepsilon>0\), and let
\(G\sim G^{(t)}(n,p)\). There exists \(C>0\) such that if
\(p \ge C n^{-t+1}\), then with high probability the following holds. For
every \(r\)-edge-colouring \(\chi\) of \(G\), there exists an \(r\)-strip
\(r\)-edge-colouring \(\psi\) of \(K_n^{(t)}\) such that
\[
\nu(\chi)\geq \nu(\psi)-\varepsilon n .
\]
\end{theorem}

Combining Theorem~\ref{thm:randomaaa0} with the strip-colouring lower
bound gives a topology-free proof of the transference AFL theorem in
\(G^{(t)}(n,p)\).

\medskip

We also apply the method to stable Kneser hypergraphs. The motivation goes
back to Schrijver's theorem~\cite{Schrijver78} in 1978: the subgraph of the Kneser
graph induced by the \(2\)-stable \(t\)-sets has the same chromatic number as
the full Kneser graph and is, moreover, vertex-critical. Thus the full
colouring obstruction survives after restricting to a much smaller,
structured family of vertices.

More generally, regard \([n]\) as cyclically ordered, and call a \(t\)-set
\(s\)-stable if every two of its elements have cyclic distance at least \(s\).
Let
$\mathrm{KG}^q(n,t)_{s\text{-stab}}$
denote the subhypergraph of \(\mathrm{KG}^q(n,t)\) induced by the
\(s\)-stable \(t\)-sets. Ziegler~\cite{Ziegler2002} conjectured that the
\(q\)-stable \(q\)-uniform Kneser hypergraph has the same chromatic number as
the full Kneser hypergraph, and Alon, Drewnowski and
{\L}uczak~\cite{ADL} made this conjecture explicit. Meunier~\cite{Meunier}
later proposed the broader formula
\[
\chi\bigl(\mathrm{KG}^q(n,t)_{s\text{-stab}}\bigr)
=
\left\lceil
\frac{n-\max\{q,s\}(t-1)}{q-1}
\right\rceil
\qquad
\text{whenever }n\ge \max\{q,s\}t.
\]
Partial results were obtained by Meunier~\cite{Meunier},
Frick~\cite{Frick}, and others~\cite{Jonsson,Chen,Jafari,Daneshpajouh}, but the general
conjecture remains open.


We determine the value predicted by Meunier's conjecture up to an additive
error of at most one in the regime where \(t\) is fixed, \(q\) is linear in
\(n\), and \(s\ll q\).


\begin{theorem}\label{thm:stable}
Let \(t\geq 2\) and \(0<c\leq 1/(2t)\), and let \(n\) be sufficiently
large with \(q=cn\). Then for every \(2\leq s=s(n)\ll q\),
\[
\bigg\lfloor\frac{n-(t-1)q}{q-1}\bigg\rfloor
\leq
\chi\bigl(\mathrm{KG}^q(n,t)_{s\text{-stab}}\bigr)
\leq
\bigg\lceil\frac{n-(t-1)q}{q-1}\bigg\rceil .
\]
\end{theorem}


\subsection{Graph tilings}

The second part of the paper shows that the LP-duality philosophy is not
limited to matchings. We apply it to monochromatic graph tilings in Ramsey
theory. For a host graph \(G\) and a fixed graph \(H\), define
\[
Rt_r(H;G)
:=
\min_{\chi}
\max_{\ell\in[r]}
\{\text{maximum number of vertex-disjoint colour-\(\ell\) copies of \(H\)}\},
\]
where the minimum is over all \(r\)-edge-colourings \(\chi\) of \(G\). We
write \(\nu_H(\chi)\) for the corresponding maximum monochromatic
\(H\)-tiling size in \(\chi\).

The study of Ramsey numbers for graph tilings was initiated by Burr,
Erd\H{o}s and Spencer~\cite{BES} in 1975. They proved that, for every fixed graph
\(H\) without isolated vertices,
\[
Rt_2(H;K_n)
=
\frac{n}{2v(H)-\alpha(H)}+O_H(1),
\]
or equivalently,
$R_2(kH)
=
\bigl(2v(H)-\alpha(H)\bigr)k+O_H(1).$
Thus the general two-colour problem is governed by  \(\alpha(H)\). Subsequent work refined the additive term and the
threshold for exactness~\cite{Burr,BS}, while special multicolour cases were
studied by Cockayne and Lorimer~\cite{CL} and Loo~\cite{Loo}. Related
questions for non-complete hosts have been investigated under minimum-degree,
random, and robust pseudorandomness assumptions
\cite{BFT,ACFMPY,FMT,GKM,KM,NW}. Nevertheless, no general multicolour
counterpart to the Burr--Erd\H{o}s--Spencer theorem was previously available,
even for complete hosts.

We study this general multicolour problem. A first indication that the
two-colour formula does not extend directly is a sharp distinction between
bipartite and non-bipartite target graphs. For connected non-bipartite \(H\),
we prove that the elementary greedy lower bound is asymptotically optimal; see
Proposition~\ref{prop:nonbipartite}. For bipartite \(H\), larger
monochromatic tilings can be forced, and the answer is governed by a finite
optimization over strip templates.

The graph analogue of
Theorem~\ref{thm:informal-hypergraph-structure} is the following structural
theorem. Recall that the maximum \(2\)-density of \(H\) is
$m_2(H)
=
\max\left\{
\frac{e(H')-1}{v(H')-2}:
H'\subseteq H,\ v(H')\ge3
\right\}.$

\begin{theorem}\label{thm:graphstructure}
Let \(r\geq2\), \(\varepsilon>0\), and let \(H\) be a connected graph with
\(v(H)\ge3\). Then there exists \(C>0\) such that, for all sufficiently
large \(n\), the following holds. If \(G\sim G(n,p)\) with
\(p\ge Cn^{-1/m_2(H)}\), then with high probability every
\(r\)-edge-colouring \(\chi\) of \(G\) admits an \(r\)-strip
\(r\)-edge-colouring \(\psi\) of \(K_n\) satisfying
\[
\nu_H(\chi)\ge \nu_H(\psi)-\varepsilon n .
\]
\end{theorem}

The graph theorem is not a formal specialization of the matching theorem to
\(t=2\). In the matching setting, once the dual vectors have been compressed,
one may assign a colour independently to each edge profile. For
\(H\)-tilings, the colours assigned to pairs of compressed parts must be
globally consistent: every monochromatic copy of \(H\) in the reconstructed
template must correspond to a genuine reduced \(H\)-pattern.

This consistency requires an additional Ramsey-theoretic ingredient. Inside
each frequent class of rounded dual vectors, we select a bounded collection of
representative clusters. A multipartite Ramsey extraction then produces
subcollections for which every pair inside or between the selected classes
has a fixed colour. These colours define the global strip template
\(\lambda(i,j)\), while the dual constraints for the representatives certify
every monochromatic \(H\)-pattern in that template. This Ramsey extraction is
the main new ingredient needed to pass from matchings to graph tilings.

For complete hosts, the structural theorem reduces the asymptotic
Ramsey-tiling problem to a finite optimization over strip templates. For each
fixed \(r\) and \(H\), only finitely many templates with at most \(r\) parts
must be considered, and the maximum monochromatic \(H\)-tiling supported by
each template is described by a finite linear program.

\begin{theorem}\label{thm:alg}
Let \(r\ge2\) and let \(H\) be a  connected graph with $v(H)\ge 3$. Then
there exists a constant \(\beta_{r, H}\), depending only on \(r\) and \(H\),
which is effectively computable from a finite collection of linear programs
whose sizes depend only on \(r\) and \(H\), such that
\[
Rt_r(H;K_n)=(\beta_{r,H}+o(1))n.
\]
Consequently, as \(k\to\infty\),
$R_r(kH)=\left(\frac1{\beta_{r,H}}+o(1)\right)k.$
\end{theorem}


Theorem~\ref{thm:alg} gives an effective asymptotic solution to the
complete-host multicolour Ramsey-tiling problem for every fixed \(H\) and
\(r\), extending the classical two-colour theorem of Burr, Erd\H{o}s and
Spencer. The finite optimization does not always yield a transparent closed
formula, but we solve it explicitly in several natural families. We determine
the value for 
\begin{itemize}
    \item connected non-bipartite graphs (
Proposition~\ref{prop:nonbipartite});

\item three-colour Hall-type bipartite
graphs (Theorem~\ref{thm:biparr=3});

\item complete bipartite graphs \(K_{a,b}\) with four and five colours (Theorems~\ref{thm:Kab-four} and~\ref{thm:Kab-five});

\item balanced Hall-type bipartite graphs for every number of colours (Theorem~\ref{thm:Kaa});

\item three-colour value for the
double star \(D_{2,3}\) (Theorem~\ref{thm:D23}).
\end{itemize}
This last example $D_{2,3}$ does not satisfy the Hall-type condition and illustrates that the finite
optimization can be solved beyond the Hall-type setting, but that the answer
may depend on finer features of \(H\) than its bipartition sizes.


\subsection{Summary of the results}

The following table summarizes the main structural reductions and their
applications.

\begin{table}[htbp]
    \centering
    \footnotesize
    \renewcommand{\arraystretch}{1.5}
    \begin{tabular}{
        >{\raggedright\arraybackslash}p{0.22\textwidth}
        >{\raggedright\arraybackslash}p{0.20\textwidth}
        >{\raggedright\arraybackslash}p{0.25\textwidth}
        >{\raggedright\arraybackslash}p{0.25\textwidth}
    }
        \toprule
        \textbf{Host setting} &
        \textbf{Target structure} &
        \textbf{Structural input} &
        \textbf{Main consequences} \\
        \midrule

        Complete or pseudo-random \(t\)-graphs &
        Monochromatic matchings &
        Theorem~\ref{thm:hypergraphstructure}: strip-colouring reduction
        via LP duality and vector compression &
        Corollary~\ref{cor:intro-asyAFL}: asymptotic AFL;
        Theorem~\ref{thm:stable}: stable Kneser bounds \\

        \midrule

        Random \(t\)-graphs \(G^{(t)}(n,p)\) &
        Monochromatic matchings &
        Theorem~\ref{thm:randomaaa0}: sparse strip-colouring reduction &
        Theorem~\ref{thm:randomAFL}: transference AFL without using exact
        AFL as a black box \\

        \midrule

        Random and complete graphs &
        Monochromatic \(H\)-tilings &
        Theorem~\ref{thm:graphstructure}: reduced \(H\)-patterns,
        LP duality, multipartite Ramsey extraction, and strip templates &
        Theorem~\ref{thm:alg}: finite LP algorithm for
        \(Rt_r(H;K_n)\) \\

        \midrule

        Complete graphs, explicit families &
        Non-bipartite and bipartite tilings &
        Strip-template analysis &
        Proposition~\ref{prop:nonbipartite};
        Theorems~\ref{thm:biparr=3}, \ref{thm:Kaa}, 
\ref{thm:Kab-four}, \ref{thm:Kab-five} and \ref{thm:D23} \\

        \bottomrule
    \end{tabular}
    \vspace{0.2cm}
    \caption{Main structural reductions and applications.}
    \label{tab:main_results}
\end{table}


\subsection{Organization}

Section~\ref{sec:proofoverview} gives a high-level overview of the LP-duality and strip-colouring framework. We then divide the rest of the paper into two halves.

The first half concerns monochromatic matchings in
uniform hypergraphs. In Section~\ref{sec:main}, we prove the hypergraph
structural theorem, Theorem~\ref{thm:hypergraphstructure}. In
Section~\ref{sec:hypapp}, we apply it to prove the asymptotic AFL theorem, the
random transference theorem, and the bounds for stable Kneser hypergraphs.

The second half concerns Ramsey graph tilings. In
Section~\ref{sec:graphstructure}, we prove the graph-tiling structural theorem,
Theorem~\ref{thm:graphstructure}. In Section~\ref{sec:graphapp}, we use it to
derive the finite LP algorithm for \(Rt_r(H;K_n)\) and to compute explicit
values for several families of graphs.

Finally, Section~\ref{sec:conclusion} contains concluding remarks and further
directions. The appendices contain the proof of the five-colour \(K_{a,b}\)-tiling
formula, the double-star calculation, and the additional hypergraph-tiling
construction.


\section{Proof overview}\label{sec:proofoverview}

We now explain the mechanism behind the main theorem. The proof of Theorem~\ref{thm:informal-hypergraph-structure} (formalized as Theorem~\ref{thm:hypergraphstructure}) consists of the following conceptual steps:
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[
  >=Stealth,
  font=\normalsize,
  node/.style={align=center, inner sep=2pt},
  mainarrow/.style={->, line width=1.2pt},
  looparrow/.style={->, line width=1.2pt},
scale=0.8]

% ----- nodes -----
\node[node] (Sk) at (0,0) {Reduced graphs};


\node[node] (Wij) at (5,0) {$S_i$};

\node[node] (Si) at (9,0) {$Y_i$};

\node[node] (strip) at (15,0) {Strip colourings};

% ----- straight main flow -----
\draw[mainarrow] (Sk) -- node[above=6pt] {LP Duality} (Wij);


% Wij -> Si (single arrow); label above arrow is "(3) Dual LP"
\draw[mainarrow] (Wij) -- node[above=6pt] {reduction} (Si);



\draw[mainarrow] (Si) -- node[above=6pt] {constructing} (strip);


\end{tikzpicture}
\caption{The structural pipeline: from coloured reduced hypergraphs to
dual cluster weights $S_i$, then to a compressed vector system $Y_i$, and finally to a
strip-colouring template.}\label{fig:toolkit}
\end{figure}


\subsection{Dual weights from reduced matchings}

Apply the multicolour weak hypergraph regularity lemma to an
\(r\)-edge-coloured pseudo-random \(t\)-graph. This gives an equipartition into
\(M=O_{r,t,\varepsilon}(1)\) clusters. For each colour \(\ell\), let
\(\mathcal X_\ell\) be the family of regular \(t\)-tuples whose colour-\(\ell\)
density is large enough. The fractional matching LP in this reduced
\(\ell\)-coloured \(t\)-graph has variables \(w_X\), \(X\in\mathcal X_\ell\),
and its dual is the fractional vertex-cover LP
\[
\begin{aligned}
\text{minimize}\quad
& \sum_{i=1}^M z_i\\
\text{subject to}\quad
& \sum_{i\in X}z_i\ge1
\qquad\text{for every }X\in\mathcal X_\ell,\\
& z_i\ge0
\qquad\text{for every }i\in[M].
\end{aligned}
\]
By the embedding lemma, a feasible fractional matching in the reduced
graph lifts to an actual monochromatic matching in the original hypergraph,
with only \(o(n)\) loss. Thus the optimal dual values control the matching
size.

For each colour \(\ell\), take an optimal dual solution and scale it by
\(t\). This gives vectors
$S_i=(s_i^{(1)},\ldots,s_i^{(r)})\in[0,t]^r,$ $i\in[M].$
For all but \(o(M^t)\) reduced \(t\)-sets \(X\), there is a colour
\(\ell\) such that
$\sum_{i\in X}s_i^{(\ell)}\ge t.$
At the same time,
$\left\|\sum_{i=1}^M S_i\right\|_\infty$
records the largest dual objective among the \(r\) colours, and hence the
matching lower bound obtained from the reduced graph. The original
colouring problem is therefore converted into a finite-dimensional problem
about a multiset of points in \([0,t]^r\) satisfying an almost-everywhere
\(t\)-wise covering condition.




\subsection{Repair and vector compression}

The vectors \(S_i\) do not immediately define a strip colouring. There
are two problems. First, the covering condition may fail on the exceptional
tuples left by regularity. Second, even if it held for every tuple, the number
of distinct vectors \(S_i\) could be as large as \(M\), whereas an \(r\)-strip
colouring must have at most \(r\) parts.

The first problem is handled by a repair step. We round all coordinates
up to a fine grid. Rare rounded values are then replaced by frequent rounded
values. Since the exceptional set has size \(o(M^t)\), this replacement turns
the almost-everywhere covering condition into a covering condition for every
\(t\)-set, while increasing $\left\|\sum_i S_i\right\|_\infty$
by only \(o(M)\). The frequency condition is important later: it ensures
that every compressed vector value can still be represented by many distinct
clusters.

The second problem is the main geometric step. Let \(S'_1,\ldots,S'_M\)
be the repaired vectors, and let \(P=\operatorname{conv}\{S'_1,\ldots,S'_M\}\).
Write \(\overline S=M^{-1}\sum_i S'_i\). Move from \(\overline S\) in the
negative all-ones direction until the ray first hits the boundary of \(P\).
Call the resulting boundary point \(J\). Since \(J\le \overline S\)
coordinatewise up to a common shift, it does not increase the relevant
\(\ell_\infty\)-norm. Moreover, \(J\) lies on a proper face of \(P\), and hence
Carath\'eodory's theorem on that face represents \(J\) as a convex combination
of at most \(r\) of the repaired vectors.

After rounding the coefficients of this convex combination to integer
multiplicities summing to \(M\), we obtain a new family
\(Y_1,\ldots,Y_M\) taking at most \(r\) distinct values and satisfying
\[
\frac{k}{t}\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty+o(n),
\]
where \(k\) is the cluster size and \(Z\) is the vector of dual optimal
values. The repaired covering condition is preserved because each compressed
value occurs frequently enough in the repaired system to realize every
required \(t\)-set by distinct original clusters.

\subsection{Constructing the strip colouring}

Let \(R_1,\ldots,R_q\) be the distinct values among the \(Y_i\), where
\(q\le r\). Merge all clusters with the same \(Y\)-value into a part
\(U_j\). To colour the complete \(t\)-graph on \(U_1\cup\cdots\cup U_q\),
consider an edge profile
$\mathbf a=(a_1,\ldots,a_q),$
$a_i\ge0,$
$a_1+\cdots+a_q=t.$
Because each value class has many representatives, we may choose a reduced
\(t\)-set using exactly \(a_j\) clusters of type \(R_j\). The repaired covering
condition supplies a colour \(\ell(\mathbf a)\) such that the corresponding
\(\ell(\mathbf a)\)-coordinate sum is at least \(t\). We then colour every
edge of profile \(\mathbf a\) by \(\ell(\mathbf a)\). This defines an
\(r\)-strip colouring \(\psi\) of \(K_n^{(t)}\).

The dual weights now bound monochromatic matchings in \(\psi\). If
\(\mathcal M\) is a colour-\(\ell\) matching in \(\psi\), then every edge of
\(\mathcal M\) consumes at least \(t\) units of \(\ell\)-weight, while the
total available \(\ell\)-weight is controlled by \(\sum_i y_i^{(\ell)}\).
Therefore
\[
\nu(\psi)
\le
\frac{k}{t}
\left\|\sum_{i=1}^M Y_i\right\|_\infty
+o(n).
\]
Combining this with the LP lower bound from the reduced hypergraph gives the desired conclusion 
$\nu(\chi)\ge \nu(\psi)-o(n).$

\subsection{Where the AFL constant enters}

The structural theorem itself does not know the AFL constant. It says
only that arbitrary colourings reduce to strip colourings. To deduce
Corollary~\ref{cor:intro-asyAFL}, one must prove a finite extremal statement: every
\(r\)-strip colouring of \(K_n^{(t)}\) contains a monochromatic matching of
size
$\frac{n}{r+t-1}-O(1).$ Indeed, in a strip colouring with parts \(U_1,\ldots,U_q\), \(q\le r\),
all edges with the same profile
$(a_1,\ldots,a_q),$
$a_i=|e\cap U_i|,$
$\sum_i a_i=t,$
have the same colour. Thus, if there is an integer profile
\(\mathbf a=(a_1,\ldots,a_q)\) satisfying
$a_i m\le |U_i|$
for every $i$,
$m=\left\lfloor\frac{n}{r+t-1}\right\rfloor,$
then we may greedily choose \(m\) disjoint edges of that profile to get a monochromatic matching.

The profile-packing lemma asserts that such a profile always exists. The
case in which one part has size at least \(tm\) is immediate, using the
profile \((t,0,\ldots,0)\). The complementary case is the only remaining
integer-combinatorial input. This is the point at which the exact denominator
\(r+t-1\) is encoded, and it should be treated explicitly rather than hidden
inside a generic pigeonhole statement. Once this strip-colouring lemma is
proved, the asymptotic AFL theorem follows immediately from
Theorem~\ref{thm:hypergraphstructure}.

\subsection{The additional Ramsey step for graph tilings}

For graph tilings, the linear programs are indexed by reduced
\(H\)-patterns
$\phi:V(H)\to[M],$
whose edges are mapped to dense regular pairs of one colour. Their duals
produce cluster-weight vectors of the same kind as in the matching problem.
After compressing these vectors to distinct values
\(R_1,\ldots,R_q\), however, the construction has an additional consistency
problem. In the matching setting, a colour may be assigned independently to
each edge profile. For \(H\)-tilings, one must choose colours
\(\lambda(i,j)\) for all pairs of compressed types so that every
monochromatic map
$f:V(H)\to[q]$
is witnessed by an actual reduced \(H\)-pattern of the same colour. An
arbitrary assignment of the pair colours need not have this property.

The repair step ensures that every compressed value \(R_i\) has many
representatives among the original clusters. From each representative class
we first select a bounded set containing no exceptional pair. A multipartite
Ramsey extraction then yields subsets \(A_i\), each of size \(v(H)\), such
that all pairs inside each \(A_i\), and all pairs between each \(A_i\) and
\(A_j\), have fixed colours. We define these colours to be
\(\lambda(i,j)\).

Now suppose that \(f:V(H)\to[q]\) is monochromatic of colour \(\ell\) in
the resulting template. Choosing distinct representatives
\[
\phi(x)\in A_{f(x)},
\qquad \text{for all }x\in V(H),
\]
produces a genuine colour-\(\ell\) reduced \(H\)-pattern. Its dual constraint
then gives
\[
\sum_{x\in V(H)}R_{f(x)}^{(\ell)}\ge1,
\]
which is exactly the inequality needed to control monochromatic \(H\)-tilings
in the constructed strip colouring. This Ramsey extraction is the additional
ingredient required in the graph setting.

