Extremal Problems in Graph Theory

and a Hypergraph Packing Problem

Attila Dankovics

A thesis submitted for the degree of

Doctor of Philosophy

Department of Mathematics
The London School of Economics
and Political Science

22nd December 2021

 Declaration

I certify that the thesis I have presented for examination for the PhD degree of
the London School of Economics and Political Science is solely my own work,
with the exceptions outlined below.
The copyright of this thesis rests with the author. Quotation from it is permitted,
provided that full acknowledgement is made. In accordance with the regulations, I
have deposited an electronic copy of it in LSE Theses Online held by the British
Library of Political and Economic Science and have granted permission for my
thesis to be made available for public reference. Otherwise, this thesis may not
be reproduced without my prior written consent. I warrant that this authorisation
does not, to the best of my belief, infringe the rights of any third party.
I declare that this thesis consists of 67742 words.

Statement of co-authored work

I confirm that Chapter 2 and Section 1.1 are based on [14], which is my own
work.
I confirm that Chapter 3 and Section 1.2 are joint work with Fabio Botler, Jan
Corsten, Nóra Frankl, Hiê.p Hàn, Andrea Jiménez and Jozef Skokan.
I confirm that Chapter 4 and Section 1.3 are joint work with Peter Allen and
Julia Böttcher.

2
 Abstract

This thesis improves on the best result of an open problem, showing that Hamilto-
nian graphs with low independence number are pancyclic. It also describes graphs
with the most triangle-free 5- and 6-colourings and generalises a packing theorem
for degenerate graphs from graphs to hypergraphs.
A graph on 𝑛 vertices is called pancyclic if it contains a cycle of every length
3 ≤ ℓ ≤ 𝑛. Given a Hamiltonian graph 𝐺 with independence number at most 𝑘,
we are looking for the minimum number of vertices 𝑓 (𝑘) that guarantees that 𝐺 is
pancyclic. The problem of finding 𝑓 (𝑘) was raised by Erdős in 1972, who showed
that 𝑓 (𝑘) ≤ 4𝑘 4 , and conjectured that 𝑓 (𝑘) = Θ(𝑘 2 ). Improving on a result of Lee
and Sudakov, we show that 𝑓 (𝑘) = 𝑂 (𝑘 11/5 ).
Let 𝑘 ≥ 3 and 𝑟 ≥ 2 be natural numbers. For a graph 𝐺, let 𝐹 (𝐺, 𝑘, 𝑟) denote
the number of colourings of the edges of 𝐺 with colours 1, . . . , 𝑟 such that, for
every colour 𝑐 ∈ {1, ..., 𝑟 }, the edges of colour 𝑐 contain no complete graph on 𝑘
vertices 𝐾 𝑘 . Let 𝐹 (𝑛, 𝑘, 𝑟) denote the maximum of 𝐹 (𝐺, 𝑘, 𝑟) over all graphs 𝐺
on 𝑛 vertices. The problem of determining 𝐹 (𝑛, 𝑘, 𝑟) was first proposed by Erdős
and Rothschild in 1974, and has so far been solved only for 𝑟 = 2, 3, and a small
number of other cases. In this thesis we consider the question for the cases 𝑘 = 3
and 𝑟 = 5 or 𝑟 = 6. We approximately determine the value 𝐹 (𝑛, 3, 5) and 𝐹 (𝑛, 3, 6)
for large values of 𝑛. We also prove a stability result for both cases. This is joint
work with F. Botler, J. Corsten, N. Frankl, H. Hàn, A. Jiménez and J. Skokan.
Given 𝐷 ≥ 1, whenever n is sufficiently large, if we are given any family of
𝑛
D-degenerate graphs of individual orders at most n, with maximum degree 𝑐 log(𝑛) ,
𝑛

and total number of edges at most (1−𝜀) 2 , then the family packs into the complete
graph 𝐾𝑛 , as proved by Allen, Böttcher, Hladký, and Piguet. If we add the condition
that a linear fraction of the degenerate graphs have linearly many leaves, we can


weaken the condition on the total number of edges to at most 𝑛2 and still obtain a
packing of the family into 𝐾𝑛 , as proved by Allen, Böttcher, Clemens and Taraz. In
this thesis we generalise both results to hypergraphs of any given uniformity. This
is joint work with P. Allen and J. Böttcher.

3
 Acknowledgements

First and foremost, I would like to thank my supervisors, Prof. Jozef Skokan,
Prof. Julia Böttcher and Prof. Peter Allen for the recommendation of problems,
their valuable advice and their continuous support during my PhD studies. I would
like to thank them for motivating me during difficult times and for encouraging me
to participate in workshops and conferences around the world.
I would like to thank LSE for providing me with a scholarship that made my
PhD studies possible. I would like to thank Kate Barker, Enfale Farooq, Rebecca
Lumb, Sarah Massey and Ed Perrin for their administrative support and for creating
a professional environment I enjoyed working in.
I would like to thank all my fellow PhD students at LSE for being great to work
with and all the fun we had together. I would especially like to thank my close
friends from the department, Jan Corsten, Keat Eng Hng and Stanislav Kucera.
I would like to thank my other collaborators, Fabio Botler, Nóra Frankl, Hiê.p
Hàn and Andrea Jiménez. They have been a pleasure to work with, and I learned
from them a lot.
I would like to thank Rita Alexiev and T Venczel for proofreading some of my
work.
Finally, I would like to thank my parents and brother for supporting me through-
out my life.

4
Contents

1 Introduction 7
1.1 Low Independence Number and Hamiltonicity Implies Pancyclicity 8
1.2 Maximum Number of Triangle-free Edge Colourings . . . . . . . 10
1.3 Packing Degenerate Hypergraphs . . . . . . . . . . . . . . . . . . 13

2 Low Independence Number and Hamiltonicity Implies Pancyclicity 18

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Definitions, earlier results . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Key lemmas, proof of the main theorem . . . . . . . . . . . . . . 23

3 Maximum Number of Triangle-free Edge Colourings 30

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 An approximate upper bound for many colours . . . . . . . . . . 34
3.4 An approximate upper bound for five and six colours . . . . . . . 35
3.4.1 Upper bound proof for five colours . . . . . . . . . . . . . 36
3.4.2 Upper bound proof for six colours . . . . . . . . . . . . . 40
3.5 Extremal configurations . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.1 Extremal configurations for five colours . . . . . . . . . . 43
3.5.2 Extremal configurations for six colours . . . . . . . . . . 44
3.6 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.6.1 Stability proof for five colours . . . . . . . . . . . . . . . 45
3.6.2 Stability proof for six colours . . . . . . . . . . . . . . . 51

4 Packing Degenerate Hypergraphs 57

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
 Contents

4.2.2 Probabilistic tools . . . . . . . . . . . . . . . . . . . . . . 64

4.2.3 Degenerate hypergraphs . . . . . . . . . . . . . . . . . . 66
4.2.4 Random matchings in bipartite graphs . . . . . . . . . . . 67
4.3 Almost perfect hypergraph packing . . . . . . . . . . . . . . . . . 70
4.3.1 Reducing the first main theorem . . . . . . . . . . . . . . 70
4.4 Staying on a diet . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5 Maintaining quasirandomness . . . . . . . . . . . . . . . . . . . 94
4.5.1 Initial coquasirandomness . . . . . . . . . . . . . . . . . 94
4.5.2 Maintaining coquasirandomness . . . . . . . . . . . . . . 95
4.6 Completing spanning embeddings . . . . . . . . . . . . . . . . . 113
4.7 Quasirandom packing . . . . . . . . . . . . . . . . . . . . . . . . 120
4.8 Perfect packings . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.8.1 The proof of Theorem 4.8.1 . . . . . . . . . . . . . . . . 143
4.8.2 Proof of the orientation lemma . . . . . . . . . . . . . . . 149
4.8.3 Proof of the almost perfect packing lemma . . . . . . . . 154

6
Introduction
1
Extremal problems in graph theory refer to problems of the following form. Given
a family of graphs G and a graph parameter 𝑇, for which graph 𝐺 ∈ G is 𝑇 (𝐺)
maximal?
We can also ask what the value of this maximal 𝑇 (𝐺) is. More broadly, any
statement of the form "For all 𝐺 ∈ G we have 𝑇 (𝐺) ≤ 𝑐" or "There exists 𝐺 ∈ G
such that 𝑇 (𝐺) ≥ 𝑐" are statements belonging to the field of extremal graph theory.
One of the most well known classical extremal graph theory results is due to
Mantel.

Theorem 1.0.1 (Mantel

j 2 k [42]). Graphs with 𝑛 vertices that contain no triangles
have no more than 𝑛4 edges.

In this case, the family of graphs is "graphs with 𝑛 vertices that contain no
triangles" and the parameter is "number of edges". Mantel also proved that this
result is sharp. Indeed the balanced
j 2 k complete bipartite graph on 𝑛 vertices fulfills
all conditions and has exactly 𝑛4 edges.
In Chapter 2 and Chapter 3, we consider questions of this format. A graph is
Hamiltonian if it contains a cycle going through all vertices. A graph on 𝑛 vertices
is called pancyclic if it contains a cycle of every length 3 ≤ ℓ ≤ 𝑛. It is widely
studied what cycle lengths must show up in graphs. One graph property of interest
is the sum of the reciprocals of cycle lengths in the graph, introduced by Erdős and
Hajnal [20].
In Chapter 2, the family of graphs is "Hamiltonian but not pancyclic graphs on
𝑛 vertices" and the parameter is independence number. This problem was also
introduced by Erdős, and we improve on the best known result. In Chapter 3, the
family is "graphs on 𝑛 vertices" while the parameter is "number of triangle free
colourings with 5 (or 6) colours".
 1 Introduction

In Section 1.1 and Section 1.2, we introduce these problems in more detail,
provide some background on the history and relevance of the problems and state
our main results on the topics.
Graph packing problems can be formulated the following way. Given a set
of graphs 𝐺 1 , . . . , 𝐺 𝑘 and a host graph 𝐻, can we embed the vertices of each of
𝐺 1 , . . . , 𝐺 𝑘 into 𝐻 such that edges are assigned to edges and each edge of 𝐻 is used
at most once. Many classical problems can be phrased as graph packing problems.
For example an equivalent definition of Hamiltonicity is the following. A graph 𝐺
on 𝑛 vertices is Hamiltonian if the complement of 𝐺 and the cycle 𝐶𝑛 pack into
the complete graph 𝐾𝑛 . Especially useful are results that not only state that such
an assignment is possible, but also provide an efficient algorithm.
A packing is perfect if each edge of the host graph is used exactly once. A
classical packing problem is whether the complete graph can be perfectly packed
with triangles. The first solution is by Kirkman [38], who proved that if the
obvious divisibility requirements are met, this is always possible. As the question
got popularised by Steiner [54], these packings are known as Steiner triples.
Hypergraphs are a generalisation of graphs. While in graphs each edge connects
exactly two vertices, in hypergraphs an edge can contain any number of vertices.
Packing questions can be more generally asked for hypergraphs. In Section 1.3,
we pose such a problem, and in Chapter 4, we provide a simple algorithm with a
complicated analysis that generates the required packing.

1.1 Low Independence Number and Hamiltonicity

Implies Pancyclicity
A Hamilton cycle of a graph is a cycle that passes through all its vertices. A graph
is Hamiltonian if it contains a Hamilton cycle as a subgraph. It is difficult to decide
whether a graph contains a Hamilton cycle, therefore it is valuable to establish
useful sufficient conditions for Hamiltonicity. The most well known sufficient
condition is by Dirac [15].

Theorem 1.1.1 (Dirac [15]). Let 𝐺 be a graph on 𝑛 vertices. If each vertex of 𝐺

has at least 𝑛2 neighbours, then 𝐺 is Hamiltonian.

A graph is pancyclic if it contains a cycle of every length 3 ≤ ℓ ≤ 𝑛, where 𝑛

denotes the number of vertices. By definition pancyclicity implies Hamiltonicity.

8
 1 Introduction

Although the converse is not true, it is often the case that conditions that imply
Hamiltonicity turn out to also imply pancyclicity. A famous meta-conjecture of
Bondy [10] states that almost all non-trivial sufficient conditions of Hamiltonicity
also imply pancyclicity with the possible exception of a few graphs. An example
of this is the following theorem by Bondy [9].

Theorem 1.1.2 (Bondy [9]). Let 𝐺 be a graph on 𝑛 vertices. If each vertex of

𝐺 has at least 𝑛2 neighbours, then 𝐺 is either pancyclic or the complete balanced
bipartite graph 𝐾 𝑛2 , 𝑛2 .

The independence number of a graph 𝐺 is the size of the largest stable set, denoted
by 𝛼(𝐺). A famous result of Chvátal and Erdős [12] states that if 𝜅(𝐺) ≥ 𝛼(𝐺),
where 𝜅(𝐺) is the vertex connectivity of 𝐺, then 𝐺 is Hamiltonian. Keevash and
Sudakov [35] showed that the similar but stronger condition 𝜅(𝐺) ≥ 600𝛼(𝐺), is
sufficient to conclude pancyclicity.
In this thesis we study a connection between Hamiltonicity, pancyclicity and
independence number. Assuming 𝐺 is a Hamiltonian graph with independence
number at most 𝑘, we are looking for the minimum number of vertices 𝑓 (𝑘) that
guarantees that 𝐺 is pancyclic. The problem of finding 𝑓 (𝑘) was raised by Erdős,
who showed that 𝑓 (𝑘) ≤ 4𝑘 4 and conjectured that a stronger statement holds.

Conjecture 1.1.3 (Erdős [19]). There are constants 𝑐 1 and 𝑐 2 such that for all 𝑘
we have 𝑐 1 𝑘 2 ≤ 𝑓 (𝑘) ≤ 𝑐 2 𝑘 2 .

The following construction due to Erdős provides a lower bound for this conjec-
ture. Let 𝑄 1 , . . . , 𝑄 𝑘 be cliques of size 𝑘 − 2 and add an edge between successive
cliques (including between the last and the first) such that these edges are independ-
ent. It is easy to check that this graph is Hamiltonian, has independence number 𝑘
and 𝑘 (𝑘 − 2) vertices, but does not contain a cycle of length 𝑘 − 1.
The result of Erdős was later improved by Keevash and Sudakov [35], who
showed that 𝑓 (𝑘) ≤ 150𝑘 3 holds and by Lee and Sudakov [40], who proved that
𝑓 (𝑘) = 𝑂 (𝑘 7/3 ) holds.
Here we improve their results.

Theorem 1.1.4. There exists 𝑐 > 0, such that if 𝐺 is a Hamiltonian graph with
𝑛 ≥ 𝑐𝑘 11/5 vertices and independence number at most 𝑘, then 𝐺 is pancyclic. In
other words 𝑓 (𝑘) ≤ 𝑐0 𝑘 11/5 .

9
 1 Introduction

1.2 Maximum Number of Triangle-free Edge

Colourings
A fundamental theorem of graph theory by Turán [55] asserts that among the graphs
on 𝑛 vertices that do not contain a complete graph on 𝑘 vertices 𝐾 𝑘 , the complete
balanced (𝑘 − 1)-partite graph, also known as the Turán graph 𝑇𝑘−1 (𝑛), has the
largest number of edges 𝑡 𝑘−1 (𝑛). Clearly, no matter what subset of edges from a
Turán graph we take, the resulting graph is also 𝐾 𝑘 -free. A natural question is:
How many 𝐾 𝑘 -free graphs are there? In 1976 Erdős, Kleitman and Rothschild
proved that the number of 𝐾 𝑘 -free graphs is asymptotically the same as the number
of subgraphs of 𝑇𝑘−1 (𝑛) [22].
Similarly, no matter how we colour the edges of the Turán graph, the edges of
the same colour form a graph with no 𝐾 𝑘 in it. We call such a colouring 𝐾 𝑘 -free.
Hence 𝑇𝑘−1 (𝑛) has 𝑟 𝑡 𝑘−1 (𝑛) 𝐾 𝑘 -free colourings with 𝑟 colours. A natural question is
whether we can find a graph with more 𝐾 𝑘 -free colourings, and if yes, what’s the
largest possible number of such colourings.
Let 𝑘 ≥ 3 and 𝑟 ≥ 2 be natural numbers. By a colouring of a graph 𝐺 = (𝑉, 𝐸)
with 𝑟 colours here we mean edge-colouring, that is a function 𝑓 : 𝐸 → {1, . . . , 𝑟 }.
In this context we refer to the numbers 1, . . . , 𝑟 as colours and by the colour class
of 𝑐 we mean 𝑓 −1 (𝑐). A colouring is 𝐾 𝑘 -free, if no colour class contains a copy of
𝐾 𝑘 . For a graph 𝐺, let 𝐹 (𝐺, 𝑘, 𝑟) denote the number of 𝐾 𝑘 -free colourings of 𝐺
with 𝑟 colours. Let 𝐹 (𝑛, 𝑘, 𝑟) denote the maximum of 𝐹 (𝐺, 𝑘, 𝑟) over all graphs
𝐺 on 𝑛 vertices. Let 𝑇𝑘−1 (𝑛) denote the Turán graph on 𝑛 vertices with clique size
𝑘 − 1. Let 𝑡 𝑘−1 (𝑛) denote the number of edges of 𝑇𝑘−1 (𝑛). Then the above lower
bound obtained from the Turán graph can be restated as

𝐹 (𝑛, 𝑘, 𝑟) ≥ 𝑟 𝑡 𝑘−1 (𝑛) . (1.2.1)

The problem of determining 𝐹 (𝑛, 𝑘, 𝑟) was first proposed by Erdős and Roth-
schild in 1974 [17, 18]. They conjectured that in the case 𝑘 = 3 and 𝑟 = 2 the lower
bound (1.2.1) is sharp for large enough 𝑛, and furthermore that 𝑇2 (𝑛) is the unique
extremal graph. Their conjecture was proved by Yuster [56] who also proved an
approximate version of the statement for general 𝑘 and 𝑟 = 2.
Improving on Yuster’s results, Alon, Balogh, Keevash and Sudakov fully resolved
the cases where 𝑟 = 2 and 𝑟 = 3 for large values of 𝑛.

10
 1 Introduction

Theorem 1.2.1 (Alon, Balogh, Keevash, Sudakov [4]). For 𝑘 ≥ 3 and 𝑛 ≥ 𝑛0 (𝑘),
we have 𝐹 (𝑛, 𝑘, 2) = 2𝑡 𝑘−1 (𝑛) and 𝐹 (𝑛, 𝑘, 3) = 3𝑡 𝑘−1 (𝑛) . Moreover, the correspond-
ing unique extremal graph is 𝑇𝑘−1 (𝑛).

In their paper [4] the authors also noted that the case 𝑟 > 3 is more challenging
as the behavior of 𝐹 (𝑛, 𝑘, 𝑟) changes. Indeed, they proved that if 𝑟 > 3 then
𝐹 (𝑛, 𝑘, 𝑟) is exponentially larger than 𝑟 𝑡 𝑘−1 (𝑛) . To prove this, they provided the fol-
lowing construction: in 𝑇𝑘−1 take two maximal independent sets, and replace them
with complete bipartite graphs. They also determined the approximate values of
𝐹 (𝑛, 3, 4) and 𝐹 (𝑛, 4, 4). Subsequently, Pikhurko and Yilma [47] improved on their
result, showing that 𝐹 (𝑛, 3, 4) = 𝐹 (𝑇4 (𝑛), 3, 4) and 𝐹 (𝑛, 4, 4) = 𝐹 (𝑇9 (𝑛), 4, 4) and
the corresponding extremal graphs are unique.
More recently, Pikhurko, Staden and Yilma [46] proved that for every 𝑛, 𝑘, 𝑟 there
is a complete multipartite graph 𝐺 such that 𝐹 (𝑛, 𝑘, 𝑟) = 𝐹 (𝐺, 𝑘, 𝑟). This graph is
not necessarily unique and not necessarily balanced. They also devised an optim-
isation problem with 𝑅𝑟 (𝑘) variables, where 𝑅𝑟 (𝑘) is the 𝑟-colour Ramsey-number
for 𝐾 𝑘 . Very recently, Pikhurko and Staden [45] proved a stability result, stating all
asymptotically optimal graphs are close to one of the solutions of their optimiza-
tion problem. Their proofs use Szemerédi’s regularity lemma and a symmetrisation
method.
In this thesis we construct many examples of approximately extremal graphs for
𝑘 = 3, 𝑟 = 5 and every 𝑛 ∈ N which are not Turán-graphs. However, all of them are
complete multipartite and for each 𝑛 there is a Turán-graph that is approximately
extremal as well. Whether there is always at least one extremal Turán-graph and
whether all extremal graphs are complete multipartite is still an open question.
Our approach and methods are different from that of Pikhurko, Staden and Yilma,
but our results can be interpreted in a way to draw parallels between the results.
The focus of Section 3.4 is to prove Theorem 3.4.3, which can be interpreted as
an optimization problem with only 2𝑟 variables (𝑟 ∈ {5, 6}) and extra constraints.
Our optimisation problem is different to the one proposed by Pikhurko et al. and
contains less variables. Then in Section 3.6, we solve this optimization problem
thus proving a stability result.
In the current thesis we consider the case 𝑘 = 3 and 𝑟 ∈ {5, 6}. Let 𝜑𝑟 (𝐺) =
𝐹 (𝐺, 3, 𝑟) and 𝜑𝑟 (𝑛) = 𝐹 (𝑛, 3, 𝑟). We find the approximate value of 𝜑5 (𝑛) and
𝜑6 (𝑛). We also prove stability results for both cases.
We use the following simplified definition of edit distance to state our stability

11
 1 Introduction

results.

Definition 1.2.2 (Edit distance). Given two graphs 𝐺 1 = (𝑉, 𝐸 1 ) and 𝐺 2 = (𝑉, 𝐸 2 )
on the same vertex set, we define their edit distance as 𝑑 (𝐺 1 , 𝐺 2 ) = |𝐸 1 4𝐸 2 |,
where 4 denotes the symmetrical difference of sets.

This definition is equivalent to only allowing edge deletion and edge insertion
of the usual edit operations, but it is sufficient for us.

Assumption 1.2.3. For simplicity we will assume that |𝑉 (𝐺)| is divisible by 24 in

the future. Thus when we define balanced graphs, we don’t have to consider parts
of different size. It will be easy to see that the error term introduced by this is
negligible.

We first state the technical stability result for 𝑟 = 6 as that is the easier of the
two.

Definition 1.2.4. Let 𝑉 be a set of vertices. We define 𝐺 6 (𝑉) as the set of balanced
complete 8-partite graphs that has vertex set 𝑉.

Theorem 1.2.5 (Stability for 𝑟 = 6). For every 0 < 𝛿 < 10−30 there is an 𝑛0
and 0 < 𝛿0 such that for all graphs 𝐺 = (𝑉, 𝐸) on 𝑛 > 𝑛0 vertices the following
2 2 2
holds. Suppose 𝜑6 (𝐺) ≥ 3𝑛 /4 43𝑛 /16−𝛿𝑛 . Then there is a 𝐺 0 ∈ 𝐺 6 (𝑉) such that
𝑑 (𝐺, 𝐺 0) < 𝛿0𝑛2 .

Calculating the number of 𝐾3 -free colourings for these graphs gives the following
corollary.
2 /4 2 /16+𝑜(𝑛2 )
Corollary 1.2.6. For every 𝑛 we have 𝜑6 (𝑛) ≤ 3𝑛 43𝑛 .

In our unpublished paper with Botler, Corsten, Frankl, Hàn, Jiménez and Skokan,
we prove that the complete balanced 8-partite graph is the unique extremal graph
for 𝑟 = 6. This is not part of the present thesis.
The case of 𝑟 = 5 is more complicated, as the structure of approximately optimal
graphs is more varied.

Definition 1.2.7. Let 𝑉 be a set of vertices with |𝑉 | = 𝑛. We define 𝐺 5 (𝑉) as the set
of complete 8-partite graphs that has vertex set 𝑉 and has the following part sizes:
(𝑛/4, 𝑛/4, 𝑎, 𝑎, 𝑏, 𝑏, 𝑛/4 − 𝑎 − 𝑏, 𝑛/4 − 𝑎 − 𝑏) for some 0 ≤ 𝑎, 𝑏 and 𝑎 + 𝑏 ≤ 𝑛/4;
or (𝑎, 𝑎, 𝑛/4 − 𝑎, 𝑛/4 − 𝑎, 𝑏, 𝑏, 𝑛/4 − 𝑏, 𝑛/4 − 𝑏) for some 0 ≤ 𝑎, 𝑏 ≤ 𝑛/4.
Note: In edge cases this might be a complete 4-partite or 6-partite graph instead
of 8-partite, these graphs are part of 𝐺 5 (𝑉) as well.

12
 1 Introduction

Theorem 1.2.8 (Stability for 𝑟 = 5). For every 0 < 𝛿 < 10−30 there is an 𝑛0 and
0 < 𝛿0 such that for all graphs 𝐺 = (𝑉, 𝐸) on 𝑛 > 𝑛0 vertices the following holds.
2 2
Suppose 𝜑6 (𝐺) ≥ 6𝑛 /4−𝛿𝑛 . Then there is a 𝐺 0 ∈ 𝐺 5 (𝑉) such that 𝑑 (𝐺, 𝐺 0) < 𝛿0𝑛2 .

Calculating the number of 𝐾3 -free colourings for these graphs gives the following
corollary.
2 /4+𝑜(𝑛2 )
Corollary 1.2.9. For every 𝑛 we have 𝜑5 (𝑛) ≤ 6𝑛 .

The fact that the aforementioned graphs all have asymptotically equal number of
𝐾3 -free colourings makes the case 𝑟 = 5 particularly difficult and interesting. We
are currently working on finding the exact solution for this case.

1.3 Packing Degenerate Hypergraphs

A packing of a family G = {𝐺 1 , . . . , 𝐺 𝑘 } of hypergraphs into a hypergraph 𝐻 is
a colouring of the edges of 𝐻 with the colours 0, 1, . . . , 𝑘 such that the edges of
colour 𝑖 form an isomorphic copy of 𝐺 𝑖 for each 1 ≤ 𝑖 ≤ 𝑘. The packing is perfect
if no edges have colour 0. We will often say an edge is covered in a packing if it
has colour at least 1, and uncovered if it has colour zero.
Packing problems have been studied for several decades. Classical theorems and
conjectures of extremal graph theory can often be written as packing problems.
For example, Turán’s theorem can be read as the statement that if the 𝑛-vertex 𝐺
does not have too many edges, then 𝐺 and 𝐾𝑟 pack into 𝐾𝑛 . However packings
in this context are usually very far from being perfect packings, with a large
fraction of 𝐸 (𝐻) uncovered. By contrast, in Chapter 4 we are interested in almost


perfect packings and perfect packings, that is, packings in which 𝑜 𝑒(𝐻) edges are
uncovered. One of the first problems asking for perfect packings in graphs is the
problem of Steiner-systems and it is over a century old. Plücker [48] in 1835 found
perfect packings of 13 𝑛𝑟 copies of 𝐾3 into 𝐾𝑛 for various values of 𝑛, and more

generally, Kirkman [38] in 1847 solved the problem for all values of 𝑛. Unaware
of Kirkman’s results, in 1853 Steiner re-asked the question [54] and popularised it.
More generally, one can ask the following question.

Question 1.3.1. Given 2 ≤ 𝑘 ≤ 𝑟, for which values of 𝑛 does the complete 𝑘-

uniform hypergraph 𝐾𝑛(𝑘) have a perfect packing with copies of 𝐾𝑟(𝑘) ?

A packing of this form is called a combinatorial design. There are some simple
divisibility conditions on 𝑛 which are necessary for an affirmative answer. Recently

13
 1 Introduction

and spectacularly, Keevash [32] proved that for sufficiently large 𝑛 these conditions
are also sufficient. He used a method called randomised algebraic construction,
setting aside a structure with algebraic properties allowing him to absorb whatever
is leftover after an almost perfect packing. Keevash reached an almost perfect
packing by using Rödl nibble.
Rödl nibble was first introduced by Rödl in 1985 [50]. Given the packing problem
in Question 1.3.1, the idea is to choose a small number of copies of 𝐾𝑟(𝑘) in the
host hypergraph, then throw out intersecting ones. Next, update our host graph and
repeat until we have an almost perfect packing. This method is more difficult to use
in problems where the graphs that we want to pack are larger, as they will almost
always intersect if we don’t take extra precautions. This is one of the reasons we
do not use Rödl nibble in our present work.
Keevash’s result was reproved independently, using a more combinatorial
method, by Glock, Kühn, Lo and Osthus [26], who were also able to extend
the result to pack with arbitrary fixed hypergraphs [26]. Their method, called
iterative absorption, was also based on absorption, but the structure they set aside
had basic combinatorial properties instead of algebraic. They applied this method
repeatedly to make the leftover part more and more structured. After the leftover
is structured enough, they can use an absorber, set aside at the start, that can ab-
sorb everything that’s left. The toroidal 𝑛-queens problem asks how many ways 𝑛
queens can be placed on a 𝑛𝑥𝑛 chessboard, where the board is considered on the
surface of the torus, such that no pair of queens attack each other. In 2021, Bowtell
and Keevash used a random greedy algorithm and a complex absorption method,
utilising ideas of the iterative absorption and randomised algebraic construction,
to assymptotically resolve the problem for 𝑛 ≡ 1, 5 (mod 6).
In general, a proof using the absorption method can be described as follows.
First, we set aside a structure in the host graph, called the absorber. Next, we
almost perfectly pack into the rest of the graph. Finally, we prove that whatever
is leftover, together with our absorber it can be perfectly packed into. This last
step usually relies on properties of the absorber established at their construction,
as well as properties of the leftover maintained through the almost perfect packing
process. Generally speaking, the stronger an absorber we can create, the less careful
we have to be in the almost perfect packing part. This method was used, described
and popularised under the name absorption by Rödl, Ruciński and Szemerédi in
2006 [51]. We note that a similar method was already used by Krivelevich in 1997

14
 1 Introduction

[39].
Intuitively, perfect packing results are hard precisely because every edge must
be used. If the hypergraphs G were embedded in order to 𝐻, on coming to the last
hypergraph of G we would need to find that a hole is left in 𝐻 of precisely the right
shape to accommodate it; this clearly requires some foresight in the packing. If
some edges remain uncovered at the end, this difficulty decreases.
In our work we consider 𝐷-degenerate hypergraphs. First, let us define the term.
An ordering of 𝑉 (𝐺) is 𝐷-degenerate if for each vertex 𝑣, there are at most 𝐷 edges
of 𝐺 whose final vertex is 𝑣. We say 𝐺 is 𝐷-degenerate if 𝑉 (𝐺) has a 𝐷-degenerate
ordering. In particular, we define trees as connected 1-degenerate hypergraphs.
We also need to define the maximum degree Δ(𝐺) := max𝑣∈𝑉 (𝐺) {𝑒 ∈ 𝐸 (𝐺) : 𝑣 ∈
𝑒} . Of course, every hypergraph of bounded maximum degree has automatically
bounded degeneracy. Note that an 𝑛-vertex 𝐷-degenerate hypergraph has less than
𝐷𝑛 edges, and so trivially has maximum degree at most 𝐷𝑛.
Our results can be interpreted as hypergraph analogues of the two widest known
𝑐𝑛
tree packing conjectures with a maximum degree log(𝑛) condition.

Conjecture 1.3.2 (Ringel’s conjecture). Given 𝑛 and a tree on 𝑛 + 1 vertices, 2𝑛 + 1

copies of this tree pack into the complete graph 𝐾2𝑛+1 .

Ringel stated this conjecture in 1968 [49]. Simple calculations show that the
packing in this conjecture is a perfect packing. Early results towards proving the
conjecture only packed simple trees, for example stars or paths. The first general
result was by Böttcher, Hladký, Piguet and Taraz in 2016 [11]. They proved an
almost perfect packing version of Ringel’s conjecture for bounded degree graphs.
Joos, Kim, Kühn and Osthus [31] proved the perfect packing result, assuming
bounded degree as well. Most recently, Montgomery, Pokrovskiy and Sudakov [43]
and later Keevash and Staden [33] proved Ringel’s conjecture for all sufficiently
large 𝑛.
Note, that both the almost perfect packing results of Böttcher, Hladký, Piguet
and Taraz [11] and the perfect packing results of Joos, Kim, Kühn and Osthus [31]
allow for much more general families of trees than Ringel’s conjecture.

Conjecture 1.3.3 (Gyárfás’ conjecture). Given 𝑛 and a family of trees 𝑇1 , . . . , 𝑇𝑛

with |𝑉 (𝑇𝑖 )| = 𝑖, the family packs into the complete graph 𝐾𝑛 .

Gyárfás stated this conjecture, also known as the tree packing conjecture, in 1978
[28]. Note, that once again the conjecture requires a perfect packing. Similarly to

15
 1 Introduction

Ringel’s conjecture, the initial results were about very specific tree-types. In 2013,
1
Balogh and Palmer [5] proved that the largest 𝑛 4 trees can be packed if none of them
are stars. This was an important step, as in general spanning or almost spanning
trees are much harder to pack than smaller trees. The first general almost perfect
packing result is again by Böttcher, Hladký, Piguet and Taraz [11] for bounded
degree graphs. Joos, Kim, Kühn and Osthus [31] in their already mentioned paper
also proved Gyárfás’ conjecture for all large values of 𝑛 and bounded degree trees.
Moving one step forward from trees, in 1967, attending a conference in Ober-
wolfach, Ringel posed the following problem.

Problem 1.3.4 (Oberwolfach problem). Given an odd number 𝑛 and a two-regular

graph 𝐹 on 𝑛 vertices, for what 𝑛 and 𝐹 can we perfectly pack copies of 𝐹 into 𝐾𝑛 ?

In 2021, Glock, Joos, Kim, Kühn and Osthus showed that this is always possible,
if 𝑛 is sufficiently large, no matter what 𝐹 is [27]. They use an absorption method,
utilising tools from several already mentioned papers, including Rödl nibble and
Keevash’s proof of the existence of designs. Later, Keevash and Staden solved
a generalised version of the Oberwolfach problem [34]. They proved that any
quasirandom dense large graph in which all degrees are equal and even can be
decomposed into any given collection of two-factors.
After this short detour to graph packing results, let us return to hypergraph
packings. In 2021, Ehard and Joos proved that a family of uniform bounded degree
hypergraphs packs into any quasirandom host graph almost perfectly. With this
result they address questions of Kim, Kühn, Osthus and Tyomkyn [37], as well
as Keevash [36]. This result is similar to our Theorem 4.7.1, which proves the
𝑐𝑛
same type of statement for bounded degeneracy and a maximum degree of log(𝑛) ,
therefore it can be applied to a wider family of graphs. Their results however apply
to the partite setting and sparser graphs.

Our results

In the entirety of Chapter 4 we consider hypergraphs that are 𝑟-uniform, that is

each edge of a hypergraph 𝐺 is an 𝑟-element subset of 𝑉 (𝐺). We will refer to
(𝑟 − 1)-element vertex sets as semi-edges.
Our first main result says that we can approximately pack the complete graph
(𝑟)
𝐾𝑛 with hypergraphs of bounded degeneracy and not too large maximum degree.

16
 1 Introduction

Theorem 1.3.5. For each 𝑟 ≥ 2, each 𝛾 > 0 and each 𝐷 ∈ N there exist 𝑐 > 0 and
a number 𝑛0 such that the following holds for each integer 𝑛 > 𝑛0 . Suppose that
(𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is a family of 𝐷-degenerate 𝑟-uniform hypergraphs, each of which has at
𝑐𝑛
most 𝑛 vertices and maximum degree at most log 𝑛 . Suppose further that the total
𝑛

number of edges of (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is at most 𝑟 − 𝛾𝑛 . Then (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] packs into the
𝑟

complete graph 𝐾𝑛(𝑟) .

Our second main result is a bit more complicated. If we insist that the graphs
𝐺 𝑡 each have linearly many vertices of degree 1, and in addition these graphs are
not too close to spanning, then we can upgrade ‘covering almost all the edges’ to a
perfect packing.

Theorem 1.3.6. For each 𝑟 ≥ 2, every 𝐷 and 𝜇 > 0 there are 𝑛0 and 𝑐 > 0 such
that for every 𝑛 ≥ 𝑛0 , the following holds. Suppose that (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is a family of
𝐷-degenerate 𝑟-uniform hypergraphs, each of which has at most 𝑛 − b𝜇𝑛c vertices,
𝑐𝑛
at least b𝜇𝑛c leaves and maximum degree at most log 𝑛 . Suppose further that the
𝑛

total number of edges of (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is exactly 𝑟 . Then (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] perfectly packs
into the complete graph 𝐾𝑛(𝑟) .

Note that with this theorem we also prove reasonable hypergraph analogues of
Ringel’s conjecture and Gyárfás’ conjecture for typical hypergraphs. Trees for us
are 1-degenerate hypergraphs. In Ringel’s conjecture all graphs are non-spanning,
which this theorem can handle. In case of Gyárfás’ conjecture, some graphs are
close to spanning, but the technical version of this theorem which is stated as
Theorem 4.1.4 can handle that. Our two extra restrictions are the requirement of
𝑐𝑛
𝜇𝑛 leaves as well as the maximum degree log 𝑛 . Both of these are true for typical
1-degenerate hypergraphs. Here by typical we mean that if we randomly generate
the back-edges in degeneracy order to get a tree, this tree will have these properties
with high probability.
Most of the work of Chapter 4 is to analyse a natural randomised algorithm
which packs almost-spanning hypergraphs. The graph version of this algorithm was
previously analysed by Allen, Böttcher, Hladký and Piguet [3] and further by Allen,
Böttcher, Clemens and Taraz [2]. They proved Theorem 1.3.5 and Theorem 1.3.6
for 𝑟 = 2 respectively. Some of the analysis carries over to hypergraphs, but there
are points where a new idea is needed which we will highlight.

17
Low Independence Number and
2
Hamiltonicity Implies Pancyclicity

2.1 Introduction
A Hamilton cycle of a graph is a cycle that passes through all its vertices. A graph
is Hamiltonian if it contains a Hamilton cycle as a subgraph. It is difficult to decide
whether a graph contains a Hamilton cycle, therefore it is valuable to establish
useful sufficient conditions for Hamiltonicity. A graph is pancyclic if it contains
a cycle of every length 3 ≤ ℓ ≤ 𝑛, where 𝑛 denotes the number of vertices. The
independence number of a graph 𝐺 is the size of the largest stable set, denoted by
𝛼(𝐺).
In this chapter we study a connection between Hamiltonicity, pancyclicity and
independence number. Assuming 𝐺 is a Hamiltonian graph with independence
number at most 𝑘, we are looking for the minimum number of vertices 𝑓 (𝑘) that
guarantees that 𝐺 is pancyclic.
In this section we introduce an equivalent theorem to Theorem 1.1.4.

Definition 2.1.1. For 𝛽 > 0, let 𝑆𝐶 (𝛽) denote the following statement, called short
cycle statement. There exists 𝑐 > 0 such that given a Hamiltonian graph 𝐺 with
𝑛 ≥ 𝑐𝑘 𝛽 vertices and independence number at most 𝑘, and any subset of vertices 𝑊
of size at most 20𝑘 2 , we can find a cycle of length 𝑛 − 1 containing all the vertices
from 𝑊.

Theorem 2.1.2 (Lee, Sudakov [40]). For all 𝛽 ≥ 2, assuming 𝑆𝐶 (𝛽) the following
statement holds. There exists 𝑐0 > 0, such that if 𝐺 is a Hamiltonian graph with
𝑛 ≥ 𝑐0 𝑘 𝛽 vertices and independence number at most 𝑘, then 𝐺 is pancyclic. In
other words 𝑓 (𝑘) ≤ 𝑐0 𝑘 𝛽 .
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

The above theorem is implicitly proved in [40]. To see this one follows their
Proof of Theorem 1.1. This gives the stronger conclusion if their Theorem 2.1 is
replaced by 𝑆𝐶 (𝛽).

Theorem 2.1.3. 𝑆𝐶 (11/5) holds.

Using Theorem 2.1.2 and Theorem 2.1.3, Theorem 1.1.4 immediately follows.
For a proof of Theorem 2.1.3 we substantially extend the methods of [40]. The
improvement comes from Lemma 2.3.9 and the inductive approach to proving
Lemma 2.3.5, which is made possible by Lemma 2.3.11. Proving these new
lemmas constitutes most of Section 2.3. Before that, we will state some definitions
and prove a basic structural proposition in Section 2.2.
The goal of Section 2.2 is to prove Theorem 2.1.3.

2.2 Definitions, earlier results

The core idea of the proof of Theorem 2.1.3, as well as those of previous results, is
the following. We break down the graph into parts along the Hamilton cycle, we call
these parts arcs. Next we show that if we have certain edge configurations between
these arcs, then we can also find a cycle of length 𝑛 − 1. Finally, we prove that a
graph contains either a sparse subgraph that contradicts the independence number
constraint, or a certain type of expander that implies the existence of one of the
aforementioned edge configurations. In this section we state the basic definitions
and prove Proposition 2.2.9, which shows the existence of the arc-system we will
use in Section 2.3.
We make no attempt to find the optimal value of 𝑐 in Theorem 1.1.4. For
this reason we can ignore small rounding errors and thus will omit all floor and
ceiling signs. We fix a large constant 𝑐, how large we actually need will come
from later calculations so we do not specify at this point. We can also assume that
𝑐𝑘 11/5 ≤ 4𝑘 4 by the result of Erdős [19], thus by setting 𝑐 large we can assume that
𝑘 is also large.

Assumption 2.2.1. From this point we assume for a contradiction to Theorem 2.1.3
the following:
• 𝐺 is a Hamiltonian graph with 𝑛 ≥ 𝑐𝑘 11/5 vertices,
• 𝐺 has independence number at most 𝑘,
• 𝑊 is a subset of 𝑉 (𝐺) with at most 20𝑘 2 vertices,

19
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

• 𝐺 has no cycle of length 𝑛 − 1 containing 𝑊,

• 𝐻 is a Hamilton cycle in 𝐺.

We call a vertex of 𝐺 problematic if it has degree at most 2𝑘 or is an element of

𝑊. There are at most 2𝑘 2 vertices with degree at most 2𝑘 (by the greedy algorithm
for finding independent sets), so there are at most 22𝑘 2 problematic vertices. Let
𝑊e denote the set of problematic vertices. The motivation for calling these vertices
problematic comes from Proposition 2.2.2.
We call a cycle 𝐶 a contradicting cycle if it has length 𝑛 − 𝑘 ≤ |𝐶 | ≤ 𝑛 − 1 and
contains all problematic vertices.
The following proposition shows that in the graphs we are considering no con-
tradicting cycle exists, thus justifying their name.

Proposition 2.2.2 ([40] Proposition 3.1). If 𝐺 satisfies Assumption 2.2.1, then there
is no contradicting cycle in 𝐺.

We say two vertices of 𝐺 are consecutive if they are neighbours in 𝐻. A set

of vertices is continuous if they form a path in 𝐻. For a subset 𝐴 ⊆ 𝑉 (𝐺) the
continuous closure of the set is the 𝐴 continuous set of minimum size that contains
it (in general this might not always be unique, but we will use it only in cases when
it is).
Next we define arc-systems, which are the objects that we will primarily use in
the rest of the section.

Definition 2.2.3. A family of subsets of 𝑉 (𝐺) (where the graph 𝐺 has a fixed
Hamilton cycle 𝐻) denoted by A is called an arc-system and its elements arcs if
the following properties hold:
(i) For all 𝐴 ≠ 𝐵 ∈ A, we have 𝐴 ∩ 𝐵 = ∅, that is, the continuous closures of
arcs are pairwise disjoint,
(ii) For all 𝐴 ∈ A, we have 𝐴 ∩ 𝑊e = ∅, that is, the continuous closure of each
arc has no problematic vertex in it,
(iii) For all 𝐴 ∈ A, no two vertices in 𝐴 are consecutive.

Remark 2.2.4. If 𝐴 is an arc and | 𝐴| ≤ 𝑘 + 2, then 𝐴 is an independent set. Indeed,

if there was an edge {𝑢, 𝑣} where 𝑢, 𝑣 ∈ 𝐴, then using the longer path between 𝑢
and 𝑣 in 𝐻 and the edge {𝑢, 𝑣} we would get a contradicting cycle (see Figure 2.1a).

Definition 2.2.5. We call an arc system A independent if it has the property that
for all 𝐴 in A we have | 𝐴| ≤ 𝑘/6.

20
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

We use 𝑘/6 in the definition instead of 𝑘 + 2 for technical reasons, as in later

proofs we will want to remove vertices from up to six arcs to find a contradicting
cycle.
We say the size of the arc-system is |A| and the length of the arc system is
min 𝐴∈A | 𝐴|.

Proposition 2.2.6. Given 𝑐 1 and 𝑐 2 , there exists 𝑐 such that if we assume Assumption
2.2.1, then there is an independent arc-system in the graph 𝐺 of size 𝑐 1 𝑘 2 and length
𝑐 2 𝑘 1/5 .

Proof. We start with the empty arc-system. Removing the problematic vertices
from 𝐻, we obtain a set P of at most 22𝑘 2 paths. From this set we will construct
an arc-system A with the desired properties.
While there is a path {𝑣 1 , 𝑣 2 , . . . , 𝑣 𝑚 } = 𝑃 ∈ P such that 𝑚 ≥ 2𝑐 2 𝑘 1/5 we do
the following. Remove 𝑃 from P. Add {𝑣 2𝑐2 𝑘 1/5 +1 , 𝑣 2𝑐2 𝑘 1/5 +2 , . . . , 𝑣 𝑚 } to P. Add
{𝑣 1 , 𝑣 3 , . . . , 𝑣 2𝑐2 𝑘 1/5 −1 } to A. In words, we remove the first 2𝑐 2 𝑘 1/5 vertices of 𝑃
and form an arc from every second vertex in it, and add that arc to A. Since we
choose such sets, by definition A fulfills properties (ii) and (iii) of arc-systems.
Also since 𝑘 is large enough we have 2𝑐 2 𝑘 1/5 ≤ 𝑘/6, so A is independent by
definition. The change to P ensures that property (i) will also hold for A.
At the end of this process we have at most (22𝑘 2 )(2𝑐 2 𝑘 1/5 ) leftover vertices
(from paths shorter than 2𝑐 2 𝑘 1/5 ), we removed 22𝑘 2 problematic vertices at the
start, and the half of the other vertices were used to form arcs in A. So we
11/5 2 ) (2𝑐 𝑘 1/5 +1)
have at least 𝑐𝑘 −(22𝑘 2𝑐 2 𝑘 1/5
2
arcs in A, which is more than 𝑐 1 𝑘 2 if 𝑐 is large
enough.

Definition 2.2.7. We denote by 𝑀2 the matching of size two, or equivalently two

independent edges. We say a graph is 𝑀2 -free if it doesn’t have two independent
edges (or equivalently in bipartite graphs, there is a vertex that is incident to all
edges).

Definition 2.2.8. We say an arc-system is simple if it is independent and for each

pair of arcs, the subgraph of 𝐺 induced by them is 𝑀2 -free.

Proposition 2.2.9. Given 𝑐 1 and 𝑐 2 , there exists 𝑐 such that if we assume Assump-
tion 2.2.1, then in the graph 𝐺 there is a simple arc-system of size 𝑐 1 𝑘 2 and length
𝑐 2 𝑘 1/5 .

21
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

𝑦2 𝑦1
𝑦2
𝑦1 𝑦4
𝑦3 𝑥4
𝑥3
𝑢 𝑣 𝑥1 𝑥2 𝑥1 𝑥2

(a) An arc with an edge (b) An intersecting 𝑀2 (c) Two non-intersecting

inside between arcs 𝑀2

Figure 2.1: Contradicting cycles implied by edge configurations

Proof. We fix an independent arc-system A provided by Proposition 2.2.6. We

draw the vertices of the graph 𝐺 on a circle in the plane in the order of the cycle 𝐻
and connect neighbouring vertices with line segments. If there are two independent
edges {𝑥 1 , 𝑦 1 }, {𝑥 2 , 𝑦 2 } between arcs 𝐴1 and 𝐴2 , then these can intersect on this
drawing or not.
If they intersect, then we immediately find a contradicting cycle the following
way. Starting from 𝐻, remove the edges of the shorter path between 𝑥 1 , 𝑥2 and,
similarly, remove the edges of the shorter path between 𝑦 1 , 𝑦 2 and instead add the
edges {𝑥 1 , 𝑦 1 } and {𝑥 2 , 𝑦 2 } (see Figure 2.1b). This gives a cycle with at least
𝑛 − 4𝑐 2 𝑘 1/5 > 𝑛 − 𝑘 vertices. By Proposition 2.2.2, this is a contradiction to
Assumption 2.2.1.
If we find two pairs of arcs each with non-intersecting 𝑀2 such that these two 𝑀2
({𝑥1 , 𝑦 1 }, {𝑥2 , 𝑦 2 } between 𝐴1 and 𝐴2 and {𝑥 3 , 𝑦 3 }, {𝑥 4 , 𝑦 4 } between 𝐴3 and 𝐴4 )
intersect each other on the drawing, then again we can find a contradicting cycle
the following way. From 𝐻 remove the edges of the shorter path between 𝑥 1 , 𝑥2 ;
𝑦 1 , 𝑦 2 ; 𝑥3 , 𝑥4 ; 𝑦 3 , 𝑦 4 and instead add the edges {𝑥1 , 𝑦 1 }, {𝑥 2 , 𝑦 2 }, {𝑥 3 , 𝑦 3 }, {𝑥4 , 𝑦 4 }
(see Figure 2.1c). This gives a cycle with at least 𝑛 − 8𝑐 2 𝑘 1/5 > 𝑛 − 𝑘 vertices. By
Proposition 2.2.2, this is a contradiction to Assumption 2.2.1.
This implies that if we look at the graph on the arcs as vertices where a pair of
arcs form an edge if the subgraph of 𝐺 induced by them contains an 𝑀2 , then this
is a planar graph. That implies 5-colourability. By taking the majority colour, we
get a simple arc-system with size 𝑐51 𝑘 2 and length 𝑐 2 𝑘 1/5 .

In some cases we want to consider an auxiliary graph, in which the arcs are the
vertices.

22
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

Definition 2.2.10. The arc-graph of an arc-system A is the graph 𝐺 A where the

vertices are the arcs in the system and there is an edge between two arcs in the
arc-graph if and only if there is an edge between the two arcs in 𝐺. We call the
edges of the arc-graph arc-edges.

2.3 Key lemmas, proof of the main theorem

The goal of this section is to prove Lemma 2.3.5, which states that arc-systems
with certain size and length contain large independent sets. This will imply
Theorem 1.1.4. To prove it, we will use induction and two structural lemmas.
Lemma 2.3.9 states that an arc system contains either a large independent set or
a subsystem of arcs that are expanding. Lemma 2.3.11 states that an expanding
arc system must have a certain edge-configuration that we will call a semi-triangle.
Finally we show that this implies the existence of a contradicting cycle, which is
impossible.
Given an arc-system A, let 𝐺 [A] denote the subgraph induced by all vertices
in the arcs of A.
Lemma 2.3.1. Given a simple arc-system A with length 𝑎, size 𝑏, and 𝑚 arc-edges
in the corresponding arc-graph, there is an independent set in 𝐺 [A] of size 𝑎𝑏 −𝑚.
Proof. Since the arc-system is simple, the edges of 𝐺 [A] corresponding to a single
arc-edge 𝑒 can be covered by a single vertex of 𝐺 [A] (that is, there is a vertex 𝑣
in 𝐺 [A] such that all the edges corresponding to 𝑒 are incident to 𝑣). Removing
these vertices we get an independent vertex set in 𝐺 [A] of size 𝑎𝑏 − 𝑚.

Now we define the function that we want to work with.

Definition 2.3.2. Let 𝑔(𝑎, 𝑏) denote the largest number such that given a simple
arc-system A of length 𝑎 and size 𝑏 there is always an independent set of size
𝑔(𝑎, 𝑏) in 𝐺 [A].
Setup 2.3.3. For all 𝑖 ∈ N we define the following constants that we will use in the
2
following section; 𝑎𝑖 = 10 · 3𝑖 , 𝑏𝑖 = 1000𝑖 · 4𝑖 .
The lemma that we want to prove in this section is the following.
Definition 2.3.4. For every 𝑝 ∈ N let 𝐴𝐼 ( 𝑝) denote the following statement, called
the arc-independence statement. For every 𝑥 ≥ 1 with 𝑎 𝑝 𝑥 ≤ 𝑘/6 we have
 
𝑝( 𝑝−1)/2
𝑔 𝑎 𝑝 𝑥, 𝑏 𝑝 𝑥 ≥ 𝑥𝑝 + 1 .

23
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

Lemma 2.3.5. For every 𝑝 ∈ N, 𝐴𝐼 ( 𝑝) holds.

 
This means that given Θ 𝑥 𝑝( 𝑝−1)/2 arcs with each having at least Θ(𝑥) vertices,
we can find an independent set of size Θ(𝑥 𝑝 ).
First, we prove Theorem 1.1.4 using Lemma 2.3.5.

Proof of Theorem 1.1.4. Using Lemma 2.3.5 for 𝑝 = 5 and 𝑥 = 𝑘 1/5 , together
with Proposition 2.2.9, we get an independent set of size 𝑘 + 1, and therefore a
contradiction. This proves Theorem 2.1.3 and thus Theorem 1.1.4.

To prove Lemma 2.3.5, we will use induction. To prepare for the induction step,
first we prove Lemmas 2.3.9 and 2.3.11.

Definition 2.3.6. Given an arc-system A of the graph 𝐺 and a subset of an arc

𝑋 ⊆ 𝐴 ∈ A, we say that the arc-neighbourhood of 𝑋 is

𝑁 A (𝑋) = {𝐵 ∈ A : (∃𝑏 ∈ 𝐵)(∃𝑥 ∈ 𝑋){𝑏, 𝑥} ∈ 𝐸 (𝐺)} .

We sometimes write 𝑁 A (𝑣) meaning 𝑁 A ({𝑣}) for simplicity. We denote by 𝑑 A (𝑋)

the size of the arc-neighbourhood, that is 𝑑 A (𝑋) = |𝑁 A (𝑋)|.

Definition 2.3.7. Using the constants from Setup 2.3.3 and given 𝑝 > 1 integer,
we say an arc 𝐴 is good in the arc-system A if to at least half of the vertices 𝑣 ∈ 𝐴
we can assign a set of
4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2

arcs in 𝑁 A (𝑣), each arc of A being assigned to at most one 𝑣 ∈ A. If an arc is not
good, we call it bad.

Also, we say that a subset 𝑋 of an arc 𝐴 ∈ A is expanding in A if (∀𝑌 ⊆

𝑋)𝑑 A (𝑌 ) ≥ |𝑌 |4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 .
The definition of good and expanding depends on 𝑝, but it will always be clear
from the context which 𝑝 is meant.
We will use the following simple proposition in the proof of Lemma 2.3.9.

Proposition 2.3.8. Given an arc-system A of a graph 𝐺, the function 𝑑 A is

submodular.

Proof. We trivially have

|𝑁 A ( 𝐴)| + |𝑁 A (𝐵)| = |𝑁 A ( 𝐴) ∪ 𝑁 A (𝐵)| + |𝑁 A ( 𝐴) ∩ 𝑁 A (𝐵)|

24
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

and |𝑁 A ( 𝐴 ∩ 𝐵)| ≤ |𝑁 A ( 𝐴) ∩ 𝑁 A (𝐵)| as the former is a subset of the latter.

Furthermore, 𝑁 A ( 𝐴) ∪ 𝑁 A (𝐵) = 𝑁 A ( 𝐴 ∪ 𝐵). Putting these together we get

𝑑 A ( 𝐴) + 𝑑 A (𝐵) ≥ 𝑑 A ( 𝐴 ∪ 𝐵) + 𝑑 A ( 𝐴 ∩ 𝐵) .

Lemma 2.3.9. Let the constants 𝑎𝑖 and 𝑏𝑖 be as defined in Setup 2.3.3. For each
integer 𝑝 > 1, given a simple arc-system A of size 𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 and length 𝑎 𝑝 𝑥,
either there is an independent set of size 𝑥 𝑝 + 1 in 𝐺 [A] or there is a non-empty
A 0 ⊆ A such that for all 𝐴 ∈ A 0, 𝐴 is good in A 0.

Proof. We observe that an arc 𝐴 is good if and only if it has an expanding subset
of size at least | 𝐴|/2, by Hall’s theorem.
We define a process, by the end of which we either have the required independent
set or the good subset. Let A0 = A be the given arc system of size 𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2
and length 𝑎 𝑝 𝑥. Let B0 be the empty system. In step 𝑡 we will define arc-systems
A𝑡 and B𝑡 as follows.
If each 𝐴 ∈ A𝑡−1 is good in A𝑡−1 , then we define A 0 = A𝑡−1 , B = B𝑡−1 and the
process terminates.
Otherwise we take a bad arc 𝐴 from A𝑡−1 and a maximal expanding set 𝑋 in
it. Note that |𝑋 | < | 𝐴|/2 = 𝑎 𝑝 𝑥/2 in this case. We say that 𝑌 ⊆ 𝑋 is tight if
𝑑 A𝑡−1 (𝑌 ) < (|𝑌 | + 1)4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 . Let 𝐵 denote 𝐴 \ 𝑋. Now for every 𝑣 ∈ 𝐵
there is a tight set 𝑇𝑣 such that 𝑑 A𝑡−1 (𝑇𝑣 ∪ {𝑣}) < (|𝑇𝑣 | + 1)4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 (by
the maximality of 𝑋). Let 𝑇 denote ∪𝑣∈𝐵𝑇𝑣 .
First we claim that if a set 𝑇𝑖 is the union of 𝑖 tight sets, then

𝑑 A𝑡−1 (𝑇𝑖 ) ≤ (|𝑇𝑖 | + 𝑖)4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 (2.3.1)

holds, which we will prove by induction. We can assume 𝑇𝑖 = 𝑇𝑖−1 ∪ 𝑇 0 where 𝑇𝑖−1
is the union of 𝑖 − 1 tight sets and 𝑇 0 is tight. Then

𝑑 A𝑡−1 (𝑇𝑖 ) ≤ 𝑑 A𝑡−1 (𝑇𝑖−1 ) + 𝑑 A𝑡−1 (𝑇 0) − 𝑑 A𝑡−1 (𝑇𝑖−1 ∩ 𝑇 0)

≤ (|𝑇𝑖−1 | + 𝑖 − 1) + (|𝑇 0 | + 1) − |𝑇𝑖−1 ∩ 𝑇 0 | 4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 ,

(2.3.2)

where the first inequality is an application of Proposition 2.3.8. The second

inequality follows by induction on 𝑇𝑖−1 , tightness of 𝑇 0 and expansion of 𝑇 0 ∩ 𝑇𝑖−1 .

25
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

This proves inequality (2.3.1).

Next, we claim that

𝑑 A𝑡−1 (𝐵) ≤ 2𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 . (2.3.3)

First we see that for each 𝑣 ∈ 𝐵 we have

|𝑁 A𝑡−1 (𝑣) \ 𝑁 A𝑡−1 (𝑇𝑣 )| ≤ 4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 ,

since 𝑇𝑣 is expanding, as it is a subset of 𝑋, and 𝑇𝑣 ∪ {𝑣} is not expanding. This

implies
|𝑁 A𝑡−1 (𝐵) \ 𝑁 A𝑡−1 (𝑇)| ≤ 𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 . (2.3.4)

By its definition, 𝑇 is the union of tight sets, therefore it is also the union of at most
|𝑇 | tight sets. As 𝑇 is a subset of 𝑋, |𝑇 | ≤ 𝑎 𝑝 𝑥/2. Thus, using inequality (2.3.1)
on 𝑇 gives us
𝑑 A𝑡−1 (𝑇) ≤ 𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 . (2.3.5)

(2.3.4) and (2.3.5) together imply (2.3.3). So in this step we define B𝑡 = B𝑡−1 ∪ {𝐵}
and A𝑡 = A𝑡−1 \ { 𝐴}. Note that the total size of A𝑡 and B𝑡 is 𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 .
If by the end of this process we have a non-empty good arc-system A 0, then we
have found what we are looking for.
If A 0 is empty, then we have an arc-system B with length at least 𝑎 𝑝 𝑥/2 and
size 𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 . We count the arc-edges of 𝐺 B in the following way. We assign
each arc-edge to the arc that was added to B in the earlier step. By the Equa-
tion (2.3.3) property of 𝐵 proven in the process, each arc will be assigned at most
2𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 arc-edges this way. Thus 𝐺 B has at most

𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 2𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2

arc-edges. This means 𝐺 B has an edge density of at most

𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 2𝑎 𝑝 𝑥4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 𝑑

≤
𝑏 𝑝 𝑥 𝑝 ( 𝑝−1)/2
𝑏 𝑝 𝑥 𝑝−2
2

where 𝑑 is a constant not depending on 𝑥 or 𝑏 𝑝 .

We take a subset C of B of size 𝑥 𝑝−1 with the minimal amount of arc-edges. 𝐺 C

26
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

𝑐1 𝑐1
𝑐2 𝑐2
𝑏2 𝑏2
𝑏1 𝑏1

𝑎2 𝑎1 𝑎2 𝑎1

(a) Type 1 (b) Type 2

Figure 2.2: Semi-triangles

will have at most the same edge density as 𝐺 B . Therefore, 𝐺 C has at most
 
𝑥 𝑝−1 𝑑 𝑑𝑥 𝑝
≤
2 𝑏 𝑝 𝑥 𝑝−2 𝑏𝑝

𝑎𝑝 𝑝 𝑑𝑥 𝑝
edges. Thus, by Lemma 2.3.1, 𝐺 [C] has an independent set of size 2 𝑥 − 𝑏𝑝 ≥
𝑥 𝑝 + 1 as 𝑎 𝑝 ≥ 6 and 𝑏 𝑝 ≥ 𝑑.

Definition 2.3.10. Let us fix a direction on the Hamilton cycle 𝐻 so we can order
the vertices of an arc 𝑢 < 𝑣 if 𝑢 is before 𝑣 in the given direction. We say that three
arcs ( 𝐴, 𝐵, 𝐶) form a semi-triangle if they are in the given order and there exists
𝑎 1 < 𝑎 2 ∈ 𝐴, 𝑏 1 < 𝑏 2 ∈ 𝐵, 𝑐 1 < 𝑐 2 ∈ 𝐶 such that one of the following conditions
holds (see Figure 2.2):
• Type 1: {𝑎 1 , 𝑐 1 }, {𝑎 2 , 𝑏 1 }, {𝑏 2 , 𝑐 2 } ∈ 𝐸 (𝐺) and 𝐴 and 𝐵 are not consecutive
arcs,
• Type 2: {𝑎 1 , 𝑏 1 }, {𝑎 2 , 𝑐 1 }, {𝑏 2 , 𝑐 2 } ∈ 𝐸 (𝐺).

Note that a Type 2 semi-triangle gives us a contradicting cycle (see Figure 2.3a),
therefore it can not exist. Using the good arc-system given by Lemma 2.3.9,
we will show the existence of certain Type 1 semi-triangles. Later, using those
semi-triangles, we find a contradicting cycle.
Given an arc 𝐴 we get the main part of it by taking the second half of it in the given
order. That is, 𝑀 ( 𝐴) ⊆ 𝐴, |𝑀 ( 𝐴)| = | 𝐴|/2 and for all 𝑣 ∈ 𝑀 ( 𝐴), 𝑢 ∈ 𝐴 \ 𝑀 ( 𝐴)
we have 𝑣 > 𝑢. We define leftover part as 𝐿 ( 𝐴) = 𝐴 \ 𝑀 ( 𝐴). For an arc-system
A, 𝑀 (A) and 𝐿 (A) are the sets of the main and leftover parts, respectively, of
each arc in the system.

Lemma 2.3.11. Let the constants 𝑎𝑖 and 𝑏𝑖 be as defined in Setup 2.3.3. For each
𝑝 > 1, if 𝐴𝐼 ( 𝑝 − 1) holds, then the following is true.

27
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

Fix a simple arc-system A of length 𝑎 𝑝 𝑥 and an arc 𝐴 ∈ A. Assume that there

is an assignment of arcs to vertices with the following properties:
• The range of the assignment is 𝐴0, a subset of 𝐴 of size at least | 𝐴|/6,
• To each 𝑣 ∈ 𝐴0 at least 𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 +1 arcs are assigned from 𝑁 𝑀 (A) (𝑣),
• None of the arcs are assigned to more than one vertex.
Then either there are arcs 𝐵, 𝐶 ∈ A such that ( 𝐴, 𝐵, 𝐶) is a semi-triangle of Type
1, or there is an independent set in 𝐺 [A] of size 𝑥 𝑝 + 1.

Proof. If any of the assigned neighbouring arcs is consecutively after 𝐴, we un-

assign it. Let us fix 𝑣 ∈ 𝐴0. We look at the corresponding leftover arcs of the
assigned neighbours of 𝑣, which form an arc-system A 𝑣 of size 𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2
and length 𝑎 𝑝 𝑥/2. Using 𝐴𝐼 ( 𝑝 − 1) and by 𝑎 𝑝 ≥ 2𝑎 𝑝−1 there is an independent
set 𝐽𝑣 in 𝐺 [A 𝑣 ] of size 𝑥 𝑝−1 + 1. Taking 𝑥 of these sets, which we can do because
𝑎 𝑝 𝑥/6 > 𝑥, we get either an independent set of size more than 𝑥 𝑝 + 1 or an edge
{𝑏, 𝑐} between 𝐽𝑤 and 𝐽𝑢 . Let 𝐵 and 𝐶 be the arcs containing 𝑏 and 𝑐, respectively.
Without loss of generality, we can assume that 𝐴, 𝐵, 𝐶 are in this order on the
Hamilton cycle. Then ( 𝐴, 𝐵, 𝐶) is a semi-triangle, because of the edge proving
that 𝑀 (𝐵) is a neighbour of 𝑤, the edge proving that 𝑀 (𝐶) is a neighbour of 𝑢
and the edge {𝑏, 𝑐} going between 𝐿(𝐵) and 𝐿(𝐶). Since Type 2 semi-triangles
cannot exist, this must be a Type 1 semi-triangle.

With this, we are ready to prove the main lemma of the section.

