Large networks and graph limits

László Lovász
Institute of Mathematics, Eötvös Loránd University, Budapest,
Hungary
2010 Mathematics Subject Classification. Primary 05C99, Secondary 05C25,
05C35, 05C80, 05C82, 05C85, 90B15
Key words and phrases. graph homomorphism, graph algebra, graph limit,
graphon, graphing, property testing, regularity lemma
 To Kati
as all my books
 Contents

Preface xi

Part 1. Large graphs: an informal introduction 1

Chapter 1. Very large networks 3

1.1. Huge networks everywhere 3
1.2. What to ask about them? 4
1.3. How to obtain information about them? 5
1.4. How to model them? 8
1.5. How to approximate them? 11
1.6. How to run algorithms on them? 18
1.7. Bounded degree graphs 22

Chapter 2. Large graphs in mathematics and physics 25

2.1. Extremal graph theory 25
2.2. Statistical physics 32

Part 2. The algebra of graph homomorphisms 35

Chapter 3. Notation and terminology 37

3.1. Basic notation 37
3.2. Graph theory 38
3.3. Operations on graphs 39

Chapter 4. Graph parameters and connection matrices 41

4.1. Graph parameters and graph properties 41
4.2. Connection matrices 42
4.3. Finite connection rank 45

Chapter 5. Graph homomorphisms 55

5.1. Existence of homomorphisms 55
5.2. Homomorphism numbers 56
5.3. What hom functions can express 62
5.4. Homomorphism and isomorphism 68
5.5. Independence of homomorphism functions 72
5.6. Characterizing homomorphism numbers 75
5.7. The structure of the homomorphism set 79

Chapter 6. Graph algebras and homomorphism functions 83

6.1. Algebras of quantum graphs 83
6.2. Reflection positivity 88
vii
viii CONTENTS

6.3. Contractors and connectors 94

6.4. Algebras for homomorphism functions 101
6.5. Computing parameters with finite connection rank 106
6.6. The polynomial method 108

Part 3. Limits of dense graph sequences 113

Chapter 7. Kernels and graphons 115

7.1. Kernels, graphons and stepfunctions 115
7.2. Generalizing homomorphisms 116
7.3. Weak isomorphism I 121
7.4. Sums and products 122
7.5. Kernel operators 124

Chapter 8. The cut distance 127

8.1. The cut distance of graphs 127
8.2. Cut norm and cut distance of kernels 131
8.3. Weak and L1 -topologies 138

Chapter 9. Szemerédi partitions 141

9.1. Regularity Lemma for graphs 141
9.2. Regularity Lemma for kernels 144
9.3. Compactness of the graphon space 149
9.4. Fractional and integral overlays 151
9.5. Uniqueness of regularity partitions 154

Chapter 10. Sampling 157

10.1. W -random graphs 157
10.2. Sample concentration 158
10.3. Estimating the distance by sampling 160
10.4. The distance of a sample from the original 164
10.5. Counting Lemma 167
10.6. Inverse Counting Lemma 169
10.7. Weak isomorphism II 170

Chapter 11. Convergence of dense graph sequences 173

11.1. Sampling, homomorphism densities and cut distance 173
11.2. Random graphs as limit objects 174
11.3. The limit graphon 180
11.4. Proving convergence 185
11.5. Many disguises of graph limits 193
11.6. Convergence of spectra 194
11.7. Convergence in norm 196
11.8. First applications 197

Chapter 12. Convergence from the right 201

12.1. Homomorphisms to the right and multicuts 201
12.2. The overlay functional 205
12.3. Right-convergent graphon sequences 207
12.4. Right-convergent graph sequences 211
 CONTENTS ix

Chapter 13. On the structure of graphons 217

13.1. The general form of a graphon 217
13.2. Weak isomorphism III 220
13.3. Pure kernels 222
13.4. The topology of a graphon 225
13.5. Symmetries of graphons 234

Chapter 14. The space of graphons 239

14.1. Norms defined by graphs 239
14.2. Other norms on the kernel space 242
14.3. Closures of graph properties 247
14.4. Graphon varieties 250
14.5. Random graphons 256
14.6. Exponential random graph models 259

Chapter 15. Algorithms for large graphs and graphons 263

15.1. Parameter estimation 263
15.2. Distinguishing graph properties 266
15.3. Property testing 268
15.4. Computable structures 276

Chapter 16. Extremal theory of dense graphs 281

16.1. Nonnegativity of quantum graphs and reflection positivity 281
16.2. Variational calculus of graphons 283
16.3. Densities of complete graphs 285
16.4. The classical theory of extremal graphs 293
16.5. Local vs. global optima 294
16.6. Deciding inequalities between subgraph densities 299
16.7. Which graphs are extremal? 307

Chapter 17. Multigraphs and decorated graphs 317

17.1. Compact decorated graphs 318
17.2. Multigraphs with unbounded edge multiplicities 325

Part 4. Limits of bounded degree graphs 327

Chapter 18. Graphings 329

18.1. Borel graphs 329
18.2. Measure preserving graphs 332
18.3. Random rooted graphs 338
18.4. Subgraph densities in graphings 344
18.5. Local equivalence 346
18.6. Graphings and groups 349

Chapter 19. Convergence of bounded degree graphs 351

19.1. Local convergence and limit 351
19.2. Local-global convergence 360

Chapter 20. Right convergence of bounded degree graphs 367

20.1. Random homomorphisms to the right 367
20.2. Convergence from the right 375
x CONTENTS

Chapter 21. On the structure of graphings 383

21.1. Hyperfiniteness 383
21.2. Homogeneous decomposition 393

Chapter 22. Algorithms for bounded degree graphs 397

22.1. Estimable parameters 397
22.2. Testable properties 402
22.3. Computable structures 405

Part 5. Extensions: a brief survey 413

Chapter 23. Other combinatorial structures 415

23.1. Sparse (but not very sparse) graphs 415
23.2. Edge-coloring models 416
23.3. Hypergraphs 421
23.4. Categories 425
23.5. And more... 429
Appendix A. Appendix 433
A.1. Möbius functions 433
A.2. The Tutte polynomial 434
A.3. Some background in probability and measure theory 436
A.4. Moments and the moment problem 441
A.5. Ultraproduct and ultralimit 444
A.6. Vapnik–Chervonenkis dimension 445
A.7. Nonnegative polynomials 446
A.8. Categories 447

Bibliography 451
Author Index 465

Subject Index 469

Notation Index 473
 Preface

Within a couple of months in 2003, in the Theory Group of Microsoft Research

in Redmond, Washington, three questions were asked by three colleagues. Michael
Freedman, who was working on some very interesting ideas to design a quantum
computer based on methods of algebraic topology, wanted to know which graph
parameters (functions on finite graphs) can be represented as partition functions
of models from statistical physics. Jennifer Chayes, who was studying internet
models, asked whether there was a notion of “limit distribution” for sequences
of graphs (rather than for sequences of numbers). Vera T. Sós, a visitor from
Budapest interested in the phenomenon of quasirandomness and its connections to
the Regularity Lemma, suggested to generalize results about quasirandom graphs
to multitype quasirandom graphs. It turned out that these questions were very
closely related, and the ideas which we developed for the answers have motivated
much of my research for the next years.
Jennifer’s question recalled some old results of mine characterizing graphs
through homomorphism numbers, and another paper with Paul Erdős and Joel
Spencer in which we studied normalized versions of homomorphism numbers and
their limits. Using homomorphism numbers, Mike Freedman, Lex Schrijver and I
found the answer to Mike’s question in a few months. The method of solution, the
use of graph algebras, provided a tool to answer Vera’s. With Christian Borgs, Jen-
nifer Chayes, Lex Schrijver, Vera Sós, Balázs Szegedy, and Kati Vesztergombi, we
started to work out an algebraic theory of graph homomorphisms and an analytic
theory of convergence of graph sequences and their limits. This book will try to
give an account of where we stand.
Finding unexpected connections between the three questions above was stim-
ulating and interesting, but soon we discovered that these methods and results are
connected to many other studies in many branches of mathematics. A couple of
years earlier Itai Benjamini and Oded Schramm had defined convergence of graph
sequences with bounded degree, and constructed limit objects for them (our main
interest was, at least initially, the convergence theory of dense graphs). Similar
ideas were raised even earlier by David Aldous. The limit theories of dense and
bounded-degree graphs have lead to many analogous questions and results, and
each of them is better understood thanks to the other.
Statistical physics deals with very large graphs and their local and global prop-
erties, and it turned out to be extremely fruitful to have two statistical physicists
(Jennifer and Christian) on the (informal) team along with graph theorists. This
put the burden to understand the other person’s goals and approaches on all of us,
but at the end it was the key to many of the results.

xi
xii PREFACE

Another important connection that was soon discovered was the theory of prop-
erty testing in computer science, initiated by Goldreich, Goldwasser and Ron sev-
eral years earlier. This can be viewed as statistics done on graphs rather than on
numbers, and probability and statistics became a major tool for us.
One of the most important application areas of these results is extremal graph
theory. A fundamental tool in the extremal theory of dense graphs is Szemerédi’s
Regularity Lemma, and this lemma turned out to be crucial for us as well. Graph
limit theory, we hope, repaid some of this debt, by providing the shortest and
most general formulation of the Regularity Lemma (“compactness of the graphon
space”). Perhaps the most exciting consequence of the new theory is that it allows
the precise formulation of, and often the exact answer to, some very general ques-
tions concerning algorithms on large graphs and extremal graph theory. Indepen-
dently and about the same time as we did, Razborov developed the closely related
theory of flag algebras, which has lead to the solution of several long-standing open
problems in extremal graph theory.
Speaking about limits means, of course, analysis, and for some of us graph the-
orists, it meant hard work learning the necessary analytical tools (mostly measure
theory and functional analysis, but even a bit of differential equations). Involving
analysis has advantages even for some of the results that can be stated and proved
purely graph-theoretically: many definitions and proofs are shorter, more trans-
parent in the analytic language. Of course, combinatorial difficulties don’t just
disappear: sometimes they are replaced by analytic difficulties. Several of these
are of a technical nature: Are the sets we consider Lebesgue/Borel measurable? In
a definition involving an infimum, is it attained? Often this is not really relevant
for the development of the theory. Quite often, on the other hand, measurability
carries combinatorial meaning, which makes this relationship truly exciting.
There were some interesting connections with algebra too. Balázs Szegedy
solved a problem that arose as a dual to the characterization of homomorphism
functions, and through his proof he established, among others, a deep connection
with the representation theory of algebras. This connection was later further de-
veloped by Schrijver and others. Another one of these generalizations has lead to
a combinatorial theory of categories, which, apart from some sporadic results, has
not been studied before. The limit theory of bounded degree graphs also found very
strong connections to algebra: finitely generated infinite groups yield, through their
Cayley graphs, infinite bounded degree graphs, and representing these as limits of
finite graphs has been studied in group theory (under the name of sofic groups)
earlier.
These connections with very different parts of mathematics made it quite diffi-
cult to write this book in a readable form. One way out could have been to focus on
graph theory, not to talk about issues whose motivation comes from outside graph
theory, and sketch or omit proofs that rely on substantial mathematical tools from
other parts. I felt that such an approach would hide what I found the most exciting
feature of this theory, namely its rich connections with other parts of mathematics
(classical and non-classical). So I decided to explain as many of these connections
as I could fit in the book; the reader will probably skip several parts if he/she does
not like them or does not have the appropriate background, but perhaps the flavor
of these parts can be remembered.
 PREFACE xiii

The book has five main parts. First, an informal introduction to the math-
ematical challenges provided by large networks. We ask the “general questions”
mentioned above, and try to give an informal answer, using relatively elementary
mathematics, and motivating the need for those more advanced methods that are
developed in the rest of the book.
The second part contains an algebraic treatment of homomorphism functions
and other graph parameters. The two main algebraic constructions (connection
matrices and graph algebras) will play an important role later as well, but they
also shed some light on the seemingly completely heterogeneous set of “graph pa-
rameters”.
In the third part, which is the longest and perhaps most complete within its
own scope, the theory of convergent sequences of dense graphs is developed, and
applications to extremal graph theory and graph algorithms are given.
The fourth part contains an analogous theory of convergent sequences of graphs
with bounded degree. This theory is more difficult and less well developed than
the dense case, but it has even more important applications, not only because
most networks arising in real life applications have low density, but also because
of connections with the theory of finitely generated groups. Research on this topic
has been perhaps the most active during the last months of my work, so the topic
was a “moving target”, and it was here where I had the hardest time drawing the
line where to stop with understanding and explaining new results.
The fifth part deals with extensions. One could try to develop a limit theory
for almost any kind of finite structures. Making a somewhat arbitrary selection,
we only discuss extensions to edge-coloring models and categories, and say a few
words about hypergraphs, to much less depth than graphs are discussed in parts
III and IV.
I included an Appendix about several diverse topics that are standard mathe-
matics, but due to the broad nature of the connections of this material in mathe-
matics, few readers would be familiar with all of them.
One of the factors that contributed to the (perhaps too large) size of this book
was that I tried to work out many examples of graph parameters, graph sequences,
limit objects, etc. Some of these may be trivial for some of the readers, others may
be tough, depending on one’s background. Since this is the first monograph on the
subject, I felt that such examples would help the reader to digest this quite diverse
material.
In addition, I included quite a few exercises. It is a good trick to squeeze a
lot of material into a book through this, but (honestly) I did try to find exercises
about which I expected that, say, a graduate student of mathematics could solve
them with not too much effort.

Acknowledgements. I am very grateful to my coauthors of those papers that

form the basis of this book: Christian Borgs, Jennifer Chayes, Michael Freedman,
Lex Schrijver, Vera Sós, Balázs Szegedy, and Kati Vesztergombi, for sharing their
ideas, knowledge, and enthusiasm during our joint work, and for their advice and
extremely useful criticism in connection with this book. The creative atmosphere
and collaborative spirit at Microsoft Research made the successful start of this
research project possible. It was a pleasure to do the last finishing touches on
the book in Redmond again. The author acknowledges the support of ERC Grant
xiv PREFACE

No. 227701, OTKA grant No. CNK 77780 and the hospitality of the Institute for
Advanced Study in Princeton while writing most of this book.
My wife Kati Vesztergombi has not only contributed to the content, but has
provided invaluable professional, technical and personal help all the time.
Many other colleagues have very unselfishly offered their expertise and advice
during various phases of our research and while writing this book. I am particularly
grateful to Miklós Abért, Noga Alon, Endre Csóka, Gábor Elek, Guus Regts, Svante
Janson, Dávid Kunszenti-Kovács, Gábor Lippner, Russell Lyons, Jarik Nešetřil,
Yuval Peres, Oleg Pikhurko, the late Oded Schramm, Miki Simonovits, Vera Sós,
Kevin Walker, and Dominic Welsh. Without their interest, encouragement and
help, I would not have been able to finish my work.
 Part 1

Large graphs: an informal

introduction
 CHAPTER 1

Very large networks

1.1. Huge networks everywhere

In the last decade it became apparent that a large number of the most inter-
esting structures and phenomena of the world can be described by networks: often
the system consists of discrete, well separable elements, with connections (or inter-
actions) between certain pairs of them. To understand the behavior of the whole
system, one has to study the behavior of the individual elements as well as the
structure of the underlying network. Let us see some examples.
• Among very large networks, probably the best known and the most studied is
the internet. Moreover, the internet (as the physical underlying network) gives
rise to many other networks: the network of hyperlinks (web, logical internet),
internet based social networks, distributed data bases, etc. The size of the
internet is growing fast: currently the number of web pages may be 30 billion
(3 · 1010 ) or more, and the number of interconnected devices is probably more
than a billion. The graph theoretic structure of the internet determines, to
a large degree, how communication protocols should be designed, how likely
certain parts get jammed, how fast computer viruses spread etc.
• Social networks are basic objects of many studies in the area of sociology,
history, epidemiology and economics. They are not necessarily formally estab-
lished, like Facebook and other internet networks: The largest social network
is the acquaintance graph of all living people, with about 7 billion nodes. The
structure of this acquaintance graph determines, among others, how fast news,
inventions, religions, diseases spread over the world, now and during history.
• Biology contributes ecological networks, networks of interactions between pro-
teins, and the human brain, just to mention a few. The human brain, a network
of neurons, is really large for its mass, having about a hundred billion nodes.
One of the greatest challenges is, of course, to understand ourselves.
• Statistical physics studies the interactions between large numbers of discrete
particles, where the underlying structure is often described by a graph. For
example, a crystal can be thought of as a graph whose nodes are the atoms
and whose edges represent chemical bonds. A perfect crystal is a rather boring
graph, but impurities and imperfections create interesting graph-theoretical
digressions. 10 gram of a diamond has about 5 × 1023 nodes. The structure of
a crystal influences important macroscopic properties like whether the material
is magnetizable, or how it melts.
• Some of the largest networks in engineering occur in chip design. There can
be more than a billion transistors on a chip nowadays. Even though these
networks are man-made and carefully designed, many of their properties, like
3
4 1. VERY LARGE NETWORKS

the exact time they will need to perform some computation, are difficult to
determine from their design, due to their huge size.
• To be pretentious, we can say that the whole universe is a single (really huge,
possibly infinite) network, where the nodes are events (interactions between
elementary particles), and the edges are the particles themselves. This is a
network with perhaps 1080 nodes. It is an ongoing debate in physics how
much additional structure the universe has, but perhaps understanding the
graph-theoretical structure of this graph can help with understanding the global
structure of the universe.
These huge networks pose exciting challenges for the mathematician. Graph
Theory (the mathematical theory of networks) has been one of the fastest develop-
ing areas of mathematics in the last decades; with the appearance of the Internet,
however, it faces fairly novel, unconventional problems. In traditional graph theo-
retical problems the whole graph is exactly given, and we are looking for relation-
ships between its parameters or efficient algorithms for computing its parameters.
On the other hand, very large networks (like the Internet) are never completely
known, in most cases they are not even well defined. Data about them can be
collected only by indirect means like random local sampling or by monitoring the
behavior of various global processes.
Dense networks (in which a node is adjacent to a positive percent of other nodes)
and very sparse networks (in which a node has a bounded number of neighbors)
show a very different behavior. From a practical point of view, sparse networks are
more important, but at present we have more complete theoretical results for dense
networks. In this introduction, most of the discussion will focus on dense graphs;
we will survey the additional challenges posed by sparse networks in Section 1.7.

1.2. What to ask about them?

Think of a really large graph, say the internet, and try to answer the following
four simple questions about it.
Question 1. Does the graph have an odd or even number of nodes?
This is a very basic property of a graph in the classical setting. For example, it
is one of the first theorems or exercises in a graph theory course that every graph
with an odd number of nodes must have a node with even degree.
But for the internet, this question is clearly nonsense. Not only does the number
of nodes change all the time, with devices going online and offline, but even if we
fix a specific time like 12:00am today, it is not well-defined: there will be computers
just in the process of booting up, breaking down etc.
Question 2. What is the average degree of nodes?
This, on the other hand, is a meaningful question. Of course, the average degree
can only be determined with a certain error, and it will change as technology or
the social composition of users change; but at a given time, a good approximation
can be sought (I am not speaking now about how to find it).
Question 3. Is the graph connected?
To this question, the answer is almost certainly no: somewhere in the world
there will be a faulty router with some unhappy users on the wrong side of it. But
this is not the interesting way to interpret the question: we should consider the
 1.3. HOW TO OBTAIN INFORMATION ABOUT THEM? 5

internet “disconnected” if, say, an earthquake combined with a sunflare severs all
connections between the Old and New worlds. So we want to ignore small com-
ponents that are negligible in comparison with the whole graph, and consider the
graph “disconnected” only if it decomposes into two parts which are commeasurable
with the whole. On the other hand, we may want to allow that the two large parts
be connected by a very few edges, and still consider the graph “disconnected”.
Question 4. Where is the largest cut in the graph?
(This means to find the partition of the nodes into two classes so as to maximize
the number of edges connecting the two classes.) This example shows that even if
the question is meaningful, it is not clear in what form can we expect the answer.
We can ask for the fraction of edges contained in the largest cut (depending on the
model, this can be determined relatively easily, with an error that is small with
high probability, although it is not easy to prove that the algorithm works). But
suppose we want to “compute” the largest cut itself; how to return the result, i.e.,
how to specify the largest cut (or even an approximate version of it)? We cannot
just list all nodes and tell on which side do they belong: this would be too much
time and memory space. Is there a better way to answer the question?

1.3. How to obtain information about them?

If we face a large network (think of the internet) the first challenge is to obtain
information about it. Often, we don’t even know the number of nodes.

1.3.1. Sampling. Properties of very large graphs can be studied by randomly

sampling small subgraphs. The theory of this, sort of a statistics where we work with
graphs instead of numbers, is called property testing in computer science. Initiated
by Goldreich, Goldwasser and Ron [1998], this theory emerged in the last 15-20
years, and will be one of the main areas of applications of the methods developed
in this book.
In the case of dense graphs G, it is simple to describe a reasonably realistic
sampling process: we select independently a fixed number k of random nodes, and
determine the edges between them, to get a random induced subgraph (Figure
1.1). We have to assume, of course, that we have methods to select a uniformly
distributed random node of the graph, and to determine whether two nodes are
adjacent. We’ll call this subgraph sampling. For each graph F on [k] = {1, 2, . . . , k},
there is a certain probability of seeing F when k nodes are sampled, which we
denote by σG,k (F ). So every graph G defines a probability distribution σG,k on all
graphs with k nodes. It turns out that this sample contains enough information to
determine many properties and parameters of the graph, with some error of course.
This error can be made arbitrarily small with high probability if we choose the
sample size k sufficiently large, depending on the error bound (and only on the
error bound, not on the graph!).
One may try to strengthen this and allow this sampling process to be repeated
a bounded number of times. This would not give anything new, however: sampling
k nodes r times gives less information than sampling kr nodes once. For clarity, it
is sometimes better to describe algorithms saying that we repeat a certain sampling
process, but this could always be replaced by taking a single sample (larger, but
still of bounded size).
6 1. VERY LARGE NETWORKS

Figure 1.1. Sampling from a dense graph and from a graph with
bounded degree.

1.3.2. Global observables. Instead of taking a random subset of the nodes

(sampling) and studying the subgraph induced by them, we can take a random
partition of the nodes into a small number of classes, and study the “quotient”, the
small graph obtained by merging the classes of the partition. (This will have very
large edge multiplicities, which we have to normalize appropriately.) These quo-
tients carry information about global measurements (like the number of stable sets,
the maximum cut, various quantities in statistical physics, etc.). The remarkable
fact is that under the right conditions, these “global” observables carry the same
information as “local” sampling (see Sections 12.3 and 20.2.)
Another source of information about a very large network is the observation of
the behavior of various processes on the graph, through a longer time interval. The
observation can be global (measurement of some global parameter), or local (at one
node, or a few neighboring nodes). Observing heat propagation through a material
is an example of the first kind of approach; web crawlers can be considered as
examples of the second, and in a sense so is our observation of the universe. There
are some sporadic results about the local observation of simple random processes
(Benjamini and Lovász [2002], Benjamini, Kozma, Lovász, Romik, and Tardos
[2006], but a general theory of such local observation of global processes has not
emerged yet.

1.3.3. Left and right homomorphisms. In theoretical studies, it is often

more convenient to talk about homomorphisms (adjacency-preserving maps) be-
tween graphs, instead of looking at randomly chosen induced subgraphs. For two
finite simple graphs F and G, let hom(F, G) denote the number of homomorphisms
of F into G (adjacency-preserving maps from V (F ) to V (G)). We often normalize
these homomorphism numbers, to get homomorphism densities:
hom(F, G)
(1.1) t(F, G) = .
v(G)v(F )
This number is the probability that a random map of V (F ) into V (G) preserves
adjacency. (We denote by V (F ) and E(F ) the sets of nodes and edges of the graph
F , respectively, and their cardinalities by v(F ) = |V (F )| and e(F ) = |E(F )|.)
Homomorphisms will be basic tools throughout the book. We introduce them
in Chapter 5 (where we survey some of our knowledge about them), but use them
all the time thereafter. There will be different versions like homomorphisms into
weighted graphs, which play an important role in statistical physics (we will return
to them at the end of the Introduction).
 1.3. HOW TO OBTAIN INFORMATION ABOUT THEM? 7

Homomorphism densities can be expressed in terms of the distribution of sam-

ples, and vice versa (at least asymptotically, as the size of G tends to infinity). For
example, let us consider the homomorphism density of the quadrilateral C4 in a
large graph G. If we map the four nodes of C4 into G, if may be that the images are
different, and so the image is a quadrilateral. It could also happen that the images
of two nodes coincide. This cannot happen to adjacent nodes, because the image
of an edge must be an edge; but it may happen to two opposite nodes. In this case,
the image is a “V” (path with two edges). Or it can happen that both pairs of
opposite nodes have the same image, and the image is a single edge. If we know the
numbers of edges, V’s and quadrilaterals in G, then we can compute the number
of homomorphisms of the quadrilateral into G. (Warning: the same quadrilateral
in G can be the image in 8 different ways, the same V can be the image in 4 ways,
and the same edge, in 2 ways!). The numbers of quadrilaterals, V’s and edges can
be estimated by sampling. In fact, the last two will not matter much for very large
graphs G, since a random map of 4 elements into v(G) elements will be one-to-one
with high probability.
So homomorphism densities and sampling distributions carry the same informa-
tion, why bother to introduce both? Using homomorphisms has several advantages
(and some disadvantages).