\section{Structural reduction for hypergraph matchings}\label{sec:main}
A $t$-graph $\cG$ is said to be \emph{$(\eta,p,d,D)$-uniform} if, for every collection of $t$ pairwise disjoint vertex sets $X_1,\dots,X_t$ with $|X_i|\ge \eta n$ for all $i$, the number of edges having one vertex in each $X_i$ lies between
$dp\cdot\prod_{i=1}^t |X_i|$
 and 
$Dp\cdot\prod_{i=1}^t |X_i|.$
Let $\nu(\chi)$ be the maximum monochromatic matching size in $\chi$.

\begin{definition}\label{def:strip}
Let $t\geq2$ and $s\geq1$, and let $\cF$ be a $t$-graph. An edge-colouring $\chi$ of $\cF$ is an \emph{$s$-strip colouring} if there exists a partition
$V(\cF)=V_1\cup\cdots\cup V_s$ such that
\[
\chi(e_1)=\chi(e_2)
\quad\text{whenever}\quad
|e_1\cap V_j|=|e_2\cap V_j|
\ \text{for every }j\in[s].
\]
\end{definition}


\begin{theorem}\label{thm:hypergraphstructure}
Let $r,t\geq2$, $\varepsilon>0$ and $0<d\leq1<D$. There exists a constant $\eta>0$ such that the following holds. Let $n$ be sufficiently large and $n^{-t}\ll p\leq 1$, and let $\cG$ be an $(\eta,p,d,D)$-uniform $t$-graph on $n$ vertices equipped with an $r$-edge colouring $\chi$. Then there exists an $r$-strip $r$-edge-colouring $\psi$ of $K_n^{(t)}$ such that 
$$\nu(\chi)\geq \nu(\psi) - \varepsilon n.$$
\end{theorem}


The proof is divided into the three steps as outlined in Section~\ref{sec:proofoverview}.


\subsection{Preliminaries: Regularity and Embeddings}\label{sec:pre}
We begin with some standard notation. 
Let $\mathcal F=\mathcal F(V_1,\dots,V_t)$ be a $t$-partite $t$-graph. 
Denote by $e(V_1,\dots,V_t)$ the number of edges having exactly one vertex in each $V_i$. 
The \emph{density} of $\mathcal F$ is defined as
\[
d(V_1,\dots,V_t)
:= \frac{e(V_1,\dots,V_t)}{|V_1|\cdots |V_t|}.
\]

\begin{definition}
Let $\varepsilon > 0$ and $p\in[0,1]$. 
We say that a $t$-partite $t$-graph $\mathcal F(V_1,\dots,V_t)$ is \emph{$(\varepsilon,p)$-regular} 
if for every choice of subsets $V_i' \subseteq V_i$ with $|V_i'| \ge \varepsilon |V_i|$ for all $i\in[t]$, we have
\[
\bigl| d(V_1',\dots,V_t') - d(V_1,\dots,V_t) \bigr|
\le \varepsilon p.
\]
\end{definition}

We first present an embedding lemma for hyperedges in regular partitions.




\begin{lemma}[Embedding matchings in regular partitions]\label{lem:distribution1_hygr}
Fix integers $2\le  t\le M$ and 
$d>10\varepsilon> 0$. Let $n$ be sufficiently large, $n^{-t}\ll p\leq 1$ and let $k=\lfloor n/M\rfloor$. 
Let $\cF=(V_1,\dots,V_M)$ be a $t$-graph on $n$ vertices with an equipartition (that $||V_i|-|V_j||\leq 1$). 
Let $\cX$ denote any collection of index sets $X=\{i_1,...,i_t\}\subset [M]$ for which  $(V_{i_1},...,V_{i_t})$ is $(\varepsilon,p)$-regular and has density at least $d p$. 
If there exists a collection of non-negative values
$\{w_X\}_{X\in\mathcal X}$
such that for every $i\in[M]$
\[
\sum_{X\in \mathcal X: i\in X} w_X
\le
1-2\varepsilon,
\]
then $\mathcal F$ contains a matching of size
\[
k\cdot\sum_{X\in\mathcal X}w_X-O(1).
\]
\end{lemma}


\begin{proof}
For every \(X\in\mathcal X\), set
\(m_X:=\lfloor kw_X\rfloor\), and process the members of
\(\mathcal X\) in an arbitrary order. For \(i\in[M]\), let
$D_i:=\sum_{\substack{X\in\mathcal X\\i\in X}}m_X.$
The capacity assumption gives
\[
D_i
\le
k\sum_{\substack{X\in\mathcal X\\i\in X}}w_X
\le
(1-2\varepsilon)k.
\]

Suppose that \(X=\{i_1,\ldots,i_t\}\) is the next tuple to be processed.
For \(i\in X\), let \(U_i\subseteq V_i\) be the set of currently unused
vertices, and write
$u_i:=|V_i\setminus U_i|.$
Choose \(i_0\in X\) such that
\(s:=|U_{i_0}|=\min_{i\in X}|U_i|\). Since every previously embedded edge
whose reduced tuple contains \(i_0\) uses exactly one vertex from
\(V_{i_0}\), the total remaining quota involving \(i_0\) is
\(D_{i_0}-u_{i_0}\). In particular,
\[
m_X
\le
D_{i_0}-u_{i_0}
\le
(1-2\varepsilon)k-u_{i_0}
\le
s-2\varepsilon k,
\]
where the last inequality uses \(|V_{i_0}|\ge k\).

For every \(i\in X\), choose a subset \(U_i'\subseteq U_i\) of size \(s\).
We greedily select disjoint edges from the \(t\)-partite hypergraph induced
by these sets. Suppose that a maximal such matching has size
\(m<m_X\). After removing its vertices, each class contains a set
\(W_i\subseteq U_i'\) of size
$|W_i|=s-m>2\varepsilon k.$
For sufficiently large \(n\), this is at least
\(\varepsilon|V_i|\). Since the original tuple indexed by \(X\) is
\((\varepsilon,p)\)-regular and has density at least \(dp\), the tuple
\((W_{i_1},\ldots,W_{i_t})\) has density at least
\((d-\varepsilon)p>0\). It therefore contains an edge, contradicting the
maximality of the matching. Hence we can embed all \(m_X\) prescribed
edges. Repeating this for every \(X\in\mathcal X\) produces one matching
in \(\mathcal F\).

Finally,
\[
\sum_{X\in\mathcal X}m_X
\ge
k\sum_{X\in\mathcal X}w_X-|\mathcal X|
\ge
k\sum_{X\in\mathcal X}w_X-\binom Mt.
\]
Since \(M\) and \(t\) are fixed, the last term is
\(O_{M,t}(1)\), as required.
\end{proof}



Next we present a multicolour  (weak) sparse regularity lemma, which is essentially  Lemma~3.1 in~\cite{GGMS}.


\begin{lemma}[multicolour  sparse regularity lemma]\label{hreg}
Let $0\le d\le D$ and
$\eta\ll 1/Z\ll 1/m\ll\varepsilon\ll 1/D,1/r,1/t$. Then the following holds for $n$ sufficiently large and $n^{-t}\ll p\leq 1$.  

Assume $\cF$ is an $(\eta,p,d,D)$-uniform $t$-graph equipped with an $r$-edge-colouring $\chi$. Then there exists an equipartition $V(\cF)=(V_1,..., V_{M})$, where $m\le M\le Z$ such that for all but at most $\varepsilon M^{t}$
$t$-tuples $X\in\binom{[M]}{t}$, the induced sub-hypergraph $\cF_\ell(V_X)$ is $(\varepsilon,p)$-regular for  any colour $\ell\in[r]$.
\end{lemma}

We remark that the embedding Lemma~\ref{lem:distribution1_hygr} will be applied to the $\ell$-colored $t$-graph $\mathcal G_\ell$, for each colour $\ell\in[r]$. Indeed,  we will choose \[
\mathcal X_\ell
:=
\left\{X\in\binom{[M]}{t}:\ \mathcal G_\ell[V_X]  \text{ is $(\varepsilon, p)$-regular with density at least $\frac{dp}{r}$}\right\}.
\]
For every non-exceptional $X\in\binom{[M]}{t}$, the clusters $V_i$, $i\in X$, have size at least $\eta n$, and hence the uniformity of $\cG$ gives
$\sum_{\ell=1}^r d_{\cG_\ell}(V_X)=d_{\cG}(V_X)\ge dp.$
Thus $d_{\cG_\ell}(V_X)\ge dp/r$ for some $\ell\in[r]$, and therefore
\[
\left|\binom{[M]}{t}\setminus\bigcup_{\ell=1}^r\mathcal X_\ell\right|
\le \varepsilon M^t.
\]
%When no confusion arises, we also denote $\mathcal X_\ell$ to be $\cG_\ell[\cX_\ell]$.







\subsection{Step I: Dual weights from reduced matchings}

Given an edge-colouring $\chi$ of an $(\eta,p,d,D)$-uniform $t$-graph $\mathcal G$, we can reduce  $\nu(\chi)$ to the set of solutions of $r$ linear programs, up to a small error. Our goal in this step is the following lemma.


\begin{lemma}\label{lem:chitoskeleton_hygr}
Let $0< d\le D$ and
$0<\eta\ll 1/Z\ll 1/m\ll\varepsilon \ll  d/r, 1/D,1/r,1/t$. Then the following holds for $n$ sufficiently large and $n^{-t}\ll p\leq 1$.  

Let $\mathcal G$ be an $(\eta,p,d,D)$-uniform $t$-graph on $n$ vertices equipped with an $r$-edge colouring $\chi$ and an equipartition $V(\cG)=(V_1,..., V_{M})$ given by Lemma~\ref{hreg}. Then for each colour $\ell \in [r]$, we have $$\nu (\mathcal G_\ell)\ge k\cdot Z_\ell-3\varepsilon n,$$ where $Z_\ell=\sum_{i=1}^M z_i$ is the optimal value of the linear program 
\begin{equation}\label{eq:DL_hygr}
\begin{aligned}
\text{minimize}\quad 
& \sum_{i=1}^M z_i\\
\text{subject to}\quad
& \sum_{i:i\in X} z_i \ge 1
\qquad\text{for every } X\in\mathcal X_\ell,\\
& z_i \ge 0
\qquad\text{for every } i\in[M].
\end{aligned}
\tag{$D_\ell$}
\end{equation}
\end{lemma}

We may choose an optimal solution with $z_i\le1$ for every $i\in[M]$, since replacing any coordinate larger than $1$ by $1$ preserves all constraints.



\begin{proof}
Consider the dual linear program for fractional vertex covers:
\begin{equation}\label{eq:PL_hygr}
\begin{aligned}
\text{maximize}\quad 
& \sum_{X\in\mathcal X_\ell} w_X\\
\text{subject to}\quad
& \sum_{\substack{X\in\mathcal X_\ell: i\in X}} w_X \le 1
\qquad\text{for every } i\in[M],\\
& w_X \ge 0
\qquad\text{for every } X\in\mathcal X_\ell .
\end{aligned}
\tag{$P_\ell$}
\end{equation}
Let $\{w_X\}_{X\in\mathcal X_\ell}$ be an 
optimal solution with optimal value $W_\ell$. By strong LP duality, $W_\ell=Z_\ell$.
Apply Lemma~\ref{lem:distribution1_hygr}, with density parameter $d/r$, to the weights $\{(1-2\varepsilon)w_X\}_{X\in\mathcal X_\ell}$. We obtain
\[
\nu(\cG_\ell)
\ge
(1-2\varepsilon)kW_\ell-O(1)
\ge
kW_\ell-3\varepsilon n
=
kZ_\ell-3\varepsilon n,
\]
where the last inequality holds for sufficiently large $n$.
\end{proof}




\subsection{Step II: Repair and vector compression}
Our goal is to replace the dual vectors by $M$ vectors in $\mathbb R^r$ taking at most $r$ distinct values, while preserving the covering constraints and controlling their total $\ell_\infty$-norm.

\begin{lemma}\label{lem:vertorrefine}
Let parameters be given in Lemma~\ref{lem:chitoskeleton_hygr}.
Then there exists a collection of vectors $\{Y_i\}_{i\in[M]}$, where $Y_i=(y_i^{(1)},\ldots,y_i^{(r)})$ for each $i\in[M]$ such that the following holds.
\begin{enumerate}
\item[(Y1)] For each $i\in[M]$ and $\ell\in [r]$, $0\le y_i^{(\ell)} \le 2t$. Moreover,
\[
\frac{k}{t}\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty
+
\left(
3\varepsilon^{1/(3tr)}
+2rt\varepsilon^{1/t}
+\frac{4r}{m}
\right)n,
\]
where $Z=(Z_1,...,Z_r).$

\item[(Y2)] For each $X\in\binom{[M]}{t}$, there exists a colour $\ell\in[r]$ such that 
$$\sum_{i\in X} y_i^{(\ell)}\geq t.$$

\item[(Y3)] $\{Y_i\}_{i\in[M]}$ takes at most $r$ distinct values in $\mathbb R^r$. And for every $i \in [M]$, there are at least $t\varepsilon^{1/t}M$ indices $j \in [M]$ such that $Y_j = Y_i$.
\end{enumerate}
\end{lemma}

We first repair the exceptional reduced $t$-sets in Lemma~\ref{lem:op1}, and then compress the repaired vectors in Lemma~\ref{lem:op2}.

For each colour $\ell\in [r]$, let $\{z^{(\ell)}_i\}_{i\in[M]}$ be a collection of weights given by Lemma~\ref{lem:chitoskeleton_hygr}.  For each $i\in [M]$ and $\ell\in [r]$, we set $s_i^{(\ell)}=t\cdot z_i^{(\ell)}(\le t)$ and
\begin{align}\label{eq:SSS}
S_i:=(s^{(1)}_i, s^{(2)}_i,..., s^{(r)}_i)    
\end{align}
to be the initial weight vector. Note that \begin{align}\label{t}
\sum_{i:i\in X}s_i^{(\ell)}=t\cdot\sum_{i:i\in X}z_i^{(\ell)}\geq t \ \text{ for at least one colour $\ell\in[r]$}
\end{align} holds for all but at most  $\varepsilon M^{t}$ sets $X\in\binom{[M]}{t}$, since $\left|\binom{[M]}{t}\setminus\bigcup_{\ell\in [r]}\mathcal X_\ell\right|\le \varepsilon M^{t}$.

To summarize, we have the following observation.

\begin{observation}
[Existence of weights]\label{lem:initialvalues}
 Let parameters be given in Lemma~\ref{lem:chitoskeleton_hygr}.
Then there exists a collection of vectors $\{S_i\}_{i\in[M]}$, where $S_i=(s_i^{(1)},\ldots,s_i^{(r)})$ for $i\in[M]$ such that the following holds.
\begin{enumerate}
\item[(i)] For each $i\in[M]$ and each $\ell\in [r]$, $0\le s_i^{(\ell)}\le t$.


\item[(ii)] For all but at most $\varepsilon M^t$ sets $X\in \binom{[M]}{t}$, there exists a colour $\ell\in [r]$ such that $\sum_{i\in X}s_i^{(\ell)}\geq t$.

\item[(iii)] For each $\ell\in[r]$, $\sum_{i=1}^M s_i^{(\ell)} = t Z_\ell$.
Consequently,
\[
\left\|\sum_{i=1}^M S_i\right\|_\infty
 \ =\
t\cdot \|Z\|_\infty.
\]
\end{enumerate}
\end{observation}

We construct $\{S'_i\}$ by adjusting $\{S_i\}$, such that the inequality~(ii) holds for all $X\in\binom{[M]}{t}$.

\begin{lemma}\label{lem:op1}
Let $S_1,\ldots,S_M\in[0,t]^r$ be defined above. Then there exist $S'_1,\ldots,S'_M\in[0,2t]^r$ such that the following holds.
\begin{enumerate}
\item For each $X\in\binom{[M]}{t}$, there exists a colour $\ell\in[r]$ such that 
$\sum_{i\in X} s_i'^{(\ell)}\geq t$.

\item For every $i \in [M]$, there are at least $t\varepsilon^{1/t}M$ indices $j \in [M]$ such that $S'_j = S'_i$.

\item The $\ell_\infty$ norm of ~$\sum S'_i$ satisfies
$$\bigg\|\sum_{i=1}^M S'_i\bigg\|_\infty
\leq\ \bigg\|\sum_{i=1}^M S_i\bigg\|_\infty +3M\cdot t\cdot \varepsilon^{1/(3tr)}.$$
\end{enumerate}
\end{lemma}
\begin{proof}
Set $\delta:=t\varepsilon^{1/(3tr)}$ and $L:=t\varepsilon^{1/t}M$.  We first replace every coordinate of $S_i$ by the smallest multiple of $\delta$ not smaller than it. For every $i\in[M]$, set the discretized vector $\tilde{S_i}:= \delta\cdot\left\lceil \frac{S_i}{\delta}\right\rceil$.

Note the vectors $\{\tilde{S_i}\}$ take at most $(2t/\delta)^r$ distinct values. Indeed, the set 
$$C=\{0,\delta,2\delta,...,\lceil t/\delta\rceil\delta\}\subset [0,2t]$$ 
 contains at most $\lceil t/\delta\rceil+1\leq (2 t/\delta)$ values, and $\tilde{S_i}\in C^r$. Moreover,
\begin{align*}
\bigg\|\sum_{1\leq i\leq M} \tilde{S}_i\bigg\|_\infty\ \le \ \bigg\|\sum_{1\leq i\leq M} S_i \bigg\|_\infty + \sum_{1\leq i\leq M} \|\tilde{S}_i-S_i\|_\infty\leq \bigg\|\sum_{1\leq i\leq M} S_i \bigg\|_\infty+ M\delta, 
\end{align*}
since $\tilde{s}^{(\ell)}_i\ge s^{(\ell)}_i$ for $i\in[M]$, $\ell\in [r]$ and $\|\tilde S_i-S_i\|_\infty \le \delta$, where $\tilde{s}^{(\ell)}_i$ is the $\ell$-th coordinate of $\tilde{S}_i$. 

We next construct $\{S'_i\}$ by selecting the vectors that appear sufficiently many times among $\{\tilde{S}_i\}$. We keep these frequently occurring vectors and use their values to replace the vectors that appear only a few times. Set
$$\cS:=\bigl\{S\in\{\tilde{S}_i\}_{i\in[M]}:\ |\{i\in[M]: \tilde{S}_i=S\}|\ge L\bigr\}.$$
For sufficiently small $\varepsilon$, we have $(2t/\delta)^rL<M$, so $\cS$ is non-empty. We define $\{S'_i\}$ as follows: For each $i\in [M]$, if $\tilde{S}_i\in \cS$, then set $S'_i:=\tilde{S}_i$; otherwise choose an arbitrary $S\in \cS$ and set $S'_i:=S$. Then (2) holds directly for $\{S'_i\}$, and it remains to check (1), (3).

For (1), suppose that some $X\in\binom{[M]}{t}$ satisfies
\[
\sum_{i\in X}s_i'^{(\ell)}<t
\qquad\text{for every }\ell\in[r].
\]
Let $R_1,\ldots,R_h$ be the distinct values in the multiset $\{S_i':i\in X\}$, with multiplicities $a_1,\ldots,a_h$, where $a_1+\cdots+a_h=t$. For each $j\in[h]$, let
\[
I_j:=|\{u\in[M]:\widetilde S_u=R_j\}|.
\]
Since every $R_j$ is frequent, $I_j\ge L$. Every $t$-set whose multiset of rounded values consists of $a_j$ copies of $R_j$ for each $j$ also violates the covering condition for the original vectors, because $S_u\le\widetilde S_u$ coordinatewise. The number of such $t$-sets is
\[
\prod_{j=1}^h\binom{I_j}{a_j}
\ge
\prod_{j=1}^h\binom{\lfloor L\rfloor}{a_j}
=
(1+o(1))\frac{L^t}{\prod_{j=1}^h a_j!}
>
\varepsilon M^t.
\]
Indeed, $L=t\varepsilon^{1/t}M$ and
$t^t/\prod_j a_j!\ge t^t/t!>1$. This contradicts Observation~\ref{lem:initialvalues}(ii).

For (3), the number of indices $i\in[M]$ for which $\widetilde S_i\notin\cS$ is at most
$\left(\frac{2t}{\delta}\right)^rL
\le
\varepsilon^{1/(2t)}M,$
provided that $\varepsilon$ is sufficiently small. Hence
\[
\max_{\ell\in[r]}
\sum_{i=1}^M
|s_i'^{(\ell)}-\widetilde s_i^{(\ell)}|
\le
2t\varepsilon^{1/(2t)}M.
\]
Together with the preceding estimate, this gives
\begin{align*}
\left\|\sum_{i=1}^M S'_i\right\|_\infty
&\le
\left\|\sum_{i=1}^M\widetilde S_i\right\|_\infty
+
2t\varepsilon^{1/(2t)}M\\
&\le
\left\|\sum_{i=1}^M S_i\right\|_\infty
+M\delta+2t\varepsilon^{1/(2t)}M\\
&\le
\left\|\sum_{i=1}^M S_i\right\|_\infty
+3Mt\varepsilon^{1/(3tr)},
\end{align*}
as required.
\end{proof}


We next present a reduction lemma for vectors in $\mathbb{R}^r$.


\begin{lemma}\label{lem:op2}
Let $M\ge r,t\geq 1$\footnote{We allow \(t=1\) here because the lemma will be applied again in the proof of Theorem~\ref{thm:graphstructure}. Note also that the case \(r=1\) is trivial.} and let $\{X_1,...,X_M\}\subset [0,2t]^r$ be a collection of $r$-dimensional vectors. Then there exist $r$ vectors $W_1,...,W_r\in \{X_1,...,X_M\}$ and integers $m_1,...,m_r\geq 0$, $m_1+...+m_r=M$ such that
$$\bigg\|\sum_{1\leq j\leq r} m_j W_j\bigg\|_\infty\ \leq\  \bigg\|\sum_{1\leq i\leq M} X_i\bigg\|_\infty\ +\ 4rt.$$
\end{lemma}



We will use the following classical result of Carathéodory concerning
faces of polytopes in $\mathbb{R}^r$; see, for instance, \cite{ziegler} for a reference.

\begin{theorem}[Carathéodory's Theorem for Faces]\label{thm:Cara}
Let $P \subseteq \mathbb{R}^r$ be a polytope with vertex set $V$.
For any $x \in \partial P$, there exist 
$v_1,\dots,v_r \in V$ and coefficients $\lambda_1,\dots,\lambda_r \ge 0$
such that
\[
x = \sum_{i=1}^r \lambda_i v_i,
\qquad
\sum_{i=1}^r \lambda_i = 1.
\]
\end{theorem}


\begin{proof}[Proof of Lemma~\ref{lem:op2}]
Denote $\cX=\{X_1,...,X_M\}$, and let $P:=\operatorname{conv}(\cX)\subseteq[0,2t]^r$ be the convex hull of $\cX\subset\mathbb{R}^r$. Set $\overline{X}=(\sum X_i)/M$, and then $\overline{X}\in P$. 