Proof of Lemma 2.3.5. We will use induction to prove the lemma. As the base
case, we observe that 𝑔(𝑥 + 1, 1) = 𝑥 + 1 by Remark 2.2.4. Also, 𝑔(2𝑥, 2𝑥) ≥ 𝑥 2 + 1,
since an arc-system of size 2𝑥 can have at most 2𝑥(2𝑥−1) 2 arc-edges, therefore by
2 2𝑥(2𝑥+1)
Lemma 2.3.1 it has an independent set of size 4𝑥 − 2 ≥ 𝑥 2 + 1.
For the induction step we can assume 𝑝 ≥ 3 and that the lemma is true for all
lower values of 𝑝.
Let A 0 be a simple arc system of size 𝑏 𝑝 𝑥 𝑝( 𝑝−1)/2 and length 𝑎 𝑝 𝑥. Applying
Lemma 2.3.9 on 𝑀 (A 0), we obtain a good subsystem M 0, or we find an independent
set of size 𝑥 𝑝 + 1 and we are done. For each arc in 𝑀 ∈ M 0, there is an arc 𝐴 ∈ A 0
such that 𝑀 ( 𝐴) = 𝑀. Let A denote the set of such 𝐴’s. Thus, 𝑀 (A) is good by
definition. Using Lemma 2.3.11, we get either a Type 1 semi-triangle or the desired
independent set. We define the length of a semi-triangle ( 𝐴, 𝐵, 𝐶) as the number
of arcs between 𝐴 and 𝐵 (note that this is at least 1 by definition ). Let ( 𝐴, 𝐵, 𝐶)

28
 2 Low Independence Number and Hamiltonicity Implies Pancyclicity

𝐶
𝐶 𝐵 𝐵
𝐵
𝐷 𝐷
𝐶
𝐴 𝐴
𝐴
(b) Type 1 semi-triangles, (c) Type 1 semi-triangles,
(a) Type 2 semi-triangle
A𝐶 case A 𝐵 case

Figure 2.3: Contradicting cycles implied by semi-triangles

be a semi-triangle of Type 1 in A with the shortest length. Let 𝐷 denote the arc
consecutively after 𝐴 and A 𝐴 those arcs between 𝐴, 𝐵; A 𝐵 those arcs between
𝐵, 𝐶; and A𝐶 those arcs between 𝐶, 𝐴.
Recall that by the definition of good, there is a subset of vertices 𝐷 0 ⊆ 𝐷, with
|𝐷 0 | ≥ |𝐷|, such that we can assign 4𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 neighbouring arcs to each.
For each vertex in 𝐷 0, using the pigeonhole principle, at least 𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 +1 of
the assigned arcs are in either A 𝐴 , A 𝐵 or A𝐶 . Using the pigeonhole principle again,
at least |𝐷|/6 vertices of 𝐷 are assigned at least 𝑏 𝑝−1 𝑥 ( 𝑝−1) ( 𝑝−2)/2 + 1 neighbouring
arcs from either A 𝐴 , A 𝐵 or A𝐶 . If this is A 𝐴 , then by applying Lemma 2.3.11 to 𝐴
and A 𝐴 we find either a shorter Type 1 semi-triangle contradicting the minimality
of ( 𝐴, 𝐵, 𝐶), or the desired independent set. If this is A 𝐵 or A𝐶 , then we get a
new Type 1 semi-triangle.
Considering this new Type 1 semi-triangle together with ( 𝐴, 𝐵, 𝐶), we get one
of the arrangements on Figures 2.3b or 2.3c. We call the vertices incident to the
edges used in the semi-triangles relevant vertices. Note that the order of these
six arcs as well as that of the relevant vertices is determined by the definition of
Type 1 semi-triangles and the definition of the sets A 𝐵 and A𝐶 . We construct a
contradicting cycle the following way: starting from the fixed Hamilton cycle, we
remove the edges between the pairs of relevant vertices in the six arcs and add the
six edges used in the two semi-triangles. Note that we removed at most 6 · 𝑘/6
edges as A is an independent arc-system. This graph is clearly 2-regular, and
connectivity can be checked using the Figures as the order of vertices used in them
is established.
As finding a contradicting cycle is impossible, we must have found the inde-
pendent set we are looking for at one of these steps.

29
Maximum Number of Triangle-free
3
Edge Colourings

3.1 Introduction
Let 𝑘 ≥ 3 and 𝑟 ≥ 2 be natural numbers. By a colouring of a graph 𝐺 = (𝑉, 𝐸)
with 𝑟 colours here we mean edge-colouring, that is a function 𝑓 : 𝐸 → {1, . . . , 𝑟 }.
In this context we refer to the numbers 1, . . . , 𝑟 as colours and by the colour class
of 𝑐 we mean 𝑓 −1 (𝑐). A colouring is 𝐾 𝑘 -free, if no colour class contains a copy of
𝐾 𝑘 . For a graph 𝐺, let 𝐹 (𝐺, 𝑘, 𝑟) denote the number of 𝐾 𝑘 -free colourings of 𝐺
with 𝑟 colours. Let 𝐹 (𝑛, 𝑘, 𝑟) denote the maximum of 𝐹 (𝐺, 𝑘, 𝑟) over all graphs
𝐺 on 𝑛 vertices.
In this chapter we consider the case 𝑘 = 3 and 𝑟 ∈ {5, 6}. Let 𝜑𝑟 (𝐺) = 𝐹 (𝐺, 3, 𝑟)
and 𝜑𝑟 (𝑛) = 𝐹 (𝑛, 3, 𝑟). We find the approximate value of 𝜑5 (𝑛) and 𝜑6 (𝑛). We
also prove stability results for both cases.
In this section we state stronger, technical versions of Theorem 1.2.8 and The-
orem 1.2.5. The reason for this is that these stronger technical theorems will be
useful when we want to find the exact extremal graph for the Erdős-Rothschild
problem. While the search for the exact result is beyond the scope of this thesis,
we state and prove the following technical results.
We first state the technical stability result for 𝑟 = 6 as that is the easier of the
two. For two bit vectors 𝜺 (1) , 𝜺 (2) ∈ {0, 1}𝑛 , with 𝜺 (1) = (𝜀1(1) , . . . , 𝜀 𝑛(1) ), we define
their distance as their Hamming-distance, which is the number of indices where
they differ.
To state our theorems, we need the following notation.

Definition 3.1.1. For a graph 𝐺, a positive integer 𝑛, a bit vector 𝜺 ∈ {0, 1}𝑛 and
 3 Maximum Number of Triangle-free Edge Colourings

vertex partitions 𝑉𝑖0 ∪ 𝑉𝑖1 = 𝑉 (𝐺) for all 𝑖 ∈ {1, . . . , 𝑛}, we define

𝜺
Ù
𝑉 (𝜺) = 𝑉𝑗 𝑗 .
𝑗 ∈[𝑛]

Theorem 3.1.2 (Technical stability for 𝑟 = 6). For every 0 < 𝛿 < 10−30 there is
an 𝑛0 such that for all graphs 𝐺 on 𝑛 > 𝑛0 vertices the following holds. Suppose
2 2 2
𝜑6 (𝐺) ≥ 3𝑛 /4 43𝑛 /16−𝛿𝑛 . Then there are balanced bipartitions 𝑉𝑖0 ∪ 𝑉𝑖1 = 𝑉 (𝐺)
for each 𝑖 ∈ [6] and E = {𝜺 (1) , . . . , 𝜺 (8) } with 𝜺 ( 𝑗) ∈ {0, 1}6 , furthermore, there is
a partition of E into E0 and E1 such that
(i) We have Ø  √ 
𝑉 (𝜺) ≥ 1 − 105 𝛿 𝑛 .
𝜺∈E

(ii) We have Ø Ø √
𝑉 (𝜺) = 𝑉 (𝜺) ± 108 𝛿𝑛 .
𝜺∈E0 𝜺∈E1

(iii) For any pair (𝜺, 𝜺0) ∈ E 2 we have

 
𝑒 𝑉 (𝜺), 𝑉 (𝜺0) ≥ 𝑉 (𝜺) · 𝑉 (𝜺0) − 107 𝛿𝑛2 .

(iv) For any 𝜺 ∈ E we have

 
𝑒 𝑉 (𝜺) ≤ 107 𝛿𝑛2 .

(v) Each pair (𝜺, 𝜺0) ∈ E0 × E1 has distance three and each pair (𝜺, 𝜺0) ∈
E0 2 ∪ E1 2 has distance four.
(vi) Moreover, for any pair (𝜺, 𝜺0) ∈ E 2 we have
√4
𝑉 (𝜺) = 𝑉 (𝜺0) ± 104 𝛿𝑛 .

Next we state the technical stability result for the more complex 𝑟 = 5 case.

Theorem 3.1.3 (Technical stability for 𝑟 = 5). For every 0 < 𝛿 < 10−30 there is
an 𝑛0 such that for all graphs 𝐺 on 𝑛 > 𝑛0 vertices the following holds. Suppose
2 2
𝜑5 (𝐺) ≥ 6𝑛 /4−𝛿𝑛 /4 . Then there are balanced bipartitions 𝑉𝑖0 ∪ 𝑉𝑖1 = 𝑉 (𝐺) for
each 𝑖 ∈ [5], an integer 𝑡 ∈ {4, 6, 8} and E = {𝜺 (1) , . . . , 𝜺 (𝑡) } with 𝜺 ( 𝑗) ∈ {0, 1}5 ,
furthermore, there is an 𝑠 ∈ [5] and a partition of E into E0 and E1 such that

31
 3 Maximum Number of Triangle-free Edge Colourings

(i) We have Ø  √4 
𝑉 (𝜺) ≥ 1 − 105 𝛿 𝑛 .
𝜺∈E

(ii) We have Ø Ø √
𝑉 (𝜺) = 𝑉 (𝜺) ± 108 𝛿𝑛 .
𝜺∈E0 𝜺∈E1

(iii) For any pair (𝜺, 𝜺0) ∈ E 2 we have

 
𝑒 𝑉 (𝜺), 𝑉 (𝜺0) ≥ 𝑉 (𝜺) · 𝑉 (𝜺0) − 107 𝛿𝑛2 .

(iv) For any 𝜺 ∈ E we have

 
𝑒 𝑉 (𝜺) ≤ 107 𝛿𝑛2 .

(v) Each pair (𝜺, 𝜺0) ∈ E0 × E1 has distance three and each pair (𝜺, 𝜺0) ∈
E02 ∪ E12 has distance two or four.
(vi) Moreover, pairs of distance four form a perfect matching in E and for each
of these matched pairs (𝜺, 𝜺0) we have
√4
𝑉 (𝜺) = 𝑉 (𝜺0) ± 104 𝛿𝑛 .

Following the implied partitions, it is easy to prove the main theorems using
these technical results. The goal of the following sections is to prove these two
theorems.
In Section 3.2 we introduce basic notation that we’ll use throughout the chapter
and cite a container theorem by Balogh et al.[6] which will be a key tool in
proving our results. In Section 3.4 we prove a technical theorem, Theorem 3.4.3,
which provides an upper bound on the number of colourings and some structural
properties of any graph and corresponding containers which comes close to this
upper bound. This result is based on the structure of edges appearing in exactly
3 containers, which is the main novelty in this chapter. In Section 3.5 we prove
matching lower bounds by providing a family of constructions and colourings on
them. In Section 3.6 we build on Theorem 3.4.3 to prove Theorem 3.1.2 and
Theorem 3.1.3.

32
 3 Maximum Number of Triangle-free Edge Colourings

3.2 Preliminaries
We denote by 𝜑𝑟 (𝐺) the number of triangle-free 𝑟-colourings of 𝐸 (𝐺), and by
𝜑𝑟 (𝑛) := max{𝜑𝑟 (𝐺) : 𝐺 is a graph with 𝑛 vertices}. We denote by 𝑡𝑟 (𝐺) the
number of triangles in 𝐺.
Suppose we are given a graph 𝐺 and subgraphs 𝐶1 , . . . , 𝐶𝑟 ⊆ 𝐺 with 𝐸 (𝐺) =
𝐸 (𝐶1 ) ∪ . . . ∪ 𝐸 (𝐶𝑟 ). We denote by Φ(𝐶1 , . . . , 𝐶𝑟 ) the family of all colourings
𝜒 in which 𝜒−1 (𝑖) ⊆ 𝐸 (𝐶𝑖 ) for all 𝑖 ∈ [𝑟]. If each 𝐶𝑖 is triangle-free, then
every such colouring is triangle-free, hence we have 𝜑𝑟 (𝐺) ≥ |Φ(𝐶1 , . . . , 𝐶𝑟 )|
in this case. Furthermore, the number of such colourings is easy to count. For
𝑖 ∈ [𝑟], let 𝑀𝑖 = 𝑀𝑖 (𝐶1 , . . . , 𝐶𝑟 ) ⊆ 𝐺 denote the graph on 𝑉 (𝐺) whose edges
are those of 𝐺 contained in exactly 𝑖 of the subgraphs 𝐶1 , . . . , 𝐶𝑟 . We denote by
𝑚𝑖 = 𝑚𝑖 (𝐶1 , . . . , 𝐶𝑟 ) the number of edges in 𝑀𝑖 . We then have
Ö Õ Õ
|Φ(𝐶1 , . . . , 𝐶𝑟 )| = 𝑖 𝑚𝑖 and 𝑖 · 𝑚𝑖 = |𝐶𝑖 |, (3.2.1)
𝑖∈[𝑟] 𝑖∈[𝑟] 𝑖∈[𝑟]

We will make use of the container theorem below proved by Mousset, Nenadov
and Steger [44] using the hypergraph container method of Balogh, Morris and
Samotij [7] and Saxton and Thomason [53]. We use an equivalent formulation of
their result as stated by Balogh et al. in [6]. This approach was introduced by Hàn
and Jiménez [29] using ideas of Clemens, Das and Tran [13].

Theorem 3.2.1 ([6, Theorem 3.2]). There exists constant 𝐶 such that for every
graph 𝐺 on 𝑛 > 𝐶 vertices there exists a collection C = C(𝐺) of subgraphs of 𝐺
such that the following holds:
(a) every triangle-free subgraph 𝐺 0 ⊆ 𝐺 is a subgraph of some 𝐶 ∈ C,
(b) 𝑡𝑟 (𝐶) ≤ 𝑛25/9 for every 𝐶 ∈ C,
(c) |C| ≤ exp 𝑛16/9 .

Theorem 3.2 from [6] is stated for 𝐺 = 𝐾𝑛 only, however, by taking the intersec-
tion of each container 𝐶 with 𝐺 we obviously obtain the family C = C(𝐺) from
above.
For a graph 𝐺 on 𝑛 vertices, let us denote by C(𝐺) the family of all graphs 𝐶 ⊆ 𝐺
on 𝑛 with 𝑡𝑟 (𝐶) ≤ 𝑛25/9 . An immediate corollary of Theorem 3.2.1 provides the
following.

33
 3 Maximum Number of Triangle-free Edge Colourings

Corollary 3.2.2. There is a constant 𝑛0 , such that for 𝑛 > 𝑛0 and all graphs 𝐺 on
𝑛 vertices and every 𝑟 ∈ N, we have
 
𝜑𝑟 (𝐺) ≤ exp 𝑟 · 𝑛16/9 · max |Φ(𝐶1 , . . . , 𝐶𝑟 )| .
𝐶1 ,...,𝐶𝑟 ∈C(𝐺)

3.3 An approximate upper bound for many colours

In this section, we showcase the strength of the container method on the Erdős-
Rothschild problem. The following theorem and proof is by Hàn and Jiménez. Note
that all previous theorems of this type made use of Szemerédi’s regularity lemma
and were much more technical. In Section 3.4 we will strengthen this theorem for
5 and 6 colours.

(1+𝑜(1)) ( 𝑛2)
Theorem 3.3.1. We have 𝜑𝑟 (𝑛) ≤ 2𝑟 .
Theorem 3.3.1 basically follows from the following claim and Corollary 3.2.2.
Claim 3.3.2. Let 𝜀 > 0, let 𝑟 ≥ 6 be an integer and let 𝐶1 , . . . , 𝐶𝑟 be graphs on


[𝑛] with 𝑒(𝐶𝑖 ) ≤ (1/2 + 𝛿) 𝑛2 . Then we have

   ( 𝑛2)
1
|Φ(𝐶1 , . . . , 𝐶𝑟 )| ≤ +𝛿 𝑟 .
2

𝑒(𝐶𝑖 )/𝑚 ≤ 𝑟 ( 12 + 𝛿) 𝑛

Í
Proof. Let 𝑚 := 𝑒(𝐺) and let 𝑖 0 = 𝑖∈[𝑟] 2 /𝑚. Note that, by
convexity, we have
Ö
|Φ(𝐶1 , . . . , 𝐶𝑟 )| ≤ 𝑖 𝑚𝑖 ≤ 𝑖0𝑚1 +...+𝑚𝑟
𝑖∈[𝑟]
   2 𝑚
1 𝑛
≤ 𝑟 +𝛿 .
2 2𝑚
2
Since 𝑟 ≥ 6 > 2𝑒, it follows from elementary calculus that (𝑟 ( 12 + 𝛿) 2𝑚
𝑛 𝑚
) is
𝑛

monotone increasing in 𝑚. The claim now follows since 𝑚 ≤ 2 .

To finish the proof of Theorem 3.3.1, we would like to bound the number of edges
based on the number of triangles. The problem of finding the best such bound is
known as the Erdős-Rademacher problem, first asked by Erdős [21] in 1955. The
problem was resolved Liu, Pikhurko and Staden [41] in 2017. Their answer is quite
complex, therefore we will make use of the following simpler theorem of Bollobás
[8] instead.

34
 3 Maximum Number of Triangle-free Edge Colourings

Theorem 3.3.3 (Bollobás, [8]). A graph with 𝑛 vertices and 𝑚 edges has at least
𝑛 2
9 (4𝑚 − 𝑛 ) triangles.

Proof of Theorem 3.3.1. Let 𝐺 be an 𝑛 vertex graph and let C(𝐺) be the collection
provided by Theorem 3.2.1. Let 𝐶1 , . . . , 𝐶𝑟 ∈ C(𝐺) with

|Φ(𝐶1 , . . . , 𝐶𝑟 )| = max |Φ(𝐶1 , . . . , 𝐶𝑟 )| .

𝐶1 ,...,𝐶𝑟 ∈C(𝐺)

Using Theorem 3.3.3, we deduce that 𝑒(𝐶𝑖 ) ≤ 𝑛2 /4 + 9/4 · 𝑛16/9 = (1/2 + 𝑜(1)) 𝑛

2
for every 𝑖 ∈ [𝑟]. Hence, using Corollary 3.2.2 and Claim 3.3.2, we deduce
 
16/9
𝜑𝑟 (𝐺) ≤ exp 𝑟 · 𝑛 · |Φ(𝐶1 , . . . , 𝐶𝑟 )|
 𝑟  (1+𝑜(1)) ( 𝑛2)
≤ .
2

3.4 An approximate upper bound for five and six

colours
In this section, we improve the upper bound from Theorem 3.3.1 for five and six
colours. These improved upper bounds are asymptotically best possible in 𝑛.

Theorem 3.4.1. For every graph 𝐺 on 𝑛 vertices, the number of 5-colourings of

𝐸 (𝐺) without monochromatic triangles is at most

2 /4+𝑜(𝑛2 )
6𝑛 .

Theorem 3.4.2. For every graph 𝐺 on 𝑛 vertices, the number of 6-colourings of

𝐸 (𝐺) without monochromatic triangles is at most

2 /4 2 /16+𝑜(𝑛2 )
3𝑛 43𝑛 .

In fact we shall prove a more general result, which will be useful for the stability
results we will discuss later.

Theorem 3.4.3. There exists 𝑐 > 0 such that for every 10−20 > 𝛿 > 0 there is an 𝑛0
such that for all 𝑛 > 𝑛0 and all graph 𝐺 on 𝑛 vertices the following holds. Suppose

35
 3 Maximum Number of Triangle-free Edge Colourings

2 /4−𝛿𝑛2
(i) 𝐶1 , . . . , 𝐶5 ⊆ 𝐺, are such that |Φ(𝐶1 , . . . , 𝐶5 )| ≥ 6𝑛 or
2 /4 2 /16−𝛿𝑛2
(ii) 𝐶1 , . . . , 𝐶6 ⊆ 𝐺, are such that |Φ(𝐶1 , . . . , 𝐶6 )| ≥ 3𝑛 43𝑛 .
and in either case 𝑡𝑟 (𝐶𝑖 ) ≤ 𝑛25/9 for all 𝐶𝑖 .
Then all the 𝐶𝑖 ’s and 𝑀3 (𝐶1 , . . . , 𝐶𝑟 ) are 𝑐𝛿𝑛2 -close to being balanced complete
bipartite. Furthermore, 𝑚 1 , 𝑚 5 , 𝑚 6 ≤ 𝑐𝛿𝑛2 . In case (i) we have 𝑚 2 + 2𝑚 4 =
𝑛2 /4 ± 𝑐𝛿𝑛2 . In case (ii) we have 𝑚 2 + 2𝑚 4 = 3𝑛2 /8 ± 𝑐𝛿𝑛2 . In particular, we have
2 2 2 2 2
|Φ(𝐶1 , . . . , 𝐶5 )| ≤ 6𝑛 /4+𝑐𝛿𝑛 and |Φ(𝐶1 , . . . , 𝐶6 )| ≤ 3𝑛 /4 43𝑛 /16+𝑐𝛿𝑛 .

Remark 3.4.4. Our proof provides the constant 𝑐 = 106 and we make no effort to
find the lowest possible 𝑐. We don’t believe the dependency on 𝛿 can be improved,
apart from the constant 𝑐. If you start with our extremal examples and remove 𝛿𝑛2
arbitrary edges, the conditions hold but you cannot hope for anything better.

Theorem 3.4.1 and Theorem 3.4.2 easily follow from Theorem 3.4.3 and Corol-
lary 3.2.2.
We will use the following theorem of Füredi [25].
𝑛2
Theorem 3.4.5 (Füredi, [25]). Every triangle-free graph with at least 4 − 𝑡 edges
2
has a bipartite subgraph with at least 𝑛4 − 2𝑡 edges.

3.4.1 Upper bound proof for five colours

We will now prove Theorem 3.4.3 for 𝑟 = 5. By the triangle removal lemma [52],
there is some 𝛿0 > 0, so that every graph with at most 𝛿0𝑛3 triangles can be made
triangle-free by removing 𝛿𝑛2 /5 edges. Hence, for every 𝑖 ∈ [5] and every large
enough 𝑛, there are triangle-free subgraphs 𝐶˜𝑖 ⊆ 𝐶𝑖 with 𝑒(𝐶˜𝑖 ) ≥ 𝑒(𝐶𝑖 ) − 𝛿𝑛2 /5.
It is easy to see now that

2 2 2
Φ(𝐶˜1 , . . . , 𝐶˜5 ) ≥ 5−𝛿𝑛 |Φ(𝐶1 , . . . , 𝐶5 )| ≥ 6𝑛 /4−2𝛿𝑛 . (3.4.1)

Define 𝛼1 , . . . , 𝛼5 ∈ R so that 𝑒(𝐶˜𝑖 ) = (1/2 − 𝛼𝑖 ) 𝑛2 and note that −1/(2𝑛 − 2) ≤

𝛼𝑖 ≤ 1/2 for every 𝑖 ∈ [5]. Furthermore, let 𝛼 = (𝛼1 + . . . + 𝛼5 )/5 and define


𝜇3 ∈ [0, 1] by 𝑚 3 = 𝜇3 𝑛2 . The theorem follows from the following three claims.

Claim 3.4.6. We have

   
3 75 75 1 1
𝜇3 ≤ + 𝛼 𝛼+ +𝑂 .
4 2 8 2 𝑛

36
 3 Maximum Number of Triangle-free Edge Colourings

Claim 3.4.7. We have

 
1 3 log(4)
𝛼≤ · (4 log(3) − 3 log(4))𝜇3 + − 2 log(3) + 16 log(6) · 𝛿 .
5 log(4) 2

Claim 3.4.8. We have 𝛼 ≤ 0.015.

Before proving the claims, we show how they imply the theorem.

Proof of Theorem 3.4.3. Putting together Claim 3.4.6 and Claim 3.4.7 we get
     
1 3 75 75 1
𝛼≤ · (4 log(3) − 3 log(4)) + 𝛼 𝛼 + 16 log(6) · 𝛿 + 𝑂 .
5 log(4) 4 2 8 𝑛
 
Since 𝑛 is large enough we can assume that 𝑂 𝑛1 ≤ 𝛿 and thus we can rearrange
the quadratic inequality to
0 ≤ 𝛼2 − 𝑏𝛼 + 𝑐𝛿,
5 log(4) 32 16 log(6) 32
where 𝑏 = −4 + 4 log(3)−3 log(4) 225 ≈ 0.18485 and 𝑐 = 1 + 4 log(3)−3 log(4) 225 ≈
18.3083.
By solving for 𝛼, we get that either
√ √
𝑏 − 𝑏 2 − 4𝑐𝛿 𝑏 + 𝑏 2 − 4𝑐𝛿
𝛼≤ or 𝛼≥ > 0.1.
2 2

By Claim 3.4.8, the second case is impossible. A simple calculation shows that
√
𝑏 − 3𝑐𝛿
𝑏 < 𝑏 2 − 4𝑐𝛿 and thus we have 𝛼 ≤ 3𝑐𝛿 2𝑏 < 150𝛿, which implies that
𝛼𝑖 < 750𝛿 for every 𝑖 ∈ [5]. Therefore, 𝐶𝑖 is 10000𝛿𝑛2 close to being balanced
complete bipartite (using Theorem 3.4.5, we get a large bipartite subgraph, which
we can balance by moving a few vertices and make it complete by adding the
missing edges). Using Claim 3.4.6 we also get that 𝜇3 < 1/2 + 100000𝛿 and
therefore 𝑀3 (𝐶1 , . . . , 𝐶5 ) is 1000000𝛿 close to being balanced complete bipartite.
By (3.4.1) and (3.2.1), we have

5
Ö
𝑛2 /4−2𝛿𝑛2
6 ≤ Φ(𝐶˜1 , . . . , 𝐶˜5 ) ≤ 𝑖 𝑚𝑖
𝑖=1

and 𝑖∈[5] 𝑖 · 𝑚𝑖 = 𝑖∈[5] 𝑒(𝐶˜𝑖 ). Using our bounds on 𝜇3 we know that 𝑚 3 ≤

Í Í

𝑛2 /4 + 400𝛿. Simple linear optimisation on 𝑚𝑖 shows that the only solutions

to these constraints are in the form 𝑚 1 , 𝑚 5 , 𝑚 6 ≤ 10000𝛿𝑛2 and 𝑚 2 + 2𝑚 4 =
𝑛2 2
4 ± 10000𝛿𝑛 .

37
 3 Maximum Number of Triangle-free Edge Colourings

It remains to prove Claims 3.4.6, 3.4.7 and 3.4.8.

Proof of Claim 3.4.6. By applying Theorem 3.4.5 we get bipartite graphs 𝐵𝑖 ⊆ 𝐶˜𝑖


such that 𝑒(𝐵𝑖 ) = (1/2 − 2𝛼𝑖 ) 𝑛2 − 𝑂 (𝑛) for every 𝑖 ∈ [5]. We will denote
by 𝑃1 (𝐵𝑖 ) and 𝑃2 (𝐵𝑖 ) the smaller and larger part of vertices respectively in the
bipartition (breaking ties arbitrarily). Let 𝐺3 = 𝑀3 (𝐶˜1 , 𝐶˜2 , 𝐶˜3 , 𝐶˜4 , 𝐶˜5 ) and define
𝑥𝑖 ∈ [0, 1/2] such that 𝑣(𝑃1 (𝐵𝑖 )) = 21 − 𝑥𝑖 𝑛.
We say an edge 𝑒 is missing, if there is an 𝑖 ∈ [5] such that 𝑒 ∉ 𝐶˜𝑖 and 𝑒 shares
a vertex with both parts of 𝐵𝑖 . We say an edge 𝑒 is extra if there is an 𝑖 ∈ [5] such
that 𝑒 ∈ 𝐶˜𝑖 and 𝑒 is contained in one part of 𝐵𝑖 .
Í5
For a triangle 𝑇 we define 𝑒𝑖 (𝑇) = |𝐸 (𝑇) ∩ 𝐸 (𝐶˜𝑖 )| and 𝑒 𝑠 (𝑇) = 𝑖=1 𝑒𝑖 (𝑇). Note
that, if a triangle 𝑇 has no missing or extra edges, then 𝑒𝑖 (𝑇) ∈ {0, 2}. Furthermore,
if a triangle 𝑇 is contained in 𝐺 3 , then 𝑒 𝑠 (𝑇) = 9. This implies that each triangle
in 𝐺 3 has a missing or an extra edge. We assign to each triangle of 𝐺 3 a missing
or an extra edge such that, if it has both, then we choose a missing one.
The number of missing edges in 𝐶˜𝑖 is at most 2(𝛼𝑖 − 𝑥𝑖2 ) 𝑛2 + 𝑂 (𝑛) and each of

them is assigned to at most 𝑛 triangles. The number of extra edges ˜

  in 𝐶𝑖 is at most
𝛼𝑖 2 + 𝑂 (𝑛) and each of them is assigned to at most 21 + 𝑥 1 𝑛 triangles as the
𝑛

third vertex must be on the same part of the bipartition (otherwise it would have a
missing edge as well, as 𝐶˜𝑖 is triangle-free).
Let 𝑡𝑟 (𝐻) denote the number of triangles in a graph 𝐻. From the above discus-
sion, we get

5       
Õ 𝑛 𝑛 1
𝑡𝑟 (𝐺 3 ) ≤ 2(𝛼𝑖 − 𝑥𝑖2 )
𝑛 + 𝛼𝑖 − 𝑥𝑖 𝑛 + 𝑂 (𝑛2 )
𝑖=1
2 2 2
5
!
𝛼𝑖  2 3𝛼𝑖2
 Õ
𝑛 15 
= 𝛼𝑖 − 6 𝑥𝑖 − + + 𝑂 (𝑛2 )
3 𝑖=1 2 4 8
5
!
3𝛼𝑖2
 Õ
𝑛 15
≤ 𝛼𝑖 + + 𝑂 (𝑛2 ).
3 𝑖=1 2 8

Í5
Since 𝛼𝑖 ≤ 5𝛼 and 𝑖=1 𝛼𝑖 = 5𝛼, we have
   
75 75 𝑛
𝑡𝑟 (𝐺 3 ) ≤ + 𝛼 𝛼 + 𝑂 (𝑛2 ) . (3.4.2)
2 8 3

38
 3 Maximum Number of Triangle-free Edge Colourings

We now apply Theorem 3.3.3 in the following equivalent form,

  
4 1 𝑛
𝑡𝑟 (𝐺 3 ) ≥ 𝜇3 − . (3.4.3)
3 2 3

Putting (3.4.2) and (3.4.3) together and organising the inequality, we get
   
3 75 75 1 1
𝜇3 ≤ + 𝛼 𝛼+ +𝑂 .
4 2 8 2 𝑛

Proof of Claim 3.4.7. By (3.4.1) and (3.2.1), we have

5
Ö
2 /4−2𝛿𝑛2
6𝑛 ≤ Φ(𝐶˜1 , . . . , 𝐶˜5 ) ≤ 𝑖 𝑚𝑖
𝑖=1

Î5 𝑚 𝑖
and 𝑖∈[5] 𝑖 · 𝑚𝑖 = 𝑖∈[5] 𝑒(𝐶˜𝑖 ). A simple optimisation shows that 𝑖=1
Í Í
𝑖 is
maximised under this constraint for fixed 𝑚 3 when 𝑚 1 = 𝑚 5 = 0 and 2𝑚 2 + 4𝑚 4 is
as large as possible. We conclude that
Í5
𝑒(𝐶˜𝑖 )−3𝑚 3 )/4
2
Ö
6 (1/4−2𝛿)𝑛 ≤ 𝑖 𝑚 𝑖 ≤ 3𝑚 3 · 4 ( 𝑖=1 (3.4.4)
𝑖∈[5]

By taking logarithms we obtain

Í5
𝑒(𝐶˜𝑖 ) − 3𝑚 3
  2
1 𝑛
log(6) · − 4𝛿 · ≤ log(3) · 𝑚 3 + log(4) · 𝑖=1
2 2 4
!
5
− 5𝛼 − 3𝜇 3 𝑛2
≤ log(3) · 𝜇3 + log(4) · 2 · .
4 2

𝑛2
By cancelling 2 and rearranging we obtain
 
5 3 5 log(4) log(6)
· log(4) · 𝛼 ≤ log(3) + · log(4) 𝜇3 + − + 4𝛿 log(6).
4 4 8 2

5 3
The claim now follows, since 2 · log(4) − 2 log(6) = 2 log(4) − 2 log(3). 

Proof of Claim 3.4.8. Note that (3.4.4) is maximised under the condition
Õ Õ
𝑚𝑖 = 𝑒(𝐶˜𝑖 )
𝑖∈[5] 𝑖∈[5]

39
 3 Maximum Number of Triangle-free Edge Colourings

if 𝑚 3 is as large as possible, i.e. when

Õ
3𝑚 3 = 𝑒(𝐶˜𝑖 )
𝑖∈[5]

or in other words when 𝜇3 = 5(1/2 − 𝛼)/3. This implies

2 2 /6
6 (1/4−2𝛿)𝑛 ≤ 35(1/2−𝛼)/3· ( 2) ≤ 35(1/2−𝛼)𝑛
𝑛
.

Taking logarithms and solving for 𝛼, we get

1 3 log(6) 12 log(6)
𝛼≤ − + · 𝛿 ≤ 0.01073 + 3.915𝛿 ≤ 0.015.
2 10 log(3) 5 log(3)

3.4.2 Upper bound proof for six colours

This proof is very similar to the 𝑟 = 5 case. We include it in this thesis, as these
calculations are essential to see that the proof actually works.
Similarly as in the 5-colour case, for large enough 𝑛, there is a triangle-free
subgraph 𝐶˜𝑖 ⊆ 𝐶𝑖 for every 𝑖 ∈ [6] with

2 /4 2 /16−2𝛿𝑛2
Φ(𝐶˜1 , . . . , 𝐶˜6 ) ≥ 3𝑛 43𝑛 . (3.4.5)

Define 𝛼1 , . . . , 𝛼6 ∈ R so that 𝑒(𝐶˜𝑖 ) = (1/2 − 𝛼𝑖 ) 𝑛2 and note that −1/(2𝑛 − 2) ≤

𝛼𝑖 ≤ 1/2 for every 𝑖 ∈ [6]. Furthermore, let 𝛼 = (𝛼1 + . . . + 𝛼6 )/6 and define


𝜇3 ∈ [0, 1] by 𝑚 3 = 𝜇3 𝑛2 . The theorem follows from the following three claims.

Claim 3.4.9. We have

   
3 27 1 1
𝜇3 ≤ 45 + 𝛼 𝛼 + + 𝑂 .
4 2 2 𝑛

Claim 3.4.10. We have

 
2 log(3) 1 1 log(3) 8
𝛼≤ − 𝜇3 + − + 𝛿.
3 log(4) 2 4 3 log(4) 3

Claim 3.4.11. We have 𝛼 ≤ 0.016.

Before proving the claims, we show how they imply the theorem.

40
 3 Maximum Number of Triangle-free Edge Colourings

Proof of Theorem 3.4.3. Putting together Claim 3.4.9 and Claim 3.4.10 and re-
arranging the quadratic inequality, we get

0 ≤ 𝛼2 − 𝑏𝛼 + 𝑐𝛿

286 log(2)−180 log(3)

where 𝑏 = 54 log(3)−81 log(2) ≈ 0.15404 and 𝑐 = 12 (for large enough 𝑛).
By solving for 𝛼, we get that either
√ √
𝑏 − 𝑏 2 − 4𝑐𝛿 𝑏 + 𝑏 2 − 4𝑐𝛿
𝛼≤ or 𝛼≥ > 0.1.
2 2

By Claim 3.4.11, the second case is impossible. A simple calculation shows that
√
𝑏 − 3𝑐𝛿
𝑏 < 𝑏 2 − 4𝑐𝛿 and thus we have 𝛼 ≤ 3𝑐𝛿 2𝑏 < 125𝛿, which implies that
𝛼𝑖 < 750𝛿 for every 𝑖 ∈ [6]. The rest of the proof is completely analogous to the
proof in the 5-colour case.

It remains to prove Claims 3.4.9, 3.4.10 and 3.4.11.

Proof of Claim 3.4.9. Analogously as in Claim 3.4.6, we derive

   
27 𝑛
𝑡𝑟 (𝐺 3 ) ≤ 45 + 𝛼 𝛼 + 𝑂 (𝑛2 ). (3.4.6)
2 3
The claim again follows from using (3.4.6) and Theorem 3.3.3 in the form (3.4.3).


Proof of Claim 3.4.10. By (3.4.5) and (3.2.1), we have

6
Ö
𝑛2 /4 3𝑛2 /16−2𝛿𝑛2
3 4 ≤ Φ(𝐶˜1 , . . . , 𝐶˜6 ) ≤ 𝑖 𝑚𝑖
𝑖=1

Î6 𝑚𝑖
and 𝑖∈[6] 𝑖 · 𝑚𝑖 = 𝑖∈[6] 𝑒(𝐶˜𝑖 ). A simple optimisation shows that 𝑖=1
Í Í
𝑖 is
maximised under this constraint for fixed 𝑚 3 when 𝑚 1 = 𝑚 5 = 𝑚 6 = 0 and
2𝑚 2 + 4𝑚 4 is as large as possible. We conclude that
Í6
𝑒(𝐶˜𝑖 )−3𝑚 3 )/4
2 /4 2 /16−2𝛿𝑛2
Ö
3𝑛 43𝑛 ≤ 𝑖 𝑚 𝑖 ≤ 3𝑚 3 · 4 ( 𝑖=1 (3.4.7)
𝑖∈[6]

41
 3 Maximum Number of Triangle-free Edge Colourings

By taking logarithms we obtain

Í6
𝑒(𝐶˜𝑖 ) − 3𝑚 3
  
log(3) 3 2
+ log(4) · − 2𝛿 𝑛 ≤ log(3) · 𝑚 3 + log(4) · 𝑖=1
4 16 4
  
𝜇3 3 3𝛼 3𝜇3
≤ log(3) · + log(4) · − − 𝑛2 .
2 8 4 8

By cancelling 𝑛2 and rearranging we obtain

   
3 𝜇3 1 3 3
· log(4) · 𝛼 ≤ log(3) − + log(4) − · 𝜇3 + 2𝛿 ,
4 2 4 16 8

which implies the claim. 

Proof of Claim 3.4.11. Note that (3.4.7) is maximised under the condition
Õ Õ
𝑚𝑖 = 𝑒(𝐶˜𝑖 )
𝑖∈[6] 𝑖∈[6]

if 𝑚 3 is as large as possible, i.e. when

Õ
3𝑚 3 = 𝑒(𝐶˜𝑖 )
𝑖∈[6]

or in other words when 𝜇3 = 2(1/2 − 𝛼). This implies

2 /4 2 /16−2𝛿𝑛2 2
≤ 32(1/2−𝛼)· ( 2) ≤ 3 (1/2−𝛼)𝑛 .
𝑛
3𝑛 43𝑛

Taking logarithms and solving for 𝛼, we get

1 3 log(4) 2 log(4)
𝛼≤ − + · 𝛿 ≤ 0.01341 + 2.524𝛿 ≤ 0.016.
4 16 log(3) log(3)

3.5 Extremal configurations

In this section we construct families of graphs that satisfy all the conditions of our
stability theorems, Theorem 3.1.2 and Theorem 3.1.3, and give a lower bound on the
triangle-free colourings these graphs have. These lower bounds are asymptotically
close to our upper bounds, proving that our upper bounds are asymptotically sharp

42
 3 Maximum Number of Triangle-free Edge Colourings

2
with an 𝑒 𝑜(𝑛 ) error factor.
Recall that we assume that the number of vertices 𝑛 is divisible by 24.
We define the following matrices

©0 0 0 0 0ª ©0 0 0 0 0ª ©0 0 0 0 0 0ª
­0 1 1 1 1®® ­0 1 1 1 1®® ­1 1 1 1 0 0®®
­ ­ ­
­ ® ­ ® ­ ®
­1 1 1 0 0®® ­1 1 1 0 0®® ­1 1 0 0 1 1®®
­ ­ ­
­1 0 0 1 1®® ­1 0 0 1 1®® ­0 0 1 1 1 1®®
𝐴1 = ­ ® 𝐴2 = ­ ® 𝐴3 = ­
­ ­ ­
®
­1 0 1 0 1®® ­1 0 1 0 1®® ­1 0 1 0 1 0®®
­ ­ ­
­1 1 0 1 0®® ­1 1 0 1 0®® ­0 1 0 1 1 0®®
­ ­ ­
­ ® ­ ® ­ ®
­1 0 1 1 0®® ­0 1 0 0 1®® ­0 1 1 0 0 1®®
­ ­ ­
«1 1 0 0 1¬ «0 0 1 1 0¬ «1 0 0 1 0 1¬

and for a matrix 𝐴 ∈ { 𝐴1 , 𝐴2 , 𝐴3 } let E 𝐴 = {𝜺 1 , . . . , 𝜺 8 } be the set of its row

vectors. In these matrices columns represent colours while rows represent disjoint
sets of vertices. The details of how exactly follows.

3.5.1 Extremal configurations for five colours

For 𝐴 = 𝐴1 and 𝑎, 𝑏, 𝑛 ∈ N with 𝑎 + 𝑏 ≤ 𝑛/4. Let 𝑐 denote 𝑛/4 − 𝑎 − 𝑏 and consider

a partition of [𝑛] into sets 𝑉 (𝜺 1 ) t · · · t 𝑉 (𝜺 8 ) such that
(i) |𝑉 (𝜺 1 )| = |𝑉 (𝜺 2 )| = 𝑛/4 and
(ii) |𝑉 (𝜺 3 )| = |𝑉 (𝜺 4 )| = 𝑎 and
(iii) |𝑉 (𝜺 5 )| = |𝑉 (𝜺 6 )| = 𝑏 and
(iv) |𝑉 (𝜺 7 )| = |𝑉 (𝜺 8 )| = 𝑐.


Let 𝐺 𝐴1 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 ) be the complete 8-partite graph with the partition
classes 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 ) and let B 𝐴(5)1 (𝑛) denote the set of all graphs obtained this
way. Note that in edge cases these graphs might be 4-partite or 6-partite instead of
8-partite. These graphs are also included in B 𝐴(5)1 (𝑛).

Proposition 3.5.1 (Lower bound for 𝑟 = 5, part 1). For each 𝐺 ∈ B 𝐴(5)1 (𝑛) we have
2
𝜑5 (𝐺) ≥ 6𝑛 /4 .

Proof. Let 𝐺 = 𝐺 𝐴1 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 𝑡 ) . Let us define 𝜋(𝜺, 𝜺0) as the set of colours

𝑐 for which 𝜺 𝑐 ≠ 𝜺0𝑐 . We colour the edges of 𝐺 according to the pattern 𝜋 = 𝜋 𝐴1 ,

i.e., for any (𝜺, 𝜺0) ∈ E 2𝐴 and 𝑢𝑣 ∈ 𝑉 (𝜺) × 𝑉 (𝜺0) we colour 𝑢𝑣 with any of the

43
 3 Maximum Number of Triangle-free Edge Colourings

colours in 𝜋(𝜺, 𝜺0). For example, an edge between vertices 𝑢, 𝑣 where 𝑢 ∈ 𝜺 3 and
𝑣 ∈ 𝜺 5 can have colours 2 or 5, as their rows differ at these coordinates.
Each of these colourings is triangle-free as 𝜋 −1
𝐴 (𝑐), for any 𝑐 ∈ [5], is bipartite.
Moreover, the number of such colourings is

𝑛2 𝑛2
Ö 0 2 +𝑏 2 +𝑐 2
|𝜋(𝜺, 𝜺0)| |𝑉 (𝜺)|·|𝑉 (𝜺 )| = 3 4 · 4 16 +𝑎 · 24𝑎𝑏+4𝑎𝑐+4𝑏𝑐
E𝐴
{𝜺,𝜺 0 }∈ ( 2 )

2 /4
which is easily verified to be 6𝑛 .

Note that this lower bound is not sharp, for example we can get further col-
ourings by permuting the colours. This also applies to Proposition 3.5.2 and
Proposition 3.5.3.
For 𝐴 = 𝐴2 and 𝑎, 𝑏, 𝑛 ∈ N with 𝑎, 𝑏 ≤ 𝑛/4. Let 𝑐 denote 𝑛/4 − 𝑎, 𝑑 denote
𝑛/4 − 𝑏 and consider a partition of [𝑛] into non-empty sets 𝑉 (𝜺 1 ) t · · · t 𝑉 (𝜺 8 )
such that
(i) |𝑉 (𝜺 1 )| = |𝑉 (𝜺 2 )| = 𝑎 and
(ii) |𝑉 (𝜺 3 )| = |𝑉 (𝜺 4 )| = 𝑏 and
(iii) |𝑉 (𝜺 5 )| = |𝑉 (𝜺 6 )| = 𝑑 and
(iv) |𝑉 (𝜺 7 )| = |𝑉 (𝜺 8 )| = 𝑐.


Let 𝐺 𝐴2 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 ) be the complete 8-partite graph with the partition
classes 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 ) and let B 𝐴(5)2 (𝑛) denote the set of all graphs obtained this
way. Note that in edge cases these graphs might be 4-partite or 6-partite instead of
8-partite. These graphs are also included in B 𝐴(5)2 (𝑛).

Proposition 3.5.2 (Lower bound for 𝑟 = 5, part 2). For each 𝐺 ∈ B 𝐴(5)2 (𝑛) we have
2
𝜑5 (𝐺) ≥ 6𝑛 /4 .

The proof is analogous to the proof of Proposition 3.5.1 and therefore omitted.

3.5.2 Extremal configurations for six colours

For 𝐴 = 𝐴3 and 𝑛 ∈ N consider a partition of [𝑛] into non-empty sets 𝑉 (𝜺 1 ) t · · · t

𝑉 (𝜺 8 ) such that |𝑉 (𝜺𝑖 )| = 𝑛/8 for all 𝑖 ∈ {1, . . . , 8}. Let 𝐺 𝐴3 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 )
be the complete 8-partite graph with the partition classes 𝑉 (𝜺 1 ), . . . , 𝑉 (𝜺 8 ) and let
B 𝐴(6)3 (𝑛) denote the set of all graphs obtained this way.

44
 3 Maximum Number of Triangle-free Edge Colourings

Proposition 3.5.3 (Lower bound for 𝑟 = 6). For each 𝐺 ∈ B 𝐴(6)3 (𝑛) we have
2 2
𝜑6 (𝐺) ≥ 3𝑛 /4 43𝑛 /16 .

The proof is analogous to the proof of Proposition 3.5.1 and therefore omitted.

3.6 Stability
In this section we prove Theorem 3.1.3 and Theorem 3.1.2 with the help of The-
orem 3.4.3.
We can assume that 𝑉 (𝐺) = [𝑛], we will do so for convenience. A ver-
tex partition 𝑉 0 ∪ 𝑉 1 = 𝑉 (𝐶) of the graph 𝐶 is called a 𝑡-bipartition if
|𝐸 (𝐶)4𝐾 (𝑉 0 , 𝑉 1 )| ≤ 𝑡, and it is called balanced if |𝑉 0 | = |𝑉 1 |±1. Furthermore, for
graphs 𝐶1 , . . . , 𝐶𝑟 with bipartitions 𝑉𝑖0 ∪ 𝑉𝑖1 = [𝑛] and 𝜺 = (𝜀1 , . . . , 𝜀𝑟 ) ∈ {0, 1}𝑟
we define the partition class
Ù
𝑉 (𝜺) = 𝑉𝑖𝜀𝑖 ,
𝑖∈[𝑟]

and set 𝑣(𝜺) = |𝑉 (𝜺)|. Let dist(𝜺, 𝜺0) denote the Hamming distance of the vectors
𝜺, 𝜺0 ∈ {0, 1} 𝑘 . Moreover, we call a pair 𝜺, 𝜺0 an 𝑚𝑖 -pair (or an 𝑚𝑖 -edge) if their
Hamming distance is 𝑖. Let us define the following constants, which will be used
in both proofs.
√ √
𝛾 = 2 · 106 𝛿 , 𝛼 = 10 𝛾 , 𝛽 = 10 𝛼 .

3.6.1 Stability proof for five colours

Proof of Theorem 3.1.3. By Theorem 3.2.1 and Corollary 3.2.2 taking

𝐶1 , . . . , 𝐶5 ∈ C(𝐺) such that |Φ(𝐶1 , . . . , 𝐶5 )| is maximal, we have 𝑡𝑟 (𝐶𝑖 ) < 𝑛25/9
2 2
and |Φ(𝐶1 , . . . , 𝐶5 )| ≥ 6𝑛 /4−2𝛿𝑛 .
By Theorem 3.4.3 there are balanced 𝛾𝑛2 -bipartitions 𝑉𝑖0 ∪ 𝑉𝑖1 = [𝑛] of 𝐶𝑖 ,
Ð
𝑖 ∈ [5], and 𝑋 ∪ 𝑌 = [𝑛] of 𝑀3 . We consider the partition 𝑉 (𝐺) = 𝜺∈{0,1}5 𝑉 (𝜺)
and let
E 𝑋 = {𝜺 ∈ {0, 1}5 : |𝑉 (𝜺) ∩ 𝑋 | > 𝛼𝑛}.
Ð
Let E𝑌 be defined analogously. Trivially 𝑉 (E 𝑋 ) = 𝜺∈E 𝑋 𝑉 (𝜺) has size

|𝑉 (E 𝑋 )| ≥ |𝑋 | − 25 𝛼𝑛 and |𝑉 (E𝑌 )| ≥ |𝑌 | − 25 𝛼𝑛. (3.6.1)

This establishes properties (i) and (ii) for E 𝑋 and E𝑌 .

45
 3 Maximum Number of Triangle-free Edge Colourings

Note that E0 will be chosen to be a subset of E 𝑋 and similarly E1 will be chosen

to be a subset of E𝑌 . For now we work on establishing the needed properties for
E 𝑋 and E𝑌 . For many of the properties this implies them trivially for subsets, for
others we will discuss the changes when we define E0 and E1 at the end of the
proof.
It will be more convenient to work only with edges of 𝐺 whose profiles are
consistent with the the partition classes containing its end vertices. More precisely,
we call 𝑢𝑣 ∈ 𝐺 a good edge (w.r.t. E 𝑋 ∪ E𝑌 ) if there are 𝜺 = (𝜀1 , . . . , 𝜀 5 ), 𝜺0 =
(𝜀01 , . . . , 𝜀05 ) ∈ E 𝑋 ∪ E𝑌 such that 𝑢 ∈ 𝑉 (𝜺) and 𝑣 ∈ 𝑉 (𝜺0) and

{𝑖 : 𝑢𝑣 ∈ 𝐶𝑖 } = {𝑖 : 𝜀𝑖 ≠ 𝜀𝑖0 } .

In other words, for each 𝑖 ∈ [5] with 𝜀𝑖 ≠ 𝜀𝑖0 the edge 𝑢𝑣 is contained in the bipartite
𝜀0 
graph 𝐶𝑖 𝑉𝑖𝜀𝑖 , 𝑉𝑖 𝑖 while 𝑢𝑣 ∉ 𝐶𝑖 if 𝜀𝑖 = 𝜀𝑖0. Note that in this case 𝑢𝑣 is an 𝑚 𝑑 -edge


where 𝑑 = dist(𝜺, 𝜺0). and a good 𝑚 𝑑 -edge is a witness for the existence of a pair
{𝜺, 𝜺0 } ∈ E 𝑋 ∪E with 𝑑 = dist(𝜺, 𝜺0).
𝑌


2
As each 𝑉𝑖0 ∪ 𝑉𝑖1 = [𝑛] is an 𝛾𝑛2 -bipartition, there are at most 5𝛾𝑛2 pairs of
vertices from 𝑉 (𝜺) × 𝑉 (𝜺0), over all (𝜺, 𝜺0) ∈ (E 𝑋 ∪ E𝑌 ) 2 , which do not form
good edges. Since 𝑋 ∪ 𝑌 is an 𝛾𝑛2 -partition of 𝑀3 and 𝑚 0 , 𝑚 1 , 𝑚 5 ≤ 2𝛾𝑛2 from
Theorem 3.4.3, we derive from the above and 𝑣(𝜺)𝑣(𝜺0) > 𝛼2 𝑛2 > 10𝛾𝑛2 that

each pair (𝜺, 𝜺0) ∈ E 𝑋 × E𝑌 has distance three and

(3.6.2)
each pair (𝜺, 𝜺0) ∈ E 𝑋2 ∪ E𝑌2 has distance two or four.

This establishes that properties (iii), (iv) and (v) hold.

Let 𝐵𝑖 be the bipartite graph induced by the good edges of 𝐺 on the bipartition
0
𝑉𝑖 ∪ 𝑉𝑖1 and let 𝐻 = 𝑖∈[5] 𝐵𝑖 . Moreover, let 𝑁𝑖 = 𝑀𝑖 (𝐵1 , . . . , 𝐵5 ), let 𝑛𝑖 = 𝑒(𝑁𝑖 )
Ð

and call an edge an 𝑛𝑖 -edge if it is contained in 𝑁𝑖 . Since all but at most 5𝛾𝑛2 +26 𝛼𝑛2
edges of 𝐺 are good we have
Õ
𝑚𝑖 − 𝑛𝑖 = 𝑒(𝐺) − 𝑒(𝐻) ≤ 27 𝛼𝑛2 (3.6.3)
𝑖∈[5]

and, for each 𝑖 ∈ [5],

|𝑚𝑖 − 𝑛𝑖 | ≤ 27 𝛼𝑛2 . (3.6.4)

46
 3 Maximum Number of Triangle-free Edge Colourings

Furthermore, as each 𝐵𝑖 is triangle-free and the restriction of each colouring

𝜑 ∈ Φ(𝐶1 , . . . , 𝐶5 ) of 𝐺 to 𝐻 is a colouring in Φ𝐻 (𝐵1 , . . . , 𝐵5 ), we have

𝑛2
Ö 7 𝛼𝑛2 7 𝛼𝑛2
𝑖 𝑛𝑖 = |Φ(𝐵1 , . . . , 𝐵5 )| ≥ |Φ(𝐶1 , . . . , 𝐶5 )|5−2 ≥ 6 4 −2 . (3.6.5)
𝑖∈[5]

We now set out to establish property (vi) and that |E | ≤ 8. From Theorem 3.4.3
and (3.6.3) we have

𝑛2
𝑛2 + 2𝑛4 ≥ 𝑚 2 + 2𝑚 4 − 28 𝛼𝑛2 ≥ − 29 𝛼𝑛2 . (3.6.6)
4

Due to (3.6.2) we have 𝑛2 (𝑋, 𝑌 ) +2𝑛4 (𝑋, 𝑌 ) = 0 and as 𝑛𝑖 = 𝑛𝑖 (𝑋) +𝑛𝑖 (𝑌 ) +𝑛𝑖 (𝑋, 𝑌 )
for each 𝑖 ∈ [5], we conclude therefore that either

𝑛2 𝑛2
𝑛2 (𝑋) + 2𝑛4 (𝑋) ≥ − 29 𝛼𝑛2 or 𝑛2 (𝑌 ) + 2𝑛4 (𝑌 ) ≥ − 29 𝛼𝑛2 .
8 8
(3.6.7)

Hence, as 𝑋 ∪ 𝑌 is a equipartition, either 𝑋 or 𝑌 will satisfy the presumption of the

following claim.
Claim 3.6.1. If 𝑛2 (𝑋) + 2𝑛4 (𝑋) > 25 |𝑋 | 2 + 26 𝛽𝑛2 then there are 𝜺 0 , 𝜺 1 ∈ E 𝑋 with
Hamming distance four and |𝑉 (𝜺 0 )|, |𝑉 (𝜺 1 )| > 𝛽𝑛. The analogous statement holds
for 𝑌 .

Proof. Let 𝐾 be the largest clique in 𝑁2 [𝑋] and consider K = {𝜺 : 𝑉 (𝐾) ∩ 𝑉 (𝜺) ≠
∅} ⊆ E 𝑋 . By possibly swapping the upper indices of 𝑉𝑖1 and 𝑉𝑖0 for some 𝑖 ∈ [5]
we may assume that (0, 0, 0, 0, 0) ∈ K. Then, by definition of 𝑁2 each of the
remaining element in K has exactly two 1-entries and two of them differ at exactly
two entries. This implies that any two elements from K \ {(0, 0, 0, 0, 0)} coincide
in (exactly) one 1-entry, hence, the family corresponds to a 2-uniform intersecting
family on a ground set of size five. By the Erdős-Ko-Rado Theorem [23] we know
that the set has size at most four. Hence, |K | ≤ 5 and Turán’s Theorem [55]
2
implies that 𝑁2 [𝑋] contains at most 45 |𝑋2| edges. From the presumption on 𝑋 we
then obtain 𝑛4 (𝑋) > 25 𝛽𝑛2 . Therefore there must be an 𝑛4 -edges in 𝑋 with ends
vertices in partition classes of size larger than 𝛽𝑛. This edge is a witness for the
desired pair in E 𝑋 with Hamming distance four. 

By possibly renaming the partition classes 𝑋 and 𝑌 suppose that 𝑋 satis-

47
 3 Maximum Number of Triangle-free Edge Colourings

fies (3.6.7). Hence, Claim 3.6.1 provides a pair 𝜺 0 , 𝜺 1 ∈ E 𝑋 with Hamming

distance four and let 𝑠 ∈ [5] be the (unique) coordinate where 𝜺 0 , 𝜺 1 agree. By
possibly reordering the 𝐶𝑖 ’s and swapping the upper indices of 𝑉𝑖1 and 𝑉𝑖0 for some
𝑖 ∈ [5] we may also assume that 𝜺 0 = (0, 0, 0, 0, 0) and 𝜺 1 = (1, 1, 1, 1, 0).
From (3.6.2) we know that the Hamming distances of any 𝜺 ∈ E𝑌 to 𝜺 0 and
to 𝜺 1 are both three. Therefore each 𝜺 = (𝜀1 , . . . , 𝜀 5 ) ∈ E𝑌 must satisfy 𝜀 5 = 1
since otherwise (𝜀1 , . . . , 𝜀 4 ) must contain three ones and three zeroes. Hence,
𝑉 (𝜺) ⊆ 𝑌 ∩ 𝑉51 for each 𝜺 ∈ E𝑌 . Recall from (3.6.1) that 𝑉 (E𝑌 ) ⊆ 𝑌 ∩ 𝑉51 covers at
least |𝑌 | − 25 𝛼𝑛 vertices of 𝑌 ∩ 𝑉51 , hence all but at most 25 𝛼𝑛 + 1 vertices of 𝑉51 as
|𝑉51 | = |𝑌 | ± 1. Thus the sizes of those 𝑉 (𝜺) with 𝜺 = (𝜀1 , . . . , 𝜀 5 ) ∈ E 𝑋 , 𝜀 5 = 1,
all together is at most |𝑉51 ∩ 𝑋 | = |𝑉51 \ 𝑌 | ≤ 25 𝛼𝑛 + 1. With this in mind we set

E𝑌0 = E𝑌 and E 0𝑋 = {(𝜀1 , . . . , 𝜀 5 ) ∈ E 𝑋 : 𝜀5 = 0} (3.6.8)

and from (3.6.1) we have |𝑉 (E 0𝑋 )| ≥ |𝑉 (E 𝑋 )| − 25 𝛼𝑛 − 1 > |𝑋 | − 27 𝛼𝑛.

Claim 3.6.2. For each 𝜺˜ ∈ E 0𝑋 there is at most one element at distance four to 𝜺˜
in E 0𝑋 . Furthermore, if 𝑣( 𝜺)
˜ > 𝛽𝑛 then there is an 𝜺ˆ ∈ E 0𝑋 at distance four to 𝜺˜
which, moreover, satisfies 𝑣( 𝜺)˜ = 𝑣( 𝜺)
ˆ ± 𝛽𝑛/2.

Proof. The first part is clear as 𝜀5 = 0 for all (𝜀 1 , . . . , 𝜀 5 ) ∈ E 0𝑋 . For a contradiction

suppose that there exists 𝜺˜ = ( 𝜀˜1 , . . . , 𝜀˜5 ) ∈ E 0𝑋 with 𝑣( 𝜺) ˜ > 𝛽𝑛 which has
Hamming distance two to all 𝜺 ∈ E 𝑋 , in particular to 𝜺 0 = (0, 0, 0, 0, 0) ∈ E 0𝑋 and
0

𝜺 1 = (1, 1, 1, 1, 0) ∈ E 0𝑋 .
Let 𝐴 = 𝑉 (𝜺 0 ), 𝐵 = 𝑉 (𝜺 1 ) and let 𝐿 ∪ 𝑅 = 𝑉 ( 𝜺) ˜ be an equitable partition
˜ We modify 𝐻 by deleting all vertices not contained in 𝑉 (E 0𝑋 ∪ E𝑌0 ) and
of 𝑉 ( 𝜺).
further, deleting the edges between 𝐴 and 𝐿 and those between 𝐵 and 𝑅, and adding
all possible edges edges between 𝐿 and 𝑅. Let 𝐻 0 denote the so obtained graph.
Furthermore, define a partition of 𝑉 (E 0𝑋 ∪ E𝑌0 ) by
• 𝑈 (𝜺 0 ) = 𝐴 ∪ 𝐿 and 𝑈 (𝜺 1 ) = 𝐵 ∪ 𝑅,
• 𝑈 ( 𝜺)
˜ = ∅ and 𝑈 (𝜺) = 𝑉 (𝜺) for each 𝜺 ∈ (E 0𝑋 ∪ E𝑌0 ) \ {𝜺 0 , 𝜺 1 , 𝜺}.
˜
Let 𝑈𝑖0 ∪ 𝑈𝑖1 = [𝑛], 𝑖 ∈ [5], be the corresponding bipartition defined by
Ø Ù
for all 𝜺 ∈ E 0𝑋 ∪ E𝑌0 .
𝑗
𝑈𝑖 = 𝑈 (𝜺), so that 𝑈 (𝜺) = 𝑈𝑖𝜀𝑖
𝜺∈E 𝑋0 ∪E𝑌0 𝑖∈[5]
𝜀𝑖 = 𝑗

For each 𝑖 ∈ [5] let 𝐵𝑖0 be the bipartite graph 𝐻𝑖0 𝑈𝑖0 ∪ 𝑈𝑖1 . Further, let
 

48
 3 Maximum Number of Triangle-free Edge Colourings

𝑁𝑖0 = 𝑀𝑖 (𝐵01 , . . . , 𝐵05 ) and 𝑛0𝑖 = 𝑒(𝑁𝑖0). Note that the edges deleted from 𝐻 either
has an end vertex in a partition class 𝑉 (𝜺) with 𝜺 ∈ E 𝑋 \ E 0𝑋 or are edges in 𝑁2 .
On the other hand, the 𝑛2 -edges between 𝐿 and 𝐵, those between 𝑅 and 𝐴 as well
as the new edges between 𝐿 and 𝑅 are now edges of 𝑁40 . All remaining edges of 𝐻
do not change their type, hence
• 𝑛03 ≥ 𝑛3 − 26 𝛼𝑛2 ,
• 𝑛02 ≥ 𝑛2 − 𝑛2 (𝐿 ∪ 𝑅, 𝐴 ∪ 𝐵) − 26 𝛼𝑛2 , and
• 𝑛04 ≥ 𝑛4 + 𝑛2 (𝐿, 𝐵) + 𝑛2 (𝑅, 𝐴) + |𝐿||𝑅| − 26 𝛼𝑛2 .
Up to renaming 𝐿 and 𝑅 suppose that 𝑛2 (𝐿, 𝐵) + 𝑛2 (𝑅, 𝐴) ≥ 𝑛2 (𝐿, 𝐴) + 𝑛2 (𝑅, 𝐵).
Then

2𝑛2 (𝐿, 𝐵) + 2𝑛2 (𝑅, 𝐴) ≥ 𝑛2 (𝐿, 𝐵) + 𝑛2 (𝑅, 𝐴) + 𝑛2 (𝐿, 𝐴) + 𝑛2 (𝑅, 𝐵)

= 𝑛2 (𝐿 ∪ 𝑅, 𝐴 ∪ 𝐵) ,

and hence 𝑛02 + 2𝑛04 ≥ 𝑛2 + 2𝑛4 + 2|𝐿||𝑅| − 28 𝛼𝑛2 . As all the 𝐵𝑖0 are triangle-free
we obtain from |𝐿||𝑅| ≥ 𝛽2 𝑛2 /5 and (3.6.5) that

0 0 0 6 𝛼𝑛2 8 𝛼𝑛2
|Φ(𝐵01 , . . . , 𝐵05 )| = 2𝑛2 3𝑛3 4𝑛4 ≥ 3𝑛3 −2 2𝑛2 +2𝑛4 +2|𝐿||𝑅|−2
𝑛2 9 𝛼𝑛2 𝑛2 2
> 6 4 −2 4|𝐿||𝑅| > 6 4 +𝛾𝑛 .

This, however, yields a contradiction to the upper bound given by Theorem 3.4.3
and hence, 𝜺˜ must be at distance four to some 𝜺ˆ ∈ E 0𝑋 .
To show that 𝑣( 𝜺) ˜ = 𝑣( 𝜺)ˆ ± 𝛽𝑛/2 we argue in a similar manner. If, say,
𝑣( 𝜺)
˜ > 𝑣( 𝜺)ˆ + 𝛽𝑛 then we make the two sets equitable by moving vertices from
𝑉 ( 𝜺)
˜ to 𝑉 ( 𝜺)
ˆ and changing the edges accordingly. This increases the number of
𝑛4 -edges by at least 𝛽2 𝑛2 /20 which would again yield a contradiction to the upper
bound given by Theorem 3.4.3. 