• Homomorphism numbers are better behaved algebraically, and they have been
used before to study various algebraic questions concerning direct product of
graphs, like cancellation laws (see Section 5.4.2). Furthermore, a lot is known
about other issues concerning homomorphisms: existence, structure, etc.

• When looking at a (large) graph G, we may try to study its local structure by
counting homomorphisms from various “small” graphs F into G; we can also
study its global structure by counting homomorphisms from G into various
small graphs H. The first type of information is closely related (in many cases,
equivalent) to sampling, while the second is related to global observables. This
way homomorphisms are pointing at a certain duality between sampling and
global observation. We can sum up our framework for studying large graphs in
the following formula:

F −→ G −→ H.
We will informally talk about “left-homomorphisms” and “right-
homomorphisms” to refer to these two kind of mappings.
• We will characterize which distributions come from sampling k nodes from a
(large) graph G, and we will characterize homomorphism densities as well. It
turns out that a characterization of sample distributions is simpler and more
natural, but putting it in another way, the characterization of homomorphism
densities is more surprising, and therefore has more interesting applications.

• Using homomorphisms leads us to looking at things through the spectacles of

category theory, and this point of view is very fruitful. For example, sometimes
one can simply “turn arrows around”, and get new results almost for free. We
will say more about this generalization to categories near the end of the book,
in Section 23.4.
8 1. VERY LARGE NETWORKS

1.4. How to model them?

1.4.1. Random graphs. We celebrated the 50th birthday of random graphs
recently: The simplest random graph model was developed by Erdős and Rényi
[1959] and Gilbert [1959]. Given a positive integer n and a real number 0 ≤ p ≤ 1,
we generate a random graph G(n, p) by taking n nodes, say [n] = {1, . . . , n}, and
connecting any two of them with probability p, making an independent decision
about each pair.
There are alternate models, which are essentially equivalent from the point of
view of many properties. Two of these were introduced in the early papers by
Erdős–Rényi [1959, 1960]: We could fix the number of edges m, and then choose a
random m-element subset of the set of pairs in [n], uniformly from all such subsets.( )
This random graph G(n, m) has very similar properties to G(n, p) when m = p n2 .
Another model, closer to some of the more recent developments, is evolving random
graphs, where edges are added one by one, always choosing uniformly from the set
of unconnected pairs. Stopping this process after m steps, we get G(n, m).
Random graphs have many interesting, often surprising properties, and a huge
literature, see Bollobás [2001], Janson, Luczak and Ruczinski [2000], or Alon and
Spencer [2000].
One conventional wisdom about random graphs with a given edge density is
that they are all alike. Their basic parameters, like chromatic number, maximum
clique, triangle density, spectra etc. are highly concentrated. This fact will be an
important motivation when defining the right measure of global similarity of two
graphs in Chapter 8.
Many generalizations of this random graph model have been studied. For ex-
ample, we can consider a “template” for the random graph in the form of a weighted
graph H on q nodes, with a weight αi > 0 associated with each node, and a weight
0 ≤ βij ≤ 1 associated with each edge ij. We assume that the nodeweights sum
to 1. We may also assume that H is complete with a loop at every node, since
the missing edges can be added with weight 0. A multitype random graph G(n; H)
with template H is generated as follows: We take [n] = {1, . . . , n} as its node set,
where we think of n as a much larger number than q. We partition [n] into q sets
V1 , . . . , Vq , by putting node u in Vi with probability αi , and connecting each pair
u ∈ Vi and v ∈ Vj with probability βij (all these decisions are made independently).
While multitype random graphs are too close to the original Erdős–Rényi model
to be useful as, say, internet models, they play an extremely important role by
serving as simple objects approximating arbitrarily large graphs (see Section 1.5.2
and Chapter 9).
More generally, one could have different probabilities assigned to different edges
(this was suggested by Erdős and Rényi in their second paper [1960] already). The
construction we will use a lot, namely constructing a random graph from a symmet-
ric measurable function [0, 1]2 → [0, 1] is a related idea, discovered independently
several times, first time, probably, by Diaconis and Freedman [1981]. These random
graphs, which we call W -random, will be discussed in Section 10.1 and will play an
important role throughout the book.

1.4.2. Quasirandom graphs. Deterministic objects that look and behave

like randomly generated ones are important in various branches of science. For
 1.4. HOW TO MODEL THEM? 9

example, pseudorandom number generators are basic algorithms in computer sci-

ence, with many applications in Monte-Carlo algorithms, computer security and
elsewhere. Exact definitions are usually difficult to give, and they vary according
to need. It is very remarkable that in graph theory it is possible to give a very ro-
bust definition of quasirandom graphs, where many related or even seemingly quite
different formalizations of properties of random graphs capture the same notion.
We know that random graphs have a variety of quite strict properties (with high
probability); it turns out that for several of these basic properties, the exceptional
graphs are the same. In other words, any of these properties implies the others,
regardless of any stochastic consideration.
A measure of quasirandomness of a graph was introduced by Thomason [1987];
the theory of quasirandom graph sequences, which has been an important example
for convergent graph sequences central to this book, was developed by Chung,
Graham and Wilson [1989].
To make this idea precise, we consider a sequence of graphs (Gn ) with v(Gn ) →
∞. For simplicity of notation, assume that v(Gn ) = n. Let 0 < p < 1 be a real
number. Consider the following properties of the sequence of graphs:
(QR1) Almost all degrees are asymptotically pn and almost all codegrees (numbers
of common neighbors of two nodes) are asymptotically p2 n.
(QR2) For every fixed graph F , the number of homomorphisms of F into Gn is
asymptotically pe(F ) nv(F ) .
(QR3) The number of edges is asymptotically pn2 /2 and the number of 4-cycles
is asymptotically p4 n4 /8. (We have to divide by 2 and 8, because we are counting
unlabeled copies rather than homomorphisms.)
(QR4) The number of edges induced by any set of n/2 nodes is asymptotically
pn2 /8.
(QR5) For any two disjoint sets X, Y of nodes, the number of edges between X
and Y is p|X||Y | + o(n2 ).
All these properties hold with probability 1 if Gn = G(n, p). However, more is
true: if a graph sequence satisfies either one of them, then it satisfies all. Perhaps
the most surprising fact along these lines is the equivalence of the second and third:
prescribing the right asymptotic number of copies in Gn for just two small graphs
(the edge and the 4-cycle) forces every other simple graph to have (asymptotically)
the right number of copies.
Such graph sequences are called quasirandom. The five properties above are
only a sampler; there are many other properties of random graphs that are also
equivalent to these (Chung, Graham and Wilson [1989], Simonovits and Sós [1991,
1997, 2003]).
Many interesting deterministic graph sequences are quasirandom. We mention
an important example from number theory:
Example 1.1 (Paley graphs). Let q be any prime congruent 1 modulo 4, and
let us define a graph on {0, . . . , q − 1} by connecting i and j if and only if i − j is
a quadratic residue modulo q. We construct this graph for every such prime, and
order them in a sequence.
This graph sequence is quasirandom with density 1/2. (To verify the first prop-
erty above is perhaps the easiest; see Exercise 1.2). This example also illustrates
how some of the equivalent conditions above may be much easier to verify than
10 1. VERY LARGE NETWORKS

others: in this case, the third would be as easy as the verification of the first, but
the second and fourth would be quite difficult: How would you count the number
of copies of, say, the Petersen graph in a Paley graph? How would you count the
number of those differences that are quadratic residues between, say, square-free
integers in {0, . . . , q − 1}? When posed directly, these questions sound formidable;
but the equivalence of the above conditions provides answers to them. 
We should emphasize that in this setting, quasirandomness is a property of a
sequence of graphs, not of a single graph. Of course, one could introduce a measure
of deviation from the “ideal” quasirandomness in each of the conditions (QR1)–
(QR5), and prove explicit relationships between them. Since our interest is the
limit theory, we will not go in this direction.
Sometimes we need to consider quasirandom bipartite graphs, which can be
defined, mutatis mutandis, by any of the properties above. More generally, just
as there are multitype random graphs, there are also multitype quasirandom graph
sequences. Similarly as for random graphs, a multitype quasirandom graph sequence
(Gn ) is defined by a “template” weighted graph H on q nodes, with a nodeweights
αi > 0 and edgeweights βij . The sequence is multitype quasirandom with template
H, if the node set V (Gn ) can be partitioned into q sets V1 , . . . , Vq such that |Vi | ∼
αi v(Gn ), the subgraphs Gn [Vi ] induced by Vi form a quasirandom sequence for
every i ∈ [q], and the bipartite subgraphs Gn [Vi , Vj ] between Vi and Vj form a
quasirandom bipartite graph sequence for each pair i ̸= j ∈ [q].
The same remark applies as for multitype random graphs: they play an ex-
tremely important role by serving as simple objects approximating arbitrarily large
graphs. The equivalence of conditions (Q1)–(Q5) can be generalized appropriately
(with a larger, but finite set of graphs in (Q3) instead of just 2), as it will be
discussed in Section 16.7.1.
The main topic of the book, the theory of convergent graph sequences, can
be considered as a further, rather far-reaching generalization of quasirandom se-
quences.
1.4.3. Randomly growing graphs. Random graph models on a fixed set of
nodes, discussed above, fail to reproduce important properties of real-life networks.
For example, the degrees of Erdős–Rényi random graphs follow a binomial distribu-
tion, and so they are asymptotically normal if the edge probability p is a constant,
and asymptotically Poisson if the expected degree is constant (i.e., p = p(n) ∼ c/n).
In either case, the degrees are highly concentrated around the mean, while the de-
grees of real life networks tend to obey the “Zipf phenomenon”, which means that
the tail of the distribution decreases according to a power law (unlike the most
familiar distributions like Gaussian, geometric or Poisson, whose tail probability
drops exponentially; Figure 1.2).
In 1999 Albert and Barabási [1999, 2002, 2002] created a new random network
model. Perhaps the main new feature compared with the Erdős–Rényi graph evo-
lution model is that not only edges, but also nodes are added by natural rules of
growing. When a new node is added, it connects itself to a given number d of old
nodes, where each neighbor is selected randomly, with probability proportional to
its degree. (This random selection is called preferential attachment.) The Albert–
Barabási graphs reproduce the “heavy tail” behavior of the degree sequences of
real-life graphs. Since then a great variety of growing networks were introduced,
reproducing this and other empirical properties of real-life networks.
 1.5. HOW TO APPROXIMATE THEM? 11

Figure 1.2. Degree distributions of an Erdős–Rényi random

graph on 100 nodes with edge density .1 (left) and of a real life
graph with similar parameters (right). The main feature to ob-
serve about the latter is not that the largest frequency is 1, but
that it is much more stretched out.

This is perhaps the first point which suggests one of our main tools, namely
assigning limits to sequences of graphs. Just as the Law of Large Numbers tells us
that adding up more and more independent random variables we get an increasingly
deterministically behaving number, these growing graph sequences tend to have
a well-defined structure, for almost all of the possible random choices along the
way. In the limit, the randomness disappears, and the asymptotic behavior of the
sequence can be described by a well-defined limit object. We will return to this in
this Introduction in Sections 1.5.3 and 11.3.
Exercise 1.2. Prove that the sequence of Paley graphs is quasirandom.

1.5. How to approximate them?

If we want to experiment with a large network (say, try out a new protocol for
the internet), then it is good to have a ”scaled down” version of it. In other words,
we want a compact approximate description of a very large network, from which
a network similar to the original, but of suitable size, can be generated. To make
this mathematically precise, we need to define what we mean by two graphs to be
“similar” or “close”, and describe what kind of structures we use for approximation.

1.5.1. The distance of two graphs. There are many ways of defining the
distance of two graphs G and G′ . Suppose that the two graphs have a common
node set [n]. Then a natural notion of distance is the edit distance, defined as the
number of edges to be changed to get from one graph to the other. This could also
be viewed as the Hamming distance |E(G)△E(G′ )| of the edge sets (△ denotes
symmetric difference). Since our graphs are very large, we want to normalize this.
If the graphs are dense, then a natural normalization is
|E(G)△E(G′ )|
d1 (G, G′ ) = .
n2
While this distance plays an important role in the study of testable graph properties,
it does not reflect structural similarity well. To raise one objection, consider two
12 1. VERY LARGE NETWORKS

random graphs on [n] with edge density 1/2. As mentioned in the introduction,
these graphs are very similar from almost every aspect, but their normalized edit
distance is large (about 1/2 with high probability). One might try to decrease this
by relabeling one of them to get the best overlay minimizing the edit distance; but
the improvement would be marginal (tending to 0 if n tends to infinity).
Another trouble with the notion of edit distance is that it is defined only when
the two graphs have the same set of nodes. We want to define a notion of distance
for two graphs that are so large that we don’t even know the number of their nodes,
and these numbers might be very different. For example, we want to find that two
large random graphs are “close” even if they have a different number of nodes.
One useful way to overcome these difficulties is to base the measurement of
distance on sampling. Recall that for a graph G, σG,k is the probability distribution
on graphs on [k] = {1, 2, . . . , k} obtained by selecting a random ordered k-subset
of nodes and taking the subgraph induced by them. Strictly speaking, this is only
defined when k ≤ v(G); but we are interested in taking a small sample from a large
graph, not the other way around. To make the definition precise, let us say that
the sampling returns the edgeless k-node graph if k > v(G). (In this case it would
be a better solution to sample with repetition, but sampling without repetition is
better in other cases, so let us stick to it.)
Now if we have two graphs G and G′ , we can compare the distributions of k-
node samples for any fixed k. We use the variation distance between distributions
α and β on the same set, defined by
dvar (α, β) = sup |α(X) − β(X)|,
X

where the supremum is taken over all measurable subsets (observable events). If
we want to measure the distance of two graphs by a single number, we use a simple
trick known from analysis: We define the sampling distance of two dense graphs G
and G′ by
∞
∑
′ 1
(1.2) δsamp (G, G ) = dvar (σG,k , σG′ ,k )
2k
k=1
k
(Here the coefficients 1/2 are quite arbitrary, they are there only to make the sum
convergent; but the above is a convenient choice.) This distance notion is very
suitable for our general goals, since two graphs are close in this distance if and
only if random sampling of “small” induced subgraphs does not distinguish them
reliably. However, sampling distance has one drawback: it does not directly reflect
any structural similarity.
In Chapter 8 we will define a notion of distance, called cut distance, between
graphs, which will be satisfactory from all these points of view: it will be defined
for two graphs with possibly different number of nodes, the distance of two random
graphs with the same edge density will be very small, and it will reflect global
structural similarity. The definition involves too many technical details to be given
here, unfortunately. But it will turn out (and this is one of the main results in this
book) that the cut distance is equivalent to the sampling distance in a topological
sense.

1.5.2. Approximation by smaller: Regularity Lemmas. Let us return

to the question of “scaling down” a huge graph, first in the dense case. The main
 1.5. HOW TO APPROXIMATE THEM? 13

tool for doing so is the “Szemerédi Partition” or “Regularity Lemma”. Szemerédi

developed the first version of the Regularity Lemma for his celebrated proof of
the Erdős–Turán Conjecture on arithmetic progressions in dense sets of integers in
1975. Since then, the Lemma has emerged as a fundamental tool in graph theory,
with many applications in extremal graph theory, combinatorial number theory,
graph property testing etc., and became a true focus of research in the past years.
Informally, the Regularity Lemma says that every graph can be approximated
by a multitype quasirandom graph, where the number of classes depends on the
error of the approximation only. This lemma can be viewed as an archetypal exam-
ple of dichotomy between randomness and structure, where we try to decompose a
(large and complicated) object A into a more highly structured object A′ with a
(quasi)random perturbation (cf. Tao [2006c]). The highly structured part may be
easier to handle because of the structure, and the quasirandom part will often be
easier to handle due to Laws of Large Numbers.
Pixel pictures. In this introductory part, we want to illustrate the idea of a
regularity partition visually. To this end, let us introduce a non-standard way of
visualizing graphs. On the left of Figure 1.3 we see a graph (the Petersen graph).
In the middle, we see its adjacency matrix. On the right, we see another version of
its adjacency matrix, where the 0’s are replaced by white squares and the 1’s are
replaced by black squares. We think of the whole picture as the unit square, so the
little squares have side length 1/n, where n is the number of nodes. The origin is
in the upper left corner, following the convention of indexing matrix elements.

0 1 0 0 1 1 0 0 0 0
1 0 1 0 0 0 1 0 0 0
0 1 0 1 0 0 0 1 0 0
0 0 1 0 1 0 0 0 1 0
1 0 0 1 0 0 0 0 0 1
1 0 0 0 0 0 0 1 1 0
0 1 0 0 0 0 0 0 1 1
0 0 1 0 0 1 0 0 0 1
0 0 0 1 0 1 1 0 0 0
0 0 0 0 1 0 1 1 0 0

Figure 1.3. The Petersen graph, its adjacency matrix, and its
pixel picture

It is not clear that this pixel picture reveals more about small graphs than
the usual way of drawing them (probably less), but it can be suggestive for large
graphs. Figure 1.4 shows the usual drawing and the pixel picture of a half-graph,
a bipartite graph defined on the set {1, . . . , n, 1′ , . . . , n′ }, where the edges are the
pairs (i, j) with i ≤ j ′ . For large n, the pixel picture of a half-graph may be more
informative, as we will see in the next section.
The left square in Figure 1.5 is the pixel picture of a (reasonably large) random
graph. We don’t see much structure—and we shouldn’t. From a distance, this
picture is more-or-less uniformly grey, similar to the second square. The 100 × 100
chessboard in the third picture is also uniformly grey, or at least it would become so
if we increased the number of pixels sufficiently. One might think that it represents
14 1. VERY LARGE NETWORKS

Figure 1.4. A half-graph and its pixel picture

a graph that is close to the random graph. But rearranging the rows and columns
so that odd indexed columns come first, we get the 2 × 2 chessboard on the right!
So wee see that both the middle and the right side pictures represent a complete
bipartite graph. The pixel picture of a graph depends on the ordering of the nodes.
We can be reassured, however, that a random graph remains random, no matter
how we order the nodes, and so the picture on the left remains uniformly grey, no
matter how the nodes are ordered.

Figure 1.5. A random graph with 100 nodes and edge density
1/2, a random graph with very many nodes and edge density 1/2,
a chessboard, and the pixel picture obtained by rearranging the
rows and columns.

Remark 1.3. Using pixel pictures to represent graphs, in particular random

graphs, goes in a sense in the opposite direction to what was studied in the psy-
chology of vision. Of course, processing images given by pixel pictures has been
a fundamental issue in connection with computer graphics and related areas, and
we are not going into this issue in this book. But we should mention the work of
Julesz, who studied the question of how well the human eye can distinguish random
noise (like Figure 1.5(a)) from images that are also uniformly grey but more struc-
tured (textured). The chessboard in Figure 1.5(b) would be a trivial example of
such an image. Disproving some of his conjectures, Diaconis and Freedman [1981]
constructed pixel pictures that are very closely related to our W -random graphs.
The Regularity Lemma. We illustrate the Regularity Lemma by Figure 1.6.
The graph on the left side (given by its pixel picture) looks quite random. In the
middle we see the same graph, with its nodes ordered differently. In this picture, we
see some structure of the graph (even though it is not as clear-cut as in Figure 1.5);
what we see is that the upper left corner is denser, and the lower right corner is
sparser. If we cut the picture into four equal parts, and average the “blackness” in
each, we get the picture on the right. Inside each of the four parts, the arrangement
 1.5. HOW TO APPROXIMATE THEM? 15

is quite random-like, and further rearrangement would not reveal any additional
structure.
Still informally, the Regularity Lemma says the following:
The nodes of every graph can be partitioned into a “small” number of “almost
equal” parts in such a way that for “almost all” pairs of partition classes, the
bipartite graph between them is “quasirandom”.

Figure 1.6. A random-looking pixel picture, an informative re-

arrangement, and its regularity partition

Some of the expressions in quotation marks are easy to explain. For the whole
theorem, we have an error bound 0 < ε < 1 specified in advance. The condition
that the parts are “almost equal” means that their sizes differ by at most one:
if the graph has n nodes partitioned into k classes, then the size of each class is
either ⌊n/k⌋ or ⌈n/k⌉. The condition that the number of classes is “small” means
that it can be bounded by an explicit function f (ε) of ε; to exclude trivialities,
we( )also assume that k ≥ 1/ε. “Almost all” pairs of classes means that we allow
ε k2 exceptional pairs about which we don’t claim anything (we can include the
subgraphs induced by the classes among these exceptions). Finally, we need to
define what it means to be “random-like”: one way to put it is that this bipartite
graph is quasirandom with some density pij (which may be different for different
pairs of classes) and with error ε, in the sense introduced (informally) in Section
1.4.2.
Regularity partitions and quasirandomness have a lot to do with each other.
Not only is quasirandomness part of the statement of the Regularity Lemma, but
the regularity lemma can be used to characterize quasirandomness: Simonovits and
Sós [1991] proved that a graph sequence is quasirandom with density p if and only
if the graphs have regularity partitions for arbitrarily small ε > 0 such that the
densities pij between the partition classes tend to p.
I have to come back to the “small” number of partition classes. The proof gives
2...
a tower 22 of height 1/ε5 , which is a very large number, and which unfortunately
cannot be improved too much, since Gowers [1997] constructed graphs for which
the smallest number of classes in a Szemerédi partition was at least a tower of
height log(1/ε). So the tower behavior is a sad fact of life. There are related
partitions with a more decent number of classes, as we shall see in Chapter 9,
where regularity partitions will be defined formally. We will also discuss situations
when the regularity partitions have a very decent size, like 1/εconst (Sections 13.4
and 16.7). Implicitly or explicitly, regularity partitions will be used throughout this
book.
16 1. VERY LARGE NETWORKS

1.5.3. Approximation by infinite: convergence and limits. This idea

can be motivated by how we look at a large piece of metal. This is a crystal, that is
a really large graph consisting of atoms and bonds between them. But from many
points of view (e.g., the use of the metal in building a bridge), it is more useful
to consider it as a continuum with a few important parameters (density, elasticity
etc.). Its behavior is governed by differential equations relating these parameters.
Can we consider a more general very large graph as some kind of a continuum?
Our way to make this intuition precise is to consider a growing sequence (Gn )
of graphs whose number of nodes tends to infinity, to define when such a sequence
is convergent, and to assign a limit object to convergent graph sequences, which
somehow incorporates all the properties we want to be remembered. (We have
mentioned this idea in connection with randomly growing graphs, but now we don’t
assume anything about how the graphs in the sequence are obtained.) This plan is
the backbone of this book: we will carry it out both for dense graphs and also for
graphs with bounded degree. There will be a good collection of applications of this
work.
Our discussion of sampling from a graph suggests a general principle leading to
a definition: we consider samples of a fixed size k from Gn , and their distribution.
We say that the sequence is locally convergent (with respect to the given sampling
method) if this distribution tends to a limit as n → ∞ for every fixed k.
For dense graphs, this notion of convergence was defined by Borgs, Chayes,
Lovász, Sós, and Vesztergombi [2006, 2008]; some elements of this definition go
back to Erdős, Lovász and Spencer [1979]. This notion has many useful properties.
Perhaps most important of these is that it can be characterized in terms of the cut
distance of graphs. It is not hard to see that the above notion of convergence is
equivalent to saying that the graph sequence is a Cauchy sequence in the sampling
distance. One of the main results presented in this book is Theorem 11.3, which
can be stated informally as follows:
The same graph sequences are convergent (Cauchy sequences) for both the cut
distance and the sampling distance.
If we have a notion of convergence, the question arises naturally: what does it
converge to? Can we describe a limit object for every convergent graph sequence?
The family of limiting sample distributions (one for each k) can be considered
as a limit object of the sequence (we call this the “weak limit”). This is not
always a helpful representation of the limit object, and a more explicit description
is desirable.
A next step is to represent the family of distributions on finite graphs (the
samples) by a single probability distribution on countable graphs: we get certain
notion of random graphs on the countable set N∗ = {1, 2, 3, . . . } (see Theorem
11.52).
More explicit descriptions of these limit objects can also be given, in the form
of a two-variable measurable function W : [0, 1]2 → [0, 1], called a graphon (Lovász
and Szegedy [2006]; see Section 7). These limit objects can be considered as
weighted graphs with an underlying set of continuum cardinality. (If you wish, you
can also think of these graphons as unweighted graphs on a non-standard model
of the unit interval, where W (x, y) is the density of edges between an infinitesimal
neighborhood of x and an infinitesimal neighborhood of y; this approach will be
explained in Section 11.3.2). Random graphs with edge density 1/2 converge to the
 1.5. HOW TO APPROXIMATE THEM? 17

identically 1/2 function (have a look at the two squares on the left of Figure 1.5).
Figure 1.7 illustrates that the sequence of half-graphs (discussed in Section 1.5.2)
converges to a limit (the function W (x, y) = 1(y ≥ x + 1/2 or x ≥ y + 1/2). It
has been observed and used before (see e.g. Sidorenko [1991]) that such functions
can be used as generalizations of graphs, and this gives certain arguments a greater
analytic flexibility.