Set
\[
\alpha_*:=\max\{\alpha\ge0:\overline X-\alpha\mathbbm{1}_r\in P\},
\qquad
J:=\overline X-\alpha_*\mathbbm{1}_r.
\]
The maximum exists because $P$ is compact, and $J\in\partial P$. Since $J\in[0,2t]^r$ and $J\le\overline X$ coordinatewise,
\[
\|J\|_\infty\le\|\overline X\|_\infty.
\]
By Theorem~\ref{thm:Cara}, there exist $W_1,...,W_r\in \cX$ and coefficients $\lambda_1,...,\lambda_r\geq0$ such that
$$ J=\sum_{1\leq j\leq r}\lambda_j W_j\qquad\text{and}\qquad \sum_{1\leq j\leq r} \lambda_j=1.$$
Set $m_j:=\lfloor\lambda_j M\rfloor$ for each $1\leq j\leq r-1$, and $m_r:=M-\sum_{j\leq r-1} m_j$. Note then $m_j\in \mathbb{Z}_{\ge 0}$ and $\sum_{j=1}^r m_j=M$.

Denote $J':=(\sum m_j W_j)/M$; then 
\begin{align*}
\|J'-J\|_\infty\leq \sum_{j\leq r}\Big|\lambda_j-\frac{m_j}{M}\Big|\cdot \|W_j\|_\infty
\end{align*}
Note that $\|W_j\|_\infty\leq 2t$, the difference $|\lambda_j-m_j/M|\leq 1/M$ for $1\leq j\leq r-1$ and $$\Big|\lambda_r-\frac{m_r}{M}\Big|=\Big|\sum_{j\leq r-1}\Big(\frac{m_j}{M}-\lambda_j\Big)\Big|\leq \frac{r-1}{M}.$$
Hence we have
$$\|J'-J\|_\infty\leq 2t\cdot\left(\frac{r-1}{M}+\frac{r-1}{M}\right)\leq \frac{4rt}{M}.$$
Therefore by multiplying by $M$,
$$\bigg\|\sum_{j\leq r} m_j W_j\bigg\|_\infty\leq M\cdot (\|J\|_\infty + \|J'-J\|_\infty)\leq M \|J\|_\infty + 4rt\leq\bigg\|\sum_{1\leq i\leq M} X_i\bigg\|_\infty\ +\ 4rt, $$
as required.
\end{proof}


Finally, gathering Lemmas~\ref{lem:initialvalues}, \ref{lem:op1} and \ref{lem:op2} together, we are ready to prove Lemma~\ref{lem:vertorrefine} as follows.


\begin{proof}[Proof of Lemma~\ref{lem:vertorrefine}]
Let $S_1,\ldots,S_M$ be the vectors from Observation~\ref{lem:initialvalues}, and let $S'_1,\ldots,S'_M$ be given by Lemma~\ref{lem:op1}. Apply Lemma~\ref{lem:op2} to $S'_1,\ldots,S'_M$. We obtain vectors $W_1,\ldots,W_r\in\{S'_1,\ldots,S'_M\}$ and integers $m_1,\ldots,m_r\ge0$ with $\sum_jm_j=M$ and
\[
\left\|\sum_{j=1}^r m_jW_j\right\|_\infty
\le
\left\|\sum_{i=1}^M S'_i\right\|_\infty+4rt.
\]
Put $L:=t\varepsilon^{1/t}M$. Since $rt\varepsilon^{1/t}<1$, some index $j_0$ satisfies $m_{j_0}\ge L$. Define
\[
\widehat W_j:=
\begin{cases}
W_j,&m_j\ge L,\\
W_{j_0},&m_j<L.
\end{cases}
\]
Then
\[
\left\|\sum_{j=1}^r m_j\widehat W_j-
\sum_{j=1}^r m_jW_j\right\|_\infty
\le
2t\sum_{j:m_j<L}m_j
\le
2rt^2\varepsilon^{1/t}M.
\]
Let $Y_1,\ldots,Y_M$ be the multiset containing $m_j$ copies of $\widehat W_j$ for each $j$. Every distinct value occurring among the $Y_i$ occurs at least $L$ times, so \textnormal{(Y3)} holds.

For any $X\in\binom{[M]}{t}$, the multiset $\{Y_i:i\in X\}$ can be realised by $t$ distinct indices in $\{S'_1,\ldots,S'_M\}$: every value used by the $Y_i$ is a value of some $S'_j$, and Lemma~\ref{lem:op1}(2) says that it occurs at least $L\ge t$ times there. Lemma~\ref{lem:op1}(1) therefore gives a colour $\ell\in[r]$ such that
$\sum_{i\in X}y_i^{(\ell)}\ge t,$
which proves \textnormal{(Y2)}.

Finally,
\begin{align*}
\left\|\sum_{i=1}^M Y_i\right\|_\infty
&\le
\left\|\sum_{i=1}^M S'_i\right\|_\infty
+2rt^2\varepsilon^{1/t}M+4rt\\
&\le
\left\|\sum_{i=1}^M S_i\right\|_\infty
+3Mt\varepsilon^{1/(3tr)}
+2rt^2\varepsilon^{1/t}M+4rt.
\end{align*}
By Observation~\ref{lem:initialvalues}(iii),
$\frac{k}{t}\left\|\sum_{i=1}^M S_i\right\|_\infty
=k\|Z\|_\infty.$
Using $Mk\le n$ and $k\le n/m$, we obtain
\[
\frac{k}{t}\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty
+
\left(
3\varepsilon^{1/(3tr)}
+2rt\varepsilon^{1/t}
+\frac{4r}{m}
\right)n,
\]
which proves \textnormal{(Y1)}.
\end{proof}



\subsection{Step III: Constructing strip colouring via adjusted vectors}
In this step, with the vectors $\{Y_i\}_{i\in[M]}$ in hand, our goal is to construct a strip colouring $\psi$ of the complete $t$-graph $K_n^{(t)}$ whose maximum monochromatic matching is controlled by the dual weights. More precisely, we prove the following lemma.

\begin{lemma}\label{lem:hygrrecon}
Let $\{Y_i\}_{i\in[M]}$ be a family of vectors satisfying \textnormal{(Y1)--(Y3)} in Lemma~\ref{lem:vertorrefine}. Let
\(
V(K_n^{(t)})=V_1\cup\cdots\cup V_M
\)
be an equipartition with $|V_i|\in\{k,k+1\}$, where $k=\lfloor n/M\rfloor$.  Then there exists an $r$-strip $r$-edge-colouring $\psi$ of $K_n^{(t)}$ such that
\[
\nu(\psi)
\le
\frac{k}{t}\cdot\bigg\|\sum_{i=1}^M Y_i\bigg\|_\infty
+
O(1).
\]
\end{lemma}

\begin{proof}
Let
\(
R_1,\ldots,R_q
\)
be the distinct values taken by the vectors $\{Y_i\}_{i\in[M]}$, where $q\le r$. For each $j\in[q]$, set
\[
\mathcal I_j:=\{i\in[M]:Y_i=R_j\},
\qquad
U_j:=\bigcup_{i\in\mathcal I_j}V_i.
\]
By \textnormal{(Y3)}, each value class satisfies
\(
|\mathcal I_j|\ge t\varepsilon^{1/t}M\ge t.
\)
We now define an edge-colouring $\psi$ of $K_n^{(t)}$. For every vector
\[
\mathbf a=(a_1,\ldots,a_q)\in\mathbb Z_{\ge0}^q
\qquad\text{with}\qquad
a_1+\cdots+a_q=t,
\]
choose a set
\[
X(\mathbf a)\in\binom{[M]}{t}
\]
by selecting exactly \(a_j\) indices from \(\mathcal I_j\) for each \(j\in[q]\). This is possible since \(|\mathcal I_j|\ge t\) for every \(j\).

By \textnormal{(Y2)}, there exists a colour \(\ell(\mathbf a)\in[r]\) such that
$\sum_{i\in X(\mathbf a)}y_i^{(\ell(\mathbf a))}\ge t.$
Since $Y_i=R_j$ for every $i\in\mathcal I_j$, this is equivalent to
\[
\sum_{j=1}^q a_jR_j^{(\ell(\mathbf a))}\ge t.
\]
If several colours satisfy this inequality, choose the smallest one. Now let \(e\in E(K_n^{(t)})\). Given the partition $V(K_n^{(t)})=U_1\cup\cdots\cup U_q$, set
\[
a_j(e):=|e\cap U_j|,
\quad j\in[q];
\qquad
\text{ and } \qquad
\psi(e):=\ell(a_1(e),\ldots,a_q(e)).
\]
Then the colour of \(e\) depends only on its intersection sizes with the parts \(U_1,\ldots,U_q\). Hence \(\psi\) is a \(q\)-strip colouring, and since \(q\le r\), it is also an \(r\)-strip colouring.
 
  We next bound the maximum monochromatic matching in \(\psi\).
We first observe that for every edge \(e\in E(K_n^{(t)})\) with
\(\psi(e)=\ell\), one has
\[
\sum_{v\in e} y_{\operatorname{part}(v)}^{(\ell)}\ge t,
\]
where \(\operatorname{part}(v)=i\) if \(v\in V_i\). 

Now let \(\mathcal M\) be an arbitrary monochromatic matching in \(\psi\),
say of colour \(\ell\). Then
\[
t|\mathcal M|
\le
\sum_{e\in\mathcal M}\sum_{v\in e}
y_{\operatorname{part}(v)}^{(\ell)}=
\sum_{i=1}^M |V_i\cap V(\mathcal M)| \cdot y_i^{(\ell)}
\le
(k+1)\sum_{i=1}^M y_i^{(\ell)}
\le
(k+1)\bigg\|\sum_{i=1}^M Y_i\bigg\|_\infty .
\]
Therefore
\[
|\mathcal M|
\le
\frac{k+1}{t}\bigg\|\sum_{i=1}^M Y_i\bigg\|_\infty
\le
\frac{k}{t}\bigg\|\sum_{i=1}^M Y_i\bigg\|_\infty+O(1),
\]
where the last term is $O(1)$ because $M$ is bounded and $Y_i\in[0,2t]^r$.
 Taking the maximum over all monochromatic matchings
\(\mathcal M\) gives
\[
\nu(\psi)
\le
\frac{k}{t}\bigg\|\sum_{i=1}^M Y_i\bigg\|_\infty+O(1),
\]
as required.
\end{proof}




\subsection{Proof of Theorem~\ref{thm:hypergraphstructure}}

Combining Lemma~\ref{lem:chitoskeleton_hygr} from Step~I,
Lemma~\ref{lem:vertorrefine} from Step~II, and Lemma~\ref{lem:hygrrecon} from Step~III,
we are ready to prove Theorem~\ref{thm:hypergraphstructure}.

\begin{proof}[Proof of Theorem~\ref{thm:hypergraphstructure}]
Fix $\varepsilon>0$. Choose $\rho>0$ and then $m$ sufficiently large so that
$3\rho
+3\rho^{1/(3tr)}
+2rt\rho^{1/t}
+\frac{4r}{m}
<\frac{\varepsilon}{2}.$
Choose the remaining parameters according to
\[
0<\eta\ll 1/Z\ll 1/m\ll\rho\ll d/r,1/D,1/r,1/t.
\]
Apply Lemma~\ref{hreg}, with $\rho$ in place of its regularity parameter, to obtain an equipartition
$V(\cG)=V_1\cup\cdots\cup V_M,$ $
m\le M\le Z,$
and put $k=\lfloor n/M\rfloor$.

Lemma~\ref{lem:chitoskeleton_hygr}, applied with $\rho$ in place of $\varepsilon$, gives a vector $Z=(Z_1,\ldots,Z_r)$ such that
\[
\nu(\chi)
=
\max_{\ell\in[r]}\nu(\cG_\ell)
\ge
k\|Z\|_\infty-3\rho n.
\]
By Lemma~\ref{lem:vertorrefine}, again with $\rho$ in place of $\varepsilon$, there are vectors $Y_1,\ldots,Y_M$ satisfying \textnormal{(Y1)--(Y3)} and
\[
\frac{k}{t}\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty
+
\left(
3\rho^{1/(3tr)}
+2rt\rho^{1/t}
+\frac{4r}{m}
\right)n.
\]
Finally, Lemma~\ref{lem:hygrrecon} gives an $r$-strip $r$-edge-colouring $\psi$ of $K_n^{(t)}$ such that
\[
\nu(\psi)
\le
\frac{k}{t}\left\|\sum_{i=1}^M Y_i\right\|_\infty+O(1).
\]
Combining these inequalities and taking $n$ sufficiently large, we obtain
\begin{align*}
\nu(\chi)
&\ge
\nu(\psi)
-
\left(
3\rho
+3\rho^{1/(3tr)}
+2rt\rho^{1/t}
+\frac{4r}{m}
\right)n
-O(1)\\
&\ge
\nu(\psi)-\varepsilon n.\qedhere
\end{align*}
\end{proof}


\section{Applications: asymptotic AFL, transference, and stable Kneser hypergraphs}
\label{sec:hypapp}

We now apply Theorem~\ref{thm:hypergraphstructure} to the complete
hypergraph, sparse random hypergraphs, and the hypergraph of stable
\(t\)-sets.


\subsection{A proof of the asymptotic Alon--Frankl--Lov\'asz theorem}
\label{subsec:asyAFL}


The standard AFL colouring, consisting of one large part and \(r-1\)
small parts, shows that the coefficient \(1/(r+t-1)\) is best possible up to
an additive constant. It remains to prove the lower bound.

Theorem~\ref{thm:hypergraphstructure} reduces the problem to the
following elementary profile-packing statement.

\begin{lemma}\label{lem:aaalower}
Let \(r,t\geq2\). Every \(r\)-strip \(r\)-edge-colouring \(\varphi\) of
\(K_n^{(t)}\) satisfies
\[
\nu(\varphi)\geq
\left\lfloor\frac{n}{t+r-1}\right\rfloor .
\]
\end{lemma}

\begin{proof}
Set
$m:=\left\lfloor\frac{n}{t+r-1}\right\rfloor.$ If \(m=0\), the assertion is trivial. Hence assume \(m\ge1\).
Let
$V(K_n^{(t)})=U_1\cup\cdots\cup U_r,$
be a strip partition for \(\varphi\), and define
\[
c_i:=\left\lfloor\frac{|U_i|}{m}\right\rfloor
\qquad (i\in[r]).
\]
Since \(m\leq n/(t+r-1)\),
$\sum_{i=1}^r c_i
>
\frac{n}{m}-r
\geq t+r-1-r
\geq t-1.$
The left-hand side is an integer, so
\[
\sum_{i=1}^r c_i\geq t.
\]
Consequently, there exist integers \(a_1,\ldots,a_r\geq0\) such that
\[
a_1+\cdots+a_r=t
\qquad\text{and}\qquad
a_i\leq c_i
\quad\text{for every }i\in[r].
\]
In particular, \(a_i m\leq |U_i|\).

For every \(i\in[r]\), choose \(m\) pairwise disjoint subsets
$A_{i,1},\ldots,A_{i,m}\subseteq U_i,$
with $|A_{i,j}|=a_i.$
Then
$e_j:=\bigcup_{i=1}^r A_{i,j},$
$j\in[m],$
are \(m\) pairwise vertex-disjoint \(t\)-edges with the same profile
\((a_1,\ldots,a_r)\). Hence they all receive the same colour under
\(\varphi\), proving the lemma.
\end{proof}



\begin{proof}[Proof of Corollary~\ref{cor:intro-asyAFL}]
Fix arbitrary \(\delta>0\). The complete \(t\)-graph is
\((\eta,1,1/2,2)\)-uniform for every \(\eta>0\), so
Theorem~\ref{thm:hypergraphstructure} gives an \(r\)-strip colouring
\(\psi\) of \(K_n^{(t)}\) such that
$\nu(\chi)\geq\nu(\psi)-\delta n.$
By Lemma~\ref{lem:aaalower},
$\nu(\chi)
\geq
\left\lfloor\frac{n}{t+r-1}\right\rfloor-\delta n.$
\end{proof}


\medskip


\subsection{A proof of the transference AFL theorem}
\label{subsec:transAFL}
We next deduce the transference AFL theorem of Gishboliner, Glock,
Michaeli, and Sgueglia~\cite{GGMS} from the structural theorem.

\begin{theorem}[Gishboliner, Glock, Michaeli and Sgueglia]
\label{thm:randomAFL}
Let \(t,r\geq2\), let \(\varepsilon>0\), and let
\(G\sim G^{(t)}(n,p)\), where every edge is present independently with
probability \(p=p(n)\). There exists \(C>0\) such that, if
\(p\geq Cn^{-t+1}\), then with high probability every \(r\)-edge-colouring
\(\chi\) of \(G\) satisfies
\[
\nu(\chi)\geq
(1-\varepsilon)\frac{n}{r+t-1}.
\]
\end{theorem}


To verify the required uniformity, we use Lemma~3.2 of~\cite{GGMS}.

\begin{lemma}\label{lem:asdf}
Let \(1/C\ll\eta,1/t\), where \(t\geq2\), and let
\(G\sim G^{(t)}(n,p)\) with \(p\geq Cn^{-t+1}\). Then with high probability,
for every collection of pairwise disjoint sets
\(X_1,\ldots,X_t\) satisfying \(|X_i|\geq\eta n\), we have
\[
\frac p2
\leq
\frac{e(X_1,\ldots,X_t)}
{|X_1|\cdots |X_t|}
\leq
\frac{3p}{2},
\]
where \(e(X_1,\ldots,X_t)\) is the number of \(t\)-edges containing exactly
one vertex from each \(X_i\).
\end{lemma}

\begin{proof}[Proof of Theorem~\ref{thm:randomaaa0}]
Fix \(\varepsilon>0\), and let \(\eta>0\) be supplied by
Theorem~\ref{thm:hypergraphstructure} with \(d=1/2\) and \(D=3/2\).
Choose \(C\) sufficiently large for Lemma~\ref{lem:asdf}. On the
high-probability event in that lemma, \(G\) is
\((\eta,p,1/2,3/2)\)-uniform, where \(p\) is the actual edge probability of
\(G\). Theorem~\ref{thm:hypergraphstructure} therefore gives the required
\(r\)-strip colouring \(\psi\) satisfying
$\nu(\chi)\geq\nu(\psi)-\varepsilon n.$
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:randomAFL}]
Apply Theorem~\ref{thm:randomaaa0} with
$\delta:=\frac{\varepsilon}{2(r+t-1)}$. With high probability, every colouring
\(\chi\) of \(G\) admits an \(r\)-strip colouring \(\psi\) of
\(K_n^{(t)}\) such that
$\nu(\chi)\geq\nu(\psi)-\delta n.$
Lemma~\ref{lem:aaalower} gives
$\nu(\psi)\geq
\left\lfloor\frac{n}{r+t-1}\right\rfloor.$
Consequently, for large \(n\),
$\nu(\chi)
\geq
\frac{n}{r+t-1}-1-\delta n
\geq
(1-\varepsilon)\frac{n}{r+t-1},$
as required.
\end{proof}


\medskip


\subsection{The chromatic number of stable Kneser hypergraphs}
\label{sec:stableKne}
Recall that a subset \(X\subseteq[n]\) is \emph{\(s\)-stable} if
$s\leq |i-j|\leq n-s$
for all distinct $i,j\in X.$
The hypergraph
\(\mathrm{KG}^q(n,t)_{s\text{-stab}}\) is the subhypergraph induced by the
\(s\)-stable \(t\)-subsets.

Let \(\mathcal S\) be the \(t\)-graph on \([n]\) whose edges are the
\(s\)-stable \(t\)-sets. A vertex-colouring of
\(\mathrm{KG}^q(n,t)_{s\text{-stab}}\) is equivalently an edge-colouring of
\(\mathcal S\), and a monochromatic matching of size \(q\) in
\(\mathcal S\) is a monochromatic hyperedge of the stable Kneser hypergraph.

\begin{lemma}\label{lem:stable}
Let \(t\geq2\) and \(0<c\leq1/(2t)\). Assume that \(n\) is sufficiently
large and set \(q=cn\). For every function \(s=s(n)\) with
\(2\leq s\ll q\), let \(\mathcal S\) be the \(t\)-graph on \([n]\) whose
edges are the \(s\)-stable \(t\)-sets. Set
\[
r=
\left\lfloor
\frac{n-(t-1)q}{q-1}
\right\rfloor-1.
\]
Then every \(r\)-edge-colouring \(\chi\) of \(\mathcal S\) satisfies
\[
\nu(\chi)
\geq
\left(\frac{c}{1-c}-o(1)\right)n>q.
\]
\end{lemma}

\begin{proof}
Set $\rho:=\frac1c-(t-1).$
Since \(q=cn\), we have
$\frac{n-(t-1)q}{q-1}
=
\frac{q\bigl(1/c-(t-1)\bigr)}{q-1}
=
\rho+\frac{\rho}{q-1}.$
As \(q\to\infty\), it follows that
$\left\lfloor
\frac{n-(t-1)q}{q-1}
\right\rfloor
=
\lfloor\rho\rfloor$
for all sufficiently large \(n\). This also holds when \(\rho\) is an
integer, since then
\(\rho<\rho+\rho/(q-1)<\rho+1\) for sufficiently large \(q\).
Consequently,
\[
r=
\left\lfloor\frac1c-(t-1)\right\rfloor-1
\]
is fixed for all sufficiently large \(n\).

Fix \(\gamma>0\), and let \(\eta>0\) be given by
Theorem~\ref{thm:hypergraphstructure} for the fixed parameters
$r, t, d=\frac12, D=2,$
and error \(\gamma n\). We claim that, for sufficiently large \(n\),
\(\mathcal S\) is \((\eta,1,1/2,2)\)-uniform.

Indeed, let \(X_1,\ldots,X_t\subseteq[n]\) be pairwise disjoint sets with
\(|X_i|\ge\eta n\). For fixed \(1\le i<j\le t\), the number of crossing
tuples
$(x_1,\ldots,x_t)\in X_1\times\cdots\times X_t$
for which \(x_i\) and \(x_j\) have cyclic distance less than \(s\) is at
most
\[
2s\,|X_i|\prod_{k\ne i,j}|X_k|
\le
\frac{2s}{\eta n}\prod_{k=1}^t|X_k|.
\]
A union bound over the \(\binom t2\) pairs gives
\[
e_{\mathcal S}(X_1,\ldots,X_t)
\ge
\left(
1-\binom t2\frac{2s}{\eta n}
\right)
\prod_{i=1}^t|X_i|.
\]
Since \(s\ll q=cn\), we have \(s/n=o(1)\). Hence, for sufficiently large
\(n\),
\[
\frac12\prod_{i=1}^t|X_i|
\le
e_{\mathcal S}(X_1,\ldots,X_t)
\le
\prod_{i=1}^t|X_i|,
\]
proving the claimed uniformity.