In view of the lemma we choose E0 to consist of those 𝜺˜ ∈ E 0𝑋 with 𝑣( 𝜺) ˜ > 𝛽𝑛

0
together with the unique corresponding 𝜺ˆ ∈ E 𝑋 at distance four. Hence, 𝑉 (E0 )
covers all but at most 25 𝛽𝑛 vertices of 𝑋. Further, note that the projections of
𝜺 ∈ E0 and of 𝜺0 ∈ E𝑌0 to the first four coordinates yields vectors with distance

E0

two, while the projections of 𝜺 and 𝜺0 with {𝜺, 𝜺0 } ∈ E20 ∪ 2𝑌 yields vectors with
distance two or four. Hence, |E0 ∪ E𝑌0 | ≤ 8 and as E𝑌0 ≠ ∅ we have |E0 | ∈ {2, 4, 6}.
Note that if E0 has size two then the largest value of 𝑛2 (𝑋) + 2𝑛4 (𝑋) is achieved by
2
a balanced complete bipartite graph. In this case 𝑛2 (𝑋) = 0 and 𝑛4 (𝑋) ≤ |𝑋4| . If

49
 3 Maximum Number of Triangle-free Edge Colourings

E0 has size four then the optimum is achieved by any four partite graph consisting
of two pairs of equitable parts, say the first pair of sizes 𝑎 and the second with
sizes 𝑏 and 𝑏 ± 1 with 2𝑎 + 2𝑏 = |𝑋 | ± 1. In this case, 𝑛2 (𝑋) ≤ 4𝑎𝑏 + 2𝑎 and
𝑛4 (𝑋) ≤ 𝑎 2 + 𝑏 2 + 𝑏, hence, 𝑛2 (𝑋) + 2𝑛4 (𝑋) ≤ 2(𝑎 + 𝑏) 2 + 2(𝑎 + 𝑏). For E 𝑋 of size
six the optimum is achieved by any six partite graph consisting of three pairs of
equitable parts, say the first of sizes 𝑎, the second of sizes 𝑏 and the third of sizes 𝑐
and 𝑐 ±1 with 2𝑎 +2𝑏 +2𝑐 = |𝑋 | ±1. In this case, 𝑛2 (𝑋) ≤ 4𝑎𝑏 +4𝑎𝑐 +4𝑏𝑐 +2𝑎 +2𝑏
and 𝑛4 (𝑋) ≤ 𝑎 2 +2𝑏 2 +2𝑐2 +𝑐, hence, 𝑛2 (𝑋) +2𝑛4 (𝑋) ≤ 2(𝑎 + 𝑏 +𝑐) 2 +2(𝑎 + 𝑏 +𝑐).
2
Hence, in any of these cases, we have 𝑛2 (𝑋) +2𝑛4 (𝑋) ≤ |𝑋2| +5|𝑋 | and by (3.6.6)
we derive that 𝑛2 (𝑌 ) + 2𝑛4 (𝑌 ) > 25 |𝑌 | 2 + 26 𝛽2 𝑛2 . By Claim 3.6.1 there is a pair
of elements at distance four in E𝑌0 and the same argument as in Claim 3.6.2 shows
that for each 𝜺˜ ∈ E𝑌0 with 𝑣( 𝜺)
˜ > 𝛽𝑛 there is an 𝜺ˆ ∈ E𝑌0 at distance four to 𝜺˜ and
𝑣( 𝜺) ˆ ± 𝛽𝑛/2. We choose E1 to consist of those 𝜺 ∈ E𝑌0 with 𝑣(𝜺) > 𝛽𝑛
˜ = 𝑣( 𝜺)
together with the unique corresponding 𝜺ˆ ∈ E𝑌0 at distance four. This establishes
property (vi).
This yields the desired families E0 and E1 . Since we only lost vertex groups of
size at most 𝛽𝑛 and at most 25 of them, properties (i) and (ii) still hold.

50
 3 Maximum Number of Triangle-free Edge Colourings

3.6.2 Stability proof for six colours

By Theorem 3.2.1 and Corollary 3.2.2 taking 𝐶1 , . . . , 𝐶5 ∈ C(𝐺) such that

|Φ(𝐶1 , . . . , 𝐶5 )| is maximal, we have 𝑡𝑟 (𝐶𝑖 ) < 𝑛25/9 and |Φ(𝐶1 , . . . , 𝐶5 )| ≥
2 2
6𝑛 /4−2𝛿𝑛 .

Proof of Theorem 3.1.2. By Theorem 3.2.1 and Corollary 3.2.2 taking

𝐶1 , . . . , 𝐶5 ∈ C(𝐺) such that |Φ(𝐶1 , . . . , 𝐶6 )| is maximal, we have 𝑡𝑟 (𝐶𝑖 ) < 𝑛25/9
2 2 2
and |Φ(𝐶1 , . . . , 𝐶6 )| ≥ 3𝑛 43𝑛 /16−2𝛿𝑛 .
By Theorem 3.4.3 there are balanced 𝛾𝑛2 -bipartitions 𝑉𝑖0 ∪ 𝑉𝑖1 = [𝑛] of 𝐶𝑖 ,
Ð
𝑖 ∈ [6], and we consider the partition 𝑉 (𝐺) = 𝜺∈{0,1}6 𝑉 (𝜺). Consider

E0 = E 𝑋 = {𝜺 ∈ {0, 1}6 : |𝑉 (𝜺) ∩ 𝑋 | > 𝛼𝑛}

Ð
and let E1 = E𝑌 be defined analogously. Trivially 𝑉 (E 𝑋 ) = 𝜺∈E 𝑋 𝑉 (𝜺) has size

|𝑉 (E 𝑋 )| ≥ |𝑋 | − 26 𝛼𝑛 and |𝑉 (E𝑌 )| ≥ |𝑌 | − 26 𝛼𝑛. (3.6.9)

This establishes properties (i) and (ii).

It will be more convenient to work only with edges of 𝐺 whose profiles are
consistent with the the partition classes containing its end vertices. More precisely,
we call 𝑢𝑣 ∈ 𝐺 a good edge (w.r.t. E 𝑋 ∪ E𝑌 ) if there are 𝜺 = (𝜀1 , . . . , 𝜀 6 ), 𝜺0 =
(𝜀01 , . . . , 𝜀06 ) ∈ E 𝑋 ∪ E𝑌 such that 𝑢 ∈ 𝑉 (𝜺) and 𝑣 ∈ 𝑉 (𝜺0) and {𝑖 : 𝑢𝑣 ∈ 𝐶𝑖 } =
{𝑖 : 𝜀𝑖 ≠ 𝜀𝑖0 }. In other words, for each 𝑖 ∈ [6] with 𝜀𝑖 ≠ 𝜀𝑖0 the edge 𝑢𝑣 is contained
𝜀0 
in the bipartite graph 𝐶𝑖 𝑉𝑖𝜀𝑖 , 𝑉𝑖 𝑖 while 𝑢𝑣 ∉ 𝐶𝑖 if 𝜀𝑖 = 𝜀𝑖0. Note that in this case


𝑢𝑣 is an 𝑚 𝑑 -edge where 𝑑 = dist(𝜺, 𝜺0) and a good 𝑚 𝑑 -edge is a witness for the
existence of a pair {𝜺, 𝜺0 } ∈ E 𝑋 ∪E with 𝑑 = dist(𝜺, 𝜺0).
𝑌


2
As each 𝑉𝑖0 ∪ 𝑉𝑖1 = [𝑛] is an 𝛾𝑛2 -bipartition, there are at most 6𝛾𝑛2 pairs of
vertices from 𝑉 (𝜺) × 𝑉 (𝜺0), over all {𝜺, 𝜺0 } ∈ E 𝑋 ∪E


2
𝑌
, which do not form good
edges. Since 𝑋 ∪ 𝑌 is an 𝛾𝑛2 -bipartition of 𝑀3 and 𝑚 0 , 𝑚 1 , 𝑚 5 , 𝑚 6 ≤ 𝛾𝑛2 , by
Theorem 3.4.3, we derive from the above and 𝑣(𝜺)𝑣(𝜺0) > 𝛼2 𝑛2 > 𝛾𝑛2 + 3𝛾𝑛2 +
6𝛾𝑛2 > 10𝛾𝑛2 that

each pair (𝜺, 𝜺0) ∈ E 𝑋 × E𝑌 has distance three and

(3.6.10)
each pair (𝜺, 𝜺0) ∈ E 𝑋2 ∪ E𝑌2 has distance two or four.

This establishes that properties (iii) and (iv) hold.

Let 𝐵𝑖 be the bipartite graph induced by the good edges of 𝐺 on the bipartition

51
 3 Maximum Number of Triangle-free Edge Colourings

𝑉𝑖0 ∪ 𝑉𝑖1 and let 𝐻 = 𝑖∈[6] 𝐵𝑖 . Moreover, let 𝑁𝑖 = 𝑀𝑖 (𝐵1 , . . . , 𝐵6 ), let 𝑛𝑖 = 𝑒(𝑁𝑖 )
Ð

and call an edge an 𝑛𝑖 -edge if it is contained in 𝑁𝑖 . Since all but at most 6𝛾𝑛2 +26 𝛼𝑛2
edges of 𝐺 are good we have
Õ
𝑚𝑖 − 𝑛𝑖 = 𝑒(𝐺) − 𝑒(𝐻) ≤ 6𝛾𝑛2 + 26 𝛼𝑛2 ≤ 27 𝛼𝑛2 (3.6.11)
𝑖∈[5]

and for each 𝑖 ∈ [6]

|𝑚𝑖 − 𝑛𝑖 | ≤ 27 𝛼𝑛2 . (3.6.12)

Furthermore, as each 𝐵𝑖 is triangle-free and the restriction of each colouring

𝜑 ∈ Φ(𝐶1 , . . . , 𝐶5 ) of 𝐺 to 𝐻 is a colouring in Φ𝐻 (𝐵1 , . . . , 𝐵5 ), we have
Ö 7 𝛼𝑛2 2 /4 2 /16 8 𝛼𝑛2
𝑖 𝑛𝑖 = |Φ(𝐵1 , . . . , 𝐵5 )| ≥ |Φ(𝐶1 , . . . , 𝐶5 )|6−2 ≥ 3𝑛 43𝑛 6−2 .
𝑖∈[5]
(3.6.13)

Due to (3.6.10) and (3.6.11), we have 𝑛3 (𝑋, 𝑌 ) = 𝑛3 (𝐻) ≥ 𝑒(𝑀3 ) − 27 𝛼𝑛2 ≥

𝑛2 /4 − 28 𝛼𝑛2 .
Our next goal is to prove |E | ≤ 8, the rest of the properties will easily follow.
Claim 3.6.3. |E 𝑋 |, |E𝑌 | ≥ 4

Proof. We first note that for each 𝑖 ∈ [6]

𝑛2 𝑛2
|𝑁3 (𝑋, 𝑌 ) ∩ 𝐸 (𝐵𝑖 )| ≥ − 29 𝛼𝑛2 ≥ − 29 𝛼𝑛2 .
8 8

Also, we have that

Õ
|𝑁3 (𝑋, 𝑌 ) ∩ 𝐸 (𝐵𝑖 )| = |{(𝑒, 𝑖) : 𝑒 ∈ 𝑁3 (𝑋, 𝑌 ) ∩ 𝐸 (𝐵𝑖 )}| ≤ 3𝑛3 (𝑋, 𝑌 )
𝑖∈[6]

≤ 3(𝑛2 /4 + 𝛾𝑛2 ) .

𝑛2 𝑛2
Hence, for each 𝑖 ∈ [6] we have that |𝑁3 (𝑋, 𝑌 ) ∩ 𝐸 (𝐵𝑖 )| ≤ 8 + 16 𝛾𝑛2 ≤ 8 + 29 𝛼𝑛2 .
Altogether, for each 𝑖 ∈ [6],

𝑛2
|𝑁3 (𝑋, 𝑌 ) ∩ 𝐸 (𝐵𝑖 )| = ± 29 𝛼𝑛2 .
8

52
 3 Maximum Number of Triangle-free Edge Colourings

Therefore, for each 𝑖 ∈ [6], the following holds

𝑛
|𝑋 ∩ 𝑉𝑖0 | = |𝑋 ∩ 𝑉𝑖1 | = |𝑌 ∩ 𝑉𝑖0 | = |𝑌 ∩ 𝑉𝑖1 | = ± 25 𝛼𝑛 , (3.6.14)
4

and hence, |E 𝑋 |, |E𝑌 | ≥ 2.

From now on in this proof, for a subset 𝑈 ⊆ 𝑉 (𝐺), we use the notation 𝑈˜ = 𝑋 ∩𝑈.
Suppose that 2 ≤ |E 𝑋 | ≤ 3. Then, for some ordering 𝜺 1 , . . . 𝜺 26 of the elements of
{0, 1}6 , we have that |𝑉˜ (𝜺𝑖 )| ≤ 𝛼𝑛 for each 𝑖 ∈ {4, . . . , 26 } and therefore

Ø
𝑉˜ (𝜺𝑖 ) ≤ 26 𝛼𝑛. (3.6.15)
𝑖≥4

Since 26 𝛼𝑛 < 𝑛4 ± 25 𝛼𝑛, due to (3.6.14) and (3.6.15), for each 𝑖 ∈ [6], there are
distinct 𝑘 𝑖 , 𝑡𝑖 ∈ {1, 2, 3} such that

𝑉˜ (𝜺 𝑘 𝑖 ) ⊆ 𝑉˜𝑖1 and 𝑉˜ (𝜺 𝑡𝑖 ) ⊆ 𝑉˜𝑖0

Hence, we can assume that for a fixed 𝑖 ∈ [6], by possibly interchanging 𝑉𝑖1 ,𝑉𝑖0 ,
that there distinct 𝑘, ℓ, 𝑡 ∈ {1, 2, 3} such that

(𝑉˜ (𝜺 𝑘 ) ∪ 𝑉˜ (𝜺 ℓ )) ⊆ 𝑉˜𝑖1 and 𝑉˜ (𝜺 𝑡 ) ⊆ 𝑉˜𝑖0 . (3.6.16)

Moreover, since 𝜺 𝑘 ≠ 𝜺 𝑙 there is some 𝑗 ∈ [6] with 𝑗 ≠ 𝑖 such that

𝑉˜ (𝜺 𝑘 ) ⊆ ∩𝑉 𝑗1 and (𝑉˜ (𝜺 𝑡 ) ∪ 𝑉˜ (𝜺 ℓ )) ⊆ 𝑉˜ 𝑗0 , (3.6.17)

by possibly interchanging 𝑉 𝑗1 ,𝑉 𝑗0 , and 𝜺 𝑘 , 𝜺 ℓ .

Due to (3.6.15), (3.6.16) and (3.6.17), we have that |𝑉˜𝑖0 | ≤ |𝑉˜ (𝜺 𝑡 )| + 26 𝛼𝑛 and
|𝑉˜ 𝑗1 | ≤ |𝑉˜ (𝜺 𝑘 )| + 26 𝛼𝑛. Hence, using (3.6.14), we obtain that

|𝑉˜ (𝜺 𝑡 )| ≥ 𝑛/4 − 27 𝛼𝑛 and |𝑉˜ (𝜺 𝑘 )| ≥ 𝑛/4 − 27 𝛼𝑛. (3.6.18)

Furthermore, by (3.6.14) and (3.6.16), we have |𝑉˜ (𝜺 𝑘 )| + |𝑉˜ (𝜺 𝑙 )| ≤ |𝑉˜𝑖0 | ≤ 𝑛/4 +

25 𝛼𝑛, which combined with (3.6.18) yields

|𝑉˜ (𝜺 𝑙 )| ≤ 28 𝛼𝑛.

Now, due to (3.6.14), (3.6.15), the previous inequality and the fact that 29 𝛼𝑛 <

53
 3 Maximum Number of Triangle-free Edge Colourings

𝑛
4± 25 𝛼𝑛, we have that for each 𝑖 ∈ [6], 𝑘 𝑖 , 𝑡𝑖 ∈ {𝑘, 𝑡}, and 𝜺 𝑘 , 𝜺 𝑡 are at distance 6.
Due to (3.6.18), 𝜺 𝑘 , 𝜺 𝑡 ∈ E 𝑋 , and we have obtained a contradiction to 3.6.10.


From Theorem 3.4.3 and (3.6.11) we have

3𝑛2
𝑛2 + 2𝑛4 ≥ 𝑚 2 + 2𝑚 4 − 27 𝛼𝑛2 ≥ − 28 𝛼𝑛2 . (3.6.19)
8

It follows from (3.6.10) that we have 𝑛2 (𝑋, 𝑌 ) + 2𝑛4 (𝑋, 𝑌 ) = 0. Since 𝑛𝑖 =

𝑛𝑖 (𝑋) + 𝑛𝑖 (𝑌 ) + 𝑛𝑖 (𝑋, 𝑌 ) for each 𝑖 ∈ [6], we conclude therefore that either

3𝑛2 3𝑛2
𝑛2 (𝑋) + 2𝑛4 (𝑋) ≥ − 27 𝛼𝑛2 or 𝑛2 (𝑌 ) + 2𝑛4 (𝑌 ) ≥ − 27 𝛼𝑛2 .
16 16
(3.6.20)

Hence, as 𝑋 ∪ 𝑌 is a equipartition, either 𝑋 or 𝑌 will satisfy the presumption of the

following claim.
5
Claim 3.6.4. If 𝑛2 (𝑋) + 2𝑛4 (𝑋) > 12 |𝑋 | 2 + 26 𝛼𝑛2 then there are 𝜺 0 , 𝜺 1 ∈ E 𝑋
with Hamming distance four and |𝑉 (𝜺 0 )|, |𝑉 (𝜺 1 )| > 𝛼𝑛. The analogous statement
holds for 𝑌 .

Proof. Let 𝐾 be the largest clique in 𝑁2 [𝑋] and consider K = {𝜺 : 𝑉 (𝐾) ∩ 𝑉 (𝜺) ≠
∅} ⊆ E 𝑋 . By possibly swapping the upper indices of 𝑉𝑖1 and 𝑉𝑖0 for some 𝑖 ∈ [6]
we may assume that (0, 0, 0, 0, 0, 0) ∈ K. Then, by definition of 𝑁2 each of the
remaining element in K has exactly two 1-entries and two of them differ at exactly
two entries. This implies that any two elements from K \{(0, 0, 0, 0, 0, 0)} coincide
in (exactly) one 1-entry, hence, the family corresponds to a 2-uniform intersecting
family on a ground set of size six. By the Erdős-Ko-Rado Theorem [23] we know
that the set has size at most five. Hence, |K | ≤ 6 and Turán’s Theorem [55] implies
5
that 𝑁2 [𝑋] contains at most 12 |𝑋 | 2 edges. From the presumption on 𝑋 we then
obtain 𝑛4 (𝑋) > 25 𝛼𝑛2 . Therefore there must be an 𝑛4 -edge in 𝑋 with end vertices
in partition classes of size larger than 𝛼𝑛. This edge is a witness for the desired
pair in E 𝑋 with Hamming distance four. 

By possibly renaming the partition classes 𝑋 and 𝑌 suppose that 𝑋 satis-

fies (3.6.20). Hence, the claim provides a pair 𝜺 0 , 𝜺 1 ∈ E 𝑋 with Hamming distance
four. By possibly reordering the 𝐶𝑖 ’s and swapping the upper indices of 𝑉𝑖1 and 𝑉𝑖0
for some 𝑖 ∈ [6] we assume that 𝜺 0 = (0, 0, 0, 0, 0, 0) and 𝜺 1 = (1, 1, 1, 1, 0, 0).

54
 3 Maximum Number of Triangle-free Edge Colourings

We now define set E 0𝑋 . If the last coordinate of 𝜺 ∈ E 𝑋 is 0, then 𝜺 ∈ E 0𝑋 and

if the last coordinate of 𝜺 ∈ E 𝑋 is 1, then 𝜺 ∈ E 0𝑋 , where 𝜺 + 𝜺 = (1, 1, 1, 1, 1, 1).
The set E𝑌0 is defined analogously. Since no pair 𝜺, 𝜺0 of elements in E 𝑋 ∪ E𝑌 is at
distance 6, we have that |E 0𝑋 | = |E 𝑋 |, |E𝑌0 | = |E𝑌 | and

|E 0𝑋 ∪ E𝑌0 | = |E 𝑋 ∪ E𝑌 | . (3.6.21)

Moreover, due to (3.6.10), we have that

each pair (𝜺, 𝜺0) ∈ E 0𝑋 × E𝑌0 has distance three and

(3.6.22)
each pair (𝜺, 𝜺0) ∈ (E 0𝑋 ) 2 ∪ (E𝑌0 ) 2 has distance two or four.

Moreover, each 𝜺 = (𝜀 1 , . . . , 𝜀 6 ) ∈ E 0𝑋 ∪ E𝑌0 satisfies 𝜀6 = 0. From (3.6.22)

we know that the Hamming distances of any 𝜺 ∈ E𝑌0 to 𝜺 0 and to 𝜺 1 are both
three. Therefore each 𝜺 = (𝜀1 , . . . , 𝜀 6 ) ∈ E𝑌0 must satisfy 𝜀5 = 1 since otherwise
(𝜀1 , . . . , 𝜀 4 ) must contain three ones and three zeroes. Also, the following claim
holds.
Claim 3.6.5. Each 𝜺 = (𝜀1 , . . . , 𝜀 6 ) ∈ E 0𝑋 satisfies 𝜀5 = 0.

Proof. For the sake of contradiction suppose that there is 𝜺 = (𝜀1 , . . . , 𝜀 6 ) ∈ E 0𝑋

with 𝜀5 = 1. Due to (3.6.22), 𝜺 is at distance 2 or 4 from 𝜺 0 and 𝜺 1 and hence,
there is 𝑖 ∈ {1, 2, 3, 4} such that 𝜀𝑖 ≠ 𝜀 𝑗 ∈ {1, 2, 3, 4} − {𝑖}. Again due to (3.6.22),
the distance between 𝜺 and 𝜺0 = (𝜀01 , . . . , 𝜀06 ) ∈ E 0𝑋 is three, and since 𝜀5 = 𝜀05
and 𝜀6 = 𝜀06 , we must have that (𝜀1 , . . . , 𝜀 4 ) and (𝜀01 , . . . , 𝜀04 ) are at distance
3. Moreover, (𝜀01 , . . . , 𝜀04 ) is at distance 2 with (0, 0, 0, 0) and (1, 1, 1, 1), which
implies that (𝜀01 , . . . , 𝜀04 ) has exactly two 1-entries. It implies that 𝜀𝑖 = 𝜀𝑖0 and there
are at most 32 = 3 elements in E𝑌0 , a contradiction to Claim 3.6.3.




Claim 3.6.5 implies that |E 0𝑋 ∪ E𝑌0 | ≤ 8. Moreover, |E 𝑋 ∪ E𝑌 | ≤ 8 by (3.6.21)

and due to Claim 3.6.3, we have |E | = |E 𝑋 ∪ E𝑌 | = 8.
We now need to show that each pair (𝜺, 𝜺0) ∈ E 𝑋2 ∪ E𝑌2 has distance four. But
this is clear from the fact that it can be realized. This establishes property (v).
Finally we show that for each (𝜺, 𝜺0) ∈ (E 𝑋 ∪ E 𝑦 ) 2 we have 𝑣(𝜺) = 𝑣(𝜺0) ± 𝛽𝑛.
This proof is analogous to the proof of (vi) in Theorem 3.1.3. Since 𝑋, 𝑌 is an
equipartition we only need to consider pairs (𝜺, 𝜺0) ∈ E 𝑋2 ∪ E𝑌2 . If, say, 𝑣(𝜺) >
𝑣(𝜺0) + 𝛽𝑛 then we make the two sets equitable by moving vertices from 𝑉 (𝜺) to
𝑉 (𝜺0) and changing the edges accordingly. This increases the number of 𝑛4 -edges

55
 3 Maximum Number of Triangle-free Edge Colourings

by at least 𝛽2 𝑛2 /5 which would yield a contradiction to the upper bound given by

Theorem 3.4.3. This establishes property (vi).

56
Packing Degenerate Hypergraphs
4
4.1 Introduction
Recall, in the entirety of Chapter 4 we will consider hypergraphs that are 𝑟-uniform,
that is each edge of a hypergraph 𝐺 is an 𝑟 element subset of 𝑉 (𝐺). We will refer
to (𝑟 − 1)-element vertex sets as semi-edges.
A packing of a family G = {𝐺 1 , . . . , 𝐺 𝑘 } of hypergraphs into a hypergraph 𝐻
is a colouring of the edges of 𝐻 with the colours 0, 1, . . . , 𝑘 such that the edges of
colour 𝑖 form an isomorphic copy of 𝐺 𝑖 for each 1 ≤ 𝑖 ≤ 𝑘. The packing is perfect
if no edges have colour 0. We will often say an edge is covered in a packing if it
has colour at least 1, and uncovered if it has colour zero.
In our work we consider 𝐷-degenerate hypergraphs. First, let us define the term.
An ordering of 𝑉 (𝐺) is 𝐷-degenerate if for each vertex 𝑣, there are at most 𝐷 edges
of 𝐺 whose final vertex is 𝑣. We say 𝐺 is 𝐷-degenerate if 𝑉 (𝐺) has a 𝐷-degenerate
ordering. In particular, we define trees as connected 1-degenerate hypergraphs.

Our results

In this section we restate Theorem 1.3.5 and Theorem 1.3.6 in stronger, more
technical forms. We will use these forms for the remainder of the chapter.
To state our more technical results, we need the following concept of quasiran-
domness.

Definition 4.1.1 (density, quasirandom). The density of a hypergraph 𝐻 is the

number 𝑒(𝐻)/ 𝑣(𝐻) 𝑟 . Suppose that 𝐻 is a hypergraph with 𝑛 vertices and with
(𝐻)

density 𝑝. We say that 𝐻 is (𝛼, 𝐿)-quasirandom if for every set 𝑆 ⊆ 𝑉𝑟−1 of size
at most 𝐿, the number of vertices 𝑣 which form an edge with every member of 𝑆 is
| N𝐻 (𝑆)| = (1 ± 𝛼) 𝑝 |𝑆| 𝑛.
 4 Packing Degenerate Hypergraphs

Our first main technical result then says that we can approximately pack any
sufficiently quasirandom 𝐻 with hypergraphs of bounded degeneracy and not too
large maximum degree. Note that in particular 𝐾𝑛(𝑟) is sufficiently quasirandom.

Theorem 4.1.2. For each 𝑟 ≥ 2, each 𝛾 > 0 and each 𝐷 ∈ N there exist 𝜉, 𝑐 > 0
and a number 𝑛0 such that the following holds for each integer 𝑛 > 𝑛0 . Suppose that
𝐻 is any 𝑛-vertex hypergraph which is (𝜉, 4(𝑟 + 1)𝑟 𝐷 + 3)-quasirandom. Suppose
that (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is a family of 𝐷-degenerate 𝑟-uniform hypergraphs, each of which
𝑐𝑛
has at most 𝑛 vertices and maximum degree at most log 𝑛 . Suppose further that the
total number of edges of (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] is at most 𝑒(𝐻) − 𝛾𝑛𝑟 . Then (𝐺 𝑡 ) 𝑡∈[𝑡 ∗ ] packs
into 𝐻, and furthermore the hypergraph of leftover edges is (𝛾, (𝑟 + 1)𝑟 𝐷 + 3)-
quasirandom.

Our second main technical result is a bit more complicated. If we insist that a
small fraction of the hypergraphs 𝐺 𝑡 each have linearly many vertices of degree 1,
and in addition those graphs are not too close to spanning, then we can upgrade
‘covering almost all the edges’ to a perfect packing. Specifically, we can perfectly
pack a collection of hypergraphs defined as follows.

Definition 4.1.3 ((𝜇, 𝑛)-sequence). We say that a sequence (𝐺 𝑖 )𝑖∈[𝑚] of hyper-

graphs is a 𝐷-degenerate (𝜇, 𝑛)-hypergraph sequence with maximum degree Δ
if
(G 1) 𝐺 𝑖 is 𝐷-degenerate and for each 𝑣 ∈ 𝑉 (𝐺 𝑖 ) we have deg(𝑣) ≤ Δ for each
𝑖 ∈ [𝑚],
(G 2) 𝑣(𝐺 𝑖 ) ≤ 𝑛 for each 1 ≤ 𝑖 ≤ 𝑚 − b𝜇𝑛𝑟−1 c, and
(G 3) 𝑣(𝐺 𝑖 ) ≤ 𝑛 − 𝜇𝑛 and 𝐺 𝑖 has at least 𝜇𝑛 leaves for each 𝑖 with 𝑚 − b𝜇𝑛𝑟−1 c <
𝑖 ≤ 𝑚.
We also call the 𝐺 𝑖 with 𝑚 − b𝜇𝑛𝑟−1 c < 𝑖 ≤ 𝑚 the special hypergraphs of the
sequence.

Theorem 4.1.4. For each 𝑟 ≥ 3, every 𝐷 and 𝜇 > 0 there are 𝑛0 and 𝜉, 𝑐 > 0
such that for every 𝑛 ≥ 𝑛0 and every 𝑚, the following holds. Suppose that 𝐻 is
any 𝑛-vertex hypergraph which is (𝜉, (𝑟 + 1)𝑟 𝐷 + 3)-quasirandom with density at
least 𝜇. Every 𝐷-degenerate (𝜇, 𝑛)-hypergraph sequence (𝐺 𝑖 )𝑖∈[𝑚] with maximum
𝑐𝑛 Í
degree log 𝑛 such that 𝑖∈𝑚 𝑒(𝐺 𝑖 ) ≤ 𝑒(𝐻) packs into 𝐻.

For future use, we will state a technical version of Theorem 4.1.2 in which
we not only pack into a quasirandom host hypergraph and obtain at the end a

58
 4 Packing Degenerate Hypergraphs

quasirandom hypergraph of leftover edges, but also obtain a collection of additional

quasirandomness conditions on the packing as a whole. Although we do consider
this theorem to be a main result of this paper, since it is rather more complicated to
state, we defer it to Section 4.7, where it is Theorem 4.7.1. We note that it is this
technical version of Theorem 4.1.2 which we need to prove Theorem 4.1.4.
We now briefly discuss the optimality of our conditions in Theorems 4.1.2
and 4.1.4. First, we observe that for packing into quasirandom hypergraphs, the
maximum degree bound is optimal up to the value of 𝑐, even if we want only to
embed one single spanning hypergraph. To see this, consider 𝐻 a typical 𝑛-vertex
hypergraph generated by selecting 𝑟-sets as edges independently with probability
1 1
2 , and 𝐺 an 𝑛-vertex hypergraph which is the disjoint union of 10 log 𝑛 stars (that
is, hypergraphs consisting of an (𝑟 − 1)-set centre which forms an edge with each
one of the remaining vertices, and no further edges) of equal size. This 𝐺 is 1-
degenerate and has maximum degree less than 20𝑟 log𝑛 𝑛 , but a standard calculation
(essentially the same as for the 2-uniform graph case) shows that 𝐺 is unlikely to
be contained in 𝐻, even though 𝐻 is very quasirandom.
Our proofs do allow 𝐷 to grow with 𝑛; we believe (but did not check carefully)
the dependency is roughly as log log log 𝑛, but this is presumably not optimal. On
the other hand, 𝐷 cannot be as big as 10 log 𝑛, since a typical random hypergraph
is unlikely to contain any given hypergraph with 9𝑛 log 𝑛 edges.
We cannot allow all hypergraphs to be spanning in Theorem 4.1.4. For example,
if 𝐻 is obtained by choosing 𝑟-sets as edges independently with probability either 12
or 21 + 𝜀, depending on whether the 𝑟-set contains one of the first 𝑛/2 vertices or not,
then 𝐻 will likely be (𝜉, 𝑄)-quasirandom for any given 𝜉 > 0 and natural number
𝑄, provided 𝜀 > 0 is sufficiently small. However, some Θ(𝑛) vertices will be in
Θ(𝑛𝑟−1 ) more edges than others, so that we certainly cannot pack any collection of
regular spanning hypergraphs. In fact, if for example each 𝐺 𝑠 is a matching, then
we need that Θ(𝑛𝑟−1 ) of the matchings are Θ(𝑛) vertices away from being spanning
to correct the imbalance in degrees. This shows that the restriction on the size of
the hypergraphs in (G 3) is optimal up to the choice of constants. However, we do
not believe that it is necessary to have many hypergraphs with many leaves. We
should note that one cannot simply omit this condition, because for example for
𝑟 = 2 no collection of cycles can perfectly pack 𝐾2𝑛 , due to a parity obstruction:
cycles use an even number of edges at each vertex, but 𝐾2𝑛 has odd degree vertices.
However for the case 𝐷 = 1 (i.e. forests) we believe one can omit the condition

59
 4 Packing Degenerate Hypergraphs

entirely (as the leaves should allow for parity correction). This is demonstrated to
be true in [1] for graphs.

Proof outline

To pack a collection of guest hypergraphs G into a host hypergraph 𝐻, in which

each guest hypergraph has order at most (1 − 𝛿)𝑛, we do the following. We take
hypergraphs in G in succession. For each 𝐺, we embed vertex by vertex into 𝐾𝑛(𝑟)
in a degeneracy order, at each time embedding to a vertex of 𝐾𝑛(𝑟) chosen uniformly
at random subject to the constraints that we do not re-use a vertex previously used
in embedding 𝐺, or an edge used in embedding a previous hypergraph. This
procedure succeeds with high probability, and after each stage of embedding a
hypergraph, the unused edges in 𝐾𝑛(𝑟) are quasirandom (in a sense we will later
make precise). The analysis of this process is what gives the quasirandomness
statements of Theorem 4.7.1.
To allow for spanning hypergraphs, we modify this slightly. We adjust the
degeneracy order so that (for some small 𝛿 > 0) the last 𝛿𝑛 vertices are strongly
independent (that is, no pair of them is contained in an edge of 𝐺) and all have
the same degree. This can be done while at worst multiplying the degeneracy of
the order by 𝑟. Then for each hypergraph we follow the above procedure to embed
the first (1 − 𝛿)𝑛 vertices, and finally complete the embedding arbitrarily using a
matching argument. We will see that this last step is with high probability always
possible. The only point to be careful about is that we have to split 𝐸 (𝐾𝑛(𝑟) ) into a
very dense bulk main part, whose edges we use only for the embedding of the first
(1 − 𝛿)𝑛 vertices, and a sparse reservoir which we use only for the completion; we
do this randomly. This is the hypergraph analogue of the main result of [3].
To obtain a perfect packing, we perform a different modification. We use the
above procedure to pack all the non-special guest graphs of our (𝜇, 𝑛)-hypergraph
sequence, and observe that the resulting leftover edges form a quasirandom graph
𝐻. We then need to pack the special graphs. We first remove from each special
graph a strongly independent set of vertices of degree 1: the omitted leaves. We
pack the resulting collection of graphs into 𝐻, using the randomised algorithm
described above. What now remains is, for each guest graph, to pack in addition
the omitted leaves. By definition, each omitted leaf is in exactly one edge of
the guest graph, together with its parent semi-edge, and all the vertices of the
parent semi-edge have been packed. This reduces our remaining packing to a

60
 4 Packing Degenerate Hypergraphs

matching problem: each semi-edge 𝑢 of 𝐾𝑛(𝑟) has some dangling leaves, namely the
collection of omitted leaves (over all special guest graphs) whose parent semi-edge
is embedded to 𝑢. We need to match the dangling leaves at 𝑢 to the vertices which
form leftover edges of 𝐾𝑛(𝑟) with 𝑢, such that we do not use any edge of 𝐻 twice
and such that we do not match dangling leaves of any graph 𝐺 𝑖 with vertices which
were used for embedding the rest of 𝐺 𝑖 , or for another dangling leaf of 𝐺 𝑖 .
It is relatively easy to ensure that we do not use leftover edges multiple times.
We orient the edges of 𝐻 by marking one vertex in each edge, and we insist on
matching dangling leaves at 𝑢 ∈ 𝑆(𝐻) to an edge containing 𝑢 where none of the
vertices in 𝑢 are marked. This reduces the problem to ensuring that we pick distinct
edges of 𝐻 to embed the dangling leaves at 𝑢. Note, that matching a dangling
leave to an edge means embedding the dangling leaf (which is a vertex in one
of the graphs we want to pack) to the marked vertex of the edge we matched to.
Therefore, we need to ensure that nothing is embedded to that vertex in the host
graph, from the relevant guest graph. In order for this to work, we of course need
that the number of dangling leaves at 𝑢 is equal to the number of edges containing
𝑢 whose marked vertex is not in 𝑢. We can obtain this by performing the marking
uniformly at random and then correcting the resulting small errors.
𝑛

We then go through all the 𝑟−1 semi-edges of 𝐻 in turn, and at each semi-edge
𝑢 embed all its dangling leaves. We do this as follows. We draw a bipartite leaf
matching graph whose parts are the dangling leaves at 𝑢 and the marked vertices
of edges containing 𝑢, putting an edge from a dangling leaf of a guest graph 𝐺 𝑖 to a
marked vertex if the marked vertex has not previously been used in the embedding
of 𝐺 𝑖 . We choose uniformly at random a perfect matching in the leaf matching
graph, and embed the dangling leaves accordingly. The difficulty with this — and
the reason we need to pick a matching uniformly at random — is that when we
embed dangling leaves at one semi-edge, we need to remove a few edges from the
leaf matching graphs of other semi-edges. We will argue that the leaf matching
graphs begin with a strong quasirandom property and are only changed slightly
through the entire process, so that they are still quasirandom at the end; this is
enough to guarantee the desired matchings always exist and hence complete a
perfect packing. This second part is the hypergraph analogue of the main result
of [2].

61
 4 Packing Degenerate Hypergraphs

Organisation of the chapter

In the following Section 4.2, we introduce notation, the probabilistic tools we need,
and state and prove a result on random matchings in bipartite graphs.
In Section 4.3, we reduce Theorem 4.1.2 to a technical statement, Theorem 4.3.1
which is convenient for our proof, formally state the algorithm sketched above
which proves it (i.e. which produces a packing), state the main lemmas which we
need to prove it, and finally give the proof of Theorem 4.3.1 from these lemmas.
In Section 4.4, we analyse the embedding of the first (1 − 𝛿)𝑛 vertices of a single
guest hypergraph into (what remains of) the bulk of the host hypergraph, and argue
that our randomised embedding is likely to succeed and maintain various desirable
properties. We build on this in Section 4.5 to analyse the sequential embedding
of the first (1 − 𝛿)𝑛 vertices of all guest hypergraphs into the bulk. Finally in
Section 4.6 we argue that it is always possible to complete the embedding in the
reservoir, which finishes the proof of Theorem 4.1.2.
In Section 4.7 we state and prove Theorem 4.7.1, the above mentioned technical
version of Theorem 4.1.2 which also gives various quasirandomness properties
of the packing. Most of the work here is done in the preceding sections, so this
section mainly consists of using established probabilistic estimates to prove the
desired quasirandomness properties are likely to hold.
In Section 4.8 we prove Theorem 4.1.4. Again, most of the work is done by
Theorems 4.1.2 and 4.7.1, and what remains to do is formalise the above sketched
packing algorithm and prove that it works with high probability.

4.2 Preliminaries
4.2.1 Notation

We will almost always work with 𝑟-uniform hypergraphs; when we write ‘hyper-
graph’ without specifying the uniformity, it should be understood that the hyper-
graph considered is 𝑟-uniform.
We refer to vertex sets of size 𝑟 − 1 as semi-edges. Let 𝑆(𝐺) denote the set
(𝐺)

of semi-edges of a hypergraph 𝐺, that is 𝑆(𝐺) = 𝑉𝑟−1 , the set of unordered
𝑟 − 1-subsets of the vertex set. Given a vertex set 𝐴 ⊆ 𝑉 (𝐺), we write N𝐺 ( 𝐴) for
the neighbourhood of 𝐴, that is, the collection of subsets 𝐵 of 𝑉 (𝐺) \ 𝐴 such that
𝐴 ∪ 𝐵 ∈ 𝐸 (𝐺). We will often write N𝐺 (𝑣) for the set N𝐺 ({𝑣}) of semi-edges. Note
that the neighbourhood of 𝐴 in a hypergraph is a (𝑟 − | 𝐴|)-uniform hypergraph on

62
 4 Packing Degenerate Hypergraphs

𝑉 (𝐺) \ 𝐴. However, when 𝐴 is a semi-edge, we will usually think of N𝐺 ( 𝐴) as

a set of vertices in 𝑉 (𝐺) \ 𝐴 (even though formally it is a set of singleton sets of
vertices). Finally, if 𝑆 is any subset of 𝑉 (𝐺)


𝑘 , we write N𝐺 (𝑆) for the intersection
Ñ
𝐴∈𝑆 N𝐺 ( 𝐴).
We call a set of vertices 𝐴 ⊆ 𝑉 (𝐺) strongly independent if for any pair of
different vertices 𝑢, 𝑣 ∈ 𝐴 we have N𝐺 ({𝑢, 𝑣}) = ∅. For any 𝑘 ≤ 𝑟 − 1 we call a
hypergraph 𝐺 a 𝑘-star with centre 𝐴 if | 𝐴| = 𝑘 and for each 𝑒 ∈ 𝐸 (𝐺) we have
𝐴 ⊆ 𝑒. The leaves of the star 𝐺 are the vertices {𝑣 ∈ 𝑉 (𝐺) : {𝑣} ∪ 𝐴 ∈ 𝐸 (𝐺)}.
For any given natural number ℓ, we write [ℓ] := {1, 2, . . . , ℓ}. The definition
of degenerate graphs naturally suggests to label the vertices of a hypergraph by
integers. Suppose that the vertex set of a hypergraph 𝐺 is 𝑉 (𝐺) = [ℓ] and 𝑖 ∈ 𝑉 (𝐺).
We write N− (𝑖) = N (𝑖) ∩ [𝑖−1] − −


𝑟−1 and deg (𝑖) = | N (𝑖)| for the left-neighbourhood
and the left-degree of 𝑖. We make use of the natural order on [ℓ] also in other
ways, like referring to sets of the form [ℓ1 ] ⊆ 𝑉 (𝐺) and {ℓ2 , ℓ2 + 1, . . . , ℓ} ⊆ 𝑉 (𝐺)
as initial vertices and final vertices, respectively. We write 𝑅(𝑣) := {𝑢 : ∃𝑦 ∈
N𝐺− (𝑣) with 𝑢 = max(𝑦)}. That is, 𝑅(𝑣) denotes the set of vertices which are the

second-to-last vertices in hyperedges of 𝐺 whose last vertex is 𝑣.

The hypergraphs to be packed are denoted 𝐺 𝑠 and referred to as guest hy-
pergraphs. By contrast, during our packing procedure, we shall work with host
hypergraphs 𝐻𝑠 which are obtained from the original 𝐾𝑛(𝑟) by removing what was
used previously.
An orientation of a hypergraph 𝐻 = (𝑉, 𝐸) is an oriented hypergraph 𝐻® on 𝑉
which contains, for each undirected edge 𝑒 ∈ 𝐸, exactly one directed edge (𝑣, 𝑒)
where 𝑣 is a vertex in 𝑒. We say the edge 𝑒 is directed towards 𝑣. The outdegree
deg+® (𝑠) of a semi-edge 𝑠 in an oriented hypergraph 𝐻® is the number of vertices 𝑣
𝐻
® such that 𝑣, 𝑠 ∪ {𝑣} is an edge of 𝐻; ® the set of these vertices 𝑣 is the


in 𝑉 ( 𝐻)
outneighbourhood 𝑁 +® (𝑠) of 𝑠. The indegree deg−® (𝑣) of a vertex 𝑣 in an oriented
𝐻 𝐻
hypergraph 𝐻® is the number of 𝑠 in 𝑆(𝐻) such that 𝑣, 𝑠 ∪ {𝑣} is an edge of 𝐻; ®

the set of these hyperedges 𝑠 is the inneighbourhood 𝑁 −® (𝑣) of 𝑣.

𝐻
We write 0 < 𝑎  𝑏 to mean that given 𝑏 > 0, any sufficiently small 𝑎 will
make our calculations work. In other words, there is a monotone increasing function
𝑓 : R>0 → R>0 with 𝑓 (𝑏) ≤ 𝑏 such that any choice 0 < 𝑎 ≤ 𝑓 (𝑏) will work. We
define longer strings similarly; thus 0  𝑎  𝑏  𝑐 means that 0  𝑏  𝑐 and
in addition 0  𝑎  𝑏, where the two monotone functions need not be the same.
We can always find constants satisfying such a sequence by choosing from the

63
 4 Packing Degenerate Hypergraphs

right-hand side: in the example above we would choose first 𝑐, then a sufficiently
small 𝑏, and given that a sufficiently small 𝑎.

4.2.2 Probabilistic tools

In this subsection we give the notation and tools we need to analyse our randomised
processes. These are taken directly from [2, Section 2.2] and we refer the reader
there for exposition and proofs.
Let Ω be a finite probability space. A filtration F0 , F1 , . . . , F𝑛 is a sequence
of partitions of Ω such that F𝑖 refines F𝑖−1 for all 𝑖 ∈ [𝑛]. In our application, the
partition F𝑖 is given by all possible histories of the run of one of our algorithms up
to time 𝑖. (For more explanation see [3].) We say that a function 𝑓 : Ω → R is
F𝑖 -measurable if 𝑓 is constant on each part of F𝑖 . Further, for any random variable
𝑌 : Ω → R the conditional expectation E(𝑌 |F𝑖 ) : Ω → R and the conditional
variance Var(𝑌 |F𝑖 ) : Ω → R of 𝑌 with respect to F𝑖 are defined by

E(𝑌 |F𝑖 )(𝑥) = E(𝑌 |𝑋),

where 𝑋 ∈ F𝑖 is such that 𝑥 ∈ 𝑋 .
Var(𝑌 |F𝑖 )(𝑥) = Var(𝑌 |𝑋),

Suppose that we have an algorithm which proceeds in 𝑚 rounds using a new

source of randomness Ω𝑖 in each round 𝑖. Then the probability space underlying
Î𝑚
the run of the algorithm is 𝑖=1 Ω𝑖 . By history up to time 𝑡 we mean a set of the
form {𝜔1 } × · · · × {𝜔𝑡 } × Ω𝑡+1 × · · · Ω𝑚 , where 𝜔𝑖 ∈ Ω𝑖 . We shall use the symbol
H𝑡 to denote any particular history of such a form. By a history ensemble up to
time 𝑡 we mean any union of histories up to time 𝑡; we shall use the symbol L to
denote any one such. Observe that there are natural filtrations associated to such
a probability space: given times 𝑡1 < 𝑡2 < . . . we let F𝑡𝑖 denote the partition of Ω
into the histories up to time 𝑡𝑖 .
Our first tool is the well known Chernoff bound about the sum of independent
Bernoulli variables.

Theorem 4.2.1 (Chernoff bounds, [30, Theorem 2.10]). Suppose 𝑋 is a random

variable which is the sum of a collection of independent Bernoulli random variables.
Then we have for 𝛿 ∈ (0, 3/2)

2 2
P 𝑋 > (1 + 𝛿)E𝑋 < 𝑒 −𝛿 E𝑋/3 P 𝑋 < (1 − 𝛿)E𝑋 < 𝑒 −𝛿 E𝑋/3 .
   
and

When analysing randomised algorithms, usually one has to deal with a sum of

64
 4 Packing Degenerate Hypergraphs

random variables which are not independent, but rather are sequentially dependent.
When we use this term, what we mean is that there is some filtration F0 , F1 , . . . ,
F𝑛 associated to the randomised algorithm such that the random variable 𝑋𝑖 is
F𝑖 -measurable for each 𝑖. The filtration in question will generally be a filtration
defined by histories up to a sequence of increasing times in the algorithm; when it is
clear from the context we will simply say that the random variables are sequentially
dependent without specifying the filtration explicitly. In this chapter, we need the
following concentration inequality for sequentially dependent random variables,
which is a corollary of Freedman’s inequality [24] deduced in [2].

Corollary 4.2.2 (Corollary 6, [2]). Let Ω be a finite probability space, and

(F0 , F1 , . . . , F𝑛 ) be a filtration. Suppose that we have 𝑅 > 0, and for each
1 ≤ 𝑖 ≤ 𝑛 we have an F𝑖 -measurable non-negative random variable 𝑌𝑖 , nonnegat-
ive real numbers 𝜇, ˜ 𝜈˜ and an event E.
Í𝑛  
(a ) Suppose that either E does not occur or we have 𝑖=1 E 𝑌𝑖 F𝑖−1 ≤ 𝜇,
˜ and
0 ≤ 𝑌𝑖 ≤ 𝑅 for each 1 ≤ 𝑖 ≤ 𝑛. Then
" 𝑛
#
Õ  𝜇˜ 
P E and 𝑌𝑖 > 2 𝜇˜ ≤ exp − .
𝑖=1
4𝑅

Í𝑛  
(b ) Suppose that either E does not occur or we have 𝑖=1 E 𝑌𝑖 F𝑖−1 = 𝜇˜ ± 𝜈,
˜
and 0 ≤ 𝑌𝑖 ≤ 𝑅 for each 1 ≤ 𝑖 ≤ 𝑛. Then for each 𝜚˜ > 0 we have
" 𝑛
#
Õ  𝜚˜ 2 
P E and 𝑌𝑖 ≠ 𝜇˜ ± ( 𝜈˜ + 𝜚)
˜ ≤ 2 exp − .
𝑖=1
2𝑅( 𝜇˜ + 𝜈˜ + 𝜚)
˜

In particular, if 𝜈˜ = 𝜚˜ = 𝜇˜ 𝜂˜ > 0 and 𝜂˜ ≤ 12 , then

" 𝑛
#
Õ 𝜇˜ 𝜂˜2  
P E and ˜ ± 2𝜂)
𝑌𝑖 ≠ 𝜇(1 ˜ ≤ 2 exp − .
𝑖=1
4𝑅

Finally, let us note that we shall be using many statements of the form

with probability at least 𝑝, provided event A we get event B. (4.2.1)

We emphasize that such statements are not statements about conditional probabil-
ities. That is, the meaning of (4.2.1) is P[A \ B] ≤ 1 − 𝑝. A prototypical example
is with probability at least 1 − 𝑜(1), if a given randomized algorithm does not fail,

65
 4 Packing Degenerate Hypergraphs

then it produces an output with certain desired properties.

4.2.3 Degenerate hypergraphs

deg(𝑥) 2 for degenerate hypergraphs 𝐺.

Í
We need to bound 𝑥∈𝑉 (𝐺)

Lemma 4.2.3. Let 𝐺 be an 𝑛-vertex hypergraph with degeneracy 𝐷 and maximum

degree Δ. Then we have
Õ
deg(𝑥) 2 ≤ 𝑟 𝐷𝑛Δ .
𝑥∈𝑉 (𝐺)

Proof. Suppose that for each 1 ≤ 𝑖 ≤ Δ there are 𝑤 𝑖 vertices in 𝐺 of degree 𝑖. We

ÍΔ Í
have 𝑖=1 𝑖𝑤 𝑖 = 𝑥∈𝑉 (𝐺) deg(𝑥) = 𝑟𝑒(𝐺) ≤ 𝑟 𝐷𝑛. By convexity of the function
ÍΔ
𝑧 ↦→ 𝑧2 , among all choices of non-negative reals 𝑤 𝑖 such that 𝑖=1 𝑖𝑤 𝑖 ≤ 2𝐷𝑛,
ÍΔ 2
the one maximising 𝑖=1 𝑖 𝑤 𝑖 is 𝑤 1 = · · · = 𝑤 Δ−1 = 0 and 𝑤 Δ = 𝑟 𝐷𝑛/Δ. The
maximum attained is 𝑟 𝐷𝑛Δ, proving the lemma.

We also need to show that degenerate hypergraphs contain large strongly inde-
pendent sets all of whose vertices have the same degree.

Lemma 4.2.4. Let 𝐺 be a 𝐷-degenerate 𝑛-vertex hypergraph. Then there exists

an integer 0 ≤ 𝑑 ≤ 𝑟 𝐷 and a set 𝐼 ⊆ 𝑉 (𝐺) with |𝐼 | ≥ (𝑟 𝐷 + 1) −4 𝑛 which is
independent, and all of whose vertices have the same degree 𝑑 in 𝐺.

Proof. We first claim that at least (𝑟 𝐷 +1) −1 𝑛 vertices of 𝐺 have degree at most 𝑟 𝐷.
Indeed, if this were false then there would be more than 𝑟 𝐷𝑛/(𝑟 𝐷 + 1) vertices
of 𝐺 all of whose degrees are at least 𝑟 𝐷 + 1, so that we obtain 𝑒(𝐺) > 𝐷𝑛,
which contradicts 𝐷-degeneracy of 𝐺. Let 0 ≤ 𝑑 ≤ 𝑟 𝐷 be chosen to maximise
the number of vertices in 𝐺 of degree 𝑑, and let 𝑆 be the set of vertices in 𝐺 with
degree 𝑑. We thus have |𝑆| ≥ (𝑟 𝐷 + 1) −2 𝑛. Now let 𝐼 be a maximal independent
subset of 𝑆. Each vertex of 𝐼 has at most 𝑑 ≤ 𝑟 𝐷 incident edges, so that the number
of vertices of 𝐺 which are either in 𝐼 or in an edge containing a vertex of 𝐼 is at
most (𝑟 − 1)𝑟 𝐷 |𝐼 | + |𝐼 | ≤ (𝑟 𝐷 + 1) 2 |𝐼 |. By maximality this set of vertices covers
𝑆, hence |𝐼 | ≥ (𝑟 𝐷 + 1) −2 |𝑆| ≥ (𝑟 𝐷 + 1) −4 𝑛, as desired.

The following auxiliary lemma takes an arbitrary family of graphs we want to

pack and produces a family with not too many members and the same bound on
maximum degree and degeneracy.

66
 4 Packing Degenerate Hypergraphs

Lemma 4.2.5 (compression lemma). Let G = (𝐺 𝑖 )𝑖∈[𝑠] be a family of 𝐷-degenerate

𝑒(𝐺 𝑖 ) ≤ 𝑛𝑟 and 𝑣(𝐺 𝑖 ) ≤ 𝑛
Í𝑠
hypergraphs with maximum degree at most Δ, with 𝑖=1
2𝑛𝑟 −1
for all 𝑖 ∈ [𝑠]. Then there is a family of hypergraphs G 0 = ( 𝐺ˇ 𝑖 )𝑖∈[𝑠∗ ] with 𝑠∗ ≤ (𝑟−1)!
Í𝑠∗
𝑒( 𝐺ˇ 𝑖 ) ≤ 𝑛𝑟 , such that for each 𝑖 ∈ [𝑠∗ ] we have 𝑣( 𝐺ˇ 𝑖 ) ≤ 𝑛,


such that 𝑖=1
Δ( 𝐺ˇ 𝑖 ) ≤ Δ, and 𝐺ˇ 𝑖 is 𝐷-degenerate, and such that G perfectly packs into G 0.
Proof. We successively modify the family G as follows. If there are two hyper-
graphs 𝐺, 𝐺 0 ∈ G with 𝑣(𝐺), 𝑣(𝐺 0) ≤ 𝑛/2, we replace 𝐺 and 𝐺 0 with the disjoint
union 𝐺 ∪ 𝐺 0. We repeat this until no further such pairs exist, giving G 0.
Observe that the maximum degree and the degeneracy of the hypergraphs in G
is the same as in G 0. Furthermore G 0 is a packing of G. Finally, there is at most one
hypergraph in G 0 with less than 𝑛/2 vertices. Hence all but at most one hypergraph
has at least 𝑛/(2𝑟) edges. We conclude that the total number 𝑠∗ of hypergraphs in
G 0 satisfies (𝑠∗ − 1)𝑛/(2𝑟) ≤ (1 − 𝛾) 𝑛𝑟 , and hence 𝑠∗ ≤ 2𝑛𝑟−1 /(𝑟 − 1)!. Finally,


𝑠∗ be obtained from the hypergraphs G 0 by adding if
we let the hypergraphs (𝐺 0𝑠 ) 𝑠=1
necessary isolated vertices to each in order to obtain 𝑛-vertex hypergraphs.

4.2.4 Random matchings in bipartite graphs

There are two places in this chapter where we reduce ‘completing’ embeddings to
a perfect matching problem in bipartite graphs. In both cases, we can show that the
bipartite graph in question is suitably quasirandom, and this not only implies that
a perfect matching exists, it also implies that a uniform random perfect matching
is about equally likely to use any given edge (which, in both cases, we need). The
following statement, which is rather similar to [2, Lemma 20], formalises this. The
proof is very similar to that given in [2], but we provide it for completeness.
Lemma 4.2.6 (matching lemma). Assume 0  𝑚1  𝜀, 𝜚  𝜇  1. Let 𝐹 =
𝐹 [𝑈, 𝑊] be a bipartite graph with |𝑈| = |𝑊 | = (1 ± 𝜀)𝑚 such that
(M 1) deg𝐹 (𝑥) = (1 ± 𝜀)𝜇𝑚 for all 𝑥 ∈ 𝑈 ∪ 𝑊, and
𝑚2
(M 2) deg𝐹 (𝑢, 𝑢0) = (1 ± 𝜀)𝜇2 𝑚 for all but at most log 𝑚 pairs {𝑢, 𝑢0 } ∈ 𝑈

2 ,
and let 𝐹 0 = 𝐹 0 [𝑈, 𝑊] be a spanning subgraph of 𝐹 [𝑈, 𝑊] such that
(M 3) deg𝐹 (𝑥) − deg𝐹 0 (𝑥) < 𝜚𝑚 for all 𝑥 ∈ 𝑈 ∪ 𝑊.
Then 𝐹 0 has a perfect matching and for a perfect matching 𝜎 chosen uniformly at
random among all perfect matchings in 𝐹 0 and for all 𝑢𝑤 ∈ 𝐸 (𝐹 0) we have

2
P[𝜎(𝑢) = 𝑤] ≤ .
𝜇𝑚

67
 4 Packing Degenerate Hypergraphs

This lemma is a straightforward consequence of a lemma (Lemma 4.2.9) on

random matchings in super-regular pairs by Felix Joos (see [37]) and the degree-
codegree characterisation of super-regular pairs (Lemma 4.2.8) provided by Duke,
Lefmann, and Rödl in [16]. The proof of this lemma is the only place in this chapter
where we use the concept of a regular pair.

Definition 4.2.7 (density, (𝜀, 𝑑)-regular, (𝜀, 𝑑)-super-regular). Let 𝐺 be a graph

and 𝑈, 𝑊 ⊆ 𝑉 (𝐺) be disjoint vertex sets. The density of the bipartite graph
𝐺 [𝑈, 𝑊] is
𝑒(𝐺 [𝑈, 𝑊])
𝑑𝐺 (𝑈, 𝑊) = .
|𝑈||𝑊 |
We say that 𝐺 [𝑈, 𝑊] is (𝜀, 𝑑)-regular if for all 𝑈 0 ⊆ 𝑈 and 𝑊 0 ⊆ 𝑊 with
|𝑈 0 | ≥ 𝜀|𝑈| and |𝑊 0 | ≥ 𝜀|𝑊 | we have

𝑑𝐺 (𝑈 0, 𝑊 0) = 𝑑 ± 𝜀 .

The graph 𝐺 [𝑈, 𝑊] is (𝜀, 𝑑)-super-regular if it is (𝜀, 𝑑)-regular and for all 𝑢 ∈ 𝑈
and for all 𝑤 ∈ 𝑊 we have

deg𝐺 [𝑈,𝑊] (𝑢) = (𝑑 ± 𝜀)|𝑊 |, and deg𝐺 [𝑈,𝑊] (𝑤) = (𝑑 ± 𝜀)|𝑈| .

It is well-known that regular pairs are forced by a degree-codegree condition;

we use the following formulation due to Duke, Lefmann, and Rödl in [16].

Lemma 4.2.8 (degree-codegree condition [16]). Assume 0 < 𝜀0 < 2−200 and let
𝐺 [𝑈, 𝑊] be a bipartite graph with parts 𝑈 and 𝑊 of size |𝑈| = |𝑊 | = 𝑛 and density
𝑑 = 𝑑𝐺 [𝑈,𝑊] (𝑈, 𝑊). If
(i ) deg𝐺 [𝑈,𝑊] (𝑢) > (𝑑 − 𝜀0)|𝑊 | for all 𝑢 ∈ 𝑈, and
(ii ) deg𝐺 [𝑈,𝑊] (𝑢, 𝑢0) < (𝑑 + 𝜀0) 2 |𝑊 | for all but at most 2𝜀0 |𝑈| 2 pairs {𝑢, 𝑢0 } ∈
𝑈

2 ,
1
then 𝐺 [𝑈, 𝑊] is (𝜀0) 6 , 𝑑 -regular.

If we choose a perfect matching uniformly at random in a super-regular pair then

each edge is roughly equally likely to appear in the matching, as was shown by Joos
(see [37]).

Lemma 4.2.9 (perfect matchings in super-regular pairs [37, Theorem 4.3]). Assume
0  𝑚10  𝜀0  𝑑  1. Let 𝐺 [𝑈, 𝑊] be an (𝜀00, 𝑑)-super-regular graph with

68
 4 Packing Degenerate Hypergraphs

|𝑈| = |𝑊 | = 𝑚0. Then 𝐺 [𝑈, 𝑊] contains a perfect matching. Moreover, for a

perfect matching 𝜎 chosen uniformly at random among all perfect matchings in
𝐺 [𝑈, 𝑊] and for all 𝑢𝑤 ∈ 𝐸 (𝐺) we have

1 1
P[𝜎(𝑢) = 𝑤] = (1 ± (𝜀00) 20 ) .
𝑑𝑚0

The proof of the matching lemma simply combines these two lemmas.

Proof of Lemma 4.2.6. By (M 1) and (M 3), for all 𝑥 ∈ 𝑈 ∪ 𝑊 we have

deg𝐹 0 (𝑥) = 𝜇 ± ( 𝜚 + 𝜀) 𝑚 (4.2.2)

𝑚2
By (M 2) and (M 3), for all but at most log 𝑚 pairs {𝑢, 𝑢0 } ∈ 𝑈

2 we have

deg𝐹 0 (𝑢, 𝑢0) = 𝜇2 ± (2𝜚 + 𝜀) 𝑚 .

(4.2.3)

We want to apply Lemma 4.2.8 with 𝑑 = 𝑑 𝐹 0 (𝑈, 𝑊) = 𝜇 ± ( 𝜚 + 𝜀) and 𝜀0 = 5( 𝜚 +

𝜀)𝜇−1 to conclude that 𝐹 0 [𝑈, 𝑊] is super-regular, and now check the conditions of
this lemma. By (4.2.2), for 𝑢 ∈ 𝑈 we have

|𝑊 |
= 𝑑±4( 𝜚+𝜀) |𝑊 | > (𝑑−𝜀0)|𝑊 | ,



deg𝐹 0 (𝑢) = 𝑑±2( 𝜚+𝜀) 𝑚 = 𝑑±2( 𝜚+𝜀)
1±𝜀

and similarly for 𝑤 ∈ 𝑊 we have deg𝐹 0 (𝑤) = 𝑑 ± 4( 𝜚 + 𝜀) |𝑈| > (𝑑 − 𝜀0)|𝑈|.

𝑚2 0 𝑈

By (4.2.3), for all but at most log 𝑚 pairs {𝑢, 𝑢 } ∈ 2 we have

|𝑊 |
deg𝐹 0 (𝑢, 𝑢0) ≤ 𝑑 2 + (2𝜚 + 𝜀) 𝑚 ≤ 𝑑 2 + (2𝜚 + 𝜀) ≤ 𝑑 2 + 4( 𝜚 + 𝜀) |𝑊 |



1−𝜀
2 0 0 2
0 2
< 𝑑 + 2𝜀 𝑑 + (𝜀 ) |𝑊 | = (𝑑 + 𝜀 ) |𝑊 | ,

where the last inequality uses 𝑑 = 𝜇 ± ( 𝜚 + 𝜀) and 𝜀0 = 5( 𝜚 + 𝜀)𝜇−1 . We conclude

𝑚2 0 |𝑈| 2 , which holds for log 𝑚 > 10/𝜀0, then 𝐹 0 is (𝜀0) 16 , 𝑑 -regular


that, if log 𝑚 ≤ 2𝜀


by Lemma 4.2.8. Since deg𝐹 0 (𝑥) = 𝑑 ± 2( 𝜚 + 𝜀) |𝑈| for all 𝑥 ∈ 𝑈 ∪ 𝑊, it follows
1
that 𝐹 0 is (𝜀0) 6 , 𝑑 -super-regular.

Hence we can apply Lemma 4.2.9 to 𝐹 0 with

61
𝑚0 = |𝑈| = (1 ± 𝑝)𝑚 , 𝜀00 = 𝜀0 , and 𝑑 = 𝜇 ± ( 𝜚 + 𝜀) ,

and conclude that 𝐹 0 has a perfect matching and that for a perfect matching 𝜎 chosen

69
 4 Packing Degenerate Hypergraphs

uniformly at random among all perfect matchings of 𝐹 0 and for all 𝑢𝑤 ∈ 𝐸 (𝐹 0) we

have 
1  1 2
P[𝜎(𝑢) = 𝑤] = 1 ± 𝜀0 120 ≤ ,
𝑑 (1 ± 𝜀)𝑚 𝜇𝑚
as required.

4.3 Almost perfect hypergraph packing

In this section we give a randomised algorithm which almost perfectly packs
almost spanning hypergraphs, and a modification which allows for almost perfect
packing of spanning hypergraphs. We state the main lemmas which show that the
randomised algorithm is likely to succeed, and assuming them prove Theorem 4.1.2.

4.3.1 Reducing the first main theorem

We deduce Theorem 4.1.2 from the following technical result.

Theorem 4.3.1 (Approximate packing technical result). For each 𝑟 ≥ 2, 𝛾 > 0 and
each 𝐷 ∈ N there exist numbers 𝑛0 ∈ N and 𝑐, 𝜉 > 0 such that the following holds
for each 𝑛 > 𝑛0 . Suppose that 𝐻 b is an (𝜉, (𝑟 + 1)𝐷 + 3)-quasirandom hypergraph
with 𝑛 vertices and density 𝑝 > 0. Suppose that 𝑠∗ ≤ 2𝑛𝑟−1 /(𝑟 − 1)!. Suppose
that for each 𝑠 ∈ [𝑠∗ ] the hypergraph 𝐺 𝑠 is a hypergraph on vertex set [𝑛], with
𝑐𝑛 −
maximum degree at most log 𝑛 , such that deg (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺 𝑠 ) and
such that the last (𝐷 + 1) −4 𝑛 vertices of [𝑛] form a strongly independent set in 𝐺 𝑠 ,
and all have the same degree 𝑑 𝑠 in 𝐺 𝑠 .


Suppose further that the total number of edges of (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] is at most ( 𝑝−3𝛾) 𝑛𝑟 .
Then (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] packs into 𝐻.
b In addition, the hypergraph of leftover edges is
(𝛾, (𝑟 + 1)𝐷 + 3)-quasirandom.

We briefly explain how to deduce Theorem 4.1.2 from this result.

Proof of Theorem 4.1.2. Given (𝐺 𝑡 )𝑡∈𝑡 ∗ to pack, we create a sequence (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ]

by applying Lemma 4.2.5 to the graphs (𝐺 𝑡 )𝑡∈[𝑡 ∗ ] . By Lemma 4.2.5, we have the
required 𝑠∗ ≤ 2𝑛𝑟−1 /(𝑟 − 1)!.
Now, for each 𝐺 0𝑠 with 𝑠 ≤ 𝑠∗ we choose an order on 𝑉 (𝐺 0𝑠 ) as follows. First,
we pick an order witnessing 𝐷-degeneracy of 𝐺 0𝑠 . Next, we pick an integer
0 ≤ 𝑑 𝑠 ≤ 2𝐷 and a strongly independent 𝐼 𝑠 set of (𝑟 𝐷 + 1) −4 𝑛 vertices each
of which has degree 𝑑 𝑠 in 𝐺 0𝑠 and change the order by moving these vertices to

70
 4 Packing Degenerate Hypergraphs

the end. Such an integer 𝑑 𝑠 and strongly independent set exist by Lemma 4.2.4.
Observe that moving vertices to the end of the order cannot increase the left-degree
of vertices which are not moved, which therefore have at most 𝐷 left-neighbours
in the new order. The moved vertices have degree at most 𝑟 𝐷, all of which are
left-neighbours. The result is an ordering of 𝑉 (𝐺 0𝑠 ) with degeneracy at most 𝑟 𝐷,
as required for Theorem 4.3.1 with input 𝛾/3 and 𝑟 𝐷. Then Theorem 4.3.1 returns
the desired packing.