Figure 1.7. A half-graph, its pixel picture, and the limit function

Let us describe another example here (more to follow in Section 11.4.2). The
picture on the left side of Figure 1.8 is the adjacency matrix of a graph G with 100
nodes, where the 1’s are represented by black squares and the 0’s, by white squares.
The graph itself is constructed by a simple randomized growing rule: Starting with
a single node, we create a new node, and connect every pair of nonadjacent nodes
with probability 1/n, where n is the current number of nodes. (This construction
will be discussed in detail in Section 11.4.2.)

Figure 1.8. A randomly grown uniform attachment graph with

100 nodes, and a (continuous) function approximating it

The picture on the right side is a grayscale image of the function U (x, y) =
1 − max(x, y). (Recall that the origin is in the upper left corner!) The similarity
with the picture on the left is apparent, and suggests that the limit of the graph
sequence on the left is this function. This turns out to be the case in a well defined
sense. It follows that to approximately compute various parameters of the graph
on the left side, we can compute related parameters of the function on the right
side. For example, the triangle density of the graph on the left tends (as n → ∞)
to the integral
∫
(1.3) U (x, y)U (y, z)U (z, x) dx dy dz
[0,1]3
18 1. VERY LARGE NETWORKS

(the evaluation of this integral is a boring but easy task). It is easy to see how to
generalize this formula to express the limiting density of any fixed graph F .
We hope that the examples above provide motivation for the following fact,
which is one of the key results to be discussed in the book (Theorem 11.21):
The limit of any convergent graph sequence can be represented by a graphon,
in the sense that the limiting density of any fixed simple graph F is given by an
integral of the type (1.3).
Of course, a graphon can be infinitely complicated; but in many cases, limits
of growing graph sequences have a limit graphon that is a continuous function
described by a simple formula (see some further examples in Section 11.4.2). Such
a limit graphon provides a very useful approximation of a large dense graph.
Graphons can be considered as generalizations of graphs, and this way of look-
ing at them is very fruitful. In fact, many results can be stated and proved for
graphons in a more natural and cleaner way. In particular, regularity lemmas can
be extended to graphons, where we will see that they are statements about approx-
imating general measurable functions by stepfunctions. Approximating graphs by
multitype quasirandom graphs is as basic a tool in graph theory as approximating
functions by stepfunctions is in analysis.
Remark 1.4. Much of this book is about finite, countable and uncountable graphs
and connections between them. There are two technical limitations of measure
theory that we have to work our way around. (a) One cannot construct more than
countably many independent random variables (in a nontrivial way, neither of them
concentrated on a single value). This is the reason while we cannot define a random
graph on an uncountable set like [0, 1], only on finite and countable subsets of it. (b)
There is no uniform distribution on a countable set (while there is one on every finite
set and then again on sets with continuum cardinality like [0, 1]). This limitation
is connected to the fact that the limit objects for convergent graph sequences will
be graphons (which could be considered as graphs defined on a continuum) rather
than graphs on a countable set as one would first expect.
I want to emphasize that these difficulties are not just annoying technicalities:
they reflect the fact, for example, that the limit object of a convergence graph se-
quence carries a lot more information than what could be squeezed into a countable
graph. Both measure theory and combinatorics force us into the same realm.

1.6. How to run algorithms on them?

1.6.1. Parameter estimation. What can we learn about a huge graph G
from sampling? There are several related questions here, depending on what we
need as a result. The easiest setup is when we want to compute a numerical
parameter of the graph; say, how large is the maximum cut, or what fraction of
the triples induce a triangle. We call this problem parameter estimation. Most of
the time we normalize the parameter to be between 0 and 1. Since, as discussed
above, we get information about the graph through random sampling, any answer
we can possibly compute will, with some probability, be in error. So we will have
to specify an error parameter ε > 0, and will have to accept an answer which, with
probability at least 1 − ε, will be closer than ε to the true value of the parameter.
(An
) easy example is to estimate the triangle density (number of triangles divided
by n3 ). A trivial algorithm is to pick many random triples of nodes independently,
 1.6. HOW TO RUN ALGORITHMS ON THEM? 19

and count how many of them form triangles in the graph. Elementary statistics
tells us that if we sample O(ε−2 | log ε|) triples, then with probability at least 1 − ε,
our estimate will be closer than ε to the truth.
A much more interesting and difficult example( ) is that of estimating the density
a of the maximum cut (its size divided by n2 ) in a graph G. One thing we can
try is to choose N random nodes (where N depends on the error bound ε), and
compute the density X of the maximum cut in the subgraph H they induce. Is X
a good estimate for a?
The inequality X ≥ a − ε (for every ε > 0 if N is large enough, with high
probability) is relatively easy to prove. The graph G has a cut C with density a,
and this cut provides a cut C ′ in the random induced subgraph H. It is easy to
see that the density of C ′ is the same as the density a of C in expectation, and
it takes some routine computation in probability theory to show that it is highly
concentrated around this value. The density X of the largest cut in H is at least
the density of C ′ , and so with high probability it is at least a − ε (Figure 1.9).

Figure 1.9. A dense cut in the large graph gives a dense cut in the sample.

The reverse inequality is much more difficult to prove, at least from scratch,
and in fact it is rather surprising. We can phrase the question like this: Suppose
that most random induced subgraphs H on N nodes have a cut that is denser than
b. Does it follow that G has a cut that is denser than b − ε? It is not clear why this
should be so: why should these cuts in these small subgraphs “line up” to give a
dense cut in G?
We will see that it does follow that the estimate is correct, once N is large
enough (about ε−4 | log ε|). In fact, one can give general necessary and sufficient
conditions under which parameters can be estimated by sampling, as we will see in
Section 15.1.
1.6.2. Property testing. A more complicated issue is property testing: we
want to determine whether the graph has some given property, for example, can
it be decomposed into two connected components of equal size, is it planar, or
does it contain any triangle. We could consider this as a 0-1 valued parameter, but
computing this parameter approximately would not make sense (or rather, it would
be requiring too much, since this would be equivalent to exact computation).
A good way of posing this problem was developed by Rubinfeld and Sudan
[1996] and Goldreich, Goldwasser and Ron [1998]. In the slightly different context of
“additive approximation”, closely related problems were studied by Arora, Karger
20 1. VERY LARGE NETWORKS

and Karpinski [1995] (see e.g. Fischer [2001] for a survey and the volume edited by
Goldreich [2010] for a collection of more recent surveys).
This approach acknowledges that any answer is only approximate. Suppose that
we want to test for a property P, and we get information about the graph by taking
a bounded size random sample of the nodes, and inspecting the subgraph induced
by them. We interpret the answer of the algorithm as follows: If it concludes that
the graph has property P, this means that we can change εn2 edges so that we get
a graph with property P; if it concludes that the graph does not have property P,
this means that we can change εn2 edges so that we get a graph without property
P.
Again, we have to specify an error parameter ε > 0 in advance, and will have to
accept an answer which may be wrong with probability ε, and even if it is “right”, it
only means that we can change εn2 edges in the graph so that the answer becomes
correct.
Sometimes we can do better and eliminate either false positives or false neg-
atives. As an example, let us try to test whether a given (dense) graph contains
a triangle. We take a sample of size f (ε) (the best function f which is known to
work is outrageously large, but let’s not worry about this), and check whether they
contain a triangle. If they do, then we know that the graph has a triangle. If they
don’t, then one can prove (see Section 15.3) that with high probability, we can
delete εn2 edges from the graph so that no triangle remains.
Remark 1.5. We will not be concerned with the sample size as a function of the
error bound ε. Sometimes it is polynomial (as in the examples above), but in other
cases one uses the Regularity Lemma, which forces tower-size samples, making the
algorithms of theoretical interest only. Goldreich [2010], in his survey of property
testing, emphasizes the importance of testing with samples of manageable size, and
I could not agree more; but this book, being about limit theory, does not address
this issue.
Another caveat: Many extensions deal with testing models where we are allowed
to sample more than a constant number of nodes of the large graph G. For this,
we have to take the number of nodes into account, but usually it is enough to know
the order of magnitude of the number of nodes, which in practical situations is easy
to do. We do not discuss these important methods in our book.
1.6.3. Computation of a structure. Perhaps the most complex algorithmic
task is the computation of a structure, where the structure is of size comparable
with the graph itself. For example, we want to find a perfect matching in the
graph, or a maximum cut (not just its density, but the cut itself), or a regularity
partition in a huge dense graph. The conceptual difficulty is that the output of
the algorithm is too large to be explicitly produced. What we can do is to carry
out some preprocessing whose result can be stored (e.g., label a bounded number
of nodes or edges), and give an algorithm which, for given input node or nodes,
determines the local part of the structure we are looking for. Usually, this algorithm
returns the “status” of a node or edge in the output structure (for example, whether
the given edge belongs to matching, or which side of the cut the given node belongs
to).
As an example, we will describe in Section 15.4.3 how to compute a maximum
cut. We can access the graph by taking a bounded size sample of the nodes,
and inspect the subgraph induced by them. For a given ε > 0, we precompute a
 1.6. HOW TO RUN ALGORITHMS ON THEM? 21

“representative set” (see next section) together with a bipartition of this set. In
addition, we describe a “Placing Algorithm” which has an arbitrary node v as its
input, and tells us on which side of the cut v is located. This Placing Algorithm
can be called any number of times with different nodes v, and the answers it gives
should be consistent with an approximately maximum cut. For example, calling
this algorithm many times, we can estimate the density of the maximum cut (but
this can be done in an easier way, as we have seen).
The parameter ε is an error bound: the cut computed may be off the true
maximum cut by εn2 edges, the precomputation may be wrong with probability at
most ε, and for each query, the answer may be in error with probability at most ε.

1.6.4. Representative set. Szemerédi partitions are closely related to the

main ingredient in these algorithms, namely a “representative set”. We want to
select a (fairly large, but bounded size) subset R of the nodes such that every node
is “similar” to one of the nodes in R. To be economical, we don’t want to include
similar points in R.
We must start with defining what “similar” means; we will do so by defining
a “similarity distance” between two nodes of a graph. A first idea would be to
use their distance in the graph (the length of the shortest path connecting them).
However, this measures something else (the prime minister and the doorman in his
office know each other, but their positions in the society are certainly not similar).
We could try considering two nodes similar, if their neighborhoods differ by
little. This is certainly a reasonable thing to do, but it is too restrictive for our
purposes. For example, if we consider a random graph on n nodes with edge density
1/2, then the neighborhoods of any two nodes are very different (they have about
n/2 elements and overlap in about n/4), but all nodes of a random graph are alike,
so we would like them to be close in the similarity distance.
It turns out (somewhat surprisingly) that it suffices to consider second neigh-
borhoods: we consider two nodes s and t similar, if for most other nodes v, the
number of paths of length two from s to v is about the same as the number of paths
of length two from t to v. The similarity distance defined this way (for the exact
definition, see Section 15.4.1) has many nice properties:
• The similarity distance can be computed by sampling.
• For every ε > 0, every graph has a “representative set” R of nodes, whose
size depends on ε only; nodes in this set are at least ε apart, and almost
every node is at a distance less than ε from the representative set.
• The representative set can be computed by sampling.
• Borrowing a phrase from geometry, we define the Voronoi cell of a node
v of the representative set R as the set of all nodes in the whole graph
that are closer to v than to any other node of R. The Voronoi cells of the
representative set give a Weak Regularity Partition, and vice versa, every
Weak Regularity Partition, after deletion of a fraction of ε of the nodes,
consists of sets with small diameter in the similarity distance.
The key to many structural computational problems is that first a represen-
tative set is computed, and then the status of any node or edge can be computed
using the representative set. For example, if we want to compute a Weak Regularity
Partition, we compute a representative set, which we consider as a set of represen-
tative nodes of the partition classes, which are the Voronoi cells of the nodes. We
22 1. VERY LARGE NETWORKS

cannot compute all the Voronoi cells; but if we want to know which class (cell) does
a given node belong to, all we need to do is to compute its distance to the nodes
in R.

1.7. Bounded degree graphs

Let us discuss briefly how, and to what degree, the above considerations carry
over to graphs with bounded degree. (We are doing injustice here to a rich and
very active research area; I hope some of this will be rectified in Part 4 of the book.
One of the reasons is that the technicalities in the bounded degree case are deeper,
and so it is more difficult to state key results, even informally.)

Sampling. In the case of graphs with bounded degree, the subgraph sampling
method gives a trivial result: the sampled subgraph will almost certainly be edge-
less. Probably the most natural way to fix this is to consider neighborhood sampling
(Figure 1.1). Let GD denote the class of finite graphs with all degrees bounded by
D. For G ∈ GD , select a random node and explore its neighborhood to a given
depth r. This provides a probability distribution ρG,r on graphs in GD , with a
specified root node, such that all nodes are at distance at most r from the root.
We will briefly refer to these rooted graphs as r-balls. Note that the number of
possible r-balls is finite if D and r are fixed.
The situation for bounded degree graphs is, however, less satisfactory than for
dense graphs, for two reasons. First, a full characterization of what distributions
of r-balls the neighborhood sampling procedure can result in is not known (cf.
Conjecture 19.8). Second, neighborhood sampling misses some important global
properties of the graph, like expansion. In Section 19.2 we will introduce a notion
of convergence, called local-global, which is better from this point of view, but it is
not based on any implementable sampling procedure.
This suggests looking at further possibilities. Suppose, for example, that in-
stead of exploring the neighborhood of a single random node, we could select two
(or more) random nodes and determine simple quantities associated with pairs of
nodes, like pairwise distances, maximum flow, electrical resistance, hitting times
of random walks (studies of this nature have been performed, for example, on the
internet, see e.g. Kallus, Hága, Mátray, Vattay and Laki [2011]). What information
can be gained by such tests? Is there a “complete” set of tests that would give
enough information to determine the global structure of the graph to a reasonable
accuracy? Such questions could lead to different theories of large graphs and their
limit objects; at this time, however, they are unexplored.

Remark 1.6. It is interesting to note that our two sampling methods correspond
to the two basic data structures for graph algorithms, adjacency matrix and neigh-
borhood lists. To be more specific, both methods assume that we can choose a
uniformly distributed random node, and repeat this a constant number of times.
In subgraph sampling, we must be able to determine whether two given nodes are
adjacent or not. For a graph that is explicitly given, this is easy if the graph is given
by its adjacency matrix. For neighborhood sampling, we have to be able to find all
neighbors of a given node. This is easy if the graph is given by neighborhood lists.
It would be very time consuming to perform these sampling operations on a graph
given by the wrong data structure.
 1.7. BOUNDED DEGREE GRAPHS 23

Sampling distance. The construction of the sampling distance can be carried over
to graphs with bounded degree, by replacing in (1.2) the sampling distributions σG,k
by the neighborhood distributions ρG,k . We must point out, however, that it seems
to be difficult to define a notion of distance between two graphs with bounded
degree (in analogy with the cut distance) that would reflect global similarity.

Regularity Lemma. This is one of the big unsolved problems for graphs with
bounded degree. If we consider regularity lemmas as providing “approximation by
the smaller”, then there is a simple non-constructive result (Proposition 19.10),
which should be proved in a constructive way to be really useful. One can start at
many other facets of the Regularity Lemma, but a satisfactory version of bounded
degree graphs has turned out most elusive.

Convergence. The notion of a convergent sequence of bounded degree graphs

was in fact the first among such convergence notions, introduced by Benjamini
and Schramm [2001], motivated in part by earlier work of Aldous [1998]. Our
discussion of local convergence of dense graphs above, based on the convergence of
the distribution of samples, was modeled on the Benjamini–Schramm definition of
convergence of bounded degree graphs.
There are, however, good reasons to try to strengthen this notion. Unlike in the
dense case, neighborhood sampling cannot distinguish between bipartite graphs and
graphs that are far from being bipartite, cannot estimate the maximum cut etc.,
which means that locally convergent graph sequences must lose this information
in the limit. We will introduce and study a stronger notion of convergence, which
we call local-global, which passes on these properties and parameters to the limit.
However, we don’t know if there is any natural and practical algorithmic setup that
would correspond to local-global convergence.

Limit objects. For bounded degree graphs, Benjamini and Schramm provide a
notion of a limit object (see Section 18). The Benjamini–Schramm limit object can
be described as a distribution on rooted countable graphs with a special property
called “involution invariance”.
Another way of describing a limit object is a “graphing”. In a sense, this latter
object is what we expect: a bounded degree graph on an infinite (typically un-
countable) set, with appropriate measurability and measure preserving conditions.
This construction was folklore in an informal way; the first exact statements were
published by Aldous and Lyons [2007] and Elek [2007a].
Graphings were invented by group theorists. The idea is to consider a finitely
generated group acting on a probability space (for example, rotations of a circle by
integer multiples of a given angle). One can construct a graph on the underlying
space, by connecting each point to its images under the generators of the group.
This construction gives a graph with bounded degree (the set of points is typically
of continuum cardinality). It is a beautiful fact that
graphings, representing groups this way, are just right to describe the limit
objects of convergent graph sequences with bounded degree.
Depending on personal taste, a graphing may be considered more complicated
or less complicated than an involution-invariant random countable rooted graph.
But graphings have an important advantage: they can express a richer structure,
the limits of graph sequences convergent in the local-global sense.
24 1. VERY LARGE NETWORKS

Algorithms. Here is finally an area where the study of bounded degree graphs
can be considered at least as advanced as the study of dense graphs. Let us discuss
the task of computing a structure.
Selecting random nodes and exploring their neighborhoods, we see (with high
probability) disjoint parts of the graph, and so there is no method to build up a
global structure. Still, very nontrivial algorithms can be designed in this model.
For example, in Section 22.3.1 we describe an algorithm due to Nguyen and Onak
[2008], that constructs an almost maximum matching. The way the output can be
described is similar to how the output of a maximum cut algorithm was described
in the dense setting: for any node we can tell which other node it is matched
to, inspecting a bounded neighborhood only; these assignments will be consistent
throughout the graph; and the difference in size from the true maximum matching
is only εn, where ε > 0 is an error bound and n is the number of nodes.
There is an equivalent way to describe such algorithms, which may be easier
to follow, and this is the model of distributed computing (going back to the 1980’s).
In this case, an agent (or processor) is sitting at each node of the graph, and they
cooperate in exploring various properties of it. They can only communicate along
the edges. In the case we are interested in (which is in a sense extreme), they are
restricted to exchange a bounded number of bits (where the bound may depend on
the degree D, on an error bound ε, and of course on the task they are performing,
but not on the number of nodes). In some other versions of the model (cellular
automata), the amount of communication is not restricted, but the computing
power of the agents is. Note that in our model communication between the agents
is restricted to a bounded number of bits, and hence they may be assumed to be
very stupid, even finite automata.
There is a large literature on distributed computing, both from the practical
and theoretical aspect. We will not be able to cover this; we will restrict ourselves
to the discussion of the strong connection of this computation model with our
approach to large graphs and graph limits.
 CHAPTER 2

Large graphs in mathematics and physics

The algorithmic treatment of very large networks is not the only area where the
notions of very large graphs and their limits can be applied successfully. Many of
the problems and methods in graph limit theory come from extremal graph theory
or from statistical physics. Let us give s very brief introduction to these theories.

2.1. Extremal graph theory

Extremal graph theory is one of the oldest areas of graph theory; it has some el-
egant general results, but also many elementary extremal problems that are still un-
solved. Graph limit theory (mostly the related theory of flag algebras by Razborov)
has provided powerful tools for the solution of some of these problems. Further-
more, graph limits, along with the algebraic tools that will be introduced soon, will
enable us to formulate and (at least partially) answer some very general questions
in extremal graph theory (similarly to the general questions for very large graphs
posed in the previous chapter).

2.1.1. Edges vs. triangles. Perhaps the first result in extremal graph theory
was found by Mantel [1907]. This says that if a graph on n nodes has more than
n2 /4 edges, then it contains a triangle. Another way of saying this is that if we
want to squeeze in the largest number of edges without creating a triangle, then we
should split the nodes into two equal classes (if n is odd, then their sizes differ by
1) and insert all edges between the two classes. As another early example, Erdős
[1938] proved a bound on the number of edges in a C4 -free bipartite graph (see
(2.9) below), as a lemma in a paper about number theory.
Mantel’s result is a special case of Turán’s Theorem [1941], which is often
considered as the work that started the systematic development of extremal graph
theory. Turán solved the generalization of Mantel’s problem for any complete graph
in place of the triangle. We define the Turán graph T (n, r) (1 ≤ r ≤ n) as follows:
we partition [n] into r classes as equitably as possible, and connect two nodes if and
only if they belong to different classes. Since we are interested in large n and fixed
r, the complication that the classes cannot be exactly equal in size (which causes
the formula for the number of edges of T (n, r) to be a bit ugly) should not worry
us. It will be enough to know that the number of edges in a Turán graph is
( )( )
( ) r n 2
e T (n, r) ∼ ,
2 r
and in terms of( the homomorphism
) densities defined in the previous chapter in
(1.1), we (have t K2 , T)(n, r) ∼ 1 − 1r . For the triangle density we have the similar
formula t K3 , T (n, r) ∼ (1 − 1r )(1 − 2r ).
25
26 2. LARGE GRAPHS IN MATHEMATICS AND PHYSICS

Theorem 2.1 (Turán’s Theorem). Among all graphs on n nodes containing no

complete k-graph, the Turán graph T (n, k − 1) has the maximum number of edges.
Let us return to triangles, however, and ask for not just their existence, but
for their number, when the number of edges is known. All of a sudden, we get to
a rather difficult problem with some unexpected complications (which makes the
subject fascinating). It is really difficult to think of a simpler question about small
subgraphs of a large graph!
Since we are interested in large n, it is natural to normalize, and use homomor-
phism densities. The Mantel–Turán Theorem says, in this language, that
(2.1) t(K2 , G) > 1/2 ⇒ t(K3 , G) > 0.
( )
Every graph G produces a pair of numbers t(K2 , G), t(K3 , G) this way, which
we can consider as a point in the plane. If we plot this point for every graph G, we
get a picture as in Figure 2.1(a). To be more precise, we get a countably infinite
set of points; the figure shows its closure, which we denote by D2,3 . (Another
motivation for introducing convergent graph sequences and their limit objects: they
give a meaning to all points of this figure.)

(a) (b) (c)

Figure 2.1. (a) The closure D2,3 of the set of pairs of edge density
and triangle density. (b) Goodman’s bound. (c) Bollobás’s’ bound.
The picture is a little distorted in order to show its special features
better.