Theorem~\ref{thm:hypergraphstructure} now gives an \(r\)-strip colouring
\(\psi\) of \(K_n^{(t)}\) such that
$\nu(\chi)\ge\nu(\psi)-\gamma n.$
By Lemma~\ref{lem:aaalower},
$\nu(\psi)
\ge
\left\lfloor\frac{n}{t+r-1}\right\rfloor.$
Write
$A:=
\frac{n-(t-1)q}{q-1}.$
Since \(r=\lfloor A\rfloor-1\),
$t+r-1
=
t+\lfloor A\rfloor-2
\le
t+A-2.$
Therefore
\[
\nu(\psi)
\ge
\frac{n}{t+A-2}-1=
\frac{n(q-1)}{n-q-t+2}-1=
\frac{n(cn-1)}{(1-c)n-t+2}-1=
\frac{c}{1-c}n-O_{c,t}(1).
\]
It follows that
\[
\nu(\chi)
\ge
\left(\frac{c}{1-c}-\gamma\right)n-O_{c,t}(1).
\]

Finally,
$\frac{c}{1-c}-c
=
\frac{c^2}{1-c}>0.$
Choosing \(\gamma\) smaller than this fixed gap shows that, for sufficiently
large \(n\),
$\nu(\chi)>cn=q.$
\end{proof}


\begin{proof}[Proof of Theorem~\ref{thm:stable}]
The upper bound follows from the upper bound on
\(\chi(\mathrm{KG}^q(n,t))\), since
\(\mathrm{KG}^q(n,t)_{s\text{-stab}}\) is a subhypergraph of
\(\mathrm{KG}^q(n,t)\).

For the lower bound, set
\[
r=
\left\lfloor
\frac{n-(t-1)q}{q-1}
\right\rfloor-1.
\]
It suffices to show that no vertex-colouring of
\(\mathrm{KG}^q(n,t)_{s\text{-stab}}\) using at most \(r\) colours is
proper. Equivalently, every \(r\)-edge-colouring of \(\mathcal S\) must
contain a monochromatic matching with at least \(q\) edges. This follows
from Lemma~\ref{lem:stable}.
\end{proof}



\section{Structural reduction for graph tilings}
\label{sec:graphstructure}

In this section, we prove Theorem~\ref{thm:graphstructure}. Throughout,
\(H\) is a fixed connected\footnote{We believe that the proof in this section works for an arbitrary graph $H$ with $m_2(H)>1$.  We impose the connectedness assumption here only because tiling problems concern vertex-disjoint copies, and it is therefore more natural to take $H$ to be connected.} graph with
\[
h:=v(H)\ge3.
\] 



\subsection{Preliminaries: reduced \(H\)-patterns}

We use the following standard consequence of the K{\L}R embedding theorem;
see~\cite{CGSS}.

\begin{lemma}[Partite K{\L}R embedding for \(H\)]
\label{lem:HpatternKLR}
Let \(H\) be a connected graph with \(v(H)\ge3\). For every \(d>0\),
there exists \(\varepsilon_{\mathrm{KLR}}>0\) such that, for every
\(\eta>0\), there exists \(C_1>0\) with the following property.
If \(G\sim G(n,p)\) with
$p\ge C_1n^{-1/m_2(H)},$
then with high probability the following holds. 

Let \(F\subseteq G\), and
let \(\{U_x:x\in V(H)\}\) be pairwise disjoint sets satisfying
$|U_x|\ge\eta n$
for every $x\in V(H).$
If, for every edge \(xy\in E(H)\), the pair \(F[U_x,U_y]\) is
\((\varepsilon_{\mathrm{KLR}},p)\)-regular
and has density at least \(dp\), then \(F\) contains a canonical copy of
\(H\), with each \(x\in V(H)\) embedded in \(U_x\).
\end{lemma}


Let \((V_1,\ldots,V_M)\) be an
\((\varepsilon_{\mathrm{reg}},p)\)-regular
equipartition of an \(r\)-edge-coloured graph \(G\). For each colour
\(\ell\in[r]\), let
$\mathcal X_\ell\subseteq\binom{[M]}2$
be the collection of pairs \(\{i,j\}\) for which
\(G_\ell[V_i,V_j]\) is
\((\varepsilon_{\mathrm{reg}},p)\)-regular
and has density at least \(p/(2r)\).

A map
$\phi:V(H)\to[M]$
is called a \emph{colour-\(\ell\) reduced \(H\)-pattern} if, for every
\(xy\in E(H)\), we have \(\phi(x)\ne\phi(y)\) and
\[
\{\phi(x),\phi(y)\}\in\mathcal X_\ell.
\]
Non-adjacent vertices of \(H\) may be mapped to the same cluster. We write
\(\Phi_\ell\) for the family of all colour-\(\ell\) reduced \(H\)-patterns.
For \(\phi\in\Phi_\ell\), define
\[
c_i(\phi):=|\phi^{-1}(i)|
\qquad (i\in[M]).
\]


The next lemma lifts a feasible fractional packing of reduced
\(H\)-patterns to an actual \(H\)-tiling.

\begin{lemma}[Lifting fractional reduced \(H\)-tilings]
\label{lem:embeddinggraph}
Fix \(r\ge2\), \(Z\in\mathbb N\), and \(\gamma>0\). There exist
\(\varepsilon_0>0\) and \(C>0\) such that, for
\(G\sim G(n,p)\) with
$p\ge Cn^{-1/m_2(H)},$
the following holds with high probability.

Let \(F\subseteq G\), and let
$V(F)=V_1\cup\cdots\cup V_M,$
$M\le Z,$
be an equipartition, where \(k=\lfloor n/M\rfloor\). Let
\(\mathcal X\subseteq\binom{[M]}2\) be a collection of pairs such that
\(F[V_i,V_j]\) is \((\varepsilon_0,p)\)-regular with density at least
\(p/(2r)\) for every \(\{i,j\}\in\mathcal X\). Let \(\Phi\) be a family
of maps \(\phi:V(H)\to[M]\) such that
$\{\phi(x),\phi(y)\}\in\mathcal X$
for every $\phi\in\Phi$
and every $xy\in E(H).$
If non-negative weights \(\{x_\phi\}_{\phi\in\Phi}\) satisfy
\[
\sum_{\phi\in\Phi}c_i(\phi)x_\phi\le1-\gamma
\qquad\text{for every }i\in[M],
\]
then \(F\) contains an \(H\)-tiling of size at least
\[
k\sum_{\phi\in\Phi}x_\phi-O_{H,Z}(1).
\]
\end{lemma}

\begin{proof}
Apply Lemma~\ref{lem:HpatternKLR} with \(d=1/(4r)\), and let
\(\varepsilon_{\mathrm{KLR}}>0\) be the resulting constant. Choose
\(\varepsilon_0>0\) sufficiently small that every restriction of an
\((\varepsilon_0,p)\)-regular pair to subsets of relative size at least
\(\gamma/(3h)\) is
\((\varepsilon_{\mathrm{KLR}},p)\)-regular and has density at least
\(p/(4r)\). This follows from the slicing lemma. Set
$\eta_0:=\frac{\gamma}{10hZ},$
and condition on the high-probability event in
Lemma~\ref{lem:HpatternKLR} with \(\eta=\eta_0\).

For each \(\phi\in\Phi\), let
$m_\phi:=\lfloor kx_\phi\rfloor.$
We process the patterns one by one and greedily embed \(m_\phi\) canonical
copies of type \(\phi\). The total number of vertices requested from a
cluster \(V_i\) is at most
\[
\sum_{\phi\in\Phi}c_i(\phi)m_\phi
\le
k\sum_{\phi\in\Phi}c_i(\phi)x_\phi
\le
(1-\gamma)k.
\]
Hence, before any prescribed copy is embedded, every cluster contains at
least \(\gamma k\) unused vertices.

Fix a pattern \(\phi\) at one step of the greedy procedure. For each
\(i\in[M]\), split the unused vertices of \(V_i\) into
\(|\phi^{-1}(i)|\) disjoint sets, one for each vertex of \(H\) mapped to
\(i\). Since \(|\phi^{-1}(i)|\le h\), each such set may be chosen to have
size at least
$\frac{\gamma k}{h}-1
\ge
\frac{\gamma k}{2h}
\ge
\eta_0 n$
for all sufficiently large \(n\). Moreover, each such set has relative
size at least \(\gamma/(3h)\) inside its cluster. For every edge
\(xy\in E(H)\), the corresponding pair is therefore
\((\varepsilon_{\mathrm{KLR}},p)\)-regular and has density at least
\(p/(4r)\). Lemma~\ref{lem:HpatternKLR} supplies the required canonical
copy of \(H\). Thus the greedy procedure succeeds.

Finally,
\[
\sum_{\phi\in\Phi}m_\phi
\ge
k\sum_{\phi\in\Phi}x_\phi-|\Phi|
\ge
k\sum_{\phi\in\Phi}x_\phi-O_{H,Z}(1),
\]
because \(|\Phi|\le M^h\le Z^h\).
\end{proof}



We apply Lemma~\ref{lem:embeddinggraph} with
\(F=G_\ell\), \(\mathcal X=\mathcal X_\ell\), and
\(\Phi=\Phi_\ell\).


\subsection{Step I: Reducing random graphs to linear programs}
\label{sec:Skeletonframeworkgraphs}

\begin{lemma}\label{lem:chitoskeleton_gr}
Fix \(H\), \(r\), and \(\xi>0\). There exists
\(\varepsilon_{\mathrm{reg}}>0\) such that, for every \(Z\), there is
\(C>0\) for which the following holds with high probability whenever
\(G\sim G(n,p)\) and
$p\ge Cn^{-1/m_2(H)}.$
Let \(\chi\) be any \(r\)-edge-colouring of \(G\), and let
$V(G)=V_1\cup\cdots\cup V_M,$
$M\le Z,$
be an equipartition. Define \(\mathcal X_\ell\) and \(\Phi_\ell\) using
\((\varepsilon_{\mathrm{reg}},p)\)-regular pairs as above. Then for every $\ell\in[r]$,
\[
\nu_H(G_\ell)\ge kZ_\ell-\xi n,
\]
where \(Z_\ell\) is the optimal value of the following linear program.

\begin{align}
\tag{LP}\label{eq:LP}
\text{minimize}\quad
& \sum_{i=1}^M z_i,\\
\text{subject to}\quad
& \sum_{x\in V(H)}z_{\phi(x)}\ge1
\qquad\text{for every }\phi\in\Phi_\ell,\notag\\
& z_i\ge0
\qquad\text{for every }i\in[M].\notag
\end{align}
\end{lemma}

\begin{proof}
The primal linear program dual to \eqref{eq:LP} is
\[
\begin{aligned}
\text{maximize}\quad
& \sum_{\phi\in\Phi_\ell}x_\phi,\\
\text{subject to}\quad
& \sum_{\phi\in\Phi_\ell}c_i(\phi)x_\phi\le1
\qquad\text{for every }i\in[M],\\
& x_\phi\ge0
\qquad\text{for every }\phi\in\Phi_\ell.
\end{aligned}
\]
By LP duality, its optimal value is \(Z_\ell\).

Choose \(\gamma>0\) sufficiently small in terms of \(\xi\), and take
\(\varepsilon_{\mathrm{reg}}\) small enough for
Lemma~\ref{lem:embeddinggraph}. Let \(\{x_\phi\}\) be an optimal primal
solution. The scaled solution
\(\{(1-\gamma)x_\phi\}\) satisfies
\[
\sum_{\phi\in\Phi_\ell}
c_i(\phi)(1-\gamma)x_\phi
\le1-\gamma
\qquad (i\in[M]).
\]
Therefore Lemma~\ref{lem:embeddinggraph} gives
\[
\nu_H(G_\ell)
\ge
k(1-\gamma)Z_\ell-O_{H,Z}(1).
\]
The constant vector \(z_i=1\) is feasible in \eqref{eq:LP}, so
\(Z_\ell\le M\). Hence
$\gamma kZ_\ell\le\gamma kM\le\gamma n.$
Taking \(\gamma\ll\xi\) and then \(n\) sufficiently large yields
$\nu_H(G_\ell)\ge kZ_\ell-\xi n.$\end{proof}


\subsection{Step II: Extracting adjusted weight vectors}

We next compress the dual vectors and, simultaneously, extract a
consistent colour for every pair of compressed types.

Let \(R_0=R_0(r,h)\) be sufficiently large that the following finite
multipartite Ramsey statement holds. Whenever \(q\le r\) and
\(B_1,\ldots,B_q\) are disjoint sets of size \(R_0\), every \(r\)-colouring
of the pairs in \(\bigcup_{j\in[q]}B_j\) contains subsets
\[
A_j\subseteq B_j,
\qquad |A_j|=h,
\]
such that, for every \(1\le i\le j\le q\), all pairs inside \(A_i\) when
\(i=j\), and all pairs between \(A_i\) and \(A_j\) when \(i<j\), have one
fixed colour.

For each colour \(\ell\in[r]\), let
\(\{s_i^{(\ell)}\}_{i\in[M]}\) be an optimal solution to
\eqref{eq:LP}, and put
\[
S_i:=\bigl(s_i^{(1)},\ldots,s_i^{(r)}\bigr),
\qquad
Z:=(Z_1,\ldots,Z_r).
\]
We may assume that
\[
0\le s_i^{(\ell)}\le1
\qquad (i\in[M],\ \ell\in[r]),
\]
since replacing a value larger than \(1\) by \(1\) preserves every
constraint in \eqref{eq:LP}. In particular,
\[
\sum_{i=1}^M S_i=Z.
\]

\begin{lemma}\label{lem:vertorrefinegraph}
For every \(\xi>0\), there exist \(m_0\in\mathbb N\) and
\(\zeta>0\) such that the following holds whenever \(M\ge m_0\). Assume
that
$\left|
\binom{[M]}2\setminus
\bigcup_{\ell\in[r]}\mathcal X_\ell
\right|
\le\zeta M^2.$ There exist vectors \(Y_1,\ldots,Y_M\in[0,2]^r\), distinct vectors
\(R_1,\ldots,R_q\in[0,2]^r\) with \(q\le r\), and colours
$\lambda(i,j)\in[r]$, $1\le i\le j\le q$
such that:
\begin{itemize}
\item[(Y1)]
$k\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty
+\xi n.$

\item[(Y2)]
For every \(\ell\in[r]\) and every map \(f:V(H)\to[q]\) satisfying
$\lambda(f(x),f(y))=\ell$
for every $xy\in E(H),$
we have
\[
\sum_{x\in V(H)}R_{f(x)}^{(\ell)}\ge1.
\]

\item[(Y3)]
$\{Y_i:i\in[M]\}\subseteq\{R_1,\ldots,R_q\}.$
\end{itemize}
\end{lemma}

\begin{proof}
Choose \(\rho>0\) sufficiently small in terms of \(\xi\) and \(r\),
and define
\[
\delta:=2\rho^{1/(6r)},
\qquad
\theta:=2\rho^{1/2},
\qquad
L:=\left(\left\lceil2/\delta\right\rceil+1\right)^r.
\]
We may assume that
$L\theta<1$
and 
$\delta+2L\theta<\frac{\xi}{3}.$
Choose \(m_0\) sufficiently large that
$\theta m_0\ge2R_0$
and 
$\frac{4r}{m_0}<\frac{\xi}{3},$
and then choose
$\zeta<\frac{\theta^2}{4R_0^2}.$


Round each coordinate of \(S_i\) upwards to the nearest multiple of
\(\delta\), obtaining \(\widetilde S_i\in[0,2]^r\). There are at most
\(L\) rounded values. Call a rounded value frequent if it occurs at least
\(\theta M\) times. Since \(L\theta<1\), at least one frequent value
exists.

Replace each non-frequent rounded value by an arbitrary frequent rounded
value, and denote the resulting vectors by \(S_1',\ldots,S_M'\). Fewer than
\(L\theta M\) indices are changed. Since all coordinates lie in \([0,2]\),
\[
\left\|\sum_{i=1}^M S_i'\right\|_\infty
\le
\left\|\sum_{i=1}^M S_i\right\|_\infty
+
\delta M+2L\theta M.
\]

Apply Lemma~\ref{lem:op2} with \(t=1\) to
\(S_1',\ldots,S_M'\). It gives selected vectors
\(W_1,\ldots,W_r\in\{S_1',\ldots,S_M'\}\) and non-negative integer
multiplicities \(m_1,\ldots,m_r\) summing to \(M\), such that
\[
\left\|\sum_{j=1}^r m_jW_j\right\|_\infty
\le
\left\|\sum_{i=1}^M S_i'\right\|_\infty+4r.
\]
Let \(Y_1,\ldots,Y_M\) be any indexed family containing exactly \(m_j\)
copies of \(W_j\), and let \(R_1,\ldots,R_q\) be its distinct values. Then
\(q\le r\), and \textnormal{(Y3)} holds. Moreover,
\[
\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
\|Z\|_\infty
+
(\delta+2L\theta)M+4r.
\]
Multiplying by \(k\), using \(kM\le n\) and \(M\ge m_0\), gives
\[
k\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty+\xi n,
\]
which proves \textnormal{(Y1)}.

Every \(R_j\) is a value of some \(S_i'\), and hence is a frequent rounded
value. Define its witness class by
\[
T_j:=\{u\in[M]:\widetilde S_u=R_j\}.
\]
Then the \(T_j\) are pairwise disjoint,
$|T_j|\ge\theta M,$ and $S_u\le R_j$
coordinatewise for every $u\in T_j.$

Let
$\mathcal B
:=
\binom{[M]}2
\setminus
\bigcup_{\ell\in[r]}\mathcal X_\ell.$
For each \(j\in[q]\), independently choose a uniformly random subset
\(B_j\subseteq T_j\) of size \(R_0\). For any fixed bad pair, the probability
that both endpoints are selected is at most
$\frac{2R_0^2}{\theta^2M^2}.$
Therefore the expected number of bad pairs selected inside or between the
sets \(B_j\) is at most
\[
|\mathcal B|
\frac{2R_0^2}{\theta^2M^2}
\le
\frac{2\zeta R_0^2}{\theta^2}
<1.
\]
Hence the sets \(B_1,\ldots,B_q\) may be chosen so that no bad pair occurs
inside any \(B_j\) or between any two of them.

For every pair \(\{u,v\}\) in their union, choose one colour
\(c(u,v)\in[r]\) such that
\[
\{u,v\}\in\mathcal X_{c(u,v)}.
\]
By the definition of \(R_0\), there are subsets
$A_j\subseteq B_j,$
with $|A_j|=h,$
such that all pairs inside each \(A_i\), and all pairs between each
\(A_i,A_j\), have fixed colours. Denote these colours by
\(\lambda(i,j)\).

It remains to prove \textnormal{(Y2)}. Let \(\ell\in[r]\), and let
\(f:V(H)\to[q]\) satisfy
$\lambda(f(x),f(y))=\ell$
for every $xy\in E(H).$
For every \(x\in V(H)\), choose
$\phi(x)\in A_{f(x)}$
so that all chosen indices are distinct. This is possible because every
\(A_j\) has \(h\) elements. If \(xy\in E(H)\), then
$\{\phi(x),\phi(y)\}\in\mathcal X_\ell,$
so \(\phi\in\Phi_\ell\). The dual constraint in \eqref{eq:LP} gives
\[
\sum_{x\in V(H)}
s_{\phi(x)}^{(\ell)}
\ge1.
\]
Since
$\phi(x)\in A_{f(x)}\subseteq T_{f(x)},$
we have \(S_{\phi(x)}\le R_{f(x)}\) coordinatewise. Consequently,
\[
\sum_{x\in V(H)}
R_{f(x)}^{(\ell)}
\ge
\sum_{x\in V(H)}
s_{\phi(x)}^{(\ell)}
\ge1.
\]
This proves \textnormal{(Y2)}.
\end{proof}


\subsection{Step III: Constructing the strip colouring}

We now construct the strip colouring from the compressed vectors and
the colours \(\lambda(i,j)\).

\begin{lemma}\label{lem:hygrrecongraph}
Let \(Y_1,\ldots,Y_M\), \(R_1,\ldots,R_q\), and
\(\lambda(i,j)\) satisfy \textnormal{(Y2)} and \textnormal{(Y3)} in
Lemma~\ref{lem:vertorrefinegraph}. Let
$V(K_n)=V_1\cup\cdots\cup V_M$
be an equipartition, where \(k=\lfloor n/M\rfloor\).
Then there exists an \(r\)-strip colouring \(\psi\) of \(K_n\) such that
\[
\nu_H(\psi)
\le
k\left\|\sum_{i=1}^M Y_i\right\|_\infty+2M.
\]
\end{lemma}

\begin{proof}
For each \(j\in[q]\), let
\[
\mathcal I_j:=\{i\in[M]:Y_i=R_j\},
\qquad
U_j:=\bigcup_{i\in\mathcal I_j}V_i.
\]
For every \(1\le i\le j\le q\), colour all edges inside \(U_i\), when
\(i=j\), and all edges between \(U_i\) and \(U_j\), when \(i<j\), with
colour \(\lambda(i,j)\). This defines an \(r\)-strip colouring \(\psi\).