Proof of Theorem 4.3.1

For the proof of Theorem 4.3.1, we need some algorithms and definitions. We
give these now along with a sketch of the proof. At this level, the proof is very
similar to that in [3]. We will do a bit of extra analysis (as compared to the proof
in [3]) in order to obtain additional properties of the packing (in order to later prove
Theorem 4.7.1), and the analysis itself will need some modifications to deal with
hypergraphs.
We prove Theorem 4.3.1 by analysing a randomised algorithm, which we call
PackingProcess, that packs the guest hypergraphs 𝐺 𝑠 into 𝐻. b We prove that this
algorithm succeeds with high probability. In this algorithm we assume that the
last 𝛿𝑛 vertices of each hypergraph 𝐺 𝑠 form an strongly independent set, where
𝛿 < (𝐷 + 1) −4 is to be chosen later.
PackingProcess begins by splitting the edges of the input hypergraph 𝐻 b into a
bulk 𝐻0 and a reservoir 𝐻0∗ by independently selecting edges into the latter with
probability chosen such that 𝑒(𝐻0∗ ) ≈ 𝛾 𝑛𝑟 . As a result, the hypergraphs 𝐻0 and

𝐻0∗ are with high probability quasirandom.

Now PackingProcess proceeds in 𝑠∗ stages. In each stage 𝑠, it runs a randomised
embedding algorithm, called RandomEmbedding and explained below, to embed
the first 𝑛 − 𝛿𝑛 vertices of 𝐺 𝑠 into the bulk 𝐻𝑠−1 . Then in the completion phase
the last 𝛿𝑛 vertices of 𝐺 𝑠 are embedded into the reservoir 𝐻𝑠−1 ∗ . Since there are

exactly 𝛿𝑛 vertices of 𝐺 𝑠 left to embed and exactly 𝛿𝑛 vertices of 𝑉 ( 𝐻)b unused so

far in this stage, we want to find a bijection between these. Since all neighbours
of each yet unembedded vertex are already embedded, this completion amounts
to choosing a system of distinct representatives. The completion phase relies on
choosing a random matching in a super-regular bipartite graph. Now 𝐻𝑠 and 𝐻𝑠∗
are defined simply by removing the edges used in this embedding.
Both RandomEmbedding and the completion phase may fail at any stage 𝑠;

71
 4 Packing Degenerate Hypergraphs

this means that it is not possible to embed a certain part of 𝐺 𝑠 . In that case
PackingProcess fails, too. If PackingProcess does not fail then it always produces
a valid packing of (𝐺 𝑠 ) into 𝐻. So, we need to show that PackingProcess (see
Algorithm 1) succeeds with positive probability.

Algorithm 1: PackingProcess
Input : hypergraphs 𝐺 1 , . . . , 𝐺 𝑠∗ , with 𝐺 𝑠 on vertex set [𝑛] such that the
last 𝛿𝑛 vertices of 𝐺 𝑠 form a strongly independent set; a hypergraph
𝐻b on 𝑛 vertices
choose 𝐻0∗ by picking edges of 𝐻 b independently with probability
𝑛

𝛾 𝑟 /𝑒( 𝐻)b ;
let 𝐻0 = 𝐻 b − 𝐻∗ ;
0
for 𝑠 = 1 to 𝑠∗ do
run RandomEmbedding(𝐺 𝑠 ,𝐻𝑠−1 ) to get an embedding 𝜑 𝑠 of 𝐺 𝑠 [ [𝑛−𝛿𝑛] ]
into 𝐻𝑠−1 ;
let 𝐻𝑠 be the hypergraph

obtained from 𝐻𝑠−1 by removing the edges of
𝜑 𝑠 𝐺 𝑠 [ [𝑛−𝛿𝑛] ] ;
choose uniformly at random an extension 𝜑∗𝑠 of
𝜑 𝑠 embedding all of
𝑉 (𝐺 𝑠 ), and embedding 𝐸 (𝐺 𝑠 ) \ 𝐸 𝐺 𝑠 [ [𝑛−𝛿𝑛] ] into 𝐸 (𝐻𝑠−1 ∗ ) ;
∗ ∗
let 𝐻𝑠 be the hypergraph obtained from 𝐻𝑠−1 by removing the edges
𝜑∗𝑠 𝐸 (𝐺 𝑠 ) \ 𝐸 𝐺 𝑠 [ [𝑛−𝛿𝑛] ] ;

end

For describing our randomised embedding algorithm RandomEmbedding we

need the following definitions. We shall use the symbol ↩→ to denote embeddings
produced by RandomEmbedding. We write 𝐺 ↩→ 𝐻 to indicate that the hypergraph
𝐺 is to be embedded into 𝐻. Also, if 𝑡 ∈ 𝑉 (𝐺), 𝑣 ∈ 𝑉 (𝐻) and 𝐴 ⊆ 𝑉 (𝐻) then
𝑡 ↩→ 𝑣 means that 𝑡 is embedded on 𝑣, and 𝑡 ↩→ 𝐴 means that 𝑡 is embedded on
a vertex of 𝐴. If 𝑡® ∈ 𝑉 (𝐺) 𝑘 , 𝑣® ∈ 𝑉 (𝐻) 𝑘 ordered vertex sets then 𝑡® ↩→ 𝑣® means
that 𝑡® is embedded on 𝑣® in the given order. If 𝜓 is an embedding 𝑉 (𝐺) → 𝑉 (𝐻)
and 𝐴 ⊆ 𝑉 (𝐺) is a vertex set then we write 𝜓( 𝐴) = {𝜓(𝑣)|𝑣 ∈ 𝐴} for the
image of 𝐴. Similarly if A ⊆ P (𝑉 (𝐺)) is a system of vertex sets then we write
𝜓(A) = {𝜓( 𝐴)| 𝐴 ∈ A}.

Definition 4.3.2 (partial embedding, candidate set). Let 𝐺 be a hypergraph with

vertex set [𝑣(𝐺)], and 𝐻 be a hypergraph with 𝑣(𝐻) ≥ 𝑣(𝐺). Further, assume
𝜓 𝑗 : [ 𝑗] → 𝑉 (𝐻) is a partial embedding of 𝐺 into 𝐻 for 𝑗 ∈ [𝑣(𝐻)], that is, 𝜓 𝑗
 
is a hypergraph embedding of 𝐺 [ 𝑗] into 𝐻. Finally, let 𝑡 ∈ [𝑣(𝐻)] be such that

72
 4 Packing Degenerate Hypergraphs

N− (𝑡) ⊆ 𝑃([ 𝑗]). Then the candidate set of 𝑡 (with respect to 𝜓 𝑗 ) is

 
𝑗 −

𝐶𝐺↩→𝐻 (𝑡) = N𝐻 𝜓 𝑗 N𝐺 (𝑡) .

RandomEmbedding (see Algorithm 2) randomly embeds a guest hypergraph 𝐺

into a host hypergraph 𝐻. The algorithm is simple: we iteratively embed the first
(1 − 𝛿)𝑛 vertices of 𝐺 randomly to one of the vertices of their candidate set which
was not used for embedding another vertex already.

Algorithm 2: RandomEmbedding
Input : hypergraphs 𝐺 and 𝐻, with 𝑉 (𝐺) = [𝑣(𝐺)] and 𝑣(𝐻) = 𝑛
𝜓0 := ∅;
𝑡 ∗ := min(𝑣(𝐺), (1 − 𝛿)𝑛);
for 𝑡 = 1 to 𝑡 ∗ do
𝑡−1 (𝑡) \ im(𝜓
if 𝐶𝐺↩→𝐻 𝑡−1 ) = ∅ then halt with failure;
choose 𝑣 ∈ 𝐶𝐺↩→𝐻 (𝑡) \ im(𝜓𝑡−1 ) uniformly at random;
𝑡−1

𝜓𝑡 := 𝜓𝑡−1 ∪ {𝑡 ↩→ 𝑣};
end
return 𝜓𝑡 ∗

To show that PackingProcess does not fail at any stage, we shall show that the
host hypergraph 𝐻𝑠 constructed in PackingProcess in embedding stage 𝑠 is quasir-
andom in the sense of Definition 4.1.1. In fact, in order to analyse the completion
phase of PackingProcess we need quasirandomness of the pair (𝐻𝑠 , 𝐻0∗ ), where
𝐻0∗ is the initial reservoir. We now define this coquasirandomness of a pair of
hypergraphs. Recall that quasirandomness of one hypergraph means that common
neighbourhoods of semi-edges are always about the size one would expect in a
random hypergraph of a similar density. Coquasirandomness of two hypergraphs
means that the intersection of a common neighbourhood in the first hypergraph
and another in the second hypergraph has about the size one would expect in two
independent random hypergraphs of the respective densities.

Definition 4.3.3 (coquasirandom). For 𝛼 > 0 and 𝐿 ∈ N, we say that a pair of

hypergraphs (𝐹, 𝐹 ∗ ), both on the same vertex set 𝑉 of order 𝑛 and semi-edge set
𝑆(𝑉) and with densities 𝑝 and 𝑝 ∗ , respectively, is (𝛼, 𝐿)-coquasirandom if for
every set 𝑆 ⊆ 𝑆(𝑉) of at most 𝐿 semi-edges and every subset 𝑅 ⊆ 𝑆 we have

| N𝐹 (𝑅) ∩ N𝐹 ∗ (𝑆 \ 𝑅)| = (1 ± 𝛼) 𝑝 |𝑅| ( 𝑝 ∗ ) |𝑆\𝑅| 𝑛 .

73
 4 Packing Degenerate Hypergraphs

The reader should make the following important observation: while we are
currently thinking of the reservoir 𝐻0∗ as being a hypergraph of small but positive
density, we will also want to make use of the same analysis when 𝐻0∗ is a zero-
edge hypergraph in order to consider the embedding of the non-spanning graphs.
Coquasirandomness of (𝐹, 𝐹 ∗ ) makes sense if 𝐹 ∗ has no edges (and so 𝑝 ∗ =
0): it reduces to quasirandomness of 𝐹. Similarly, in the following setting, and
consequently in most of the following lemmas, we assume that 𝐹 ∗ has density
either bounded away from zero or equal to zero. If the final 𝛿𝑛 vertices of each
𝐺 𝑠 are isolated vertices, and 𝐻0∗ has density zero, then the embedding loop of
PackingProcess effectively just runs RandomEmbedding repeatedly to embed the
first 𝑛 − 𝛿𝑛 vertices of each 𝐺 𝑠 into 𝐻𝑠−1 and then removes the used edges to form
𝐻𝑠 . We will see this algorithm explicitly (as PackingProcess2) in Section 4.7, and
will use the lemmas below and in the following two sections for its analysis.
With this we can state the setting of our main lemmas and fix various constants
which we will use throughout this chapter.

Setting 4.3.4. Let 𝐷, 𝑛, 𝑟 ∈ N and 𝛾 > 0 be given. Without loss of generality, we

may suppose 𝛾 is sufficiently small to play the rôle of 𝜇 in Lemma 4.2.6. We define

2𝑛𝑟−1 𝛾𝑄
𝑄 = 𝐷 (𝑟 + 1) + 3 , 𝑠max = , 𝜂≤ ,
(𝑟 − 1)! 200𝑄
𝛾 10𝑄 𝜂
𝐶 = 40𝑄 exp 1000𝑄𝛿−2 𝛾 −2𝑄−10 ,


𝛿= ,
106 𝑄 4
𝛿2  108𝐶𝐷 3 𝑄 3𝑟 3𝑟!𝛿−2 𝛾 −2𝑄 (𝑥 − 𝑠 ) 
max
𝛼𝑥 = 8 3 exp for each 𝑥 ∈ R,
10 𝐶 𝑄 𝑛 𝑟−1

𝛼0 𝛿8 𝛾 10𝑄
𝜀= , 𝑐 = 10−8𝐶 −2𝑟 −2 2−2𝑄 𝑄 −4 𝜀 8 and 𝜉 = 𝛼0 /100 ,
1000𝐶𝑄𝑟
(4.3.1)

and we require that 10𝐷𝜂 to be small enough to play the rôle of 𝜀 in Lemma 4.2.6
for input 𝜇 = 𝛾 𝐷 .
Let 𝐺 1 , 𝐺 2 , . . . , 𝐺 𝑠∗ (for some 𝑠∗ ≤ 𝑠max ) be hypergraphs on [𝑛], such that for
−
each 𝑠 and 𝑥 ∈ 𝑉 (𝐺 𝑠 ) we have deg𝐺 𝑠
(𝑥) ≤ 𝐷, such that Δ(𝐺 𝑠 ) ≤ 𝑐𝑛/log 𝑛,
and such that the final 𝛿𝑛 vertices of 𝐺 𝑠 all have degree 𝑑 𝑠 and form an strongly
independent set.
Let 𝐻0 and 𝐻0∗ be two edge-disjoint hypergraphs on the same vertex set of order 𝑛
such that (𝐻0 , 𝐻0∗ ) is ( 41 𝛼0 , 𝑄)-coquasirandom, and 𝑠∈[𝑠∗ ] 𝑒(𝐺 𝑠 ) ≤ 𝑒(𝐻0 ) − 𝛾𝑛𝑟 .
Í

Suppose that 𝐻0∗ has either zero edges or at least (𝛾 − 𝛼0 ) 𝑛𝑟 edges.

74
 4 Packing Degenerate Hypergraphs

Note that in (4.3.1) we give numbers 𝛼𝑥 which we call ‘constant’ even though
𝑛 appears in their definition. Observe that 𝛼𝑥 is strictly increasing in 𝑥. We will
be interested only in values 0 ≤ 𝑥 ≤ 𝑠max (though it is technically convenient to
have the definition for all 𝑥 ∈ R), and it is easy to check that neither 𝛼0 nor 𝛼𝑠max
depends on 𝑛.

Remark 4.3.5. Note that for the proof to work it is sufficient to choose the constants
in the following manner.

𝐷 −1 , 𝛾, 𝑟 −1  𝜂  𝛿  𝐶 −1  𝛼0  𝜀  𝑐 (4.3.2)

We give exact formulas so our calculations can be properly checked, but most of
the time one should simply think of the constants as being of very different orders
of magnitude.

The next lemma justifies splitting our host graph into bulk and reservoir, main-
taining quasirandomness.

Lemma 4.3.6. For each 𝑟 ≥ 2, 𝐷 ∈ N and each 𝛾 > 0, and for each 𝑛 sufficiently
large, let us suppose that the constants 𝑄, 𝛼0 and 𝜉 are as in Setting 4.3.4.
Suppose that 𝐻 b is a (𝜉, 𝑄)-quasirandom hypergraph of order 𝑛 and density
𝑝 ≥ 3𝛾. Let 𝐻0∗ be a 𝑞-random subgraph of 𝐻, b where 𝑞 = 𝛾/𝑝. Let 𝐻0 be the
complement of 𝐻0∗ in 𝐻. b Then with probability at least 1 − 𝑛−5𝐶 , we have that
𝑒(𝐻0∗ ) = (1 ± 𝛼0 )𝛾 𝑛𝑟 and the pair (𝐻0 , 𝐻0∗ ) is 14 𝛼0 , 𝑄 -coquasirandom.

The next lemma states that coquasirandomness of (𝐻𝑠 , 𝐻0∗ ) is preserved when
we embed into the bulk.

Lemma 4.3.7. For each 𝑟 ≥ 2, 𝐷 ∈ N and each 𝛾 > 0, and for each 𝑛 sufficiently
large, the following holds with probability at least 1 − 𝑛−4𝐶 . Suppose that the
constants and 𝐺 1 , 𝐺 2 , . . . , 𝐺 𝑠∗ and the hypergraph 𝐻0 ∪ 𝐻0∗ = 𝐻 and the constant
𝑄 are as in Setting 4.3.4. When PackingProcess is run, for each 𝑠 ∈ [𝑠∗ ] either
PackingProcess fails before completing stage 𝑠, or the pair (𝐻𝑠 , 𝐻0∗ ) is (𝛼𝑠 , 𝑄)-
coquasirandom.

The next lemma estimates the probability that a single execution of Ran-
domEmbedding succeeds.

Lemma 4.3.8. For each 𝑟 ≥ 2, 𝐷 ∈ N, each 𝛾 > 0, and any sufficiently large 𝑛, let
𝑄, 𝛿, 𝜂, 𝛼0 , 𝛼𝑠max , 𝜀 and 𝑐 be as in Setting 4.3.4. Given any 𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max , let 𝐺

75
 4 Packing Degenerate Hypergraphs

be a hypergraph on vertex set [𝑛] with maximum degree at most 𝑐𝑛/log 𝑛 such that
deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺), and let 𝐻 be any (𝛼, 𝑄)-quasirandom 𝑛-vertex


hypergraph with at least 𝛾 𝑛𝑟 edges. When RandomEmbedding is run then it fails
with probability at most 𝑛−5𝐶 .

For the following lemma we need to define the following property which states
that the random choice of the extension has desirable random properties.
Our final two main lemmas concern the completion phase of PackingProcess.
The first states that the completion phase is likely to delete very few edges at any
vertex of 𝐻0∗ .

Lemma 4.3.9. For each 𝑟 ≥ 2, 𝐷 ∈ N and 𝛾 > 0, let 𝑛 be sufficiently large.

Suppose that the constants and 𝐺 1 , 𝐺 2 , . . . , 𝐺 𝑠∗ and 𝐻 are as in Setting 4.3.4, and
suppose 𝐻0∗ has at least (𝛾 − 𝛼0 ) 𝑛𝑟 edges. When PackingProcess is run, with

probability at least 1 − 𝑛−3𝐶 one of the following three events occurs.

• PackingProcess fails.
• There is some 𝑠 ∈ [𝑠∗ ] such that (𝐻𝑠 , 𝐻0∗ ) is not (𝛼𝑠 , 𝑄)-coquasirandom.
• For each 𝑠 ∈ [𝑠∗ ] and 𝑦 ∈ 𝑆(𝐻𝑠∗ ) we have deg𝐻 ∗ (𝑦) − deg𝐻𝑠∗ (𝑦) ≤
0
10𝑟!𝛾 −𝐷 𝐷𝛿𝑛, and (𝐻𝑠 , 𝐻𝑠∗ ) is (𝜂, 𝑄)-coquasirandom.

We will show in the proof of Theorem 4.3.1 that the first two events are unlikely,
so that the likely event is the last one.
Our last lemma states that with high probability, at any stage 𝑠, provided
∗ ) is sufficiently coquasirandom, running RandomEmbedding to par-
(𝐻𝑠−1 , 𝐻𝑠−1
tially embed 𝐺 𝑠 into 𝐻𝑠−1 is likely to give a partial embedding which can be
completed to an embedding of 𝐺 𝑠 using 𝐻𝑠∗ .

Lemma 4.3.10. For each 𝑟 ≥ 2, 𝐷 ∈ N and each 𝛾 > 0, and for each 𝑛 sufficiently
large, let the constants be as in Setting 4.3.4. Suppose that 𝐺 is a hypergraph on
[𝑛], such that we have deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺), we have Δ(𝐺) ≤ 𝑐𝑛/log 𝑛,
and such that the final 𝛿𝑛 vertices of 𝐺 form a strongly independent set, and all
have degree 𝑑. Suppose (𝐻, 𝐻 ∗ ) are a pair of (𝜂, 𝑄)-coquasirandom hypergraphs
on 𝑛 vertices, and 𝐻 is (𝛼𝑠∗ , 𝑄)-quasirandom, with 𝑒(𝐻) = 𝑝 𝑛𝑟 and 𝑒(𝐻 ∗ ) =




(1 ± 𝜂)𝛾 𝑛𝑟 , where 𝑝 ≥ 𝛾. When RandomEmbedding is run to embed 𝐺 [ [𝑛−𝛿𝑛] ]
into 𝐻, with probability at least 1 − 𝑛−2𝐶 it returns a partial embedding 𝜑 which
can be extended to an embedding 𝜑∗ of 𝐺 into 𝐻 ∪ 𝐻 ∗ , with all the edges using a
vertex in {𝑛 − 𝛿𝑛 + 1, . . . , 𝑛} mapped to 𝐻 ∗ .

76
 4 Packing Degenerate Hypergraphs

Let us briefly explain why we cannot simply perform the whole embedding
in the quasirandom 𝐻, b but have to split it into a bulk and a reservoir. In order
to analyse RandomEmbedding, we require that the bulk is very quasirandom, but
RandomEmbedding is very well-behaved and preserves this good quasirandomness.
In contrast, we are not able to show that the completion stage, where we choose a
system of distinct representatives for the remaining vertices, is so well-behaved. If
we used the bulk for this embedding the errors would rapidly become unacceptably
large. However, to show that choosing such a system of distinct representatives
is possible, we do not need much quasirandomness. Thus the reservoir 𝐻𝑠∗ does
rapidly lose its quasirandomness (compared to 𝐻𝑠 ), but it is sufficient for the
completion.
We now argue that our main lemmas imply Theorem 4.3.1.

Proof of Theorem 4.3.1. We can assume that 𝑝 > 3𝛾 as the statement is vacuous
otherwise.
Suppose that we run PackingProcess on the input hypergraphs 𝐺 1 , . . . , 𝐺 𝑠∗ . For
the course of the analysis of this run, we shall first ignore possible failures during
the completion phase. That is, if any failure during the completion phase occurs,
we ignore it and continue embedding using RandomEmbedding into the bulk.
Clearly, this does not change behaviour of future rounds of RandomEmbedding or
the evolution of the bulk.
It is clear that PackingProcess does not fail (in the RandomEmbedding stage)
unless at least one of the following exceptional events occurs:
(i) (𝐻0 , 𝐻0∗ ) is not ( 41 𝛼0 , 𝑄)-coquasirandom.
(ii) RandomEmbedding proceeded through stages 𝑠 = 1, . . . , 𝑠0 (for some 𝑠0 ∈
[𝑠∗ − 1]) without failure, the pairs (𝐻𝑠 , 𝐻0∗ ) are (𝛼𝑠 , 𝑄)-coquasirandom
for 𝑠 < 𝑠0 , and (𝐻𝑠0 , 𝐻0∗ ) is not an (𝛼𝑟 , 𝑄)-coquasirandom pair.
(iii) RandomEmbedding proceeded through stages 𝑠 = 1, . . . , 𝑠0 (for some
𝑠0 ∈ {0, . . . , 𝑠∗ − 1}) without failure, the hypergraphs 𝐻𝑠 are (𝛼𝑠 , 𝑄)-
quasirandom for 𝑠 ≤ 𝑠0 . Then, in stage 𝑠0 + 1, RandomEmbedding fails.
Lemma 4.3.6 gives an upper bound on the probability of the event in (i).
Lemma 4.3.7 gives an upper bound on the probability of all the events in (ii). For
each fixed 𝑠0 ∈ {0, . . . , 𝑠∗ −1}, the event in (iii) can be bounded using Lemma 4.3.8.
Thus, the probability that PackingProcess fails in the RandomEmbedding part is at
most 𝑛−5𝐶 + 𝑛−4𝐶 + 𝑛−5𝐶 .

77
 4 Packing Degenerate Hypergraphs

Let us now analyse the completion phases of PackingProcess. If PackingProcess

fails in one of the completion phases then one of the following events occurs:
(iv) One of the events described under (i)-(iii).
(v) None of (i)-(iii) occurs. RandomEmbedding and the completion phase pro-
ceed successfully through the first 𝑠0 stages (for some 𝑟 ∈ {1, . . . , 𝑠∗ − 1}.
For 𝑠 ∈ [𝑠0 ] all the pairs (𝐻𝑠 , 𝐻0∗ ) are (𝛼𝑠 , 𝑄)-coquasirandom. However,
there is a stage 𝑠 ∈ [𝑟] where (𝐻𝑠 , 𝐻𝑠∗ ) is not (𝜂, 𝑄)-coquasirandom.
(vi) None of (i)-(iii) occurs. RandomEmbedding and the completion phase pro-
ceeds successfully through the first 𝑠0 stages (for some 𝑠0 ∈ {0, . . . , 𝑠∗ −
1}, and throughout all the pairs (𝐻𝑠 , 𝐻0∗ ) and (𝐻𝑠 , 𝐻𝑠∗ ) are (𝛼𝑠 , 𝑄)-
coquasirandom and (𝜂, 𝑄)-coquasirandom, respectively. In stage 𝑠0 + 1,
RandomEmbedding successfully embeds but the completion phase fails.
Lemma 4.3.9 bounds the probability of the event in (v) by 𝑛−3𝐶 . Finally,
Lemma 4.3.10 bounds the probability of events in (vi) for each given 𝑠0 by 𝑛−2𝐶 .
Thus, the total probability of failure due to (v) or (vi) is at most 𝑛−2𝐶 + 𝑛−3𝐶 .
We conclude that PackingProcess packs the hypergraphs 𝐺 1 , . . . , 𝐺 𝑠∗ into 𝐻, b
and none of the above bad events occur, with probability at least 1 − 𝑛−𝐶 . This
in particular gives the desired packing. In addition, we see that (𝐻𝑠∗ , 𝐻𝑠∗∗ ) is
(𝜂, 𝑄)-coquasirandom. Let 𝐻 0 be the hypergraph of leftover edges, i.e. 𝐸 (𝐻 0) =
𝐸 (𝐻𝑠∗ ) ∪ 𝐸 (𝐻𝑠∗∗ ). Given a set 𝑆 of semi-edges of size at most 𝑄, the set N𝐻 0 (𝑆) is
partitioned into sets
N𝐻𝑠∗ (𝑅) ∩ N𝐻 ∗∗ (𝑆 \ 𝑅)
𝑠

as 𝑅 runs over all subsets of 𝑆. Since these sets have sizes controlled by (𝐻𝑠∗ , 𝐻𝑠∗∗ )
being (𝜂, 𝑄)-coquasirandom, we see that 𝐻 0 is (2𝑄 𝜂, 𝑄)-quasirandom, as desired.

It thus remains to prove all the main lemmas from this section. Lemmas 4.3.6
and 4.3.7 are proven in Section 4.5. We actually prove a stronger statement than
Lemma 4.3.8 in Section 4.4. This stronger form, Lemma 4.4.3 is also needed for
proving Lemma 4.3.7. Lemmas 4.3.9 and 4.3.10 are proven in Section 4.6.

4.4 Staying on a diet

In this section we consider the running of RandomEmbedding to embed one de-
generate hypergraph 𝐺 𝑠 into a quasirandom hypergraph 𝐻𝑠−1 . Since we will only

78
 4 Packing Degenerate Hypergraphs

consider one stage 𝑠, to avoid a profusion of subscripts we write 𝐺 in place of 𝐺 𝑠

and 𝐻 in place of 𝐻𝑠−1 .
The basic strategy here is broadly similar to that in [3]. Our main aim is to
show that during the embedding of 𝐺 into 𝐻, if 𝑋 is the set of vertices which have
been used in the embedding at any given time 𝑡, then 𝑋 looks like a random set of
vertices in that it intersects any given common neighbourhood in about as many
vertices as a uniform random set of the same size would do.

Definition 4.4.1 (diet condition, codiet condition). Let 𝐻 be a hypergraph on 𝑛

vertices and 𝑝 𝑛𝑟 edges, and let 𝑋 ⊆ 𝑉 (𝐻) be any vertex set. We say that the pair
(𝐻, 𝑋) satisfies the (𝛽, 𝐿)-diet condition if for every set 𝑆 ⊆ 𝑆(𝐻) of at most 𝐿
semi-edges we have | N𝐻 (𝑆) \ 𝑋 | = (1 ± 𝛽) 𝑝 |𝑆| (𝑛 − |𝑋 |).
Let 𝐻, 𝐻 ∗ be two hypergraphs on vertex set 𝑉 of order 𝑛 vertices with semi-edge
set 𝑆(𝑉) and 𝑝 𝑛𝑟 and 𝑝 ∗ 𝑛𝑟 edges, respectively, and let 𝑋 ⊆ 𝑉 be any vertex set.

We say that the triple (𝐻, 𝐻 ∗ , 𝑋) satisfies the (𝛽, 𝐿)-codiet condition if for every
set 𝑆 ⊆ 𝑆(𝑉) of at most 𝐿 semi-edges and for every subset 𝑅 ⊆ 𝑆 we have

N𝐻 (𝑅) ∩ N𝐻 ∗ (𝑆 \ 𝑅) \ 𝑋 = (1 ± 𝛽) 𝑝 |𝑅| ( 𝑝 ∗ ) |𝑆\𝑅| (𝑛 − |𝑋 |) .

Observe that the (𝛽, 𝐿)-diet condition holding for (𝐻, ∅) is simply the statement
that 𝐻 is (𝛽, 𝐿)-quasirandom, and similarly for the codiet condition. To see why
it is enough to show the diet condition holds for (𝐻, 𝑋) where 𝑋 is the set of
vertices used up to some given time 𝑡 in the embedding, consider the embedding
of vertex 𝑡 + 1 of 𝐺. The only way RandomEmbedding can fail is if there is no
vertex in the candidate set which is not contained in 𝑋. But the candidate set is
precisely a common neighbourhood of some at most 𝐷 semi-edges in 𝐻, namely
the semi-edges to which were embedded the left-neighbourhood of 𝑡 + 1. So the
diet condition tells us how many vertices in the candidate set are not covered by 𝑋,
and in particular that that number is not zero.
In order to argue that we maintain the diet condition, we introduce the cover
condition. Roughly speaking, this states that for any given 𝑣 in 𝐻 and any short
interval of vertices (of length 𝜀𝑛) of 𝐺, about the ‘right fraction’ of vertices 𝑥 in
the interval have 𝑣 in their final candidate set when RandomEmbedding is run. To
make precise what we mean by ‘the right fraction’ some care is needed. How likely
it is that 𝑣 is in the final candidate set of 𝑥 depends on the number of semi-edges
in the left-neighbourhood of 𝑥. Therefore we will partition 𝑉 (𝐺) according to this

79
 4 Packing Degenerate Hypergraphs

number of previous neighbours. Thus we define

𝑋𝑖,𝑑 := {𝑥 ∈ 𝑉 (𝐺) : 𝑖 ≤ 𝑥 < 𝑖 + 𝜀𝑛, | N− (𝑥)| = 𝑑} .

When 𝐺 is given with a 𝐷-degenerate ordering it is enough to consider 𝑑 ∈

{0, 1, . . . , 𝐷}. So if 𝐻 is quasirandom and has 𝑝 𝑛𝑟 edges, then for an arbitrary
𝑣 ∈ 𝑉 (𝐻), we would expect that about a 𝑝 𝑑 -fraction of vertices 𝑥 in each 𝑋𝑖,𝑑 have
𝑣 in their final candidate sets (at time 𝑥 − 1).

Definition 4.4.2 (cover condition). Suppose that 𝐺 and 𝐻 are two hypergraphs
such that 𝐻 has order 𝑛, the vertex set of 𝐺 is [𝑛], and 𝐺 has density 𝑝. Suppose
that numbers 𝛽, 𝜀 > 0 and 𝑖 ∈ [𝑛 − 𝜀𝑛] are given. We say that a partial embedding
𝜓 of 𝐺 into 𝐻, which embeds N− (𝑥) for each 𝑖 ≤ 𝑥 < 𝑖 + 𝜀𝑛, satisfies the
Ð

(𝜀, 𝛽, 𝑖)-cover condition if for each 𝑣 ∈ 𝑉 (𝐻) \ im 𝜓, and for each 𝑑 ∈ N, we have

𝑥 ∈ 𝑋𝑖,𝑑 : 𝑣 ∈ N𝐻 𝜓( N− (𝑥)) = (1 ± 𝛽) 𝑝 𝑑 |𝑋𝑖,𝑑 | ± 𝜀 2 𝑛 .



Note that a corresponding condition for 𝑑 = 0 is trivial, even with zero error
parameters.

The main idea of the analysis is to show that if the diet condition holds up to
some given time 𝑡, then it is unlikely that the cover condition fails at or before
time 𝑡, and similarly if the diet and cover conditions hold up to time 𝑡 then the diet
condition is unlikely to fail at time 𝑡 + 1 (and so RandomEmbedding does not fail
before this time either). We wrap this up in the following lemma.

Lemma 4.4.3 (Diet-and-cover lemma). For each 𝐷 ∈ N, each 𝛾 > 0, and any
sufficiently large 𝑛, let 𝑠max , 𝑄, 𝛿, 𝜂, 𝛼0 , 𝛼𝑠max , 𝜀 and 𝑐, 𝐶 be as in Setting 4.3.4.
Let 𝛼 ∈ [𝛼0 , 𝛼𝑠max ] be arbitrary. Let 𝐺 be a hypergraph on vertex set [𝑛] with
maximum degree at most 𝑐𝑛/log 𝑛 such that deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺), and


let 𝐻 be any (𝛼, 𝑄)-quasirandom 𝑛-vertex hypergraph with at least 𝛾 𝑛𝑟 edges.
Suppose in addition that 𝐻 ∗ is a hypergraph on 𝑉 (𝐻) with 𝑝ˆ 𝑛𝑟 edges such that

(𝐻, 𝐻 ∗ ) is (𝜂, 𝑄)-coquasirandom and either 𝑝ˆ ≥ (1 − 𝜂)𝛾 or 𝑝ˆ = 0. Finally, fix

any set 𝑍 of vertices of 𝐻 with |𝑍 | ≥ 𝑐−1 log 𝑛. Then with probability at least
1 − 𝑛−5𝐶 all of the following good events hold.
(a) When RandomEmbedding is run it does not fail and generates a sequence
(𝜓𝑖 )𝑖∈[𝑛−𝛿𝑛] of partial embeddings of 𝐺 into 𝐻.
(b) For each 𝑡 ∈ [𝑛−𝛿𝑛] the pair (𝐻, im 𝜓𝑡 ) satisfies the (𝐶𝛼, 𝑄)-diet condition.

80
 4 Packing Degenerate Hypergraphs

(c) For each 1 ≤ 𝑡 ≤ 𝑛 + 1 − 𝜀𝑛, the partial embedding 𝜓𝑡+𝜀𝑛−2 of 𝐺 into 𝐻

satisfies the (𝜀, 𝐶𝛼, 𝑡)-cover condition.
(d) For each 𝑡 ∈ [𝑛 − 𝛿𝑛], the triple (𝐻, 𝐻 ∗ , im 𝜓𝑡 ) satisfies the (2𝜂, 𝑄)-codiet
condition.
(e) For each 𝑡 ∈ [𝑛 − 𝛿𝑛], we have |𝑍 \ im 𝜓𝑡 | = (1 ± 𝐶𝛼) 𝑛−𝑡
𝑛 |𝑍 |.

Observe that conclusion (a) of this lemma is the conclusion of Lemma 4.3.8, so
that proving Lemma 4.4.3 also proves Lemma 4.3.8.
The main difficulty is to establish that the cover and diet conditions hold. We
will see that the other two conditions are easy byproducts. The reason for the
difficulty is that the error terms in the cover and diet conditions for small times 𝑡
feed back into the calculations which will establish the cover and diet conditions
for larger times 𝑡, so that the errors grow. To bound their growth, we define a
new sequence of error terms, which we need only in the proof of Lemma 4.4.3.
There are three important points to note about the following sequence 𝛽𝑡 . It is an
increasing sequence, we have 𝛽0 = 2𝛼, and 𝛽𝑛 /𝛽0 is bounded by a constant which
does not depend on 𝛼 (though it does depend on 𝐷, 𝛾 and 𝛿). This last observation
will turn out to be crucial for the analysis of PackingProcess.

Definition 4.4.4. Given 𝑄 and 𝛼, 𝛿, 𝛾 > 0, we define

1000𝑄𝛿−2 𝛾 −2𝑄−10 𝑡

𝛽𝑡 := 2𝛼 exp 𝑛 . (4.4.1)

We will mainly take 𝑡 integer in the range [0, 𝑛], but it is convenient to allow 𝑡 to
be any real number. In particular, for each 𝑡 ≥ 0, we have
∫ 𝑡
1
𝑛 1000𝑄𝛿−2 𝛾 −2𝑄−10 𝛽𝑖 d𝑖
𝑖=0
(4.4.2)
1000𝑄𝛿−2 𝛾 −2𝑄−10
∫ 𝑡
1000𝑄𝛿−2 𝛾 −2𝑄−10 𝑖

≤2𝛼 exp 𝑛 d𝑖 = 𝛽𝑡 .
𝑖=−∞ 𝑛

One should read (4.4.2) as saying that when we want to estimate a parameter of
the process RandomEmbedding at some time 𝑡, even if we make in each step 𝑖 an
error which is a rather large multiple of the current error 𝛽𝑖 , the total error is still
bounded by 𝛽𝑡 .
We split the proof of Lemma 4.4.3 into two parts. The cover lemma
(Lemma 4.4.5) states that if the (𝛽𝑡 , 𝑄)-diet condition holds for (𝐻, im 𝜓𝑖 ) for
each 𝑖 ∈ [𝑡 − 1], then it is very unlikely that the (𝜀, 20𝑄 𝛽𝑡 , 𝑡)-cover condition fails

81
 4 Packing Degenerate Hypergraphs

for 𝜓𝑡+𝜀𝑛−2 . Note that the time 𝑡 + 𝜀𝑛 − 2 is the first time at which the (𝜀, 20𝑄 𝛽𝑡 , 𝑡)-
cover condition is guaranteed to be determined, since at this time all left-neighbours
of all vertices 𝑡, 𝑡 + 1, . . . , 𝑡 + 𝜀𝑛 − 1 have certainly been embedded.

Lemma 4.4.5 (Cover lemma). For each 𝐷, each 𝛾 > 0 and sufficiently large 𝑛, let
𝑄, 𝑠max , 𝛼0 , 𝛼𝑠max , 𝜀, 𝛿 and 𝑐 be as in Setting 4.3.4. Suppose that 𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max
and 𝐺 is a hypergraph on vertex set [𝑛], with deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ [𝑛], with
maximum degree at most 𝑐𝑛/log 𝑛, and suppose that 𝐻 is an 𝑛-vertex hypergraph
of density at least 𝛾. Let 𝛽𝑡 for 0 ≤ 𝑡 ≤ 𝑛 be defined as in (4.4.1) and assume that
1
𝛽𝑛 ≤ 10 . Let 𝑡 with 1 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛 − 𝜀𝑛 + 1 be fixed.
When RandomEmbedding is run to embed 𝐺 into 𝐻, with probability at most
𝑛−6𝐶 the following holds. For each 0 ≤ 𝑖 ≤ 𝑡 − 1 the (𝛽𝑡 , 𝑄)-diet condition holds
for (𝐻, im 𝜓𝑖 ), but the (𝜀, 20𝑄 𝛽𝑡 , 𝑡)-cover condition does not hold for 𝜓𝑡+𝜀𝑛−2 .

Note that for any 0 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛 − 𝜀𝑛, if RandomEmbedding runs up to time 𝑡 and

the (𝛽𝑡 , 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡 ), then by choice of 𝜀 the (2𝛽𝑡 , 𝑄)-diet
condition holds deterministically for (𝐻, im 𝜓 𝑗 ) for each 𝑡 + 1 ≤ 𝑗 ≤ 𝑡 + 𝜀𝑛. In
particular RandomEmbedding cannot fail before time 𝑡 + 𝜀𝑛.
The diet lemma (Lemma 4.4.6) states that when the (𝛽𝑖 , 𝑄)-diet condition holds
for (𝐻, im 𝜓𝑖 ) for each 𝑖 ∈ [𝑡 − 1], and the (𝜀, 20𝑄 𝛽𝑖 , 𝑖)-cover condition holds for
𝜓𝑖+𝜀𝑛−2 for each 𝑖 ∈ [𝑡 + 1 − 𝜀𝑛], then it is unlikely that the (𝛽𝑡 , 𝑄)-diet condition
fails for (𝐻, im 𝜓𝑡 ). We also obtain the desired codiet condition.

Lemma 4.4.6 (Diet lemma). For each 𝑟 ≥ 2, 𝐷 ∈ N, each 𝛾 > 0, and any
sufficiently large 𝑛, let 𝑄, 𝛼0 , 𝛼2𝑛 , 𝜀, 𝛿 and 𝜂 be as in Setting 4.3.4. For any
𝑡 ≤ (1−𝛿)𝑛, and 𝛼0 ≤ 𝛼 ≤ 𝛼2𝑛 the following holds. Suppose that 𝐺 is a hypergraph
on [𝑛] such that deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ [𝑛], and 𝐻 is an (𝛼, 𝑄)-quasirandom


hypergraph with 𝑛 vertices with 𝑝 𝑛𝑟 edges, with 𝑝 ≥ 𝛾. Suppose furthermore that
𝐻 ∗ is a hypergraph on 𝑉 (𝐻) and 𝑝ˆ 𝑛𝑟 edges with either 𝑝ˆ ≥ (1 − 𝜂)𝛾 or 𝑝ˆ = 0,

such that (𝐻, 𝐻 ∗ ) satisfies the (𝜂, 𝑄)-coquasirandomness condition. Let 𝛽𝑡 for
1
0 ≤ 𝑡 ≤ 𝑛 be defined as in (4.4.1) and assume that 𝛽𝑛 ≤ 10 . Finally let 𝑍 be any
−1
subset of 𝑉 (𝐻) of size at least 𝑐 log 𝑛.
When RandomEmbedding is run to embed 𝐺 into 𝐻, with probability at most
𝑛 −6𝐶 the following event occurs.
• For each 1 ≤ 𝑗 ≤ 𝑡 − 1 the (𝛽 𝑗 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑗 ), and
• for each 1 ≤ 𝑗 ≤ 𝑡 + 1 − 𝜀𝑛 the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition holds for
𝜓 𝑗+𝜀𝑛−2 , and

82
 4 Packing Degenerate Hypergraphs

• the (𝛽𝑡 , 𝑄)-diet condition does not hold for (𝐻, im 𝜓𝑡 ), or the (2𝜂, 𝑄)-codiet
condition does not hold for (𝐻, 𝐻 ∗ , im 𝜓𝑡 ), or |𝑍 \ im 𝜓𝑡 | ≠ (1 ± 𝛽𝑡 ) 𝑛−𝑡
𝑛 |𝑍 |.

Since the hypergraphs 𝐺 and 𝐻 are fixed in the proof of Lemma 4.4.3, in this
𝑗
section we drop the subscript in the notation 𝐶𝐺↩→𝐻 (𝑥) and write simply 𝐶 𝑗 (𝑥).
We now show that Lemmas 4.4.5 and 4.4.6, whose proofs are deferred to later in
this section, imply Lemma 4.4.3.

Proof of Lemma 4.4.3. Given 𝑄 and 𝛾, 𝛿 > 0, we have 𝐶 as in Setting 4.3.4.

Given 𝛼 > 0, we define 𝛽𝑡 for each 0 ≤ 𝑡 ≤ 𝑛 as in (4.4.1). We claim that
with high probability, when we run RandomEmbedding, the following event A
occurs: the algorithm RandomEmbedding does not fail, for each 1 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛
the pair (𝐻, im 𝜓𝑡 ) satisfies the (𝛽𝑡 , 𝑄)-diet condition and the triple (𝐻, 𝐻 ∗ , im 𝜓𝑡 )
satisfies the (2𝜂, 𝑄)-codiet condition, and for each 𝜀𝑛 − 1 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛 the
(𝜀, 20𝑄 𝛽𝑡−𝜀𝑛+2 , 𝑡 − 𝜀𝑛 + 2)-cover condition holds for 𝜓𝑡 .
Indeed, if the event A does not occur then there is a first time witnessing
its failure. Let us calculate what is the probability that this first time is 𝑡. We
can thus assume A does not fail before time 𝑡. This in particular means that
the (𝛽 𝑗 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑗 ) for each 1 ≤ 𝑗 < 𝑡, and the
(𝜀, 20𝑄 𝛽 𝑗−𝜀𝑛+2 , 𝑗 − 𝜀𝑛 + 2)-cover condition holds for 𝜓 𝑗 for each 𝜀𝑛 − 1 ≤ 𝑗 < 𝑡.
Firstly, we show that RandomEmbedding cannot fail at time 𝑡, then we use
Lemma 4.4.5 to show that with high probability the (𝜀, 20𝑄 𝛽𝑡−𝜀𝑛+2 , 𝑡 − 𝜀𝑛 + 2)-
cover condition holds for 𝜓𝑡 . Because the (𝛽𝑡−1 , 𝑄)-diet condition holds for (𝐻,
𝜓𝑡−1 ), picking 𝑆 = 𝜓𝑡−1 ( N− (𝑡)), we have 𝐶 𝑡−1 (𝑡)\im 𝜓𝑡−1 = N𝐻 (𝑆)\im 𝜓𝑡−1 > 0.
It follows that RandomEmbedding cannot fail at time 𝑡. Now by Lemma 4.4.5, the
probability of the (𝜀, 20𝑄 𝛽𝑡−𝜀𝑛+2 , 𝑡 −𝜀𝑛 +2)-cover condition failing is at most 𝑛−6𝐶 .
Secondly, we use Lemma 4.4.6 to show that with high probability neither diet
condition fails at time 𝑡. More precisely, by Lemma 4.4.6, the probability that the
(𝛽𝑡 , 𝑄)-diet condition fails for (𝐻, im 𝜓𝑡 ), or the (2𝜂, 𝑄)-codiet condition fails for
(𝐻, 𝐻 ∗ , im 𝜓𝑡 ), is at most 𝑛−6𝐶 .
We conclude that the probability that a given 𝑡 is the first time that we witness
event A failing is at most 2𝑛−6𝐶 . Taking a union bound over the at most 𝑛 choices
of 𝑡, we see that with probability at least 1 − 𝑛−5𝐶 the good event from the statement
of Lemma 4.4.3 holds, i.e., that RandomEmbedding does not fail, and by the choice
of 𝐶 and by (4.4.1), for each 1 ≤ 𝑡 ≤ (1 − 𝛿)𝑛 the pair (𝐻, im 𝜓𝑡 ) satisfies the
(𝐶𝛼, 𝑄)-diet condition and the triple (𝐻, 𝐻 ∗ , im 𝜓𝑡 ) satisfies the (2𝜂, 𝑄)-codiet

83
 4 Packing Degenerate Hypergraphs

condition, and for each 1 ≤ 𝑡 ≤ 𝑛 + 1 − 𝜀𝑛 the embedding 𝜓𝑡+𝜀𝑛−2 satisfies the

(𝜀, 𝐶𝛼, 𝑡)-cover condition, as desired.

For the next proof we need the following notation. Given 1 ≤ ℓ ≤ 𝑑 ≤ 𝐷 we

call a sequence (𝑎 1 , . . . , 𝑎 ℓ ) whose entries are in [𝑑] an (ℓ, 𝑑)-pattern if we have
Íℓ
𝑖=1 𝑎 𝑖 = 𝑑.
Next for a vertex 𝑥 ∈ 𝑋𝑡,𝑑 we define a pattern. Let 𝐸 𝑥 denote the following

hyperedge set: 𝑒 ∈ 𝐸 (𝐺) : 𝑒 ⊆ [𝑥], 𝑥 ∈ 𝑒}, that is the hyperedges that contain
𝑥 as last vertex. For each vertex 𝑦 ∈ [𝑥 − 1] let 𝑓 (𝑦) denote the number of edges
𝑒 ∈ 𝐸 𝑥 such that 𝑦 ∈ 𝑒 and 𝑦 = max(𝑒 \ {𝑥}), i.e. 𝑦 is the second largest vertex in
𝑒. We need to remove the zeroes from this sequence in order to obtain the desired
pattern. To that end, let ℓ denote the number of 𝑦 ∈ [𝑥 − 1] such that 𝑓 (𝑦) is
non-zero, and let 𝑎 1 , . . . , 𝑎 ℓ be the nonzero values of 𝑓 (𝑦) as 𝑦 runs from 1 to
𝑥 − 1 in order. By construction, this is a (ℓ, |𝐸 𝑥 |)-pattern which we call the pattern
associated to 𝑥.
If a is any (ℓ, 𝑑)-pattern with 1 ≤ ℓ ≤ 𝑑 ≤ 𝐷, we let 𝑋𝑡,a be defined as the set
of vertices in 𝑋𝑡,𝑑 whose associated pattern is a. We note that for each 𝑑 there are
2𝑑−1 possible patterns.
We now prove the cover lemma.



Proof of Lemma 4.4.5. Let 𝑒(𝐺) = 𝑝 𝑛𝑟 ≥ 𝛾 𝑛𝑟 . Let D be the event that the
(𝛽𝑡 , 𝑄)-diet condition holds for each (𝐻, im 𝜓𝑖 ) with 1 ≤ 𝑖 ≤ 𝑡 − 1. We fix a vertex
𝑣 ∈ 𝑉 (𝐻). Observe that if 𝑣 ∈ im 𝜓𝑡+𝜀𝑛−2 then there is nothing to prove, so when
it is necessary we will assume 𝑣 ∉ im 𝜓𝑡+𝜀𝑛−2 . We also fix 1 ≤ ℓ ≤ 𝑑 ≤ 𝐷 and an
Í
(ℓ, 𝑑)-pattern a. Let 𝑠𝑖 denote 𝑖𝑗=1 𝑎 𝑗 and 𝑠0 = 0. Define B𝑣,a as the event that D
holds, and that 𝑣 and a witness the failure of the following condition,

𝑥 ∈ 𝑋𝑡,a : 𝑣 ∈ N𝐻 𝜓( N− (𝑥)) = (1 ± 20𝑄 𝛽𝑡 ) 𝑝 𝑑 |𝑋𝑡,a | ± 𝜀 2 𝑛/2𝑑 .



Our aim is to show that

P B𝑣,a ≤ 𝑛−7𝐶 𝐷 −1 2−𝐷 .

 
(4.4.3)

A union bound over the choices of 𝑣, 𝑑 and a then gives the lemma.
Our strategy for proving (4.4.3) is as follows. Ideally, we would like to assert
that for each 𝑥 ∈ 𝑋𝑡,a the probability of 𝑣 ∈ 𝐶 𝑥−1 (𝑥) is roughly 𝑝 𝑑 and apply
Corollary 4.2.2 to bound the probability of the bad event B𝑣,a . Note that at time

84
 4 Packing Degenerate Hypergraphs

𝑖 = 0, we have 𝑣 ∈ 𝐶 𝑖 (𝑥), and as 𝑖 increases, the set 𝐶 𝑖 (𝑥) changes exactly at times
when a vertex 𝑦 ∈ 𝑅(𝑥) (i.e. a vertex which is second-to-last in a hyperedge with
last vertex 𝑥) is embedded. This ideal strategy is not possible because the indicator
variables 1{𝑣∈𝐶 𝑥−1 (𝑥)} may not be sequentially dependent as 𝑥 ranges over 𝑋𝑡,a : the
sets 𝑅(𝑥) may interleave each other. For a vertex 𝑦 ∈ 𝑅(𝑥), we write 𝐹𝑥 (𝑦) for the
set of semi-edges 𝑒 ∈ N𝐺 − (𝑥) such that 𝑦 = max(𝑒) (i.e. 𝐹 (𝑦) is the set of edges
𝑥
which witness 𝑦 ∈ 𝑅(𝑥)). We write 𝑊𝑥 (𝑦) for the event that for each semi-edge
𝑒 ∈ 𝐹𝑥 (𝑦) we have {𝑣} ∪ 𝜓(𝑒) ∈ 𝐸 (𝐻). That is, the edges that have 𝑦 as their
second to last vertex do not stop 𝑥 from having 𝑣 in their final candidate set. We
refine our previous strategy as follows. Let 𝑦 1 , . . . , 𝑦 𝑙 be the vertices of 𝑅(𝑥) in
increasing order, and we define

ℓ
Ù


e𝑥 (𝑘) = 𝑊𝑥 (𝑦 1 ) ∩ 𝑊𝑥 (𝑦 2 ) ∩ . . . ∩ 𝑊𝑥 (𝑦 𝑘 )
𝑊 so 𝑣 ∈ 𝐶 𝑥−1 (𝑥) = e𝑥 (𝑘) .
𝑊
𝑘=1
(4.4.4)
The event 𝑊 e𝑥 (ℓ), of course, equals the entire intersection (4.4.4). However, this
more complicated way of expressing (4.4.4) gets us into the setting of sequential
dependence as in Corollary 4.2.2.
More formally, given 1 ≤ 𝑘 ≤ ℓ and 𝑦 ∈ 𝑉 (𝐺), we define random variables
𝑌𝑘,1 , . . . , 𝑌𝑘,𝑡+𝜀𝑛−2 as follows. Let 𝑌𝑘,𝑦 be the number of vertices 𝑥 ∈ 𝑋𝑡,a such that
𝑦 is the 𝑘th vertex of 𝑅(𝑥) in increasing order and 𝑊 e𝑥 (𝑘) holds. In other words,
𝑌𝑘,𝑦 counts the number of 𝑥 ∈ 𝑋𝑡,a where 𝑦 is the 𝑘th element of 𝑅(𝑥) and which
immediately after embedding 𝑦 could still be embedded to 𝑣 (i.e. all the semi-edges
of N− (𝑥) which are contained in [𝑦] have been embedded to semi-edges that make
a hyperedge with 𝑣). Observe that 𝑌𝑘,1 , . . . , 𝑌𝑘,𝑡+𝜀𝑛−2 are by definition sequentially
dependent.
𝑎𝑖 for the number of semi-edges of N− (𝑥) which are fully
Í𝑘
We write 𝑠 𝑘 = 𝑖=1
embedded once we embed the 𝑘th element of 𝑅(𝑥). These quantities are by
definition of a pattern the same for each 𝑥 ∈ 𝑋𝑡,a . For each 0 ≤ 𝑘 ≤ ℓ, we let Y𝑘
be the event that (1 ± 10𝛽𝑡 ) 𝑠 𝑘 𝑝 𝑠 𝑘 |𝑋𝑡,a | ± 𝑠 𝑘 𝜀 2 𝑛/(𝑑2𝑑 ) vertices 𝑥 ∈ 𝑋𝑡,a have the
property that 𝑊 e𝑥 (𝑘) holds. Observe that the event Y𝑘 is precisely the statement
that
𝑡+𝜀𝑛−2
Õ
𝑌𝑘,𝑦 = (1 ± 10𝛽𝑡 ) 𝑠 𝑘 𝑝 𝑠 𝑘 |𝑋𝑡,a | ± 𝑠 𝑘 𝜀 2 𝑛/(𝑑2𝑑 ) . (4.4.5)
𝑦=1

85
 4 Packing Degenerate Hypergraphs

Our bad event then satisfies

B𝑣,a ⊆ D and not Y𝑑 ,

because we have 𝑠 𝑘 ≤ 𝑑 and (1 ± 10𝛽𝑡 ) 𝑑 = 1 ± 20𝑄 𝛽𝑡 . In order to bound the

probability of B𝑣,a we cover B𝑣,a with ℓ events, each of whose probabilities we can
bound with Corollary 4.2.2. For this purpose we define the event

E 𝑘 = Y𝑘−1 and D

for each 1 ≤ 𝑘 ≤ ℓ. Note that E1 = D since Y0 holds trivially with probability

one. We thus have

Ø

B𝑣,a ⊆ D and not Y𝑑 ⊆ E 𝑘 and not Y𝑘 .
1≤𝑘 ≤ℓ

Our aim then is to show that for each 1 ≤ 𝑘 ≤ ℓ we have

P[E 𝑘 and not Y𝑘 ] ≤ 𝑛−7𝐶 ℓ −1 𝐷 −1 2−𝐷 . (4.4.6)

Note that this and a union bound over the ℓ choices of 𝑘 gives (4.4.3).
To establish (4.4.6) we would like to apply Corollary 4.2.2. Hence we need to
E[𝑌𝑘,𝑦 |H𝑦−1 ], where H𝑦−1 is
Í
argue that either E 𝑘 fails, or we can estimate 𝑡+𝜀𝑛−2
𝑦=1
the history of embedding decisions taken in RandomEmbedding up to and including
the embedding of vertex 𝑦 − 1. To this end, for 𝑦 ∈ [𝑡 + 𝜀𝑛 − 2] let 𝑍 𝑘,𝑦 be the
number of vertices 𝑥 ∈ 𝑋𝑡,a such that 𝑦 is the 𝑘th vertex of 𝑅(𝑥) and 𝑊 e𝑥 (𝑘 − 1)
holds. Also let 𝑌𝑘,𝑦,𝑥 and 𝑍 𝑘,𝑦,𝑥 be indicator variables that 𝑥 is counted in 𝑌𝑘,𝑦 and
𝑍 𝑘,𝑦 respectively. Then the quantity 𝑍 𝑘,𝑦,𝑥 is determined by H𝑦−1 and
 
E[𝑌𝑘,𝑦,𝑥 |H𝑦−1 ] = 𝑍 𝑘,𝑦,𝑥 · P 𝑊𝑥 (𝑦)|H𝑦−1 . (4.4.7)

Observe further that

𝑡+𝜀𝑛−2
Õ 𝑡+𝜀𝑛−2
Õ
𝑍 𝑘,𝑦 = 𝑌𝑘−1,𝑦 , (4.4.8)
𝑦=1 𝑦=1

because both sums count the number of vertices 𝑥 ∈ 𝑋𝑡,𝑎 such that the first 𝑘 − 1
e𝑥 (𝑘 −1) holds, in the first sum grouped
vertices of 𝑅(𝑥) are embedded to 𝐻 so that 𝑊
by the 𝑘th vertex of 𝑅(𝑥) and in the second sum by the (𝑘 − 1)st vertex.
Assume now that 𝑥, 𝑦 ∈ 𝑉 (𝐺) are fixed and that H𝑦−1 does not witness that E 𝑘

86
 4 Packing Degenerate Hypergraphs

fails, and let us bound P 𝑊𝑥 (𝑦)|H𝑦−1 . Since H𝑦−1 does not witness that E 𝑘 fails
 

and D ⊆ E 𝑘 , the (𝛽𝑡 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑦−𝜀𝑛 ), where we have to
subtract 𝜀𝑛 in the index of 𝜓 𝑦−𝜀𝑛 because 𝑦 could be as large as 𝑡 + 𝜀𝑛 − 2 (and we
only know that the diet condition holds up to time 𝑡 − 1). This implies that for each
set 𝑆 of semi-edges in 𝐻 with |𝑆| ≤ 𝑄 we have

N𝐻 (𝑆) \ im 𝜓 𝑦−1 = (1 ± 𝛽𝑡 ) 𝑝 |𝑆| (𝑛 − 𝑦 + 𝜀𝑛) ± 𝜀𝑛

= (1 ± 𝛽𝑡 ) 𝑝 |𝑆| (𝑛 − 𝑦 + 1) ± 2𝜀𝑛 = (1 ± 2𝛽𝑡 ) 𝑝 |𝑆| (𝑛 − 𝑦 + 1) ,

where the last inequality follows from 𝛾 ≤ 𝑝 and 𝜀 ≤ 𝛼𝛾 𝑄 ≤ 12 𝛽𝑡 𝛾 𝑄 . We conclude

that the (2𝛽𝑡 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑦−1 ).
Given 𝜓 𝑦−1 , we define a set 𝑁 ⊆ 𝐶 𝑦−1 (𝑦) \ im 𝜓 𝑦−1 of vertices of 𝐺 with the
following property: if we embed 𝑦 to any 𝑢 ∈ 𝑁, then we can still embed 𝑥 to
𝑣. That is, we put 𝑢 ∈ 𝐶 𝑦−1 (𝑦) \ im 𝜓 𝑦−1 in 𝑁 if and only if all the semi-edges
N− (𝑥) ∩ [𝑦] are embedded by 𝜓 𝑦−1 ∪ {𝑦 ↩→ 𝑢} to semi-edges of 𝐻 which form
edges together with 𝑣. Note that this definition forces 𝑁 ∩ im 𝜓 𝑦−1 = ∅.
We can write 𝑁 differently: 𝑁 = N𝐻 (𝑆) \ im 𝜓 𝑦−1 where 𝑆 is the collection
of semi-edges 𝜓 𝑦−1 N− (𝑦) together with semi-edges of the form {𝑣} ∪ 𝜓 𝑦−1 ( 𝑓 )

where 𝑦 ∉ 𝑓 and 𝑓 ∪ {𝑦} is a semi-edge of 𝐹𝑥 (𝑦). There are deg− (𝑦) semi-edges
of the first type, and 𝑎 𝑘 of the second type. Using this observation, we can estimate
|𝑁 | using the diet condition. Observe that these semi-edges are all distinct: by
definition the semi-edges 𝜓 𝑦−1 N− (𝑦) contain only vertices in im 𝜓 𝑦−1 , while 𝑣 is

not in im 𝜓 𝑦−1 . Furthermore they are genuinely semi-edges, i.e. they have 𝑟 − 1
vertices: for the first type this is obvious since 𝜓 𝑦−1 is injective, while for the
second type we need to observe that any such 𝑓 has 𝑟 − 2 vertices by definition and
𝜓 𝑦−1 ( 𝑓 ) does not contain 𝑣.
Since deg− (𝑦) ≤ 𝐷 we have

−
(𝑦)
𝐶 𝑦−1 (𝑦) \ im 𝜓 𝑦−1 = (1 ± 2𝛽𝑡 ) 𝑝 deg (𝑛 − 𝑦 + 1) and
−
|𝑁 | = (1 ± 2𝛽𝑡 ) 𝑝 𝑎 𝑘 +deg (𝑦)
(𝑛 − 𝑦 + 1) .

Therefore we have
  |𝑁 |
P 𝑊𝑥 (𝑦)|H𝑦−1 = = (1 ± 10𝛽𝑡 ) 𝑝 𝑎 𝑘 .
𝐶 (𝑦) \ im 𝜓 𝑦−1
𝑦−1

87
 4 Packing Degenerate Hypergraphs

We conclude from (4.4.7) that

𝑡+𝜀𝑛−2
Õ 𝑡+𝜀𝑛−2
Õ
E(𝑌𝑘,𝑦 |H𝑦−1 ) = (1 ± 10𝛽𝑡 ) 𝑝 𝑍 𝑘,𝑦 , (4.4.9)
𝑦=1 𝑦=1

unless E 𝑘 fails. Further, unless E 𝑘 fails, we have

𝑡+𝜀𝑛−2 𝑡+𝜀𝑛−2
𝑠 𝑘−1 𝜀 2 𝑛
Õ (4.4.8)
Õ (4.4.5)
𝑍 𝑘,𝑦 = 𝑌𝑘−1,𝑦 = (1 ± 10𝛽𝑡 ) 𝑠 𝑘−1 𝑝 𝑠 𝑘−1 |𝑋𝑡,a | ± 𝑑2𝑑
.
𝑦=1 𝑦=1

Plugging this into (4.4.9), and noting 𝑠 𝑘−1 + 1 ≤ 𝑠 𝑘 , we get that E 𝑘 fails or we have

𝑡+𝜀𝑛−2
(𝑠 𝑘−1 +0.5)𝜀 2 𝑛
Õ
E[𝑌𝑘,𝑦 |H𝑦−1 ] = (1 ± 10𝛽𝑡 ) 𝑠 𝑘 𝑝 𝑠 𝑘 |𝑋𝑡,a | ± 𝑑2𝑑
.
𝑦=1

Since 0 ≤ 𝑌𝑘,𝑦 ≤ deg(𝑦) for each 𝑦, we can thus apply Corollary 4.2.2 with
the event E = E 𝑘 , with 𝑅 = Δ(𝐺), with 𝜇˜ ± 𝜈˜ = (1 ± 10𝛽𝑡 ) 𝑠 𝑘 𝑝 𝑠 𝑘 |𝑋𝑡,a | ± 𝑠 𝑘−1 +
0.5 𝜀 2 𝑛𝑑 −1 2−𝑑 , and with 𝜚˜ = 21 𝜀 2 𝑛𝑑 −1 2−𝑑 to conclude that


 𝑡+𝜀𝑛−2
Õ 

P [E 𝑘 and not Y𝑘 ] = P E 𝑘 and
 𝑌𝑘,𝑦 ≠ 𝜇 ± (𝜈 + 𝜚) 
 𝑦=1 
  
− 𝜚˜ 2

≤ 2 exp .
2𝑅( 𝜇˜ + 𝜈˜ + 𝜚)˜

Substituting Δ(𝐺) ≤ 𝑐𝑛/log 𝑛, and because 𝑐 ≤ 10−4𝐶 −1𝑟 −1 2−2𝐷 𝐷 −4 𝜀 4 /(100 + 𝑟)

and 𝑑 ≤ 𝐷, we obtain (4.4.6) as desired.

We deduce the diet lemma from the following simplified version.

Lemma 4.4.7. For each 𝑟 ≥ 2, 𝐷 ∈ N, each 𝛾 > 0, and any sufficiently large 𝑛,
let 𝑄, 𝛼0 , 𝛼2𝑛 , 𝜀, 𝛿 and 𝜂 be as in Setting 4.3.4. For any 𝑡 ≤ (1 − 𝛿)𝑛, and
𝛼0 ≤ 𝛼 ≤ 𝛼2𝑛 the following holds. Suppose that 𝐺 is a hypergraph on [𝑛] such
that deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ [𝑛], and 𝐻 is an (𝛼, 𝑄)-quasirandom hypergraph
with 𝑛 vertices with 𝑝 𝑛𝑟 edges, with 𝑝 ≥ 𝛾. Suppose furthermore that 𝐻 ∗ is a




hypergraph on 𝑉 (𝐻) and 𝑝ˆ 𝑛𝑟 edges with either 𝑝ˆ ≥ (1 − 𝜂)𝛾 or 𝑝ˆ = 0, such that
(𝐻, 𝐻 ∗ ) satisfies the (𝜂, 𝑄)-coquasirandomness condition. Let 𝛽𝑡 for 0 ≤ 𝑡 ≤ 𝑛 be
1
defined as in (4.4.1) and assume that 𝛽𝑛 ≤ 10 . Finally let 𝑍 0 be any subset of 𝑉 (𝐻)
of size at least 𝑐−1 log 𝑛.

88
 4 Packing Degenerate Hypergraphs

When RandomEmbedding is run to embed 𝐺 into 𝐻, with probability at most

𝑛−7𝐶 the following event occurs.
• For each 1 ≤ 𝑗 ≤ 𝑡 − 1 the (𝛽 𝑗 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑗 ), and
• for each 1 ≤ 𝑗 ≤ 𝑡 + 1 − 𝜀𝑛 the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition holds for 𝜓 𝑗 ,
and
• we have |𝑍 0 \ im 𝜓𝑡 | ≠ 1 ± 41 𝛽𝑡 𝑛−𝑡 0


𝑛 |𝑍 |.

Before we prove this lemma, let us briefly observe that it implies Lemma 4.4.6.