Some features of this picture are easy to explain. The lower edge means that
there are triangle-free graphs with edge density up to 1/2, and the Mantel–Turán
Theorem says that for larger edge density, the triangle density must be positive. A
lower bound for the triangle density was proved by Goodman [1959],
(2.2) t(K3 , G) ≥ t(K2 , G)(2t(K2 , G) − 1),
which corresponds to the parabola shown in 2.1(b).
The upper boundary curve turns out to be given by the equation y = x3/2 ,
which is a very special case of the Kruskal–Katona Theorem in extremal hypergraph
theory (the full theorem gives the precise value, not just asymptotics, and concerns
uniform hypergraphs, not just graphs). In other words, this says that
(2.3) t(K3 , G) ≤ t(K2 , G)3/2 .
 2.1. EXTREMAL GRAPH THEORY 27

Both (2.2) and (2.3) are sharp in a sense: Goodman’s Theorem is sharp if the
edge density is of the form 1/2, 2/3, 3/4, . . . (Turán graphs give equality). In this
form of the Kruskal–Katona Theorem equality is not attained except at the points
(0, 0) and (1, 1), but for every point (x, x3/2 of the upper boundary curve there
are points representing a graph arbitrarily close (just use graphs consisting of a
complete graph and isolated nodes).
From our perspective, there is nothing to improve on the upper bound, but
can we get arbitrarily close to the lower bound between two special edge density
values 1 − 1/k? Surprisingly, the answer is no. Bollobás [1976] proved in 1976 that
the triangle density for a graph with edge density x ∈ (1 − k−1 1
, 1 − k1 ) is not only
above the parabola, but also above the chord of the parabola connecting the special
points corresponding to T (n, k − 1) and T (n, k).
Lovász and Simonovits [1976, 1983] formulated a conjecture about the exact
bounding curve, and proved it in very small neighborhoods of the special edge
density values above. One way to state this is that the minimum number of triangles
is attained by a complete k-partite graph with unequal color classes. The sizes of
the color classes can be determined by solving an optimization problem, which
leads to a cubic concave curve connecting the two special points. This conjecture
turned out quite hard. Lovász and Simonovits proved it in the special case when
the edge density x was close to one of the endpoints of the interval. Fisher [1989]
proved the conjecture for the first interval (1/2, 2/3). After quite a while, Razborov
[2007, 2008] proved the general conjecture. His work was extended by Nikiforov
[2011] to bounding the number of complete 4-graphs, and by Reiher [2012] to all
complete graphs.
So we know what the lower and upper bounding curves are. Luckily, math
plays no further tricks on us: it is easy to see that for every point between the two
curves there are points representing graphs arbitrarily close.
I dwelt quite long on this very simple special problem not only to show how
complicated it gets (and yet solvable), but also because Razborov’s methods for the
solution fit quite well in the framework developed in this book, and they will be
presented in Chapter 16.
2.1.2. A sampler of classical results. Let us start with some remarks to
simplify and to some degree unify the statements of these results. Every algebraic
inequality between subgraph densities can be “linearized”, using the following mul-
tiplicativity of t(., G):
(2.4) t(F1 F2 , G) = t(F1 , G)t(F2 , G),
where F1 F2 denotes the disjoint union of F1 and F2 . (This property will play a very
important role in the sequel, but right now it is just a convenient simplification.)
For example, we can replace (2.2) by
(2.5) t(K3 , G) ≥ 2t(K2 K2 , G) − t(K2 , G).
We can make the statements (and their proofs, as we will see below) more
transparent by two further tricks: first, if a linear inequality between the densities
of certain subgraphs F1 , . . . , Fk holds for all graphs, then we write it as an inequal-
ity between F1 , . . . , Fk ; and for specific small graphs Fi , we use little pictograms.
Goodman’s Inequality (2.2) can be expressed as follows:
(2.6) K3 ≥ 2K2 2 − K2
28 2. LARGE GRAPHS IN MATHEMATICS AND PHYSICS

or
(2.7) ≥2 − .
The Kruskal–Katona Theorem for triangles is:

(2.8) ≤ .
Let us describe some further classical results. Instead of counting complete
graphs, we can consider the density of some other graph F in G. Erdős proved the
inequality
(2.9) t(C4 , G) ≥ t(K2 , G)4 ,
or in pictograms

(2.10) ≥ .
Graphs with asymptotic equality here are quasirandom graphs (Section 1.4.2).
Bounding from below the homomorphism density of paths is a more difficult
question, but it turns out to be equivalent to theorems of Mulholland and Smith
[1959], Blakley and Roy [1965], and London [1966] in matrix theory (applied to the
adjacency matrix). If Pk denotes the path with k nodes, then for all k ≥ 2,
(2.11) t(Pk , G) ≥ t(K2 , G)k−1 .
Regular graphs give equality here. The first nontrivial case of inequality (2.11) is

(2.12) ≥ .
Translating to homomorphisms, this means that
v(G)hom(P3 , G) ≥ hom(K2 , G)2 .
If we count the homomorphisms on the left side by the image of the middle node,
we see that it is the sum of the squared degrees of G. Since hom(K2 , G) = 2e(G)
is the sum of the degrees, this inequality is just the inequality between arithmetic
and quadratic means, applied to the sequence of degrees.
Bounding the P3 -density from above in terms of the edge density is more diffi-
cult, but it was solved by Ahlswede and Katona [1978]; we formulate this as Exercise
2.4 below.
The next case of inequality (2.11) is

(2.13) ≥ ,
and this is already quite hard, although short proofs with a tricky application of
the Cauchy–Schwarz inequality are known.
In Chapter 16 we will return to the question of how far the application of such
elementary inequalities takes us in proving inequalities between subgraph densities.

2.1.3. An algebraic “proof ” of an extremal theorem. We illustrate the

use of the formalism with pictograms for an algebraic proof of Goodman’s Inequality
2.2. This will motivate a basic tool to be introduced in Chapter 6, namely graph
algebras.
To describe this proof, we extend the pictogram formalism from Section 2.1.2.
If we fill a node, this indicates that this node is labeled . We should write the label
on the node, but to keep the picture simple, let us agree that the labels are 1, 2, . . . ,
 2.1. EXTREMAL GRAPH THEORY 29

starting from the lower left corner, and going counterclockwise. (It does not really
matter.)
The role of the labels is that when taking a “product” of two graphs, we take
the disjoint union, but identify nodes with the same label. With this convention,
it is easy to check that
( )2
− − + = − − +
(this combination is “idempotent”) and
( )2
− = −2 +
Forgetting the labels, adding up, and deleting isolated nodes, we get
( )2 ( )2
− − + +2 − = −2 + .
So the right side is a sum of squares, which implies that it is nonnegative:
−2 + ≥ 0,
which is just (2.7).
Is this a valid argument? It turns out that it is, and the method can be
formalized using the notion of graph algebras. These will be very useful tools in
the proofs of characterization theorems of homomorphism functions, and also in
some other studies of graph parameters.
2.1.4. General results. Moving from special extremal graph problems to
the more general, let us describe some quite general results about extremal graphs,
which were obtained quite a long time ago in several papers of Erdős, Stone and
Simonovits [1946, 1966, 1968]. We exclude an arbitrary graph L as subgraph of a
simple graph G, and want to determine the maximum number of edges of G, given
the number of nodes n. Turán’s Theorem 2.1 is a special case when L is a complete
graph. It turns out that the key quantity that governs the answer is the chromatic
number r = χ(G).
The Turán graph T (n, r − 1) is certainly one of the candidates for the extremal
graph, since it cannot contain any graph as a subgraph that has chromatic number
r. For certain excluded graphs L it is easy to construct examples that have slightly
more edges than this Turán graph; however, the gain is negligible: for every graph
G on n nodes that does not contain L as a subgraph, we have
( 1 )( n )
(2.14) e(G) ≤ (1 + o(1))e(T (n, r − 1)) = 1 − + o(1) .
r−1 2
There is also a “stability” result: For every ε > 0 there is an ε′ > 0 (depending
on
( L and ε, but not ) G) such that if G is a graph not containing
)( on ( ) L with at least
1 − 1/(r − 1) − ε′ n2 edges, then we can change at most ε n2 edges of G to get a
Turán graph T (n, r − 1).
We will see that graph limit theory gives very short and elegant proofs for
these facts. The idea that extremal graph problems have “continuous versions” (in
a sense quite similar to our use of graphons), which are often cleaner and easier to
handle, goes back to around 1980, when Katona [1978, 1980, 1985] and Sidorenko
[1980, 1982] used this method to generalize graph and hypergraph problems, and
also to give applications in probability theory.
30 2. LARGE GRAPHS IN MATHEMATICS AND PHYSICS

Remark 2.2. If r = 2 (which means that L is bipartite), then the main term in
(2.14) disappears, and all we get is that the number of edges is o(n2 ). Of course,
one would like to know the precise order of magnitude of the best upper bound.
This is known in several cases (e.g., small complete bipartite graphs and cycles),
but in general it seems to be a difficult unsolved problem. The extremal graphs in
this case are sparse, and quite complex: for example, C4 -free graphs with maximum
edge density are constructed from finite projective planes. Extremal problems for
graphs with excluded bipartite graphs do not seem to fit in with the framework
developed in this book, but perhaps they can serve as motivation for extending it
to sparser graphs.
2.1.5. General questions. We have brought up the idea of introducing
graphons (graph limits) in Section 1.5.3 motivated by the goal to approximate
very large networks by simpler analytic objects. We have seen that graphons pro-
vide cleaner formulations, with no error terms, of some results in graph theory (for
example, about quasirandom graphs). We will see in Section 16.7 that extremal
graph theory provides another, also quite compelling motivation: Graphons pro-
vide a way to state, in an exact way, general questions about the nature of extremal
graphs, and also help answering them, at least in some cases. (They have similar
uses in the theory of computing; cf. Chapter 15).
Which inequalities between subgraph densities are valid? Given a linear
inequality between subgraph densities (like (2.7) above), is it valid for all graphs
G? Hatami and Norine [2011] proved recently that this question is algorithmically
undecidable. We will describe the proof of this fundamental result in Section 16.6.1.
On the other hand, it follows from the results of Lovász and Szegedy [2012a] that if
we allow an arbitrarily small “slack”, then it becomes decidable (see Section 16.6.2).
Can all linear inequalities between subgraph densities be proved using
just Cauchy–Schwarz? We described above a proof of the simple inequality
(2.12) using the inequality between arithmetic and quadratic means, or equivalently,
the Cauchy–Schwarz Inequality. Many other extremal problems can be proved by
using the Cauchy–Schwarz Inequality (often repeatedly and in nontrivial ways).
Exercise 2.5 shows that Goodman’s Inequality can also be proved by this method.
How general a tool is the Cauchy–Schwarz Inequality in this context?
Using the notions of graphons and graph algebras we will be able to give an
exact formulation of this question. It will turn out that the answer is negative
(Hatami and Norine [2011], Section 16.6.1), but it becomes positive if we allow an
arbitrarily small error (Lovász and Szegedy [2012a], Section 16.6.2).
Is there always an extremal graph? Let us consider extremal problems of
the form “maximize a linear combination of subgraph densities, subject to fixing
other such combinations”. For example, “maximize the triangle density subject
to a given edge density” (the answer is given by the first nontrivial case of the
Kruskal–Katona Theorem (2.8)).
To motivate our approach, consider the following two optimization problems.
Classical optimization problem. Find the minimum of x3 − 6x over all
numbers x ≥ 0.
Graph optimization problem. Find the minimum of t(C4 , G) over all
graphs G with t(K2 , G) ≥ 1/2.
 2.1. EXTREMAL GRAPH THEORY 31

√
The solution of the classical optimization problem is of course x = 2. This
means that it has no solution in rationals, but we can find rational numbers that
are arbitrarily close to being optimal. If we want an exact solution, we have to go
to the completion of the rationals, i.e., to the reals.
The graph optimization problem may take a bit more effort to solve, but (2.9)
shows that if the edge density is 1/2, then the 4-cycle density is at least 1/16. With
a little effort one can show that equality is never attained here. Furthermore, the
4-cycle-density gets arbitrarily close to 1/16 for appropriate families of graphs: the
simplest example is a random graph with edge density 1/2 (cf. also Section 1.4.2).
The analogy with the classical optimization problem above suggests that we
should try to enlarge the set of (finite) graphs with new objects so that the appro-
priate extension of our optimization problem has a solution among the new objects.
Furthermore, we want that these new objects should be approximable by graphs,
just like real numbers are approximable by rationals. As it turns out, graphons are
just the right objects for this.
One can prove that there is always an extremal graphon, which then gives a
“template” for asymptotically extremal graphs. This follows from another fact that
can be considered one of the basic results treated in this book:
The space of graphons is compact in the cut-distance metric.
(This notion of distance was mentioned in Section 1.5.1, and will be defined in
Chapter 8; the compactness of the graphon space will be proved in Section 9.3).
Which graphs are extremal? This is not a good question (every graph is
extremal for some sufficiently complicated extremal graph problem), but replac-
ing “graph” by “graphon” makes it mathematically meaningful. Every extremal
graphon gives a “template” for asymptotically extremal graphs.

Figure 2.2. Templates for optimal solutions to some classical ex-

tremal graph results: (a) Turán’s Theorem 2.1 and Goodman’s In-
equality (2.2); (b) the Kruskal–Katona Theorem (2.3); (c) Erdős’s
inequality (2.9)

In classical extremal graph results, these templates are quite simple (Figure
2.2). A natural guess would be that all templates have the form of a stepfunction,
like the rightmost square in Figure 1.6. All of these are indeed templates for
appropriate extremal problems, but they is not all the templates: we will see that
the limit of half-graphs (the rightmost square in Figure 1.7) is also the template for
the extremal graph of a quite simple extremal problem, and there are many other,
more complicated, templates. We will prove several results about the structure of
these extremal templates (Section 16.7), but no full characterization is known.
32 2. LARGE GRAPHS IN MATHEMATICS AND PHYSICS

Exercise 2.3. Prove inequality (2.13)

Exercise 2.4. Let G be a simple graph with edge density d = t(K2 , G). Prove
that t(P3 , G) ≤ max(d3/2 , 1 − 2d + d3/2 ).
Exercise 2.5. Translate the “proof” of Goodman’s Inequality 2.2 above into a
valid proof using the Cauchy–Schwarz inequality twice.

2.2. Statistical physics

One area of research where graph homomorphisms play an important role, and
the study of the asymptotic behavior of parameters when tending to infinity with
the size of a graph is a main goal, is statistical physics. I am afraid this book will
not do justice to this connection; my excuse is that statistical physics is such a
large area, with so advanced special methods, that any reasonable treatment would
double the size of the book. Nevertheless, I must give a very short introduction to
the subject here.
To describe a basic model in statistical physics, suppose that we have a piece
of a crystal, where the spin of every atom can point either up or down. We model
this by an n × n grid G = Gn×n (for simplicity, in two dimensions). If we assign
to every node of G (every “site”) a “state”, which can be UP or DOWN, we get
a “configuration”. The atoms are changing their spins randomly all the time, but
not independently of each other: depending on the spins of adjacent atoms, one
direction of the spin of an atom may be less likely then the other, or even entirely
impossible. We would like to know how a typical configuration looks like: is it
random-like as the first picture in Figure 2.3, is it homogeneous as the second
(well, maybe with a few exceptions here and there), or is it structured in other
ways, as the third?

Figure 2.3. Three configurations of the Ising model.

Two atoms that are adjacent in the grid have an “interaction energy”, which
depends on their states. In the simplest version of the basic Ising model, the
interaction energy is some number −J if the atoms are in the same state, and J
if they are not. The states of an atom can be described by the integers 1 and −1,
and so a configuration is a mapping σ : V (G) → {1, −1}. If σu denotes the state
of atom u, then the total energy of a given configuration is
∑
H(σ) = − Jσu σv .
uv∈E(G)

Basic physics (going back to Boltzmann) tells us that the system is more likely to
be in states with low energy. In formula, the probability of a given configuration
 2.2. STATISTICAL PHYSICS 33

is proportional to e−H(σ)/T , where T is the temperature (from the point of view of

the mathematician, just a parameter). Since probabilities must add up to 1, these
values must be normalized:
e−H(σ)/T
P(σ) = ,
Z
where the normalizing factor Z is called the partition function of the system (it
is called a “function” because it depends on the temperature). This is perhaps
the most important quantity to know, which contains implicitly many important
physical parameters. The partition function is simple to describe:
∑ ∑ (1 ∑ )
Z= e−H(σ)/T = exp Jσu σv ,
σ σ
T
uv∈E(G)

but since the number of terms is enormous, partition functions can be very hard to
compute or analyze.
The behavior of the system depends very much on the sign of J. If J > 0, then
adjacent pairs that are in the same state contribute less to the total energy than
those that are in different state, and so the configuration with the lowest energy
is attained when all atoms are in the same state. The typical configuration of the
system will be close to this, at least as long as the temperature T is small. This is
called the ferromagnetic Ising model, because it gives an explanation how materials
like iron get magnetized. If J < 0 (the antiferromagnetic case), then the behavior is
different: the chessboard-like pattern minimizes the energy, and no magnetization
occurs at any temperature.
One may notice that the temperature T emphasizes the difference between the
energy of different configurations when T → 0 (and de-emphasizes it when T → ∞).
In the limit when T → 0, all the probability will be concentrated on the states with
minimum energy, which are called ground states. In the simplest ferromagnetic
Ising model, there are two ground states: either all atoms are in state UP, or all of
them are in state DOWN. If the temperature increases, disordered states like the
left picture in Figure 2.3 become more likely. The transition from the ordered state
to the disordered may be gradual (in dimension 1), or it may happen suddenly at a
given temperature (in dimensions 2 and higher, for large graphs G); this is called a
phase transition. This leads us to one of the central problems in statistical physics;
alas, we cannot go deeper into the discussion of this issue in our book.
To make the connection to graph homomorphisms, we generalize the Ising
model a little. First, we replace the grid by an arbitrary graph G. (From the point
of view of physics, other lattices, corresponding to crystals with other structure, are
certainly natural. Other materials don’t have a simple periodic crystal structure.)
Second, we introduce a “magnetic ∑ field”, which prefers one state over the other: in
the simplest case it adds − u hσu to the energy function, with some parameter h.
Third, we consider not two, but q possible states for every atom, which we label by
1, 2, . . . , q (unlike 1 and −1 before, these should not be considered as numbers: they
are just labels). We have to specify an interaction energy Jij for any two states i
and j, and a magnetic field energy hi for every state i. A configuration is now a
map σ : V (G) → [q], and the energy of it is
∑ ∑
H(σ) = − hσ(v) − Jσ(u),σ(v) .
v∈V (G) uv∈E(G)
34 2. LARGE GRAPHS IN MATHEMATICS AND PHYSICS

The partition function is

∑ ( 1( ∑ ∑ ))
Z= exp − hσ(v) + Jσ(u),σ(v) .
T
σ: V (G)→[q] v∈V (G) uv∈E(G)

We are almost at homomorphisms! For i, j ∈ [q], let

( 1 ) ( 1 )
αi = exp − hi , and βij = exp − Jij ,
T T
then the partition function can be expressed as
∑ ∏ ∏
(2.15) Z= ασ(v) βσ(u)σ(v) .
σ:V (G)→[q] v∈V (G) uv∈E(G)

Consider the case when αi = 1 for all i, and βij is 0 or 1 (in the Ising model βij
cannot be zero, but (2.15) allows this substitution). Then every term in (2.15) is
either 0 or 1, and a term is 1 if and only if βσ(u)σ(v) = 1 for every uv ∈ E(G). Let
us build a graph H with node set V (H) = [q], in which i, j ∈ [q] are adjacent if
and only if βij = 1. Then a term in (2.15) is 1 if and only if σ is a homomorphism
G → H, and so the sum simply counts these homomorphisms, and gives the value
Z = hom(G, H).
In the case of general values for the α and β, we can define a weighted graph
H with nodeweights αi and edgeweights βij . Formula (2.15) can then serve as the
definition of hom(G, H), which will be very important for us.
We don’t discuss the connections between statistical physics and graph theory
(homomorphisms and limits) any further; for an introduction to the connections
between statistical physics and graph theory, with more examples, see de la Harpe
and Jones [1993].
Exercise 2.6. Define a model in statistical physics in which the ground state
corresponds to the maximum cut of a graph.
 Part 2

The algebra of graph

homomorphisms
 CHAPTER 3

Notation and terminology

In this book, different areas of mathematics come together (graph theory, prob-
ability, algebra, functional analysis), and this makes it difficult to find good nota-
tion, and impossible in some cases to stick to standard notation. I tried to find
notation that helps readability. For example, when doing computations with small
graphs, I often use pictograms instead of introducing dozens of notations for them.
When labeling one or more nodes of a graph G, I use G• or G•• , and when adding
some loops at the nodes, I use G◦ . These graphs must be still defined, but perhaps
the meaning of the notation is easier to remember keeping this in mind.

3.1. Basic notation

Let R, C, Z denote the sets of real, complex and integer numbers. We denote
by N the set of nonnegative integers, by N∗ , the set of positive integers, by Zq ,
the set of integers modulo q, and by R+ , the set of nonnegative reals. We use the
notation [n] = {1, 2, . . . , n} and (n)k = n(n − 1) . . . (n − k + 1).
If A is a statement, then
{
1, if A is true,
1(A) =
0, otherwise.

If A is a set, then 1A is its indicator function: 1A (x) = 1(x ∈ A).

If A is a real matrix, then A ≽ 0 means that A is positive semidefinite (in
particular, symmetric), while A ≥ 0 means that all its elements are nonnegative.
For two matrices A, B ∈ Rm×n , their dot product is defined by
∑
m ∑
n
A·B = Aij Bij .
i=1 j=1

The natural logarithm will be denoted by ln; logarithm of base 2, by log. (There
is a recurring dilemma about which logarithm to use. Base 2 is used in information
theory, and it is often better suited for combinatorial problems; natural logarithm
has simpler analytical formulas. Luckily, the two differ in a constant factor only so
the difference is usually irrelevant.) We denote by log∗ x the least n for which the
n-times iterated logarithm of x is less than 1. The Lebesgue measure on R will be
denoted by λ.
We will consider partitions of both finite sets and the interval [0, 1]. A partition
of [0, 1] will be called an equipartition, if it has a finite number of measurable classes
with the same measure. A partition of a finite set V will be called equitable, if
|S| − |T | ≤ 1 for any two partition classes.

37
38 3. NOTATION AND TERMINOLOGY

3.2. Graph theory

We denote by v(G) = |V (G)| the number of nodes and by e(G) = |E(G)|, the
number of edges. The subgraph induced by S ⊆ V (G) is denoted by G[S]. For
X, Y ⊆ V (G), let eG (X, Y ) denote the number of edges with one endnode in X and
another in Y ; edges with both endnodes in X ∩ Y are counted twice. We denote
by NG (v) the set of neighbors of v in the graph G (which we abbreviate as N (v) if
the graph G is understood). We denote by ∇(v) the set of edges incident with the
node v. For every r ≥ 0 and v ∈ V (G), we denote by BG,r (v) the subgraph of G
induced by those nodes that are at a distance at most r from v. We also call this
graph the r-ball about v.
For any family C of sets, we denote by L(C) the intersection graph of C, i.e.
the graph with node set C, where two nodes (sets in C) are connected if and only
if they have a nonempty intersection. As a special case, the intersection graph of
E(G) (where G is any multigraph) is called the line-graph of G, and denoted by
L(G).
We have to introduce many types of graphs. A graph!simple is a finite graph
without loops and multiple edges. A looped-simple graph is a finite graph without
multiple edges, in which any subset of the nodes can carry a loop; equivalently, this
is a symmetric binary relation on a finite set. A multigraph is a finite graph (in
which loops and multiple edges are allowed). Let Fksimp denote the set of simple
graphs on node set [k], and Fkmult , the set of multigraphs on node set [k].
We denote by Gsimp the simple graph obtained from a multigraph G by deleting
loops as well as all but one edge from every parallel class.
Some special graphs need special names: Pn denotes the path with n nodes
(note the somewhat unusual indexing; we usually put the number of nodes in the
subscript); Cn denotes the cycle with n nodes (this is mostly used for n ≥ 3, but C2
and even C1 (a node with a loop) will be useful occasionally); Kn is the complete
graph with n nodes (including the graph K0 with no nodes and edges); Kn◦ is the
complete graph with n nodes, with a loop added at every node; Sn is the star with
n nodes; On is the graph on [n] with no edges. The m-bond B m consists of two
nodes connected by m edges.
For a simple graph G, we denote by Conn(G) the set of connected subgraphs ,
and by Csp(G), connected spanning subgraphs (note: spanning, not induced!).

Weighted graphs. A weighted graph H is a looped-simple graph, with a

positive real weight αi (H) associated with each node i and a real weight βi,j (H)
associated with each edge ij.
It is often convenient to assume that H is a complete graph with a loop at all
nodes; the missing edges can be added with weight 0. Then the weighted graph H
is completely described (up to isomorphism) by a nonnegative integer q = v(H), a
positive real vector a = (α1 , . . . , αq ) ∈ Rq of nodeweights and the real symmetric
matrix B = (βij ) ∈ Rq×q of edgeweights. We denote this weighted graph by
H(a, B). An edge-weighted graph is a weighted graph with nodeweights 1.
A simple graph can be considered as a special edge-weighted graph in which all
edge-weights are 0 or 1, and all loops have weight 0. Multigraphs can be considered
as edge-weighted graphs in which the nodeweights are 1 and the edgeweights are
nonnegative integers (but this is not always the best).
 3.3. OPERATIONS ON GRAPHS 39

Signed graphs. Suppose that the edges of a graph F are partitioned into two
sets E+ and E− . The triple F = (V, E+ , E− ) will be called a signed graph. (We
don’t consider this as a weighted graph with edge weights ±1, because these signs
will play a quite different role!)
Partially labeled graphs. This less standard type of graphs will play a
crucial role in this book. A simply k-labeled graph is a graph in which k of the
nodes are labeled by 1, . . . , k (there may be any number of unlabeled nodes). A
k-multilabeled graph is a graph in which labels 1, . . . , k are attached to some nodes;
the same node may carry more than one label (but a label occurs only once). So
a k-multilabeled graph is a graph F together with a map [k] → V (F ), and this is
k-labeled if this map is injective. We omit “simply” from k-labeled, unless we want
to emphasize that it is simply k-labeled. The set of isomorphism types of k-labeled
multigraphs will be denoted by Fk• .
More generally, for every finite set S ⊆ N of labels we can talk about S-labeled
and S-multilabeled graphs. A partially labeled graph is an S-labeled graph for
some finite set S. A 0-labeled graph (or equivalently an ∅-labeled graph) is just
an unlabeled graph. The set of S-labeled multigraphs will be denoted by FS• . A
partially labeled graph in which all nodes are labeled will be called flat or fully
labeled or flat.
For every partially labeled graph G and S ⊆ N, let [[G]]S denote the partially
labeled graph obtained by removing the labels not in S . For S = ∅, we denote
[[G]]∅ simply by [[G]]; this is the unlabeled version of the graph G.
We need some notation for differently labeled versions of some basic graphs
(Figure 3.1). We denote by Kn , Kn• , Kn•• , . . . the complete graph with 0, 1, 2, . . .
nodes labeled. . We denote by Pn , Pn• , Pn•• the path on n nodes with 0, 1, 2 endnodes
labeled. . The m-bond labeled at both nodes will be denoted by B m•• . . We de-
• ••
note by Ka,b , Ka,b , Ka,b the complete bipartite graph with a nodes in the “first”
bipartition class and b nodes in the “second”, with no node labeled, the first bi-
partition class labeled, and all nodes labeled, respectively. In figures, the labeled
nodes are denoted by black circles, the labels ordered left-to-right or up-down. The
2-multilabeled graph consisting of a single node will be denoted by K1•• .
The adjacency matrix of a multigraph G is the V (G) × V (G) matrix AG where
(AG )ij is the number of edges connecting node i and j. In the case of a simple
graph, this is a 0-1 matrix. For a weighted graph, we let (AG )ij denote the weight
of the edge ij (the nodeweights can be encoded in a separate vector in RV (G) ).
Colored graphs. We will use graphs in which all the edges and all the nodes
are colored (so they are colorful objects indeed). To be precise, a colored graph of
type (b, c) (where b and c are positive integers) is a multigraph (possibly with loops)
G = (V, E), which is node-colored with b colors and edge-colored with c colors.