Let \(Q\) be a colour-\(\ell\) copy of \(H\) in \(\psi\). Identify
\(V(Q)\) with \(V(H)\), and define \(f:V(H)\to[q]\) by letting \(f(x)=j\)
when the vertex corresponding to \(x\) lies in \(U_j\). Then
$\lambda(f(x),f(y))=\ell$
for every $xy\in E(H).$
By \textnormal{(Y2)},
\[
\sum_{x\in V(H)}
R_{f(x)}^{(\ell)}
\ge1.
\]

Now let \(\mathcal M\) be a colour-\(\ell\) \(H\)-tiling in \(\psi\).
Applying the preceding inequality to every copy in \(\mathcal M\), and using
the fact that these copies are vertex-disjoint, gives
\[
|\mathcal M|
\le
\sum_{j=1}^q
|U_j\cap V(\mathcal M)|R_j^{(\ell)}
\le
\sum_{j=1}^q|U_j|R_j^{(\ell)}.
\]
Since \(|V_i|\le k+1\),
\[
|\mathcal M|
\le
(k+1)\sum_{j=1}^q
|\mathcal I_j|R_j^{(\ell)}
=
(k+1)\sum_{i=1}^M y_i^{(\ell)}
\le
(k+1)
\left\|\sum_{i=1}^M Y_i\right\|_\infty.
\]
Because \(Y_i\in[0,2]^r\),
$\left\|\sum_{i=1}^M Y_i\right\|_\infty\le2M.$
Therefore
\[
\nu_H(\psi)
\le
k\left\|\sum_{i=1}^M Y_i\right\|_\infty+2M.\qedhere
\]
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:graphstructure}]
Fix the error \(\varepsilon>0\) in the theorem, and set
$\xi:=\frac{\varepsilon}{4}.$
Apply Lemma~\ref{lem:vertorrefinegraph} with error \(\xi\), obtaining
\(m_0\) and \(\zeta\). Choose the regularity parameter
\(\varepsilon_{\mathrm{reg}}>0\) sufficiently small that
$\varepsilon_{\mathrm{reg}}\le\zeta$
and Lemma~\ref{lem:chitoskeleton_gr} applies with error \(\xi\). Apply the multicolour sparse regularity lemma
(Lemma~\ref{hreg}, with \(t=2\), \(d=1/2\), and \(D=3/2\)), using lower
bound \(m_0\) and regularity parameter
\(\varepsilon_{\mathrm{reg}}\). Let \(\eta>0\) and \(Z\) be the resulting
constants. Choose \(C\) sufficiently large that, with high probability, both of the
following hold for \(G\sim G(n,p)\):
\(G\) is \((\eta,p,1/2,3/2)\)-uniform; and 
the universal embedding event required in
Lemma~\ref{lem:chitoskeleton_gr} holds for every partition with at most
\(Z\) clusters.

Condition on these events, and let \(\chi\) be an arbitrary
\(r\)-edge-colouring of \(G\). Apply Lemma~\ref{hreg} to obtain an
equipartition
$V(G)=V_1\cup\cdots\cup V_M,$
$m_0\le M\le Z,$
such that all but at most
\(\varepsilon_{\mathrm{reg}}M^2\) cluster pairs are
\((\varepsilon_{\mathrm{reg}},p)\)-regular in every colour. For every non-exceptional pair \(\{i,j\}\), uniformity of \(G\) gives
$d_G(V_i,V_j)\ge\frac p2.$
Since the colour densities sum to \(d_G(V_i,V_j)\), some colour
\(\ell\in[r]\) has density at least \(p/(2r)\). Therefore
$\left|
\binom{[M]}2
\setminus
\bigcup_{\ell\in[r]}\mathcal X_\ell
\right|
\le
\varepsilon_{\mathrm{reg}}M^2
\le
\zeta M^2.$

Let \(Z_\ell\) be the LP optimum in colour \(\ell\), and put
$Z=(Z_1,\ldots,Z_r).$
Lemma~\ref{lem:chitoskeleton_gr} gives
\[
\nu_H(\chi)
=
\max_{\ell\in[r]}\nu_H(G_\ell)
\ge
k\|Z\|_\infty-\xi n.
\]
Lemma~\ref{lem:vertorrefinegraph} gives vectors \(Y_1,\ldots,Y_M\) and
colours \(\lambda(i,j)\) satisfying
\[
k\left\|\sum_{i=1}^M Y_i\right\|_\infty
\le
k\|Z\|_\infty+\xi n.
\]
Finally, Lemma~\ref{lem:hygrrecongraph} gives an \(r\)-strip colouring
\(\psi\) of \(K_n\) such that
\[
\nu_H(\psi)
\le
k\left\|\sum_{i=1}^M Y_i\right\|_\infty+2M.
\]
Consequently,
$\nu_H(\chi)
\ge
\nu_H(\psi)-2\xi n-2M.$
Since \(M\le Z\), for all sufficiently large \(n\) we have
\(2M\le\xi n\). Hence
$\nu_H(\chi)
\ge
\nu_H(\psi)-3\xi n
\ge
\nu_H(\psi)-\varepsilon n,$
as required.
\end{proof}


\section{Ramsey graph tilings: finite optimization and explicit values}
\label{sec:graphapp}

Recall that 
$Rt_r(H;G):=\min_{\chi:E(G)\to[r]}\nu_H(\chi).$
We first prove that the asymptotic value of \(Rt_r(H;K_n)\) is determined by
a finite optimization over strip templates. We then derive explicit values
for connected non-bipartite graphs, Hall-type bipartite graphs with three
colours, and balanced Hall-type bipartite graphs.


\subsection{Finite optimization for the asymptotic constant}

Theorem~\ref{thm:graphstructure} was stated for connected graphs with
at least three vertices, whereas Theorem~\ref{thm:alg} allows arbitrary fixed
graphs with at least one edge. For complete hosts, the required extension is
immediate from the proof of the structural theorem. We record it explicitly.

\begin{lemma}[Complete-host structural reduction]
\label{lem:complete-host-graphstructure}
Let \(r\ge2\), let \(H\) be any fixed graph with at least one edge, and let
\(\varepsilon>0\). For all sufficiently large \(n\), every
\(r\)-edge-colouring \(\chi\) of \(K_n\) admits an \(r\)-strip colouring
\(\psi\) of \(K_n\) such that
\[
\nu_H(\chi)\ge \nu_H(\psi)-\varepsilon n.
\]
\end{lemma}

\begin{proof}
For connected \(H\) with \(v(H)\ge3\), this is
Theorem~\ref{thm:graphstructure} with \(p=1\). The same proof applies to
disconnected \(H\): in the lifting lemma, split each available cluster into
disjoint sets corresponding to the vertices of \(H\), and embed each
nontrivial connected component separately. Components with at least three
vertices are embedded by Lemma~\ref{lem:HpatternKLR} with \(p=1\);
\(K_2\)-components are embedded in their positive-density regular pairs; and
isolated vertices are selected greedily. Since all vertex sets used for
different components are disjoint, the resulting embeddings form a copy of
\(H\). The LP-duality, vector-compression, and Ramsey-extraction steps are
unchanged.
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:alg}]

Fix \(1\le q\le r\). An \(r\)-colour \(q\)-strip template is a map
\[
\lambda:\{(i,j):1\le i\le j\le q\}\to[r].
\]
For \(x=(x_1,\ldots,x_q)\) in the simplex
\(\Delta_q:=\{x_i\ge0:\sum_i x_i=1\}\), and for
\(\ell\in[r]\), let
\[
\Phi_\ell(\lambda)
:=
\{f:V(H)\to[q]:
\lambda(f(u),f(v))=\ell
\text{ for every }uv\in E(H)\}.
\]
For \(f\in\Phi_\ell(\lambda)\), put \(a_i(f):=|f^{-1}(i)|\). Define
\(\tau_\ell(\lambda,x)\) as the optimum of
\[
\begin{aligned}
\text{maximize}\quad
& \sum_{f\in\Phi_\ell(\lambda)}w_f\\
\text{subject to}\quad
& \sum_{f\in\Phi_\ell(\lambda)}a_i(f)w_f\le x_i
\qquad (i\in[q]),\\
& w_f\ge0
\qquad (f\in\Phi_\ell(\lambda)).
\end{aligned}
\]
Finally, set
\[
\beta_{r,H}
:=
\min_{1\le q\le r}
\ \min_{\lambda}
\ \min_{x\in\Delta_q}
\ \max_{\ell\in[r]}\tau_\ell(\lambda,x).
\]
The minimum is attained because there are finitely many templates and
\(\Delta_q\) is compact.

We first interpret this optimization. Consider the strip colouring
corresponding to \((q,\lambda)\), with part sizes
\(|U_i|=x_i n\). Up to rounding the part sizes, its maximum
colour-\(\ell\) \(H\)-tiling has size
\[
\tau_\ell(\lambda,x)n+O_{r,H}(1).
\]
Indeed, if a tiling uses \(N_f\) copies of type \(f\), then the numbers
\(N_f/n\) form a feasible solution to the LP. Conversely, from a feasible
solution \(\{w_f\}\), take \(\lfloor nw_f\rfloor\) copies of each type.
The capacity constraints allow the required vertices to be allocated
disjointly inside each part, and the template guarantees all required edges
have colour \(\ell\). Since the number of types depends only on \(r\) and
\(H\), rounding loses only \(O_{r,H}(1)\) copies.

Choosing a minimizing template and part-size vector, and rounding its part
sizes, gives
\[
Rt_r(H;K_n)\le(\beta_{r,H}+o(1))n.
\]
For the reverse inequality, fix \(\delta>0\) and let \(\chi\) be an arbitrary
\(r\)-colouring of \(K_n\). By
Lemma~\ref{lem:complete-host-graphstructure}, there is an \(r\)-strip
colouring \(\psi\) such that
\(\nu_H(\chi)\ge\nu_H(\psi)-\delta n\). If \(\psi\) has template
\(\lambda\) and part proportions \(x\), then
\[
\nu_H(\psi)
=
\max_{\ell\in[r]}\tau_\ell(\lambda,x)n+O_{r,H}(1)
\ge
\beta_{r,H}n-O_{r,H}(1).
\]
Thus
\[
\nu_H(\chi)\ge\beta_{r,H}n-\delta n-O_{r,H}(1).
\]
As \(\delta>0\) is arbitrary, this proves
\(Rt_r(H;K_n)=(\beta_{r,H}+o(1))n\).

It remains to justify the claimed finite algorithm. For fixed
\((q,\lambda,\ell)\), the dual feasible region is
\[
P_\ell:=
\left\{
z\in[0,1]^q:
\sum_{i=1}^q a_i(f)z_i\ge1
\text{ for every }f\in\Phi_\ell(\lambda)
\right\}.
\]
The restriction \(z_i\le1\) is harmless: truncating a larger coordinate to
\(1\) preserves every constraint. Let \(E_\ell\) be the finite set of
extreme points of \(P_\ell\). Strong duality gives
\[
\tau_\ell(\lambda,x)
=
\min_{z\in E_\ell}x\cdot z.
\]

For each tuple
\(\mathbf z=(z_1,\ldots,z_r)\in E_1\times\cdots\times E_r\), consider the
polytope of \(x\in\Delta_q\) satisfying
\[
x\cdot z_\ell\le x\cdot z
\qquad
\text{for every }\ell\in[r]\text{ and }z\in E_\ell.
\]
On this polytope, \(\tau_\ell(\lambda,x)=x\cdot z_\ell\) for every
\(\ell\). The minimum of
\(\max_\ell x\cdot z_\ell\) there is obtained from the linear program
\[
\text{minimize }y
\quad\text{subject to}\quad
x\in\Delta_q,\qquad
y\ge x\cdot z_\ell\quad(\ell\in[r]),
\]
together with the preceding active-region constraints. Enumerating the
finitely many tuples \(\mathbf z\), templates \(\lambda\), and integers
\(q\le r\) therefore computes \(\beta_{r,H}\) by finitely many linear
programs depending only on \(r\) and \(H\).
\end{proof}


\subsection{Explicit values for \(Rt_r(H;K_n)\)}
\label{subsec:application:3colour}

We now extract closed formulas from the finite optimization in three
natural cases.


\subsubsection{Non-bipartite tilings}

\begin{proposition}\label{prop:nonbipartite}
Let \(r\ge3\) and let \(H\) be a connected non-bipartite graph. Then
\[
Rt_r(H;K_n)=\frac{n}{r\,v(H)}+O_{r,H}(1).
\]
\end{proposition}

\begin{observation}\label{obs:lownonbipar}
For every \(r\)-edge-colouring \(\chi\) of \(K_n\),
\[
\nu_H(\chi)\ge\frac{n}{r\,v(H)}-O_{r,H}(1).
\]
\end{observation}

\begin{proof}
Greedily remove monochromatic copies of \(H\) until fewer than
\(R_r(H)\) vertices remain. This produces at least
\((n-R_r(H))/v(H)\) disjoint copies in total, so one colour contains at
least \(n/(r\,v(H))-O_{r,H}(1)\) of them.
\end{proof}

\begin{lemma}\label{lem:uppernonbipar}
There exists an \(r\)-edge-colouring \(\chi\) of \(K_n\) such that
\[
\nu_H(\chi)\le\frac{n}{r\,v(H)}+O_{r,H}(1).
\]
\end{lemma}

\begin{proof}
Let \(V(K_n)=V_1\cup\cdots\cup V_r\) be an equipartition. Suppose first that \(r\) is odd. Identify the indices and colours with
\(\mathbb Z_r\). Colour the edges inside \(V_i\) with colour \(i\), and
for \(i\ne j\), colour all edges between \(V_i\) and \(V_j\) with
\((i+j)2^{-1}\in\mathbb Z_r\). For a fixed colour \(c\), the
colour-\(c\) pairs between distinct parts form a matching on
\(\mathbb Z_r\setminus\{c\}\), while no such pair is incident with \(V_c\).
Hence the colour-\(c\) graph is the disjoint union of the clique on \(V_c\)
and complete bipartite graphs. Since \(H\) is connected and non-bipartite,
every colour-\(c\) copy of \(H\) lies inside \(V_c\).

Now suppose that \(r\) is even. Index \(V_1,\ldots,V_{r-1}\) and the first
\(r-1\) colours by \(\mathbb Z_{r-1}\), and denote the final part and colour
by \(\infty\). Use the preceding odd construction on the first \(r-1\)
parts, colour the edges inside \(V_\infty\) with \(\infty\), and colour the
edges between \(V_\infty\) and \(V_i\) with \(i+1\pmod{r-1}\). For a fixed
\(c\in\mathbb Z_{r-1}\), the colour-\(c\) pairs between parts form a
matching together with at most one additional adjacent edge, and hence form
a forest; moreover, \(V_c\) is isolated from all colour-\(c\) cross-pairs.
Thus the colour-\(c\) graph outside \(V_c\) is bipartite, while colour
\(\infty\) occurs only inside \(V_\infty\). Again, every monochromatic copy
of \(H\) lies inside one designated part.

Consequently, every monochromatic \(H\)-tiling has at most
\(\max_i\lfloor |V_i|/v(H)\rfloor=n/(r\,v(H))+O_H(1)\) copies.
\end{proof}

Proposition~\ref{prop:nonbipartite} now follows from
Observation~\ref{obs:lownonbipar} and Lemma~\ref{lem:uppernonbipar}.


\subsubsection{Hall-type bipartite tilings for \(r=3\)}

Let \(H\) be a connected bipartite graph with a fixed bipartition
\(A\cup B\), where \(|A|=a\le b=|B|\). We call \(H\) \emph{Hall-type} if
it contains a matching saturating \(A\). Equivalently, by K\"onig's theorem,
the vertex-cover number of \(H\) is \(a\).

\begin{theorem}\label{thm:biparr=3}
Let \(b\ge a\ge1\), and let \(H\) be a connected Hall-type bipartite graph
with bipartition sizes \(a\) and \(b\). Then
\[
Rt_3(H;K_n)=
\begin{cases}
\displaystyle
\left(\frac{1}{3a+b}+o(1)\right)n,
& b\le3a,\\[2ex]
\displaystyle
\left(\frac{2}{3(a+b)}+o(1)\right)n,
& b\ge3a.
\end{cases}
\]
\end{theorem}

\begin{lemma}\label{lem:upper3colour}
Let \(H\) be as in Theorem~\ref{thm:biparr=3}.
\begin{enumerate}
\item If \(b\ge3a\), then there exists a \(3\)-colouring \(\chi\) of
\(K_n\) such that
\[
\nu_H(\chi)\le\frac{2}{3(a+b)}n+O_{a,b}(1).
\]

\item If \(b\le3a\), then there exists a \(3\)-colouring \(\chi\) of
\(K_n\) such that
\[
\nu_H(\chi)\le\frac{1}{3a+b}n+O_{a,b}(1).
\]
\end{enumerate}
\end{lemma}

\begin{proof}
If \(b\ge3a\), partition \(V(K_n)\) into three parts of size
\(n/3+O(1)\). Colour the edges inside \(V_i\) and between \(V_i,V_{i+1}\)
with colour \(i\), with indices modulo \(3\). Each colour is incident with
at most \(2n/3+O(1)\) vertices, and hence supports at most
\(2n/[3(a+b)]+O_{a,b}(1)\) disjoint copies of \(H\).

Now assume \(b\le3a\). Take
\(|V_2|=|V_3|=\lfloor an/(3a+b)\rfloor\), and let \(V_1\) contain the
remaining vertices. Colour all edges inside \(V_3\) and between \(V_3\) and
\(V_1\cup V_2\) with colour \(3\); colour all edges inside \(V_2\) and
between \(V_2,V_1\) with colour \(2\); and colour all edges inside \(V_1\)
with colour \(1\).

In the colour-\(i\) graph, for \(i\in\{2,3\}\), the vertices outside
\(V_i\) form an independent set. Hence the vertices of any colour-\(i\)
copy of \(H\) lying in \(V_i\) form a vertex cover of \(H\), and therefore
there are at least \(a\) of them. Such a tiling has at most
\(|V_i|/a\le n/(3a+b)+O(1)\) copies. Colour \(1\) occurs only inside
\(V_1\), so a colour-\(1\) tiling has at most
\(|V_1|/(a+b)\le n/(3a+b)+O(1)\) copies.
\end{proof}

\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=0.75]

\definecolor{EdgeLR}{RGB}{235,140,135}
\definecolor{EdgeTR}{RGB}{140,165,230}
\definecolor{EdgeTL}{RGB}{140,205,165}

\begin{scope}[xshift=-3.36cm]
  \coordinate (T2) at (0,1.76);
  \coordinate (L2) at (-1.6,-0.8);
  \coordinate (R2) at ( 1.6,-0.8);

  \tikzset{
    partBig/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=24mm, inner sep=0pt, font=\small
    },
    partSmall/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=12.8mm, inner sep=0pt, font=\small
    }
  }

  \draw[line width=12.8pt, line cap=round, color=EdgeTL] (R2) -- (L2);

  \foreach \i in {0,...,29} {
    \pgfmathsetmacro{\t}{\i/30}
    \pgfmathsetmacro{\tt}{(\i+1)/30}
    \pgfmathsetmacro{\w}{35.2 - (35.2-9.6)*\i/29}
    \draw[line cap=round, color=EdgeTL, line width=\w pt]
      ($(T2)!\t!(R2)$) -- ($(T2)!\tt!(R2)$);
  }

  \foreach \i in {0,...,29} {
    \pgfmathsetmacro{\t}{\i/30}
    \pgfmathsetmacro{\tt}{(\i+1)/30}
    \pgfmathsetmacro{\w}{35.2 - (35.2-9.6)*\i/29}
    \draw[line cap=round, color=EdgeTR, line width=\w pt]
      ($(T2)!\t!(L2)$) -- ($(T2)!\tt!(L2)$);
  }

  \node[partBig,   fill=EdgeLR] at (T2) {$V_1$};
  \node[partSmall, fill=EdgeTR] at (L2) {$V_2$};
  \node[partSmall, fill=EdgeTL] at (R2) {$V_3$};

  \node[font=\small] at (0,-1.96) {$b\le 3a$};
\end{scope}

\begin{scope}[xshift=3.36cm]
  \coordinate (T1) at (0,1.76);
  \coordinate (L1) at (-1.6,-0.8);
  \coordinate (R1) at ( 1.6,-0.8);

  \tikzset{
    part/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=19.2mm, inner sep=0pt, font=\small
    }
  }

  \draw[line width=22.4pt, line cap=round, color=EdgeLR] (T1) -- (L1);
  \draw[line width=22.4pt, line cap=round, color=EdgeTR] (L1) -- (R1);
  \draw[line width=22.4pt, line cap=round, color=EdgeTL] (R1) -- (T1);

  \node[part, fill=EdgeLR] at (T1) {$V_1$};
  \node[part, fill=EdgeTR] at (L1) {$V_2$};
  \node[part, fill=EdgeTL] at (R1) {$V_3$};

  \node[font=\small] at (0,-1.96) {$b\ge 3a$};
\end{scope}

\end{tikzpicture}
\caption{Extremal \(3\)-colourings for Lemma~\ref{lem:upper3colour}.}
\label{fig:3strip}
\end{figure}

\begin{lemma}\label{lem:Kab-bipartite-packing}
Let \(b\ge a\ge1\) and \(N_1\ge N_2\). The complete bipartite graph
\(K_{N_1,N_2}\) contains a \(K_{a,b}\)-tiling of size at least
\[
\min\left\{
\frac{N_1+N_2}{a+b},
\frac{N_2}{a}
\right\}
-O_{a,b}(1).
\]
\end{lemma}

\begin{proof}
If \(a=b\), the assertion follows by greedily taking
\(\lfloor N_2/a\rfloor\) copies. Assume \(b>a\). If
\(aN_1>bN_2\), orient every copy with \(b\) vertices in the first class
and \(a\) in the second; this gives \(\lfloor N_2/a\rfloor\) copies.

Suppose instead that \(aN_1\le bN_2\). Set
\[
x:=\frac{bN_1-aN_2}{b^2-a^2},
\qquad
y:=\frac{bN_2-aN_1}{b^2-a^2}.
\]
Both quantities are non-negative. Taking \(\lfloor x\rfloor\) copies with
\(b\) vertices in the first class and \(\lfloor y\rfloor\) copies with
the opposite orientation respects both vertex capacities and gives
\[
x+y-O(1)=\frac{N_1+N_2}{a+b}-O(1)
\]
copies.
\end{proof}

\begin{lemma}\label{lem:lower3colour}
Let \(b\ge a\ge1\), and let \(\chi\) be a \(3\)-strip \(3\)-colouring of
\(K_n\). Then
\[
\nu_{K_{a,b}}(\chi)\ge
\begin{cases}
\displaystyle
\frac{1}{3a+b}n-O_{a,b}(1),
& b\le3a,\\[2ex]
\displaystyle
\frac{2}{3(a+b)}n-O_{a,b}(1),
& b\ge3a.
\end{cases}
\]
\end{lemma}

\begin{proof}
Pad the strip partition with empty parts if necessary, and write
\(V(K_n)=V_1\cup V_2\cup V_3\), where
\(n_1:=|V_1|\ge n_2:=|V_2|\ge n_3:=|V_3|\). Each graph induced by one
part, and each bipartite graph between two parts, is monochromatic.