Proof of Lemma 4.4.6. Observe that the difference between Lemmas 4.4.6
and 4.4.7 is that the probability bound in Lemma 4.4.7 is stronger and the er-
ror term on the size of |𝑍 ∩ im 𝜓𝑡 | is smaller, and that there are a few more ways
in which we can enter the unlikely event of Lemma 4.4.6, namely there can be a
failure of the diet or codiet conditions at time 𝑡.
Suppose that 𝑅 ⊆ 𝑆 ⊆ 𝑆(𝐻) are sets of semi-edges, with |𝑆| ≤ 𝑄. Then 𝑅
witnesses a failure of the (𝛽𝑡 , 𝑄)-diet condition for (𝐻, im 𝜓𝑡 ) if and only if we
have
N𝐻 (𝑅) \ im 𝜓𝑡 ≠ (1 ± 𝛽𝑡 ) 𝑝 |𝑅| (𝑛 − 𝑡) .

Observe that by (𝛼, 𝑄)-quasirandomness of 𝐻 we have N𝐻 (𝑅) = (1 ± 𝛼) 𝑝 |𝑅| 𝑛.

Letting 𝑍 0 = N𝐻 (𝑅), if we have |𝑍 0 ∩ im 𝜓𝑡 | = 1 ± 14 𝛿𝛽𝑡 𝑛𝑡 |𝑍 |, then we have

 
N𝐻 (𝑅) \ im 𝜓𝑡 = (1 ± 𝛼) 𝑝 |𝑅| 𝑛 1 − 1 ± 14 𝛿𝛽𝑡 𝑛𝑡 = (1 ± 𝛼) 𝑝 |𝑅| 𝑛 − 𝑡 ± 41 𝛿𝛽𝑡 𝑛

= (1 ± 𝛼) 𝑝 |𝑅| 1 ± 14 𝛽𝑡 (𝑛 − 𝑡) = (1 ± 𝛽𝑡 ) 𝑝 |𝑅| (𝑛 − 𝑡)

where the first equality on the second line uses 𝑛 − 𝑡 ≥ 𝛿𝑛 and the second that
𝛼 = 21 𝛽0 ≤ 21 𝛽𝑡 and that 𝛽𝑡 is sufficiently small. In particular, what this cal-
culation establishes is that if 𝑅 witnesses a failure of the (𝛽𝑡 , 𝑄)-diet condition
for (𝐻, im 𝜓𝑡 ), then the corresponding 𝑍 0 = N𝐻 (𝑅) witnesses the low-probability
event of Lemma 4.4.7 occurring. Note that since |𝑍 0 | ≥ (1 − 𝛼) 𝑝 |𝑅| 𝑛 ≥ 12 𝛾 𝑄 𝑛, the
condition on |𝑍 0 | of Lemma 4.4.7 is indeed satisfied.
If 𝑝ˆ = 0, then the (2𝜂, 𝑄)-codiet condition for (𝐻, 𝐻, ˆ im 𝜓𝑡 ) is implied by
the (𝛽𝑡 , 𝑄)-diet condition for (𝐻, im 𝜓𝑡 ) since 𝛽𝑡 < 𝜂. If 𝑝ˆ ≥ (1 − 𝜂)𝛾, then a
similar calculation shows that if 𝑅 and 𝑆 witness a failure of the (2𝜂, 𝑄)-codiet
condition for (𝐻, 𝐻 ∗ , im 𝜓𝑡 ) then the corresponding 𝑍 0 = N𝐻 (𝑅) ∩ N𝐻 ∗ (𝑆 \ 𝑅)
witnesses the low-probability event of Lemma 4.4.7 occurring. This calculation
holds with rather more room to spare since 𝜂 is much larger than 𝛼, and we omit

89
 4 Packing Degenerate Hypergraphs

the details. Taking a union bound over the at most 𝑛2𝑟𝑄 + 1 choices of 𝑅 and
𝑆, and of 𝑍 in Lemma 4.4.6, we observe that the probability that any one of the
corresponding 𝑍 0 for Lemma 4.4.7 witnesses the low-probability event occurring
is at most 2𝑛2𝑟𝑄 𝑛−7𝐶 < 𝑛−6𝐶 . This is the required upper bound on the probability
of the unlikely event of Lemma 4.4.6.

We now prove Lemma 4.4.7.

Proof of Lemma 4.4.7. Observe that if 𝜓𝑡−1 satisfies the (𝛽𝑡−1 , 𝑄)-diet condition,
RandomEmbedding cannot fail at time 𝑡, so 𝜓𝑡 exists. To begin with, we show that
in any short interval of time, not too many vertices can be embedded to 𝑍 0. This
analysis is not particularly accurate; we need it for the more accurate analysis that
follows.
Claim 4.4.8. For every 0 ≤ 𝑗 ≤ 𝑡 − 1, if the (𝛽 𝑗 , 𝑄)-diet condition holds for
(𝐻, im 𝜓 𝑗 ), then with probability at least 1 − 𝑛−10𝐶 we have

𝑍 0 ∩ (im 𝜓min( 𝑗+𝜀𝑛,𝑡) \ im 𝜓 𝑗 ) ≤ 4𝜀𝛾 −𝐷 𝛿−1 |𝑍 0 | .

Proof. At each time 𝑗 + 1 ≤ 𝑖 ≤ min( 𝑗 + 𝜀𝑛, 𝑡), we embed the vertex 𝑖 to the
set 𝐶 𝑖−1 (𝑖) \ im 𝜓𝑖−1 . This set is a common neighbourhood of some at most
𝐷 semi-edges, from which we remove im 𝜓 𝑗 and a further at most 𝜀𝑛 vertices.
Since (𝐻, im 𝜓 𝑗 ) satisfies the diet condition, we conclude that 𝐶 𝑖−1 (𝑖) \ im 𝜓𝑖−1 ≥
3 𝐷 1 𝐷 0
4 𝛾 𝛿𝑛 − 𝜀𝑛 ≥ 2 𝛾 𝛿𝑛. The probability of embedding 𝑖 to 𝑍 is thus at most
2𝛾 −𝐷 𝛿−1 |𝑍 0 |𝑛−1 . By Corollary 4.2.2, the probability that more than 4𝛾 −𝐷 𝛿−1 𝜀|𝑍 0 |
vertices 𝑖 with 𝑗 + 1 ≤ 𝑖 ≤ min( 𝑗 + 𝜀𝑛, 𝑡) are embedded to 𝑍 0, is at most exp −
1 −𝐷 −1
𝛿 𝜀|𝑍 0 | ≤ 𝑛−9𝐶 since |𝑍 0 | ≥ 𝑐−1 log 𝑛 and by choice of 𝑐.


2𝛾 

We now state a claim that if the diet condition holds up to time 𝑡 − 𝜀𝑛, then for
any given large set 𝑇 ⊆ 𝑍 0, with high probability either the cover condition fails at
some time before 𝑡 − 𝜀𝑛, or 𝜓𝑡 embeds about the expected fraction of each interval
of 𝜀𝑛 vertices to 𝑇.
Claim 4.4.9. For every 1 ≤ 𝑗 ≤ 𝑡 − 𝜀𝑛 + 1, and for every 𝑇 ⊆ 𝑉 (𝐻) \ im 𝜓 𝑗
with |𝑇 | ≥ 21 𝛾 𝐷 𝛿|𝑍 0 |, if the (𝛽 𝑗 , 𝑄)-diet condition holds for (𝐻, im 𝜓 𝑗 ), then with
probability at least 1 − 𝑛−9𝐶 , one of the following occurs.
(a) 𝜓𝑡 does not have the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition, or
(b) {𝑥 : 𝑗 ≤ 𝑥 < 𝑗 + 𝜀𝑛, 𝜓𝑡−1 (𝑥) ∈ 𝑇 } = (1 ± 40𝑄 𝛽 𝑗 ) |𝑇𝑛−|𝜀𝑛𝑗 .

90
 4 Packing Degenerate Hypergraphs

We defer the proof of this claim to later, and move on to state a second claim,
which we will deduce from Claim 4.4.9. Let ℓ = b 𝜀𝑛𝑡 c. We claim it is likely that
either we witness a failure of the diet or cover conditions before time 𝑡, or the set
𝑍 0 \ im 𝜓ℓ𝜀𝑛 has about the expected size.
Claim 4.4.10. With probability at least 1 − 𝑛−8𝐶 , one of the following holds.
(a) The (𝛽 𝑗 , 𝑄)-diet condition fails for (𝐻, im 𝜓 𝑗 ) for some 1 ≤ 𝑗 ≤ 𝑡 − 1, or
(b) the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition fails for 𝜓𝑡−1 for some 1 ≤ 𝑗 ≤ 𝑡 + 1 − 𝜀𝑛,
or
(c) we have

ℓ−1 
Ö 
0 0
𝜀𝑛
𝑍 \ im 𝜓ℓ𝜀𝑛 = |𝑍 | 1 − 1 ± 40𝑄 𝛽 𝑘𝜀𝑛 𝑛−𝑘𝜀𝑛 . (4.4.10)
𝑘=0

Before proving these claims, we show that they imply the lemma. Suppose that
the likely event of Claim 4.4.10 holds, and, if ℓ𝜀𝑛 < 𝑡, that the likely event of
Claim 4.4.8 with 𝑗 = ℓ𝜀𝑛 holds. Taking logs, we have

log 𝑍 0 \ im 𝜓ℓ𝜀𝑛
ℓ−1
Õ  
0 𝜀𝑛
= log 𝑍 + log 1 − (1 ± 40𝑄 𝛽 𝑘𝜀𝑛 ) 𝑛−𝑘𝜀𝑛
𝑘=0
ℓ−1 
Õ 
40𝑄 𝛽 𝑘 𝜀𝑛 𝜀𝑛

= log 𝑍 0 + log 𝑛−(𝑘+1)𝜀𝑛
𝑛−𝑘𝜀𝑛 + log 1 ± 𝑛−(𝑘+1)𝜀𝑛
𝑘=0
ℓ−1
Õ
0
40𝑄 𝛽 𝑘 𝜀𝑛 𝜀
= log 𝑍 + log 1 − ℓ𝜀 ± 2 1−(𝑘+1)𝜀 ,
𝑘=0

where the final equality holds since 1 − (𝑘 + 1)𝜀 ≥ 𝛿, and hence by choice of 𝜀
the quantity 40𝑄 𝛽 𝑘 𝜀𝑛 𝜀
1−(𝑘+1)𝜀 is close to 0. By the likely event of Claim 4.4.8 we assumed,
if ℓ𝜀𝑛 < 𝑡 then the number of vertices embedded to 𝑍 0 \ im 𝜓ℓ𝜀𝑛 by 𝜓𝑡 is at most
4𝜀𝛿−1 𝛾 −𝑄 |𝑍 0 \𝜓ℓ𝜀𝑛 |. If ℓ𝜀𝑛 = 𝑡 then the same estimate holds trivially. We conclude
|𝑍 0 \ 𝜓ℓ𝜀𝑛 | = |𝑍 0 \ 𝜓𝑡 | ± 4𝜀𝛿−1 𝛾 −𝑄 |𝑍 0 |, and so

ℓ−1
𝑛 − 𝑡 ± 𝜀𝑛  Õ 
𝑍 0 \ im 𝜓𝑡 = |𝑍 0 | · · exp ± 80𝑄𝛿−1 𝜀 𝛽 𝑘𝜀𝑛 ± 4𝜀𝛿−1 𝛾 −𝑄 |𝑍 0 | .
𝑛 𝑘=0
(4.4.11)

91
 4 Packing Degenerate Hypergraphs

Now since 𝛽𝑥 is increasing in 𝑥, we can estimate

ℓ−1 ℓ𝑛−1 ∫ ℓ𝑛
Õ Õ (4.4.2)
−1 −1
80𝑄𝛿 𝜀 𝛽 𝑘𝜀𝑛 ≤ 1
𝑛 80𝑄𝛿 𝛽𝑥 ≤ 1
𝑛 80𝑄𝛿−1 𝛽𝑥 d𝑥 ≤ 1
16 𝛽ℓ𝑛 ≤ 1
16 𝛽𝑡 .
𝑘=0 𝑥=0 𝑥=−∞

1 1
= 1 ± 81 𝛽𝑡 . Plugging this into (4.4.11)


Since 16 𝛽𝑡 is small, we have exp ± 16 𝛽𝑡
we get
 𝑛 − 𝑡 ± 𝜀𝑛 
0 0 −1 −𝑄
1
= 1 ± 41 𝛽𝑡 |𝑍 0 | 𝑛−𝑡



𝑍 \ im 𝜓𝑡 = |𝑍 | · · 1± 8 𝛽𝑡 ± 4𝜀𝛿 𝛾 𝑛 ,
𝑛

where the final equality uses the fact that 𝜀 is tiny compared to 𝛽𝑡 , 𝛿2 and 𝛾 𝑄 .
This concludes the proof of the lemma, modulo the proofs of Claim 4.4.9 and
Claim 4.4.10, which we now provide.

Proof of Claim 4.4.9. Let 𝑗 and 𝑇 be as in the statement, and suppose that the
likely event of Claim 4.4.8 holds for 𝑍 0 and 𝑗. Fix 0 ≤ 𝑑 ≤ 𝐷. We want to show
how to make use of the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition for 𝜓 𝑗 (which we have when
Part (a) fails) to deduce that the assertion of Part (b) holds with high probability.
That is, we consider the number of vertices in 𝑋 𝑗,𝑑 embedded to 𝑇. In order to
apply Corollary 4.2.2, we want to estimate the sum over 𝑥 ∈ 𝑋 𝑗,𝑑 of the probability
that 𝑥 is embedded to 𝑇, conditioning on 𝜓𝑥−1 , that is, we need to estimate the
number
𝑇 ∩ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1
. (4.4.12)
𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1
By the diet condition, we have 𝐶 𝑥−1 (𝑥) \ im 𝜓 𝑗 = (1 ± 𝛽 𝑗 ) 𝑝 𝑑 (𝑛 − 𝑗). Since
𝑗 < 𝑡 ≤ (1 − 𝛿)𝑛, since 𝑥 ≤ 𝑗 + 𝜀𝑛, since 𝑝 ≥ 𝛾, and by choice of 𝜀, we have

𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 = (1 ± 2𝛽 𝑗 ) 𝑝 𝑑 (𝑛 − 𝑗) , (4.4.13)

thus providing a bound on the denumerator in (4.4.12). (Note that this bound on
the denumerator does not depend on the choice of 𝑥 ∈ 𝑋 𝑗,𝑑 .) Now 𝑥 is embedded
uniformly at random into 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 , so it remains to determine the sum of
the numerators in (4.4.12). We rewrite the sum as
Õ Õ
𝑇 ∩ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 = {𝑥 ∈ 𝑋 𝑗,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 } .
𝑥∈𝑋 𝑗,𝑑 𝑣∈𝑇

We split this sum into two cases. For 𝑣 ∉ im 𝜓 𝑗+𝜀𝑛 , by definition 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \

92
 4 Packing Degenerate Hypergraphs

im 𝜓𝑥−1 holds if and only if 𝑣 ∈ 𝐶 𝑥−1 (𝑥), and by the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition
we have

{𝑥 ∈ 𝑋 𝑗,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 } = (1 + 20𝛽 𝑗 ) 𝑝 𝑑 |𝑋 𝑗,𝑑 | ± 𝜀 2 𝑛 .

Observe that this quantity is bounded between 0 and 𝜀𝑛. For 𝑥 ∈ im 𝜓 𝑗+𝜀𝑛 , we use
the trivial bound 0 ≤ {𝑥 ∈ 𝑋 𝑗,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 } ≤ 𝜀𝑛, which we write as

{𝑥 ∈ 𝑋 𝑗,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 } = (1 + 20𝛽 𝑗 ) 𝑝 𝑑 |𝑋 𝑗,𝑑 | ± 𝜀 2 𝑛 ± 𝜀𝑛 .

Since there are by the good event of Claim 4.4.8 at most 4𝜀𝛾 −𝐷 𝛿−1 |𝑍 0 | vertices
𝑥 ∈ 𝑇 such that 𝑥 ∈ im 𝜓 𝑗+𝜀𝑛 , we obtain the estimate
Õ Õ
𝑇 ∩ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 = {𝑥 ∈ 𝑋 𝑗,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 }
𝑥∈𝑋 𝑗,𝑑 𝑣∈𝑇
 
= |𝑇 | (1 + 20𝛽 𝑗 ) 𝑝 |𝑋 𝑗,𝑑 | ± 𝜀 𝑛 ± 4𝜀𝛾 −𝐷 𝛿−1 |𝑍 0 | · 𝜀𝑛
𝑑 2
 
= |𝑇 | (1 + 20𝛽 𝑗 ) 𝑝 𝑑 |𝑋 𝑗,𝑑 | ± 10𝜀 2 𝛾 −2𝐷 𝛿−2 𝑛 , (4.4.14)

where the second line uses the bounds we calculated above and the third our
assumption |𝑇 | ≥ 12 𝛾 𝐷 𝛿|𝑍 0 |.
We can thus apply Corollary 4.2.2, setting E to be the event that the
(𝜀, 20𝐷 𝛽 𝑗 , 𝑗)-cover condition holds for 𝜓 𝑗 . Combining (4.4.13) and (4.4.14),
the expected number of vertices of 𝑋 𝑗,𝑑 embedded to 𝑇 is

(1 ± 20𝑄 𝛽 𝑗 ) 𝑝 𝑑 |𝑇 ||𝑋 𝑗,𝑑 | ± 10𝜀 2 𝛾 −2𝐷 𝛿−2 𝑛|𝑇 |

=
(1 ± 2𝛽 𝑗 ) 𝑝 𝑑 (𝑛 − 𝑗)
|𝑇 ||𝑋 𝑗,𝑑 |
= (1 ± 30𝑄 𝛽 𝑗 ) ± 20𝜀 2 𝛾 −3𝐷 𝛿−3 |𝑇 | ,
𝑛− 𝑗

where we use 𝑛− 𝑗 ≥ 𝛿𝑛 and 𝑝 ≥ 𝛾. By Corollary 4.2.2, with 𝑅 = 1, the probability

that the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition holds for 𝜓 𝑗 and the outcome differs from
this by more than 𝜀 2 |𝑇 | is at most 2 exp(−𝜀 4 |𝑇 |) ≤ 𝑛−10𝐶 , where the inequality
holds by our assumptions on |𝑇 | and |𝑍 0 | and choice of 𝑐. Taking the union bound
over the 𝐷 + 1 choices of 𝑑 and the unlikely events of Claim 4.4.8, we conclude
that with probability at most 𝑛−9𝐶 the (𝜀, 20𝑄 𝛽 𝑗 , 𝑗)-cover condition holds for 𝜓 𝑗

93
 4 Packing Degenerate Hypergraphs

and the number of vertices 𝑥 with 𝑗 ≤ 𝑥 < 𝑗 + 𝜀𝑛 embedded to 𝑇 is not equal to

|𝑇 |𝜀𝑛 |𝑇 |𝜀𝑛
(1 ± 30𝑄 𝛽 𝑗 ) ± 40(𝐷 + 1)𝜀 2 𝛾 −3𝐷 𝛿−3 |𝑇 | = (1 ± 40𝑄 𝛽 𝑗 ) ,
𝑛− 𝑗 𝑛− 𝑗

where the final equality uses our choice of 𝜀 tiny compared to 𝐷 −1 , 𝛽 𝑗 , 𝛾 3𝐷 and
𝛿4 . This is what we wanted to show. 

Proof of Claim 4.4.10. We set 𝑇𝑘 = 𝑍 0 \ im 𝜓 𝑘𝜀𝑛 . Suppose that |𝑇𝑘 | ≥ 21 𝛾 𝐷 𝛿|𝑍 0 |,

and note that this assumption holds for 𝑘 = 0 trivially. On this assumption, we can
apply Claim 4.4.9 with 𝑇 = 𝑇𝑘 and obtain that with probability at least 1 − 𝑛−9𝐶
either a failure of the diet or the cover condition is witnessed before time 𝑘, or we
have 
𝜀𝑛 
|𝑇𝑘+1 | = |𝑇𝑘 | 1 − 1 ± 40𝑄 𝛽 𝑘𝜀𝑛 .
𝑛 − 𝑘𝜀𝑛
Inductively, taking the union bound over 0 ≤ 𝑘 ≤ ℓ − 1, with probability at least
1 − 𝑛1−9𝐶 > 1 − 𝑛−8𝐶 one of the following occurs. Either we witness a failure
of the diet or cover condition before time ℓ𝜀𝑛, or for each 1 ≤ 𝑘 ≤ ℓ we have
|𝑇𝑘−1 | ≥ 12 𝛾 𝐷 𝛿|𝑍 0 | and hence

𝑘−1 
Ö 
𝑇𝑘 = |𝑍 0 | 𝜀𝑛
1 − (1 ± 40𝑄 𝛽𝑖𝜀𝑛 ) 𝑛−𝑖𝜀𝑛 ≥ 12 𝛾 𝐷 𝛿|𝑍 0 | . 
𝑖=0

4.5 Maintaining quasirandomness

In this section we provide the proofs of Lemma 4.3.6 and Lemma 4.3.7.

4.5.1 Initial coquasirandomness

We begin with the easy proof of Lemma 4.3.6, which states that splitting the edges
of a quasirandom hypergraph randomly gives a coquasirandom pair with high
probability.

Proof of Lemma 4.3.6. Using Theorem 4.2.1 we see that the densities 𝑝 0 and 𝑝 ∗0
of 𝐻0 and 𝐻0∗ satisfy

𝑝 0 = (1 ± 𝛼0
1000𝑄 )( 𝑝 − 𝛾) and 𝑝 ∗0 = (1 ± 𝛼0
1000𝑄 )𝛾 (4.5.1)

94
 4 Packing Degenerate Hypergraphs

with probability at least 1 − 𝑛−6𝐶 , giving the first part of Lemma 4.3.6.
Now, let 𝑅 ⊆ 𝑆 ⊆ 𝑆( 𝐻) b be two sets of size at most 𝑄. By quasirandomness of 𝐻 b
we have | N𝐻b (𝑆)| = (1 ± 𝜉) 𝑝 |𝑆| 𝑛. Observe that each vertex of N𝐻b (𝑆) appears with
probability 𝑞 |𝑅| (1 − 𝑞) |𝑆\𝑅| in N𝐻0∗ (𝑅) ∩ N𝐻0 (𝑆 \ 𝑅). Hence,
h i
E N (𝑅) ∩ N𝐻0 (𝑆 \ 𝑅)
𝐻0∗ = 𝑞 |𝑅| (1 − 𝑞) |𝑆\𝑅| (1 ± 𝜉) 𝑝 |𝑆| 𝑛 .

Observe also that for distinct vertices in N𝐻b (𝑆) the events whether these appear in
N𝐻0∗ (𝑅) ∩ N𝐻0 (𝑆 \ 𝑅) are independent. Using again Theorem 4.2.1, with probability
at least 1 − 𝑛−𝑄−10 we have that

N𝐻0∗ (𝑅) ∩ N𝐻0 (𝑆 \ 𝑅) = 𝑞 |𝑅| (1 − 𝑞) |𝑆\𝑅| (1 ± 2𝜉) 𝑝 |𝑆| 𝑛 . (4.5.2)

Taking the union bound we conclude that (4.5.2) holds for all 𝑆 ⊆ 𝑆( 𝐻)
b with
|𝑆| ≤ 𝑄 and 𝑅 ⊆ 𝑆 with probability at least 1 − 𝑛−6𝐶 .
Now, assume that (4.5.1) holds. Then the right-hand side of (4.5.2) can be
rewritten as
! |𝑅| ! |𝑆\𝑅|
𝑝 ∗0
(1 ± 2𝜉)𝛾 |𝑅| ( 𝑝 − 𝛾) |𝑆\𝑅| 𝑛 = (1 ± 2𝜉) 𝜉0
𝑝0
𝛼0 𝑛
1± 1000𝑄 1± 1000𝑄

= (1 ± 2𝜉)(1 ± 100𝛼0
)( 𝑝 ∗0 ) |𝑅| 𝑝 0|𝑆\𝑅|
1

∗ |𝑅| |𝑆\𝑅|
= 1± 10 𝛼0 ( 𝑝 0 ) 𝑝 0 .

1
We conclude that (𝐻0∗ , 𝐻0 ) is 10 𝛼0 , 𝑄)-coquasirandom with probability at least
1 − 𝑛−5𝐶 .

4.5.2 Maintaining coquasirandomness

In this subsection we prove Lemma 4.3.7. We need to show that, provided coquasir-
andomness is maintained up to stage 𝑠 −1 and RandomEmbedding does not fail, it is
likely that coquasirandomness holds after stage 𝑠, when 𝐺 𝑠 is embedded into 𝐻𝑠−1
and we obtain 𝐻𝑠 . Let us briefly sketch the idea (for convenience focusing only on
quasirandomness of 𝐻𝑠 ). We fix a set 𝑅 ⊆ 𝑆( 𝐻) b with |𝑅| ≤ 𝑄, and consider the
running of PackingProcess up to stage 𝑠. We want to show that it is very unlikely
that 𝑅 witnesses the failure of 𝐻𝑠 to be quasirandom, since then the union bound
over choices of 𝑅 tells us that it is likely that 𝐻𝑠 is quasirandom. In other words,

95
 4 Packing Degenerate Hypergraphs

we want to know that 𝑁 𝐻𝑠 (𝑅) is very likely close to the expected size. We write

𝑁 𝐻𝑠 (𝑅) = 𝑁 𝐻0 (𝑅) − 𝑌1 − · · · − 𝑌𝑠 ,

where 𝑌𝑖 = 𝑁 𝐻𝑖 −1 (𝑅) − 𝑁 𝐻𝑖 (𝑅) is the change at step 𝑖, and apply Corollary 4.2.2
to show that the sum 𝑌1 + · · · + 𝑌𝑠 is very likely to be close to its expectation. So
proving Lemma 4.3.7 boils down to estimating accurately E(𝑌𝑖 |𝐻𝑖−1 ). We now
sketch how this goes.
Observe that 𝑌𝑖 is equal to the number of 1-stars in 𝐻𝑖−1 whose leaves are 𝑅
with centre 𝑣 ∈ N𝐻𝑖−1 (𝑅) (that is, the hypergraph obtained from 𝑅 by adding to
each semi-edge of 𝑅 the vertex 𝑣 to make a hyperedge of 𝐻𝑖−1 ), at least one of
whose hyperedges is used in embedding 𝐺 𝑖 to 𝐻𝑖−1 . By linearity of expectation,
E(𝑌𝑖 |𝐻𝑖−1 ) is equal to the sum, over 1-stars in 𝐻𝑖−1 whose leaves are 𝑅, of the
probability that at least one edge in the star is used in embedding 𝐺 𝑖 . We will see
that this probability is about the same for any given star 𝑆, and the problem is to
calculate it. To do this we need to consider the running of RandomEmbedding.
First, we find accurate bounds on the probability any one of a small set 𝑈 of
vertices being used in a short time interval in RandomEmbedding. That is, we con-
dition on a give 𝜓𝑡−1 and suppose 𝑈 ∩ im 𝜓𝑡−1 = ∅. We then let RandomEmbedding
embed the following vertices 𝑡, 𝑡 + 1, . . . , 𝑡 + 𝜀𝑛 − 1. If we embedded each such
|𝑈|
vertex to an unused vertex uniformly at random, it would have a roughly 𝑛−𝑡 chance
𝜀𝑛|𝑈|
of being embedded to 𝑈, and we would estimate a probability 𝑛−𝑡 of using any
vertex of 𝑈 during all the 𝜀𝑛 embeddings. Note that we do not consider in this heur-
istic either the fact that the number of used vertices is decreasing or that we might
embed to two different vertices of 𝑈. These two facts do affect the probability, but
(because 𝜀 is tiny) the effect is negligible. What is important is that of course we
do not embed a vertex 𝑥 uniformly to the unused vertices, but rather to the unused
vertices in its candidate set, which intersects 𝑈 in an unknown amount. We use the
diet condition to control the number of vertices in the candidate set of 𝑥 that are
unused, and the cover condition to control (on average) the size of the intersection
with 𝑈. A complication is that at time 𝑡 − 1 the (𝜀, 𝐶𝛼, 𝑡)-cover condition which
we want to use has not been decided: to get around this, we suppose it is unlikely
to fail.
Lemma 4.5.1. For each 𝑟 ≥ 2, 𝐷 ∈ N and 𝛾 > 0, let 𝑄, 𝛿, 𝛼0 , 𝛼𝑠max , 𝐶, 𝜀 be as in
Setting 4.3.4. The following holds for any 𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max and all sufficiently large
𝑛. Suppose that 𝐺 is a hypergraph on [𝑛] such that deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺),

96
 4 Packing Degenerate Hypergraphs

and 𝐻 is an (𝛼, 𝑄)-quasirandom hypergraph with 𝑛 vertices and 𝑝 𝑛𝑟 edges, with
𝑝 ≥ 𝛾. Let 1 ≤ 𝑘 ≤ 2𝐶 be fixed and 𝑈 ⊆ 𝑉 (𝐻) with |𝑈| = 𝑘 be arbitrary. When
RandomEmbedding is run to embed 𝐺 into 𝐻, for any 1 ≤ 𝑡 ≤ 𝑛 + 1 − (𝛿 + 𝜀)𝑛, if
the history H𝑡−1 up to and including embedding 𝑡 − 1 is such that |𝑈 ∩ im 𝜓𝑡−1 | = 0,
the (𝐶𝛼, 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡−1 ), and the probability, conditioned
on H𝑡−1 , of the (𝜀, 𝐶𝛼, 𝑡)-cover condition failing is at most 𝑛−3 , we have
 
P 𝑈 ∩ im 𝜓𝑡+𝜀𝑛−1 | ≥ 1 H𝑡−1 = (1 ± 10𝐶𝛼) 𝑘𝜀𝑛
𝑛−𝑡 .

Furthermore, for any 0 ≤ 𝑞 < 𝜀𝑛, we have

 
P 𝑈 ∩ im 𝜓𝑡+𝑞−1 | ≥ 1 H𝑡−1 ≤ 2𝜀𝛾 𝐷 𝛿𝑘 .

The main technical difficulty in this proof is to rigorously allow for embedding
to multiple vertices of 𝑈. To deal with this, we define the following Modified
RandomEmbedding, which generates a sequence of embeddings with an identical
distribution to RandomEmbedding, but which in addition generates a sequence of
reported vertices. The modification we make is simple: at each time 1 ≤ 𝑡 0 ≤
𝑛 − 𝛿𝑛, RandomEmbedding chooses a vertex of 𝐶𝐺↩→𝐻 𝑡 0 −1 (𝑡 0) \ im 𝜓 0 . In Modified
𝑡 −1
0 −1 0
RandomEmbedding, we instead choose a vertex 𝑤 of 𝐶𝐺↩→𝐻 (𝑡 ) \ (im 𝜓𝑡 0−1 \𝑈), and
𝑡

report this vertex. If the reported vertex 𝑤 is not in im 𝜓𝑡 0−1 , we set 𝜓𝑡 0 = 𝜓𝑡 0−1 ∪
{𝑡 0 ↩→ 𝑤}, as in RandomEmbedding. If the reported vertex is in im 𝜓𝑡 0−1 (which
happens only if 𝑤 ∈ 𝑈) we choose 𝑤 0 uniformly at random in 𝐶𝐺↩→𝐻 𝑡 0 −1 (𝑡 0) \ im 𝜓 0 ,
𝑡 −1
0 0
and set 𝜓𝑡 0 = 𝜓𝑡 0−1 ∪ {𝑡 ↩→ 𝑤 }. It is not too hard to estimate the expected number
of vertices of 𝑈 reported, and easy to show that the contribution due to multiple
reports of vertices of 𝑈 is tiny. The probability of RandomEmbedding using 𝑣
is the same as the probability that Modified RandomEmbedding reports 𝑣 at least
once, which we can thus calculate.

Proof of Lemma 4.5.1. Instead of RandomEmbedding, we consider Modified Ran-

domEmbedding as defined above, which creates the same embedding distribution.
For each 𝑖, let 𝑟 (𝑖) be the vertex reported by Modified RandomEmbedding at time
𝑖.
Note that since the (𝐶𝛼, 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡−1 ), for each 𝑡 ≤

97
 4 Packing Degenerate Hypergraphs

𝑥 < 𝑡 + 𝜀𝑛, setting 𝑆 = 𝜓𝑥−1 ( N− (𝑥)), we have

𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 ± 𝑘 = N𝐻 (𝑆) \ im 𝜓𝑡−1 ± 𝜀𝑛 ± 𝑘

− (𝑥)|
= (1 ± 𝐶𝛼) 𝑝 |N (𝑛 − 𝑡) ± 𝜀𝑛 ± 𝑘 (4.5.3)
− (𝑥)|
= (1 ± 2𝐶𝛼) 𝑝 |N (𝑛 − 𝑡) .

To begin with, we obtain an easy (but not very accurate) upper bound on the
probability of a vertex 𝑣 ∈ 𝑈 being used by a vertex 𝑥 with 𝑡 ≤ 𝑥 < 𝑡 + 𝜀𝑛.
By (4.5.3), the probability that any such 𝑥 is embedded to 𝑣, conditioned on 𝜓𝑥−1 ,
is at most 2𝛾 𝐷 𝛿𝑛−1 , and so by the union bound the probability that some vertex 𝑥
with 𝑡 ≤ 𝑥 < 𝑡 + 𝜀𝑛 is embedded to 𝑣 is at most 2𝜀𝛾 𝐷 𝛿. Summing over 𝑣 ∈ 𝑈 we
obtain the ‘furthermore’ statement of the lemma.
We shall use the following two auxiliary claims.
Define 𝐸 as the random variable counting the times when a vertex in 𝑈 is reported
by Modified RandomEmbedding in the interval 𝑡 ≤ 𝑥 < 𝑡 + 𝜀𝑛,

𝐸= 𝑥 ∈ [𝑡, 𝑡 + 𝜀𝑛) : 𝑟 (𝑥 − 1) ∈ 𝑈 .

The probability that RandomEmbedding uses vertices of 𝑈 in the interval 𝑡 ≤

𝑥 < 𝑡 + 𝜀𝑛, conditioning on H𝑡−1 , is equal to the probability that Modified Ran-
domEmbedding reports a vertex of 𝑈 at least once in that interval, which probability
is by definition at least

h i Õ𝜀𝑛  
E 𝐸 | H𝑡−1 − ≥ ℓ | H𝑡−1 .

P 𝑥 ∈ [𝑡, 𝑡 + 𝜀𝑛) : 𝑟 (𝑥 − 1) = 𝑣
ℓ=2

Our first claim estimates E 𝐸 | H𝑡−1 .

 

Claim 4.5.2. We have that

h i  𝜀𝑛 
E 𝐸 | H𝑡−1 = 𝑘 (1 ± 4𝐶𝛼) 2 −2𝐷 −2
± 4(𝐷 + 1)𝜀 𝛾 𝛿 .
𝑛−𝑡

Our second claim is that the sum in the expression above is small.

98
 4 Packing Degenerate Hypergraphs

Claim 4.5.3. We have that

𝜀𝑛 
Õ 
≥ ℓ | H𝑡−1 ≤ 8𝑘 2 𝜀 2 𝛾 −2𝐷 𝛿−2

P 𝑥 ∈ [𝑡, 𝑡 + 𝜀𝑛) : 𝑟 (𝑥 − 1) = 𝑣
ℓ=2

≤ 8𝑘𝐶𝜀 2 𝛾 −2𝐷 𝛿−2 .

By choice of 𝜀, we have 64(𝐷 + 1)𝐶𝜀 2 𝛾 −2𝐷 𝛿−2 < 𝐶𝛼𝜀𝛿−1 , so putting the two
claims together we have a proof of the Lemma.

Proof of Claim 4.5.2. Let us fix a vertex 𝑣 ∈ 𝑈. By linearity of expectation, it is

enough to estimate the random variable 𝐸 0 counting the number of times when 𝑣
is reported by Modified RandomEmbedding in the interval 𝑡 ≤ 𝑥 < 𝑡 + 𝜀𝑛.
By linearity of expectation, we have

h i 𝑡+𝜀𝑛−1
Õ
E 𝐸 0 | H𝑡−1 = P 𝑣 is reported at time 𝑥 H𝑡−1


𝑥=𝑡
𝑡+𝜀𝑛−1
1{𝑣 ∈ 𝐶 𝑥−1 (𝑥)}
Õ  
= E H𝑡−1 (4.5.4)
𝑥=𝑡
|𝐶 𝑥−1 (𝑥) \ (im 𝜓𝑥−1 \ {𝑣})|
𝑡+𝜀𝑛−1
1{𝑣 ∈ 𝐶 𝑥−1 (𝑥)}
Õ  
= E H𝑡−1 .
𝑥=𝑡
|𝐶 𝑥−1 (𝑥) \ im 𝜓𝑥−1 | ± 𝑘

Using (4.5.3), we get

P 𝑣 ∈ 𝐶 𝑥−1 (𝑥) H𝑡−1

h i 𝑡+𝜀𝑛−1

Õ
E 𝐸 0 | H𝑡−1 = − (𝑥)| .
𝑥=𝑡 (1 ± 2𝐶𝛼) 𝑝 |N (𝑛 − 𝑡)

Splitting this sum up according to | N− (𝑥)|, and again using linearity of expectation,
we have

E |{𝑥 ∈ 𝑋𝑡,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥)}| H𝑡−1

h i 𝐷

Õ
E 𝐸 | H𝑡−1 =
0
.
𝑑=0
(1 ± 2𝐶𝛼) 𝑝 𝑑 (𝑛 − 𝑡)

We now need to estimate E |{𝑥 ∈ 𝑋𝑡,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥)}| H𝑡−1 . We do this by

separating two cases. In the case that the (𝜀, 𝐶𝛼, 𝑡)-cover condition holds and
𝑣 ∉ im 𝜓𝑡+𝜀𝑛−1 , the cover condition gives

|{𝑥 ∈ 𝑋𝑡,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥)}| = (1 ± 𝐶𝛼) 𝑝 𝑑 |𝑋𝑖,𝑑 | ± 𝜀 2 𝑛 .

99
 4 Packing Degenerate Hypergraphs

If either the (𝜀, 𝐶𝛼, 𝑡)-cover condition fails, or 𝑣 ∈ im 𝜓𝑡+𝜀𝑛−1 , we use the trivial
bound that the given set size is in [0, 𝜀𝑛] since |𝑋𝑡,𝑑 | ≤ 𝜀𝑛, and write

|{𝑥 ∈ 𝑋𝑡,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥)}| = (1 ± 𝐶𝛼) 𝑝 𝑑 |𝑋𝑖,𝑑 | ± 𝜀 2 𝑛 ± 𝜀𝑛 .

The second case holds with probability at most 𝑛−3 + 2𝜀𝛾 𝐷 𝛿, by our assumption
on the likelihood of the cover condition failing and by our easy upper bound on the
probability of embedding some 𝑥 to 𝑣. Putting these together, we get

E |{𝑥 ∈ 𝑋𝑡,𝑑 : 𝑣 ∈ 𝐶 𝑥−1 (𝑥)}| H𝑡−1 = (1 ± 𝐶𝛼) 𝑝 𝑑 |𝑋𝑖,𝑑 | ± 𝜀 2 𝑛 ± 𝑛−3 + 2𝜀𝛾 𝐷 𝛿 𝜀𝑛

= (1 ± 𝐶𝛼) 𝑝 𝑑 |𝑋𝑖,𝑑 | ± 3𝛾 𝐷 𝛿𝜀 2 𝑛 .

Substituting this in, we have

(1 ± 𝐶𝛼) 𝑝 𝑑 |𝑋𝑡,𝑑 | ± 3𝛾 𝐷 𝛿𝜀 2 𝑛
h i Õ𝐷
E 𝐸 | H𝑡−1
0
=
𝑑=0
(1 ± 2𝐶𝛼) 𝑝 𝑑 (𝑛 − 𝑡)
𝜀𝑛
= (1 ± 4𝐶𝛼) 𝑛−𝑡 ± 4(𝐷 + 1)𝜀 2 𝛾 −2𝐷 𝛿−2 ,

where the last equality uses 𝑝 ≥ 𝛾 and 𝑛 − 𝑡 ≥ 𝛿𝑛.

By linearity of expectation, we have
h i h i
E 𝐸 | H𝑡−1 = 𝑘E 𝐸 | H𝑡−1 , 0

which proves the claim. 

Proof of Claim 4.5.3. Since the (𝐶𝛼, 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡−1 ),
since 𝑝 ≥ 𝛾, and since 𝑛 − 𝑡 ≥ 𝛿𝑛, for each 𝑥 ∈ [𝑡, 𝑡 + 𝜀𝑛), when we embed
𝑥 we report a uniform random vertex from a set of size at least 21 𝛾 𝐷 𝛿𝑛. The prob-
ability of reporting a vertex in 𝑈 when we embed 𝑥 is thus at most 2𝑘𝛾 −𝐷 𝛿−1 𝑛−1 ,
conditioning on H𝑡−1 and any embedding of the vertices [𝑡, 𝑥). Since the condi-
tional probabilities multiply, the probability that at each of a given ℓ-set of vertices
in [𝑡, 𝑡 + 𝜀𝑛) we report 𝑣 is at most 2ℓ 𝑘 ℓ 𝛾 −ℓ𝐷 𝛿−ℓ 𝑛−ℓ . Taking the union bound over

100
 4 Packing Degenerate Hypergraphs

choices of ℓ-sets, we have

𝜀𝑛 
Õ 
≥ ℓ | H𝑡−1 ≤

P 𝑥 ∈ [𝑡, 𝑡 + 𝜀𝑛) : 𝑟 (𝑥 − 1) = 𝑣
ℓ=2
𝜀𝑛   𝜀𝑛
𝜀𝑛 ℓ ℓ −ℓ𝐷 −ℓ −ℓ Õ 4𝑘 2 𝜀 2 𝛾 −2𝐷 𝛿−2
Õ
2𝑘𝜀𝛾 −𝐷 𝛿−1 ≤

ℓ
≤ 2 𝑘 𝛾 𝛿 𝑛 ≤ 1−2𝜀𝛾 −𝐷 𝛿−1
ℓ=2
ℓ ℓ=2
2 2 −2𝐷 −2
≤8𝑘 𝜀 𝛾 𝛿 ,

𝜀𝑛

where we use the bound ℓ ≤ (𝜀𝑛) ℓ and sum the resulting geometric series. 

We now extend this to deal with intervals of any length, and replace the single
history in which the cover condition is not too likely to fail with a sufficiently large
collection of histories.

Lemma 4.5.4. Given 𝐷 ∈ N and 𝛾 > 0, let 𝑄, 𝛿, 𝛼0 , 𝛼𝑠max , 𝐶, 𝜀 be as in Set-

ting 4.3.4. Then the following holds for any 𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max and all sufficiently
large 𝑛. Suppose that 𝐺 is a hypergraph on [𝑛] such that deg− (𝑥) ≤ 𝐷 for each


𝑥 ∈ 𝑉 (𝐺), and 𝐻 is an (𝛼, 𝑄)-quasirandom hypergraph with 𝑛 vertices and 𝑝 𝑛𝑟
edges, with 𝑝 ≥ 𝛾. Let 0 ≤ 𝑡0 < 𝑡1 ≤ 𝑛 − 𝛿𝑛. Let L be a history ensemble
of RandomEmbedding up to time 𝑡 0 , and suppose that P(L ) ≥ 𝑛−4𝐶 . Then the
following hold for any fixed 1 ≤ 𝑘 ≤ 2𝐶 and set of vertices 𝑈 ⊆ 𝑉 (𝐻) with |𝑈| = 𝑘.
If |𝑈 ∩ im 𝜓𝑡0 | = 0 then we have

P(|𝑈 ∩ im 𝜓𝑡1 | = 0|L ) = (1 ± 100𝐶 𝑘𝛼𝛿−1 ) 𝑛−1−𝑡 1
𝑘

𝑛−𝑡0 .

We should note that the numerator 𝑛 − 1 − 𝑡1 might look strange, since 𝜓𝑡1 has
𝑛 − 𝑡1 unused vertices. The extra −1 is absorbed in the error term, and will make
for neater cancellation in a future lemma.

Proof. We divide the interval (𝑡0 , 𝑡1 ] into ℓ := d(𝑡1 − 𝑡0 )/𝜀𝑛e intervals, all but
the last of length 𝜀𝑛. Let L0 := L . Let, for each 1 ≤ 𝑖 < ℓ, the set L𝑖 be
the embedding histories up to time 𝑡0 + 𝑖𝜀𝑛 of RandomEmbedding which extend
histories in L𝑖−1 and are such that |𝑈 ∩ 𝜓𝑡0 +𝑖𝜀𝑛 | = 0. Let Lℓ be the embedding
histories up to time 𝑡1 extending those in Lℓ−1 such that |𝑈 ∩ 𝜓𝑡1 | = 0. Thus we
have
P(|𝑈 ∩ im 𝜓𝑡1 | = 0|L ) = P(Lℓ )/P(L0 ) .

101
 4 Packing Degenerate Hypergraphs

Finally, for each 1 ≤ 𝑖 ≤ ℓ, let the set L𝑖−1

0 consist of all histories in L
𝑖−1 such that
the (𝐶𝛼, 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡0 +(𝑖−1)𝜀𝑛 ) and the probability that the
(𝜀, 𝐶𝛼, 𝑡 0 + 1 + (𝑖 − 1)𝜀𝑛)-cover condition fails, conditioned on 𝜓𝑡0 +(𝑖−1)𝜀𝑛 , is at
most 𝑛−3 . In other words, L𝑖0 is the subset of L𝑖 consisting of typical histories,
satisfying the conditions of Lemma 4.5.1.
We now determine P(Lℓ ) in terms of P(L0 ), and in particular we show induct-
ively that P(L𝑖 ) > 𝑛−9𝐶/2 for each 𝑖. Observe that for any time 𝑡, the probability
(not conditioned on any embedding) that either the (𝐶𝛼, 𝑄)-diet condition fails for
(𝐻, im 𝜓 𝑗 ) for some 𝑗 ≤ 𝑡 or that the (𝜀, 𝐶𝛼, 𝑡 + 1)-cover condition has probability
greater than 𝑛−3 of failing, is at most 𝑛−5𝐶 by Lemma 4.4.3. In other words, for
each 𝑖 we have P(L𝑖 \ L𝑖0) ≤ 𝑛−5𝐶 . Thus by Lemma 4.5.1 we have

0
) ± 𝑛−5𝐶
𝜀𝑛


P(L𝑖 ) = 1 − 𝑘 (1 ± 10𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 P(L𝑖−1
P(L𝑖−1 ) ± 𝑛−5𝐶 ± 𝑛−5𝐶
𝜀𝑛



= 1 − 𝑘 (1 ± 10𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛
𝜀𝑛


= 1 − 𝑘 (1 ± 20𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 P(L𝑖−1 ) ,

where the final equality uses the lower bound P(L𝑖−1 ) ≥ 𝑛−9𝐶/2 . Similarly, using
the ‘furthermore’ statement of Lemma 4.5.1, we have


P(Lℓ ) = 1 ± 4𝜀𝛾 𝐷 𝛿𝑘 P(Lℓ−1 ) .

Putting these observations together, we can compute P(Lℓ ):

ℓ−1 
Ö 
𝐷
𝜀𝑛
P(Lℓ ) = 1 ± 4𝜀𝛾 𝛿𝑘 P(L0 ) 1 − 𝑘 (1 ± 20𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 .
𝑖=1

Observe that the approximation log(1 + 𝑥) = 𝑥 ± 𝑥 2 is valid for all sufficiently small
𝑥. In particular, since 𝑛 − 𝑡0 − (𝑖 − 1)𝜀𝑛 ≥ 𝑛 − 𝑡1 ≥ 𝛿𝑛 and by choice of 𝜀, for each
𝑖 we have
 
𝜀𝑛 𝜀𝑛
log 1 − 𝑘 (1 ± 20𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 = −𝑘 (1 ± 30𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 .

102
 4 Packing Degenerate Hypergraphs

Thus we obtain
ℓ−1
Õ
𝐷 𝜀𝑛
log P(Lℓ ) = log P(L0 ) ± 8𝜀𝛾 𝛿𝑘 − 𝑘 (1 ± 30𝐶𝛼) 𝑛−𝑡0 −(𝑖−1)𝜀𝑛
𝑖=1
∫ (ℓ−1)𝜀𝑛
𝐷 1
= log P(L0 ) ± 8𝜀𝛾 𝛿𝑘 − 𝑘 (1 ± 40𝐶𝛼) 𝑛−𝑡0 −𝑥 d𝑥
𝑥=0


= log P(L0 ) ± 8𝜀𝛾 𝐷 𝛿𝑘 − 𝑘 (1 ± 50𝐶𝛼) log(𝑛 − 𝑡0 ) − log(𝑛 − 1 − 𝑡 1 )
−1
= log P(L0 ) + 𝑘 log 𝑛−1−𝑡 1 𝐷
𝑛−𝑡 0 ± 8𝜀𝛾 𝛿𝑘 ± 50𝑘𝐶𝛼 log 𝛿 , (4.5.5)

where we use 𝑡1 ≤ 𝑛 − 𝛿𝑛, and we justify that the integral and sum are close by
observing that for each 𝑖 in the summation, if (𝑖 − 1)𝜀𝑛 ≤ 𝑥 ≤ 𝑖𝜀𝑛 then we have

1 1 1
𝑛−𝑡0 −𝑖𝜀𝑛 ≤ 𝑛−𝑡0 −𝑥 ≤ 𝑛−𝑡0 −(𝑖−1)𝜀𝑛 ≤ (1 + 𝛼) 𝑛−𝑡01−𝑖𝜀𝑛 ,

where the final inequality uses 𝑛 − 𝑡 0 − 𝑖𝜀𝑛 ≤ 𝑛 − 𝑡1 ≤ 𝛿𝑛 and the choice of 𝜀. By

choice of 𝜀, and since 𝛿−1 > log 𝛿−1 , this gives the lemma. Furthermore, (4.5.5),
and the fact 𝑡1 ≤ 𝑛−𝛿𝑛, imply that P(Lℓ ) ≥ 𝑛−9𝐶/2 . Since the L𝑖 form a decreasing
sequence of events the same bound holds for each L𝑖 .

Next, suppose that 𝐹 is a small labelled induced subhypergraph of 𝐺 whose

vertices are in [𝑛 − 𝛿𝑛], and suppose that we have a not necessarily induced copy
of 𝐹 in 𝐻, with 𝜚 : 𝑉 (𝐹) → 𝑉 (𝐻) the embedding witnessing this. Suppose in
addition 𝑍 is a small set of vertices of 𝐻 disjoint from 𝜚(𝑉 (𝐹)). We now estimate
the probability that RandomEmbedding embeds 𝐹 to 𝐹 ∗ , preserving the labels (i.e.
we produce an embedding extending 𝜚) and does not use any vertex of 𝑍. For the
moment, the reader should think of 𝐹 as being a single vertex.
The idea is the following. Suppose 𝐹 = {𝑢} is a single vertex. We embed 𝑢
to 𝜚(𝑢) if and only if, as we run through 𝑉 (𝐺), for each semi-edge 𝑆 ∈ N𝐺 − (𝑢),

we embed 𝑆 to a semi-edge which forms an edge of 𝐻 together with 𝜚(𝑢), all the
while not using the vertices 𝜚(𝑢) or 𝑍. We then embed 𝑢 to 𝜚(𝑢), and complete the
embedding without using 𝑍. The point of phrasing it like this is that we can estimate
accurately (using the diet condition) the probability of embedding 𝑆 ∈ N𝐺 − (𝑢) to

a semi-edge which forms an edge of 𝐻 with 𝜚(𝑢): wherever we embed the first
𝑟 − 2 vertices of 𝑆, what matters is where we embed the final vertex to complete
the embedding of 𝑆. So we can split up the embedding of 𝐺 into a collection of
intervals (in which we embed vertices which are neither 𝑢 nor in the set 𝑅(𝑢) of
maximum vertices of 𝑆) together with the embeddings of the vertices of 𝑅(𝑢) and

103
 4 Packing Degenerate Hypergraphs

finally of 𝑢. We can use Lemma 4.5.4 to estimate the chance that 𝜚(𝑢) or 𝑍 is
used in an interval. Finally, if we know that N𝐺− (𝑢) is embedded to a collection of

semi-edges which all form edges with 𝜚(𝑢) and 𝜚(𝑢) is not used when we come to
embed 𝑢, then we can also estimate accurately the probability of embedding 𝑢 to
𝜚(𝑢).
For more general 𝐹, there are two points to change in the above sketch. First,
there are several different vertices in 𝑉 (𝐹) and consequently we need to split the
embedding up into more intervals. Second, it is no longer necessarily the case that
for each 𝑆 ∈ N𝐺− (𝑢) the embedding of the last vertex is what matters. Rather, what

matters is the embedding of the last vertex of 𝑆 \ 𝑉 (𝐹). Note that if 𝑆 \ 𝑉 (𝐹) = ∅,
then 𝑆 ∪ {𝑢} is an edge of 𝐹.
In what follows, we also slightly extend the rôle of 𝑍: we allow for considering
only part of the run of RandomEmbedding up to some point 𝑛 − 𝛿𝑛, ˜ where 𝛿 ≤ 𝛿˜ ≤
1 − 𝛿.

Lemma 4.5.5. For each 𝑟 ≥ 2, 𝐷 ∈ N, and 𝛾 > 0, let constants 𝑄, 𝛿, 𝜀, 𝐶, 𝛼0 , 𝛼𝑠max

be as in Setting 4.3.4. Then the following holds for any 𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max and all
sufficiently large 𝑛. Suppose that 𝐺 is a hypergraph on [𝑛] such that deg− (𝑥) ≤ 𝐷
for each 𝑥 ∈ 𝑉 (𝐺), and 𝐻 is an (𝛼, 𝑄)-quasirandom hypergraph with 𝑛 vertices
and 𝑝 𝑛𝑟 edges, with 𝑝 ≥ 𝛾. Let 𝛿 ≤ 𝛿˜ ≤ 1− 𝛿. Let 𝐹 be an induced subhypergraph

˜
of 𝐺 with 0 ≤ |𝑉 (𝐹)| ≤ 𝐶, and suppose 𝑉 (𝐹) ⊆ [𝑛 − 𝛿𝑛]. Let 𝜚 : 𝑉 (𝐹) → 𝑉 (𝐻)
be an embedding of 𝐹. Finally let 𝑍 ⊆ 𝑉 (𝐻) \ 𝜚(𝑉 (𝐹)) be of size at most 𝐶. When
RandomEmbedding is run to embed 𝐺 into 𝐻, the probability that 𝜓𝑛−𝛿𝑛 ˜ extends
𝜚 and im 𝜓𝑛−𝛿𝑛 ˜ ∩ 𝑍 = ∅ is

(2+2𝐷)|𝑉 (𝐹)|+1
1 ± 200𝐶 (|𝑉 (𝐹)| + 2|𝑍 |)𝛼𝛿−1 𝑝 −|𝐸 (𝐹)| 𝑛−|𝑉 (𝐹)| 𝛿˜|𝑍 | .

Observe that the number of labelled copies of 𝐹 in 𝐻 is easily seen (by quasiran-
domness of 𝐻) to be roughly 𝑝 |𝐸 (𝐹)| 𝑛 |𝑉 (𝐹)| , so this lemma states that we embed 𝐹 in
𝐺 to a copy chosen (roughly) uniformly at random and this is roughly independent
of the set of vertices left unused.

Proof. If |𝑉 (𝐹)| = 0, the lemma statement follows directly from Lemma 4.5.4 with
˜ so we now assume |𝑉 (𝐹)| ≥ 1. Let 𝑉 (𝐹) = {𝑥 1 , . . . , 𝑥ℓ } in
𝑡 0 = 0 and 𝑡1 = 𝑛 − 𝛿𝑛,
increasing order.
We define relevant vertices in the following way: for each edge 𝑒 ∈ 𝐸 (𝐺) \ 𝐸 (𝐹)
whose final vertex is in 𝑉 (𝐹), we let the last vertex in 𝑒 \ 𝑉 (𝐹) be a relevant vertex.

104
 4 Packing Degenerate Hypergraphs

Let 𝑧1 , . . . , 𝑧 𝑘 be the increasing sequence of relevant vertices. Note that 𝑘 ≤ 𝐷ℓ as

each vertex of 𝐹 has at most 𝐷 left neighbours. Let 𝑆𝑖 denote the set of hyperedges


𝑒 ∈ 𝐸 (𝐺) such that max(𝑒) is in 𝑉 (𝐹), and 𝑧𝑖 = max 𝑒 \ 𝑉 (𝐹) . We say that a
partial embedding 𝜓 𝑧𝑖 of [1, . . . , 𝑧𝑖 ] is good for 𝑆𝑖 if for each hyperedge 𝑒 ∈ 𝑆𝑖 , the



set 𝜓 𝑧𝑖 𝑒 \ 𝑉 (𝐹) ∪ 𝜚 𝑒 ∩ 𝑉 (𝐹) is a hyperedge of 𝐻.
Let 𝑗0 = 0, and for each 1 ≤ 𝑖 ≤ ℓ − 1, let 𝑗𝑖 be the index such that 𝑧 𝑗𝑖 <
˜ is split by removing
 
𝑥𝑖 < 𝑧 𝑗𝑖+1 . Let 𝑗ℓ = 𝑘. Observe that the interval 1, 𝑛 − 𝛿𝑛
{𝑥 1 , . . . , 𝑥ℓ , 𝑧1 , . . . , 𝑧 𝑘 } into 𝑘 + ℓ + 1 intervals 𝐼1 , . . . , 𝐼 𝑘+ℓ+1 in increasing order.
For convenience, when two vertices in {𝑥 1 , . . . , 𝑥ℓ , 𝑧1 , . . . , 𝑧 𝑘 } are consecutive in
𝑉 (𝐺), we say there is an interval between them, but it contains zero vertices
(in order to justify having exactly 𝑘 + ℓ intervals). Thus the interval 𝐼1 is either
[1, 𝑧1 − 1] or [1, 𝑥1 − 1], depending on whether 𝑧1 < 𝑥1 or not, and so on, and the
final interval 𝐼 𝑘+ℓ+1 is the interval [max(𝑉 (𝐹)) + 1, 𝑛 − 𝛿𝑛] ˜ (or the empty interval
if max(𝑉 (𝐹)) = 𝑛 − 𝛿𝑛). ˜
We consider the run of RandomEmbedding split up in this way, with the various
intervals and the embeddings of the individual vertices {𝑥 1 , . . . , 𝑥ℓ , 𝑧1 , . . . , 𝑧 𝑘 }.
We obtain an increasing sequence 𝑡0 , . . . , 𝑡 2𝑘+2ℓ+1 of critical times, where 𝑡0 = 0,
𝑡1 = max(𝐼1 ) is the last vertex of 𝐼1 (or 0 if 𝐼1 = ∅), 𝑡2 is whichever is smaller from
𝑥 1 and 𝑧1 , and so on. In general, 𝑡2𝑖−1 will be the last vertex of 𝐼𝑖 (if this interval
is empty, we set 𝑡2𝑖−1 = 𝑡2𝑖−2 ) and 𝑡2𝑖 is the 𝑖th vertex of {𝑥 1 , . . . , 𝑥ℓ , 𝑧1 , . . . , 𝑧 𝑘 }.
Correspondingly, we define a nested collection of events L0 , . . . , L2𝑘+2ℓ+1 . The
first, L0 , is the sure event, and the last is the event that 𝜓𝑛−𝛿𝑛 ˜ extends 𝜚 and has
image disjoint from 𝑍 whose probability we want to estimate. In general, we let
L𝑖 be the event that after embedding 𝑡𝑖 it is not yet impossible to have the desired
event. More formally, L2𝑖−1 is the event that L2𝑖−2 occurs and furthermore no
vertex of 𝐼𝑖 was embedded to 𝜚 𝑉 (𝐹) ∪ 𝑍, and L2𝑖 is the event that L2𝑖−1 occurs

and furthermore, if 𝑡2𝑖−1 = 𝑥 𝑗 then 𝑥 𝑗 is embedded to 𝜚(𝑥 𝑗 ), while if 𝑡2𝑖−1 = 𝑧 𝑗 then

𝜓 𝑧 𝑗 is good for 𝑆 𝑗 and 𝑧 𝑗 is not embedded to any vertex of 𝜚 𝑉 (𝐹) ∪ 𝑍.
Since these events are nested, we have

2𝑘+2ℓ+1
Ö
P L2𝑘+2ℓ+1 = P[L0 ] P L𝑖 L𝑖−1 ,
   
(4.5.6)
𝑖=1

and we will see that we can estimate each of the conditional probabilities in
this product accurately. These estimates will depend on an assumption that
P[L𝑖 ] ≥ 𝑛−2𝐶 , which we will justify by establishing this bound for the smal-

105
 4 Packing Degenerate Hypergraphs

lest event L2𝑘+2ℓ+1 at the end of the proof. We first state as a claim the bounds
on the various P L𝑖 L𝑖−1 which we need: there are three types, depending on
 

whether 𝑡𝑖 is the end of an interval, or some 𝑥 𝑗 , or some 𝑧 𝑗 . Though the formulae are
somewhat complicated, we will see that most of the main terms end up cancelling.
We need one more definition: for each 𝑖, we let ℓ𝑖 := 𝑉 (𝐹) \ [1, 𝑡𝑖 ] be the number
of vertices of 𝐹 still to embed once 𝑡𝑖 is embedded.
Claim 4.5.6. If 𝑡𝑖 is the end of the interval 𝐼 𝑗 , then we have

  ℓ𝑖 +|𝑍 |
P L𝑖 L𝑖−1 = 1 ± 200𝐶 (ℓ𝑖 + |𝑍 |)𝛼𝛿−1 𝑛−1−𝑡
 
𝑛−𝑡𝑖−1
𝑖
. (4.5.7)

If 𝑡𝑖 = 𝑧 𝑗 , then we have

P L𝑖 L𝑖−1 = 1 ± 100𝐶ℓ𝑖 𝛼 𝑝 |𝑆 𝑗 | .
 

(4.5.8)

If 𝑡𝑖 = 𝑥 𝑗 , then we have

P L𝑖 L𝑖−1 = 1 ± 100𝐶𝛼 1
 

deg− ( 𝑥 𝑗 ) . (4.5.9)
𝑝 𝐺 (𝑛−𝑡𝑖 )

Before proving this claim, we use it to prove Lemma 4.5.5. We first point
out which terms cancel in the product (4.5.6). If 𝑡𝑖 is the end of the interval 𝐼 𝑗 ,
and so 𝑡𝑖+2 is the end of the interval 𝐼 𝑗+1 , observe that 𝑡𝑖+1 = 𝑡𝑖 + 1 and so the
numerator of the fraction in (4.5.7) for 𝑖 is the same as the denominator in (4.5.7)
for 𝑖 + 2. Thus these terms either cancel exactly (if ℓ𝑖 = ℓ𝑖+2 ), or with an extra factor
𝑛 − 1 − 𝑡𝑖 = 𝑛 − 𝑡𝑖+1 if ℓ𝑖 = ℓ𝑖+2 + 1. The latter case occurs exactly when 𝑡𝑖+1 = 𝑥 𝑗 ,
in which case the factor 𝑛 − 𝑡𝑖+1 cancels with the same factor in the denominator
of (4.5.9) for 𝑖 + 1. Putting these observations together, when we multiply together
all the terms from the various cases of (4.5.7) and of (4.5.9) we have a collection
−
of error terms, the term 𝑛−ℓ−|𝑍 | from the term P L1 L0 , a term ℓ𝑗=1 𝑝 − deg𝐺 (𝑥 𝑗 )
  Î

˜ − 1 |𝑍 | from the final case

from the various cases of (4.5.9), and finally a term 𝛿𝑛
𝑡 2𝑘+2ℓ+1 . The various terms from (4.5.8) simply give us a collection of error terms

106
 4 Packing Degenerate Hypergraphs

Î𝑘
and the term 𝑗=1 𝑚 𝑗 . Putting all this together, from (4.5.6) we have

2𝑘+2ℓ+1
!
Ö
P L2𝑘+2ℓ+1 = 1 · −1

𝑛−ℓ−|𝑍 |
 
1 ± 200𝐶 (ℓ𝑖 + |𝑍 |)𝛼𝛿
𝑖=1
ℓ 𝑘
− deg−𝐺 (𝑥 𝑗 )
Ö Ö
|𝑍 |
· 𝑝 𝑝 |𝑆 𝑗 | 𝛿𝑛
˜ −1
𝑗=1 𝑗=1
2𝑘+2ℓ+1
!
Ö
−1

=1· 1 ± 200𝐶 (ℓ𝑖 + 2|𝑍 |)𝛼𝛿 𝑛−ℓ
𝑖=1
ℓ 𝑘
−
Ö Ö
· 𝑝 − deg𝐺 (𝑥 𝑗 ) 𝑝 |𝑆 𝑗 | 𝛿˜|𝑍 | .
𝑗=1 𝑗=1

Observe now that if 𝑒 is an edge whose last vertex is in 𝑉 (𝐹), then it is counted
Í −
exactly once in ℓ𝑗=1 deg𝐺 (𝑥 𝑗 ), namely for 𝑥 𝑗 = max(𝑒). If 𝑒 \ 𝑉 (𝐹) is non-empty,
Í

it is also counted exactly once in 𝑘𝑗=1 |𝑆 𝑗 |, namely for 𝑧 𝑗 = max 𝑒 \ 𝑉 (𝐹) . If
Í
𝑒 \ 𝑉 (𝐹) = ∅ then 𝑒 is not counted in 𝑘𝑗=1 |𝑆 𝑗 |, and if 𝑒 is any edge whose
last vertex is not in 𝑉 (𝐹), then it is counted in neither sum. Thus we have
Íℓ − Í𝑘
𝑗=1 deg𝐺 (𝑥 𝑗 ) − 𝑗=1 |𝑆 𝑗 | = 𝑒(𝐹), and we get

2𝑘+2ℓ+1
!
Ö
P L2𝑘+2ℓ+1 = 1 · −1

𝑛−ℓ 𝑝 −𝑒(𝐹) 𝛿˜|𝑍 | ,
 
1 ± 200𝐶 (ℓ + 2|𝑍 |)𝛼𝛿
𝑖=1

which is as claimed in the lemma statement. What remains is to prove the Claim.

Proof of Claim 4.5.6. We begin with (4.5.7), which is given immediately by

Lemma 4.5.4 with 𝑈 = 𝜚 𝑉 (𝐹) \ [1, 𝑡𝑖−1 ] , by our assumption P[L𝑖−1 ] ≥ 𝑛−2𝐶 .