3.3. Operations on graphs

For the standard notions of edge-deletion, contraction, subdivision, and minor,
we refer to any textbook. We will need some less standard operations on graphs.
Twin reduction. Let H be a weighted graph and let i, j ∈ V (H) be two nodes
such that βik = βjk for every node k; in particular, this includes that βii = βjj = βij
(but we allow that αi ̸= αj ). Such a pair of nodes will be called twin nodes. In
the case when H is a simple graph, twin nodes are nonadjacent and have the same
neighborhood. Interchanging a twin pair is an automorphism in the unweighted
40 3. NOTATION AND TERMINOLOGY

Figure 3.1. The most often used partially labeled graphs

case, but not necessarily in the weighted case, since their nodeweights may be
different.
Let H ′ be obtained by identifying two twin nodes i and j, which means that
we delete j, and add αj (H) to αi (H). We can repeat this operation until we get a
weighted graph with no twins. The construction leading to this twin-free weighted
graph is called twin reduction. It is not hard to see that the twin-free graph obtained
from a given graph by twin reduction is uniquely determined.
Quotient. Let P be a partition of V (G). We denote by G/P the graph obtained
by merging each class of P into a single node. This definition is not precise; in
different parts of the book, we need it with edge multiplicities summed, averaged,
or maximized. If G is a simple graph (or a looped-simple graph, then one natural
interpretation is that G/P is a looped-simple graph, in which two nodes are adjacent
if and only if they have adjacent pre-images. we will call this the simple quotient..
For other versions of the quotient construction, instead of introducing a differ-
ent notation for each of these versions, we will define how the edges are mapped
whenever we use this notation.
Blow-up. We define the m-blowup G(m) of a graph G if it is obtained by replacing
each node of G by m twin copies (m ≥ 1). Sometimes we need a blow-up of G with
a given number of nodes, and so we need a little more general notion: we say that
a graph G′ is a near-blowup of G if it is obtained by replacing each node of G by
m or m + 1 twin copies for some m ≥ 1.
Product of graphs. For two looped-simple graphs G1 and G2 , their categorical
(weak){( product G1 × G2 )is defined by V (G1 × G2 ) = V (G}1 ) × V (G2 ), and E(G1 ×
G2 ) = (u1 , u2 ), (v1 , v2 ) : u1 v1 ∈ E(G1 ), u2 v2 ∈ E(G1 ) . We denote by G×k the
k-fold categorical product of G with itself.
If G1 and G2 are simple, then so is G1 × G2 . The strong product G1  G2 of two
simple graphs can be defined by adding a loop at every node, taking the categorical
product, and then removing the loops.
A further operation on graphs is the Cartesian {( sum G1 G2 ,) defined by
V (G1 G2 ) = V (G1 ) × V (G2 ) and E(G1 G2 ) =} (u1 , u2 ), (v1 , v2 ) : u1 v1 ∈
E(G1 ) and u2 = v2 , or u1 = v1 and u2 v2 ∈ E(G1 ) .
 CHAPTER 4

Graph parameters and connection matrices

4.1. Graph parameters and graph properties

A graph parameter is a function defined on isomorphism types of multigraphs
with loops. We will mostly consider real valued graph parameters; we’ll say explic-
itly when complex values are also allowed. A graph parameter f is called simple if
its value is not changed when loops are removed and edge multiplicities are reduced
to 1. Equivalently, we can think of a simple graph parameter as being defined on
simple graphs only, but it is often convenient to extend it to multigraphs G by
f (G) = f (Gsimp ).
A graph parameter f is additive, if f (G) = f (G1 ) + f (G2 ) whenever G is the
disjoint union of G1 and G2 ; it is multiplicative if f (G) = f (G1 )f (G2 ), and maxing
if f (G) = max{f (G1 ), f (G2 )}. We say that a graph parameter is normalized if its
value on K1 , the graph with one node and no edge, is 1. Note that if a graph
parameter is multiplicative and not identically 0, then its value on K0 (the graph
with no nodes and no edges) is 1. We call a graph parameter isolate-indifferent if
its value does not change when isolated nodes are removed. Every multiplicative
and normalized graph parameter is isolate-indifferent.
There are, of course, too many graph parameters to be treated in a unified way.
We will need (and say a few words about) those which are everybody’s favorites:
• the maximum size of a stable set of nodes α(G): this is additive;
• the maximum number of independent edges (matching number) ν(G) (addi-
tive);
• the number of perfect matchings pm(G) (multiplicative);
• the size of the maximum clique ω(G) (maxing);
• the chromatic number χ(G) (maxing);
Our main objects of study will be graph parameters defined by counting ho-
momorphisms into, and from, given graphs. We discuss those in detail in the next
chapter.
A graph property is a class of graphs that is invariant under isomorphism. We
identify every graph property P with its indicator function 1P , so for us, graph
properties are just 0-1 valued graph parameters.
Of course, there are almost as many graph properties in the literature as there
are graph parameters. For our purposes, some “properties of properties” will be
important. In particular we will often consider the following special types of prop-
erties.
• Monotone property: inherited by subgraphs, i.e., G ∈ P implies that G′ ∈ P
for every subgraph G′ of G. To be bipartite, triangle-free, or planar are examples
of monotone properties.
41
42 4. GRAPH PARAMETERS AND CONNECTION MATRICES

• Hereditary property: inherited by induced subgraphs. All monotone proper-

ties are also hereditary. Further (non-monotone) hereditary properties are being
perfect, or triangulated, or a line-graph.
• Minor-closed property: inherited by minors. Being planar, or series-parallel,
or linklessly embedable in 3-space are such properties.
These monotonicity conditions can be extended to real valued graph parameters
in a natural way. For example, a graph parameter f is called minor-monotone, if
f (G′ ) ≤ f (G) whenever G′ is a minor of G.
An important operation on graph parameters is the Möbius transformation.
It is best to introduce this here, because we will need to use more than one kind.
Appendix A.1 introduces the Möbius transformation on a general finite lattice. We
will need three special cases for graphs:
• The upper Möbius inverse of a simple graph parameter f (with respect to the
lattice of simple graphs on the given nodes set) is defined by
∑ ′
(4.1) f ↑ (F ) = (−1)e(F )−e(F ) f (F ′ )
F′

(the summation ranges over all simple graphs F ′ ⊇ F with V (F ′ ) = V (F )).

• The lower Möbius inverse of a multigraph parameter f is defined by
∑ ′
(4.2) f ↓ (F ) = (−1)e(F )−e(F ) f (F ′ )
F′

(the summation ranges over all subgraphs F ′ ⊆ F with V (F ′ ) = V (F )).

• The Möbius inverse of a graph parameter f , relative to the partition lattice,
is defined by
∑
(4.3) f ⇓ (F ) = µP f (F/P ),
P

where P ranges over all partitions of V (F ), and µ is the Möbius function of the
partition lattice, given by (A.2) (the actual value of µP will not be important to
us, just that such integers exist). This is in fact the “lower” Möbius inverse on the
partition lattice, but thankfully we don’t need the upper one in this book. By the
general properties of Möbius inversion, we have the relations
(4.4) ∑ ∑ ∑
f (F ) = f ↑ (F ′ ), f (F ) = f ↓ (F ′ ), f (F ) = f ⇓ (F ′ ).
F ′ ⊇F F ′ ⊆F P
V (F ′ )=V (F ) V (F ′ )=V (F )

Exercise 4.1. Let f be a multiplicative graph parameter. Prove that f ↓ is

multiplicative as well.
Exercise 4.2. Let f be an additive graph parameter. Prove that f ↓ (G) = 0 if G
is disconnected with at least two non-singleton components.

4.2. Connection matrices

Let F1 and F2 be two partially multilabeled graphs. We define their gluing
product (or often just product, if there is no danger of confusion with any other
product notion) F1 F2 by taking their disjoint union, and then identifying nodes
with the same label. Note that this may force further identifications, since labels
i and j may occur on the same node u in F1 , but on different nodes v and v ′ in
 4.2. CONNECTION MATRICES 43

F2 , in which case u, v and v ′ must be identified (Figure 4.1). (If F1 and F2 are
simply labeled, then this does not happen, and F1 F2 is also simply labeled. If
F1 and F2 are k-labeled, then F1 F2 is also k-labeled.) Another way to describe
this construction: form the disjoint union of F1 and F2 , add edges between nodes
with the same label, and contract the new edges. So the new labeled nodes will
correspond to the connected components of the graph on the original labeled nodes,
formed by the new edges.
Even if F1 and F2 are simple graphs, which are k-multilabeled, their product
may have loops and parallel edges. If F1 and F2 are simply k-labeled and have no
loops, then F1 F2 has no loops, but may have multiple edges. For two 0-labeled
(i.e., unlabeled) graphs, F1 F2 is their disjoint union. Clearly this multiplication is
associative and commutative.

Figure 4.1. Top: the product of two simply partially labeled

graphs. Bottom: the product of two 4-multilabeled graphs

Example 4.3. Consider edgeless fully k-multilabeled graphs. Such a graph is given
by a partition of the label set [k]. The product of two such graphs is also edgeless,
and it corresponds to the join of the partitions in the partition lattice (see Appendix
A.1). The (unique) simply labeled graph in this class corresponds to the discrete
partition; the graph with one node corresponds to the indiscrete partition. 
Our basic tool to study a graph parameter will be the sequence of its con-
nection matrices: These are infinite matrices, one for every integer k ≥ 0, whose
linear algebraic properties are closely related to graph-theoretic properties of graph
parameters.
Let f be any multigraph parameter and fix an integer k ≥ 0. We define the
k-th multilabeled connection matrix of the graph parameter f as the (infinite) sym-
metric matrix M mult (f, k), whose rows and columns are indexed by (isomorphism
types of) k-multilabeled multigraphs, and the entry in the intersection of the row
corresponding to F1 and the column corresponding to F2 is f ([[F1 F2 ]]). The sub-
matrix corresponding to the simply k-labeled graphs is denoted by M simp (f, k) or
just M (f, k), and will be called simply the k-th connection matrix (Figure 4.2).
The submatrix of M (f, k) formed by rows and columns that are fully labeled (flat)
will be called the flat connection matrix and denoted by M flat (f, k). If the graph
parameter f is a simple graph parameter, then in M (f, k) those rows that corre-
spond to rows indexed by graphs with loops and/or multiple edges are just copies
of rows indexed by simple graphs, and similarly for the columns.
44 4. GRAPH PARAMETERS AND CONNECTION MATRICES

Sometimes it is convenient to work with a single connection matrix M (f, N),

whose rows and columns are indexed by all partially labeled graphs, and (as before)
the entry in row F1 and column F2 is f ([[F1 F2 ]]). Trivially, M (f, N) contains all
the connection matrices M (f, k) as submatrices, but it does not carry substantially
more information, at least for parameters that will be most interesting for us: if f
is isolate-indifferent, then every finite submatrix of M (f, N) is also a submatrix of
M (f, k) for every sufficiently large k.

 
0 1 0 1 2 ...
 
1 2 1 2 3 . . .
 
 
0 1 0 1 2 . . .
 
 
1 2 1 2 3 . . .
 
 
2 3 2 3 4 . . .
 
 
.. .. .. .. .. ..
. . . . . .

Figure 4.2. Some rows and columns of the second connection

matrix. The entries are obtained by applying a graph parameter
to the graphs shown. If the graph parameter is the number of
edges, we get the infinite matrix on the right.

Two possible properties of connection matrices will be particularly important

for us. We call the graph parameter f reflection positive if all the corresponding
connection matrices M (f, k) are positive semidefinite. For isolate-indifferent pa-
rameters, this is equivalent to saying that M (f, N) is positive semidefinite. (To be
precise, we have to talk about reflection positivity with respect to multilabeled or
simply labeled connection matrices. If not said explicitly, we mean simply labeled.)
We call the parameter flatly reflection positive if its flat connection matrices are
positive semidefinite. For simple graphs, these matrices are finite for every fixed
k, so flat reflection positivity is a much friendlier notion than general reflection
positivity. Nevertheless, they will turn out to be equivalent under mild conditions
(Proposition 14.60).
We define ( the connection
) rank function of a graph parameter as the rank
r(f, k) = rk M (f, k) as a function of k (again, simply/multilabeled and simple
graph/multigraph versions can be defined). This number is infinite in general, but
it is finite in a surprisingly large number of cases. Those parameters for which it
is finite for all k, which we call parameters of finite rank, are of particular interest,
and will be discussed next.

Exercise 4.4. (a) Show that a multigraph parameter f is multiplicative if and

only if M (f, 0) is positive semidefinite and has rank 0 or 1. (b) Characterize graph
parameters for which M (f, 0) has rank 1.
 4.3. FINITE CONNECTION RANK 45

4.3. Finite connection rank

Connection matrices are infinite, and typically they have infinite rank. How-
ever, surprisingly many multigraph parameters, including some large classes, have
finite connection rank, which makes this finiteness a combinatorially important
property.
Finite connection rank will have an important algorithmic consequence: every
such parameter can be computed efficiently (in polynomial time) for graphs with
bounded treewidth. The exact statement of this fact and the description of the
algorithm will be given in Section 6.5.
As a warm-up, we make a few simple observations about operations on multi-
graph parameters preserving finite connection rank. (These also hold for simple
graph parameters, which can be considered as a special case here.)
Lemma 4.5. Let f and g be graph parameters and c ∈ C. If f and g have finite
connection rank, then so do cf , f + g and f g.
Proof. The first two assertions are trivial. The third one follows from the
observation that every connection matrix M (f g, k) is a submatrix of the Kronecker
product of M (f, k) and M (g, k), and so its rank is at most the product of the ranks
of them. 
Lemma 4.6. Let f and g be graph parameters with finite rank, and suppose that
f − g has finite range. Then max(f, g) and min(f, g) have finite connection rank.
Proof. First, we prove the case when both f and g have finite range. Lemma
4.5 implies that p(f, g) has finite connection rank for every polynomial p in two
variables. Since over a finite range, every function of f and g can be expressed as
a polynomial of them, the assertion follows.
Second, in the general case we use that max(f, g) = f + max(0, g − f ). By
what we just proved, max(0, g − f ) is a parameter with finite rank, and hence so is
max(f, g). The assertion for the minimum follows similarly. 
4.3.1. Many parameters with finite connection rank. In this section,
we describe connection matrices for a variety of multigraph parameters. Our main
concern will be their rank (and sometimes whether they are semidefinite).
Example 4.7 (Nodes and edges). The number of edges e(G) in G is an additive
parameter: e(F1 F2 ) = e(F1 ) + e(F2 ) for two unlabeled multigraphs F1 and F2 . In
fact, this holds for two k-labeled graphs as well, and so M (e, k) is the sum of two
matrices of rank 1. Thus M (e, k) has rank 2, so r(e, k) = 2 for all k ≥ 0. Similarly,
the number of nodes has finite connection rank r(v, k) = 2 for all k. 
Example 4.8 (Non-parallel edges). Let e′ (G) = e(Gsimp ) denote the number of
different (i.e., non-parallel) edges in G. For two k-labeled graphs G1 and G2 , we
have e′ (G1 G2 ) = e′ (G1 ) + e′ (G2 ) − 12 A1 · A1 , where Ai is the adjacency matrix of
the subgraph of Gi induced by the labeled nodes. Hence M (e′ , k) can be ( )written
as the sum of two matrices of rank 1 and one matrix of rank at most k2 . Thus
( )
r(e′ , k) ≤ k2 + 2, and one can check that this is the exact value. 
Example 4.9 (Subgraphs). Let sg(G) = 2e(G) denote the number of spanning
subgraphs of G. Then sg(G1 G2 ) = sg(G1 )sg(G2 ), and so M (sg, k) has rank 1. Thus
r(sg, k) = 1 for all k. These matrices are trivially positive semidefinite. 
46 4. GRAPH PARAMETERS AND CONNECTION MATRICES

′
Example 4.10 (Simple subgraphs). Let sg′ (G) = 2e (G) denote the number of
simple subgraphs of G. Then
1
sg′ (G1 G2 ) = sg′ (G1 )sg′ (G2 ) ′ .
sg (G1 ∩ G2 )
The first two factors don’t change the rank, and the rows of the matrix given by
the third factor are determined by the edges induced by the labeled nodes, so the
k k
corresponding matrix has at most 2(2) different rows. Hence r(sg′ , k) ≤ 2(2) . Again
one can check that this is the exact value. 
Next we look at some of the less trivial but still common graph parameters,
which make more complicated examples.
Example 4.11 (Stability number). The maximum size α(G) of a stable set of
nodes is additive, and has finite connection rank (Godlin, Kotek and Makowski
[2008]). This is more difficult to prove. First, we split the rows of the matrix
k
M (α, k) into 2(2) classes, according to the subgraph Hi of Fi induced by the labeled
nodes. This splits the matrix M (α, k) into 2k(k−1) submatrices, and it suffices to
show that each of these has finite rank. So let us fix H1 and H2 . Let I denote the
set of stable sets of nodes in H1 ∪ H2 , and let Fi′ = Fi \ [k] and FiS = Fi′ \ NFi (S).
For two k-labeled graphs F1 and F2 with Fi [k] = Hi , we have α(F1 F2 ) =
maxS∈I αS (F1 F2 ), where
( (
αS (F1 F2 ) = |S| + α F1S ) + α F2S )
is the maximum( size of
) a stable set in F1 F2 intersecting [k] in S. The rank of
the matrix αS (F1 F2 ) is at most 3. Unfortunately, we cannot apply Lemma
4.6 directly, since αS (F1 F2 ) is not bounded. But we can use that α(F1 F2 ) ≥
α∅ (F1 F2 ) = α(F1′ ) + α(F2′ ), and hence those sets S for which α(F1S ) < α(F1′ ) − k or
α(F2S ) < α(F2′ ) − k play no role in the maximum. In other words, we can replace
αS by
{ } { }
αS′ (F1 F2 ) = |S| + max α(F1S ), α(F1′ ) − k + max α(F2S ), α(F2′ ) − k ,
( )
and still have that α(F1 F2 ) = maxS αS′ (F1 F2 ). The matrices αS′ (F1 F2 ) have
rank at most 3, and for different sets S the corresponding entries differ by at most
3k, so the same argument as in the proof of Lemma 4.6 implies that α has finite
connection rank. 
Example 4.12 (Node cover number). The minimum number of nodes covering
all edges, τ (G) = v(G) − α(G), has finite connection rank as well, since every
connection matrix of τ is the difference of the corresponding connection matrices
of the parameters v and α, which both have finite rank. 
Example 4.13 (Number of stable sets). Let stab(G) denote the number of
stable sets in G. This parameter is multiplicative, and has finite connection rank:
this can be verified easily by distinguishing stable sets according to their intersection
with the set of labeled nodes. 
Example 4.14 (Number of perfect matchings). Let pm(G) denote the number
of perfect matchings in the multigraph G. It is trivial that pm is multiplicative.
Let G be a k-labeled multigraph, let X ⊆ [k], and let pm(G, X) denote the
number of matchings in G that match all the unlabeled nodes and the nodes with
 4.3. FINITE CONNECTION RANK 47

label in X, but not any of the other labeled nodes. Then we have for any two
k-labeled multigraphs G1 and G2
∑
pm(G1 G2 ) = pm(G1 , X)pm(G2 , [k] \ X).
X⊆[k]

Hence the matrix M (pm, k) can be written as the sum of 2k matrices of rank 1,
and its rank is at most 2k (it is not hard to see that in fact equality holds).
If we consider the matching number as a simple graph parameter (in terms of
multigraphs, this means that we don’t care which edge in a parallel class matches
a given pair of nodes), then the above argument has to be modified, to arrive at a
similar conclusion. The details of this are left to the reader as an exercise. 
Example 4.15 (Number of Hamilton cycles). Let ham(G) be the number of
Hamilton cycles in G. For two k-labeled multigraphs G1 and G2 , every Hamilton
cycle H in G1 G2 defines a cyclic ordering (i1 , . . . , ik ) of the nodes in [k], and for
any two consecutive nodes ir and ir+1 , it defines an index jr ∈ [2] which tells us
whether the arc of H between ir and ir+1 uses G1 or G2 . Let us call the cyclic
ordering (i1 , . . . , ik ), together with the indices (j1 , . . . , jk ), the trace of H on the
labeled nodes. (If you are living in the set of labeled nodes, and cannot see farther
than a small neighborhood of the nodes, then the trace is all that you can see from
a Hamilton cycle.)
Given a possible trace T = (i1 , . . . , ik ; j1 , . . . , jk ), we denote by ham(Gj ; T ) the
number of systems of edge-disjoint paths in Gj which connect ir and ir+1 for all r
with jr = j, and which cover all unlabeled nodes in Gj . Then
∑
ham(G1 G2 ) = ham(G1 ; T )ham(G2 ; T ),
T

showing that the rank of M (ham, k) is bounded by the number of possible traces
(which is 2k−1 (k − 1)! by standard combinatorial calculation). 
Example 4.16 (Chromatic polynomial). Every substitution into the chromatic
polynomial chr(G, x) gives a graph parameter (see Appendix A.2). If we substitute
a nonnegative integer q for the variable x, we get the number of q-colorings, which
is a special case of homomorphism functions (to be discussed in the next Chapter).
What about evaluations at other values? The rank of the connection matrices
for the general case was determined by Freedman, Lovász and Welsh (see Lovász
[2006a]).
Let Bk denote the number of partitions of a k-set (the k-th Bell number), and
let Bk,q denote the number of its partitions into at most q parts.
Proposition 4.17. For every fixed x, chr(., x) is a multiplicative graph parameter.
For every k ≥ 0,
{
Bk,x if x is a nonnegative integer,
r(chr(., x), k) =
Bk otherwise.
Furthermore, M (chr(., x), k) is positive semidefinite if and only if x is a nonnegative
integer or x ≥ k − 1.
Proof. We prove that the right hand side is an upper bound even for the
rank of the multi-connection matrix, and a lower bound for the rank of the simple
48 4. GRAPH PARAMETERS AND CONNECTION MATRICES

connection matrix. The case x = 0 is trivial, so suppose that x ̸= 0. Using the

deletion-contraction relation
(4.5) chr(F, x) = chr(F − e, x) − chr(F/e, x),
we see that for every k-multilabeled multigraph F with at least one edge, the row
of M (chr(., x), k) corresponding to F is the difference of two earlier rows (where we
order the rows so that the number of edges is non-decreasing). So the rank of the
whole matrix is the same as the rank of the submatrix formed by k-multilabeled
graphs with no edges. Since deleting an unlabeled isolated node just divides the
row by x, we may assume that all nodes are labeled. So the rows and columns
of the remaining matrix M ′ correspond to partitions of the label set [k]. In the
intersection of the row indexed by P ∈ Π(k) and column indexed by Q ∈ Π(k), we
find the k-multilabeled graph corresponding to P ∨ Q. The chromatic polynomial
for this graph is x|P∨Q| . Let D denote the diagonal matrix in which the entry in row
and column P is (x)|P| , then by identities (A.3) and (A.1), we have M ′ = ZDZ T .
This implies that the rank of M ′ is the same as the rank of D, and it is positive
semidefinite if and only if D is positive semidefinite. If x is not a nonnegative
integer, then D has full rank. The same conclusion holds for positive integers
x ≥ k. If x < k is a nonnegative integer, then the number of nonzero diagonal
entries in D is Bk,x . Finally, D is positive semidefinite if and only if (x)j ≥ 0
for all 1 ≤ j ≤ k, which is clearly equivalent to x being a nonnegative integer or
x ≥ k − 1. 