Suppose first that \(b\ge3a\). If \(n_1\ge2n/3\), the monochromatic clique
on \(V_1\) contains \(2n/[3(a+b)]-O(1)\) disjoint copies of
\(K_{a,b}\). Otherwise \(n_1<2n/3\). Since
\(n=n_1+n_2+n_3\le n_1+2n_2\), we have \(n_2>n/6\). By
Lemma~\ref{lem:Kab-bipartite-packing}, the monochromatic graph between
\(V_1,V_2\) contains at least
\[
\min\left\{
\frac{n_1+n_2}{a+b},
\frac{n_2}{a}
\right\}
-O(1)
\]
copies. Here \(n_1+n_2\ge2n/3\), and
\(n_2/a>n/(6a)\ge2n/[3(a+b)]\), using \(b\ge3a\). This proves the first
case.

Now suppose that \(b\le3a\). If
\(n_1\ge(a+b)n/(3a+b)\), the clique on \(V_1\) gives
\(n/(3a+b)-O(1)\) copies. Otherwise the inequality
\(n\le n_1+2n_2\) implies \(n_2\ge an/(3a+b)\). Again applying
Lemma~\ref{lem:Kab-bipartite-packing}, the graph between \(V_1,V_2\)
contains at least the minimum of
\((n_1+n_2)/(a+b)\) and \(n_2/a\), up to \(O(1)\).
The second term is at least \(n/(3a+b)\), while
\(n_1+n_2\ge2n/3\) and \(b\le3a\) imply
\[
\frac{n_1+n_2}{a+b}
\ge
\frac{2n}{3(a+b)}
\ge
\frac{n}{3a+b}.
\]
This proves the second case.
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:biparr=3}]
Lemma~\ref{lem:upper3colour} gives the upper bounds. For the lower
bounds, apply Lemma~\ref{lem:complete-host-graphstructure} with an error
tending to zero. It gives a \(3\)-strip colouring \(\psi\) such that
\(\nu_H(\chi)\ge\nu_H(\psi)-o(n)\). Since \(H\subseteq K_{a,b}\), every
\(K_{a,b}\)-tiling contains an \(H\)-tiling of the same size, and hence
\(\nu_H(\psi)\ge\nu_{K_{a,b}}(\psi)\). The result follows from
Lemma~\ref{lem:lower3colour}.
\end{proof}


\subsubsection{A non-Hall bipartite example}
\label{subsec:D23-main}

The Hall-type condition in Theorem~\ref{thm:biparr=3} leads to a uniform
formula, but it is not necessary for an explicit evaluation. Let \(D_{2,3}\)
be the double star whose two centres are adjacent, with two leaves attached
to one centre and three leaves attached to the other. Its bipartition classes
have sizes \(3\) and \(4\), but its maximum matching has size \(2\).
Consequently, \(D_{2,3}\) does not satisfy the Hall-type condition.

Nevertheless, a direct analysis of its three-part strip templates determines
the asymptotic constant exactly.

\begin{theorem}\label{thm:D23}
$Rt_3(D_{2,3};K_n)
=
\left(\frac{7}{78}+o(1)\right)n.$
\end{theorem}

\begin{figure}[htbp]
\centering
\begin{tikzpicture}[scale=0.75]

\definecolor{EdgeOne}{RGB}{235,140,135}
\definecolor{EdgeTwo}{RGB}{140,165,230}
\definecolor{EdgeThree}{RGB}{140,205,165}

\begin{scope}[xshift=0cm]
  \coordinate (T) at (0,2.35);
  \coordinate (L) at (-2.25,-1.0);
  \coordinate (R) at ( 2.25,-1.0);

  \tikzset{
    partBig/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=33mm, inner sep=0pt, font=\small, align=center
    },
    partMid/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=24mm, inner sep=0pt, font=\small, align=center
    },
    partSmall/.style={
      circle, draw=black, line width=1.2pt,
      minimum size=17mm, inner sep=0pt, font=\small, align=center
    }
  }

  \foreach \i in {0,...,29} {
    \pgfmathsetmacro{\t}{\i/30}
    \pgfmathsetmacro{\tt}{(\i+1)/30}
    \pgfmathsetmacro{\w}{44 - (44-12)*\i/29}
    \draw[line cap=round, color=EdgeOne, line width=\w pt]
      ($(T)!\t!(L)$) -- ($(T)!\tt!(L)$);
  }

  \foreach \i in {0,...,29} {
    \pgfmathsetmacro{\t}{\i/30}
    \pgfmathsetmacro{\tt}{(\i+1)/30}
    \pgfmathsetmacro{\w}{44 - (44-26)*\i/29}
    \draw[line cap=round, color=EdgeTwo, line width=\w pt]
      ($(T)!\t!(R)$) -- ($(T)!\tt!(R)$);
  }

  \foreach \i in {0,...,29} {
    \pgfmathsetmacro{\t}{\i/30}
    \pgfmathsetmacro{\tt}{(\i+1)/30}
    \pgfmathsetmacro{\w}{26 - (26-12)*\i/29}
    \draw[line cap=round, color=EdgeTwo, line width=\w pt]
      ($(R)!\t!(L)$) -- ($(R)!\tt!(L)$);
  }

  \node[partSmall, fill=EdgeOne] at (L)
    {$V_1$\\[-1mm]\scriptsize \(4n/39\)};

  \node[partBig, fill=EdgeThree] at (T)
    {$V_2$\\[-1mm]\scriptsize \(49n/78\)};

  \node[partMid, fill=EdgeOne] at (R)
    {$V_3$\\[-1mm]\scriptsize \(7n/26\)};

\end{scope}

\end{tikzpicture}
\caption{The \(D_{2,3}\) construction.}
\label{fig:D23upper}
\end{figure}

For comparison, the complete bipartite graph \(K_{3,4}\), which has the same
bipartition sizes, has three-colour constant \(1/13\) by
Theorem~\ref{thm:biparr=3}. Thus the tiling constant is not determined by the
bipartition sizes alone once the Hall-type condition is dropped. The proof of
Theorem~\ref{thm:D23} is given in Appendix~\ref{app:D23}.


\subsubsection{Complete bipartite tilings with four and five colours}

The four-colour case is the first in which the optimal value has more than
two regimes. It also illustrates how the finite strip-template optimization
can be solved without enumerating all templates.

\begin{theorem}\label{thm:Kab-four}
Let \(b\ge a\ge1\). Then
\[
Rt_4(K_{a,b};K_n)=
\begin{cases}
\displaystyle
\left(\frac{1}{4a+b}+o(1)\right)n,
& b\le3a,\\[2ex]
\displaystyle
\left(\frac{2}{5a+3b}+o(1)\right)n,
& 3a\le b\le5a,\\[2ex]
\displaystyle
\left(\frac{3}{5(a+b)}+o(1)\right)n,
& b\ge5a.
\end{cases}
\]
\end{theorem}

The upper bounds are given by the following three strip templates:
\[
\Lambda_1=
\begin{pmatrix}
1&2&3&4\\
2&2&3&4\\
3&3&3&4\\
4&4&4&4
\end{pmatrix},
\qquad
\Lambda_2=
\begin{pmatrix}
1&1&3&4\\
1&2&2&4\\
3&2&3&4\\
4&4&4&4
\end{pmatrix},
\qquad
\Lambda_3=
\begin{pmatrix}
1&1&2&3\\
1&4&4&4\\
2&4&4&4\\
3&4&4&4
\end{pmatrix}.
\]
Here the corresponding normalized part-size vectors are
\[
\frac{(a+b,a,a,a)}{4a+b},
\qquad
\frac{(a+b,a+b,a+b,2a)}{5a+3b},
\qquad
\frac{(2,1,1,1)}5,
\]
respectively.

\begin{figure}[htbp]
\centering
\begin{tikzpicture}[
  scale=0.72,
  transform shape,
  line cap=round,
  line join=round
]

\definecolor{KabRed}{RGB}{235,140,135}
\definecolor{KabBlue}{RGB}{140,165,230}
\definecolor{KabGreen}{RGB}{140,205,165}
\definecolor{KabPurple}{RGB}{190,120,230}

\tikzset{
  KabBig/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=28mm, inner sep=0pt,
    font=\small, align=center
  },
  KabMedium/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=21mm, inner sep=0pt,
    font=\small, align=center
  },
  KabSmall/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=14mm, inner sep=0pt,
    font=\small, align=center
  }
}

% Tapered band from #1 to #2, with colour #3,
% starting width #4 and ending width #5.
\def\KabTaper#1#2#3#4#5{%
  \foreach \i in {0,...,19} {
    \pgfmathsetmacro{\KabStart}{\i/20}
    \pgfmathsetmacro{\KabEnd}{(\i+1)/20}
    \pgfmathsetmacro{\KabWidth}{#4-(#4-#5)*\i/19}
    \draw[color=#3,line width=\KabWidth pt]
      ($(#1)!\KabStart!(#2)$) --
      ($(#1)!\KabEnd!(#2)$);
  }
}

% ---------------------------------------------------------
% b <= 3a
% Part sizes: (a+b,a,a,a)/(4a+b)
% ---------------------------------------------------------
\begin{scope}[xshift=-7.1cm]
  \coordinate (A1) at (0,2.15);
  \coordinate (A2) at (-2.25,-0.75);
  \coordinate (A3) at (0,-0.75);
  \coordinate (A4) at (2.25,-0.75);

  \draw[color=KabGreen,line width=11pt] (A2)--(A3);
  \draw[color=KabPurple,line width=11pt] (A3)--(A4);
  \draw[color=KabPurple,line width=11pt]
    (A2) to[bend right=55] (A4);

  \KabTaper{A1}{A2}{KabBlue}{28}{8}
  \KabTaper{A1}{A3}{KabGreen}{30}{8}
  \KabTaper{A1}{A4}{KabPurple}{28}{8}

  \node[KabBig,fill=KabRed] at (A1) {$V_1$};
  \node[KabSmall,fill=KabBlue] at (A2) {$V_2$};
  \node[KabSmall,fill=KabGreen] at (A3) {$V_3$};
  \node[KabSmall,fill=KabPurple] at (A4) {$V_4$};

  \node[align=center] at (0,-2.75)
  {$b\le3a$\\[-1mm]
   \scriptsize $(a+b,a,a,a)/(4a+b)$};
\end{scope}

% ---------------------------------------------------------
% 3a <= b <= 5a
% Part sizes: (a+b,a+b,a+b,2a)/(5a+3b)
% ---------------------------------------------------------
\begin{scope}[xshift=0cm,yshift=0.25cm]
  \coordinate (B1) at (0,2.05);
  \coordinate (B2) at (-1.75,0);
  \coordinate (B3) at (1.75,0);
  \coordinate (B4) at (0,-1.65);

  \draw[color=KabRed,line width=14pt] (B1)--(B2);
  \draw[color=KabGreen,line width=14pt] (B1)--(B3);
  \draw[color=KabBlue,line width=14pt] (B2)--(B3);

  \KabTaper{B1}{B4}{KabPurple}{20}{8}
  \KabTaper{B2}{B4}{KabPurple}{20}{8}
  \KabTaper{B3}{B4}{KabPurple}{20}{8}

  \node[KabMedium,fill=KabRed] at (B1) {$V_1$};
  \node[KabMedium,fill=KabBlue] at (B2) {$V_2$};
  \node[KabMedium,fill=KabGreen] at (B3) {$V_3$};
  \node[KabSmall,fill=KabPurple] at (B4) {$V_4$};

  \node[align=center] at (0,-3.1)
  {$3a\le b\le5a$\\[-1mm]
   \scriptsize $(a+b,a+b,a+b,2a)/(5a+3b)$};
\end{scope}

% ---------------------------------------------------------
% b >= 5a
% Part sizes: (2,1,1,1)/5
% ---------------------------------------------------------
\begin{scope}[xshift=7.1cm]
  \coordinate (C1) at (0,2.15);
  \coordinate (C2) at (-2.25,-0.75);
  \coordinate (C3) at (0,-0.75);
  \coordinate (C4) at (2.25,-0.75);

  \draw[color=KabPurple,line width=11pt] (C2)--(C3);
  \draw[color=KabPurple,line width=11pt] (C3)--(C4);
  \draw[color=KabPurple,line width=11pt]
    (C2) to[bend right=55] (C4);

  \KabTaper{C1}{C2}{KabRed}{28}{8}
  \KabTaper{C1}{C3}{KabBlue}{30}{8}
  \KabTaper{C1}{C4}{KabGreen}{28}{8}

  \node[KabBig,fill=KabRed] at (C1) {$V_1$};
  \node[KabSmall,fill=KabPurple] at (C2) {$V_2$};
  \node[KabSmall,fill=KabPurple] at (C3) {$V_3$};
  \node[KabSmall,fill=KabPurple] at (C4) {$V_4$};

  \node[align=center] at (0,-2.75)
  {$b\ge5a$\\[-1mm]
   \scriptsize $(2,1,1,1)/5$};
\end{scope}

\end{tikzpicture}
\caption{Extremal four-colour strip templates for
\(K_{a,b}\)-tilings.}
\label{fig:Kab4-extremal}
\end{figure}

For \(\Lambda_1\), colour \(1\) occurs only inside the first part, while
every edge of colour \(j\in\{2,3,4\}\) meets the \(j\)-th part. For
\(\Lambda_2\), each of colours \(1,2,3\) is supported on two of the three
large parts, while every colour-\(4\) edge meets the fourth part. For
\(\Lambda_3\), every colour is supported on parts of total normalized size
at most \(3/5\). These observations give the three upper bounds.

We now prove the matching lower bounds simultaneously.

\begin{lemma}\label{lem:Kab-four-forcing}
Let \(\lambda>0\), and put
$A:=(a+b)\lambda,$
$B:=a\lambda.$
Suppose that
$A+3B\le1,$
$1-\frac{3A}{2}\ge B,$
and 
$A\le\frac35.$
Then every strip colouring with at most four parts and at most four colours
on its cross-pairs contains a monochromatic \(K_{a,b}\)-tiling of size
$\lambda n-O_{a,b}(1).$
The colours inside the individual parts may be arbitrary.
\end{lemma}

\begin{proof}
Pad the strip partition with empty parts if necessary, and let its normalized
part sizes be
$x_1\ge x_2\ge x_3\ge x_4,$
$x_1+x_2+x_3+x_4=1.$

Suppose first that \(x_1\ge(b/a)x_2\). If \(x_1\ge A\), the monochromatic
clique on the first part gives the required tiling. If \(x_2\ge B\), the
complete bipartite graph between the first two parts gives it. Otherwise,
$1\le x_1+3x_2<A+3B\le1,$
a contradiction.

We may therefore assume that \(x_1<(b/a)x_2\). By
Lemma~\ref{lem:Kab-bipartite-packing}, the first two parts give the required
tiling unless
$x_1+x_2<A.$
Since \(x_3\le x_2\le(x_1+x_2)/2\), we have
\[
x_4
=
1-x_1-x_2-x_3
>
1-\frac{3A}{2}
\ge B.
\]
If \(x_1\ge(b/a)x_4\), the first and fourth parts give the required
tiling. Hence we may assume
$x_1<\frac ba x_4.$
Thus every pair of parts has size ratio smaller than \(b/a\).

Moreover,
$2x_1+1
=
3x_1+x_2+x_3+x_4
\le
3(x_1+x_2)
<
3A.$
It follows that every three parts have total normalized size greater than
\[
1-\frac{3A-1}{2}
=
\frac32(1-A)
\ge A.
\]

Among the six cross-pairs, some colour occurs at least twice. If two
same-coloured pairs are adjacent in the auxiliary \(K_4\), the common part
and the union of the other two parts form a monochromatic complete bipartite
graph. Both sides have size at least \(B\), and their total size is at least
\(A\), so Lemma~\ref{lem:Kab-bipartite-packing} gives the required tiling.

If the two pairs are disjoint, each pair can be tiled using all its vertices
up to \(O_{a,b}(1)\), because its part-size ratio is smaller than \(b/a\).
The two tilings have the same colour and together contain
$\frac{n}{a+b}-O_{a,b}(1)
\ge
\lambda n-O_{a,b}(1)$
copies.
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:Kab-four}]
The three templates above give the upper bounds.

For the lower bounds, fix \(\delta>0\). By
Lemma~\ref{lem:complete-host-graphstructure}, every four-colouring of
\(K_n\) admits a four-strip colouring \(\psi\) such that
$\nu_{K_{a,b}}(\chi)
\ge
\nu_{K_{a,b}}(\psi)-\delta n.$

Apply Lemma~\ref{lem:Kab-four-forcing} with
$\lambda=
\frac1{4a+b},$
$\lambda=
\frac2{5a+3b},$
$\lambda=
\frac3{5(a+b)}$
in the three respective ranges. Direct substitution shows that the three
hypotheses of the lemma hold, respectively, when
$b\le3a,$
$3a\le b\le5a,$
$b\ge5a.$
Therefore \(\psi\), and hence \(\chi\), contains a monochromatic
\(K_{a,b}\)-tiling of the claimed asymptotic size. Since \(\delta>0\) is
arbitrary, the result follows.
\end{proof}


\begin{theorem}\label{thm:Kab-five}
Let \(b\ge a\ge1\). Then
\[
Rt_5(K_{a,b};K_n)=
\begin{cases}
\displaystyle
\left(\frac{1}{5a+b}+o(1)\right)n,
& b\le3a,\\[2ex]
\displaystyle
\left(\frac{2}{7a+3b}+o(1)\right)n,
& 3a\le b\le5a,\\[2ex]
\displaystyle
\left(\frac{3}{8a+5b}+o(1)\right)n,
& 5a\le b\le\frac{13}{2}a,\\[2ex]
\displaystyle
\left(\frac{5}{9(a+b)}+o(1)\right)n,
& b\ge\frac{13}{2}a.
\end{cases}
\]
\end{theorem}

\begin{figure}[htbp]
\centering
\begin{tikzpicture}[
  scale=0.57,
  transform shape,
  line cap=round,
  line join=round
]

\definecolor{KabRed}{RGB}{235,140,135}
\definecolor{KabBlue}{RGB}{140,165,230}
\definecolor{KabGreen}{RGB}{140,205,165}
\definecolor{KabPurple}{RGB}{190,120,230}
\definecolor{KabOrange}{RGB}{245,190,140}

\tikzset{
  KabFiveLarge/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=27mm, inner sep=0pt,
    font=\large, align=center
  },
  KabFiveMedium/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=20mm, inner sep=0pt,
    font=\large, align=center
  },
  KabFiveSmall/.style={
    circle, draw=black, line width=1.2pt,
    minimum size=13mm, inner sep=0pt,
    font=\large, align=center
  }
}

\def\KabFiveTaper#1#2#3#4#5{%
  \foreach \i in {0,...,19} {
    \pgfmathsetmacro{\KabFiveStart}{\i/20}
    \pgfmathsetmacro{\KabFiveEnd}{(\i+1)/20}
    \pgfmathsetmacro{\KabFiveWidth}{#4-(#4-#5)*\i/19}
    \draw[color=#3,line width=\KabFiveWidth pt]
      ($(#1)!\KabFiveStart!(#2)$) --
      ($(#1)!\KabFiveEnd!(#2)$);
  }
}

% ---------------------------------------------------------
% b <= 3a
% Part sizes: (a+b,a,a,a,a)/(5a+b)
% ---------------------------------------------------------
\begin{scope}[xshift=-10.6cm]
  \coordinate (A1) at (0,2.2);
  \coordinate (A2) at (-3.05,-0.6);
  \coordinate (A3) at (-1.02,-0.6);
  \coordinate (A4) at (1.02,-0.6);
  \coordinate (A5) at (3.05,-0.6);

  \draw[color=KabGreen,line width=9pt] (A2)--(A3);
  \draw[color=KabPurple,line width=9pt] (A3)--(A4);
  \draw[color=KabPurple,line width=9pt]
    (A2) to[bend right=52] (A4);

  \KabFiveTaper{A1}{A2}{KabBlue}{23}{7}
  \KabFiveTaper{A1}{A3}{KabGreen}{25}{7}
  \KabFiveTaper{A1}{A4}{KabPurple}{23}{7}
  \KabFiveTaper{A1}{A5}{KabOrange}{23}{7}

  \draw[color=KabOrange,line width=9pt] (A4)--(A5);
  \draw[color=KabOrange,line width=9pt]
    (A3) to[bend right=50] (A5);
  \draw[color=KabOrange,line width=9pt]
    (A2) to[bend right=58] (A5);

  \node[KabFiveLarge,fill=KabRed] at (A1) {$V_1$};
  \node[KabFiveSmall,fill=KabBlue] at (A2) {$V_2$};
  \node[KabFiveSmall,fill=KabGreen] at (A3) {$V_3$};
  \node[KabFiveSmall,fill=KabPurple] at (A4) {$V_4$};
  \node[KabFiveSmall,fill=KabOrange] at (A5) {$V_5$};

  \node[align=center] at (0,-3)
  {$b\le3a$\\[-1mm]
   \scriptsize $(a+b,a,a,a,a)/(5a+b)$};
\end{scope}

% ---------------------------------------------------------
% 3a <= b <= 5a
% Part sizes: (a+b,a+b,a+b,2a,2a)/(7a+3b)
% ---------------------------------------------------------
\begin{scope}[xshift=-3.3cm,yshift=0.25cm]
  \coordinate (B1) at (0,2.15);
  \coordinate (B2) at (-2.0,0);
  \coordinate (B3) at (2.0,0);
  \coordinate (B4) at (-0.9,-2.15);
  \coordinate (B5) at (0.9,-2.15);

  \draw[color=KabRed,line width=11pt] (B1)--(B2);
  \draw[color=KabGreen,line width=11pt] (B1)--(B3);
  \draw[color=KabBlue,line width=11pt] (B2)--(B3);

  \KabFiveTaper{B1}{B4}{KabPurple}{18}{7}
  \KabFiveTaper{B2}{B4}{KabPurple}{18}{7}
  \KabFiveTaper{B3}{B4}{KabPurple}{18}{7}

  \KabFiveTaper{B1}{B5}{KabOrange}{18}{7}
  \KabFiveTaper{B2}{B5}{KabOrange}{18}{7}
  \KabFiveTaper{B3}{B5}{KabOrange}{18}{7}

  \draw[color=KabOrange,line width=9pt] (B4)--(B5);