To prove the other two statements of this Claim, we will first consider only
histories in which the (𝐶𝛼, 𝑄)-diet condition holds for (𝐻, im 𝜓𝑡 ) for each time
𝑡, and the probability of the (𝜀, 𝐶𝛼, 𝑡)-cover condition failing for any 𝜓𝑡+𝜀𝑛−2 is
less than 𝑛−4 . Observe that by Lemma 4.4.3, the total probability of histories not
satisfying these conditions is at most 𝑛−4𝐶 , which we will see can be absorbed in
the error term.
We now prove (4.5.8). By the (𝐶𝛼, 𝑄)-diet condition, we embed 𝑧 𝑗 uniformly
−
at random to a set of size (1 ± 𝐶𝛼) 𝑝 deg𝐺 (𝑧 𝑗 ) (𝑛 − 𝑡𝑖 − 1). We need to estimate the


number of these vertices which are good for 𝑆 𝑗 and not in 𝜚 𝑉 (𝐹) . Observe that
𝑣 is in 𝐶 𝑧 𝑗 −1 (𝑧 𝑗 ) and good for 𝑆 𝑗 if and only if it is in the common neighbourhood

107
 4 Packing Degenerate Hypergraphs

− (𝑧 ), and also of all the semi-edges in the set

of the semi-edge N𝐺 𝑗

n

o
𝑊 := 𝜓𝑧 𝑗 − 1 𝑒 \ (𝑉 (𝐹) ∪ {𝑧 𝑗 }) ∪ 𝜚 𝑒 ∩ 𝑉 (𝐹) : 𝑒 ∈ 𝑆 𝑗 .

Observe that, since we are in L𝑖−1 , the elements of 𝑊 are sets each of 𝑟 − 1 vertices.
−
This collection of semi-edges has size deg𝐺 (𝑧 𝑗 ) + |𝑆 𝑗 |: that is, no two semi-edges
−
listed are equal. This is automatic for N𝐺 (𝑧 𝑗 ) and for 𝑊, because both of these sets
are obtained from sets of edges of 𝐺 by removing 𝑧 𝑗 , which is in all of them. So
we just need to check that no edge in N𝐺 − (𝑧 ) is also in {𝑆 \ {𝑧 } : 𝑆 ∈ 𝑆 }. Notice
𝑗 𝑗 𝑗
that every edge 𝑒 of 𝑆 𝑗 has by definition its largest vertex in 𝑉 (𝐹), and since 𝑧 𝑗 is
not in 𝑉 (𝐹) in particular the largest vertex of 𝑒 is not embedded by 𝜓 𝑧 𝑗 −1 , whereas
− (𝑧 ) are embedded by 𝜓 −
all vertices of each edge of N𝐺 𝑗 𝑧 𝑗 −1 . Thus 𝑒 is not in N𝐺 (𝑧 𝑗 )
as required.
Using the (𝐶𝛼, 𝑄)-diet condition with the set N𝐺 − (𝑧 )∪𝑊, we see that the number
𝑗
−
of vertices of 𝐶 𝑧 𝑗 −1 (𝑧 𝑗 )\im 𝜓 𝑧 𝑗 −1 which are good for 𝑆 𝐽 is (1±𝐶𝛼) 𝑝 deg𝐺 (𝑧 𝑗 )+|𝑆 𝑗 | (𝑛−
𝑡𝑖 − 1). Taking account finally of the histories in which the diet or cover condition
fails, we see
−
(1 ± 𝐶𝛼) 𝑝 deg𝐺 (𝑧 𝑗 )+|𝑆 𝑗 | (𝑛 − 𝑡𝑖 − 1)
P L𝑖 L𝑖−1 ± 𝑛−4𝐶 P[L𝑖−1 ] −1 ,
 
= −
1 ± 𝐶𝛼 𝑝 deg𝐺 (𝑧 𝑗 ) (𝑛 − 𝑡𝑖 + 1)

which by the assumption P[L𝑖−1 ] ≥ 𝑛−2𝐶 gives (4.5.8).

Finally, we prove (4.5.9). When we embed 𝑥 𝑗 , provided L𝑖−1 holds, we have
𝜚(𝑥 𝑗 ) ∉ im 𝜓𝑥 𝑗 −1 and 𝜚(𝑥 𝑗 ) ∈ 𝐶 𝑥 𝑗 −1 (𝑥 𝑗 ). Thus we simply need to estimate the
size of |𝐶 𝑥 𝑗 −1 (𝑥 𝑗 ) \ im 𝜓𝑡𝑖 −1 | (to which we embed 𝑥 𝑗 uniformly), which we can do
−
using the diet condition: it has size (1 ± 𝐶𝛼) 𝑝 deg𝐺 (𝑥 𝑗 ) (𝑛 − 𝑡𝑖 + 1). Taking account
of the possibility of the cover or diet conditions failing, we get

1
P L𝑖 L𝑖−1 = ± 𝑛−4𝐶 P[L𝑖−1 ] −1 ,
 
−
𝑝 deg𝐺 (𝑥 𝑗 ) (𝑛


1 ± 𝐶𝛼 − 𝑡𝑖 + 1)

which by the assumption P[L𝑖−1 ] ≥ 𝑛−2𝐶 gives (4.5.9). 

The immediate case of this lemma which we need is the case that 𝐹 is an edge
of 𝐺 contained in [𝑛 − 𝛿𝑛] and 𝑍 = ∅. We use this to estimate the probability
that, again for fixed 𝐺 and 𝐻, at least one edge in a given 1-star in 𝐻 is used by
RandomEmbedding.

108
 4 Packing Degenerate Hypergraphs

Lemma 4.5.7. For each 𝑟 ≥ 2, 𝐷 ∈ N, and 𝛾 > 0, let the constants

𝑄, 𝛿, 𝜀, 𝛼0 , 𝛼𝑠max , 𝐶 be as in Setting 4.3.4. Then the following holds for any
𝛼0 ≤ 𝛼 ≤ 𝛼𝑠max and all sufficiently large 𝑛. Suppose that 𝐺 is a hypergraph
on [𝑛] such that deg− (𝑥) ≤ 𝐷 for each 𝑥 ∈ 𝑉 (𝐺), with at least 𝑛/4 edges and
maximum degree Δ(𝐺) ≤ 𝑛/log 𝑛, and 𝐻 is an (𝛼, 𝑄)-quasirandom hypergraph


with 𝑛 vertices and 𝑝 𝑛𝑟 edges, where 𝑝 ≥ 𝛾. Let 𝑢 1 , . . . , 𝑢 𝑘 be semi-edges and 𝑣
a vertex of 𝐻 for some 𝑘 ≤ 𝑄, and suppose 𝑢𝑖 𝑣 is an edge of 𝐻 for each 𝑖. When
RandomEmbedding is run to embed 𝐺 into 𝐻, the probability that there is at least
one 𝑢𝑖 𝑣 to which some edge of 𝐺 is embedded is

2(𝐷+1)𝑟+1
1 ± 1000𝐶𝑟𝛼𝛿−1 𝑝 −1 𝑛−𝑟 · 𝑟!𝑘𝑒(𝐺 0) ,

where 𝐺 0 is the subhypergraph of 𝐺 induced by [𝑛 − 𝛿𝑛].

Proof. Given 𝑢 1 , . . . , 𝑢 𝑘 , 𝑣 and 𝐺 and 𝐻, let 𝑆 be the event that there is at least one
𝑢𝑖 𝑣 to which some edge of 𝐺 is embedded.
Fix any given 𝑢𝑖 𝑣. By Lemma 4.5.5, the probability that any one edge of 𝐺 0
is embedded to 𝑢𝑖 𝑣 in any given order is (1 ± 200𝐶𝑟𝛼𝛿−1 ) 2(𝐷+1)𝑟+1 𝑝 −1 𝑛−𝑟 . Since
at most one edge in at most one order is so embedded, these events are disjoint,
and summing over all edges of 𝐺 0 and all of the 𝑟! possible orders, we see that the
probability that some edge of 𝐺 0 is embedded to 𝑢𝑖 𝑣 in some order is

2(𝐷+1)𝑟+1
1 ± 200𝐶𝑟𝛼𝛿−1 𝑝 −1 𝑛−𝑟 · 𝑟!𝑒(𝐺 0) .

By linearity of expectation, the expected number of edges 𝑢𝑖 𝑣 embedded to by

RandomEmbedding is, by Lemma 4.5.5 and linearity of expectation,

2(𝐷+1)𝑟+1
𝐸 := 1 ± 200𝐶𝑟𝛼𝛿−1 𝑝 −1 𝑛−𝑟 · 𝑟!𝐸 (𝐺 0) · 𝑘 ,

and by inclusion-exclusion, we have

Õ  
𝐸− P 𝑢𝑖 𝑣 and 𝑢𝑖 0 𝑣 are embedded to by RandomEmbedding ≤ P[𝑆] ≤ 𝐸 .
1≤𝑖<𝑖 0 ≤𝑘

We thus simply have to show that the above sum, which has 2𝑘 ≤ 𝑄2 terms,
is small. We will show that the probability of RandomEmbedding embedding to
any two fixed edges 𝑢𝑣, 𝑢0𝑣 is small. This probability is equal to the sum over
triples 𝑥, 𝑥 0 ∈ 𝑆(𝐺), 𝑦 ∈ 𝑉 (𝐺) such that 𝑥𝑦, 𝑥 0 𝑦 ∈ 𝐸 (𝐺) of the probability that

109
 4 Packing Degenerate Hypergraphs

𝑥 ↩→ 𝑢, 𝑥 0 ↩→ 𝑢0 and 𝑦 ↩→ 𝑣. For any given 𝑦 ∈ 𝑉 (𝐺) there are at most deg𝐺 (𝑦) 2
choices of (𝑥, 𝑥 0). If 𝑟 = 2 then by Lemma 4.2.3, there are at most 2𝐷𝑛Δ(𝐺) such
triples. If 𝑟 > 2 we use the trivial upper bound (𝑟𝑒(𝐺)) 2 . It is now enough to
make the estimate for one such triple. Assuming the (𝐶𝛼, 𝑄)-diet condition holds
throughout RandomEmbedding, we embed each vertex of 𝑥, 𝑥 0 and 𝑦 uniformly
at random into a set of size at least 12 𝑝 𝐷 𝛿𝑛 ≥ 12 𝛾 𝐷 𝛿𝑛, so the probability of the
event 𝑥 ↩→ 𝑢, 𝑥 0 ↩→ 𝑢0, 𝑦 ↩→ 𝑣 is at most 8𝛾 −(2𝑟−1)𝐷 𝛿−(2𝑟−1) 𝑛−(2𝑟−1) . Finally,
the probability of the (𝐶𝛼, 𝑄)-diet condition failing for some (𝐻, im 𝜓𝑖 ) is by
Lemma 4.4.3 at most 𝑛−5𝐶 . Putting this together, we have

2(𝐷+1)𝑟+1 −1 −𝑟
P[𝑆] = 1 ± 200𝐶𝑟𝛼𝛿−1 𝑝 𝑛 · 𝑟!𝑘𝑒(𝐺 0)
Õ
deg𝐺 (𝑦) 2 · 8𝛾 −(2𝑟−1)𝐷 𝛿−(2𝑟−1) 𝑛−(2𝑟−1) ± 𝑛−5𝐶 . (4.5.10)


± 𝑄2
𝑦∈𝑉 (𝐺)

Because 𝑒(𝐺 0) ≥ 𝑛/(2𝑟) the first term in the above is Θ(𝑛−(𝑟−1) ), while since
Δ(𝐺) ≤ 𝑛/log 𝑛 the other two terms are of asymptotically smaller order. Since 𝑛
is sufficiently large, this gives the desired result.

Finally, we use Lemma 4.5.7 to control the way in which subsets of common
neighbourhoods shrink during the embedding process. This in particular is what
we need to prove Lemma 4.3.7, though we will use the more general statement
again to prove Theorem 4.7.1.

Lemma 4.5.8. For each 𝑟 ≥ 2, 𝐷 ∈ N and 𝛾 > 0, let the constants

𝑄, 𝛿, 𝜀, 𝛼0 , 𝛼2𝑛 , 𝐶 be as in Setting 4.3.4. Then the following holds for any
0 ≤ 𝑠 ≤ 𝑠0 ≤ 𝑠∗ . Suppose that when PackingProcess is run up to and in-
cluding stage 𝑠. Let 𝑘 ≤ 𝑄, and let 𝑢 1 , . . . , 𝑢 𝑘 be semi-edges of 𝐻𝑠 . Fix any
𝑋 ⊆ N𝐻𝑠 (𝑢 1 , . . . , 𝑢 𝑘 ) satisfying |𝑋 | ≥ 𝜀𝑛. Then with probability at least 1 − 𝑛−5𝐶 ,
when PackingProcess is run further up to and including stage 𝑠0, if it does not fail
and if 𝐻𝑖 is (𝛼𝑖 , 𝑄)-quasirandom for each 𝑠 ≤ 𝑖 ≤ 𝑠0, we have

 𝑝𝑠0  𝑘
𝑋 ∩ N𝐻𝑠 0 (𝑢 1 , . . . , 𝑢 𝑘 ) = 1 ± 21 𝛼𝑠 0 𝑝𝑠 |𝑋 | .

Proof. We claim that, for a given 𝑢 1 , . . . , 𝑢 𝑘 , 𝑠 and 𝑋, the probability of 𝑠0 being

the first time after 𝑠 at which the given equation fails is at most 1 − 𝑛−6𝐶 , and by
taking a union bound over the at most 𝑠max choices of 𝑠0 we conclude the lemma
statement.

110
 4 Packing Degenerate Hypergraphs

For each 𝑠 ≤ 𝑖 ≤ 𝑠0, let 𝑋𝑖 := 𝑋 ∩ N𝐻𝑖 (𝑢 1 , . . . , 𝑢 𝑘 ), and for each 𝑠 + 1 ≤ 𝑖 ≤ 𝑠0

let 𝑌𝑖 = 𝑋𝑖 \ 𝑋𝑖−1 . Thus 𝑋𝑠 = 𝑋, and we want to know how much smaller 𝑋𝑠 0 is
than 𝑋𝑠 . This quantity is simply 𝑌𝑠+1 + · · · + 𝑌𝑠 0 . We aim to apply Corollary 4.2.2
with E being the event that after each stage 𝑖 with 𝑠 ≤ 𝑖 ≤ 𝑠0, the hypergraph 𝐻𝑖 is
(𝛼𝑖 , 𝑄)-quasirandom, and we have

 𝑝𝑖  𝑘
|𝑋𝑖 | = (1 ± 12 𝛼𝑖 𝑝𝑠 |𝑋 | .

By definition, the failure event of the above claim is contained in E, and so it suffices
to estimate the probability that E occurs and the given equation does not hold. We
therefore assume in the estimates which follow that we are in E. The probability
space in which we work is the set of all possible histories of RandomEmbedding,
and the sequence of partitions required by Corollary 4.2.2 is given by the histories
up to increasing times 𝑠 ≤ 𝑖 ≤ 𝑠0 of RandomEmbedding.
For each 𝑖, let 𝐺 0𝑖 be the subhypergraph of 𝐺 𝑖 induced by [𝑛 − 𝛿𝑛], and let 𝑝𝑖
be such that 𝑝𝑖 𝑛𝑟 = 𝑒(𝐻𝑖 ) = 𝑒(𝐻0 ) − 𝑖𝑗=1 𝑒(𝐺 0𝑗 ). Then by Lemma 4.5.7 and

Í

linearity of expectation, we have

 𝑝𝑖−1  𝑘
(2𝐷+2)𝑟+1
E[𝑌𝑖 |𝐻𝑖−1 ] =(1 ± 12 𝛼𝑖−1 𝑝𝑠 |𝑋 | 1 ± 1000𝐶𝑟𝛼𝑖−1 𝛿−1
−1 −𝑟
· 𝑝𝑖−1 𝑛 𝑟!𝑘𝑒(𝐺 0𝑖 )

(2𝐷+2)𝑟+1
= 1 ± 2000𝐶𝑟𝛼𝑖−1 𝛿−1 𝑘−1 −𝑘 −𝑟
|𝑋 | 𝑝𝑖−1 𝑝 𝑠 𝑛 𝑟!𝑘𝑒(𝐺 0𝑖 ) .

Í𝑠 0
We now need to estimate the sum 𝑖=𝑠+1 E(𝑌𝑖 |𝐻𝑖−1 ), on the assumption that each
𝐻𝑖−1 is (𝛼𝑖−1 , 𝑄)-quasirandom. We first estimate the sum of the main terms, using
the facts that 𝑒(𝐺 0𝑖 ) = ( 𝑝𝑖−1 − 𝑝𝑖 ) 𝑛𝑟 , that 𝑖 ≤ 𝑠max , that for any 𝑥 ∈ [0, 1], any

0 < ℎ < 1 and any integer 𝑎 ≥ 1 we have (𝑥 + ℎ) 𝑎 − 𝑥 𝑎 = 𝑎ℎ𝑥 𝑎−1 ± 2𝑎 ℎ2 , and that
𝑝𝑖−1 − 𝑝𝑖 ≤ 4𝐷𝑛1−𝑟 :

𝑠0
Õ 𝑠0
Õ
𝑘−1 −𝑘 −𝑟
|𝑋 | 𝑝𝑖−1 𝑝 𝑠 𝑛 𝑟!𝑘𝑒(𝐺 0𝑖 ) = 𝑘−1
𝑘 𝑝𝑖−1 ( 𝑝𝑖−1 − 𝑝𝑖 ) 𝑝 −𝑘 −1 𝑟
𝑠 |𝑋 |(1 ± 𝑟𝑛 )
𝑖=𝑠+1 𝑖=𝑠+1
𝑠0 
Õ 
2 −1
) 𝑝 −𝑘 − 𝑝𝑖𝑘 ± 16𝐷 2 2 𝑘 𝑛2−2𝑟
𝑘

= |𝑋 |(1 ± 2𝑟 𝑛 𝑠 𝑝𝑖−1
𝑖=𝑠+1

= |𝑋 | 𝑝 −𝑘 𝑘 𝑘
2 2 𝑘 −1
−𝑘
𝑠 𝑝 𝑠 − 𝑝 𝑠 0 ± 64𝐷 𝑟 2 𝑛 |𝑋 | 𝑝 𝑠 .

111
 4 Packing Degenerate Hypergraphs

Next, we estimate the sum of the error terms:

𝑠0
Õ
106𝐶𝐷 2 𝑄𝑟 2𝑟!𝛿−1 𝛼𝑖−1 𝑝𝑖−1
𝑘−1 −𝑘
𝑝 𝑠 𝑒(𝐺 0𝑖 )𝑛−𝑟 |𝑋 | ≤
𝑖=𝑠+1
∫ 𝑠0
≤ 107𝐶𝐷 3 𝑄𝑟 2𝑟!𝛿−1 𝛾 −2𝑄 𝑛1−𝑟 |𝑋 |𝛼𝑥 d𝑥
−∞
1 0 𝑄
≤ 8 𝛼𝑠 𝛾 |𝑋 | ,

where we use 𝑒(𝐺 0𝑖 ) ≤ 𝐷𝑛 and the final inequality is by (4.3.1). Putting these two
estimates together, we have

𝑠0
Õ
E(𝑌𝑖 |𝐻𝑖−1 ) = 𝑝 𝑠𝑘 − 𝑝 𝑠𝑘0 𝑝 −𝑘 1

𝑄
𝑠 |𝑋 | ± 4 𝛼𝑠 0 𝛾 |𝑋 | := 𝜇˜ ± 𝜈˜ .
𝑖=𝑠+1

𝑐𝑄𝑛
The range of each 𝑌𝑖 is at most 𝑘Δ(𝐺 𝑖 ) ≤ log 𝑛 . We apply Corollary 4.2.2 with
𝜚˜ = 𝜀𝛾 𝑄 |𝑋 | and E as above. We obtain that the probability that

𝑠0
Õ
𝑌𝑖 ≠ 𝑝 𝑠𝑘 −𝑝 𝑠𝑘0 𝑝 −𝑘 𝑘
−𝑘 1

𝑄 𝑄 𝑘 𝑄
𝑠 |𝑋 |±(𝛼 𝑠 0 𝛾 |𝑋 |/4+𝜀𝛾 |𝑋 |) = 𝑝 −𝑝 0 𝑝
𝑠 𝑠 𝑠 |𝑋 |± 2 𝛼𝑠 0 𝛾 |𝑋 |
𝑖=𝑠+1

is at most
 −𝜀 2 𝛾 𝑄 |𝑋 | 2 
2 exp 𝑐𝑄𝑛
< 𝑛−6𝐶 ,
2 log 𝑛 ( 𝜇˜ + 𝜈˜ + 𝜚)
˜
where the last inequality is by choice of |𝑋 | and 𝑐.
Observe that if the above likely event occurs, we have
 𝑝 0 𝑘
|𝑋𝑠 0 | = |𝑋 | − 𝑝 𝑠𝑘 − 𝑝 𝑠𝑘0 𝑝 −𝑘 1 0 𝑄
|𝑋 | ± 12 𝛼𝑠 0 𝛾 𝑄 |𝑋 | ,

𝑠
𝑠 |𝑋 | ± 𝛼
2 𝑠 𝛾 |𝑋 | =
𝑝𝑠

which implies the desired likely equation of the lemma. Taking the union bound
over the at most 𝑠max choices of 𝑠0 we obtain the lemma statement.

We are now in a position to prove Lemma 4.3.7.

Proof of Lemma 4.3.7. We define 𝑝ˆ by 𝑒(𝐻0∗ ) = 𝑝ˆ 𝑛𝑟 . By assumption we have

either 𝑝ˆ = (1 ± 𝜂)𝛾 or 𝑝ˆ = 0.
Our aim is to show that with high probability, for a given 𝑠, either PackingProcess
fails before completing stage 𝑠 or the pair (𝐻𝑠 , 𝐻0∗ ) is (𝛼𝑠 , 𝑄)-coquasirandom. Let
𝑆 be a set of at most 𝑄 vertices in 𝑉 (𝐻0∗ ), and let 𝑅 ⊆ 𝑆. Suppose that these two

112
 4 Packing Degenerate Hypergraphs

sets witness a first failure at some time 𝑠0 of (𝛼𝑠 0 , 𝑄)-coquasirandomness. Recall

that for (𝐻𝑠 0 , 𝐻0∗ ) to be (𝛼𝑠 0 , 𝑄)-coquasirandom means that N𝐻𝑠 0 (𝑅) ∩ N𝐻0∗ (𝑆 \ 𝑅)
has about the size one would expect if both hypergraphs were random. Observe
that if 𝑝ˆ = 0 and 𝑆 \ 𝑅 ≠ ∅ then N𝐻0∗ (𝑆 \ 𝑅) is empty, and consequently there is
nothing to check. We therefore assume that either 𝑝ˆ = (1 ± 𝜂)𝛾 or 𝑆 \ 𝑅 = ∅ in
what follows (and in the latter case we interpret the expression 𝑝ˆ |𝑆\𝑅| as 1).
We apply Lemma 4.5.8 with 𝑠 = 0, with 𝑠0 as above, with 𝑢 1 , . . . , 𝑢 𝑘 the
semi-edges of 𝑅, and with 𝑋 = N𝐻0 (𝑅) ∩ N𝐻0∗ (𝑆 \ 𝑅). Since (𝐻0 , 𝐻0∗ ) is ( 14 𝛼0 , 𝑄)-
coquasirandom, this set 𝑋 satisfies |𝑋 | ≥ 𝜀𝑛 as required. We see that with
probability at least 1 − 𝑛−5𝐶 we have

1 0

 𝑝 𝑠 0  |𝑅|
N𝐻𝑠 0 (𝑅) ∩ N 𝐻0∗ (𝑆 \ 𝑅) = 1 ± 2 𝛼𝑠 𝑝0 |𝑋 |
1 0

1

 𝑝 𝑠 0  |𝑅| |𝑅| |𝑆\𝑅|
= 1 ± 2 𝛼𝑠 1 ± 4 𝛼0 𝑝0 𝑝 0 𝑝ˆ0 𝑛

= 1 ± 𝛼𝑠 0 𝑝 𝑠|𝑅| ˆ0|𝑆\𝑅| 𝑛 ,


0 𝑝

which precisely says that we do not have a witness to a failure of (𝛼𝑠 0 , 𝑄)-
coquasirandomness of (𝐻𝑠 0 , 𝐻0∗ ).
Taking the union bound over all choices of 𝑅 ⊆ 𝑆 and 𝑆 of size at most 𝑄,
and applying Lemma 4.4.3, we see that the following event has probability at
most 𝑛2𝑟𝑄−5𝐶 . The pair (𝐻𝑖 , 𝐻0∗ ) is (𝛼𝑖 , 𝑄)-coquasirandom for each 0 ≤ 𝑖 ≤
𝑠 − 1, but either RandomEmbedding fails to embed 𝐺 𝑠 or (𝐻𝑠 , 𝐻0∗ ) is not (𝛼𝑠 , 𝑄)-
coquasirandom. Taking now the union bound over all choices of 1 ≤ 𝑠 ≤ 𝑠max , and
recalling that (𝐻0 , 𝐻0∗ ) is by assumption 14 𝛼0 , 𝑄 -coquasirandom, we conclude

that the probability that for any 1 ≤ 𝑠 ≤ 𝑠max , RandomEmbedding fails to embed
𝐺 𝑠 or the pair (𝐻𝑠 , 𝐻0∗ ) fails to be (𝛼𝑠 , 𝑄)-coquasirandom is at most 𝑛−4𝐶 . This
completes the proof.

4.6 Completing spanning embeddings

Recall that we complete the embedding of each hypergraph 𝐺 𝑠 by embedding the
∗ . From Setting 4.3.4, these vertices form
final 𝛿𝑛 vertices using only edges of 𝐻𝑠−1
a strongly independent set and each of them has degree 𝑑 𝑠 . Lemma 4.3.9 states
that it is very likely, provided PackingProcess does not fail and provided (𝐻𝑠 , 𝐻0∗ )
is coquasirandom for each 𝑠, that only a few edges of 𝐻0∗ are used at any given
semi-edge to form 𝐻𝑠∗ , and hence (𝐻𝑠 , 𝐻𝑠∗ ) is also coquasirandom. Complementing

113
 4 Packing Degenerate Hypergraphs

this, Lemma 4.6.1 states that this coquasirandomness guarantees that completing
the embedding is possible. We prove these two lemmas in this section, beginning
with Lemma 4.3.10.
Recall that Lemma 4.3.10 states that it is likely that the partial embedding 𝜑 𝑠
of each 𝐺 𝑠 provided by RandomEmbedding can be extended to an embedding 𝜑∗𝑠
of 𝐺 𝑠 , with the completion edges used for the extension lying in 𝐻 ∗ . Since the
neighbours of each of the last 𝛿𝑛 vertices of 𝐺 𝑠 are embedded by 𝜑 𝑠 , the set of
candidate vertices

𝐶𝑠∗ (𝑥) := 𝑣 ∈ 𝑉 (𝐻𝑠−1

∗

) \ im 𝜑 𝑠 : 𝜑 𝑠 (𝑦) ∈ N𝐻𝑠−1
∗ (𝑣) for each 𝑦 ∈ N𝐺 (𝑥)
𝑠

for each 𝑥 of these last 𝛿𝑛 vertices in 𝑉 (𝐻𝑠−1 ∗ ) \ im 𝜑 are already fixed, and the
𝑠
∗
desired 𝜑 𝑠 exists if and only if there is a system of distinct representatives for the
𝐶𝑠∗ (𝑥) as 𝑥 ranges over the last 𝛿𝑛 vertices of 𝐺 𝑠 . Recall that Lemma 4.4.3 states in
particular that (𝐻 ∗ , im 𝜑 𝑠 ) is likely to satisfy the (2𝜂, 𝑄)-codiet condition, which
implies both that 𝐶𝑠∗ (𝑥) is of size roughly 𝑝 𝑑𝑠 𝛿𝑛 for each of these last 𝑥, and also
that the collection of sets is well-distributed (in a sense we will make precise later).
We would like to apply Lemma 4.2.6, for which we need to prove that our bipartite
graph is super-regular. The codiet condition provides the degree and codegree
condition on one half of the graph, and the following lemma provides the other
half.

Lemma 4.6.1. For each 𝑟 ≥ 2, 𝐷 ∈ N and 𝛾 > 0, let the constants be as in

Setting 4.3.4. Suppose that 𝐺 is a hypergraph on vertex set [𝑛], with deg− (𝑥) ≤ 𝐷
for each 𝑥 ∈ 𝑉 (𝐺), with maximum degree at most 𝑐𝑛/log 𝑛 and whose last 𝛿𝑛
vertices all have degree 𝑑, where 0 ≤ 𝑑 ≤ 𝐷, and form a strongly independent set.


Suppose that 𝐻 is an (𝛼𝑠∗ , 𝑄)-quasirandom 𝑛-vertex hypergraph with 𝑝 𝑛𝑟 edges,
where 𝑝 ≥ 𝛾, and that 𝐻 ∗ is a hypergraph on 𝑉 (𝐻) with (1 ± 𝜂)𝛾 𝑛𝑟 edges such

that (𝐻, 𝐻 ∗ ) forms an (𝜂, 𝑄)-coquasirandom pair. When RandomEmbedding is

run to embed the first 𝑛 − 𝛿𝑛 vertices of 𝐺 into 𝐻, with probability at least 1 − 𝑛−4𝐶
we have that for all 𝑣 ∈ 𝑉 (𝐻 ∗ ) \ im 𝜓𝑛−𝛿𝑛

𝑥 ∈ 𝑉 (𝐺) : 𝑛 − 𝛿𝑛 < 𝑥 ≤ 𝑛, 𝜓𝑛−𝛿𝑛 ( N− (𝑥)) ⊆ N𝐻 ∗ (𝑣)


= (1 ± 10𝐷𝜂)𝛾 𝑑 𝛿𝑛 .

The proof of this lemma is similar to the proof of Lemma 4.4.5. We use (ℓ, 𝑑)-
patterns as described before the proof of Lemma 4.4.5.

114
 4 Packing Degenerate Hypergraphs

Proof. Fix 𝑣 ∈ 𝑉 (𝐻 ∗ ), fix 1 ≤ ℓ ≤ 𝑑 and an (ℓ, 𝑑)-pattern a. Let 𝑠 𝑘 denote

Í𝑘
𝑖=1 𝑎 𝑖 . Let 𝐼 be the last 𝛿𝑛 vertices of 𝐺, which by assumption form a strongly
independent set. Let 𝑋a denote the set of vertices from the last 𝛿𝑛 vertices of 𝐺
that have associated pattern a. For 𝑥 ∈ 𝑋a let 𝑅(𝑥) denote the set of vertices 𝑡, for
which there is an edge 𝑒 ∈ 𝐺 such that 𝑥 is the last vertex in it and 𝑡 is second to
last. Let for each such 𝑡 the set 𝐸 𝑡,𝑥 be the set of edges that caused 𝑡 to be in 𝑅(𝑥).
Given 𝑡 ∈ 𝑅(𝑥) and an embedding 𝜓𝑡−1 of {1, . . . , 𝑡 − 1}, we say the correct set of
𝑡 is
© Ù
ª
𝐴𝑥,𝑡 = ­ N𝐻 ∗ 𝜓𝑡−1 (𝑒 \ {𝑡, 𝑥}) ∪ {𝑣} ® ∩ (𝑉 (𝐻) \ im 𝜓𝑡−1 ) .
«𝑒∈𝐸 𝑥,𝑡 ¬
In other words, if at time 𝑡 − 1 the embedded part of N− (𝑥) has been embedded to
N𝐻 ∗ (𝑣), then 𝐴𝑥,𝑡 is the set of vertices to which we can embed 𝑡 and maintain this
property.
Given 𝑥 among the last 𝛿𝑛 vertices of 𝐺 and 1 ≤ 𝑘 ≤ ℓ, suppose that 𝑡 is the
𝑘th vertex of 𝑅(𝑥). We let N−𝑘 (𝑥) denote the set of semi-edges in N (𝑥) which are
contained in [𝑡]. Let Y𝑘 denote the event that N−𝑘 (𝑥) is embedded to N𝐻 ∗ (𝑣) for
about as many 𝑥 as one would expect. More formally, Y𝑘 is the event that

𝑥 ∈ 𝐼 : 𝜓𝑛−𝛿𝑛 N−𝑘 (𝑥) ⊆ N𝐻 ∗ (𝑣)



= (1 ± 10𝑘𝜂)𝛾 𝑠 𝑘 𝛿𝑛 . (4.6.1)

Observe that what we want is to show that Yℓ holds with sufficiently high probability
for each choice of 𝑣 ∉ im 𝜓𝑛−𝛿𝑛 , since 𝑠ℓ = 𝑑 and ℓ ≤ 𝐷.
Let B be the event that the (2𝜂, 𝑄)-codiet condition fails at some time 𝑡 ≤ 𝑛 − 𝛿𝑛.
We define the Bernoulli random variable 𝑍 𝑘,𝑥,𝑡 to be equal to 1 if 𝑡 is the 𝑘th
vertex of 𝑅(𝑥) and we have 𝜓𝑛−𝛿𝑛 N−𝑘−1 (𝑥) ⊆ N𝐻 ∗ (𝑣), and zero otherwise. Note

that 𝑍 𝑘,𝑥,𝑡 depends only on 𝜓𝑡−1 and not the later part of the embedding. Let
𝑌𝑘,𝑥,𝑡 := 𝑍 𝑘,𝑥,𝑡 · 1𝜓𝑛− 𝛿𝑛 (𝑡)∈𝐴 𝑥,𝑡 . That is, 𝑍 𝑘,𝑥,𝑡 tells us before embedding 𝑡 if we have a
chance for 𝑥 to contribute to the set in (4.6.1), and 𝑌𝑘,𝑥,𝑡 tells us if, having embedded
𝑡, it does contribute or not.
We want to show that if Y𝑘−1 occurs, then Y𝑘 is very likely to occur. We will
then show this implies the lemma. Observe that Y𝑘 is the event that

𝑛−𝛿𝑛
Õ 𝑛
Õ
𝑌𝑘,𝑥,𝑡 = (1 ± 10𝑘𝜂)𝛾 𝑠 𝑘 𝛿𝑛 .
𝑡=1 𝑥=𝑛−𝛿𝑛+1

Í𝑛−𝛿𝑛 Í𝑛
Furthermore, Y𝑘−1 implies that 𝑡=1 𝑥=𝑛−𝛿𝑛+1 𝑍 𝑘,𝑥,𝑡 = (1 ± 10(𝑘 − 1)𝜂)𝛾 𝑠 𝑘−1 𝛿𝑛.

115
 4 Packing Degenerate Hypergraphs

𝑥=𝑛−𝛿𝑛+1 E(𝑌𝑘,𝑥,𝑡 H𝑡−1 ), where H𝑡−1 denotes

Í𝑛−𝛿𝑛 Í𝑛
We would like to calculate 𝑡=1
the embedding history of RandomEmbedding up to and including embedding 𝑡 − 1.
Given a time 𝑡, if 𝑡 is the 𝑘th vertex of 𝑅(𝑥), then at time 𝑡 − 1 the first 𝑘 − 1
vertices of 𝑅(𝑥) have already
 been embedded,
 so 𝑍 𝑘,𝑥,𝑡 is determined. Thus we
have E(𝑌𝑘,𝑥,𝑡 H𝑡−1 ) = P 𝜓𝑡 (𝑡) ∈ 𝐴𝑥,𝑡 H𝑡−1 · 𝑍 𝑘,𝑥,𝑡 . Suppose that at time 𝑡 − 1 we
have not seen a witness that B fails. If 𝑣 ∈ im 𝜓𝑡−1 , then 𝑣 ∈ im 𝜓𝑛−𝛿𝑛 and there
is nothing to prove. So we may suppose 𝑣 ∉ im 𝜓𝑡−1 . Using the (2𝜂, 𝑄)-codiet
condition once with 𝑆 = N− (𝑡) ∪ {𝜓𝑡−1 (𝑒 \ {𝑡, 𝑥}) ∪ {𝑣} : 𝑒 ∈ 𝐸 𝑥,𝑡 } (these two sets
are disjoint since 𝑣 ∉ im 𝜓𝑡−1 ) and 𝑅 = N− (𝑡) ⊆ 𝑆 and once with 𝑆 = 𝑅 = N− (𝑡),
we obtain
  (1 ± 2𝜂)(1 ± 𝜂)𝛾 𝑎 𝑘 𝑝 |N− (𝑡)| (𝑛 − 𝑡 + 1)
P 𝜓𝑡 (𝑡) ∈ 𝐴𝑥,𝑡 H𝑡−1 = | N − (𝑡)| = (1 ± 6𝜂)𝛾 𝑎 𝑘 .
(1 ± 2𝜂) 𝑝 (𝑛 − 𝑡 + 1)

Therefore, if B and Y𝑘−1 hold, we have

𝑛−𝛿𝑛
Õ 𝑛
Õ
E(𝑌𝑘,𝑥,𝑡 H𝑡−1 ) = (1 ± 10(𝑘 − 1)𝜂)(1 ± 6𝜂)𝛾 𝑠 𝑘 𝛿𝑛 .
𝑡=1 𝑥=𝑛−𝛿𝑛+1

Applying Corollary 4.2.2 with 𝜚˜ = 𝜂𝛾 𝑘 𝛿𝑛, we deduce that the probability

that Y𝑘 fails is exponentially small. Indeed the probability that B holds but
𝜂2 𝛾 2𝑘 𝛿2 𝑛2 log 𝑛

≤ 𝑛−5𝐶 ,
Í𝑛−𝛿𝑛 Í𝑛
𝑥=𝑛−𝛿𝑛+1 𝑌𝑘,𝑥,𝑡 ≠ (1 ± 10𝑘𝜂)𝛾 𝛿𝑛 is at most 2 exp −
𝑘
𝑡=1 2𝐷𝑐𝑛2
Í𝑛 𝑐𝑛
where we use that 𝑥=𝑛−𝛿𝑛+1 𝑌𝑘,𝑥,𝑡 ≤ deg(𝑡) ≤ log 𝑛.
As Y0 holds trivially with probability one, by a union bound over the choices of
ℓ, a, 𝑘 and 𝑣 we obtain that the probability that B holds but there is some 1 ≤ 𝑘 ≤ 𝑑
for which Y𝑘 fails is at most 𝑛10−5𝐶 . Finally, Lemma 4.4.3 states that B holds with
probability at most 𝑛−5𝐶 , giving the lemma statement by the union bound.

Given a hypergraph 𝐺 on [𝑛] whose last 𝛿𝑛 vertices are strongly independent,

if RandomEmbedding is run to give an embedding 𝜑 of the first 𝑛 − 𝛿𝑛 vertices
of 𝐺 into the 𝑛-vertex hypergraph 𝐻, and 𝐻 ∗ is a hypergraph on 𝑉 (𝐺), we define
the completion graph 𝐶 (𝐺, 𝐻 ∗ , 𝜑) to be the bipartite graph with parts 𝑋 = 𝑉 (𝐺) \
[𝑛 − 𝛿𝑛] and 𝑌 = 𝑉 (𝐻) \ im 𝜑, and an edge 𝑥𝑦 for 𝑥 ∈ 𝑋 and 𝑦 ∈ 𝑌 whenever


we have 𝜑 N𝐺 (𝑥) ⊆ 𝑁 𝐵𝐻𝐻 ∗ (𝑦). Recall that each stage of PackingProcess first
uses RandomEmbedding to embed the first 𝑛 − 𝛿𝑛 vertices of 𝐺 𝑠 into 𝐻𝑠 , then
picks a uniform random extension to complete the embedding of 𝐺 𝑠 using edges
of 𝐻𝑠∗ . In other words, what is picked is a uniform random perfect matching in

116
 4 Packing Degenerate Hypergraphs

𝐶 (𝐺 𝑠 , 𝜑, 𝐻𝑠∗ ). The following lemma guarantees that 𝐶 (𝐺 𝑠 , 𝜑, 𝐻𝑠∗ ) satisfies the

conditions of Lemma 4.2.6 which tell us how the uniform random perfect matching
behaves.

Lemma 4.6.2. For each 𝑟 ≥ 2, 𝐷 ∈ N and 𝛾 > 0, let the constants be as in

Setting 4.3.4. Suppose that 𝐺 is a hypergraph on vertex set [𝑛], with deg− (𝑥) ≤ 𝐷
for each 𝑥 ∈ 𝑉 (𝐺), with maximum degree at most 𝑐𝑛/log 𝑛 and whose last 𝛿𝑛
vertices all have degree 𝑑, where 0 ≤ 𝑑 ≤ 𝐷, and form a strongly independent set.


Suppose that 𝐻 is an (𝛼𝑠∗ , 𝑄)-quasirandom 𝑛-vertex hypergraph with 𝑝 𝑛𝑟 edges,
where 𝑝 ≥ 𝛾, and that 𝐻 ∗ is a hypergraph on 𝑉 (𝐻) with (1 ± 𝜂)𝛾 𝑛𝑟 edges such

that (𝐻, 𝐻 ∗ ) forms an (𝜂, 𝑄)-coquasirandom pair. When RandomEmbedding is

run to give an embedding 𝜑 of the first 𝑛 − 𝛿𝑛 vertices of 𝐺 into 𝐻, with probability
at least 1 − 2𝑛−4𝐶 it succeeds and the completion graph 𝐶 (𝐺, 𝜑, 𝐻 ∗ ) satisfies the
following. For every 𝑥 ∈ 𝑋, we have deg(𝑥) = (1 ± 2𝜂)𝛾 𝑑 𝛿𝑛, and for all but at
𝐷𝑐𝑛 0 0 2𝑑
most log 𝑛 choices of 𝑥 we have in addition deg(𝑥, 𝑥 ) = (1 ± 2𝜂)𝛾 𝛿𝑛. For every
𝑦 ∈ 𝑌 , we have deg(𝑦) = (1 ± 10𝐷𝜂)𝛾 𝑑 𝛿𝑛.

Proof. When RandomEmbedding is run to produce a partial embedding 𝜑 of 𝐺 into

𝐻, by Lemma 4.4.3 with probability at least 1−𝑛−5𝐶 the algorithm succeeds and the
triple (𝐻, 𝐻 ∗ , im 𝜑) satisfies the (2𝜂, 𝑄)-codiet condition. By Lemma 4.6.1, with
probability at least 1 − 𝑛−4𝐶 in addition we have, for every vertex 𝑣 of 𝑉 (𝐻 ∗ ) \ im 𝜑,

𝑥 ∈ 𝑉 (𝐺) : 𝑛−𝛿𝑛 < 𝑥 ≤ 𝑛, 𝜑( N− (𝑥)) ⊆ N𝐻 ∗ (𝑣)


= (1±10𝐷𝜂)𝛾 𝑑 𝛿𝑛 . (4.6.2)

Suppose that both good events occur, which happens with probability at least
1 − 2𝑛−4𝐶 . Let 𝐼 denote the last 𝛿𝑛 vertices of 𝐺. Since the triple (𝐻, 𝐻 ∗ , im 𝜑)
satisfies the (2𝜂, 𝑄)-codiet condition we get, for every vertex 𝑥 of 𝐼,

𝑣 ∈ 𝑉 (𝐻 ∗ ) \ im 𝜑 : 𝜑( N− (𝑥)) ⊆ N𝐻 ∗ (𝑣)

= (1 ± 2𝜂)𝛾 𝑑 𝛿𝑛 , (4.6.3)

and for every pair of vertices 𝑥, 𝑥 0 of 𝐼 such that N𝐺 (𝑥) ∩ N𝐺 (𝑥 0) = ∅ we get

𝑣 ∈ 𝑉 (𝐻 ∗ ) \ im 𝜑 : 𝜑( N− (𝑥)) ∪ 𝜑( N− (𝑥 0)) ⊆ N𝐻 ∗ (𝑣)

= (1 ± 2𝜂)𝛾 2𝑑 𝛿𝑛 .


(4.6.4)
For any given 𝑥 ∈ 𝐼, the set N𝐺 (𝑥) contains at most 𝐷 semi-edges. Since no
𝑐𝑛 𝐷𝑐𝑛
semi-edge of 𝐺 is in more than log 𝑛 edges of 𝐺, we see that there are at most log 𝑛
vertices 𝑥 0 ∈ 𝐼 such that N𝐺 (𝑥) ∩ N𝐺 (𝑥 0) ≠ ∅.

117
 4 Packing Degenerate Hypergraphs

By definition of 𝐶 (𝐺, 𝜑, 𝐻 ∗ ), (4.6.2) gives the statement about degrees in 𝑌 ,

while (4.6.3) and (4.6.4) give the degree and codegree statements for 𝑋.

The completion lemma, Lemma 4.3.10, now follows easily.

Proof of Lemma 4.3.10. The assumptions of Lemma 4.3.10 match those of

Lemma 4.6.2, so with probability at least 1 − 2𝑛−4𝐶 RandomEmbedding succeeds
in producing an embedding 𝜑 of the first 𝑛 − 𝛿𝑛 vertices of 𝐺 into 𝐻, and the
completion graph 𝐶 (𝐺, 𝜑, 𝐻 ∗ ) satisfies the degree and codegree conditions of that
lemma.
We apply Lemma 4.2.6 with 𝐹 = 𝐹 0 = 𝐶 (𝐺, 𝜑, 𝐻 ∗ ), with 𝑚 = 𝛿𝑛, with 𝜇 = 𝛾 𝑑 ,
𝑚2 𝛿 2 𝑛2 𝐷𝑐𝑛2
with 𝜀 = 10𝐷𝜂, and any valid 𝜚 > 0. Since log 𝑚 ≥ log 𝑛 > log 𝑛 , the degree and
codegree conditions from Lemma 4.6.2 give us (M 1) and (M 1), while (M 3) is
trivially true. Thus 𝐶 (𝐺, 𝜑, 𝐻 ∗ ) contains a perfect matching, which corresponds
to the existence of an extension of 𝜑 to the desired 𝜑∗ .

Finally, we prove Lemma 4.3.9. The idea is as follows. We fix a semi-edge 𝑦, give
an upper bound for the expected number of edges used at 𝑦 in each stage, and apply
Corollary 4.2.2 to show that the actual outcome is with high probability not much
larger than this upper bound. In order to obtain this upper bound on expectation,
we will need to use again the observation just made that with probability at least
1−2𝑛−4𝐶 the output of RandomEmbedding gives us a completion graph 𝐶 (𝐺, 𝜑, 𝐻 ∗ )
which satisfies the conditions of Lemma 4.2.6.
For each 𝑧 ∈ 𝑆(𝐺 𝑠 ), we define the completion degree of 𝑧, written deg∗ (𝑧), to
be the degree of 𝑧 in the hypergraph 𝐺 𝑐𝑠 whose vertex set is 𝑉 (𝐺 𝑠 ) and whose edge
set is all edges of 𝐺 𝑠 which have exactly one vertex in the final 𝛿𝑛 vertices and the
other 𝑟 − 1 in the first 𝑛 − 𝛿𝑛 vertices. Then the number of edges of 𝐻0∗ at 𝑦 used in
stage 𝑠 is deg𝐺 𝑠 (𝑧) where 𝑧 is the semi-edge of 𝐺 𝑠 embedded to 𝑦. Note that since
every edge of 𝐺 𝑐𝑠 has exactly one vertex among the final 𝛿𝑛 vertices, and each of
these final 𝛿𝑛 vertices is in 𝑑 𝑠 edges, we have
Õ
deg∗ (𝑦) = 𝑟𝛿𝑛𝑑 𝑠 . (4.6.5)
𝑦∈𝑆(𝐺 𝑐𝑠 )

Let 𝑆0 denote the semi-edges in 𝐺 𝑐𝑠 that have no vertices in the last 𝛿𝑛, 𝑆1 those
that have one vertex in the last 𝛿𝑛 and 𝑆2 those that have at least two.

Proof of Lemma 4.3.9. Fix 𝑦 ∈ 𝑆(𝐻0∗ ). For each 𝑠 ∈ [𝑠∗ ], let 𝑌𝑠 be the number of
edges of 𝐻0∗ at 𝑦 used in stage 𝑠. For 𝑧 ∈ 𝑆2 we have deg∗ (𝑧) = 0 since the last 𝛿𝑛

118
 4 Packing Degenerate Hypergraphs

vertices of 𝐺 𝑐𝑠 are strongly independent, so we have

deg∗ (𝑧) 1𝑧↩→𝑦 = deg∗ (𝑧) 1𝑧↩→𝑦 + deg∗ (𝑧) 1𝑧↩→𝑦 + 0 · 1𝑧↩→𝑦
Õ Õ Õ Õ
𝑌𝑠 =
𝑧∈𝑆(𝐺 𝑐𝑠 ) 𝑧∈𝑆0 𝑧∈𝑆1 𝑧∈𝑆2
(4.6.6)
We define E 𝑠 to be the event that PackingProcess completes stages up to and
including 𝑠 − 1, that (𝐻𝑠−1 , 𝐻0∗ ) is (𝛼𝑠−1 , 𝑄)-coquasirandom, and that for each
𝑧 ∈ 𝑆(𝐻0∗ ) we have deg𝐻 ∗ (𝑧) − deg𝐻 ∗ (𝑧) ≤ 10𝑟!𝛾 −𝐷 𝐷𝛿𝑛. Observe if E 𝑠 holds,
0 𝑠−1
that if 𝑅 ⊆ 𝑆 are any two collections of semi-edges of 𝐻𝑠−1 ∗ with |𝑆| ≤ 𝑄, since
(𝐻𝑠−1 , 𝐻0∗ ) is (𝛼𝑠−1 , 𝑄)-coquasirandom, we have

|𝑅|
N𝐻𝑠−1 (𝑅) ∩ N𝐻0∗ (𝑆 \ 𝑅) = (1 ± 𝛼𝑠−1 ) 𝑝 𝑠−1 𝛾 |𝑆\𝑅| 𝑛 .

We therefore have

|𝑆\𝑅| |𝑅|
N𝐻𝑠−1 (𝑅) ∩ N𝐻𝑠−1
∗ (𝑆 \ 𝑅) = (1 ± 𝜂) 𝑝
𝑠−1 𝛾 𝑛,

since the neighbourhood in question is smaller in size by at most 10𝑟!𝛾 −𝐷 𝐷 2 𝛿𝑛;

by choice of 𝛿 this is tiny compared to 𝜂𝛾 𝑄 𝑛. Thus (𝐻𝑠−1 , 𝐻𝑠−1 ∗ ) is also (𝜂, 𝑄)-

coquasirandom when E 𝑠 holds. Thus E 𝑠 implies the required coquasirandomness.

Suppose that H𝑠−1 is an arbitrary history of PackingProcess up to and including
stage 𝑠 − 1 which is in E 𝑠 . We begin by estimating E(𝑌𝑠 |H𝑠−1 ).
To estimate the desired expectation, we use Lemma 4.5.5. If 𝑧 ∈ 𝑆0 , then we
need to consider the (𝑟 − 1)! possible ways that 𝑧 can be mapped to 𝑦, each of
which mappings actually occurs with probability at most 2𝑛𝑟−1 . If 𝑧 ∈ 𝑆1 , then let
𝑧0 denote the 𝑟 − 2 vertices of 𝑍 which are in the first 𝛿𝑛 vertices and let 𝑧00 = 𝑧 \ 𝑧0.
There are (𝑟 − 1)! different ways to map 𝑧0 to 𝑦; fix one such and let 𝑦00 be the
vertex of 𝑦 not mapped to. By Lemma 4.5.5, the probability that 𝑧0 is embedded
to 𝑦 and 𝑦00 is not in the image of 𝜑 when we run RandomEmbeddingis at most
2𝑛𝑟−2 𝛿. In order for 𝑧 to then be mapped to 𝑦, we need that the final vertex 𝑧00 of 𝑧 is
embedded by the random matching to the final vertex 𝑦00 of 𝑦. If 𝑦00 𝑧00 is not an edge
∗ ), then this probability is zero. Otherwise, it is the probability
of 𝐶 (𝐺 𝑠 , 𝜑 𝑠 , 𝐻𝑠−1
that 𝑦00 𝑧00 is used in a uniform random perfect matching of 𝐶 (𝐺 𝑠 , 𝜑 𝑠 , 𝐻𝑠−1
∗ ). Since

H𝑠−1 is in E 𝑠 , as in the proof of Lemma 4.3.10, with probability at least 1 − 2𝑛−4𝐶

RandomEmbedding produces 𝜑 𝑠 such that the conditions of Lemma 4.2.6 are met
∗ ), and in this case the probability that the chosen perfect matching
for 𝐶 (𝐺 𝑠 , 𝜑 𝑠 , 𝐻𝑠−1
uses 𝑥 00 𝑦00 is at most 𝛾 𝑑2𝑠 𝛿𝑛 , where 𝑑 𝑠 is the degree of the last 𝛿𝑛 vertices of 𝐺 𝑠 .

119
 4 Packing Degenerate Hypergraphs

Putting this together, we obtain

P[𝑧 ↩→ 𝑦|H𝑠−1 ] ≤ 2(𝑟 − 1)!𝑛−(𝑟−1) if 𝑧 ∈ 𝑆0 , and (4.6.7)

P[𝑧 ↩→ 𝑦|H𝑠−1 ] ≤ 4(𝑟 − 1)!𝛾 −𝑑𝑠 𝑛−(𝑟−1) + 2𝑛−4𝐶
≤ 5(𝑟 − 1)!𝛾 −𝑑𝑠 𝑛−(𝑟−1) if 𝑧 ∈ 𝑆1 . (4.6.8)

For (4.6.8), the term 𝑛−4𝐶 covers the possibility that RandomEmbedding pro-
duces 𝜑 𝑠 which does not give a completion graph with the required degree and
codegree conditions.
We now put (4.6.5), (4.6.7) and (4.6.8) into (4.6.6). It follows that when H𝑠−1
is in E 𝑠 we have

E[𝑌𝑠 |H𝑠−1 ] ≤ 5(𝑟 − 1)!𝛾 −𝑑𝑠 𝑛−(𝑟−1) · 𝑟𝛿𝑛𝑑 𝑠 ≤ 5𝑟!𝐷𝛾 −𝐷 𝛿𝑛−(𝑟−2) .

Given any 1 ≤ 𝑠0 ≤ 𝑠∗ , we estimate the probability that E 𝑠 holds for each

1 ≤ 𝑠 ≤ 𝑠0 but we have deg𝐻 ∗ (𝑦) − deg𝐻 ∗0 (𝑦) > 10𝑟!𝛾 −𝐷 𝐷𝛿𝑛.
0 𝑠
Since 0 ≤ 𝑌𝑠 ≤ Δ(𝐺 𝑠 ) holds for each 𝑠, and since 𝑠0 ≤ 𝑠max , we can apply
Ñ0
Corollary 4.2.2, with E = 𝑠𝑠=1 E 𝑠 , to give

h 𝑠0 i
20𝑟!𝐷𝛾 −𝐷 𝑛−(𝑟 −2) 𝑠max

Õ
P E and 𝑌𝑠 > 10𝑟!𝐷𝛾 −𝐷 𝑛−(𝑟−2) · 𝑠max ≤ exp − Δ(𝐺 𝑠 ) < 𝑛−4𝐶 ,
𝑖=1

where the final inequality is since Δ(𝐺 𝑠 ) ≤ 𝑐𝑛/log 𝑛 and by choice of 𝑐. Taking
the union bound over all choices of 𝑠0 and 𝑦, we see that the probability that neither
of the first two events of the lemma statement occurs, and yet E 𝑠 fails for any 𝑠, is
at most 𝑛−3𝐶 as required.

4.7 Quasirandom packing

The main theorem of this section is the following, which states that if we want to
pack almost-spanning hypergraphs, we can obtain a packing with several additional
quasirandomness properties. To state these conveniently, we need a few definitions.
We first define weight functions on our guest graphs 𝐺 𝑠 . Given a nonnegative
integer 𝑞 and an 𝑟-uniform hypergraph 𝐺 𝑠 whose vertex set 𝑉 (𝐺 𝑠 ) is [𝑣(𝐺 𝑠 )] (i.e.
it is the first 𝑣(𝐺 𝑠 ) natural numbers), we say 𝜔 𝑠 is a 𝑞-weight function on 𝐺 𝑠 when
𝜔 𝑠 is a function whose domain is 𝑉 (𝐺 𝑠)


𝑞 , the 𝑞-vertex subsets of 𝑉 (𝐺 𝑠 ), and whose

120
 4 Packing Degenerate Hypergraphs

codomain is R≥0 . The support of 𝜔 𝑠 is the set f ∈ 𝑉 (𝐺 𝑠)



𝑞 : 𝜔 𝑠 (f) > 0 , and we
say that 𝜔 𝑠 is free if all 𝑞-sets in its support contain no edges of 𝐺 𝑠 . We will say
that the support of 𝜔 𝑠 is contained in a set 𝑋 ⊆ 𝑉 (𝐺 𝑠 ) if each 𝑞-set in the support
is in 𝑋. We write
Õ Õ
Δ(𝜔 𝑠 ) := max 𝜔 𝑠 (f) and sum(𝜔 𝑠 ) := 𝜔 𝑠 (f) .
𝑥∈𝑉 (𝐺 𝑠 )
𝑉 (𝐺𝑠 ) 𝑉 (𝐺𝑠 )
f∈ ( 𝑞 ) f∈ ( 𝑞 )
𝑥∈f

Given an embedding 𝜑 𝑠 : 𝑉 (𝐺 𝑠 ) → 𝑉 (𝐻), where 𝐻 is the 𝑟-uniform host hyper-

graph, and an ordered 𝑞-set e of distinct vertices of 𝐻, we define 𝜔 𝑠 (e) to be the
weight 𝜑 𝑠 assigns to e, with the natural order on 𝐺 𝑠 corresponding to the order of
e. More formally, if 𝜑−1 −1 −1
𝑠 (𝑒 1 ) < 𝜑 𝑠 (𝑒 2 ) < · · · < 𝜑 𝑠 (𝑒 𝑞 ), we set

𝜔 𝑠 (e) = 𝜔 𝑠 {𝜑−1 −1

𝑠 (𝑒 1 ), . . . , 𝜑 𝑠 (𝑒 𝑞 )} .

If some 𝑒𝑖 is not in the image of 𝜑 𝑠 , or if one of the above inequalities does not
hold, we set 𝜔 𝑠 (e) = 0.
Given an 𝑟-uniform hypergraph 𝐻, a vertex 𝑣 ∈ 𝑉 (𝐻), and a collection
𝑋1 , . . . , 𝑋𝑟−1 of subsets of 𝑉 (𝐻), we write N𝐻 (𝑣; 𝑋1 , . . . , 𝑋𝑟−1 ) for the collec-
tion of ordered (𝑟 − 1)-sets whose 𝑖th element is in 𝑋𝑖 for each 𝑖 ∈ [𝑟 − 1] and
which are (as unordered sets) contained in N𝐻 (𝑣).

Theorem 4.7.1. For each 𝑟 ≥ 2, each 𝛾 > 0 and each 𝐷, 𝐾 ∈ N there exist
numbers 𝑛0 ∈ N and 𝛿, 𝑐, 𝜉 > 0 with 𝛿 ≤ 𝛾 such that the following holds for
each 𝑛 > 𝑛0 , with 𝑄 = (𝑟 + 1)𝐷 + 3. Suppose that 𝐻 b is an (𝜉, 𝑄)-quasirandom
𝑛

𝑟-uniform hypergraph with 𝑛 vertices and 𝑝ˆ 𝑟 edges, where 𝑝ˆ > 0. Suppose that
𝑠∗ ≤ 2𝑛𝑟−1 /(𝑟 − 1)! and that for each 𝑠 ∈ [𝑠∗ ] the 𝑟-uniform hypergraph 𝐺 𝑠 is on
𝑐𝑛 −
vertex set [𝑛 − 𝛿𝑛], with maximum degree at most log 𝑛 , such that deg (𝑥) ≤ 𝐷


Í∗
for each 𝑥 ∈ 𝑉 (𝐺 𝑠 ). Let 𝑝 satisfy 𝑝 𝑛𝑟 = 𝑝ˆ 𝑛𝑟 − 𝑠𝑠=1 𝑒(𝐺 𝑠 ), and suppose 𝑝 ≥ 𝛾.
Given any 0 ≤ 𝑞 ≤ 𝑄, let 𝜔 𝑠 : 𝑉 (𝐺 𝑠)

 𝑐𝑛 
𝑞 → 0, log 𝑛 be a 𝑞-weight function on 𝐺 𝑠
∗
for each 𝑠 ∈ [𝑠 ]. Let 𝑋, 𝑋1 , . . . , 𝑋𝑞 ⊆ 𝑉 ( 𝐻) satisfy |𝑋 |, |𝑋1 |, . . . , |𝑋𝑞 | ≥ 𝛾𝑛, and
b


suppose that for each 1 ≤ 𝑖 ≤ 𝑞 the pair 𝐻, b 𝑉 ( 𝐻)
b \ 𝑋𝑖 satisfies the (𝜉, 𝑄)-diet
condition.
There is a randomised algorithm (whose running is unaffected by the choice
of 𝑞, 𝜔 𝑠 , 𝑋, 𝑋1 , . . . , 𝑋𝑞 ) which with probability at least 1 − 𝑛−𝐾 returns maps
𝜑𝑠 : 𝐺 𝑠 → 𝐻 b for each 𝑠 ∈ [𝑠∗ ] which form a packing of (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] into 𝐻, b such
that the hypergraph 𝐻 of leftover edges is (𝛾, 𝑄)-quasirandom, and in addition the

121
 4 Packing Degenerate Hypergraphs

following hold for every 𝛿 ≤ 𝛿˜ ≤ 1 − 𝛿, every 𝑈 ⊆ 𝑉 ( 𝐻) b with |𝑈| ≤ 𝑄, and every

collection 𝑇1 , . . . , 𝑇ℓ of distinct semi-edges in 𝑉 ( 𝐻)
b with 0 ≤ ℓ ≤ 𝑄.
(QUASI 1) For every 𝑆 ⊆ [𝑠∗ ] with |𝑆| ≤ 𝑄, if 𝑋 ∩ N𝐻b (𝑇1 , . . . , 𝑇ℓ ) ≥ 21 𝛾 𝑄 𝑛,
we have
Ø
˜


𝑋 ∩ N𝐻 (𝑇1 , . . . , 𝑇ℓ ) \ 𝜑 𝑠 [𝑛 − 𝛿𝑛] =
𝑠∈𝑆

= (1 ± 𝛾) 𝛿˜|𝑆| 𝑝
ℓ
𝑝ˆ 𝑋 ∩ N𝐻b (𝑇1 , . . . , 𝑇ℓ ) .

(QUASI 2) Suppose that for each 𝑠 ∈ [𝑠∗ ] the weight function 𝜔 𝑠 is free. Then
b \ 𝑈, if Í𝑠∗ sum(𝜔 𝑠 ) ≥ 𝛾𝑛𝑞+1 , we
for every ordered 𝑞-set e in 𝑉 ( 𝐻) 𝑠=1
have
𝑠∗ 𝑠∗
𝜔 𝑠 (e) 1𝑈∩𝜑𝑠 ( [𝑛−𝛿𝑛])=∅
Õ Õ
˜ = (1 ± 𝛾) 𝛿˜|𝑈| 𝑛−𝑞 sum(𝜔 𝑠 ) .
𝑠=1 𝑠=1

(QUASI 3) Suppose that 𝑞 = 𝑟 − 1 ≥ 2. For every 𝑠 ∈ [𝑠∗ ] and 𝑣 ∈ 𝑉 ( 𝐻), b

˜
if sum(𝜔 𝑠 ) ≥ 𝛾𝑛, the support of 𝜔 𝑠 is contained in [𝑛 − 𝛿𝑛], and
𝑐𝑛
Δ(𝜔 𝑠 ) ≤ log ˜
𝑛 , then we have either 𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]) or

Õ 𝑝
𝜔 𝑠 (e) = (1 ± 𝛾) · N𝐻b (𝑣; 𝑋1 , . . . , 𝑋𝑟−1 )
𝑝ˆ
e∈N 𝐻 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )

· 𝑛1−𝑟 sum(𝜔 𝑠 ) .

Observe that (QUASI 2) for 𝑞 = 0 does have content. There is only one ordered
0-set of vertices of 𝐺 𝑠 ; if for example we choose to give it weight 1 in each 𝐺 𝑠 ,
then (QUASI 2) counts the number of 𝑠 such that 𝑈 is disjoint from 𝜑 𝑠 ([𝑛 − 𝛿𝑛]).˜
By taking the union bound, we can ask for a packing in which our properties hold
for polynomially many different choices of 𝑞, 𝜔 𝑠 , 𝑋, 𝑋1 , . . . , 𝑋𝑞 simultaneously.
The proof of Theorem 4.7.1 mainly consists of using the tools developed in
Sections 4.4 and 4.5. As mentioned in the sketch in Section 4.3.1, we use the
cases of the various lemmas in which 𝐻0∗ has no edges, and 𝐻0 = 𝐻. b That is, the
randomised algorithm of Theorem 4.7.1 is the following PackingProcess2 and the
various lemmas of the preceding three sections apply to it.

Proof of Theorem 4.7.1. To begin with, we prove that PackingProcess2 is likely to

succeed. This is much the same as the corresponding proof for Theorem 4.3.1,

122
 4 Packing Degenerate Hypergraphs

Algorithm 3: PackingProcess2
Input : hypergraphs 𝐺 1 , . . . , 𝐺 𝑠∗ , with 𝐺 𝑠 on vertex set [𝑛 − 𝛿𝑛]; a
hypergraph 𝐻b on 𝑛 vertices
let 𝐻0 = 𝐻b;
for 𝑠 = 1 to 𝑠∗ do
run RandomEmbedding(𝐺 𝑠 ,𝐻𝑠−1 ) to get an embedding 𝜑 𝑠 of 𝐺 𝑠
into 𝐻𝑠−1 ;
let 𝐻𝑠 be

the hypergraph obtained from 𝐻𝑠−1 by removing the edges of
𝜑𝑠 𝐺 𝑠 ;
end

so we omit the details. In order that PackingProcess2 does not succeed, there
must exist some smallest 1 ≤ 𝑠 ≤ 𝑠∗ such that PackingProcess2 runs up to and
including stage 𝑠 − 1, and 𝐻𝑠 0 is (𝛼𝑠 0 , 𝑄)-quasirandom for each 1 ≤ 𝑠0 ≤ 𝑠 − 1, but
either RandomEmbedding fails to embed 𝐺 𝑠 , or 𝐻𝑠 is not (𝛼𝑠 , 𝑄)-quasirandom.
By Lemma 4.3.8, the probability of RandomEmbedding failing to embed 𝐺 𝑠 is at
most 𝑛−5𝐶 , and by Lemma 4.3.7, if 𝐺 𝑠 is successfully embedded, the probability
that 𝐻𝑠 is not (𝛼𝑠 , 𝑄)-quasirandom is at most 𝑛−4𝐶 . Taking the union bound over
the choices of 𝑠, we see that the probability of either failure at any stage is at most
𝑛−3𝐶 . In what follows, we refer to the likely event as PackingProcess2 behaves and
we will generally assume it occurs.
We next prove (QUASI 1), with the error bound 21 𝛾 in place of 𝛾. Note that the
claim that 𝐻 is (𝛾, 𝑄)-quasirandom is a special case of this statement obtained by
taking 𝑋 = 𝑉 ( 𝐻) b and 𝑆 = ∅, since for any given collection of distinct (𝑟 − 1)-sets
𝑇1 , . . . , 𝑇ℓ , where ℓ ≤ 𝑄, we have N𝐻b (𝑇1 , . . . , 𝑇ℓ ) = (1 ± 𝜉) 𝑝ˆℓ 𝑛 > 12 𝛾 𝑄 𝑛.
Given a set 𝑋, we fix 𝑆 = {𝑠1 , . . . } in increasing order, ℓ and 𝑇1 , . . . , 𝑇ℓ , and
˜ as in the statement; we assume that 𝑋 ∩ N b (𝑇1 , . . . , 𝑇ℓ ) ≥ 1 𝛾 𝑄 𝑛 otherwise
𝛿, 𝐻 2
there is nothing to prove. The idea is the following. We split up the running of
PackingProcess2 into several parts: the interval before embedding 𝐺 𝑠1 , embedding
𝐺 𝑠1 , the interval between 𝐺 𝑠1 and 𝐺 𝑠2 , and so on. We use Lemma 4.5.8 to determine
how 𝑋 ∩ N𝐻 𝑗 (𝑇1 , . . . , 𝑇ℓ ) shrinks as 𝑗 increases through the first interval, how
˜
𝑋 ∩ N𝐻 𝑗 (𝑇1 , . . . , 𝑇ℓ ) \ 𝜑 𝑠1 ([𝑛 − 𝛿𝑛]) shrinks as 𝑗 increases through the second
interval, and so on. We use Lemma 4.4.3 to determine how much of each set is
covered by the next im 𝜑 𝑠𝑖 ([𝑛 − 𝛿𝑛]), ˜ and observe that embedding any one graph
can only change N𝐻 𝑗 (𝑇1 , . . . , 𝑇ℓ ) by a tiny amount.