Note the nontrivial fact that the rank is always finite. If x is a nonnegative
integer, then the connection rank is bounded by xk , but otherwise, as a function of
k, it grows faster than ck for every c. 
Example 4.18 (Tutte polynomial). The cluster expansion version cep(G; u, v)
of the Tutte polynomial generalizes the chromatic polynomial (see again Appendix
A.2), and it behaves similarly. It is not hard to show that for v ̸= 0,
{
Bk,u if u is a nonnegative integer,
r(cep, k) =
Bk otherwise
(the case v = 0 is trivial). Furthermore, cep(G; u, v) is reflection positive if and only
if u is a nonnegative integer. For other versions of the Tutte polynomial (e.g., tut)
similar conclusions hold, since they are related to cep by scaling and substitution
in the variables (except when the expressions we scale with are 0). 
Example 4.19 (Number of spanning trees). The number of spanning trees
tree(G) of a graph G is obtained by substitution into the Tutte polynomial tut with
x = y = 1. Since u = (x − 1)(y − 1) = 0, this falls under the exception at the end of
the last example. Nevertheless, the arguments can be adjusted appropriately, and
we get that r(tree, k) = Bk . 
We conclude with a couple of examples of parameters whose connection matri-
ces have infinite rank, but they are still “interesting”.
Example 4.20 (Maximum clique). The size of a maximum clique, ω(G), is
maxing. It does not have finite connection rank. In fact, consider the connection
matrix M (ω, 0), and its submatrix M whose rows and columns are indexed by
 4.3. FINITE CONNECTION RANK 49

cliques K1 , K2 , . . . . This looks like

 
1 2 3 4 ...
2 2 3 4 . . .
 
 . . .
M = 3 3 3 4 
4 4 4 4 . . .
 
.. .. .. .. ..
. . . . .
and has infinite rank. Similar argument shows that no unbounded maxing graph
parameter has finite connection rank (Exercise 4.34). In particular, the chromatic
number has infinite connection rank. 
Example 4.21 (Eulerian orientations). For an undirected multigraph F , let
−
→
eul(F ) denote the number of eulerian orientations of F (i.e., orientations in which
every node has the same outdegree as indegree; for simplicity, let’s exclude loops).
−
→
By Euler’s theorem, eul(F ) = 0 if and only if F has a node with odd degree. It is
−
→
clear that the parameter eul is multiplicative.
−→
Let us define eul(G; a1 . . . ak ) as the number of orientations of a k-labeled graph
G such that the unlabeled nodes have equal outdegree and indegree, while for a
node i ∈ [k], the difference between its indegree and outdegree is ai . Then we have:
−→ ∑− → −
→
eul(G1 G2 ) = eul(G1 ; a1 . . . ak )eul(G2 ; −a1 · · · − ak )
a
∑−
→ −
→
(4.6) = eul(G1 ; a1 . . . ak )eul(G2 ; a1 . . . ak ).
a
(The sum is finite for every G1 and G2 , but the number of nonzero terms is not
−
→
bounded). This implies that the connection matrices M (eul, k) are positive semi-
definite, but it does not follow that they have finite rank; and in fact, they have
infinite rank for k ≥ 2 (see Exercise 4.33). 
4.3.2. Minor-closed graph properties. We have seen many examples of
graph parameters with finite connection rank. In the next sections, we will describe
some very general classes of such parameters. A challenging problem is to determine
all such graph parameters.
Theorem 4.22. Every minor-closed multigraph property has finite connection
rank.
Proof. We use the very deep theorem of Robertson and Seymour [2004] that
such a property can be characterized by a finite number of excluded minors. Let
P be the multigraph property that G does not contain any of H1 , . . . , Hm as a
minor, and let Pi be the property that G does not contain Hi as a minor. Then
1P = 1P1 . . . 1Pm , and so it suffices to prove that Pi has finite connection rank. In
other words, we may assume that P is the property of not containing a given graph
H as a minor.
Fix k, and consider all graphs H ′ on at most v(H) + k − 1 nodes that can be
contracted to H. In H ′ , select a subset of at most k nodes in all possible ways,
and label them by different numbers from [k]. Finally, 2-color the edges of H ′ red
and blue so that only labeled nodes can be incident with edges of both colors. Call
every partially labeled 2-colored graph obtained this way a pre-minor. It is clear
that the number of pre-minors is finite.
50 4. GRAPH PARAMETERS AND CONNECTION MATRICES

Let G and G′ be two partially labeled edge-colored graphs. We say that G′ is

a minor of G, if G′ can be obtained from G by deleting edges and/or nodes and
contracting edges so that no edge with both endnodes labeled is ever contracted.
(The remaining edges keep their colors and the remaining labeled nodes keep their
labels.)
Consider a product G1 G2 of two k-labeled graphs, and color the edges of G1
red and the edges of G2 blue. It is easy to see that G1 G2 , as an unlabeled uncolored
graph, contains H as a minor if and only if it contains, as a labeled edge-colored
graph, at least one pre-minor as a minor.
For a pre-minor H ′ , let H1′ be the subgraph of H ′ formed by all red edges,
their endpoints, and the labeled nodes. We define H2′ similarly using the blue
edges. Then G1 G2 contains H ′ as a minor if and only if G1 contains H1′ and G2
contains H2′ as a minor.
Let G1 and G′1 be two k-labeled graphs and suppose that for every pre-minor
H , H1′ is a minor of G1 if and only if it is a minor of G′1 . Then the rows of
′

M (1P , k) indexed by G1 and G′1 are equal. This means that M (1P , k) has only a
finite number of different rows, and hence its rank is finite. 
Corollary 4.23. Every nonnegative integer valued bounded minor-monotone multi-
graph parameter has finite connection rank.
Proof. Let f be such a parameter, and assume that f ≤ K. Then
f (G) = 1(f (G) ≤ 1) + 1(f (G) ≤ 2) + · · · + 1(f (G) ≤ K).
Since the graph property that f (.) ≤ i is minor-closed, each parameter 1(f (G) ≤ i)
has finite connection rank by Theorem 4.22, and hence so does f . 
4.3.3. Monadic second order formulas. To describe a very rich class of
graph properties with finite connection rank (at least for looped-simple graphs),
we consider properties defined by certain logical formulas. A first order formula
in graph theory is composed of primitives “x = y” and “x ∼ y”, using logical
operations “∧” (AND), “∨” (OR) and “¬” (NEGATION), and logical quantifiers
“∀” and “∃”. Every such formula, properly composed, with all variables quantified,
defines a property of looped simple graphs, if we interpret the quantified variables
as nodes, and the relation x ∼ y as x and y being adjacent. For example, the
property of being a 2-regular loopless graph can be expressed as
(∀x)(∀y)(x = y ⇒ x ̸∼ y)
( )
∧ (∀x)(∃y)(∃z) y ̸= z ∧ x ∼ y ∧ x ∼ z ∧ (∀u)(x ∼ u ⇒ (u = y ∨ u = z))
(to facilitate reading these formulas, we will use some standard conventions like
writing A ⇒ B instead of ¬A ∨ B and x ̸= y instead of ¬(x = y)).
First order formulas can define rather simple graph properties only, but we get
a real jump in generality if we allow quantifying over subsets of nodes and edges.
A monadic second order formula has three types of variables, which we distinguish
using different fonts. Lower case letters denote nodes, upper case letters denote
subsets of nodes, and upper case boldface letters denote subsets of the edges. The
primitives then also include x ∈ X and xy ∈ Y. We call the formula node-monadic
if quantifying over subsets of edges is not allowed.
This way we get a quite powerful language to express graph formulas, as the
following examples show (see also the exercises at the end of the section).
 4.3. FINITE CONNECTION RANK 51

Example 4.24. The property of being bipartite (2-colorable) can be expressed as

( )
(∃X)(∃Y ) (∀x)(x ∈ X ∨ x ∈ Y ) ∧ ¬(x ∈ X ∧ x ∈ Y )
( )
∧ (∀x)(∀y) x ∼ y ⇒ ((x ∈ X ∧ y ∈ Y ) ∨ (x ∈ Y ∧ y ∈ X)) .
Colorability by any given number of colors can be expressed similarly. 
Example 4.25. The existence of a perfect matching can be expressed as
( ( )
(∃M) (∀x)(∀y) (xy ∈ M) ⇒ (x ∼ y) ∧ (∀x)(∃y)(xy ∈ M)
( ))
∧ (∀x)(∀y)(∀z) ((xy ∈ M) ∧ (xz ∈ M) ⇒ z = y) .
The existence of a Hamilton cycle can also be expressed (see Exercise 4.38). 
Example 4.26. Planarity of a graph can be expressed by a node-monadic second-
order formula. First we construct a formula expressing the property of a graph
G that it contains a subdivision of K5 . One way to do so is to look for 5 nodes
v1 , . . . , v5 and 10 subsets of nodes P12 , P13 , . . . , P45 such that every Pij induces a
connected subgraph containing vi and vj , and the sets Pij \ {vi , vj } are disjoint.
For a given i < j, the required properties of Pij can be expressed by
(( )
Ψij : (vi ∈ Pij ) ∧ (vj ∈ Pij ) ∧ (∀S) (vi ∈ S) ∧ (vj ∈ / S)
( ))
⇒ (∃u)(∃w) (u ∈ Pij ) ∧ (w ∈ Pij ) ∧ (u ∈ S) ∧ (w ∈ / S) ∧ (u ∼ w) .
For every pair of pairs {{i, j}, {k, l}} with i < j, k < l, and {i, j} ∩ {k, l} = ∅ we
write ( )
Φi,j,k,l : (∀u) (u ∈
/ Pij ) ∨ (u ∈
/ Pkl ) .
and for every pair of pairs {{i, j}, {k, l}} with i < j, k < l, and {i, j} ∩ {k, l} = {m}
we write (( ) )
Φi,j,k,l : (∀u) (u ∈ Pij ) ∧ (u ∈ Pkl ) ⇒ (u = vm ) .
Then the formula
(∧ ∧ )
Θ1 = (∃v1 ) . . . (∃v5 )(∃P12 )(∃P13 ) . . . (∃P45 ) Ψij ∧ Φi,j,k,l
i<j i<j,k<l
{i,j}̸={k,l}

expresses that the graph contains a subdivision of K5 . We can construct similarly

a formula Θ2 expressing that the graph contains a subdivision of K3,3 . Then
¬Θ1 ∧ ¬Θ2 means that the graph is planar.
In a similar way, every minor-monotone property can be expressed by a monadic
second order formula, at least for looped-simple graphs (Exercise 4.39). 
Godlin, Kotek and Makowski [2008, 2009] proved the following very general
sufficient condition for a graph property P to have finite connection rank.
Theorem 4.27. Every looped-simple graph property definable by a monadic second
order formula has finite connection rank.
In fact, they prove a more general result about monadic second order definable
graph polynomials, which we do not introduce in this book.
In order to prove Theorem 4.27, we have to look at structures that are more
general than simple graphs. A gaudy graph of type (a, b, c) (a, b, c ∈ Z, a ≥ 0,
b, c ≥ 1) is a looped-simple graph G = (V, E) that is a-multilabeled, node-colored
52 4. GRAPH PARAMETERS AND CONNECTION MATRICES

with b colors, and edge-colored with c colors. In other words, a gaudy graph is a
5-tuple (V, E, α, β, γ), where (V, E) is an underlying multigraph and α : [a] → V ,
β : V → [b] and γ : E → [c] are maps. The labels of a gaudy graph will play a
role different from the role labels play in most of our discussions, and we will call
them badges.
Isomorphism of two gaudy graphs of the same type is defined in the natural
way, as an isomorphism between the underlying graphs that preserves the badge as-
signment and coloring maps. A gaudy graph parameter of type (a, b, c) is a complex
valued function defined on gaudy graphs of type (a, b, c), invariant under isomor-
phism.
To extend the notion of connection matrices to gaudy graphs is a bit tedious
but necessary. Let f be a gaudy graph parameter of type (a, b, c). We will define
infinitely many connection matrices (just like in the case of ordinary graph parame-
ters), but these will not be indexed by just a number k, as for ordinary graphs, but
we will have a connection matrix M (f ; H, a1 , a2 ) for every gaudy graph H of type
(a0 , b, c), where a0 +a1 +a2 = a. Its rows and columns are indexed by gaudy graphs
G of type (a0 + a1 , b, c) and (a0 + a2 , b, c), respectively, with a fixed embedding of H
into G which preserves badges, node colors and edge colors. The product G1 G2 of
two such structures is obtained by taking their disjoint union and then identifying
the copies of H in them. The badges and colors are defined in G1 G2 in a natural
way. The entry of M (f ; H, a1 , a2 ) in row G1 and column G2 is f (G1 G2 ).
A gaudy graph parameter f of type (a, b, c) has finite connection rank if all
connection matrices M (f ; H, a1 , a2 ) have finite rank.
The two observations in Lemmas 4.5 and 4.6 and their proofs remain valid
for gaudy graphs. We now formulate more involved operations, manipulating the
badges and colors. Let f be a gaudy graph parameter of type (a + 1, b, c), and
define the gaudy graph parameter f ∗ of type (a, b, c) by
(4.7) f ∗ (G; α, β, γ) = max f (G; α′ , β, γ).
α′ :[a+1]→V (G)
α′ |[a] =α

Let f be a gaudy graph parameter of type (a, 2b, c), and let φ : [2b] → [b]. Define
the gaudy graph parameter f ∗∗ of type (a, b, c) by
(4.8) f ∗∗ (G; α, β, γ) = max f (G; α, β ′ , γ).
β ′ :φ◦β ′ =β

Finally, to formulate an analogous construction for edge-colorings, let f be a gaudy

graph parameter of type (a, b, 2c), and let ψ : [2c] → [c]. Define gaudy graph
parameter f ∗∗∗ of type (a, b, c) by
(4.9) f ∗∗∗ (G; α, β, γ) = max f (G; α, β, γ ′ ).
γ ′ :ψ◦γ ′ =γ

The following lemma asserts that these operations preserve finite rank:
Lemma 4.28. If f is a gaudy graph parameter with finite connection rank and
finite range, then f ∗ , f ∗∗ and f ∗∗∗ have finite connection rank.
Proof. We describe the proof for f ∗ ; the proof for f ∗∗ and f ∗∗∗ is similar.
Consider a connection matrix M = M (f ∗ ; H, a1 , a2 ), where H is a gaudy graph
of type (a0 , b, c), and a0 + a1 + a2 = a − 1. Consider a general entry of M , defined
by a row index G1 and column index G2 , where G1 G2 = (G, α, β, γ):
MG1 ,G2 = f ∗ (G1 G2 ) = ′max f (G; α′ , β, γ).
α |[a]=α
 4.3. FINITE CONNECTION RANK 53

We split this maximum into v(H)+2 parts, according to α′ (a) = v ∈ V (H), α′ (a) ∈
V (G1 )\V (H) and α′ (a) ∈ V (G2 )\V (H). This defines M = M (f ∗ ; H, a1 , a2 ) as the
maximum of v(H) + 2 matrices Av (v ∈ V (H)), B b and C.
b By the same argument
as in the proof of Lemma 4.6, it suffices to prove that these matrices have finite
rank. We show that these matrices can be expressed by the connection matrices of
f in a way that finite rank is preserved.
First, each of the matrices Av (v ∈ V (H)) is a connection matrix for f itself
(with a new badge added to v), and so it has finite rank.
Second, each entry B bG G is obtained as the maximum of the entries NG′ G of
1 2 1 2
the matrix N = M (f ; H, a1 + 1, a2 ), where G′1 is obtained from G1 by attaching
badge a to one additional node. Let L(G1 ) denote the set of these rows. This is a
finite set but there is no common bound on its size. However, we may notice that
for a fixed G1 , the columns have a basis with at most rk(N ) elements, and if two
rows agree on these basis columns then they agree everywhere. The range of f is
finite, say it consists of r elements; hence there are at most K = rrk(N ) different
rows in N . Let L1 (G1 ) ⊆ L(G1 ) be a maximal set of different rows.
Now we create K matrices B b1 , . . . , B
bK , which are of the same shape as B,
b the
b
row of Bi corresponding to G1 is the i-th row in L1 (G1 ) (we repeat the last row
if we run out of rows). These are all submatrices of N , so they have rank at most
rk(N ). Furthermore, B b is obtained by taking their maximum entry-by-entry, and
hence it has finite rank by the argument used in the proof of Lemma 4.6.
It follows by a similar argument that C has finite rank. 

Proof of Theorem 4.27. We prove the theorem more generally for gaudy
graph properties definable by a monadic second order formula F, by induction on
the number of quantifiers. If there are no quantifiers in F, then the assertion is
easy. Else, F can be written in one of the following forms:
(∀x)Fx , (∃x)Fx , (∀S)FS , (∃S)FS , (∀Ψ)FΨ , (∃Ψ)FΨ ,
Here Fx is a monadic second order formula in the language of gaudy graphs with
an additional badge x; FS is obtained from F using twice as many node colors
{1, . . . , 2b}, where color i means “colored i and not in S”, and color i + b means
“colored i and in S”. FΨ is defined analogously, splitting edge colors according
to containment in the edge subset Ψ. Replacing the property by its negation if
necessary, we can forget about the three versions starting with ∀ (see Exercise
4.37).
First, we consider the case when F = (∃x)Fx . The indicator function 1Fx of
the gaudy graph property defined by Fx has finite connection rank by the induction
hypothesis. For the indicator functions, we have 1F = 1∗Fx , so Lemma 4.28 implies
that 1F has finite connection rank. The two remaining cases follow similarly. 

Remark 4.29. Further variants of the problem of characterizing graph parameters

with finite connection rank ask for the characterization of graph parameters with
exponentially bounded connection rank, or polynomially bounded connection rank.
Not much is known in this direction. For the analogous “edge-connection” problem,
see Schrijver’s Theorem 23.6.

Exercise 4.30. Prove that for every isolate-indifferent graph parameter f , the
connection rank r(f, k) is a monotone non-decreasing function of k.
54 4. GRAPH PARAMETERS AND CONNECTION MATRICES

Exercise 4.31. Let χ(G) denote the minimum number of cliques in the graph G
covering all nodes (the chromatic number of the complement). Prove that χ has
finite connection rank.
Exercise 4.32. Prove that the graph parameters 2α(G) and 2τ (G) are multiplica-
tive and have finite connection rank.
−→
Exercise 4.33. Prove that the multigraph parameter eul(G) has infinite connec-
−→
tion rank for k ≥ 2. Hint: for k = 2, consider the submatrix of M (eul, k) formed
by rows and columns indexed by (2i −[(1)-bonds, )]n i = 1, 2, . . . , n. Verify that this
2i−1
can be written as 2AAT , where A = i+j−1 i,j=1
is a lower triangular matrix
with 1-s in the diagonal.
Exercise 4.34. (a) Prove that no unbounded maxing graph parameter has finite
connection rank. (b) Show by an example that a bounded maxing parameter can
have finite connection rank.
Exercise 4.35. Prove that if P is a minor-closed property, then the property P
defined by P(G) = P(G) has finite connection rank.
Exercise 4.36. The Hadwiger number had(G) if a graph G is the largest n for
which Kn is a minor of G. Prove that the parameter had is minor-monotone, but
its connection rank is not finite.
Exercise 4.37. Prove that if a graph property has finite connection rank, then
so does its negation.
Exercise 4.38. Show that the following graph properties can be expressed by
monadic second order formulas: (a) G is connected; (b) G is a tree; (c) G is 3-
degenerate (i.e., its nodes can be ordered so that every node is connected to no
more than 3 earlier nodes); (d) G is Hamiltonian.
Exercise 4.39. Show that every minor-monotone graph property can be ex-
pressed by a node-monadic second-order formula.
Exercise 4.40. The property that the number of nodes is even cannot be ex-
pressed by a monadic second order formula.
Exercise 4.41. Prove that (a) the property that G is Hamiltonian has finite
connection rank; (b) the property that the complement G is Hamiltonian does
not have finite connection rank. (c) The property that G is Hamiltonian can not
be expressed by a node-monadic second order formula.
Exercise 4.42. Show that the following graph properties cannot be expressed by
monadic second order formulas: (a) G has a nontrivial automorphism; (b) G has
a node-transitive automorphism group.
Exercise 4.43. Let f be an integer valued graph parameter with finite connection
rank, and let m be a positive integer. Prove that f mod m has finite connection
rank.
Exercise 4.44. Let f be a bounded graph parameter with ( )finite connection rank.
Let g : R → R be an arbitrary function. Prove that g f (.) has finite connection
rank.
Exercise 4.45. Let f (G; x) be a graph parameter whose values are analytic
functions of a variable x (defined for |x| ≤ 1). Suppose that for every real x,
f (., x) has finite connection rank.
(a) Prove that the k-th connection rank of f (.; x) is uniformly bounded in x.
d
(b) Prove that dx f (., x) has finite connection rank for all x.
Exercise 4.46. From a gaudy graph parameter f , define the new parameters f ′ ,
f ′′ and f ′′′ by replacing the “max” by “sum” in (4.7), (4.8) and (4.9). Prove that
if f has finite connection rank, then so do f ′ , f ′′ and f ′′′ (no finiteness assumption
for the range is needed).
 CHAPTER 5

Graph homomorphisms

5.1. Existence of homomorphisms

Let G and H be two simple graphs. An adjacency-preserving map from V (G)
to V (H) is called a homomorphism. We write G → H if there is a homomorphism
from G into H. The existence of homomorphisms between graphs is an important
and highly non-trivial question. For example, Kn → G means that G contains a
clique with n nodes; G → Kn means that G is n-colorable.
If we fix G, then G → H can be decided in time that is polynomial in v(H), by
checking all possible maps V (G) → V (H). What happens if we fix H? If H has no
edge, then G → H if and only if G has no edge. If H is a bipartite graph with at
least one edge, then H → K2 and K2 → H. Hence G → H if and only if G → K2 ,
which is equivalent to saying that G is bipartite.
Unfortunately, these are the only cases when the existence of a homomorphism
into a fixed graph H is polynomial time decidable (at least if P ̸= N P ). It was
proved by Hell and Nešetřil [1990] that for a (fixed) nonbipartite graph H, the
problem whether G → H is NP-complete. So among the infinitely many problems
G → H (one for each H), some are solvable in polynomial time and the others are
NP-hard; there is no case whose complexity would be inbetween these extremes.
This is an interesting phenomenon called “dichotomy”.
If we consider directed graphs, then the problem becomes more challenging.
We don’t go into the details in this book; The monograph of Hell and Nešetřil
[2004] provides an in-depth treatment of many questions on the existence of homo-
morphisms; we give a few exercises below to illustrate these questions.