  \node[KabFiveMedium,fill=KabRed] at (B1) {$V_1$};
  \node[KabFiveMedium,fill=KabBlue] at (B2) {$V_2$};
  \node[KabFiveMedium,fill=KabGreen] at (B3) {$V_3$};
  \node[KabFiveSmall,fill=KabPurple] at (B4) {$V_4$};
  \node[KabFiveSmall,fill=KabOrange] at (B5) {$V_5$};

  \node[align=center] at (0,-3.25)
  {$3a\le b\le5a$\\[-1mm]
   \scriptsize $(a+b,a+b,a+b,2a,2a)/(7a+3b)$};
\end{scope}

% ---------------------------------------------------------
% 5a <= b <= 13a/2
% Part sizes:
% (2(a+b),a+b,a+b,a+b,3a)/(8a+5b)
% ---------------------------------------------------------
\begin{scope}[xshift=3.65cm]
  \coordinate (C1) at (-1,2.35);
  \coordinate (C2) at (-2.35,-0.85);
  \coordinate (C3) at (0,-0.85);
  \coordinate (C4) at (2.35,-0.85);
  \coordinate (C5) at (1.75,1.75);

  \draw[color=KabPurple,line width=9pt] (C2)--(C3);
  \draw[color=KabPurple,line width=9pt] (C3)--(C4);
  \draw[color=KabPurple,line width=9pt]
    (C2) to[bend right=52] (C4);

  \KabFiveTaper{C1}{C2}{KabRed}{23}{7}
  \KabFiveTaper{C1}{C3}{KabBlue}{25}{7}
  \KabFiveTaper{C1}{C4}{KabGreen}{23}{7}

  \KabFiveTaper{C1}{C5}{KabOrange}{24}{7}
  \KabFiveTaper{C2}{C5}{KabOrange}{18}{7}
  \KabFiveTaper{C3}{C5}{KabOrange}{18}{7}
  \KabFiveTaper{C4}{C5}{KabOrange}{18}{7}

  \node[KabFiveLarge,fill=KabRed] at (C1) {$V_1$};
  \node[KabFiveMedium,fill=KabPurple] at (C2) {$V_2$};
  \node[KabFiveMedium,fill=KabPurple] at (C3) {$V_3$};
  \node[KabFiveMedium,fill=KabPurple] at (C4) {$V_4$};
  \node[KabFiveSmall,fill=KabOrange] at (C5) {$V_5$};

  \node[align=center] at (0,-2.85)
  {$5a\le b\le13a/2$\\[-1mm]
   \scriptsize
   $(2(a+b),a+b,a+b,a+b,3a)/(8a+5b)$};
\end{scope}

% ---------------------------------------------------------
% b >= 13a/2
% Part sizes: (3,2,2,1,1)/9
% ---------------------------------------------------------
\begin{scope}[xshift=10.35cm]
  \coordinate (D1) at (0,2.2);
  \coordinate (D2) at (-2,0);
  \coordinate (D3) at (2,0);
  \coordinate (D4) at (-0.9,-2);
  \coordinate (D5) at (0.9,-2);

  \KabFiveTaper{D1}{D2}{KabRed}{26}{8}
  \KabFiveTaper{D1}{D3}{KabGreen}{26}{8}

  \draw[color=KabBlue,line width=11pt] (D2)--(D3);

  \KabFiveTaper{D1}{D4}{KabPurple}{22}{7}
  \KabFiveTaper{D2}{D4}{KabBlue}{17}{7}
  \KabFiveTaper{D3}{D4}{KabBlue}{17}{7}

  \KabFiveTaper{D1}{D5}{KabPurple}{22}{7}
  \KabFiveTaper{D2}{D5}{KabOrange}{17}{7}
  \KabFiveTaper{D3}{D5}{KabOrange}{17}{7}

  \draw[color=KabPurple,line width=9pt] (D4)--(D5);

  \node[KabFiveLarge,fill=KabRed] at (D1) {$V_1$};
  \node[KabFiveMedium,fill=KabBlue] at (D2) {$V_2$};
  \node[KabFiveMedium,fill=KabGreen] at (D3) {$V_3$};
  \node[KabFiveSmall,fill=KabBlue] at (D4) {$V_4$};
  \node[KabFiveSmall,fill=KabOrange] at (D5) {$V_5$};

  \node[align=center] at (0,-3.05)
  {$b\ge13a/2$\\[-1mm]
   \scriptsize $(3,2,2,1,1)/9$};
\end{scope}

\end{tikzpicture}
\caption{Extremal five-colour strip templates for
\(K_{a,b}\)-tilings.}
\label{fig:Kab5-extremal}
\end{figure}

The five-colour optimization has one additional regime and its proof uses
Theorem~\ref{thm:Kab-four} recursively inside a four-part subtemplate. Since
the resulting case analysis is substantially longer, we defer it to
Appendix~\ref{subsec:5colour}.


These formulas
suggest that, although Theorem~\ref{thm:alg} gives a finite optimization for
every fixed number of colours, extracting a uniform closed formula for
\(K_{a,b}\) requires a more conceptual understanding of the extremal strip
templates.


\subsubsection{Hall-type balanced bipartite tilings}

\begin{theorem}\label{thm:Kaa}
Let \(r\ge1\), and let \(H\) be a connected bipartite graph whose two
bipartition classes both have size \(a\), and which contains a matching of
size \(a\). Then
\[
Rt_r(H;K_n)=
\left(\frac{1}{a(r+1)}+o(1)\right)n.
\]
\end{theorem}

In particular, this applies to the \(d\)-dimensional hypercube.

\begin{corollary}\label{cor:Qd}
For every \(r\ge1\) and \(d\ge2\),
\[
Rt_r(Q_d;K_n)
=
\left(\frac{1}{2^{d-1}(r+1)}+o(1)\right)n.
\]
\end{corollary}

\begin{proof}
The two bipartition classes of \(Q_d\) both have size \(2^{d-1}\), and
\(Q_d\) has a perfect matching. Apply Theorem~\ref{thm:Kaa}.
\end{proof}

\begin{lemma}\label{lem:upperKaa}
Let \(r\ge1\), and let \(H\) be as in Theorem~\ref{thm:Kaa}. There exists
an \(r\)-colouring \(\chi\) of \(K_n\) such that
\[
\nu_H(\chi)\le\frac{1}{a(r+1)}n+O_a(1).
\]
\end{lemma}

\begin{proof}
Partition \(V(K_n)\) into \(V_1,\ldots,V_r\), with
\(|V_i|=n/(r+1)+O(1)\) for \(i<r\) and
\(|V_r|=2n/(r+1)+O(1)\). Colour the edges inside \(V_i\) with colour
\(i\), and colour all edges between \(V_i,V_j\), where \(i<j\), with
colour \(i\).

For \(i<r\), the vertices outside \(V_i\) form an independent set in
colour \(i\). Hence every colour-\(i\) copy of \(H\) uses a vertex cover
of \(H\) inside \(V_i\). Since \(H\) has a matching of size \(a\) and a
bipartition class of size \(a\), its vertex-cover number is exactly \(a\).
Thus a colour-\(i\) tiling has at most \(|V_i|/a\) copies. Colour \(r\)
occurs only inside \(V_r\), so a colour-\(r\) tiling has at most
\(|V_r|/(2a)\) copies. Both bounds are
\(n/[a(r+1)]+O_a(1)\).
\end{proof}

\begin{lemma}\label{lem:lowerKaa}
Let \(r\ge1\), \(a\ge1\), and let \(\chi\) be an \(r\)-colouring of
\(K_n\). Then
\[
\nu_{K_{a,a}}(\chi)\ge
\left(\frac{1}{a(r+1)}-o(1)\right)n.
\]
\end{lemma}

\begin{proof}
The case \(r=1\) is immediate, so assume \(r\ge2\). Apply
Lemma~\ref{lem:complete-host-graphstructure} with an error tending to zero.
It suffices to prove that every \(r\)-strip colouring contains
\(\lfloor n/[a(r+1)]\rfloor\) disjoint monochromatic copies of
\(K_{a,a}\).

Let \(U_1,\ldots,U_q\), where \(q\le r\), be the strip parts, and set
\(x_i:=(r+1)|U_i|/n\). If some \(x_i\ge2\), the monochromatic clique on
\(U_i\) contains at least \(\lfloor |U_i|/(2a)\rfloor\ge
\lfloor n/[a(r+1)]\rfloor\) disjoint copies of \(K_{a,a}\).

We may therefore assume \(x_i<2\) for every \(i\). Since
\(\sum_i x_i=r+1\), the quantity
\[
\sum_{i=1}^q\bigl(x_i-\lfloor x_i\rfloor\bigr)
=
r+1-\sum_{i=1}^q\lfloor x_i\rfloor
\]
is an integer smaller than \(q\le r\). Hence
\(\sum_i\lfloor x_i\rfloor\ge2\). There are therefore two parts,
say \(U_1,U_2\), each of size at least \(n/(r+1)\). All edges between
them have one colour, so greedily taking \(a\) vertices from each part
produces at least
\[
\min\left\{
\left\lfloor\frac{|U_1|}{a}\right\rfloor,
\left\lfloor\frac{|U_2|}{a}\right\rfloor
\right\}
\ge
\left\lfloor\frac{n}{a(r+1)}\right\rfloor
\]
disjoint monochromatic copies of \(K_{a,a}\).
\end{proof}

Theorem~\ref{thm:Kaa} now follows directly from
Lemmas~\ref{lem:upperKaa} and~\ref{lem:lowerKaa}.



\section{Concluding remarks and further directions}\label{sec:conclusion}

In this paper we introduced a strip-colouring framework for reducing arbitrary
edge-colourings to highly structured colourings while preserving, up to
\(o(n)\), the maximum size of a monochromatic matching or, in the graph case,
a monochromatic bipartite tiling. This gives topology-free asymptotic proofs
of AFL-type results, applications to stable Kneser hypergraphs, and a finite
LP-based description of monochromatic graph tilings. We conclude by recording
several directions for future research.

\subsection{Toward an exact topology-free AFL theorem}

The most immediate gap between Corollary~\ref{cor:intro-asyAFL} and the exact
Alon--Frankl--Lov\'asz theorem is the \(o(n)\) error term. This loss enters
through the regularity reduction, the LP rounding step, and the passage from
the original colouring to a strip colouring. Removing this loss would amount
to a genuinely sharper version of the structural theorem. A starting point is to improve the error term in Corollary~\ref{cor:intro-asyAFL} 
from \(o(n)\) to \(O_{r,t}(1)\).


A second limitation is that our proof is inherently a bounded-uniformity
argument. The weak regularity input and the finite-dimensional LP reduction
are effective when \(t\) is fixed, but do not address regimes in which the
uniformity grows with \(n\).
Is there a topology-free proof of the AFL theorem, or even of its asymptotic
form, which remains effective when \(t=t(n)\) grows with \(n\)?

\subsection{Stable Kneser hypergraphs}

Theorem~\ref{thm:stable} determines the chromatic number of
\(\mathrm{KG}^q(n,t)_{s\text{-stab}}\) up to an additive error of at most one
in the regime \(t=O(1)\), \(q=cn\), and \(s\ll q\). It would be natural to
remove this final ambiguity.

The condition \(s\ll q\) is also a natural boundary of the present argument:
it ensures that the hypergraph of \(s\)-stable \(t\)-sets behaves sufficiently
pseudorandomly on large vertex subsets. When \(s\) is comparable to \(q\), the
geometry of stable sets becomes much more visible. Understanding this regime would require either a more refined structural
analysis of stable sets or a replacement for the pseudorandomness input used
in the proof of Theorem~\ref{thm:stable}.

\subsection{Hypergraph tilings beyond matchings}

The matching problem treated in the hypergraph part of the paper is the
special case of tiling by a single \(t\)-edge. The graph part of the paper shows that, when \(t=2\), the same
LP-duality philosophy can handle much more general tiling problems. It is
therefore natural to ask how far this can be extended for \(t\)-uniform
hypergraphs with \(t\ge3\).

\begin{problem}\label{prob:hypergraph-tiling-strip}
Let \(r,t\ge3\). For which \(t\)-uniform hypergraph \(\mathcal H\) does there exist a finite strip-template
optimization determining
$Rt_r(\mathcal H; K_n^{(t)})=(\beta_{r,\mathcal H}+o(1))n?$
\end{problem}

We believe that our method should also handle the case where $\mathcal H$ is the balanced $t$-partite $t$-graph $K^{(t)}_{a,\ldots,a}$, since $K_n^{(t)}$ has sufficient edge density to allow the use of a Turán-type theorem to embed $\mathcal H$.
In the case $r\ge 3$ and $\cH=K_p^{(t)}$ for any $p>t$, an optimal monochromatic tiling can always be obtained by a greedy strategy.  We provide the proof in the Appendix.

\begin{proposition}\label{prop:clique-tiling-upper}
Let $t\ge 3$, $r\ge 3$, and $p\ge t+1$. Then
$Rt_r\left(K_p^{(t)};K_n^{(t)}\right)
=
\frac{n}{rp}+O_{r,t,p}(1).$
\end{proposition}

There is a further distinction between complete hosts and sparse pseudorandom
hosts. For matchings, weak regularity is sufficient. For general
\(\mathcal H\)-tilings, however, if the host hypergraph is sparse, weak regularity does not provide counting
lemmas for non-linear hypergraphs, that is, hypergraphs containing two edges
with more than one common vertex. This obstruction already appears for simple
examples such as \(K_{2,2,2}^{(3)}\). It would be interesting to develop a structural reduction for monochromatic \(\mathcal H\)-tilings in
sparse \(t\)-graphs, for fixed \(t\ge3\), which retains a finite
optimization description of the asymptotic tiling size.

\subsection{Explicit bipartite tiling constants}

Theorem~\ref{thm:alg} reduces the evaluation of
\(\beta_{r,H}\) to a finite optimization for every fixed \(r\) and \(H\),
but extracting a transparent closed formula remains a separate extremal
problem. For complete bipartite graphs, the two-colour formula follows from
the theorem of Burr, Erd\H{o}s and Spencer, while
Theorems~\ref{thm:biparr=3},~\ref{thm:Kab-four} and~\ref{thm:Kab-five} determine the cases
\(r=3,4,5\).

Already these first cases display increasing complexity. As the number of
colours increases from \(3\) to \(5\), the number of parameter regimes in the
ratio \(b/a\) increases from two to three and then four. Each breakpoint
corresponds to a change in the optimal strip template. Although the finite
algorithm applies for every fixed \(r\), enumerating all templates for larger
\(r\) becomes unwieldy and does not reveal why particular templates are
extremal.

\begin{problem}\label{prob:Kab-explicit}
For fixed \(r\ge6\), determine \(a\,\beta_{r,K_{a,b}}\) as a function of
the ratio \(b/a\). In particular, how many breakpoints can occur as a
function of \(r\), and can the extremal strip templates be characterized
without exhaustive case analysis?
\end{problem}

The double-star result shows that explicit evaluation is not confined to
Hall-type graphs. The graph \(D_{2,3}\) has bipartition classes of sizes
\(3\) and \(4\), but its maximum matching has size \(2\), so it lies outside
the scope of Theorem~\ref{thm:biparr=3}. Nevertheless,
Theorem~\ref{thm:D23} gives
$\beta_{3,D_{2,3}}=\frac{7}{78}.$
Since \(K_{3,4}\) has the same bipartition sizes but a different tiling
constant, the value of \(\beta_{3,H}\) for non-Hall bipartite graphs depends
on finer structural properties of \(H\).

\begin{problem}\label{prob:nonHall-bipartite}
Determine \(\beta_{3,H}\) for connected bipartite graphs \(H\) which do not
contain a matching saturating a smallest bipartition class. Which structural
parameters of \(H\), beyond its bipartition sizes and independence number,
govern the answer?
\end{problem}





\section*{Acknowledgement}
We thank Rob Morris for helpful discussions. HL, LW and ZY are supported by the Institute for Basic Science (IBS-R029-C4), and LW is also supported by National Key R\&D Program of China under grant number 2024YFA1013900, NSFC under grant number 12471327 and by China Scholarship Council.






%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


\bibliographystyle{abbrv}
%\addcontentsline{toc}{chapter}{Bibliography}
\bibliography{multicolourtiling}

\appendix


\section{Complete bipartite tilings in five colours}
\label{subsec:5colour}

We prove Theorem~\ref{thm:Kab-five}. We repeatedly use
Lemma~\ref{lem:Kab-bipartite-packing}, as well as the fact that tilings
supported on disjoint same-coloured pairs may be combined.

\subsection{Extremal templates}

The four upper-bound templates are
\[
\Lambda_1^{(5)}=
\begin{pmatrix}
1&2&3&4&5\\
2&2&3&4&5\\
3&3&3&4&5\\
4&4&4&4&5\\
5&5&5&5&5
\end{pmatrix},
\qquad
\Lambda_2^{(5)}=
\begin{pmatrix}
1&1&3&4&5\\
1&2&2&4&5\\
3&2&3&4&5\\
4&4&4&4&5\\
5&5&5&5&5
\end{pmatrix},
\]
\[
\Lambda_3^{(5)}=
\begin{pmatrix}
1&1&2&3&5\\
1&4&4&4&5\\
2&4&4&4&5\\
3&4&4&4&5\\
5&5&5&5&5
\end{pmatrix},
\qquad
\Lambda_4^{(5)}=
\begin{pmatrix}
1&1&3&4&4\\
1&2&2&2&5\\
3&2&3&2&5\\
4&2&2&4&4\\
4&5&5&4&5
\end{pmatrix}.
\]
Their corresponding unnormalized part-size vectors are
\[
(a+b,a,a,a,a),
\quad
(a+b,a+b,a+b,2a,2a),
\quad
(2(a+b),a+b,a+b,a+b,3a),
\quad
(3,2,2,1,1),
\]
respectively. See Figure~\ref{fig:Kab5-extremal}.

For \(\Lambda_1^{(5)}\), colour \(1\) is supported only on the first part,
while each later colour has a designated part meeting all its edges. For
\(\Lambda_2^{(5)}\), each of the first three colours is supported on two
large parts, while colours \(4\) and \(5\) have designated vertex-cover
parts. For \(\Lambda_3^{(5)}\), each of colours \(1,2,3,4\) is supported
on parts of total normalized size \(3(a+b)/(8a+5b)\), while every
colour-\(5\) copy uses at least \(a\) vertices of the fifth part. Finally,
in \(\Lambda_4^{(5)}\), every colour is supported on parts of total
normalized size at most \(5/9\). These observations give all four upper
bounds in Theorem~\ref{thm:Kab-five}.

\subsection{The five-part forcing lemma}

Define
\[
c_4(a,b):=
\begin{cases}
\displaystyle\frac1{4a+b},
& b\le3a,\\[1ex]
\displaystyle\frac2{5a+3b},
& 3a\le b\le5a,\\[1ex]
\displaystyle\frac3{5(a+b)},
& b\ge5a.
\end{cases}
\]


\begin{lemma}\label{lem:Kab-five-forcing}
Let \(\lambda>0\), and put
$A:=(a+b)\lambda,$ $B:=a\lambda$.
Suppose that
\begin{align}
A+4B&\le1,\label{eq:five-force-1}\\
\frac{1-\frac32A}{2}&\ge B,\label{eq:five-force-2}\\
A&\le\frac59,\label{eq:five-force-3}\\
c_4(a,b)(1-B)&\ge\lambda,\label{eq:five-force-4}\\
\frac ba\cdot\frac{4-7A}{3}&\ge2A-1.\label{eq:five-force-5}
\end{align}
Then every strip colouring with at most five parts and at most five colours
on its cross-pairs contains a monochromatic \(K_{a,b}\)-tiling of size
$\lambda n-O_{a,b}(1).$
\end{lemma}

\begin{proof}
Let the normalized part sizes be
$x_1\ge x_2\ge x_3\ge x_4\ge x_5,$
$\sum_{i=1}^5x_i=1.$
As in the proof of Lemma~\ref{lem:Kab-four-forcing},
\eqref{eq:five-force-1} implies that we may assume
\[
x_1<\frac ba x_2
\qquad\text{and}\qquad
x_1+x_2<A.
\]
Indeed, otherwise either the clique on the first part, the pair formed by the
first two parts, or the inequality
\(1\le x_1+4x_2<A+4B\) gives the result.

Since \(x_3\le x_2\le(x_1+x_2)/2\),
$x_4+x_5
=
1-x_1-x_2-x_3
>
1-\frac{3A}{2}.$
Thus \(x_4>B\) by \eqref{eq:five-force-2}. We may further assume
$x_1<\frac ba x_4,$
since otherwise the pair formed by the first and fourth parts gives the
required tiling.

Put
$T:=x_2+x_3+x_4.$
If \(T\ge A\), two of the six cross-pairs among the first four parts have the
same colour. Adjacent pairs yield a same-coloured complete bipartite graph
on three parts, while disjoint pairs yield two compatible tilings on four
parts. Since all first four parts have size at least \(B\) and size ratio
smaller than \(b/a\), either configuration gives the required tiling.
We may therefore assume
$T<A.$

Let
$S:=x_2+x_3+x_4+x_5.$
Since \(x_2\ge(x_3+x_4+x_5)/3\),
\[
S>\frac{4(1-A)}3
\qquad\text{and hence}\qquad
x_5=S-T>\frac{4-7A}{3}.
\]
Furthermore, the inequalities
\(x_1+x_2<A\), \(x_2+x_3+x_4<A\), and \(x_5\le x_4\) imply
$x_2+x_3<2A-\frac23.$
Consequently,
\[
x_1+x_4+x_5
=
1-x_2-x_3
>
\frac53-2A
\ge A
\]
by \eqref{eq:five-force-3}. Thus every three-part set containing the first
part has total normalized size at least \(A\).

The cross-pairs \(12,13,14\) must now have distinct colours; otherwise two
same-coloured adjacent pairs already give the desired tiling. Denote these
colours by \(c_2,c_3,c_4\).

Suppose first that \(x_5\ge B\). If \(x_1\ge(b/a)x_5\), the pair \(15\)
gives the desired tiling. Otherwise all five parts have pairwise size ratio
smaller than \(b/a\). The colour \(c_5\) of \(15\) must be distinct from
\(c_2,c_3,c_4\), since a repeated colour would give a same-coloured
three-part configuration of total size at least \(A\).

Let \(c_0\) be the only possible remaining colour. Every edge \(ij\) among
\(2,3,4,5\) must have colour \(c_0\): any colour \(c_i\) or \(c_j\) gives
two adjacent same-coloured pairs, while a colour \(c_k\) with
\(k\notin\{i,j\}\) gives two disjoint same-coloured pairs. Hence all
cross-pairs among the final four parts have colour \(c_0\). Taking one of
these parts against the union of the other three, and using \(S>A\), gives
the required tiling.