123
 4 Packing Degenerate Hypergraphs

Claim 4.7.2. For each −1 ≤ 𝑖 ≤ |𝑆| we have with probability at least 1−3(𝑖+1)𝑛−5𝐶

𝑖
Ø
𝑋 ∩ N𝐻𝑠𝑖+1 −1 (𝑇1 , . . . , 𝑇ℓ ) \ ˜
𝜑 𝑠 𝑗 ([𝑛 − 𝛿𝑛])
𝑗=1
𝑝
(𝑠𝑖+1 −1) ℓ

3𝑖+3 ˜max(0,𝑖)
= (1 ± 𝐶𝛼𝑠𝑖+1 −1 ) 𝛿 · 𝑋 ∩ N𝐻b (𝑇1 , . . . , 𝑇ℓ ) ,
𝑝ˆ

where we set 𝑠 |𝑆|+1| := 𝑠∗ + 1 and 𝑠0 = 1.

Proof. The statement for 𝑖 = −1 is a triviality. Suppose now that 𝑖 ≥ 0 and the
statement holds for 𝑖 − 1. Let
𝑖−1
Ø
𝑍 = 𝑋 ∩ N𝐻𝑠𝑖 −1 (𝑇1 , . . . , 𝑇ℓ ) \ ˜ ,
𝜑 𝑠 𝑗 ([𝑛 − 𝛿𝑛])
𝑗=1
𝑖
Ø
𝑍 0 = 𝑋 ∩ N𝐻𝑠𝑖 −1 (𝑇1 , . . . , 𝑇ℓ ) \ ˜
𝜑 𝑠 𝑗 ([𝑛 − 𝛿𝑛]) and
𝑗=1
𝑖
Ø
𝑍 00 = 𝑋 ∩ N𝐻𝑠𝑖 (𝑇1 , . . . , 𝑇ℓ ) \ ˜ .
𝜑 𝑠 𝑗 ([𝑛 − 𝛿𝑛])
𝑗=1

If 𝑖 = 0, then we have 𝑍 = 𝑍 0. If 𝑖 > 0, by Lemma 4.4.3(e) with the given 𝑍, with

˜
probability at least 1 − 𝑛−5𝐶 , we have |𝑍 0 | = 𝑍 \ 𝜑 𝑠𝑖 ([𝑛 − 𝛿𝑛]) ˜ |.
= (1 ± 𝐶𝛼𝑠𝑖 ) 𝛿|𝑍
In either case, we conclude that with probability at least 1 − 𝑛−5𝐶 we have

𝑠𝑖 −1 ℓ
𝑝 
|𝑍 0 | = (1 ± 𝐶𝛼𝑠𝑖 ) 3𝑖+1 𝛿˜max(0,𝑖) · 𝑋 ∩ N𝐻b (𝑇1 , . . . , 𝑇ℓ ) .
𝑝ˆ
𝑐𝑛
Observe that since 𝐺 𝑠𝑖 has no vertex of degree more than log 𝑛 , the number of
𝑐𝑛
vertices in N𝐻𝑠𝑖 −1 (𝑇1 , . . . , 𝑇ℓ ) which are not in N𝐻𝑠𝑖 (𝑇1 , . . . , 𝑇ℓ ) is at most 𝑄 log 𝑛,
and since 𝑒(𝐺 𝑠 ) ≤ 𝐷𝑛 we have 𝑝 𝑠𝑖 −1 − 𝑝 𝑠𝑖 < 𝑛−0.5 . Putting these together we have
with probability at least 1 − 𝑛−5𝐶
 𝑝 ℓ
|𝑍 00 | = (1 ± 𝐶𝛼𝑠𝑖 ) 3𝑖+2 𝛿˜max(0,𝑖)
𝑠𝑖
· 𝑋 ∩ N𝐻b (𝑇1 , . . . , 𝑇ℓ ) .
𝑝ˆ

Finally, we consider the embeddings of the graphs 𝐺 𝑠 with 𝑠𝑖 < 𝑠 ≤ 𝑠𝑖+1 − 1. By

Lemma 4.5.8, with probability at least 1 − 𝑛−5𝐶 , if for each 𝑠 with 𝑠𝑖 ≤ 𝑠 ≤ 𝑠𝑖+1 − 1
the graph 𝐻𝑠 is (𝛼𝑠 , 𝑄)-coquasirandom, we have

 𝑝 𝑠𝑖+1 −1  ℓ
|𝑍 00 ∩ N𝐻𝑠𝑖+1 −1 (𝑇1 , . . . , 𝑇ℓ ) = 1 ± 12 𝛼𝑠𝑖+1 −1 𝑝 𝑠𝑖 |𝑍 00 | ,

124
 4 Packing Degenerate Hypergraphs

and this implies the desired equation for 𝑖. Summing the probability bounds we
obtain the claimed probability of this equation holding. 

Taking the 𝑖 = |𝑆| case of Claim 4.7.2, and a union bound over choices of 𝑆,
˜ by choice of 𝛼𝑠max we see that (QUASI 1), with error bound
ℓ, 𝑇1 . . . , 𝑇ℓ and 𝛿,
1 𝑄𝑟+𝑄𝑟 2 +1 · 3𝑛−5𝐶 ≥ 1 − 𝑛−4𝐶 provided
2 𝛾, holds with probability at least 1 − 𝑄𝑛
PackingProcess2 behaves. Taking account of the possibility of misbehaviour, we
see that (QUASI 1) holds with probability at least 1 − 𝑛−2𝐶 . Note that since 𝑛 − 𝛿𝑛˜
is required to be in [𝑛], the number of choices of 𝛿˜ over which we take a union
bound is at most 𝑛.
We next prove (QUASI 2). Given 𝑞 ≤ 𝑄 and free 𝑞-weight functions 𝜔 𝑠 on 𝐺 𝑠
for each 𝑠 ∈ [𝑠∗ ], fix 𝑈 ⊆ 𝑉 ( 𝐻),
b 𝛿,˜ and an ordered 𝑞-set e in 𝑉 ( 𝐻)
b \𝑈. Let for each
1 ≤ 𝑠 ≤ 𝑠∗ the random variable 𝑌𝑠 := 𝜔 𝑠 (e) by running PackingProcess2 to obtain
embeddings 𝜑 𝑠 of each 𝐺 𝑠 . Let E be the event that PackingProcess2 succeeds and
that 𝐻𝑠 is (𝛼𝑠 , 𝑄)-quasirandom for each 𝑠 ∈ [𝑠∗ ]. Given 𝐹 in the support of 𝜔 𝑠 ,
since 𝜔 𝑠 is free, 𝐹 is a 𝑞-set which contains no edges of 𝐺 𝑠 . By Lemma 4.5.5,
provided that 𝐻𝑠−1 is (𝛼𝑠−1 , 𝑄)-quasirandom, the probability that the 𝑖th vertex of
𝐹 (in the natural order on [𝑛 − 𝛿𝑛]) is embedded to the 𝑖th vertex of e, and in
˜
addition im 𝜑 𝑠 ([𝑛 − 𝛿𝑛]) ∩ 𝑈 = ∅, is

(2+2𝐷)𝑄+1
1 ± 600𝐶𝑄𝛼𝑠−1 𝛿−1 𝑝 0𝑠−1 𝑛−𝑞 𝛿˜|𝑈| = 1 ± 41 𝛾 𝑛−𝑞 𝛿˜|𝑈| .

By linearity of expectation, if 𝐻𝑠−1 is (𝛼𝑠−1 , 𝑄)-quasirandom we have

E[𝑌𝑠 |𝐻𝑠−1 ] = sum(𝜔 𝑠 ) · 1 ± 41 𝛾 𝑛−𝑞 𝛿˜|𝑈| .

𝑐𝑛
Since 0 ≤ 𝑌𝑠 ≤ log 𝑛 , by Corollary 4.2.2 we have

𝑠∗
Õ 𝑠∗
Õ
sum(𝜔 𝑠 ) · 1 ± 12 𝛾 𝑛−𝑞 𝛿˜|𝑈|


𝑌𝑠 =
𝑠=1 𝑠=1

Í∗
with probability at least 1 − 𝑛−5𝐶 . Here we use the fact 𝑠𝑠=1 sum(𝜔 𝑠 ) ≥ 𝛾𝑛𝑞+1 and
choice of 𝑐. Taking the union bound over choices of 𝑈, 𝛿˜ and e, and taking account
of the possibility of PackingProcess2 not behaving, we see that (QUASI 2) holds
with probability at least 1 − 𝑛−2𝐶 as desired.

We now complete the proof of Theorem 4.7.1 by proving (QUASI 3). Our

125
 4 Packing Degenerate Hypergraphs

calculations in the preceding parts go through for 𝑟 = 2, but for this part we need
to assume 𝑟 ≥ 3. Fix 𝑠 and 𝑣 ∈ 𝑉 ( 𝐻), b and sets 𝑋1 , . . . , 𝑋𝑟−1 in 𝑉 ( 𝐻)
b each of size
at least 𝛾𝑛. Suppose that 𝜔 𝑠 is a (𝑟 − 1)-weight function on 𝐺 𝑠 whose support is
˜ such that sum(𝜔 𝑠 ) ≥ 𝛾𝑛. Recall that we need to prove that it is
contained in [𝑛− 𝛿𝑛]
˜
likely that either 𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]),
Í
or we have control of e∈N𝐻 (𝑣;𝑋1 ,...,𝑋𝑟 −1 ) 𝜔 𝑠 (e).
We will assume for now that PackingProcess2 behaves, and take account of its
˜
small failure probability later. Since there is nothing to prove if 𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]),
we will assume this event does not occur.
The proof of (QUASI 3) is fairly involved, so we now sketch the idea. First,
we argue that when RandomEmbedding embeds the first 𝑛 − 𝛿𝑛 ˜ vertices of 𝐺 𝑠
Í
into 𝐻𝑠−1 , we can estimate e∈N𝐻 (𝑣;𝑋1 ,...,𝑋𝑟 −1 ) 𝜔 𝑠 (e) fairly accurately. We then
𝑠−1
argue that the embedding of what remains of 𝐺 𝑠 does not change the result much;
Í
that is, we obtain an estimate for e∈N𝐻𝑠 (𝑣;𝑋1 ,...,𝑋𝑟 −1 ) 𝜔 𝑠 (e). Finally, we argue that
as the remaining hypergraphs 𝐺 𝑠+1 , . . . , 𝐺 𝑠∗ are embedded, we can maintain an
Í
estimate of e∈N𝐻 0 (𝑣;𝑋1 ,...,𝑋𝑟 −1 ) 𝜔 𝑠 (e). It is easy to estimate (using Lemma 4.5.5)
𝑠
the conditional expected change 𝑌𝑖 when any one 𝐺 𝑖 is embedded, and what we
Í𝑠∗
would like to do is to show that 𝑖=𝑠+1 𝑌𝑖 is close to the sum of the conditional
expected changes. However the range of these 𝑌𝑖 is too large for Corollary 4.2.2
to give useful results. What we do to get around this is to define capped random
variables; that is, we define a random variable 𝑌𝑖0 which is equal to the minimum
of the actual change when 𝐺 𝑖 is embedded and a fixed number, the cap. The cap is
Í𝑠∗
small enough that we can apply Corollary 4.2.2 to show concentration of 𝑖=𝑠+1 𝑌𝑖0,
and we then argue that it is very likely that 𝑌𝑖 = 𝑌𝑖0 for each 𝑖 and that the conditional
expectations of the two random variables are very close. This final part is where
we use our assumption 𝑟 ≥ 3 and where we need to do most work.
The following claim establishes the first step in the above argument. In this claim
we will write 𝜑 𝑠 although we formally do not assume RandomEmbedding succeeds
in embedding all of 𝐺 𝑠 but only the first 𝑛 − 𝛿𝑛 ˜ vertices and should really therefore
write instead the partial embedding of these first vertices.
Claim 4.7.3. When PackingProcess2 is run up to and including the embedding of
˜ vertices of 𝐺 𝑠 , with probability at least 1 − 3𝑛−3𝐶 , we have either
the first 𝑛 − 𝛿𝑛
˜
𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]) or
Õ
𝑤 𝑠 (e) = 1 ± 5𝑟𝐶𝛼𝑠 𝑝 𝑠−1 · |𝑋1 | . . . |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 ) .


e∈N 𝐻𝑠−1 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )

126
 4 Packing Degenerate Hypergraphs

Proof. We prove this claim by arguing that it is likely that the total weight of
semi-edges in 𝐺 𝑠 whose first element is mapped to 𝑋1 is roughly |𝑋𝑛1 | sum(𝜔 𝑠 ), and
inductively prove that the total weight of semi-edges in 𝐺 𝑠 whose first 𝑖 elements
are mapped to 𝑋1 , . . . , 𝑋𝑖 respectively is likely to be roughly |𝑋1 |...|𝑋
𝑛𝑖
𝑖|
sum(𝜔 𝑠 ), for
each 2 ≤ 𝑖 ≤ 𝑟 − 2. We finally use this to prove the claim statement. In each of
these steps, the critical point is that we can keep control of the probability that any
given vertex 𝑥 of 𝐺 𝑠 is embedded to any given set 𝑋𝑖 : we know how large is the set
of vertices to which we will embed 𝑥 from the diet condition, and we will prove a
version of the diet condition for each set 𝑋𝑖 . We begin by stating and proving this
last condition.
We first claim that with probability at least 1 − 2𝑛−3𝐶 , when PackingProcess2
is run up to and including the embedding of 𝐺 𝑠−1 , it succeeds, 𝐻𝑠−1 is (𝛼𝑠−1 , 𝑄)-
quasirandom, and for any set 𝑅 of distinct semi-edges of 𝐻𝑠−1 with |𝑅| ≤ 𝑄, and
|𝑅|
any 1 ≤ 𝑖 ≤ 𝑞 we have N𝐻𝑠−1 (𝑅) ∩ 𝑋𝑖 = (1 ± 𝛼𝑠−1 ) 𝑝 𝑠−1 |𝑋𝑖 |. The first two parts of
this are part of the event that PackingProcess2 behaves, whose failure probability
is at most 𝑛−3𝐶 , so we need to establish the final part. We use Lemma 4.5.8, once
for each 𝑋𝑖 and for each choice of 𝑅, to establish that (for each fixed such choice)
we have with probability at least 1 − 𝑛−5𝐶 that either PackingProcess2 does not
behave, or

𝑝 𝑠−1
|𝑅| |𝑅|
N𝐻𝑠−1 (𝑅) ∩ 𝑋𝑖 = 1 ± 21 𝛼𝑠−1 𝑝ˆ · N𝐻𝑠−1 (𝑅) ∩ 𝑋𝑖 = (1 ± 𝛼𝑠−1 ) 𝑝 𝑠−1 |𝑋𝑖 | ,

where the second equality is since ( 𝐻,b 𝑉 ( 𝐻)

b \ 𝑋𝑖 ) satisfies the (𝜉, 𝑄)-diet condition.
Taking the union bound over choices of 𝑖 and 𝑅 we obtain the desired claim.
We now build on this to prove the desired version of the diet condition. That
is, we prove that assuming that the likely event of the first claim occurs, when
RandomEmbedding is run to embed the first 𝑛 − 𝛿𝑛 ˜ vertices of 𝐺 𝑠 into 𝐻𝑠−1 ,
with probability at least 1 − 𝑛−4𝐶 it succeeds, for each 1 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛 ˜ the pair


𝐻𝑠−1 , 𝜑 𝑠 ([𝑡]) satisfies the (𝐶𝛼𝑠 , 𝑄)-diet condition, and for any set 𝑅 of distinct
semi-edges of 𝐻𝑠−1 with |𝑅| ≤ 𝑄, and any 1 ≤ 𝑖 ≤ 𝑞 we have

|𝑅|
N𝐻𝑠−1 (𝑅) ∩ 𝑋𝑖 \ 𝜑 𝑠 ([𝑡]) = (1 ± 2𝐶𝛼𝑠 ) 𝑝 𝑠−1 𝑛−𝑡
𝑛 |𝑋𝑖 | .

This follows directly from the first claim and Lemma 4.4.3, with 𝑍 = N𝐻𝑠−1 (𝑅) ∩ 𝑋𝑖 ,
taking the union bound over 𝑡 and choices of 𝑅 and 𝑖.
We now prove by induction on 𝑖 the following. Suppose that the likely event

127
 4 Packing Degenerate Hypergraphs

of the first claim above holds. For each 0 ≤ 𝑖 ≤ 𝑟 − 2, with probability at least
˜ vertices
1 − 𝑛−4𝐶 − 𝑖𝑛−5𝐶 , when RandomEmbedding is run to embed the first 𝑛 − 𝛿𝑛
of 𝐺 𝑠 into 𝐻𝑠−1 , we have

𝑖
|𝑋 𝑗|
Õ Ö
𝜔 𝑠 ({𝑥 1 , . . . , 𝑥𝑟−1 }) = (1 ± 5𝑖𝐶𝛼𝑠 )sum(𝜔 𝑠 ) 𝑛 ,
(𝑥 1 ,...,𝑥𝑟 −1 ) 𝑗=1

where the sum runs over vectors with 𝑥1 < 𝑥 2 < · · · < 𝑥𝑟−1 members of [𝑛 − 𝛿𝑛],˜
such that 𝜑 𝑠 (𝑥 𝑗 ) ∈ 𝑋 𝑗 for each 1 ≤ 𝑗 ≤ 𝑖.
The induction statement is trivial for 𝑖 = 0. Suppose now 𝑖 ≥ 1, and the statement
holds for 𝑖 − 1. For each 1 ≤ 𝑡 ≤ 𝑛 − 𝛿𝑛, ˜ let

𝜔 𝑠 ({𝑥 1 , . . . , 𝑥𝑟−1 }) 1𝜑𝑠 (𝑡)∈𝑋𝑖 ,

Õ
𝑌𝑡 =
(𝑥1 ,...,𝑥𝑟 −1 )

where the sum runs over vectors with 𝑥 1 < 𝑥 2 < · · · < 𝑥𝑟−1 members of [𝑛 −
˜
𝛿𝑛], such that 𝜑 𝑠 (𝑥 𝑗 ) ∈ 𝑋 𝑗 for each 1 ≤ 𝑗 ≤ 𝑖 and 𝑥𝑖 = 𝑡. Observe that 0 ≤
𝑛 by definition. Letting H𝑡−1 denote the embedding history of
𝑐𝑛
|𝑌𝑡 | ≤ Δ(𝜔 𝑠 ) ≤ log
RandomEmbedding up to and including the embedding of 𝑡 − 1, we have

deg− (𝑡) 𝑛+1−𝑡

(1 ± 2𝐶𝛼𝑠 ) 𝑝 𝑠−1 𝑛 |𝑋𝑖 |
P[𝜑 𝑠 (𝑡) ∈ 𝑋𝑖 |H𝑡−1 ] = −
deg (𝑡)
= (1 ± 4𝐶𝛼𝑠 ) |𝑋𝑛𝑖 | ,
(1 ± 𝐶𝛼𝑠 ) 𝑝 𝑠−1 (𝑛 + 1 − 𝑡)


where we use the (𝐶𝛼𝑠 , 𝑄)-diet condition for 𝐻𝑠−1 , 𝜑 𝑠 ([𝑡 − 1]) to estimate the
denominator and the second claim for the numerator. Assuming the likely event of
the 𝑖 − 1 induction statement, we conclude

𝑛− ˜
𝛿𝑛 𝑖
|𝑋 𝑗|
Õ Ö
E[𝑌𝑡 |H𝑡−1 ] = (1 ± 4𝐶𝛼𝑠 ) |𝑋𝑛𝑖 | · (1 ± 5(𝑖 − 1)𝐶𝛼𝑠 )sum(𝜔 𝑠 ) 𝑛 ,
𝑡=1 𝑗=1

and by Corollary 4.2.2, with E the event that the diet condition, and the variant diet
conditions of the second claim, and the equation of the 𝑖 − 1 statement, all hold,
we see that the probability that E holds and the equality of the 𝑖 statement fails is
at most 𝑛−5𝐶 . This gives the desired induction statement for 𝑖.
Finally, we prove the main Claim. The argument is very similar to the above
induction, except that we want to consider (𝑥 1 , . . . , 𝑥𝑟−1 ) such that not only
𝜑 𝑠 (𝑥𝑟−1 ) ∈ 𝑋𝑟−1 , but also 𝜑 𝑠 ({𝑥 1 , . . . , 𝑥𝑟−1 }) ∈ N𝐻𝑠−1 (𝑣). Here we need to as-

128
 4 Packing Degenerate Hypergraphs

˜
sume 𝑣 ∉ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]). We define similarly as above

𝜔 𝑠 ({𝑥 1 , . . . , 𝑥𝑟−1 }) 1𝜑𝑠 (𝑡)∈𝑋𝑖 ,

Õ
𝑌𝑡0 =
(𝑥 1 ,...,𝑥𝑟 −1 )

where the sum runs over vectors with 𝑥1 < 𝑥 2 < · · · < 𝑥𝑟−1 members of [𝑛 − 𝛿𝑛], ˜
such that 𝜑 𝑠 (𝑥 𝑗 ) ∈ 𝑋 𝑗 for each 1 ≤ 𝑗 ≤ 𝑟 − 1, and 𝜑 𝑠 ({𝑥1 , . . . , 𝑥𝑟−1 }) ∈ N𝐻𝑠−1 (𝑣),
and 𝑥𝑟−1 = 𝑡. For any 𝑡, and any 𝑥 1 < 𝑥2 < · · · < 𝑥𝑟−2 < 𝑡, we have

P[𝜑 𝑠 (𝑡) ∈ 𝑋𝑟−1 and 𝜑 𝑠 ({𝑥 1 , . . . , 𝑥𝑟−2 , 𝑡}) ∈ N𝐻𝑠−1 (𝑣)] =

deg− (𝑡)+1 𝑛+1−𝑡
(1 ± 2𝐶𝛼𝑠 ) 𝑝 𝑠−1 𝑛 |𝑋𝑟−1 |
= −
deg (𝑡)
= (1 ± 4𝐶𝛼𝑠 ) 𝑝 𝑠−1 |𝑋𝑟𝑛−1 | ,
(1 ± 𝐶𝛼𝑠 ) 𝑝 𝑠−1 (𝑛 + 1 − 𝑡)

where we obtain the extra factor of 𝑝 𝑠−1 in the numerator by looking at the common
− (𝑡) and the extra semi-edge {𝜑 (𝑥 ), . . . , 𝜑 (𝑥


neighbourhood of 𝜑 𝑠 N𝐺 𝑠
𝑠 1 𝑠 𝑟−2 ), 𝑣}.
−


Note that this extra semi-edge is not in the set 𝜑 𝑠 N𝐺 𝑠 (𝑡) since it contains 𝑣 and
˜
𝑣 is not in 𝜑 𝑠 ([𝑛 − 𝛿𝑛]). An essentially identical application of Corollary 4.2.2 as
above now gives us the claim. 

Next, we show that the embedding of the rest of 𝐺 𝑠 does not have much effect.
Claim 4.7.4. When PackingProcess2 is run up to and including the embedding of
˜
𝐺 𝑠 , with probability at least 1 − 4𝑛−3𝐶 , we have either 𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]) or
Õ
𝑤 𝑠 (e) = 1 ± 10𝑟𝐶𝛼𝑠 𝑝 𝑠 · |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 ) .


e∈N 𝐻𝑠 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )

Proof. Let E be the intersection of the likely event of Claim 4.7.3 and the event
that RandomEmbedding, run to embed 𝐺 𝑠 into 𝐻𝑠−1 , succeeds and for each


1 ≤ 𝑡 ≤ 𝑣(𝐺 𝑠 ) the pair 𝐻𝑠−1 , 𝜑 𝑠 ([𝑡]) satisfies the (𝐶𝛼𝑠 , 𝑄)-diet condition.
By Lemma 4.4.3 and Claim 4.7.3, E occurs with probability at least 1 − 4𝑛−3𝐶 .
Suppose that when RandomEmbedding embeds 𝐺 𝑠 into 𝐻𝑠−1 , at some time 𝑡 we
˜ then the claim statement is true trivially, so we may
have 𝜑 𝑠 (𝑡) = 𝑣. If 𝑡 ≤ 𝑛 − 𝛿𝑛,
assume 𝑡 > 𝑛 − 𝛿𝑛.˜ We now consider the embeddings of vertices from 𝑡 onwards.
Each vertex that we embed completes the embedding of at most 𝐷 edges of 𝐺 𝑠 , and
therefore is responsible for at most 𝐷 semi-edges of N𝐻𝑠−1 (𝑣) not being in N𝐻𝑠 (𝑣).
𝐷𝑐𝑛
These 𝐷 semi-edges have weight at most log 𝑛.
𝑐𝑛 𝑐𝑛
Since 𝑡 has degree at most log 𝑛 in 𝐺 𝑠 , there are at most log 𝑛 vertices of 𝐺 𝑠

129
 4 Packing Degenerate Hypergraphs

which are in an edge with 𝑡, and only when we embed one of these vertices can
we remove a semi-edge from N𝐻𝑠−1 (𝑣). Provided that we are in the event E, at
each time 𝑡 0 when we embed such a vertex, by the diet condition we do so to a set
of size at least 21 𝛾 𝐷 𝛿𝑛. The expected weight of the semi-edges we remove from
N𝐻𝑠−1 (𝑣) by one such vertex embedding, conditioning on the previous embedding
history, is thus at most 2𝛾 −𝐷 𝛿−1 sum(𝜔 𝑠 )𝑛−1 . We apply Corollary 4.2.2, with
𝜇˜ = 𝜈˜ = 2𝛾 −𝐷 𝛿−1 sum(𝜔 𝑠 )𝑛−1 · log
𝑐𝑛
𝑛 and 𝜚˜ = 𝜀sum(𝜔 𝑠 ), to bound the probability
that E occurs and the total weight removed, i.e.
Õ
𝜔 𝑠 (e)
e∈N 𝐻𝑠−1 (𝑣)\N 𝐻𝑠 (𝑣)

exceeds 2𝜀sum(𝜔 𝑠 ). We obtain the upper bound

𝜚˜ 2
< 𝑛−5𝐶 ,


2 exp − 𝐷𝑐𝑛
2 log 𝑛 (2 𝜇˜ + 𝜚)
˜

where the inequality uses the choice of 𝑐. This bound, together with the choice
of 𝜀, and the facts 𝑝 𝑠−1 − 𝑝 𝑠 ≤ 𝐷𝑛/ 𝑛𝑟 , |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟 ≥ 𝛾 𝑟−1 , and 𝑝 𝑠 ≥ 𝛾,

establish the claim. 

We now need to deal with the embedding of 𝐺 𝑠+1 , . . . , 𝐺 𝑠∗ . What we want to do

is establish the following claim.
Claim 4.7.5. With probability at least 1 − 𝑛−𝐾−2𝑟 , for each 𝑠 ≤ 𝑠0 ≤ 𝑠∗ , either
˜
𝑣 ∈ 𝜑 𝑠 ([𝑛 − 𝛿𝑛]) or we have
Õ
𝑤 𝑠 (e) = 1 ± 30𝑟𝐶𝛼𝑠 0 𝑝 𝑠 0 · |𝑋1 | . . . |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 ) .


e∈N 𝐻𝑠 0 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )

Proof. Observe that to establish this claim, by the union bound it is enough to show
that, for a given 𝑠0 with 𝑠 < 𝑠0 ≤ 𝑠∗ , with probability at most 𝑛−𝐾−5𝑟 we have that 𝑠0
is minimal such that (4.7.1) fails, since Claim 4.7.4 already establishes this bound
( 𝐻)
for 𝑠0 = 𝑠. Furthermore, defining a weight function 𝜔0 on 𝑉𝑟−1
b

by
Õ
𝜔0 (𝑦) = 𝜔 𝑠 (e)
e∈N 𝐻𝑠 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )
e is an ordering of 𝑦

130
 4 Packing Degenerate Hypergraphs

we see that the equation of Claim 4.7.5 is equivalent to

Õ
𝜔0 (𝑦) = 1 ± 30𝑟𝐶𝛼𝑠 0 𝑝 𝑠 0 · |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 ) ,


(4.7.1)
𝑦∈N 𝐻𝑠 0 (𝑣)

and we will aim to prove this version, in which we simply work with weights on
semi-edges of 𝐻.
b
Let, for each 𝑠 < 𝑖 ≤ 𝑠∗ ,
Õ
𝑌𝑖 = 𝜔0 (𝑦) .
𝑦∈N 𝐻𝑖−1 (𝑣)\N 𝐻𝑖 (𝑣)

To begin with, we observe that, provided (4.7.1) holds for 𝑖−1 and 𝐻𝑖−1 is (𝛼𝑖−1 , 𝑄)-
quasirandom, we can calculate E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0].
Claim 4.7.6. Suppose that for some 𝑠 < 𝑖 ≤ 𝑠∗ , (4.7.1) holds for 𝑖 − 1 and 𝐻𝑖−1 is
(𝛼𝑖−1 , 𝑄)-quasirandom. Then we have

E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] = 1 ± 104𝐶𝑟𝛼𝑖−1 𝑄𝛿−1 𝑟!|𝑋1 | · · · |𝑋𝑟−1 |𝑛1−2𝑟 sum(𝜔 𝑠 )𝑒(𝐺 𝑖 ) .

Proof. By definition of 𝑌𝑖 we have

𝜔0 (𝑦) · 1𝑦𝑣 is used when embedding 𝐺 𝑖

Õ
𝑌𝑖 =
𝑦∈N 𝐻𝑖−1 (𝑣)

and therefore
Õ
E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] = 𝜔0 (𝑦) · P[𝑦𝑣 is used when embedding 𝐺 𝑖 | 𝐻𝑖−1 , 𝜑 𝑠 ] .
𝑦∈N 𝐻𝑖−1 (𝑣)

Since 𝐻𝑖−1 is (𝛼𝑖−1 , 𝑄)-quasirandom, Lemma 4.5.5 yields

Õ 𝑟!𝑒(𝐺 𝑖 )
E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] = 𝜔0 (𝑦) · (1 ± 200𝐶𝑟𝛼𝑖−1 𝛿−1 ) 2+2𝐷 ,
𝑝𝑖−1 𝑛𝑟
𝑦∈N 𝐻𝑖−1 (𝑣)

131
 4 Packing Degenerate Hypergraphs

and since (4.7.1) holds for 𝑖 − 1 we get

E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] = 1 ± 30𝑟𝐶𝛼𝑖−1 𝑝𝑖−1 |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 )

𝑟!𝑒(𝐺 𝑖 )
· (1 ± 200𝐶𝑟𝛼𝑖−1 𝛿−1 ) 2+2𝐷
𝑝𝑖−1 𝑛𝑟
= 1 ± 104𝐶𝑟𝛼𝑖−1 𝑄𝛿−1 𝑟!|𝑋1 | · · · |𝑋𝑟−1 |𝑛1−2𝑟 sum(𝜔 𝑠 )𝑒(𝐺 𝑖 ) .




Í𝑠∗
What we would like to do now is to use Corollary 4.2.2 to estimate 𝑖=𝑠+1 𝑌𝑖 .
However the range of the 𝑌𝑖 might be too large for this. To that end, we define, for
each 𝑠 + 1 ≤ 𝑖 ≤ 𝑠∗ , the random variables

0 sum(𝜔 𝑠 )
 
𝑌𝑖0 := min 𝑌𝑖 , 2𝐾 𝑐 with 𝐾 0 = 1010 𝐾𝛾 −1𝑟 2𝐶𝐷 3 𝛿−1 .
log 𝑛

The ‘capped’ random variable 𝑌𝑖0 trivially does not have an excessively large range,
Í𝑠∗
and so we can use Corollary 4.2.2 to show that 𝑖=𝑠+1 𝑌𝑖0 is very likely to be
close to the corresponding sum of conditional expectations. For 𝑟 ≥ 3, we will
now prove that for each 𝑖, with very high probability we have 𝑌𝑖0 = 𝑌𝑖 , so that
in particular (by the union bound) the statement holds simultaneously for all 𝑠 <
𝑖 ≤ 𝑠∗ and E[𝑌𝑖0 |𝐻𝑖−1 , 𝜔0] is very close to E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0]. This is what we need
to establish (QUASI 3). We should note that our proof does not go through for
𝑟 = 2, and indeed for 𝑟 = 2 one can find examples where it is likely that 𝑌𝑖0 ≠ 𝑌𝑖 for
some 𝑖. For a specific class of weight functions 𝜔 𝑠 , and with 𝑋1 = 𝑉 ( 𝐻),b in [2] a
different proof of (QUASI 3) is given; the approach there should generalise to prove
the entire 𝑟 = 2 case of (QUASI 3), but we have not checked the details.
𝑠)
Let us define the event CapE (𝑖) as 𝑌𝑖 > 2𝐾 0 𝑐 sum(𝜔
log 𝑛 . This means that the set
N𝐻𝑖−1 (𝑢) \ N𝐻𝑖 (𝑢) of semi-edges has a large total weight. We want to separate cases
when this comes from few edges and when this comes from many edges. Therefore
we define the set
 
0 sum(𝜔 𝑠 )
𝑊 = 𝑦 ∈ N𝐻𝑖−1 (𝑣) 𝜔 (𝑦) >
(log 𝑛) 10

as the heavy semi-edges and

 
sum(𝜔 𝑠 )
𝐿 = 𝑦 ∈ N𝐻𝑖−1 (𝑣) ≥ 𝜔0 (𝑦) > 0
(log 𝑛) 10

𝑠)
𝜔0 (𝑦) > 𝐾 0 𝑐 sum(𝜔
Í
as the light semi-edges. We define CapE1(𝑖) as 𝑦∈𝑊\N 𝐻𝑖 (𝑣) log 𝑛

132
 4 Packing Degenerate Hypergraphs

meaning that the heavy edges hit the cap and CapE2(𝑖) as 𝑦∈𝐿\N𝐻𝑖 (𝑣) 𝜔0 (𝑦) >
Í
𝑠)
𝐾 0 𝑐 sum(𝜔
log 𝑛 meaning that the light edges hit the cap. Clearly CapE (𝑖) implies at
least one of CapE1(𝑖) and CapE2(𝑖) holds. By definition of sum(𝜔 𝑠 ), we have
|𝑊 | ≤ (log 𝑛) 10 .
In order to prove CapE (𝑖) is unlikely, it is enough to prove that each of CapE1(𝑖)
and CapE2(𝑖) is unlikely. We begin with CapE1(𝑖). The idea here is that in order
for CapE (𝑖) to occur, a reasonably large matching of semi-edges of 𝑊 must form
edges with 𝑣 which are used in embedding 𝐺 𝑖 ; we prove that any one such large
matching is unlikely and take the union bound over all such matchings. The next
claim justifies that a large matching must be chosen.
Claim 4.7.7. Let 𝑆 be a set of semi-edges of 𝐻 b such that Í𝑦∈𝑆 𝜔0 (𝑦) ≥ 𝑎(𝑟 −
1)Δ(𝜔 𝑠 ). There is a subset 𝑆0 ⊆ 𝑆 such that |𝑆0 | = 𝑎 and 𝑆0 is a matching.

Proof. We greedily pick semi-edges 𝑦 from 𝑆, add them to 𝑆0 and remove any
semi-edge from 𝑆 that shares at least one vertex with 𝑦. Since a semi-edge has 𝑟 − 1
vertices, in each step we remove semi-edges with total weight at most (𝑟 − 1)Δ(𝜔 𝑠 ),
and therefore the number of steps before removing all semi-edges from 𝑆 is at least
𝑎. If we stop after 𝑎 steps, we have |𝑆0 | = 𝑎, and by construction 𝑆0 is a matching.


We now show that CapE1(𝑖) is unlikely.

Claim 4.7.8. Suppose that 𝐻𝑖−1 is (𝛼𝑖−1 , 𝑄)-quasirandom and 𝑟 > 2. Then we
have
P CapE1(𝑖) 𝐻𝑖−1 , 𝜔0 ≤ 𝑛−𝐾−10𝑟 .
 

Proof. Let 𝜓1 , . . . , 𝜓𝑛−𝛿𝑛 the embedding sequence generated by RandomEmbed-

ding when called to embed 𝐺 𝑖 into 𝐻𝑖−1 . Let E be the event that for each


1 ≤ 𝑗 ≤ 𝑛 − 𝛿𝑛 the (𝐶𝛼, 𝑄)-diet condition holds for 𝐻𝑖−1 , im(𝜓 𝑗 ) . By
Lemma 4.4.3 we have
P E 𝐻𝑖−1 , 𝜔0 ≥ 1 − 𝑛−5𝐶 .
 

For the following calculations we assume E holds.

𝑠)
By Claim 4.7.7, since 𝐾 0 𝑐 sum(𝜔 log 𝑛 > (𝐾 + 20𝑟)(𝑟 − 1)Δ(𝜔 𝑠 ), if CapE1(𝑖) holds
then there is a matching of 𝑎 := 𝐾+20𝑟 semi-edges 𝑦 1 = (𝑦 1,1 , . . . , 𝑦 1,𝑟−1 ), . . . , 𝑦 𝑎 =
(𝑦 𝑎,1 , . . . , 𝑦 𝑎,𝑟−1 ) ∈ 𝑊 such that 𝑦 𝑗 𝑣 is an edge of 𝐻𝑖−1 which is used in the
embedding of 𝐺 𝑖 for each 1 ≤ 𝑗 ≤ 𝑎. That is, there exists 𝑥 ∈ 𝑉 (𝐺 𝑖 ) and
a matching of semi-edges 𝑧 1 = (𝑧1,1 , . . . , 𝑧 1,𝑟−1 ), . . . , 𝑧 𝑎 = (𝑧 𝑎,1 , . . . , 𝑧 𝑎,𝑟−1 ) in

133
 4 Packing Degenerate Hypergraphs

N𝐺 𝑖 (𝑥), such that 𝜑𝑖 maps 𝑥 to 𝑣, and 𝑧 𝑗,ℓ to 𝑦 𝑗,ℓ for each 𝑗 ∈ [𝑎] and ℓ ∈ [𝑟 − 1].
We fix such a choice, and will later take the union bound over all possible such
choices. Let 𝐸 denote the event that 𝜑𝑖 (𝑧 𝑗 , ℓ) = 𝑦 𝑗,ℓ for each 𝑗 ∈ [𝑎] and ℓ ∈ [𝑟 −1].
We order the pairs (𝑧 𝑗,𝑘 , 𝑦 𝑗,𝑘 ) by increasing 𝑧 𝑗,𝑘 and for convenience rename
them to pairs ((𝑥 1 , 𝑦01 ), . . . , (𝑥 𝑎(𝑟−1) , 𝑦0𝑎(𝑟−1) )). Let, for each ℓ ∈ [𝑎(𝑟 − 1)], the
event 𝐸 ℓ be that 𝜑𝑖 (𝑥ℓ ) = 𝑦0ℓ . Let 𝐸˜ ℓ denote ∩ℓℎ=1 𝐸 ℎ . Since E holds, we have

P 𝐸 ℓ 𝐻𝑖−1 , 𝜔0, 𝐸˜ ℓ−1 ≤ 2𝛾 −𝐷 𝛿−1 𝑛−1 .

 

This implies

𝑎(𝑟−1)
Ö
0
P 𝐸 ℓ 𝐻𝑖−1 , 𝜔0, 𝐸˜ ℓ−1
 
P[𝐸 |𝐻𝑖−1 , 𝜔 ] = 1 − P[E |𝐻 𝑖−1,𝜔 0 ]+
ℓ=1
−5𝐶 −𝐷 −1 −1
𝑎(𝑟−1) 1 𝐷

−𝑎(𝑟−1)
≤𝑛 + 2𝛾 𝛿 𝑛 ≤ 4 𝛾 𝛿𝑛 .

Taking the union bound over choices of 𝑥, 𝑦 𝑗,ℓ and 𝑧 𝑗,ℓ , we see

  
0 |𝑊 | N𝐺 𝑖 (𝑥)
−𝑎(𝑟−1)
((𝑟 − 1)!) 𝑎 · 14 𝛾 𝐷 𝛿𝑛
 
P CapE1(𝑖) 𝐻𝑖−1 , 𝜔 ≤𝑛
𝑎 𝑎
−𝐷 (𝑟−1) 1−𝑟
𝑎
10𝑎 𝑐𝑛 𝑎 𝑟−1
· 𝑛−𝑎(𝑟−1)


≤ 𝑛(log 𝑛) · log 𝑛 4 𝑟!𝛾 𝛿
≤ 𝑛2−𝑎(𝑟−2) ≤ 𝑛−𝐾−10𝑟 ,

as desired. 

To deal with CapE2(𝑖), we also consider the vertex-by-vertex embedding of

𝐺 𝑖 . The critical point is that each vertex embedding embeds at most 𝐷 edges of
𝐺 𝑖 , and therefore is responsible for removing at most 𝐷 light semi-edges from
N𝐻𝑖 (𝑣). Since these semi-edges are light, the maximum weight removed is at
most 𝐷 · sum(𝜔 𝑠 )(log 𝑛) −10 , and the extra log factors here (as compared to heavy
semi-edges) give us better concentration.
Claim 4.7.9. Suppose that 𝐻𝑖−1 is (𝛼𝑖−1 , 𝑄)-quasirandom and 𝑟 > 2. Then we
have
P CapE2(𝑖) 𝐻𝑖−1 , 𝜔0 ≤ 𝑛−𝐾−10𝑟 .
 

Proof. Again, let 𝜓1 , . . . , 𝜓𝑛−𝛿𝑛 the embedding sequence generated by Ran-

domEmbedding when called to embed 𝐺 𝑖 into 𝐻𝑖−1 . Let E be the event that

134
 4 Packing Degenerate Hypergraphs

for each 1 ≤ 𝑗 ≤ 𝑛 − 𝛿𝑛 the (𝐶𝛼, 𝑄)-diet condition holds for 𝐻𝑖−1 , im(𝜓 𝑗 ) . By
Lemma 4.4.3 we have
P E 𝐻𝑖−1 , 𝜔0 ≥ 1 − 𝑛−5𝐶 ,
 

and for the following calculations we assume E holds.

Let us fix 𝑥 ∈ 𝑉 (𝐺 𝑖 ) and assume 𝜑𝑖 (𝑥) = 𝑣; we will later take a union bound
over the choices of 𝑥.
Let 𝑥 1 < · · · < 𝑥ℓ be the vertices of 𝐺 𝑖 which are maximum vertices of some
semi-edge in N𝐺 𝑖 (𝑥) and which are greater than 𝑥. Note that there can be some at
most 𝐷 semi-edges of N𝐺 𝑖 (𝑥) all of whose vertices are smaller than 𝑥. Trivially
ℓ ≤ Δ. For each 𝑥 𝑗 let 𝑆 𝑗 be the set of semi-edges in N𝐺 𝑖 (𝑥) whose last element is
𝑥 𝑗 , and note |𝑆 𝑗 | ≤ 𝐷 since 𝐺 𝑖 is 𝐷-degenerate. We define
Õ
𝑋 𝑗 := 𝜔0 (𝑦) .
𝑦∈𝜑𝑖 (𝑆 𝑗 )∩𝐿

We have 0 ≤ 𝑋 𝑗 ≤ 𝐷 · sum(𝜔 𝑠 )(log 𝑛) −10 for each 𝑗. Taking account of the at

most 𝐷 light semi-edges which come entirely before 𝑥, we see that in order for
CapE2(𝑖) to occur, we need

ℓ
Õ
𝑠)
𝐷 · sum(𝜔 𝑠 )(log 𝑛) −10
+ 𝑋 𝑗 > 𝐾 0 𝑐 sum(𝜔
log 𝑛 ,
𝑗=1

and so ℓ𝑗=1 𝑋 𝑗 > 12 𝐾 0 𝑐 · sum(𝜔 𝑠 )(log 𝑛) −1 . Our aim is now to show that this is
Í

unlikely.
When we embed 𝑥 𝑗 , by the diet condition we embed it uniformly at random to
a set of size at least 12 𝛾 𝐷 𝛿𝑛 vertices; at most, all of sum(𝜔 𝑠 ) could be distributed
over these vertices (making semi-edges with the remaining embedded parts of 𝑆 𝑗 )
and so we have

E[𝑋 𝑗 |𝐻𝑖−1 , 𝜔0, 𝜓𝑥 𝑗 −1 ] ≤ sum(𝜔 𝑠 ) · 2𝛾 −𝐷 𝛿−1 𝑛−1 .

Summing over 𝑗, within E we get

ℓ
Õ 𝑐𝑛 𝑠)
E[𝑋 𝑗 |𝐻𝑖−1 , 𝜔0, 𝜓𝑥 𝑗 −1 ] ≤ · sum(𝜔 𝑠 ) · 2𝛾 −𝐷 𝛿−1 𝑛−1 sum(𝜔 𝑠 ) ≤ 41 𝐾 0 𝑐 sum(𝜔
log 𝑛
𝑗=1
log 𝑛

by choice of 𝐾 0.

135
 4 Packing Degenerate Hypergraphs

𝑠)
Applying Corollary 4.2.2(a ) with the given E, with 𝜇˜ = 14 𝐾 0 𝑐 sum(𝜔
log 𝑛 , and with
𝑅 = 𝐷 · sum(𝜔 𝑠 )(log 𝑛) −10 we then obtain that conditioning on 𝐻𝑖−1 and 𝜔 𝑠 we
have
 ℓ  1 0 sum(𝜔 𝑠 )
Õ
sum(𝜔 ) 4 𝐾 𝑐 log 𝑛
0
 1

P E and 𝑋 𝑗 ≥ 4 𝐾 𝑐 log 𝑛  ≤ 2 exp ­−
𝑠 © ª
−10
®
 𝑗=1  4𝐷 · sum(𝜔 𝑠 )(log 𝑛)
  « ¬
−𝐾−20𝑟
≤𝑛 ,

and by the union bound over choices of 𝑥 we have the claim. 

By Claims 4.7.8 and 4.7.9, we see that for each 𝑠 < 𝑖 ≤ 𝑠∗ , provided that
PackingProcess2 behaves, we have P[ CapE (𝑖)|𝐻𝑖−1 , 𝜔0] ≤ 2𝑛−𝐾−10𝑟 . Thus, by the
union bound, with probability at least 1 − 𝑛−3𝐶 − 2𝑛−𝐾−9𝑟 we have 𝑌𝑖0 = 𝑌𝑖 for each
𝑠 < 𝑖 ≤ 𝑠∗ . In addition, since 𝑌𝑖 ≤ sum(𝜔 𝑠 ) trivially, provided PackingProcess2
behaves we have

E[𝑌𝑖0 |𝐻𝑖−1 , 𝜔0] > E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] + (𝑛−3𝐶 − 2𝑛−𝐾−9𝑟 ) · sum(𝜔 𝑠 )
> (1 + 𝑛−4 )E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0] .

By Corollary 4.2.2, and since sum(𝜔 𝑠 ) ≥ 𝛾𝑛, with probability at least 1−𝑛−𝐾−6𝑟 ,
conditioning on 𝜔0 and 𝐻𝑠 , if PackingProcess2 behaves and the good event of
Claim 4.7.4 occurs, we have

𝑠0
Õ 𝑠0
Õ
𝑌𝑖0 = (1 ± 𝜀) E[𝑌𝑖0 |𝐻𝑖−1 , 𝜔0] ,
𝑖=𝑠+1 𝑖=𝑠+1

and we claim that this equation is inconsistent with 𝑠0 being the first failure time
of (4.7.1). Summing up the failure probabilities, this then establishes the Claim.
Supposing that (4.7.1) holds up to and including 𝑠0 − 1 and the likely events

136
 4 Packing Degenerate Hypergraphs

above occur, we have

Õ
𝜔0 (𝑦)
𝑦∈N 𝐻𝑠 0 (𝑣)

Õ 𝑠0
Õ
0
= 𝜔 (𝑦) − 𝑌𝑖
𝑦∈N 𝐻𝑠 (𝑣) 𝑖=𝑠+1

Õ 𝑠0
Õ
0
= 𝜔 (𝑦) − (1 ± 𝜀) E[𝑌𝑖0 |𝐻𝑖−1 , 𝜔0]
𝑦∈N 𝐻𝑠 (𝑣) 𝑖=𝑠+1

Õ 𝑠0
Õ
0 −4
= 𝜔 (𝑦) − (1 ± 𝜀)(1 ± 𝑛 ) E[𝑌𝑖 |𝐻𝑖−1 , 𝜔0]
𝑦∈N 𝐻𝑠 (𝑣) 𝑖=𝑠+1
Õ
= 𝜔0 (𝑦)
𝑦∈N 𝐻𝑠 (𝑣)
𝑠0
Õ
1 ± 104𝐶𝑟𝛼𝑖−1 𝑄𝛿−1 𝑟!|𝑋1 | · · · |𝑋𝑟−1 |𝑛1−2𝑟 sum(𝜔 𝑠 )𝑒(𝐺 𝑖 ) ,


− (1 ± 2𝜀)
𝑖=𝑠+1

and the main term of this last line is, by the good event of Claim 4.7.4,

𝑠0
Õ
1−𝑟 1−2𝑟
𝑝 𝑠 · |𝑋1 | · · · |𝑋𝑟−1 |𝑛 sum(𝜔 𝑠 ) − 𝑟!|𝑋1 | · · · |𝑋𝑟−1 |𝑛 sum(𝜔 𝑠 ) 𝑒(𝐺 𝑖 )
𝑖=𝑠+1
1−𝑟
= (1 ± 𝜀) 𝑝 𝑠 0 |𝑋1 | · · · |𝑋𝑟−1 |𝑛 · sum(𝜔 𝑠 ) ,

which is as desired. What remains is to estimate the error term

10𝑟𝐶𝛼𝑠 𝑝 𝑠 |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟 sum(𝜔 𝑠 )

𝑠0
Õ
+ 105𝐶𝑟𝛼𝑖−1 𝑄𝛿−1𝑟!|𝑋1 | · · · |𝑋𝑟−1 |𝑛1−2𝑟 sum(𝜔 𝑠 )𝑒(𝐺 𝑖 )
𝑖=𝑠+1
 𝑠0
Õ 
≤|𝑋1 | · · · |𝑋𝑟 |𝑛 1−𝑟
sum(𝜔 𝑠 ) · 10𝑟𝐶𝛼𝑠 𝑝 𝑠 + 105𝐶𝐷𝑟𝛼𝑖−1 𝑄𝛿−1𝑟!𝑛1−𝑟
𝑖=𝑠+1
 ∫ 𝑠0 
1−𝑟 5 −1 1−𝑟
≤|𝑋1 | · · · |𝑋𝑟 |𝑛 sum(𝜔 𝑠 ) · 10𝑟𝐶𝛼𝑠 𝑝 𝑠 + 10 𝐶𝐷𝑟𝛼𝑥 𝑄𝛿 𝑟!𝑛 d𝑥
𝑥=−∞
 
≤|𝑋1 | · · · |𝑋𝑟 |𝑛1−𝑟 sum(𝜔 𝑠 ) 10𝑟𝐶𝛼𝑠 𝑝 𝑠 + 𝛾𝛼𝑠 0

≤20𝑟𝐶𝛼𝑠 𝑝 𝑠 |𝑋1 | · · · |𝑋𝑟 |𝑛1−𝑟 sum(𝜔 𝑠 ) ,

where the penultimate line is by choice of the 𝛼𝑥 .

137
 4 Packing Degenerate Hypergraphs

Putting these together, the above likely events imply

Õ
𝜔0 (𝑦) =(1 ± 𝜀) 𝑝 𝑠 0 |𝑋1 | · · · |𝑋𝑟−1 |𝑛1−𝑟
𝑦∈N 𝐻𝑠 0 (𝑣)

· sum(𝜔 𝑠 ) ± 20𝑟𝐶𝛼𝑠 𝑝 𝑠 |𝑋1 | · · · |𝑋𝑟 |𝑛1−𝑟 sum(𝜔 𝑠 ) ,

which is precisely the statement that (4.7.1) holds for 𝑠0. This completes the proof
of Claim 4.7.5. 

Finally, Claim 4.7.5, for 𝑠0 = 𝑠∗ , together with the union bound over 𝑣 and
𝑠 ∈ [𝑠∗ ], establishes (QUASI 3).

4.8 Perfect packings

In this section we prove Theorem 4.1.4. We deduce Theorem 4.1.4 from The-
orem 4.1.2 and the following technical result.
Theorem 4.8.1. For every 𝐷 ≥ 1 and 𝑟 ≥ 3, and every sufficiently small 𝜇 > 0
there are 𝑛0 and 𝜉, 𝑐 > 0 such that for every 𝑛 ≥ 𝑛0 the following holds. Suppose
b is a 𝑟-uniform (𝜉, 𝐷 (𝑟 + 1) + 3)-quasirandom hypergraph with 𝑛 vertices.
that 𝐻
Suppose that 𝑠∗ = b𝜇𝑛𝑟−1 c, and that the hypergraphs (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] all have maximum
𝑐𝑛
degree log 𝑛 , have exactly 𝑛 − b𝜇𝑛c vertices, and have at least 𝜇𝑛 vertices of degree
Í b Then (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] packs into 𝐻.
1. Suppose further that 𝑠∈[𝑠∗ ] 𝑒(𝐺 𝑠 ) = 𝑒( 𝐻). b

Before sketching the proof of Theorem 4.8.1, we show that it, together with
Theorem 4.1.2, implies Theorem 4.1.4.

Proof of Theorem 4.1.4. Without loss of generality, we may assume 𝜇 is sufficiently

small for Theorem 4.8.1 for input 𝐷 and 𝑟. Given 𝐷, 𝑟, 𝜇, let 𝛾 0 be small enough
to play the rôle of 𝜉 in Theorem 4.8.1 for input 𝐷, 𝜇. Suppose that 𝑛0 is also large
enough, and 𝜉, 𝑐 > 0 are small enough, for this application of Theorem 4.8.1 and
also for Theorem 4.1.2 for input 𝑟, 𝐷 and 𝛾 = min(𝛾 0, 12 𝜇2 ).
Let 𝑛 and the hypergraphs 𝐻 b and (𝐺 𝑖 )𝑖∈[𝑚] be given. Without loss of generality,
Í
we may assume 𝑖∈[𝑚] 𝑒(𝐺 𝑖 ) = 𝑒( 𝐻),
b since otherwise we can add new hypergraphs
consisting of single edges to our family until this condition is satisfied. We can
further assume that the b𝜇𝑛𝑟−1 c special hypergraphs come last in the sequence.
We apply Theorem 4.1.2, with input as above, to pack the hypergraphs 𝐺 𝑖 with
1 ≤ 𝑖 ≤ 𝑠∗ − b𝜇𝑛𝑟−1 c into 𝐻,b and let 𝐻e be the hypergraph of leftover edges. By
Theorem 4.1.2 such a packing exists and 𝐻 e is (𝛾, 4𝐷 (𝑟 + 1)𝑟 + 3)-quasirandom.

138
 4 Packing Degenerate Hypergraphs

We now create a sequence 𝐺 01 , . . . , 𝐺 0𝑠 0 of hypergraphs with 𝑠0 := b𝜇𝑛𝑟−1 c by

taking the special hypergraphs 𝐺 𝑠∗ −b𝜇𝑛𝑟 −1 c+1 , . . . , 𝐺 𝑠∗ and adding to each isol-
ated vertices until they have 𝑛 − b𝜇𝑛c vertices. By Theorem 4.8.1 we can pack
(𝐺 0𝑖 )𝑖∈[𝑠 0] into 𝐻.e Ignoring the extra isolated vertices, this gives us a packing of
𝐺 𝑠∗ −b𝜇𝑛𝑟 −1 c+1 , . . . , 𝐺 𝑠∗ into 𝐻
e and therefore a packing of 𝐺 1 , . . . , 𝐺 𝑠∗ into 𝐻
b as
desired.

We now sketch the proof of Theorem 4.8.1. For this, we need to define various
constants, which we will take as fixed throughout this section. We do not provide
explicit dependencies for all constants (mainly because Lemma 4.2.9, which we
used to prove Lemma 4.2.6, does not provide explicit dependencies), but explain
now how to choose them.

Definition 4.8.2 (Choice of constants). The constants 𝐷, 𝑟 and 𝜇 are provided

as input to Theorem 4.8.1, with the assumption that 𝜇 is sufficiently small for
Lemma 4.2.6. We let 𝜚 be returned by Lemma 4.2.6 for input 𝜇. We choose
𝜈 ≤ 10−3𝑟 −2 𝜚𝜇 sufficiently small to play the rôle of 𝜀 in Lemma 4.2.6.
1
We choose 𝛾 ≤ 100 (𝑟!) −2𝑟 𝜇3𝑟 𝜈 2𝑟 2−3𝑟 𝑟 −3𝑟 sufficiently small for Lemma 4.8.5
below, and given 𝛾 we let 𝜉 and 𝑐 be sufficiently small for Lemma 4.8.5 with input
𝐷, 𝑟, 𝜇, 𝜈 and 𝛾, and in addition we let 𝑐 be sufficiently small for various explicit
calculations in this section. Finally, we suppose 𝑛0 is large enough for Lemma 4.8.5
1
below for the given inputs, and also large enough such that 8𝑟!𝑟 𝜇𝜈𝑛0 is large enough
to play the rôle of 𝑚 in Lemma 4.2.6 with inputs as above, and sufficiently large
for various explicit calculations in this section.
Given 𝑛 ≥ 𝑛0 , the density 𝑝 is determined by
  −1
𝑛
= 1 ± 21 𝜇𝜈
𝑟−1

𝑝 = b𝜇𝑛 c b𝜈𝑛c 𝑟! . (4.8.1)
𝑟

In what follows, we will assume 𝑠∗ := b𝜇𝑛𝑟−1 c.

We start by creating an almost perfect packing, which omits ℓ leaves (i.e. vertices
of degree 1) in each hypergraph.

Definition 4.8.3 (corresponding subgraph sequence). Given a sequence (𝐺 𝑠 ) 𝑠∈[𝑠∗ ]

of hypergraphs, we say that (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] is a corresponding subgraph sequence
omitting ℓ leaves if for each 𝑠 ∈ [𝑠∗ ] we have 𝐺 0𝑠 = 𝐺 𝑠 − 𝑉𝑠 + 𝐼 𝑠 for a strongly
independent set 𝑉𝑠 of leaves in 𝐺 𝑠 with |𝑉𝑠 | = ℓ, and a set 𝐼 𝑠 of new and isolated
vertices with |𝐼 𝑠 | = ℓ.

139
 4 Packing Degenerate Hypergraphs

The addition of the isolated set 𝐼 𝑠 is purely for technical reasons: it guarantees
that the 𝐺 0𝑠 have exactly 𝑛 − 𝜇𝑛 vertices, which allows for easier formulae than the
𝑛 − 𝜇𝑛 − ℓ we would otherwise have. The restriction that the set 𝑉𝑠 is strongly
independent means that there is a bijection between the set 𝑉𝑠 and the set of edges
of 𝐺 𝑠 which are not in 𝐺 0𝑠 ; we do not pick two leaves which lie in the same edge of
𝐺 𝑠.
We first find a packing (𝜑0𝑠 ) 𝑠∈[𝑠∗ ] of the corresponding subgraph sequence
(𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] with the additional properties guaranteed by Lemma 4.8.5 below. We
will apply Lemma 4.8.5 (which is a corollary of Theorem 4.7.1) with the weight
function on 𝐺 0𝑠 defined as follows.

Definition 4.8.4 (weights). Let (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] be a sequence of 𝑛-vertex hypergraphs

each with a strongly independent set 𝑉𝑠 of ℓ leaves, and (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] be a correspond-
ing subgraph sequence, 𝐻 be an 𝑛-vertex hypergraph, and 𝜑0𝑠 : 𝑉 (𝐺 0𝑠 ) → 𝑉 (𝐻)
be an injection for each 𝑠 ∈ [𝑠∗ ]. For 𝑠∗ − b𝜇𝑛𝑟−1 c < 𝑠 ≤ 𝑠∗ we define for each
𝑥 ∈ 𝑆(𝐺 𝑠 ) the weight

𝑤 𝑠 (𝑥) = {𝑦 ∈ N𝐺 𝑠 (𝑥) : 𝑦 is a leaf of 𝐺 𝑠 in 𝑉𝑠 } ,

and for each 𝑦 ∈ 𝑆(𝐻) the weight

𝑤 𝑠 (𝑦) = 𝑤 𝑠 𝜑0−1


𝑠 (𝑣) .

Further, for each 𝑦 ∈ 𝑆(𝐻) we define

Õ
𝑤(𝑣) = 𝑤 𝑠 (𝑦) .
𝑠∈[𝑠∗ ]

Since each set 𝑉𝑠 of omitted leaves is a strongly independent set in 𝐺 𝑠 , the weight
of any semi-edge containing an omitted leaf is 0. Thus the entire weight of 𝐺 𝑠 ,
which is ℓ, the number of omitted leaves, is supported on the semi-edges on the
vertices of 𝐺 0𝑠 . The packing of the 𝐺 0𝑠 is then guaranteed by the following lemma.

Lemma 4.8.5 (almost perfect packing lemma). Given 𝐷 ≥ 1, 𝑟 ≥ 2 and 𝜇 > 0,

provided we have
1 1
0 < 𝑐  𝜉  𝛾  𝜈, 𝜇, ,
𝐷 𝑟
the following holds. Let 𝐻 b be a (𝜉, (𝑟 + 1)𝐷 + 3)-quasirandom hypergraph with
𝑛 vertices. Let (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] be a 𝐷-degenerate hypergraph sequence with maximum

140
 4 Packing Degenerate Hypergraphs

𝑐𝑛 Í
degree log 𝑛 in which each graph contains at least 𝜇𝑛 leaves and 𝑠∈[𝑠∗ ] 𝑒(𝐺 𝑠 ) =
b and let (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] be a corresponding subgraph sequence omitting b𝜈𝑛c
𝑒( 𝐻),
leaves. Then there exists a packing (𝜑0𝑠 ) 𝑠∈[𝑠∗ ] of (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] into 𝐻
b with leftover 𝐻

−1
which is (𝛾 3 , 2)-quasirandom such that for 𝑝 = b𝜇𝑛𝑟−1 c b𝜈𝑛c 𝑛𝑟 and for all
0 ∗
𝑦 ∈ 𝑆(𝐻) and 𝑠, 𝑠 ∈ [𝑠 ] we have
(P 1) 𝑤(𝑦) = (1 ± 𝛾 3 ) 𝑝𝑛
𝑟 ,
(P 2) N𝐻 (𝑦) \ im 𝜑0𝑠 = (1 ± 𝛾 3 )𝜇𝑝𝑛,
(P 3) N𝐻 (𝑦) \ (im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) = (1 ± 𝛾 3 )𝜇2 𝑝𝑛 if 𝑠 ≠ 𝑠0,
for all 𝑢 ∈ 𝑉 (𝐻) and 𝑦 ∈ 𝑆(𝐻) with 𝑢 ∉ 𝑦 we have,
(P 4) 𝑠 𝑤 𝑠 (𝑦) 1𝑢∉im 𝜑𝑠0 = (1 ± 𝛾 3 )𝜇 𝑝𝑛
Í
𝑟 ,
and for all 𝑢 ∈ 𝑉 (𝐻) we have
10𝑟!𝑝 2 𝑛
(P 5) If 𝑢 ∉ im 𝜑0𝑠 then 𝑦 : 𝑦𝑢∈𝐸 (𝐻) 𝑤 𝑠 (𝑦) <
Í
𝜇 .

This lemma is deduced from Theorem 4.7.1 in Section 4.8.3. Given this packing
of the 𝐺 0𝑠 , what remains is to extend each 𝜑0𝑠 by packing the remaining leaves of 𝐺 𝑠 .
At each semi-edge 𝑦 ∈ 𝑆(𝐻), we have a collection of 𝑤(𝑦) leaves to pack coming
from the various hypergraphs 𝐺 𝑠 ; we say these are the leaves dangling at 𝑦. We
next choose an orientation 𝐻® of 𝐻 such that 𝑁 +® (𝑦) = 𝑤(𝑦) for each semi-edge 𝑦
𝐻
in 𝐻. The idea here is that we will then embed the leaves dangling at 𝑦 to the out-
neighbours of 𝑦. The following lemma, which we prove in Section 4.8.2, tells us
that a random orientation can be slightly modified to obtain the desired orientation:
this will allow us to show that 𝐻® inherits the quasirandomness properties of 𝐻
guaranteed by Lemma 4.8.5.

Lemma 4.8.6 (orientation lemma). Given 𝑝 > 0, 𝑟 ≥ 2 and 0 < 𝛾 < 2−2𝑟−3 𝑝 2𝑟 𝑟 −3𝑟 ,
let 𝐻 be an 𝑛-vertex 𝑟-uniform (𝛾 3 , 2)-quasirandom hypergraph of density 𝑝 with
semi-edge weights 𝑤 : 𝑉 (𝐻) → N0 such that 𝑤(𝑦) = (1 ± 𝛾 3 ) 𝑝𝑛
𝑟 for all 𝑦 ∈ 𝑆(𝐻)
Í ®
and such that we have 𝑦∈𝑉 𝑤(𝑦) = 𝑒(𝐻). If 𝐻0 is a uniform random orientation
of 𝐻, then with probability tending to 1 as 𝑛 → ∞ there is an orientation 𝐻® of 𝐻
such that for all 𝑦 ∈ 𝑆(𝐻) we have
(O 1) deg+® (𝑦) = 𝑤(𝑦), and
𝐻
® : 𝑦𝑢 ∈ 𝐸 (𝐻) and 𝑦𝑢 is oriented differently in 𝐻® and 𝐻®0 } ≤
(O 2) {𝑢 ∈ 𝑉 ( 𝐻)
𝛾 2 𝑛.