Exercise 5.1. Verify that allowing looped-simple graphs would not give any
interesting new cases of the homomorphism existence problem.
−
→ −
→
Exercise 5.2. Let C n denote the directed n-cycle, P n , the directed path on n
−
→
nodes, and K n , the transitive tournament on n nodes. For any cycle or path in a
digraph G, define its gain as the difference between the numbers of forward and
backward edges (if the cycle or path is traversed in the opposite direction, this
number changes sign). Prove that
−
→
(a) G → C n if and only if the gain of every cycle is a multiple of n;
−
→
(b) G → P n if and only if the gain of every cycle is 0 and the gain of every path
is bounded by n − 1;
−
→ −
→
(c) G → K n if and only if P n+1 ̸→ G.
−
→ −
→ −
→
Exercise 5.3. Prove that G → C n , G → P n and G → K n are polynomial time
decidable.
55
56 5. GRAPH HOMOMORPHISMS

5.2. Homomorphism numbers

5.2.1. Versions of homomorphism numbers. Now we come to one of the
main notions for this book. Unfortunately, we have to discuss the definition for
the different types of graphs introduced in Section 3.2. We could probably give a
single very general definition, but this would be too awkward. Rather, we start
with homomorphism numbers between simple graphs, and talk about the issues for
more complicated types as they arise.
Simple graphs. For two finite simple graphs F and G, let Hom(F, G) denote the
set of homomorphisms of F into G, and let hom(F, G) be their number. Let inj(F, G)
be the number of injective homomorphisms of F into G, and ind(F, G), the number
of embeddings of F into G as an induced subgraph. In other words, ind(F, G)
counts injective homomorphisms that also preserve non-adjacency. Occasionally,
one needs also surj(F, G), the number of surjective homomorphisms from F to G;
here we mean surjective on both the nodes and edges, and refrain from introducing
a notation for counting those which are only (say) node-surjective.
Multigraphs. If we want to extend the definition of hom(F, G) to multigraphs,
then there is no agreement any more what it should mean. Most often we will use
the notion that a homomorphism must specify not only which node goes on which
node, but also which edge goes on which edge. So if i, j ∈ V (F ) are connected by aij
edges, and u, v ∈ V (G) are connected by buv edges, then after specifying that i 7→ u
a
and j 7→ v, we have buvij ways of mapping the i-j edges to the u-v edges. In other
words, a node-and-edge homomorphism F → G is a pair of maps φ : V (F ) → V (G)
and ψ : E(F ) → E(G) such that if e ∈ E(G) connects i and j, then ψ(e) connects
a
φ(i) and φ(j). If buv , aij ∈ {0, 1}, then buvij is 1 unless aij = 1 and buv = 0, and so
for simple graphs this specializes to the earlier definition. Unless otherwise stated,
the number of node-and-edge homomorphisms is what is meant by hom(F, G).
Another possibility is to define a node-homomorphism as a map φ : V (F ) →
V (G) such that the multiplicity of φ(i)φ(j) ∈ E(G) is at least as large as the
multiplicity of ij ∈ E(F ). In the case of simple graphs, this too specializes to the
old notion. These versions will play an important role in Section 17.1.
We can also consider special node-homomorphisms that preserve edge multi-
plicities. In this case the role of multiplicities is reduced to distinguishing edges
with different multiplicities. These homomorphisms generalize “induced” homo-
morphisms between simple graphs that preserve both adjacency and non-adjacency,
and will be called induced homomorphisms
Weighted graphs. The definition of homomorphism numbers can be extended to
the case when G is a weighted looped-simple graph with nodeweights αv (G) and
edgeweights βuv (G). To every map φ : V (F ) → V (G), we assign the weights
∏
(5.1) αφ = αφ(u) (G),
u∈V (F )

and
∏
(5.2) homφ (F, G) = βφ(u)φ(v) (G).
uv∈E(F )

We define
∑
(5.3) hom(F, G) = αφ homφ (F, G),
φ: V (F )→V (G)
 5.2. HOMOMORPHISM NUMBERS 57

and
∑
(5.4) inj(F, G) = αφ homφ (F, G).
φ: V (F )→V (G)
φ injective

For these definitions to make sense, αv (G) and βuv (G) can be from any com-
mutative ring; we will, however, never need any field other than R and C, and most
of the time αv (G) will be positive and βuv (G) real, and often itself positive.
This definition of hom(F, G) makes sense if F is a multigraph and G is a
weighted graph. If G is an (unweighted) multigraph, then we can consider the
weighted simple graph G′ in which each edge is weighted by its multiplicity in G.
Then hom(F, G) = hom(F, G′ ) (in the node-and-edge sense).
One can also define hom(F, G) when both F and G are weighted, provided
these weights satisfy some reasonable conditions. Let us give the formula first. To
every map φ : V (F ) → V (G), we define the weights
∏
(5.5) αφ = αφ(u) (G)αu (F ) .
u∈V (F )

and
∏
(5.6) homφ (F, G) = βφ(u)φ(v) (G)βuv (F )
uv∈E(F )

We then define
∑
(5.7) hom(F, G) = αφ homφ (F, G).
φ: V (F )→V (G)

The exponential βφ(u)φ(v) (G)βuv (F ) may not be well defined. Mostly (and even this
is not very often), we will need this definition when the nodeweights and edgeweights
of F are nonnegative integers; then (with the usual convention that 00 = 1) the
definition is meaningful. Another case when the homomorphism number is well
defined is when all the edgeweights in G are positive.
Note that in the case when F is an unweighted multigraph, we can replace it
with a weighted graph on the same set of nodes where the nodeweights are 1 and
the edgeweights are equal to the corresponding multiplicities in F . This does not
change the homomorphism numbers hom(F, G).
Signed graphs. There is a convenient way to treat conditions on preservation of
edges and also preservation of non-edges together. Let F and G be simple graphs,
where F = (V, E+ , E− ) is signed. We define hom(F, G) as the number of maps
V (F ) → V (G) where edges in E+ must be mapped onto adjacent pairs, and edges
in E− must be mapped onto nonadjacent pairs. The quantity inj(F, G) is defined
analogously. If F is an unsigned simple graph, we construct the signed graph Fb
from the complete graph on V (F ) by signing the edges of F positive and the edges
not in F negative. Then for every simple graph G,

(5.8) inj(Fb, G) = ind(F, G).

The definition of homomorphism numbers from signed graphs into sim-
ple graphs can be extended to homomorphisms into weighted graphs H. For
58 5. GRAPH HOMOMORPHISMS

φ : V (F ) → V (H), we define αφ by (5.1) as before,

∏ ∏ ( )
(5.9) homφ (F, H) = βφ(u)φ(v) (H) 1 − βφ(u)φ(v) (H) ,
uv∈E+ uv∈E−

and then hom(F, H) by (5.3) as before.

Partially labeled graphs. Let F and G be a simple graph, S ⊆ V (F ), and
φ : S → V (G). We denote by homφ (F, G) the number of homomorphisms F → G
that extend the mapping φ. Most of the time we use this notation when F is
[k]-labeled, and S = [k] is the set of labeled nodes. If φ(i) = vi (i ∈ [k]), then we
denote homφ also by homv1 ...vk .
Let H be a weighted graph, and φ : S → V (H) for some S ⊆ V (F ). Then
homφ (F, H) is already defined when S = ∅ (which is just hom(F, H)), and when
S = V (G) (by (5.2)). Extension to the general case is rather natural (but note that
nodeweights are used for the unlabeled nodes only):
∑ ∏
(5.10) homφ (F, H) = αψ(j) (H)homψ (F, H).
ψ:V (F )→V (H) j∈V (F )\S
ψ⊇φ

Extension of this formula to the case when F is itself weighted or signed is left to
the reader.

5.2.2. Homomorphism densities. We often normalize these homomor-

phism numbers, to get homomorphism densities. Setting n = v(G) and k = v(F ),
we define
hom(F, G)
(5.11) t(F, G) = ,
nk
which is the probability that a random map of V (F ) → V (G) is a homomorphism.
We define similarly
inj(F, G)
(5.12) tinj (F, G) =
(n)k
(the probability that a random injection V (F ) → V (G) is a homomorphism), and
ind(F, G)
(5.13) tind (F, G) =
(n)k
(the probability that a random injection V (F ) → V (G) preserves both adjacency
and non-adjacency).
For a weighted graph H, we define
∑
αH = αv (H),
v∈V (H)

and
hom(F, H)
t(F, H) = v(F )
.
αH
Note that t(F, H) = hom(F, H 0 ), where H 0 is obtained from H by dividing all
node weights by αH , so that αH 0 = 1. The nodeweights in H 0 form a probability
distribution, and t(F, H) = hom(F, H 0 ) is the expectation of homφ (F, H), where φ
is the random map V (F ) → V (H) in which the image of each v ∈ V (F ) is chosen
independently from the distribution α(H0 ).
 5.2. HOMOMORPHISM NUMBERS 59

Bounded degree graphs. The above definition of homomorphism densities works

well if the graphs are dense, but they are more-or-less meaningless for sparse graphs.
Homomorphism numbers are of course defined in the same way independently of
the densities of the graphs, but to introduce meaningful homomorphism densities
in the case of bounded degree graphs we must normalize them differently. The best
analogue of the dense homomorphism density t(F, G) is the number

hom(F, G)
(5.14) t∗ (F, G) = ,
v(G)

which we consider for connected graphs F . We call this the homomorphism fre-
quency of F in G, to distinguish it from the homomorphism densities that are used
in the dense case.
We can interpret the homomorphism frequencies as follows. Let us label any
node of F by 1, to get a 1-labeled graph F1 . For v ∈ V (G), the quantity homv (F1 , G)
denotes the number of homomorphisms φ of F1 into G with φ(1) = v. Now we select
a uniform random node v of G. Then t∗ (F, G) is the expectation of homv (F1 , G).
We can interpret the injective and induced homomorphism frequencies

inj(F, G) ind(F, G)
t∗inj (F, G) = , t∗ind (F, G) =
v(G) v(G)

similarly.
For general (not necessarily connected) bounded degree graphs, the order of
magnitude of hom(F, G) (where F is fixed and v(G) tends to infinity) is v(G)c(F ) ,
where c(F ) is the number of connected components of F . But since hom(F, G) is
multiplicative over the connected components of F , we don’t lose any information
if we restrict the definition of t∗ (F, G) to connected graphs F .

Remark 5.4. Normalizing homomorphism densities as above is not the only rea-
sonable choice. For example, if F is a bipartite graph, then hom(F, G) will be
positive for graphs G with at least one edge, and we may be interested in the order
of magnitude of, say, hom(C4 , G), given hom(K2 , G). Hence we might look at quo-
tients log hom(F, G)/ log v(G), or more generally, log hom(F1 , G)/ log hom(F2 , G).
Such quantities were studied by Kopparty and Rossman [2011] and Nešetřil and
Ossona de Mendez [2011]. However, it is fair to say that this interesting area is
largely unexplored.

5.2.3. Relations between homomorphism numbers.

Injective and induced homomorphism numbers. These quantities are closely

related. For two simple graphs F and G, we have
∑
(5.15) inj(F, G) = ind(F ′ , G),
F ′ ⊇F

where F ′ ranges over all simple graphs obtained from F by adding edges, and
∑
(5.16) hom(F, G) = inj(F/P, G),
P
60 5. GRAPH HOMOMORPHISMS

where P ranges over all partitions of V (F ), and F/P is the simple quotient graph.
Conversely, ind can be expressed by inj using inclusion-exclusion:
∑ ′
(5.17) ind(F, G) = (−1)e(F )−e(F ) inj(F ′ , G).
F ′ ⊇F
V (F ′ )=V (F )

We can also express this using the Möbius inverse (recall the definitions (4.1)–(4.3)):
ind(., G) = inj↑ (., G).
The inj function, in turn, can be expressed by hom, by considering the values
inj(F ′ , G) in the equations (5.16) as unknowns and solving the system. To give an
explicit expression, we use the Möbius inverse of the partition lattice:
∑
(5.18) inj(F, G) = µP hom(F/P, G),
P

where P ranges over all partitions of V (F ), and µP is defined by (A.2). In other

words, inj(., G) = hom⇓ (., G).
Injective and induced homomorphism densities. We have
∑
(5.19) tinj (F, G) = tind (F ′ , G)
F ′ ⊇F

and the inversion formula

∑ ′
(5.20) tind (F, G) = (−1)e(F )−e(F )
tinj (F ′ , G) = t↑inj (F, G).
F ′ ⊇F

For t and tinj the relationship is not so simple, due to the different normalizations
in their definitions, but recalling that mostly we are interested in large graphs G,
the following inequality is usually enough to relate them:
( )
1 v(F )
(5.21) |tinj (F, G) − t(F, G)| ≤ .
v(G) 2
(The proof of this inequality is left to the reader as an exercise.) It follows that (for
large graphs G, when the error in (5.21) is negligible) subgraph sampling provides
the same information as any of the homomorphism densities t, tinj or tind .
Complementation. Taking induced subgraphs commutes with complementation,
which implies
(5.22) ind(F, G) = ind(F , G),
Counting maps into the complement of a graph can be expressed, via inclusion-
exclusion, by numbers of maps into the graph itself. Applying this idea to injective
homomorphisms, we get the following identities for every simple graph F :
∑ ′
(5.23) hom(F, G) = (−1)e(F ) hom(F ′ , G)
F ′ ⊆F
V (F ′ )=V (F )

and
∑ ′
(5.24) inj(F, G) = (−1)e(F ) inj(F ′ , G).
F ′ ⊆F
V (F ′ )=V (F )
 5.2. HOMOMORPHISM NUMBERS 61

Averaging. If we know homomorphism numbers from graphs of a given size, then

we get homomorphism numbers from smaller graphs; also, we get homomorphism
numbers into larger graphs. This idea has many formulations, of which we state a
couple. Let v(F0 ) = k ≤ t ≤ v(G) = n, then
1 ∑ ind(F0 , F )ind(F, G)
(5.25) ind(F0 , G) = (n−k) ,
t−k
aut(F )
F : v(F )=t

Let v(F ) = k ≤ t ≤ v(G) = n, then

1 ∑
(5.26) inj(F, G) = (n−k) inj(F, G[S]).
t−k S∈(V (G)
t )

We get similar expressions for the induced homomorphism numbers ind. We can
also express this in terms of homomorphism densities:
1 ∑
(5.27) tinj (F, G) = (n) tinj (F, G[S]).
t S∈ V (G)
( t )
Graph operations. If F1 and F2 are node-disjoint, then
(5.28) hom(F1 ∪ F2 , G) = hom(F1 , G)hom(F2 , G).
If F is connected and G1 and G2 are node-disjoint, then
(5.29) hom(F, G1 ∪ G2 ) = hom(F, G1 ) + hom(F, G2 ).
About homomorphisms into a product, we have
(5.30) hom(F, G1 × G2 ) = hom(F, G1 )hom(F, G2 ).
All these identities are straightforward to verify.
5.2.4. Homomorphism numbers and sampling. Our basic way of obtain-
ing information about a graph is sampling (Section 1.3.1): subgraph sampling in
the dense case and neighborhood sampling in the bounded degree case. Homo-
morphism densities and frequencies carry the same information as the appropriate
sample distributions. For the dense case, the connection is straightforward:
Proposition 5.5. For two simple graphs F and G, tind (F, G) is the probability that
sampling V (F ) nodes of G (ordered, without repetition), they induce the graph F
(with a fixed labeling of the nodes). 
In the bounded degree case, homomorphism frequencies contain the same in-
formation as the distribution of neighborhood samples, but the proof of this equiv-
alence is a bit trickier. Let us recall that ρG,r is a probability distribution on rooted
r-balls: ρG,r (B) denotes the probability that selecting a uniform random node of
the graph G, its neighborhood of radius r is isomorphic with the ball B.
Proposition 5.6. Let us fix an upper bound D for the degrees of the graphs we
consider.
(a) Each density t∗ (F, G) can be expressed as a linear combination (with co-
efficients independent of G) of the neighborhood sample densities ρG,r with r =
v(F ) − 1.
(b) For every r ≥ 0 there are a finite number of connected simple graphs
F1 , . . . , Fm such that ρG,r can be expressed as a linear combination (with coeffi-
cients independent of G) of the densities t∗ (Fi , G).
62 5. GRAPH HOMOMORPHISMS

Proof. (a) From the interpretation of t∗ (F, G) given above, we see that it can
be obtained as the expectation of the number of homu→v (F, B), where u is any
fixed node of F , and B is a random ball from the neighborhood sample distribution
ρG,r , with center v and radius r = v(F ) − 1. This gives the formula
∑
t∗ (F, G) = ρG,r (B)homu→v (F, B),
B
where the summation extends over all possible r-balls.
(b) To compute the neighborhood sample distributions from the quantities
t∗ (F, G), we first express the quantities t∗inj (F, G) via inclusion-exclusion. By a
similar argument, we can express the induced densities t∗ind (F, G). (Since we are
normalizing by v(G) in all cases, we avoid here the difficulty we had in the dense
case with expressing tinj by t.)
Next, we count copies of F in G where we also prescribe the degree of each
node of F in the whole graph G. To be precise, we consider graphs F together with
maps δ : V (F ) → {0, . . . , D}, and we determine the numbers
ind(F, δ, G)
t∗ind (F, δ, G) = ,
v(G)
where ind(F, δ, G) is the number injections φ : V (F ) → V (G) which embed F in
G as an induced subgraph, so that the degree of φ(v) is δ(v). This is again done
by an inclusion-exclusion argument.
For a ball B of radius r, we have
∑ t∗ (B, δ, G)
ind
ρG,r (B) = ,
aut(B)
δ
where the summation extends over all functions δ which assign the degree in B to
each node of B at distance less than r from the root, and an arbitrary integer from
[D] to those nodes at distance r. This proves that homomorphism densities and
neighborhood sampling are equivalent. 

Exercise 5.7. Find formulas similar to (5.16) and (5.18), relating hom and surj.
Exercise 5.8. Which of the relations (5.15)–(5.30) generalize to weighted graphs?

5.3. What hom functions can express

Homomorphisms of “small” graphs into G are related to sampling, as men-
tioned earlier. There are many other applications of homomorphism numbers. We
start with a number of graph parameters, some unexpected, some not, that can be
expressed as homomorphism numbers into and from the given graph.
Example 5.9 (Walks). A walk in G is a homomorphism of a path into G, so
hom(Pk , G) counts the number of walks with k − 1 steps in G. Note that we can
express this as the sum of the entries of Ak−1 , where A is the adjacency matrix of
G. 
Example 5.10 (Stars and degrees). Homomorphisms from stars into G give the
moments of the degree sequence:
∑
hom(Sk , G) = deg(i)k−1 .
i∈V (G)
 5.3. WHAT HOM FUNCTIONS CAN EXPRESS 63


Example 5.11 (Cycles and spectrum). If Ck denotes the cycle on k nodes, then
hom(Ck , G) is the trace of the k-th power of the adjacency matrix of the graph G.
In other words,
∑n
(5.31) hom(Ck , G) = tr(Ak ) = λki ,
i=1
where λ1 , . . . , λn are the eigenvalues of the adjacency matrix of G. Knowing this
homomorphism number for sufficiently many values of k, the eigenvalues of G re-
covered; eigenvalues with large absolute value are easier to express. For example,
hom(C2k , G)1/(2k) tends to the largest eigenvalue of G as k → ∞. 
Several important graph parameters can be expressed in terms of homomor-
phisms into fixed “small” graphs.
Example 5.12 (Colorings). If Kq denotes the complete graph with q nodes (no
loops), then hom(G, Kq ) is the number of colorings of the graph G with q colors,
satisfying the usual condition that adjacent nodes must get different colors. 
Example 5.13 (Stable sets). Let H = be obtained from K2 by adding a loop
at one of the nodes. Then hom(G, H) is the number stab(G) of stable sets of nodes
in G. 
Example 5.14 (Eulerian property). Recall that a graph is eulerian, if all degrees
are even. For every loopless graph G, let Eul(G) = 1(G is eulerian). This 0-1 valued
graph parameter can be represented as a homomorphism function hom(., H), where
H = (a, B) is a weighted graph with two nodes, given by
( ) ( )
1/2 1 −1
a= , B=
1/2 −1 1
This was first noted by de la Harpe and Jones [1993]. (By Theorem 5.54 it will
follow that this function is reflection positive, and r(Eul, k) ≤ 2k .) 
Example 5.15 (Nowhere-zero flows). The number of nowhere-zero q-flows is
denoted by flo(G, q) (see Appendix A.2). The choice q = 2 gives the special case
in example 5.14 (indicator function of eulerian graphs). The parameter flo can be
described as a homomorphism function, as will be demonstrated in larger generality
in the next example. 
Example 5.16 (S-Flows). Let Γ be a finite abelian group (written additively),
and let S ⊆ Γ be a subset such that −S = S, and let G be a graph. An S-flow is
an assignment of an element f (uv) ∈ S to each∑ edge uv with a specified orientation
such that f (uv) = −f (vu) for each edge, and u∈N (v) f (uv) = 0 for each node v.
Let flo(G; Γ, S) denote the number of S-flows. The special case when Γ = Zq and
S = Γ \ {0} gives the number of nowhere zero q-flows in Example 5.15.
For a fixed Γ and S, the graph parameter flo(G; Γ, S) is defined in terms of
mappings from the edge set; it is therefore surprising that it can be described as a
homomorphism number (which is defined via a function on the node set). Let Γ∗
be the character group of Γ, and let H be the complete looped directed graph on
Γ∗ . Let αχ = 1/|Γ| for each χ ∈ Γ∗ , and let
∑
βχ,χ′ = χ(−s)χ′ (s),
s∈S
64 5. GRAPH HOMOMORPHISMS

for any two characters χ, χ′ ∈ Γ∗ . It follows from our assumption that −S = S

that βχ,χ′ is real and βχ′ ,χ = βχ,χ′ . It takes a straightforward computation to show
(Freedman, Lovász and Schrijver [2007]) that
(5.32) flo(G; Γ, S) = hom(G, H).

Example 5.17 (Tutte polynomial). The Tutte polynomial is not a homomor-
phism function in general; Theorem 5.54 will imply that if q is not a positive integer,
then it is not. But if q is a positive integer, then it has such a representation: Using
the connection between the Tutte polynomial and the Potts model in statistical
physics (Welsh and Merino [2000]), one can prove that
cep(G; q, v) = hom(G, Hq,v )
(here Hq,v is a complete graph on q nodes, with a loop added at each node; every
node has weight 1, the ordinary edges too have weight 1, but the loops have weight
1 + v). 
Example 5.18 (Maximum cut). An important graph parameter is the maximum
cut Maxcut(G), the maximum number of edges between a set S ⊆ V (G) of nodes
and its complement. While finding minimum cuts is perhaps a more natural task at
the first sight, the maximum cut problem comes up when we want to approximate
general graphs by bipartite graphs, when computing ground states in statistical
physics (see Section 2.2), and in many other applications. For our purposes, it will
be more convenient to consider the normalized maximum cut, defined by
Maxcut(G) eG (S, V \ S)
maxcut(G) = = max .
|V |2 S⊆V |V |2
The maximum cut cannot be expressed as a homomorphism number, but the
following easy fact relates maximum cuts and homomorphism numbers. Let H be
the edge-weighted graph on {1, 2} with edgeweights 1 except for the non-loop edge,
which has weight 2. Then we have the trivial inequalities
2Maxcut(G) ≤ hom(G, H) ≤ 2v(G) 2Maxcut(G) ,
which upon taking the logarithm and dividing by v(G)2 become
log hom(G, H) 1
(5.33) maxcut(G) ≤ 2
≤ maxcut(G) + .
v(G) v(G)
So the homomorphism number into this simple 2-node graph determines maxcut(G)
with an additive error that tends to 0 as v(G) → ∞. 
Example 5.19 (Multicuts). A natural extension of the maximum cut problem
involves partitions into q ≥ 1 classes instead of 2. Instead of just counting edges
between different classes, we specify in advance a symmetric matrix B of coefficients
Bij (i, j ∈ [q]). We define the maximum multicut density (with target weights B)
as
1 ∑
cut(G, B) = max Bij eG (Si , Sj ),
v(G)2 i,j
where the maximum is(taken)over all partitions {S1 , . . . , Sq } of V (G). As a special
0 1
case, the matrix B = defines the maximum cut.
1 0
 5.3. WHAT HOM FUNCTIONS CAN EXPRESS 65

Similarly as above, we construct an edge-weighed graph H on [q] with

edgeweights 2Bij . Then
log hom(G, H) log q
(5.34) cut(G, B) ≤ ≤ cut(G, B) + .
v(G)2 v(G)
In the terminology of statistical physics, the negative of the value cut(G, B) would
be called the ground state energy . 
5.3.1. Graph polynomials and homomorphisms. We describe more com-
plex relations between homomorphism functions and some important graph poly-
nomials. These relations will not be used until Chapter 19, but they illustrate
the many subtle connections between homomorphism functions and other graph
theoretic constructions.
Let G be a graph and let I(G) denote the set of stable (independent) subsets
of V (G). ∏
We assign a variable xi to each node i. For every multiset S of the nodes,
let xS = i∈S xi . We define the multivariate stable set polynomial as
∑
stab(G, x) = xS .
S∈I(G)

Note that stab(G, 1, . . . , 1) = stab(G) = hom(G, H), where H is the graph on two
adjacent nodes, with a loop at one of them (all weights being 1).
We have seen that both the chromatic polynomial and the stable set polynomial
can be expressed, at least for special substitutions, as homomorphism numbers. We
show that, conversely, homomorphism numbers between graphs can be expressed in
terms of the stable set polynomial and also in terms of a the chromatic polynomials
of related graphs. Our first lemma expresses the logarithm of the stable set poly-
nomial in terms of the coefficient of the linear term of the chromatic polynomial
(see Appendix A.2 for the relevant definitions). We need a natural extension of
the notion for an induced subgraph: For a multiset Z of nodes, let G[Z] denote
the graph whose nodes are the elements of the multiset Z, and two of them are
adjacent iff the corresponding nodes of G are adjacent.
Lemma 5.20. Assuming that the series below is absolute convergent, we have
∑∞ ∑
(−1)m
(5.35) ln stab(G, x) = cri(G[v1 , . . . , vm ])xv1 . . . xvm .
m!
m=1 v1 ,...,vm ∈V (G)

Proof. Let ∑I+ denote the set of non-empty stable subsets of G. Writing
stab(G, x) = 1 + A∈I+ xA , we get
(5.36)
∞ ( )( ∑
∑ )k ∞ ( )
∑ ∑
y y
stab(G, x)y = 1 + xA =1+ xA1 . . . xAk .
k k
k=1 A∈I+ k=1 A1 ,...,Ak ∈I+

Let Z = {v1 , . . . , vm } denote the union of the Ai as multiset, so that correspond-

ing term in the last sum is xZ . Any choice of S = {A1 , . . . , Ak } that results
in the same multiset Z gives rise to a coloring of the graph G[Z] with exactly
k colors, and vice versa. We have to take into account that this coloring is not
unique: Let r be the number of different nodes vj , and let m1 , . . . mr be the mul-
tiplicities, then we can associate m1 ! . . . mr ! colorings of G[Z] with the same fam-
ily S = {A1 , . . . , Ak } of stable sets. So the coefficient of xZ in the last sum in
66 5. GRAPH HOMOMORPHISMS

(5.36) is chr0 (G[Z], k)/(m1 ! . . . mr !). We get a nicer formula if instead of multi-
sets, we sum over sequences (v1 , . . . , vm ) of nodes. Then our multiset Z is counted
m!/(m1 ! . . . mr !) times, so we have to divide by this, to get that the contribution
of a sequence (v1 , . . . , vm ) is
1
chr0 (G[v1 , . . . vm ], k)xv1 . . . xvm ,
m!
and summing over k, we get that the contribution of (v1 , . . . , vm ) is
∞ ( )
∑ y 1 1
chr0 (G[v1 , . . . vm ], k)xv1 . . . xvm = chr(G[v1 , . . . vm ], y)xv1 . . . xvm .
k m! m!
k=1

For m = 0, we get 1. This implies that

∞
∑ ∑ 1
(5.37) stab(G, x)y = chr(G[v1 , . . . vm ], y)xv1 . . . xvm .
m=0 v1 ,...,vm ∈V (G)
m!