It remains to consider \(x_5<B\). If the six cross-pairs among the first
four parts use at most four colours, then
Lemma~\ref{lem:Kab-four-forcing}, applied on their
\((1-x_5)n\) vertices, gives at least
\[
c_4(a,b)(1-x_5)n-O(1)
>
c_4(a,b)(1-B)n-O(1)
\ge
\lambda n-O(1)
\]
copies, by \eqref{eq:five-force-4}.

We may therefore assume that those six pairs use all five colours. Let
\(c_5\) be the colour of \(15\). If \(c_5\in\{c_2,c_3,c_4\}\), two
same-coloured adjacent pairs again give the required tiling. Hence \(c_5\)
is distinct from \(c_2,c_3,c_4\). Since all five colours occur among the
first four parts, some edge \(ij\), with \(i,j\in\{2,3,4\}\), has colour
\(c_5\). Thus \(15\) and \(ij\) are disjoint same-coloured pairs.

If \(x_1<(b/a)x_5\), both pairs can tile all their vertices up to
\(O_{a,b}(1)\), and the four parts involved have total size at least
\(S>A\). We may therefore assume \(x_1\ge(b/a)x_5\). The two pairs then
give at least
\[
\frac{x_5}{a}n
+
\frac{x_i+x_j}{a+b}n
-O(1)
\]
copies. If this were smaller than \(\lambda n-O(1)\), then
\[
x_i+x_j
<
A-\frac{a+b}{a}x_5.
\]
Since \(x_3+x_4\le x_i+x_j\), we would obtain
\[
x_1+x_2
>
1-A+\frac ba x_5
>
1-A+\frac ba\cdot\frac{4-7A}{3}
\ge A
\]
by \eqref{eq:five-force-5}. But \(x_1<(b/a)x_2\), so the pair \(12\)
would then give the desired tiling, a contradiction.
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:Kab-five}]
The four templates above give the upper bounds.

For the lower bounds, apply Lemma~\ref{lem:Kab-five-forcing} with
\[
\lambda=
\frac1{5a+b},
\qquad
\frac2{7a+3b},
\qquad
\frac3{8a+5b},
\qquad
\frac5{9(a+b)}
\]
in the four respective ranges.

Conditions \eqref{eq:five-force-1}--\eqref{eq:five-force-3} follow directly
from the stated bounds on \(b/a\). In the first three ranges,
\[
c_4(a,b)(1-B)=\lambda,
\]
while in the fourth range condition \eqref{eq:five-force-4} is equivalent
to \(2b\ge13a\).

Writing \(z=b/a\), condition \eqref{eq:five-force-5} is automatic in the
first range and reduces in the remaining ranges to
\[
2z(7-z)\ge3(z-3),
\qquad
z(11-z)\ge3(z-2),
\qquad
z\ge3,
\]
respectively. These inequalities hold throughout the indicated intervals.

Thus every five-strip colouring contains a monochromatic
\(K_{a,b}\)-tiling of the asserted size. The complete-host structural
reduction transfers the result to arbitrary five-colourings of \(K_n\).
\end{proof}




\section{The double star \(D_{2,3}\)}
\label{app:D23}

In this section, we prove Theorem~\ref{thm:D23}. Write
\(D:=D_{2,3}\). Thus \(v(D)=7\), its bipartition classes have sizes \(3\)
and \(4\), and its vertex-cover number is \(2\).

We begin with the extremal construction.


\begin{lemma}\label{lem:D23upper}
There exists a \(3\)-colouring \(\chi\) of \(E(K_n)\) such that
$\nu_D(\chi)\le \frac{7}{78}n+O(1).$
\end{lemma}

\begin{proof}
Let \(V(K_n)=V_1\cup V_2\cup V_3\), where
\[
|V_1|=\left\lfloor\frac{4n}{39}\right\rfloor,\qquad
|V_3|=\left\lfloor\frac{7n}{26}\right\rfloor,\qquad
|V_2|=n-|V_1|-|V_3|=\frac{49n}{78}+O(1).
\]

Colour the edges inside \(V_1\) and \(V_3\), and the edges between
\(V_1\) and \(V_2\), red. Colour the edges between \(V_3\) and
\(V_1\cup V_2\) blue, and colour the edges inside \(V_2\) green; see
Figure~\ref{fig:D23upper}.

The red graph has two components: the clique on \(V_3\), and the graph
on \(V_1\cup V_2\) in which every edge meets \(V_1\). Since \(D\) has
vertex-cover number \(2\), every red copy in \(V_1\cup V_2\) uses at least
two vertices of \(V_1\), while every red copy in \(V_3\) uses seven vertices
there. Hence a red tiling has at most
$\frac{|V_1|}{2}+\frac{|V_3|}{7}
=
\frac{7n}{78}+O(1)$
copies.

The blue graph is complete bipartite with classes \(V_3\) and
\(V_1\cup V_2\). Since the bipartition classes of \(D\) have sizes \(3\)
and \(4\), every blue copy uses at least three vertices of \(V_3\).
Thus a blue tiling has at most
$\frac{|V_3|}{3}
=
\frac{7n}{78}+O(1)$
copies. Finally, green occurs only inside \(V_2\), so a green tiling has at
most
$\frac{|V_2|}{7}
=
\frac{7n}{78}+O(1)$
copies.
\end{proof}


We next prove the matching lower bound for strip colourings. We use the
finite pattern-packing LP from the proof of Theorem~\ref{thm:alg}. For fixed
part proportions, its integral optimum differs from its fractional optimum by
\(O_D(1)\).

\begin{lemma}[Three-part packing calculation]
\label{lem:D23-threepart-LP}
Let \(A,U,W\) be three strip parts of normalized sizes \(x,u,v\),
respectively. Suppose that
$AU,\ UW$  are blue,
$AW$  is red,
and that the edges inside \(A\) have a third colour. According to the colours
inside \(U\) and \(W\), the normalized fractional red and blue
\(D\)-tiling numbers are given by the following table:
\[
\begin{array}{c|c|c}
\text{colours inside }U,W
&
\tau_{\mathrm{red}}
&
\tau_{\mathrm{blue}}
\\ \hline
\mathrm{R},\mathrm{R}
&
\displaystyle
\min\left\{\frac{u}{7}+\frac{v}{2},\frac17\right\}
&
\displaystyle
\min\left\{\frac{u}{3},\frac{x+v}{3},\frac17\right\}
\\[3ex]
\mathrm{R},\mathrm{B}
&
\displaystyle
\min\left\{\frac{u}{7}+\frac{v}{3},
            \frac{x}{3}+\frac{u}{7},
            \frac17\right\}
&
\displaystyle
\min\left\{\frac{4u+v}{7},
            \frac{u}{3}+\frac{2v}{9},
            \frac{x}{3}+\frac{v}{2},
            \frac17\right\}
\\[3ex]
\mathrm{B},\mathrm{R}
&
\displaystyle
\min\left\{\frac{v}{2},\frac{x+v}{7}\right\}
&
\displaystyle
\min\left\{\frac{u}{2},\frac17\right\}
\\[3ex]
\mathrm{B},\mathrm{B}
&
\displaystyle
\min\left\{\frac{v}{3},\frac{x}{3},\frac{x+v}{7}\right\}
&
\displaystyle
\min\left\{\frac{4u+v}{7},
            \frac{u}{2}+\frac{v}{6},
            \frac17\right\}.
\end{array}
\]
\end{lemma}

\begin{proof}
For a fixed colour, the pattern LP has one capacity constraint for each of
\(A,U,W\). A copy of \(D\) contributes its numbers of vertices in these
three parts. Enumerating the possible locations of the two centres, and then
of their five leaves, gives the finite primal LP.

The coordinatewise-minimal vertices of the corresponding three-variable
dual LP are as follows:
\[
\begin{array}{c|c|c}
\text{colours inside }U,W
&
\text{red dual vertices}
&
\text{blue dual vertices}
\\ \hline
\mathrm{R},\mathrm{R}
&
(0,\tfrac17,\tfrac12),\
(\tfrac17,\tfrac17,\tfrac17)
&
(0,\tfrac13,0),\
(\tfrac13,0,\tfrac13),\
(\tfrac17,\tfrac17,\tfrac17)
\\[1ex]
\mathrm{R},\mathrm{B}
&
(0,\tfrac17,\tfrac13),\
(\tfrac13,\tfrac17,0),\
(\tfrac17,\tfrac17,\tfrac17)
&
(0,\tfrac47,\tfrac17),\
(0,\tfrac13,\tfrac29),\
(\tfrac13,0,\tfrac12),\
(\tfrac17,\tfrac17,\tfrac17)
\\[1ex]
\mathrm{B},\mathrm{R}
&
(0,0,\tfrac12),\
(\tfrac17,0,\tfrac17)
&
(0,\tfrac12,0),\
(\tfrac17,\tfrac17,\tfrac17)
\\[1ex]
\mathrm{B},\mathrm{B}
&
(0,0,\tfrac13),\
(\tfrac13,0,0),\
(\tfrac17,0,\tfrac17)
&
(0,\tfrac47,\tfrac17),\
(0,\tfrac12,\tfrac16),\
(\tfrac17,\tfrac17,\tfrac17).
\end{array}
\]
Here the coordinates are ordered as \(A,U,W\). Taking the scalar product
with \((x,u,v)\), and applying strong duality, gives the table.
\end{proof}

\begin{lemma}\label{lem:D23lower}
Every \(3\)-strip \(3\)-colouring \(\psi\) of \(K_n\) satisfies
$\nu_D(\psi)\ge \frac{7}{78}n-O(1).$
\end{lemma}

\begin{proof}
Put
$\alpha:=\frac{7}{78}.$
Since the number of strip parts and copy types is fixed, it is enough to
prove that the fractional monochromatic tiling number is at least \(\alpha\).

Pad the strip partition by empty parts if necessary, and let its normalized
part sizes be
\[
x\ge y\ge z,
\qquad
x+y+z=1.
\]
Suppose, for a contradiction, that every colour has fractional
\(D\)-tiling number smaller than \(\alpha\).

The monochromatic clique on the largest part gives \(x/7<\alpha\), so
$x<7\alpha.$
The pair formed by the two largest parts is a monochromatic complete
bipartite graph. Its \(D\)-tiling number is at least
$\min\left\{\frac{x+y}{7},\frac{y}{3}\right\}.$
Since \(x+y=1-z\ge2/3\), the first term is larger than \(\alpha\).
Consequently,
$y<3\alpha.$
It follows that
\[
z=1-x-y>1-10\alpha=\frac8{78}
\qquad\text{and}\qquad
x\ge1-2y>1-6\alpha.
\]

Let \(c_0\) be the colour inside the largest part. We claim that \(c_0\)
occurs on no other pair of strip parts, including the two remaining diagonal
pairs.

If \(c_0\) also occurs inside another part, the two monochromatic cliques
have total size at least \(x+z=1-y>7\alpha\). If \(c_0\) occurs between the
largest part and another part, the resulting monochromatic split graph
contains a \(D\)-tiling of normalized size at least
\[
\min\left\{\frac{x}{2},\frac{x+z}{7}\right\}>\alpha.
\]
Finally, if \(c_0\) occurs between the two smaller parts, we may combine a
tiling inside the largest part with a tiling in the resulting complete
bipartite graph. Its normalized size is at least
\[
\frac{x}{7}
+
\min\left\{\frac{y+z}{7},\frac{z}{3}\right\}
>
\alpha,
\]
using \(x>1-6\alpha\) and \(z>1-10\alpha\). This proves the claim.

Thus all five remaining entries of the strip template use only the other two
colours, which we call red and blue. The two cross-pairs incident with the
largest part must have different colours. Otherwise they form a monochromatic
complete bipartite graph between the largest part and the union of the other
two parts; both sides have size greater than \(3\alpha\), so it contains an
\(\alpha n-O(1)\)-sized \(D\)-tiling.

The third cross-pair agrees in colour with one of these two pairs. Relabel the
two smaller parts as \(U,W\) so that
$AU,\ UW$  are blue,
$AW$  is red,
where \(A\) is the largest part. Write the corresponding normalized sizes as
\(x,u,v\). We have
\[
x<7\alpha,\qquad
u,v<3\alpha,\qquad
x>1-6\alpha.
\]
The edges inside \(U\) and \(W\) are each red or blue. We consider the four
possibilities in Lemma~\ref{lem:D23-threepart-LP}.

\smallskip

\noindent\textbf{Case 1: the two diagonal colours are
\(\mathrm{R},\mathrm{R}\).}
The blue tiling bound gives \(u<3\alpha\), while the red bound gives
$\frac{u}{7}+\frac{v}{2}<\alpha.$
Therefore
\[
1=x+u+v
<
7\alpha+u+2\alpha-\frac{2u}{7}
<
9\alpha+\frac{15\alpha}{7}
=
1,
\]
a contradiction.

\smallskip

\noindent\textbf{Case 2: the diagonal colours are
\(\mathrm{B},\mathrm{R}\).}
The blue and red bounds imply, respectively,
\(u<2\alpha\) and \(v<2\alpha\). Hence
\[
1=x+u+v<7\alpha+4\alpha=11\alpha<1,
\]
a contradiction.

\smallskip

\noindent\textbf{Case 3: the diagonal colours are
\(\mathrm{B},\mathrm{B}\).}
The red bound gives \(v<3\alpha\). From the blue bound, either
$4u+v<7\alpha$
 or 
$3u+v<6\alpha.$
In the first case,
\[
u+v
=
\frac{4u+v}{4}+\frac{3v}{4}
<
\frac{7\alpha}{4}+\frac{9\alpha}{4}
=
4\alpha.
\]
The same conclusion follows in the second case from
\[
u+v
=
\frac{3u+v}{3}+\frac{2v}{3}
<
2\alpha+2\alpha.
\]
Thus \(1=x+u+v<11\alpha<1\), again a contradiction.

\smallskip

\noindent\textbf{Case 4: the diagonal colours are
\(\mathrm{R},\mathrm{B}\).}
The red bound gives
$3u+7v<21\alpha.$
The blue bound gives either \(4u+v<7\alpha\) or
\(3u+2v<9\alpha\). In the first case,
\[
u+v
=
\frac4{25}(4u+v)+\frac3{25}(3u+7v)
<
\frac{91}{25}\alpha
<
4\alpha.
\]
In the second case,
\[
u+v
=
\frac4{15}(3u+2v)+\frac1{15}(3u+7v)
<
\frac{19}{5}\alpha
<
4\alpha.
\]
Once more, \(1=x+u+v<11\alpha<1\), a contradiction.

All possibilities lead to a contradiction, proving that some colour has
fractional \(D\)-tiling number at least \(\alpha\). Rounding the fixed
pattern LP loses only \(O(1)\) copies.
\end{proof}

\begin{proof}[Proof of Theorem~\ref{thm:D23}]
Lemma~\ref{lem:D23upper} gives the upper bound. For the lower bound, let
\(\chi\) be an arbitrary \(3\)-colouring of \(K_n\). By
Theorem~\ref{thm:graphstructure}, there exists a \(3\)-strip colouring
\(\psi\) of \(K_n\) such that
$\nu_D(\chi)\ge\nu_D(\psi)-o(n).$
Lemma~\ref{lem:D23lower} therefore gives
$\nu_D(\chi)
\ge
\frac{7}{78}n-o(n).$
\end{proof}


\section{Clique tilings in \(t\)-graphs}
\label{subsec:3-clique}

We next prove Proposition~\ref{prop:clique-tiling-upper}. That is, for clique tilings in \(t\)-graphs, with at least
three colours and clique order strictly larger than the uniformity, the
elementary greedy bound is asymptotically sharp up to an additive constant.

\begin{lemma}\label{lem:pair-colouring}
For every \(r\ge3\), there exists a map
$\rho:\binom{[r]}2\to[r]$
such that:
\begin{enumerate}
\item
\(\rho(\{i,j\})\notin\{i,j\}\) for all distinct \(i,j\in[r]\);

\item
for every three distinct \(a,b,c\in[r]\), the values
$\rho(\{a,b\})$, $\rho(\{a,c\})$,
 $\rho(\{b,c\})$
are not all equal.
\end{enumerate}
\end{lemma}

\begin{proof}
If \(r\) is odd, identify \([r]\) with \(\mathbb Z_r\) and define
$\rho(\{i,j\})=\frac{i+j}{2}\pmod r.$
Since \(2\) is invertible in \(\mathbb Z_r\), this is well-defined and
\(\rho(\{i,j\})\notin\{i,j\}\). For a fixed colour \(c\), each vertex
\(a\) has at most one neighbour \(j\) satisfying
\((a+j)/2=c\). Thus every colour class is a matching, proving the second
property.

Now suppose that \(r\) is even. Identify the colour set with
\(\mathbb Z_{r-1}\cup\{\infty\}\). For distinct
\(i,j\in\mathbb Z_{r-1}\), define
$\rho(\{i,j\})=\frac{i+j}{2}\pmod{r-1},$
and define
$\rho(\{\infty,i\})=i+1\pmod{r-1}.$
The first property is immediate. A triangle contained in
\(\mathbb Z_{r-1}\) is not monochromatic by the odd case, while a triangle
\(\{\infty,a,b\}\) is not monochromatic because its two edges incident with
\(\infty\) have the distinct colours \(a+1\) and \(b+1\).
\end{proof}

\begin{lemma}\label{lem:shadow-separating-colouring}
Let \(t\ge3\) and \(r\ge3\). There exists a map \(\gamma\) from the family
of size-\(t\) multisets on \([r]\) to \([r]\) such that:
\begin{enumerate}
\item
\(\gamma(i^t)=i\) for every \(i\in[r]\), where \(i^t\) denotes the
multiset consisting of \(t\) copies of \(i\);

\item
if \(A\) is a size-\((t+1)\) multiset which is not of the form
\(i^{t+1}\), then the values
\[
\gamma(A-x),
\qquad x\in A,
\]
are not all equal, where \(A-x\) means that one occurrence of \(x\) is
deleted.
\end{enumerate}
\end{lemma}

\begin{proof}
Fix the usual order on \([r]\). For a multiset \(M\), let
\[
O(M):=
\{i\in[r]:\text{the multiplicity of \(i\) in \(M\) is odd}\}.
\]

Suppose first that \(t\) is odd. Then \(O(M)\ne\varnothing\) for every
size-\(t\) multiset \(M\), so define
\(\gamma(M):=\min O(M)\). Clearly \(\gamma(i^t)=i\).

Let \(A\) be a nonconstant size-\((t+1)\) multiset. If \(O(A)=\varnothing\),
choose distinct \(x,y\in\operatorname{supp}(A)\). Then
\(\gamma(A-x)=x\ne y=\gamma(A-y)\). If \(O(A)\ne\varnothing\), its
cardinality is even and at least two. Let \(x<y\) be its two smallest
elements. Then
\(\gamma(A-x)=y\ne x=\gamma(A-y)\).

Now suppose that \(t\) is even, and let \(\rho\) be given by
Lemma~\ref{lem:pair-colouring}. Define
\[
\gamma(M)=
\begin{cases}
\min\operatorname{supp}(M),
& O(M)=\varnothing,\\
\rho(O(M)),
& |O(M)|=2,\\
\min O(M),
& |O(M)|\ge4.
\end{cases}
\]
Since \(|O(M)|\) is even, this is well-defined, and
\(\gamma(i^t)=i\).

Let \(A\) be a nonconstant size-\((t+1)\) multiset. Then \(|O(A)|\) is
odd. If \(O(A)=\{u\}\), put
\(c:=\gamma(A-u)=\min\operatorname{supp}(A-u)\). If \(c\ne u\), then
\[
\gamma(A-c)=\rho(\{u,c\})\ne c=\gamma(A-u).
\]
If \(c=u\), choose \(y\in\operatorname{supp}(A)\setminus\{u\}\); then
\[
\gamma(A-y)=\rho(\{u,y\})\ne u=\gamma(A-u).
\]

If \(|O(A)|=3\), say \(O(A)=\{a,b,c\}\), deleting \(a,b,c\) produces the
three colours
$\rho(\{b,c\})$, 
$\rho(\{a,c\})$, 
$\rho(\{a,b\})$,
which are not all equal by Lemma~\ref{lem:pair-colouring}. Finally, if
\(|O(A)|\ge5\), let \(x<y\) be its two smallest elements. Then
\(\gamma(A-x)=y\ne x=\gamma(A-y)\).
\end{proof}

\begin{proof}[Proof of Proposition~\ref{prop:clique-tiling-upper}]
Let \(R:=R_r(K_p^{(t)})\). Greedily remove monochromatic copies of
\(K_p^{(t)}\) until fewer than \(R\) vertices remain. This produces at least
\((n-R)/p\) vertex-disjoint copies in total, so one colour contains at least
$\frac{n}{rp}-O_{r,t,p}(1)$
of them. This proves the lower bound.

For the upper bound, let \(\gamma\) be given by
Lemma~\ref{lem:shadow-separating-colouring}, and partition
$V(K_n^{(t)})=V_1\cup\cdots\cup V_r$
as evenly as possible. For a \(t\)-edge \(e\), let \(M(e)\) be the multiset
of its part labels, and define
\[
\chi(e):=\gamma(M(e)).
\]

We claim that no monochromatic \(K_{t+1}^{(t)}\) meets more than one part.
Indeed, let \(S\) be a \((t+1)\)-set meeting at least two parts, and let
\(A(S)\) be the multiset of its part labels. The colours of the
\(t\)-subsets of \(S\) are precisely
$\gamma(A(S)-x),$
as one occurrence \(x\) ranges over \(A(S)\). These colours are not all
equal by Lemma~\ref{lem:shadow-separating-colouring}.

Since \(p\ge t+1\), every set of \(p\) vertices meeting at least two parts
contains a \((t+1)\)-subset meeting at least two parts. Consequently, every
monochromatic \(K_p^{(t)}\) is contained in one \(V_i\). Moreover, all
\(t\)-edges inside \(V_i\) have colour
\(\gamma(i^t)=i\), so every monochromatic \(K_p^{(t)}\)-tiling of colour
\(i\) is contained in \(V_i\). It therefore has at most
$\left\lfloor\frac{|V_i|}{p}\right\rfloor
\le
\frac{n}{rp}+O_{r,p}(1)$
copies. This proves the upper bound.
\end{proof}
\end{document}