To help with packing the leaves dangling at each semi-edge 𝑦, we define the
following auxiliary bipartite graphs.

141
 4 Packing Degenerate Hypergraphs

® we define the leaves at

Definition 4.8.7 (leaf matching graphs). Given 𝑦 ∈ 𝑆( 𝐻),
𝑦 to be the set
n o
𝐿 𝑦 := 𝑥 : ∃𝑠 such that 𝑥 ∈ 𝑉 (𝐺 𝑠 ) \ 𝑉 (𝐺 0𝑠 ) and 𝑥𝜑0−1
𝑠 (𝑦) ∈ 𝐸 (𝐺 𝑠 )

Let the leaf matching graph 𝐹𝑦 be the bipartite graph with parts 𝐿 𝑦 and 𝑁 +® (𝑦),
𝐻
and edges 𝑥𝑢 with 𝑥 ∈ 𝐿 𝑦 and 𝑢 ∈ 𝑁 +® (𝑦) whenever 𝑢 ∉ im 𝜑0𝑠 for the 𝑠 such that
𝐻
𝑥 ∈ 𝑉 (𝐺 0𝑠 ).

Observe that a perfect matching in 𝐹𝑦 defines an assignment of the leaves at (all

preimages of) 𝑦 to 𝑁 +® (𝑦) which extends the packing of (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] . In what follows,
𝐻
we will be able to choose 𝐻® such that for each 𝑦, the graph 𝐹𝑦 has equal parts of
size roughly 1𝑟 𝑝𝑛 and every vertex has degree roughly 1𝑟 𝜇𝑝𝑛. We will further see
that each 𝐹𝑦 satisfies a codegree condition which by Lemma 4.2.6 implies that 𝐹𝑦
has a perfect matching.
Ð
If we simply chose a perfect matching in each 𝐹𝑦 to embed all the leaves 𝑦 𝐿 𝑦 ,
then we would almost have a perfect packing—each edge of 𝐻 b would be used
exactly once—but it could be the case that multiple leaves of some 𝐺 𝑠 (not in
the same 𝐿 𝑦 ) are embedded to a single 𝑢 ∈ 𝑉 ( 𝐻). b To avoid this, we find perfect
matchings in each 𝐹𝑦 one at a time and update the leaf matching graphs by removing
edges which are no longer useable. We choose these perfect matchings uniformly
at random at each step, and we will argue (using Lemma 4.2.6) that our updated
leaf matching graphs lose only very few (at most 𝜚 𝑝𝑛/𝑟) edges at any given vertex,
and thus continue to satisfy the conditions of Lemma 4.2.6.
Making this precise, assume 𝑉 ( 𝐻) ® = {1, . . . , 𝑛}, and set 𝐹𝑦(0) := 𝐹𝑦 for each
® We use the following algorithm MatchLeaves.
𝑦 ∈ 𝑆( 𝐻).
Assuming MatchLeaves succeeds, we then, for each 𝑠 ∈ [𝑠∗ ] and each 𝑥 ∈ 𝑉 (𝐺 𝑠 ),
set
 𝜑0𝑠 (𝑥) if 𝑥 ∈ dom(𝜑0𝑠 )



𝜑 𝑠 (𝑥) = (4.8.2)
 𝜎𝑦 (𝑥) if 𝑥 ∈ 𝐿 𝑦 .


This is a perfect packing of (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] into 𝐻
b by construction. Thus to prove
Theorem 4.8.1, what we need to do is prove Lemmas 4.8.5 and 4.8.6 above, and
argue that MatchLeaves succeeds with positive probability. We do the last of these
steps first, in the following subsection.

142
 4 Packing Degenerate Hypergraphs

Algorithm 4: MatchLeaves
Input : a hypergraph sequence (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] , a corresponding subgraph
sequence (𝐺 0𝑠 ) 𝑠∈[𝑠∗ ] omitting ℓ leaves, and associated leaf
matching graphs (𝐹𝑦(0) ) 𝑦∈𝑆( [𝑛])
Output : matchings (𝜎𝑦 ) 𝑦∈𝑆( [𝑛]) of the omitted leaves to feasible image
vertices as given by the leaf matching graphs
pick an arbitrary order 𝑆([𝑛]) = {𝑦 1 , . . . , 𝑦 ( 𝑛 ) } ;
𝑟 −1
𝑛

for 𝑖 = 1 to 𝑟−1 do
let 𝑦 = 𝑦𝑖 ;
let 𝜎𝑦 be a uniform
random perfect matching in 𝐹𝑦(𝑖−1) ;
𝑛
for 𝑗 = 𝑖 + 1 to 𝑟−1 do
let 𝐵 𝑦 𝑗 := 𝑥𝑢 ∈ 𝐸 (𝐹𝑦(𝑖−1) ) : ∃𝑠 such that 𝑥 ∈ 𝑉 (𝐺 0𝑠 ) and 𝜎𝑦−1 (𝑢) ∈

𝑗
𝑉 (𝐺 𝑠 ) ;
let 𝐹𝑦(𝑖)𝑗 := 𝐹𝑦(𝑖−1)
𝑗
− 𝐵𝑦 𝑗 ;
end
end
return (𝜎𝑦 ) 𝑦∈𝑆( [𝑛]) ;

4.8.1 The proof of Theorem 4.8.1

We are now in a position to prove (assuming Lemmas 4.8.5 and 4.8.6) The-
orem 4.8.1. As mentioned above, what this boils down to is, given the leftover 𝐻
from Lemma 4.8.5, showing that we can choose 𝐻® to inherit the required quasir-
andomness properties (which imply that the leaf matching graphs 𝐹𝑦 satisfy the
degree-codegree condition of Lemma 4.2.6) and arguing that during the running
of MatchLeaves, it is unlikely that more than 𝜚 𝑝𝑛/𝑟 edges are lost at any vertex of
any 𝐹𝑦 , which (by Lemma 4.2.6) implies that MatchLeaves is likely to complete
® we use Chernoff’s
and therefore the desired packing exists. For the choice of 𝐻,
inequality and union bounds together with Lemma 4.8.6. For the analysis of
MatchLeaves, we use Corollary 4.2.2 together with union bounds.

Proof of Theorem 4.8.1. Given 𝑟, 𝐷 and 𝜇 > 0 as in the statement of The-

orem 4.8.1, we choose constants as described at the beginning of this subsection.
We set 𝑄 = 𝐷 (𝑟 + 1) + 3.
Suppose that 𝑛 ≥ 𝑛0 . Let 𝐻 b be a (𝜉, 𝑄)-quasirandom hypergraph with 𝑛 vertices,
and let (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] be a sequence of 𝐷-degenerate hypergraphs, each on 𝑛 − b𝜇𝑛c
𝑐𝑛
vertices, with maximum degree at most log 𝑛 , and containing at least 𝜇𝑛 leaves.
Í
Suppose further that 𝑠∈[𝑠∗ ] 𝑒(𝐺 𝑠 ) = 𝑒( 𝐻).
b

143
 4 Packing Degenerate Hypergraphs

For each 𝑠 ∈ [𝑠∗ ], we choose a strongly independent set 𝑉𝑠 of b𝜈𝑛c leaves of 𝐺 𝑠 ,

and let 𝐺 0𝑠 be the corresponding hypergraph with vertices (𝑉 (𝐺 𝑠 ) \ 𝑉𝑠 ) ∪ 𝐼 𝑠 , where
𝐼 𝑠 is a set of |𝑉𝑠 | isolated vertices. We define weights 𝑤 𝑠 (𝑦) as in Definition 4.8.4.
Let for each 𝑠 the map 𝜑0𝑠 be the embedding of 𝐺 0𝑠 into 𝐻 b provided by
Lemma 4.8.5, with 𝐻 the (𝛾 3 , 2)-quasirandom hypergraph of leftover edges, and
suppose that the properties (P 1)–(P 5) hold, with the weight function 𝑤 as in


Definition 4.8.4. By construction, 𝐻 has b𝜇𝑛𝑟−1 c b𝜈𝑛c = 𝑝 𝑛𝑟 edges, so has density
𝑝.
The following claim establishes the existence of the desired orientation of 𝐻.
Claim 4.8.8. There exists an orientation 𝐻® of 𝐻 such that for each semi-edge
𝑦 ∈ 𝑆(𝐻) we have 𝑤(𝑦) = deg+® (𝑦), and in addition for each 𝑠, 𝑠0 ∈ [𝑠∗ ] we have
𝐻
(P’ 2) 𝑁 +® (𝑦) \ im 𝜑0𝑠 = (1 ± 𝛾) 𝜇𝑝𝑛
𝑟 , and
𝐻
2
(P’ 3) 𝑁 +® (𝑦) \ (im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) = (1 ± 𝛾) 𝜇 𝑟𝑝𝑛 if 𝑠 ≠ 𝑠0.
𝐻

Proof. 𝐻 is (𝛾 3 , 2)-quasirandom and of density 𝑝, and by (P 1) we have 𝑤(𝑦) =

(1 ± 𝛾 3 ) 𝑝𝑛
𝑟 for all 𝑦 ∈ 𝑆(𝐻). This, together with our choice of 𝛾, verifies that 𝐻
satisfies the conditions of the orientation lemma (Lemma 4.8.6).
Let 𝐻®0 be a uniform random orientation of 𝐻. Given 𝑦 ∈ 𝑆(𝐻) and 𝑠 ∈ [𝑠∗ ],
6 𝜇𝑝𝑛

by (P 2) and Theorem 4.2.1, with probability at least 1 − exp − 𝛾 12 we have

𝑁 +® (𝑦) \ im 𝜑0𝑠 = (1 ± 3𝛾 3 ) 𝜇𝑝𝑛

𝑟 .
𝐻0

Similarly, given 𝑦 ∈ 𝑆(𝐻) and 𝑠, 𝑠0 ∈ [𝑠∗ ] with 𝑠 ≠ 𝑠0, by (P 3) and Theorem 4.2.1,
6 2

with probability at least 1 − exp − 𝛾 𝜇12𝑝𝑛 we have
2
𝑁 +® (𝑦) \ (im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) = (1 ± 3𝛾 3 ) 𝜇 𝑟𝑝𝑛 .
𝐻0

Taking the union bound, and by Lemma 4.8.6, with probability at least 1 −
6 2
2𝑛3𝑟+6 exp − 𝛾 𝜇12𝑝𝑛 − 𝑜(1) each of the above good events holds for each 𝑦 ∈ 𝑆(𝐻)

and each 𝑠, 𝑠0 ∈ [𝑠∗ ] with 𝑠 ≠ 𝑠0, and in addition there is an orientation 𝐻® of 𝐻

satisfying conclusions (O 1) and (O 2) of Lemma 4.8.6.
6 2
For sufficiently large 𝑛 we have 1 − 2𝑛3𝑟+6 exp − 𝛾 𝜇12𝑝𝑛 − 𝑜(1) > 0, so we

fix 𝐻®0 and 𝐻® satisfying all these properties. By (O 1) the orientation 𝐻® satisfies
deg+® (𝑦) = 𝑤(𝑦) for each 𝑦 ∈ 𝑆(𝐻), as desired. Given 𝑦 ∈ 𝑆(𝐻) and 𝑠, 𝑠0 ∈ [𝑠∗ ]
𝐻

144
 4 Packing Degenerate Hypergraphs

with 𝑠 ≠ 𝑠0, by (O 2) we have

𝑁 +® (𝑦) \ im 𝜑0𝑠 = 𝑁 +® (𝑦) \ im 𝜑0𝑠 ± 𝛾 2 𝑛 = (1 ± 3𝛾 3 ) 𝜇𝑝𝑛 2 𝜇𝑝𝑛

𝑟 ± 𝛾 𝑛 = (1 ± 𝛾) 𝑟 ,
𝐻 𝐻0

where the final inequality is by choice of 𝛾. This verifies (P’ 2). Similarly, we have

𝑁 +® (𝑦) \ (im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) = 𝑁 +® (𝑦) \ (im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) ± 𝛾 2 𝑛

𝐻 𝐻0
2 2
= (1 ± 3𝛾 3 ) 𝜇 𝑟𝑝𝑛 ± 𝛾 2 𝑛 = (1 ± 𝛾) 𝜇 𝑟𝑝𝑛 ,

giving (P’ 3). 

From now on, we assume 𝐻® satisfies the properties of the above claim, which
we use instead of (P 2) and (P 3). What remains is to analyse MatchLeaves,
which means showing that the leaf matching graphs 𝐹𝑦(𝑡−1) satisfy the conditions
of Lemma 4.2.6 at every step 𝑡. We first show that the leaf matching graphs 𝐹𝑦(0)
satisfy the conditions (M 1) and (M 2) required to play the rôle of 𝐹 in Lemma 4.2.6,
with 𝑚 := 𝑝𝑛/𝑟. The main work is then to argue that sufficiently few edges are
deleted from 𝐹𝑦(0) to form 𝐹𝑦(𝑡−1) that it can play the rôle of 𝐹 0, i.e. that (M 3) is
satisfied. For this last part, we actually need only to show specifically that the
condition holds for 𝐹𝑦(𝑡−1)
𝑡
, where 𝑦 𝑡 is the 𝑡th semi-edge in our arbitrary order on
𝑆(𝐻).
Property (M 1): Given any 𝑦 ∈ 𝑆(𝐻) and any 𝑥 ∈ 𝑉 𝐹𝑦(0) , we separate two

cases. If 𝑥 ∈ 𝐿 𝑦 is in the hypergraph 𝐺 𝑠 , then by (P’ 2) we have deg𝐹 (0) (𝑥) =

𝑦
N+® (𝑦) \ im 𝜑0𝑠 = (1 ± 𝛾) 𝑟 . If 𝑥 ∈ 𝑁 +® (𝑦), then by (P 4) we have deg𝐹 (0) (𝑥) =
𝜇𝑝𝑛
𝐻 𝐻 𝑦
𝑤 𝑠 (𝑦) 1𝑥∉im 𝜑𝑠0 = (1 ± 𝛾 3 ) 𝜇𝑝𝑛
Í
𝑠 𝑟 . In either case, since 𝑝 > 𝛾 this verifies (M 1).
Property (M 2): Given any 𝑦 ∈ 𝑆(𝐻) and any 𝑢, 𝑢0 ∈ 𝐿 𝑦 , if 𝑢 ∈ 𝑉 (𝐺 𝑠 ) and
𝑢0 ∈ 𝑉 (𝐺 𝑠 0 ), where 𝑠 ≠ 𝑠0, then by (P’ 3) we have deg𝐹 (0) (𝑢, 𝑢0) = N+® (𝑦) \
𝑦 𝐻
2
(im 𝜑0𝑠 ∪ im 𝜑0𝑠 0 ) = (1 ± 𝛾) 𝜇 𝑟𝑝𝑛 .
Again since 𝛾 < 𝑝 this is as required by (M 2), and
we only need to show that the number of 𝑢, 𝑢0 ∈ 𝐿 𝑦 which are both in 𝐺 𝑠 for some
𝑝 2 𝑛2
𝑠 ∈ [𝑠∗ ] is at most 4 log( 𝑐𝑛
𝑝𝑛/2) . But any given 𝐺 𝑠 has at most 𝑤 𝑠 (𝑦) ≤ log 𝑛 vertices in
𝑐𝑛 0 0
𝐿 𝑦 , so that for a given 𝑢 there are at most log 𝑛 choices of 𝑢 with 𝑢, 𝑢 ∈ 𝑉 (𝐺 𝑠 ) for
2 2 2
some 𝑠 ∈ [𝑠∗ ]. Since |𝐿 𝑦 | ≤ 𝑛 we conclude that there are at most log𝑐𝑛 𝑝 𝑛
𝑛 < 4 log( 𝑝𝑛/2)
pairs 𝑢, 𝑢0 ∈ 𝐿 𝑦 such that 𝑢, 𝑢0 ∈ 𝑉 (𝐺 𝑠 ) for some 𝑠 ∈ [𝑠∗ ]. This completes the
verification of (M 2).
Property (M 3): This property does not hold deterministically, but we shall show
that it holds for all 𝑡 with high probability. For this purpose we define the following

145
 4 Packing Degenerate Hypergraphs

events. For each 𝑡 ∈ 𝑛

𝑟−1 , let E𝑡 be the event that for each 𝑥 ∈ 𝑉 (𝐹𝑦(0)
𝑡
) we have

1
deg𝐹 (0) (𝑥) − deg𝐹 (𝑡−1) (𝑥) ≤ 2𝑟 𝜚 𝑝𝑛 , (4.8.3)
𝑦𝑡 𝑦𝑡

that is, E𝑡 implies the event that (M 3) holds for 𝐹 = 𝐹𝑡(0) and 𝐹 0 = 𝐹𝑡(𝑡−1) . We shall
prove the following claim below, but first show how it implies the theorem.
Claim 4.8.9. With probability at least 1 − 𝑛−1 , for every 𝑡 ∈ 𝑟−1
 𝑛

the event E𝑡
holds.
If E𝑡 holds then all conditions of Lemma 4.2.6 are satisfied for 𝐹 = 𝐹𝑡(0) ,
𝐹 0 = 𝐹𝑡(𝑡−1) . In particular, the perfect matching in 𝐹𝑡(𝑡−1) required by MatchLeaves
exists, and so MatchLeaves does not fail at time 𝑡. Thus Claim 4.8.9 states that
with high probability, MatchLeaves does not fail at any time, and thus the maps 𝜑 𝑠
defined in (4.8.2) form (as explained there) the required perfect packing, proving
Theorem 4.8.1.
To prove Claim 4.8.9, let for each 𝑡 ∈ 𝑟−1 𝑛

the history H𝑡−1 consist of the
collection of matchings 𝜎1 , . . . , 𝜎𝑡−1 obtained by running MatchLeaves up to and
including step 𝑡 − 1. Note that giving H𝑡−1 determines whether E𝑡 holds or not.
, each 𝑢 ∈ 𝑁 +® (𝑦 𝑡 ) and each 𝑠 ∈ [𝑠∗ ] the
 𝑛

Claim 4.8.10. For each 𝑡 ∈ 𝑟−1
𝐻
following holds. Either E𝑡 does not occur or a random perfect matching 𝜎𝑡 in
𝐹𝑡(𝑡−1) satisfies
h i
2𝑟𝑤 𝑠 (𝑦 𝑡 )
P 𝜎𝑡−1 (𝑢) ∈ 𝑉 (𝐺 𝑠 )|H𝑡−1 ≤ 𝜇𝑝𝑛 .

Proof. If E𝑡 occurs, then all properties (M 1)–(M 3) from Lemma 4.2.6 are satisfied
with 𝐹 = 𝐹𝑡(0) and 𝐹 0 = 𝐹𝑡(𝑡−1) , and thus, a random matching 𝜎𝑡 in 𝐹 0 satisfies for
any given edge 𝑥𝑢 ∈ 𝐸 𝐹𝑡(𝑡−1)

P 𝑥𝑢 ∈ 𝜎𝑡 H𝑡−1 ≤ 2 2𝑟
 
𝜇𝑚 = 𝜇𝑝𝑛 .

Taking the union bound over the 𝑤 𝑠 (𝑦 𝑡 ) choices of 𝑥 ∈ 𝐿 𝑦 𝑡 in 𝐺 𝑠 , the claim follows.


We now verify Claim 4.8.9. We shall first argue that the claimed probability
bound follows from a probability bound, given in (4.8.4), which is of the right form
to use Corollary 4.2.2. Indeed, let A𝑡 be the event that E𝑖 holds for each 1 ≤ 𝑖 < 𝑡
but E𝑡 does not hold. Observe that if for each 𝑡 the event A𝑡 does not hold, then

146
 4 Packing Degenerate Hypergraphs

 𝑛
  𝑛

E𝑡 holds for each 𝑡 ∈ 𝑟−1 . In particular, by the union bound over 𝑡 ∈ 𝑟−1 it
suffices to show that for each fixed 𝑡 ∈ 𝑟−1 we have P[A𝑡 ] ≤ 𝑛−𝑟−1 . Further, by
 𝑛


another union bound over the at most 𝑣(𝐹𝑦(0)𝑡

) = 2𝑤(𝑦 𝑡 ) ≤ 2𝑛 different 𝑥 ∈ 𝑉 (𝐹𝑦(0)
𝑡
)
Ñ (0)
and since A𝑡 ⊆ 1≤𝑖≤𝑡−1 E𝑖 it is enough to show that for a fixed 𝑡 and 𝑥 ∈ 𝑉 (𝐹𝑦 𝑡 )
h Ù i
P E𝑖 and deg𝐹 (0) (𝑥) − deg𝐹 (𝑡−1) (𝑥) > 1
2𝑟 𝜚 𝑝𝑛 ≤ 𝑛−𝑟−3 . (4.8.4)
𝑦𝑡 𝑦𝑡
1≤𝑖≤𝑡−1

The remainder of this proof is devoted to establishing this bound. We will use
Ñ
Corollary 4.2.2 for this purpose, with the good event 1≤𝑖≤𝑡−1 E𝑖 . To that end,
define for each 1 ≤ 𝑖 ≤ 𝑡 − 1 the random variable

𝑌𝑖 := deg𝐹 (𝑖−1) (𝑥) − deg𝐹 (𝑖) (𝑥)

𝑦𝑡 𝑦𝑡

and observe that

𝑡−1
Õ
deg𝐹 (0) (𝑥) − deg𝐹 (𝑡−1) (𝑥) = 𝑌𝑖 .
𝑦𝑡 𝑦𝑡
𝑖=1

To apply Corollary 4.2.2 we need to find the range of each 𝑌𝑖 and the expectation
of each 𝑌𝑖 , conditioned on the history H𝑖−1 . This is encapsulated in Claim 4.8.11.
𝑐𝑛
Claim 4.8.11. For each 1 ≤ 𝑖 ≤ 𝑡 − 1, we have 0 ≤ 𝑌𝑖 ≤ log 𝑛 . Furthermore,
Í𝑡−1
either some E𝑖 with 1 ≤ 𝑖 ≤ 𝑡 − 1 does not occur, or we have 𝑖=1 E[𝑌𝑖 |H𝑖−1 ] ≤
40𝑝 2𝑟!𝑟 𝜇−2 𝑛.
𝑐𝑛
Proof. We first show 0 ≤ 𝑌𝑖 ≤ log 𝑛 . There are two cases to consider. First, if
𝑥 ∈ 𝐿 𝑦 𝑡 , then 𝑥 is in 𝐺 𝑠 for some 𝑠 ∈ An edge 𝑥𝑢 of 𝐹𝑦(𝑖−1)
[𝑠∗ ]. 𝑟
is removed to form
(𝑖) 𝑐𝑛
𝐹𝑦𝑟 only if 𝑢 is assigned a leaf of 𝐺 𝑠 in 𝜎𝑖 . Since there are at most 𝑤 𝑠 (𝑦𝑖 ) ≤ log 𝑛
such leaves, we have 𝑌𝑖 ≤ log 𝑐𝑛
in this case. Second, if 𝑥 ∈ 𝑁 + (𝑦 ), and 𝑥 is
𝑛 ® 𝑡 𝐻
assigned a leaf of 𝐺 𝑠 in 𝜎𝑖 , then we remove all edges of 𝐹𝑡(𝑖−1) from 𝑥 to leaves
of 𝐺 𝑠 to form 𝐹𝑡(𝑖) . Since 𝜎𝑖 is a matching, this happens for at most one 𝑠 ∈ [𝑠∗ ].
𝑐𝑛
There are at most 𝑤 𝑠 (𝑦 𝑡 ) ≤ log 𝑛 such leaves of 𝐺 𝑠 , so also in this case we have
𝑐𝑛
𝑌𝑖 ≤ log 𝑛.
We now bound above the sum of conditional expectations. Again, there are two
cases to consider. First, if 𝑥 ∈ 𝐿 𝑦 𝑡 , then let 𝑠 be such that 𝑥 ∈ 𝑉 (𝐺 𝑠 ). Suppose
that H𝑖−1 is a history up to and including 𝜎𝑖−1 such that E𝑖 holds. Recall that
𝜎𝑖−1 (𝑢) ∈ 𝑉 (𝐺 𝑠 ) means that some leaf of 𝐺 𝑠 is matched to 𝑢 in 𝜎𝑖 , and that we
® to mean that the edge 𝑢𝑦𝑖 , directed towards 𝑢, is in 𝐻.
write {𝑢𝑦𝑖 , 𝑢} ∈ 𝐸 ( 𝐻) ® By

147
 4 Packing Degenerate Hypergraphs

linearity of expectation, we have

Õ
E 𝑌𝑖 H𝑖−1 = P 𝜎𝑖−1 (𝑢) ∈ 𝑉 (𝐺 𝑠 ) H𝑖−1
   

𝑢∈N ®
(𝑖−1) (𝑥) : {𝑢𝑦 𝑖 ,𝑢}∈𝐸 ( 𝐻)
𝐹𝑦𝑡
Õ
≤ 2𝑟
𝑤 𝑠 (𝑦𝑖 ) 𝜇𝑝𝑛 ≤ 2𝑝 2 𝑛 · 𝑤 𝑠 (𝑦𝑖 ) · 2𝑟
𝜇𝑝𝑛 ,
𝑢∈N ®
(𝑖−1) (𝑥) : {𝑢𝑦 𝑖 ,𝑢}∈𝐸 ( 𝐻)
𝐹𝑦𝑡

where the first inequality is by Claim 4.8.10 and the second holds since if
® then in particular 𝑢 ∈ N𝐻 (𝑦𝑖 , 𝑦 𝑡 ), and since
𝑢 ∈ N𝐹 (𝑖−1) (𝑥) and {𝑢𝑦𝑖 , 𝑢} ∈ 𝐸 ( 𝐻)
𝑦𝑡
N𝐻 (𝑦𝑖 , 𝑦 𝑡 ) ≤ 2𝑝 2 𝑛 by (𝛾 3 , 2)-quasirandomness of 𝐻. Summing over 𝑖, either
some E𝑖 with 𝑖 ∈ [𝑡 − 1] does not hold, or we have

𝑡−1
Õ 𝑡−1 𝑡−1 (Õ
𝑟 −1)
𝑛
 Õ 4𝑟 𝑝
Õ
4𝑟 𝑝
E 𝑌𝑖 H𝑖−1 ≤ 2𝑝 2 𝑛 · 𝑤 𝑠 (𝑦𝑖 ) · 2𝑟

𝜇𝑝𝑛 = 𝜇 𝑤 𝑠 (𝑦𝑖 ) ≤ 𝜇 𝑤 𝑠 (𝑦𝑖 ) .
𝑖=1 𝑖=1 𝑖=1 𝑖=1

Í ( 𝑟 −1
𝑛
)
Since 𝑖=1 𝑤 𝑠 (𝑖) = b𝜈𝑛c counts the number of leaves removed from 𝐺 𝑠 to form
0
𝐺 𝑠 , we obtain that either some E𝑖 with 𝑖 ∈ [𝑡 − 1] does not hold, or

𝑡−1
40𝑝 2
Õ
E 𝑌𝑖 H𝑖−1 ≤
 
𝜇2
𝑛,
𝑖=1

as desired.
Finally, we consider the case 𝑥 ∈ 𝑁 +® (𝑦 𝑡 ). Suppose that a leaf of 𝐺 𝑠 is assigned
𝐻
to 𝑥 by 𝜎𝑖 , and that 𝑥 is adjacent to 𝑤 𝑠 (𝑦 𝑡 ) leaves of 𝐺 𝑠 in 𝐹𝑦(𝑖−1)
𝑡
. Then the edges to
these leaves are exactly the edges at 𝑥 removed from 𝐹𝑦 𝑡 to form 𝐹𝑦(𝑖)𝑡 . Suppose
(𝑖−1)

that H𝑖−1 is a history up to and including 𝜎𝑖−1 such that E𝑖 holds. Recall that a
leaf of 𝐺 𝑠 can only be assigned to 𝑥 by 𝜎𝑖 if {𝑥𝑦𝑖 , 𝑥} ∈ 𝐸 ( 𝐻) ® and 𝑥 ∉ im 𝜑0𝑠 . By
linearity of expectation, we have

1𝑥∈𝑁 + (𝑦𝑖 ) P 𝜎𝑖−1 (𝑥) ∈ 𝑉 (𝐺 𝑠 ) H𝑖−1 · 𝑤 𝑠 (𝑦 𝑡 )

Õ
E 𝑌𝑖 H𝑖−1 =
   
®
𝐻
𝑠∈[𝑠∗ ]

1𝑥∈𝑁 +® (𝑦𝑖 ) 1𝑥∉im 𝜑𝑠0 𝑤 𝑠 (𝑦 𝑡 )𝑤 𝑠 (𝑦𝑖 ) 𝜇𝑝𝑛

Õ
2𝑟
≤ ,
𝐻
𝑠∈[𝑠∗ ]

where the second line follows by Claim 4.8.10. Summing over 𝑖, either some E𝑖

148
 4 Packing Degenerate Hypergraphs

with 𝑖 ∈ [𝑟 − 1] does not hold, or we have

𝑡−1 𝑡−1 Õ
1𝑥∈𝑁 + (𝑦𝑖 ) 1𝑥∉im 𝜑𝑠0 𝑤 𝑠 (𝑦 𝑡 )𝑤 𝑠 (𝑦𝑖 ) 𝜇𝑝𝑛
Õ  Õ
E 𝑌𝑖 H𝑖−1 ≤ 2𝑟

®
𝐻
𝑖=1 𝑖=1 𝑠∈[𝑠∗ ]

Õ (Õ
𝑟 −1)
𝑛

≤ 1𝑥∈𝑁 +® (𝑦𝑖 ) 1𝑥∉im 𝜑𝑠0 𝑤 𝑠 (𝑦 𝑡 )𝑤 𝑠 (𝑦𝑖 ) 𝜇𝑝𝑛

2𝑟
𝐻
𝑠∈[𝑠∗ ] 𝑖=1

1𝑥∉im 𝜑𝑠0 𝑤 𝑠 (𝑦 𝑡 )𝑤 𝑠 (𝑦𝑖 ) 𝜇𝑝𝑛

Õ Õ
2𝑟
=
𝑠∈[𝑠∗ ] 𝑖:{𝑥𝑦 𝑖 ,𝑥}∈𝐸 ( 𝐻)
®

𝜇𝑝𝑛 1𝑥∉im 𝜑 𝑠
Õ Õ
2𝑟𝑤 𝑠 (𝑦 𝑡 )
≤ 0 𝑤 𝑠 (𝑦𝑖 )
𝑠∈[𝑠∗ ] 𝑖:𝑥𝑦 𝑖 ∈𝐸 (𝐻)
10𝑟!𝑝 2 𝑛
Õ Õ
2𝑟𝑤 𝑠 (𝑦 𝑡 ) 20𝑝𝑟!𝑟
≤ 𝜇𝑝𝑛 · 𝜇 = 𝜇2
𝑤 𝑠 (𝑦 𝑡 ) ,
𝑠∈[𝑠∗ ] 𝑠∈[𝑠∗ ]

where the last inequality is by (P 5). By definition of 𝑤(𝑦 𝑡 ), by (P 1) and by choice

of 𝛾 we have 𝑠∈[𝑠∗ ] 𝑤 𝑠 (𝑦 𝑡 ) = 𝑤(𝑦 𝑡 ) ≤ 2𝑟 𝑝𝑛, so we conclude that either some E𝑖
Í

with 𝑖 ∈ [𝑡 − 1] does not hold, or we have

𝑡−1
Õ
20𝑝𝑟!𝑟
E 𝑌𝑖 H𝑖−1 ≤ · 2𝑟 𝑝𝑛 ,
 
𝜇2
𝑖=1

as desired. 

Recall that 𝑝 ≤ 2𝜇𝜈/𝑟! by (4.8.1). Thus 40𝑝𝑟!𝑟 𝜇−2 ≤ 160𝜈𝜇−1 ≤ 4𝑟1 𝜚, where
the final inequality is by choice of 𝜈. It follows that 40𝑝 2𝑟!𝑟 𝜇−2 ≤ 4𝑟1 𝜚 𝑝. Thus
𝑐𝑛 1
using Claim 4.8.11 and Corollary 4.2.2, with 𝑅 = log 𝑛 , with 𝜇˜ = 4𝑟 𝜚 𝑝𝑛, and with
Ñ𝑟−1
the event E = 𝑖=1 E𝑖 , we get

h Ù 𝑟−1
Õ i
𝜇˜

P E𝑖 and 𝑌𝑖 > 1
2𝑟 𝜚 𝑝𝑛 ≤ exp − 4𝑅 ≤ 𝑛−𝑟−3 ,
1≤𝑖≤𝑟−1 𝑖=1

where the final inequality is by choice of 𝑐. This establishes (4.8.4), so completes

the proof.

4.8.2 Proof of the orientation lemma

In this subsection we prove Lemma 4.8.6. For this we need the following two
® to vertex 𝑣 ∈ 𝑦
definitions. The directed neighbourhood of a semi-edge 𝑦 ∈ 𝑆( 𝐻)

149
 4 Packing Degenerate Hypergraphs

in a directed hypergraph 𝐻® is

® ,

N𝑣® (𝑦) = 𝑢 : {𝑦𝑢, 𝑣} ∈ 𝐸 ( 𝐻)
𝐻

i.e. all vertices that form an edge with 𝑦, which is directed towards 𝑣. Also, if
𝑦 1 , . . . , 𝑦 𝑘 ∈ 𝑉 (𝐻) is a sequence of vertices, and 1 ≤ 𝑖 ≤ 𝑗 ≤ 𝑘 are integers, then
we write 𝑦𝑖.. 𝑗 for the set {𝑦𝑖 , . . . , 𝑦 𝑗 }.
The idea of the proof is the following. We first argue that a uniform random
orientation 𝐻®0 of 𝐻 is likely to have N𝐻®0 (𝑦) very close to 𝑤(𝑦) for every semi-edge
𝑦, and in addition it is likely to inherit the quasirandomness of 𝐻. Assuming both
these likely events hold, we then iteratively modify 𝐻®0 as follows. If there is no
𝑦 ∈ 𝑆(𝐻) whose out-degree is currently larger than 𝑤(𝑦), we are done. If there is
such a 𝑦, there exists also 𝑦0 ∈ 𝑆(𝐻) whose out-degree is smaller than 𝑤(𝑦0). We
identify (using the quasirandomness) a short chain of edges whose orientations we
can change in order that overall the out-degree of 𝑦 is decreased by one, that of 𝑦0 is
increased by one, and all other semi-edges remain unchanged. We repeat this until
we obtain the desired 𝐻. ® It remains to argue that we do not alter the orientation of
too many edges at any given semi-edge: to that end, we choose the chain of edges
at each step randomly and argue that it is unlikely any given semi-edge appears too
often in the chosen chains.

Proof of Lemma 4.8.6. By the given quasirandomness of 𝐻 we know that

deg𝐻 (𝑦) = (1 ± 𝛾 3 ) 𝑝𝑛 and | N𝐻 (𝑦) ∩ N𝐻 (𝑧)| = (1 ± 𝛾 3 ) 𝑝 2 𝑛 for every 𝑦 ≠ 𝑧 ∈ 𝑆(𝐻).
Using Theorem 4.2.1, we see that a.a.s. for every 𝑦 ≠ 𝑧 ∈ 𝑆(𝐻) and 𝑣 ∈ 𝑧 we have

𝑝𝑛 𝑝2𝑛
deg+® (𝑦) = (1 ± 2𝛾 3 ) and N+® (𝑦) ∩ N𝑣® (𝑧) = (1 ± 2𝛾 3 ) 2 .
𝐻0 𝑟 𝐻0 𝐻0 𝑟

From now on, we fix an arbitrary orientation 𝐻®0 satisfying these two properties,
and we will show that for any such 𝐻®0 there is an orientation 𝐻® of 𝐻 satisfying (O 1)
and (O 2).
Starting with 𝐻®0 , we successively switch the orientations of sets of 2𝑟 − 2 edges,
thus producing a sequence of oriented hypergraphs ( 𝐻®𝑖 )0≤𝑖≤𝑡 , with 𝐻®𝑡 being the
® For any such oriented hypergraph 𝐻®𝑖 and every semi-edge 𝑦 ∈ 𝑆(𝐻)
desired 𝐻.
we define the potential 𝜑𝑖 (𝑦) := deg+® (𝑦) − 𝑤(𝑦) and
𝐻𝑖
Õ
𝜑( 𝐻®𝑖 ) := |𝜑𝑖 (𝑦)| .
𝑦∈𝑆(𝐻)

150
 4 Packing Degenerate Hypergraphs

Observe that 𝜑( 𝐻®𝑖 ) is always nonnegative and even, and it is equal to zero exactly
when (O 1) is satisfied. To begin with, we have |𝜑0 (𝑦)| ≤ 3𝛾 3 𝑝𝑛 for every semi-
edge 𝑦, so that 𝜑( 𝐻®0 ) ≤ 3𝛾 3 𝑝𝑛 𝑟−1
𝑛

.
The algorithm PathSwitch, with input two semi-edges 𝑦 and 𝑧, finds a way of
changing the orientations of some 2𝑟 − 2 edges such that the out-degree of 𝑦 is
increased by 1 and that of 𝑧 is decreased by 1, with all other semi-edges keeping
their out-degree the same.

Algorithm 5: PathSwitch
Input : A directed hypergraph 𝐻® 0 and two disjoint semi-edges 𝑦, 𝑧 ∈ 𝑆( 𝐻® 0).
let 𝑦 1 , . . . , 𝑦𝑟−1 be a uniform random order of the vertices in 𝑦 ;
let 𝑧1 , . . . , 𝑧𝑟−1 be a uniform random order of the vertices in 𝑧 ;
let 𝐻®00 := 𝐻® 0 ;
for 𝑖 = 1 to 𝑟 − 1 do
let 𝑎 := 𝑦𝑖..𝑖+𝑟−2 ;
let 𝑏 := 𝑧𝑖..𝑖+𝑟−2 ;
let 𝑋 := N ®𝑖 0 (𝑎) ∩ N+® 0 (𝑏) \ 𝑦 1..𝑖+𝑟−2 ∪ 𝑧1..𝑖+𝑟−2 ;
𝑦

𝐻𝑖−1 𝐻𝑖−1
halt with failure if 𝑋 = ∅ ;
let 𝑣 be a uniform random vertex from 𝑋 ;
let 𝑦𝑖+𝑟−1 := 𝑣 ;
let 𝑧𝑖+𝑟−1 := 𝑣 ;
let 𝐻®𝑖0 be 𝐻®𝑖−1
0 − {𝑎𝑣, 𝑦 } − {𝑏𝑣, 𝑣} + {𝑎𝑣, 𝑣} + {𝑏𝑣, 𝑧 } ;
𝑖 𝑖
end
return 𝐻®𝑟−10 ;

Observe that PathSwitch either halts with failure, or performs as described: that
is, in the returned hypergraph the out-degree of 𝑦 is increased by one, that of 𝑧 is
decreased by one, and all other semi-edges have the same out-degree as before,
with the orientation of in total 2𝑟 − 2 edges changed. To see this, observe first
that the definition of 𝑋 ensures that all the edges whose orientations we choose
are distinct, so that we indeed change the orientation of 2𝑟 − 2 edges. Consider
now the set of vertices 𝑦 1..2𝑟−2 . In step 𝑖, we change the orientation of the edge
𝑦𝑖..𝑖+𝑟−1 from being directed towards 𝑦𝑖 , to being directed towards 𝑦𝑖+𝑟−1 . Thus
this change increases the out-degree of 𝑦𝑖..𝑖+𝑟−2 , and decreases that of 𝑦𝑖+1..𝑖+𝑟−1 ,
by one, and affects no other semi-edge. Overall, the result of these changes among
the vertices 𝑦 1..2𝑟−2 is to increase the out-degree of 𝑦 in 𝐻®𝑟−1
0 , and decrease that of

𝑦𝑟..2𝑟−2 , by one. Similarly, the changes among the vertices 𝑧1..2𝑟−2 have the effect
of decreasing the out-degree of 𝑧 in 𝐻®𝑟−1 0 , and increasing that of 𝑧
𝑟..2𝑟−2 , by one.

151
 4 Packing Degenerate Hypergraphs

But 𝑦𝑟..2𝑟−2 = 𝑧𝑟..2𝑟−2 by construction, so that this semi-edge overall has unchanged
out-degree in 𝐻®𝑟−1
0 as required.
Next, using the algorithm PathSwitch as a subroutine, the algorithm Orient-
ationSwitch describes how orientations are switched. In every iteration of this
algorithm, we change the potential of two semi-edges 𝑦, 𝑧 ∈ 𝑆(𝐻) with 𝜑𝑖 (𝑦) < 0
and 𝜑𝑖 (𝑧) > 0 by applying PathSwitch. In case that

{𝑢𝑠 ∈ 𝐸 (𝐻) : 𝑢𝑠 is oriented differently in 𝐻®𝑖 and 𝐻®0 }

gets too large in some round 𝑖 and for some semi-edge 𝑠, we let the algorithm
halt with failure. However, we will see in the following that this happens with
probability tending to 0.

Algorithm 6: OrientationSwitch
let 𝑡 := 𝜑( 𝐻®0 )/2 ;
for 𝑖 = 0 to 𝑡 − 1 do
if ∃ 𝑠 ∈ 𝑆( 𝐻®0 ) with
{𝑢𝑠 ∈ 𝐸 (𝐻) : 𝑢𝑠 is oriented differently in 𝐻®𝑖 and 𝐻®0 } > 𝛾 2 𝑛
then halt with failure;
choose semi-edges 𝑦, 𝑧 ∈ 𝑆( 𝐻®𝑖 ) with 𝜑𝑖 (𝑦) < 0 and 𝜑𝑖 (𝑧) > 0 ;
let 𝐻®𝑖+1 be the result of PathSwitch on 𝐻®𝑖 , 𝑦, 𝑧 ;
end
return 𝐻® := 𝐻®𝑡 ;

Observe that by our assumption on 𝛾, provided 𝑛 is sufficiently large PathSwitch

does not halt with failure when called by OrientationSwitch, because the relevant
directed (common) neighbourhoods can only be of size less than 4𝑟 if the failure
condition of OrientationSwitch is satisfied for some semi-edge. Furthermore, by
construction if OrientationSwitch does not halt with failure, then at each step 𝑖 we
have 𝜑( 𝐻®𝑖+1 ) = 𝜑( 𝐻®𝑖 ) − 2, so the returned 𝐻® has 𝜑( 𝐻)
® = 0 and thus satisfies (O 1).
Finally, by definition, not failing implies that 𝐻® also satisfies (O 2). Thus what
remains to prove is that OrientationSwitch succeeds with positive probability.
Note that when PathSwitch is run, it changes the orientation of an edge contained
in a semi-edge 𝑥 only if 𝑥 is contained in either 𝑦 1..2𝑟−2 or 𝑧1..2𝑟−2 (as constructed
by PathSwitch), and the number of edges containing 𝑥 whose orientation changes
is at most four (for the semi-edge 𝑦𝑟..2𝑟−2 = 𝑧𝑟..2𝑟−2 ). Thus in order to prove that
OrientationSwitch succeeds with positive probability, it is enough to upper bound

152
 4 Packing Degenerate Hypergraphs

the number of times any given 𝑥 ∈ 𝑆( 𝐻®0 ) is a subset of 𝑦 1..2𝑟−2 or 𝑧1..2𝑟−2 . This is
given by the following claim.
√
Claim 4.8.12. Given 𝑥 ∈ 𝑆( 𝐻®0 ), with probability at least 1 − 2𝑟 exp(− 𝑛), the
number of times 𝑥 is a subset of 𝑦 1..2𝑟−2 or 𝑧1..2𝑟−2 is at most 22𝑟+3 𝑝 −2𝑟 𝑟 3𝑟 𝛾 3 𝑛.

Proof. We fix 𝑥 ∈ 𝑆( 𝐻®0 ). To begin with, we upper bound the probability that 𝑥 is
a subset of 𝑦 1..2𝑟−2 for call 𝑖 to PathSwitch with semi-edges 𝑦 and 𝑧, working in the
graph 𝐻®𝑖−1 . This probability depends on |𝑥 ∩ 𝑦|, but not on 𝑧 or on the graph 𝐻®𝑖−1 .
If OrientationSwitch fails, the probability is zero, so we will from now assume that
OrientationSwitch has not failed.
When we run PathSwitch, we choose 𝑦𝑟 , . . . , 𝑦 2𝑟−2 in order, at each step (we
claim) choosing from a set of size at least 21 𝑝 2 𝑛/𝑟 2 . To see that the claimed set size
is valid, observe that the set 𝑋 of PathSwitch is obtained from a set of the form
N+® (𝑦0) ∩ N𝑣® (𝑧0) by removing vertices in edges at 𝑦0 or 𝑧0 whose orientation is
𝐻0 𝐻0
changed in 𝐻®𝑖−1 from that in 𝐻®0 , and a further at most 3𝑟 − 3 vertices previously
used in PathSwitch. Since OrientationSwitch has not failed, the number of vertices
removed due to orientation changes is at most 2𝛾 2 𝑛, so we obtain the claimed set
size by choice of 𝛾.
For each 𝑟 ≤ 𝑖 ≤ 2𝑟 − 2, when we choose 𝑦𝑖 the probability of choosing a vertex
2
of 𝑥 is by the above claim at most (𝑟 − 1) · 𝑝2𝑟2 𝑛 , and the probability that we end up
choosing all of 𝑥 is thus at most
  (𝑟−1)−|𝑥∩𝑦|   (𝑟−1)−|𝑥∩𝑦|
2𝑟 2 2𝑟 3
(𝑟 − 1) · 𝑝2 𝑛
≤ 𝑝2 𝑛
.

We now use this bound to estimate how often 𝑥 is a subset of 𝑦 1..2𝑟−2 over all
calls to PathSwitch by OrientationSwitch. For each 0 ≤ 𝑗 ≤ 𝑟 − 1, the number of

(𝑟−1)− 𝑗
choices of a semi-edge 𝑦 such that |𝑥 ∩ 𝑦| = 𝑗 is at most 𝑟−1
𝑗 ·𝑛 ≤ 2𝑟 𝑛𝑟−1− 𝑗 .
The number of times that any given 𝑦 is input to PathSwitch by OrientationSwitch
is at most |𝜑0 (𝑦)| ≤ 3𝛾 3 𝑝𝑛. Thus the total number of calls to PathSwitch with
some semi-edge 𝑦 which intersects 𝑥 in 𝑗 vertices is at most 3𝛾 3 𝑝𝑛 · 2𝑟 𝑛𝑟−1− 𝑗 =
3𝛾 3 2𝑟 𝑝𝑛𝑟− 𝑗 . We see that the number of times that 𝑥 is a subset of 𝑦 1..2𝑟−2 from
calls to PathSwitch with input 𝑦 where |𝑥 ∩ 𝑦| = 𝑗, is stochastically dominated
by a binomial random variable with 3𝛾 3 2𝑟 𝑝𝑛𝑟− 𝑗 tries and success probability
2𝑟 3 (𝑟−1)− 𝑗


2
𝑝 𝑛
. By the Chernoff bound, Theorem 4.2.1, the probability that the

153
 4 Packing Degenerate Hypergraphs

number of successes exceeds

  (𝑟−1)− 𝑗
2𝑟 3
3 𝑟
2 · 3𝛾 2 𝑝𝑛 𝑟− 𝑗
· 𝑝2 𝑛
= 3 · 22𝑟− 𝑗 𝛾 3 𝑝 −2𝑟+3+2 𝑗 𝑟 3𝑟−3−3 𝑗 𝑛 ≤ 22𝑟+2 𝑝 −2𝑟 𝑟 3𝑟−1 𝛾 3 𝑛 .

√
is at most exp(− 𝑛). Taking the union bound over choices of 𝑗, we see that the total
√
number of times that we have 𝑥 ⊆ 𝑦 1..2𝑟−2 is with probability at least 1−𝑟 exp(− 𝑛)
bounded above by 22𝑟+2 𝑝 −2𝑟 𝑟 3𝑟 𝛾 3 𝑛
We obtain the same estimate for the number of times that 𝑥 ⊆ 𝑧 1..2𝑟−2 by the
same argument replacing 𝑦 with 𝑧 throughout, and the claim follows. 

We can now complete the proof that OrientationSwitch succeeds with positive
probability. Indeed, if it fails there is some 𝑥 ∈ 𝑆( 𝐻®0 ) which witnesses the
failure. Since by choice of 𝛾 we have 22𝑟+3 𝑝 −2𝑟 𝑟 3𝑟 𝛾 3 𝑛 ≤ 𝛾 2 𝑛, by Claim 4.8.12
√
the probability that a given 𝑥 is such a witness is at most 2𝑟 exp(− 𝑛). Taking
𝑛
√
the union bound, the probability of a witness existing is at most 2𝑟 𝑟−1 exp(− 𝑛),
which tends to zero as 𝑛 → ∞, as desired.

4.8.3 Proof of the almost perfect packing lemma

In this section we prove Lemma 4.8.5, which is an easy corollary of Theorem 4.7.1.

Proof of Lemma 4.8.5. Given 𝐷, 𝑟, 𝜇, 𝜈, set 𝛾, we choose 𝛾 > 0 such that 𝛾 3 ≤

1 1 3
2 𝜇𝜈. We let 𝛿, 𝑐, 𝜉 > 0 be returned by Theorem 4.7.1 with input 𝑟, 2 𝛾 , 𝐷, 𝐾 = 1,
and suppose 𝑛 is sufficiently large for this theorem. Recall Theorem 4.7.1 provides
1 3
that 𝛿 ≤ 𝛾, and we without loss of generality also suppose 𝜉 ≤ 100 𝛾 .
Given 𝐻 0
b and (𝐺 𝑠 ) 𝑠∈[𝑠∗ ] as in Lemma 4.8.5, let 𝑝ˆ be the density of 𝐻, b i.e.
0
𝑛
𝑛
𝑛
Í
𝑟 edges. We have 𝑝 𝑟 = 𝑝ˆ 𝑟 − 𝑠∈[𝑠∗ ] 𝑒(𝐺 𝑠 ) = b𝜇𝑛
𝐻b has 𝑝ˆ 𝑟−1 c b𝜈𝑛c as in

Definition 4.8.2.
For each 𝑠 ∈ [𝑠∗ ] we create an 𝑛 − 𝛿𝑛-vertex hypergraph 𝐺 00𝑠 by adding 𝑛 − 𝛿𝑛 −
𝑣(𝐺 0𝑠 ) isolated vertices to 𝐺 0𝑠 , which we put at the end of the degeneracy order.
We apply Theorem 4.7.1 to pack the hypergraphs (𝐺 00𝑠 ) 𝑠∈[𝑠∗ ] into 𝐻, b with input
as above, with 𝑋 = 𝑋1 = · · · = 𝑋𝑟−1 = 𝑉 ( 𝐻), b and with 𝜔 𝑠 the (𝑟 − 1)-weight
function obtained from 𝑤 𝑠 of Lemma 4.8.5 for each 𝑠 ∈ [𝑠∗ ] as follows: we give
each ordered (𝑟 − 1)-set the weight 𝑤 𝑠 of the underlying unordered set. Note that
the required diet conditions reduce to (𝜉, 𝑄)-quasirandomness of 𝐻. b Suppose that
the likely outcome of Theorem 4.7.1 occurs, with 𝜑00𝑠 being the packing of 𝐺 00𝑠 for
each 𝑠 ∈ [𝑠∗ ]. Then the hypergraph 𝐻 of leftover edges is ( 21 𝛾 3 , 𝑄)-quasirandom.
In addition, setting 𝛿˜ = 𝜇, we have (QUASI 1)–(QUASI 3).

154
 4 Packing Degenerate Hypergraphs

Observe that for each 𝑠 ∈ [𝑠∗ ] we have sum(𝜔 𝑠 ) = (𝑟 − 1)!b𝜈𝑛c, because each
omitted leaf is counted exactly (𝑟 − 1)! times, once for each ordering of its parent
semi-edge. We set 𝜑0𝑠 to be the restriction of 𝜑00𝑠 to the first 𝑛 − 𝜇𝑛 vertices of 𝐺 00𝑠 ,
so that the 𝜑0𝑠 form a packing of the 𝐺 0𝑠 into 𝐻,
b also with leftover edges 𝐻. Note
that im 𝜑0𝑠 = 𝜑00𝑠 ([𝑛 − 𝜇𝑛]) for each 𝑠.
We now verify (P 1). Given any semi-edge 𝑦 ∈ 𝑆( 𝐻), b we let e be an arbitrary
ordering of 𝑦. By (QUASI 2) with 𝑈 = ∅, we have

𝑠∗
Õ 𝑠∗
Õ
1 3 −(𝑟−1)
𝜔 𝑠 (e) = (1 ± 2 𝛾 )𝑛 sum(𝜔 𝑠 )
𝑠=1 𝑠=1
1 3 −(𝑟−1)
= (1 ± 2 𝛾 )𝑛 b𝜇𝑛𝑟−1 c (𝑟 − 1)!b𝜈𝑛c
 
𝑛
= (1 ± 12 𝛾 3 )𝑛−(𝑟−1) (𝑟 − 1)!𝑝
𝑟
= (1 ± 𝛾 3 ) 𝑝𝑛
𝑟 ,

𝑟
where the final equality uses the approximation 𝑛𝑟 = (1 ± 𝑛−0.5 ) 𝑛𝑟! , valid for all

sufficiently large 𝑛. This verifies (P 1) since 𝑤 𝑠 (𝑦) = 𝜔 𝑠 (e).

For (P 2) and (P 3), which we need to verify for each semi-edge 𝑦 and each
𝑠 ≠ 𝑠0 ∈ [𝑠∗ ], we use (QUASI 1), with ℓ = 1 and 𝑇1 = 𝑦, with respectively 𝑆 = {𝑠}
and 𝑆 = {𝑠, 𝑠0 }. By (𝜉, 𝑄)-quasirandomness of 𝐻, b we have N b (𝑦) = (1 ± 𝜉) 𝑝𝑛,
𝐻 ˆ
and so

N𝐻 (𝑦) \ im 𝜑0𝑠 = (1 ± 12 𝛾 3 )𝜇 𝑝ˆ · (1 ± 𝜉) 𝑝𝑛
ˆ = (1 ± 𝛾 3 )𝜇𝑝𝑛 ,
𝑝

as required for (P 2). Essentially the same calculation gives (P 3) and we omit the
details.
For (P 4), with a given 𝑢 ∈ 𝑉 ( 𝐻)
b and semi-edge 𝑦 ∈ 𝑆( 𝐻)b that does not contain
𝑢, we apply (QUASI 2), this time with 𝑈 = {𝑢}, and with an arbitrary ordering e of
𝑦. We obtain

𝑠∗ 𝑠∗
𝜔 𝑠 (e) 1𝑈∩im 𝜑𝑠0 =∅ = (1 ±
Õ Õ
1 3 −(𝑟−1)
2 𝛾 )𝜇𝑛 sum(𝜔 𝑠 ) = (1 ± 𝛾 3 )𝜇 𝑝𝑛
𝑟
𝑠=1 𝑠=1

by the same calculation as for (P 1), which as with (P 1) gives (P 4).

155
 4 Packing Degenerate Hypergraphs

Finally we verify (P 5). Given 𝑣 ∈ 𝑉 ( 𝐻),

b note that
 
𝑛−1 1
N𝐻b (𝑣; 𝑋1 , . . . , 𝑋𝑟−1 ) = (𝑟 − 1)! N𝐻b (𝑣) = (1 ± 𝜉)(𝑟 − 1)! 𝑝𝑛
ˆ · ,
𝑟 −2 𝑟−1

where the final inequality is by (𝜉, 𝑄)-quasirandomness of 𝐻b and the observation

𝑟−1

that each semi-edge in N𝐻b (𝑣) contains 𝑟−2 = 𝑟 − 1 sets of size 𝑟 − 2 which make
a semi-edge with 𝑣. From (QUASI 3), we have that either 𝑣 ∈ im 𝜓 0𝑠 or
 
Õ
1 3 𝑝 𝑛−1
𝜔 𝑠 (e) = (1 ± 2 𝛾 ) 𝑝ˆ · (1 ± 𝜉)(𝑟 − 1)! 𝑝𝑛
ˆ 1
· 𝑛1−𝑟 sum(𝜔 𝑠 )
𝑟 −2 𝑟−1
e∈N 𝐻 (𝑣;𝑋1 ,...,𝑋𝑟 −1 )
𝑟 −1
= (1 ± 𝛾 3 ) 𝑝(𝑟 − 1)! (𝑟−1)!
𝑛
𝑛1−𝑟 (𝑟 − 1)!b𝜈𝑛c
𝑟!𝑝 2 𝑛
= (1 ± 𝛾 3 ) 𝑝(𝑟 − 1)! · (1 ± 34 ) 𝑝𝑟!𝑛
𝜇 ≤ 2(𝑟 − 1)! ,
𝜇

where the final line uses (4.8.1). Observe that each semi-edge contributing to (P 5)
contributes (𝑟 − 1)! times in the above sum (once for each ordering) and so this
gives (P 5).

156
Bibliography

[1] P. Allen, J. Böttcher, D. Clemens, J. Hladký, D. Piguet and A. Taraz,

The tree packing conjecture for trees of almost linear maximum degree,
arXiv:2106.11720.
[2] P. Allen, J. Böttcher, D. Clemens and A. Taraz, Perfectly packing graphs
with bounded degeneracy and many leaves, arXiv:1906.11558.
[3] P. Allen, J. Böttcher, J. Hladký and D. Piguet, Packing degenerate graphs,
Adv. Math. 354 (2019), 106739.
[4] N. Alon, J. Balogh, P. Keevash and B. Sudakov, The number of edge colorings
with no monochromatic cliques, J. London Math. Soc. (2) 70.2 (2004), 273–
288.
[5] J. Balogh and C. Palmer, On the Tree Packing Conjecture, SIAM J. Discrete
Math. 27.4 (2013), 1995–2006.
[6] J. Balogh, N. Bushaw, M. Collares, H. Liu, R. Morris and M. Sharifzadeh,
The typical structure of graphs with no large cliques, Combinatorica 37.4
(2017), 617–632.
[7] J. Balogh, R. Morris and W. Samotij, Independent sets in hypergraphs, J.
Amer. Math. Soc. 28.3 (2015), 669–709.
[8] B. Bollobás, Extremal graph theory, vol. 11, London Mathematical Society
Monographs, Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers],
London-New York, 1978, xx+488.
[9] J. A. Bondy, Pancyclic graphs, Proceedings of the Second Louisiana Con-
ference on Combinatorics, Graph Theory and Computing (Louisiana State
Univ., Baton Rouge, La., 1971), 1971, 167–172.
 Bibliography

[10] J. A. Bondy, Pancyclic graphs: recent results, Infinite and finite sets (Colloq.,
Keszthely, 1973; dedicated to P. Erdős on his 60th birthday), Vol. I, 1975,
Colloq. Math. Soc. János Bolyai, Vol. 10, 181–187.
[11] J. Böttcher, J. Hladký, D. Piguet and A. Taraz, An approximate version of
the tree packing conjecture, Israel J. Math. 211.1 (2016), 391–446.
[12] V. Chvátal and P. Erdős, A note on Hamiltonian circuits, Discrete Math. 2
(1972), 111–113.
[13] D. Clemens, S. Das and T. Tran, Colourings without monochromatic disjoint
pairs, European J. Combin. 70 (2018), 99–124.
[14] A. Dankovics, Low independence number and Hamiltonicity implies pan-
cyclicity, J. Graph Theory 95.2 (2020), 181–191.
[15] G. A. Dirac, Some theorems on abstract graphs, Proc. London Math. Soc.
(3) 2 (1952), 69–81.
[16] R. A. Duke, H. Lefmann and V. Rödl, A fast approximation algorithm for
computing the frequencies of subgraphs in a given graph, SIAM J. Comput.
24.3 (1995), 598–620.
[17] P. Erdős, Some new applications of probability methods to combinatorial
analysis and graph theory, Proceedings of the Fifth Southeastern Conference
on Combinatorics, Graph Theory and Computing (1974), 39–51. Congressus
Numerantium, No. X.
[18] P. Erdős, Some of my favourite problems in various branches of combinator-
ics, Matematiche (Catania) 47.2 (1992), Combinatorics 92 (Catania, 1992),
231–240 (1993).
[19] P. Erdős, Some problems in graph theory (1974), 187–190. Lecture Notes in
Math., Vol. 411.
[20] P. Erdős, Some recent progress on extremal problems in graph theory, Pro-
ceedings of the Sixth Southeastern Conference on Combinatorics, Graph
Theory and Computing (Florida Atlantic Univ., Boca Raton, Fla., 1975),
1975, 3–14. Congressus Numerantium, No. XIV.
[21] P. Erdős, Some theorems on graphs, Riveon Lematematika 9 (1955), 13–17.
[22] P. Erdős, D. J. Kleitman and B. L. Rothschild, Asymptotic enumeration of
𝐾𝑛 -free graphs, Colloquio Internazionale sulle Teorie Combinatorie (Rome,
1973), Tomo II, 1976, 19–27. Atti dei Convegni Lincei, No. 17.

158
 Bibliography

[23] P. Erdős, C. Ko and R. Rado, Intersection theorems for systems of finite sets,
Quart. J. Math. Oxford Ser. (2) 12 (1961), 313–320.
[24] D. A. Freedman, On tail probabilities for martingales, Ann. Probability 3
(1975), 100–118.
[25] Z. Füredi, A proof of the stability of extremal graphs, Simonovits’ stability
from Szemerédi’s regularity, J. Combin. Theory Ser. B 115 (2015), 66–71.
[26] S. Glock, D. Kühn, A. Lo and D. Osthus, The existence of designs via iterative
absorption: hypergraph 𝐹-designs for arbitrary 𝐹, arXiv:1611.06827.
[27] S. Glock, F. Joos, J. Kim, D. Kühn and D. Osthus, Resolution of the Ober-
wolfach problem, J. Eur. Math. Soc. (JEMS) 23.8 (2021), 2511–2547.
[28] A. Gyárfás and J. Lehel, Packing trees of different order into 𝐾𝑛 , Combin-
atorics (Proc. Fifth Hungarian Colloq., Keszthely, 1976), vol. 18, Colloq.
Math. Soc. János Bolyai, North-Holland, Amsterdam, 1978, 463–469.
[29] H. Hàn and A. Jiménez, Improved Bound on the Maximum Number of Clique-
Free Colorings with Two and Three Colors, SIAM J. Discrete Math. 32.2
(2018), 1364–1368.
[30] S. Janson, T. Łuczak and A. Ruciński, Random graphs, Wiley-Interscience,
2000.
[31] F. Joos, J. Kim, D. Kühn and D. Osthus, Optimal packings of bounded degree
trees, J. Eur. Math. Soc. (JEMS) 21.12 (2019), 3573–3647.
[32] P. Keevash, The existence of designs, arXiv:1401.3665.
[33] P. Keevash and K. Staden, Ringel’s tree packing conjecture in quasirandom
graphs, arXiv:2004.09947.
[34] P. Keevash and K. Staden, The generalised Oberwolfach problem,
arXiv:2004.09937.
[35] P. Keevash and B. Sudakov, Pancyclicity of Hamiltonian and highly connec-
ted graphs, J. Combin. Theory Ser. B 100.5 (2010), 456–467.
[36] P. Keevash, Hypergraph matchings and designs, Proceedings of the Interna-
tional Congress of Mathematicians—Rio de Janeiro 2018. Vol. IV. Invited
lectures, World Sci. Publ., Hackensack, NJ, 2018, 3113–3135.

159
 Bibliography

[37] J. Kim, D. Kühn, D. Osthus and M. Tyomkyn, A blow-up lemma for ap-
proximate decompositions, Trans. Amer. Math. Soc. 371.7 (2019), 4655–
4742.
[38] T. P. Kirkman, On a problem in combinations, Cambridge and Dublin Math.
J. 2 (1847), 191–204.
[39] M. Krivelevich, Triangle factors in random graphs, Combin. Probab. Com-
put. 6.3 (1997), 337–347.
[40] C. Lee and B. Sudakov, Hamiltonicity, independence number, and pancycli-
city, European J. Combin. 33.4 (2012), 449–457.
[41] H. Liu, O. Pikhurko and K. Staden, The exact minimum number of triangles
in graphs with given order and size, Forum Math. Pi 8 (2020), e8, 144.
[42] W. Mantel, Problem 28, Wiskundige Opgaven 10.60-61 (1907), 320.
[43] R. Montgomery, A. Pokrovskiy and B. Sudakov, A proof of Ringel’s conjec-
ture, Geom. Funct. Anal. 31.3 (2021), 663–720.
[44] F. Mousset, R. Nenadov and A. Steger, On the number of graphs without
large cliques, SIAM J. Discrete Math. 28.4 (2014), 1980–1986.
[45] O. Pikhurko and K. Staden, Stability for the Erdős-Rothschild problem,
arXiv:2105.09991.
[46] O. Pikhurko, K. Staden and Z. B. Yilma, The Erdős-Rothschild problem on
edge-colourings with forbidden monochromatic cliques, Math. Proc. Cam-
bridge Philos. Soc. 163.2 (2017), 341–356.
[47] O. Pikhurko and Z. B. Yilma, The maximum number of 𝐾3 -free and 𝐾4 -free
edge 4-colorings, J. Lond. Math. Soc. (2) 85.3 (2012), 593–615.
[48] J. Plücker, System der analytischen Geometrie, auf neue Betrachtungsweisen
gegründet, und insbesondere eine ausführliche Theorie der Kurven dritter
Ordnung enthaltend, Duncker und Humboldt, Berlin, 1835.
[49] G. Ringel, Problem 25, Theory of Graphs and its Applications (Proc. Int.
Symp. Smolenice 1963), Prague: Czech. Acad. Sci., 1963.
[50] V. Rödl, On a packing and covering problem, European J. Combin. 6.1
(1985), 69–78.

160
 Bibliography

[51] V. Rödl, A. Ruciński and E. Szemerédi, Perfect matchings in uniform hy-

pergraphs with large minimum degree, European J. Combin. 27.8 (2006),
1333–1349.
[52] I. Z. Ruzsa and E. Szemerédi, Triple systems with no six points carrying
three triangles, Combinatorics (Proc. Fifth Hungarian Colloq., Keszthely,
1976), Vol. II, vol. 18, Colloq. Math. Soc. János Bolyai, North-Holland,
Amsterdam-New York, 1978, 939–945.
[53] D. Saxton and A. Thomason, Hypergraph containers, Invent. Math. 201.3
(2015), 925–992.
[54] J. Steiner, Combinatorische Aufgaben, J. Reine Angew. Math. 45 (1853),
181–182.
[55] P. Turán, Eine Extremalaufgabe aus der Graphentheorie, Mat. Fiz. Lapok
48 (1941), 436–452.
[56] R. Yuster, The number of edge colorings with no monochromatic triangle, J.
Graph Theory 21.4 (1996), 441–452.

161