Differentiating (5.37) according to y and substituting y = 0, we get the formula in

the lemma. 

Next, we express hom(G, H) in terms of the intersection graph G of connected

subgraphs of G with at least two nodes. For a weighted graph H, let H denote the
weighted graph on the same node set and with the same nodeweights as H, and
with edgeweights 1 − βij (H). Recall that L(C) denotes the intersection graph of a
family C of sets, and Conn(G) is the set of connected subgraphs of G with at least
one edge.
Lemma 5.21. For a simple graph G = (V, E) and weighted graph
( H, and the vector
)
t ∈ RConn(G) defined by tF = t(F, H), we have t(G, H) = stab L(Conn(G)), t .
Proof. By (5.22), we have
∑ ′
t(G, H) = (−1)|E | t(G′ , H),
E ′ ⊆E

where G′ = (V, E ′ ). Using that t(G′ , H) is multiplicative over the components of

G′ and that singleton components give a factor of 1, we get
∑ ∏
(5.38) t(G, H) = (−1)e(F ) t(F, H),
E ′ ⊆E F

where the product extends over all connected components of G′ with at least one
edge. These components form a stable set in L(Conn(G)), and vice versa, every
stable
( set in L(Conn(G))
) corresponds to a subgraph G′ . Hence the last sum is just
stab L(Conn(G)), t . 

Combining this lemma with the previous one, we get a very useful relationship
between homomorphism densities and the chromatic invariant.
Corollary 5.22. Assuming the series below is absolute convergent, we have
∑∞ ∑
(−1)m ( )∏
m
ln t(G, H) = cri L(F1 , . . . , Fm ) t(Fj , H).
m!
m=1 F1 ,...,Fm ∈Conn(G) j=1
 5.3. WHAT HOM FUNCTIONS CAN EXPRESS 67

What about all these assumptions about convergence? In fact, can we take
the logarithm in Lemma 5.20 at all? A fundamental result about the roots of the
stable set polynomial, Dobrushin’s Theorem [1996], gives us a sufficient condition
for this. Dobrushin’s Theorem has many statements in the literature, which are
more-or-less equivalent (but not quite), and we choose one that is convenient for
our purposes; see e.g. Scott and Sokal [2006] and Borgs [2006].
Theorem 5.23 (Dobrushin’s Theorem). Let G = (V, E) be a simple graph, and
let z ∈ CV and b ∈ RV+ satisfy
∑
(5.39) |zj |ebj ≤ bi
j∈{i}∪N (i)

for every node i. Then stab(G, z) ̸= 0, and

∑
| ln stab(G, z)| ≤ |zj |ebj .
j

Condition (5.39) defines a multidisc in which ln stab(G, z) is analytic, and so

by elementary properties of convergence of power series it follows that the series
in (5.35) is convergent inside this multidisc. We remark that to get good bounds
here, one can combine those terms in (5.38) for which the components of G′ give
the same partition of V (G); then we get an expression in terms of the intersection
graph of connected induced subgraphs of G, which has a smaller number of terms;
see e.g. Borgs, Chayes, Kahn and Lovász [2012].
Another important graph polynomial is the characteristic polynomial of the
adjacency matrix, A(G, x) = det(xI − AG ). We state a formula from Lyons [2005]
for the case when G is a connected D-regular non-bipartite graph.
Proposition 5.24. Let G be a connected D-regular non-bipartite graph on n nodes.
Then for every x > D, we have
∞
∑ hom(Cr , G)
(5.40) ln A(G, x) = n ln x − ,
r=1
rxr

and
∞
∑ hom(Cr , G) − Dr
(5.41) ln tree(G) = (n − 1) ln D − ln n − .
r=1
rDr

The formula for the number of trees extends easily to non-regular graphs, since
we can add loops to the nodes to make the graph regular, and adding loops does
not change the number of spanning trees (adding a loop to a node increases its
degree by 1 in this case). This expression seems to have been first formulated by
Lyons [2005].

Proof. We can write this polynomial as

∏
n
(5.42) A(G, x) = (x − λk ),
k=1

where λ1 ≥ λ2 ≥ · · · ≥ λn are the eigenvalues of the matrix AG . It is well known

that we have λ1 = D > λ2 (as G is connected) and λn > −D (as G is non-bipartite).
68 5. GRAPH HOMOMORPHISMS

To handle this product, we take the logarithm and expand it:

∏
n ∑
n ∑ ∞
n ∑
1 λr
ln (x − λk ) = n ln x + ln(1 − λk /x) = n ln x − k
r xr
k=1 k=1 k=1 r=1
∑∞
1 ∑ r
n
(5.43) = n ln x − λk .
r=1
rxr
k=1

We can express the last sum using (5.31), to get (5.40).

By the Matrix Tree Theorem, ntree(G) is the coefficient of the linear term in
the determinant det(yI + DI − A), and hence
( )
ln tree(G) = lim ln det(yI + DI − A) − ln y − ln n .
y→0

Using (5.40),
∞
∑ hom(Cr , G)
ln det(yI + DI − A) = n ln(y + D) −
r=1
r(y + D)r
∑∞
hom(Cr , G) − Dr y
= n ln(y + D) − + ln
r=1
r(y + D)r y+D
∞
∑ hom(Cr , G) − Dr
= (n − 1) ln(y + D) − + ln y.
r=1
r(y + D)r

Substituting this in the formula for ln tree(G) and letting y → 0, we get (5.41). 

Exercise 5.25. Prove identity (5.32).

Exercise 5.26. Let H = H(a, B) be a weighted graph, where
( ) ( )
1 2 1
a= , B=
−1 1 1
(illegal weighting, because there is a negative nodeweight, but the formula defining
the hom function makes sense). Prove that hom(F, H) is the number of those
subsets of edges that cover every node.
Exercise 5.27. Verify that (5.41) yields the Cayley formula tree(Kn ) = nn−2 .
∏
Exercise 5.28. Prove that ntree(G) = n k=2 (D − λk ), and show that this implies
(5.41).

5.4. Homomorphism and isomorphism

5.4.1. Homomorphism–profiles. We start with a simple but useful obser-
vation (Lovász [1967]); various less trivial extensions and generalizations of this fact
will play an important role a number of times (cf. Theorems 5.33, 13.9 and 17.5,
and Corollaries 5.45 and 10.34).
Theorem 5.29. Either one of the simple graph parameters hom(., G) and hom(G, .)
determines a simple graph G.
By the same argument, these parameters defined on looped-simple graphs de-
termine a looped-simple graph G.
 5.4. HOMOMORPHISM AND ISOMORPHISM 69

Proof. We prove that hom(., G) determines G; the argument for hom(G, .) is

similar. The analogous statement for injective homomorphisms is trivial: if G and
G′ are two simple graphs such that inj(F, G) = inj(F, G′ ) for every simple graph
F , then in particular inj(G′ , G) = inj(G′ , G′ ) > 0 and inj(G, G′ ) = inj(G, G) > 0,
so G and G′ have injective homomorphisms into each other, and hence they are
isomorphic.
Now (5.18) expresses injective homomorphism numbers in terms of ordinary
homomorphism numbers, which implies that if hom(F, G) = hom(F, G′ ) for every
simple graph F , then inj(F, G) = inj(F, G′ ) for every simple graph F , and hence
G∼ = G′ . 

We see from the proof of Theorem 5.29 that in fact G is determined by the
values hom(F, G) where v(F ) ≤ v(G), as well as by the values hom(G, F ) where
v(F ) ≤ v(G). It is a long-standing open problem whether, up to trivial exceptions,
strictly smaller graphs F are enough:
Conjecture 5.30 (Reconstruction Conjecture). If G is a simple graph with
v(G) ≥ 3, then the numbers hom(F, G) with v(F ) < v(G) determine G.
There is a weaker version, which is also unsolved:
Conjecture 5.31 (Edge Reconstruction Conjecture). If G is a simple graph
with e(G) ≥ 4, then the numbers hom(F, G) with e(F ) < e(G) determine G.
It is known that the Edge Reconstruction Conjecture holds for graphs G with
e(G) ≥ v(G) log v(G) (Müller [1977]). We will prove an “approximate” version of
the Reconstruction Conjecture (Theorem 10.32): for an arbitrarily large graph
√ G,
the numbers hom(F, G) with v(F ) ≤ k determine G up to an error of O(1/ log k)
(measured in the cut distance, which was mentioned in the Introduction but will
be formally defined in Chapter 8). Unfortunately, this does not seem to bring us
closer to the resolution of the Reconstruction Conjecture.
The normalized homomorphism density function t(., G) does not determine a
simple graph(G: If G(p)) is obtained from G by replacing every node by p twin
nodes, then t F, G(p) = t(F, G). But this is all that can go wrong:
Theorem 5.32. If G1 and G2 are simple graphs such that t(F, G1 ) = t(F, G2 ) for
every simple graph F , then there is a third simple graph G and positive integers
p1 , p2 such that G1 ∼
= G(p1 ) and G2 ∼
= G(p2 ).
Proof. Let ni = v(Gi ), and consider the blowups G′1 = G1 (n2 ) and G′2 =
G2 (n1 ). These have the same number of nodes, and hence t(F, G′1 ) = t(F, G1 ) =
t(F, G2 ) = t(F, G′2 ) implies that hom(F, G′1 ) = hom(F, G′2 ). So by Theorem 5.29, we
have G′1 ∼ = G′2 . It follows that the number of elements in every class of twin nodes of
′ ∼ ′
G1 = G2 is divisible by both n1 and n2 , and so it also divisible by m = lcm(n1 , n2 ).
So G′1 ∼= G′2 ∼= G(m) for some simple graph G, and hence pi = m/ni satisfies the
requirements in the theorem. 

For weighted graphs, one must be a little careful. Let H be a weighted graph
and let H ′ be obtained from H by twin reduction. Then hom(F, H ′ ) = hom(F, H)
for every multigraph F , even though H and H ′ are not isomorphic. Restricting
our attention to twin-free graphs, we have an analogue of Theorem 5.29 (Lovász
[2006b]):
70 5. GRAPH HOMOMORPHISMS

Theorem 5.33. Let H1 and H2 be twin-free weighted graphs such that

hom(F, H1 ) = hom(F, H2 ) holds for all simple graphs F with at most 2(v(H1 ) +
v(H2 ) + 3)8 nodes. Then H1 ∼
= H2 .
The proof of this theorem is substantially more complicated then that of its
unweighted version. A proof without the bound on the size of F will be described
in Section 6.4.1, where it will follow easily from the general tools developed there,
and the full proof will be postponed until Section 6.4.2
5.4.2. Algebraic properties of graph multiplication. The fact that ho-
momorphism numbers (into it or from it) determine the graph will motivate much
in the sequel. Here we describe a few old applications of this fact to some basic
algebraic properties of categorical product (Lovász [1967, 1971]).
First, we show that taking k-th root is unique (if it exists at all).
Theorem 5.34. If G1 and G2 are looped-simple graphs such that G×k ∼ ×k for
1 = G2
some k ≥ 1, then G1 ∼
= G2 .
Proof. For every looped-simple graph F , we have
hom(F, G1 ) = hom(F, G×k
1 )
1/k
= hom(F, G×k
2 )
1/k
= hom(F, G2 ),
∼
whence by Theorem 5.29, G1 = G2 . 
Next we turn to the question of Cancellation Law: does G1 × H ∼ = G2 × H
(where G1 , G2 and H are looped-simple graphs) imply that G1 ∼
= G2 ? This is false
in general:
(5.44) K2 × C6 ∼
= K2 × (K3 K3 ).
But the proof method of Theorem 5.34 almost goes through: for every looped-simple
graph F , we have
hom(F, G1 )hom(F, H) = hom(F, G1 × H) = hom(F, G2 × H)
= hom(F, G2 )hom(F, H);
if hom(F, H) ̸= 0, then this implies that hom(F, G1 ) = hom(F, G2 ). What to do if
hom(F, H) = 0? We can find several simple conditions under which this difficulty
can be handled:
Proposition 5.35. Let G1 , G2 and H be looped-simple graphs such that G1 × H ∼
=
G2 × H.
(a) If H has a loop, then G1 ∼= G2 .
(b) If both G1 and G2 have a homomorphism into H, then G1 ∼ = G2 .
(c) If a looped-simple graph H has a homomorphism into H, then G1 × H ′
′ ∼
=
G2 × H ′ .
Since strong product corresponds to having a loop at every node, we get:
Corollary 5.36. Let G1 , G2 and H be simple graphs such that G1  H ∼
= G2  H.
Then G1 ∼
= G2 .
The proof of Proposition 5.35 is left to the reader as an exercise. With a little
more effort, we can characterize cancelable graphs (Lovász [1971]). If H is bipartite,
then hom(H, K2 ) > 0, and so (5.44) and Proposition 5.35 imply
H × C6 ∼ = H × (K3 ∪ K3 ),
 5.4. HOMOMORPHISM AND ISOMORPHISM 71

so H is not cancelable. On the other hand, nonbipartite graphs are cancelable:

Theorem 5.37. Let G1 , G2 and H be looped-simple graphs such that G1 × H ∼
=
G2 × H. If H is not bipartite, then G1 ∼
= G2 .
The proof depends on the following lemma:
Lemma 5.38. Suppose that G1 × H ∼ = G2 × H. Then there is an isomorphism
σ : G1 ×H → G2 ×H such that σ(V (G1 )×{v}) = V (G2 )×{v} for every v ∈ V (H).
Proof. We consider graphs G together with a homomorphism π : G → H.
We call the pair G = (G, π) an H-colored graph. For every graph G, the product
G × H is H-colored in the natural way by the projection onto H. We denote this
H-colored graph by GH .
Two H-colored graphs F = (F, ρ) and G = (G, π) are isomorphic, if there
is(an isomorphism
) σ : G1 → G2 which commutes with the projections to H, i.e.,
π η(i) = ρ(i) for every i ∈ V (F ). In this language, we want to prove that G1 ×H ∼
=
G2 × H implies that (G1 )H ∼ = (G2 )H .
For two H-colored graphs F = (F, ρ) and G = (G, π), let hom(F, ( G)) denote
the number of those homomorphisms η from F to G that satisfy π η(i) = ρ(i)
for every i. Let inj(F, G) denote the number of injective homomorphisms with this
property.
We can define the product F × G of two H-colored graphs F = (F, ρ) and
G = (G, π) as the subgraph of F ×G induced by those nodes (i, j) with ρ(i) = π(j),
together with the homomorphism σ(i, j) = ρ(i) into H.
The case when H consists of a single node with a loop is equivalent to just
ordinary homomorphism numbers.
Two identities extend quite easily to this more general notion:
(5.45) hom(F, G1 × G2 ) = hom(F, G1 )hom(F, G2 ),
and
∑
(5.46) inj(F, G) = µ(F, F′ )hom(F′ , G),
F′

where we sum over all H-colored graphs F′ on at most v(F) nodes, with appropriate
coefficients µ(F, F′ ). Let us add the easy identity
(5.47) (F × G)H ∼= FH × GH .
From G1 × H ∼= G2 × H it follows that (G1 × H)H =∼ (G2 × H)H (as H-colored
graphs). By (5.47), this implies that (G1 )H × HH ∼
= (G2 )H × HH , and hence by
(5.45),
hom(F, (G1 )H )hom(F, HH ) = hom(F, (G2 )H )hom(F, HH ).
But notice that hom(F, HH ) > 0: if F = (F, σ), then (σ, σ) is a homomorphism
F → HH . Thus we can divide by hom(F, HH ) to get
hom(F, (G1 )H ) = hom(F, (G2 )H )
for every H-colored graph F. From here (G1 )H = ∼ (G2 )H follows just like in the
proof of Theorem 5.29. 
Proof of Theorem 5.37. By Proposition 5.35, we may assume that H is an
odd cycle with V (H) = [2r + 1] and E(H) = {ij : j ≡ i + 1 (mod 2r + 1)}. By
Lemma 5.38 there exist bijections φ1 , . . . , φ2r+1 : V (G1 ) → V (G2 ) such that for
72 5. GRAPH HOMOMORPHISMS

every ij ∈ E(H), φi (u)φj (v) ∈ E(G2 ) if and only if uv ∈ E(G1 ). (Note that this
means a different condition if we interchange i and j.)
We show that φ1 is an isomorphism between G1 and G2 . Indeed, we have
( )
φ1 (u)φ1 (v) ∈ E(G2 ) ⇐⇒ φ−1
2 φ1 (u) v ∈ E(G1 ) ⇐⇒ φ1 (u)φ3 (v) ∈ E(G2 )
( )
⇐⇒ φ−1
4 φ1 (u) v ∈ E(G1 ) ⇐⇒ φ1 (u)φ5 (v) ∈ E(G2 )
⇐⇒ . . . ⇐⇒ φ1 (u)φ2r+1 (v) ∈ E(G2 ) ⇐⇒ uv ∈ E(G1 )
This completes the proof. 
Remark 5.39. You may have noticed that the proof of Lemma 5.38 followed the
lines of the proof of Proposition 5.35, only restricting the notion of homomorphisms
to those respecting the H-coloring. This suggests that there is a more general
formulation for categories. This is indeed the case, as we will see in Section 23.4.
A further natural question about multiplication is whether prime factorization
is unique. This is clearly a stronger property than the Cancellation Law, so let us
restrict our attention to the strong product, which satisfies the Cancellation Law.
The following example shows that prime factorization is not unique in general. We
start with an algebraic identity:
(5.48) (1 + x + x2 )(1 + x3 ) = (1 + x)(1 + x2 + x4 ).
If we substitute any connected graph G for x, and interpret “+” as disjoint union,
we get a counterexample. For example,
(5.49) (K1 ∪ K2 ∪ K4 )  (K1 ∪ K8 ) = (K1 ∪ K2 )  (K1 ∪ K4 ∪ K16 ).
But there is a very nice positive result of Dörfler and Imrich [1970] and McKen-
zie [1971]. (The proof uses different techniques, and we don’t reproduce it here.)
Theorem 5.40. Prime factorization is unique for the strong product of connected
graphs. 

Exercise 5.41. (a) Prove that the strong product of two graphs is connected if
and only if both graphs are connected. (b) Show by an example that the categor-
ical product of two connected graphs is not always connected. (c) Characterize
all counterexamples in (b).
Exercise 5.42. Given two looped-simple digraphs F and G, we define the
digraph GF ( as follows:) V (GF ) =( V (G)V (F ) , E(G) ) = {(φ, ψ) : φ, ψ ∈
F

V (G) V (F )
, φ(u), ψ(v) ∈ E(G) ∀(u, v) ∈ E(F ) }. (a) Prove the following
identities:
(G1 × G2 )F ∼= G1 × G2 ,
F F
GF1 ×F2 ∼ F F
= (G 1 ) 2 , GF1 F2 ∼
= G 1 × G 2.
F F

(b) Show that hom(F, G) is the number of loops in GF . (c) Prove that if adjacency
is symmetric both in G and in F , then it is also symmetric in GF (Lovász [1967]).

5.5. Independence of homomorphism functions

How independent are homomorphism functions hom(F, .) (in an algebraic
sense)? We know that hom(F1 F2 , G) = hom(F1 , G)hom(F2 , G) for two (unlabeled)
graphs F1 and F2 ; is this the only identity relating these functions?
We start with excluding linear relations. For a set of (non-isomorphic) simple
graphs A = {F1 , . . . , Fm }, we define the matrix
A
( )m
Mhom = hom(Fi , Fj ) i,j=1 .
 5.5. INDEPENDENCE OF HOMOMORPHISM FUNCTIONS 73

A A A
The matrices Minj and Msurj are defined analogously. Finally, we also define Maut as
the matrix with aut(Fi ) = surj(Fi , Fi ) = inj(Fi , Fi ) in the i-th entry of the diagonal
and 0 outside the diagonal.
A
Clearly, Maut is a diagonal matrix; if we order the graphs Fi according to
increasing number of edges (and arbitrarily for graphs with the same number of
A A
edges), then the matrices Minj and Msurj become triangular. All diagonal entries
A A A
are positive in each case. Hence the matrices Maut , Minj and Msurj are nonsingular.
A
With Mhom the situation is more complicated: it may be singular (Exercise 5.46).
However, we have the following simple but useful fact, observed by Borgs, Chayes,
Kahn and Lovász [2012]:

Proposition 5.43. Let A be a family of simple graphs closed under surjective

A
homomorphisms. Then Mhom is nonsingular.

In particular, this holds if A consists of all graphs with at most k nodes, or at

most k edges, for some k ≥ 0.

Proof. Under the conditions of the Proposition, the matrices introduced above
are related by the following identity:
A A A −1 A
(5.50) Mhom = Msurj (Maut ) Minj .
Indeed, every homomorphism can be decomposed as a surjective homomorphism
followed by an (injective) embedding. By our assumption, the image F of the
surjective homomorphism is in A. The decomposition is uniquely determined except
for the automorphisms of F . This gives the equation
∑
m
surj(Fi , Fk )inj(Fk , Fj )
hom(Fi , Fj ) = ,
aut(Fk )
k=1

which is just 5.50 written out in coordinates.

A
It follows that Mhom is the product of three nonsingular matrices, and hence it
is also nonsingular. 

If A is just an arbitrary set of simple graphs, we can still create a nonsingular

A
matrix related to Mhom (Erdős, Lovász and Spencer [1979]).

Proposition 5.44. Let F1 , . . . , Fk be nonisomorphic simple graphs.

(a) Let Hi be obtained from Fi by weighting its nodes, and suppose that all the
[ ]k
weights used are algebraically independent. Then the matrix hom(Fi , Hj ) i,j=1 is
nonsingular.
[ ]k
(b) There are simple graphs G1 , . . . , Gk such that the matrix hom(Fi , Gj ) i,j=1
is nonsingular.
(c) If F1 , . . . , Fk have no isolated nodes, then there are simple graphs G1 , . . . , Gk
[ ]k
such that the matrix t(Fi , Gj ) i,j=1 is nonsingular.

We could use nodeweights in (a) chosen randomly and independently from the
uniform distribution on [0, 1] (or form any other atomfree distribution); the matrix
will be nonsingular with probability 1.
74 5. GRAPH HOMOMORPHISMS

Proof. (a) Considering the node weights as variables, the determinant of the
[ ]k
matrix hom(Fi , Hj ) i,j=1 is a polynomial p with integral coefficients. The mul-
[ ]k
tilinear part of p is just the determinant of inj(Fi , Hj ) i,j=1 , which is non-zero,
since this matrix is upper triangular and the diagonal entries are nonzero polyno-
mials. Hence p is not the zero polynomial, which shows that for an algebraically
independent substitution it does not vanish.
(b) Instead of algebraically independent weights, we can also substitute appro-
[ ]k
priate positive integers in p to get a nonsingular matrix hom(Fi , Hj ) i,j=1 , since
a nonzero polynomial cannot vanish for all positive integer substitutions. For a
graph Hj and a node v ∈ V (Hj ) with weight mv , we replace v by mv twin copies of
weight 1. Let Gj be the graph obtained this way, then hom(Fi , Gj ) = hom(Fi , Hj )
[ ]k [ ]k
for all i, and hence hom(Fi , Gj ) i,j=1 = hom(Fi , Hj ) i,j=1 is nonsingular.
(c) Let n = maxi v(Fi ), and let us add n − v(Fi ) isolated nodes to every
Fi . The resulting graphs Fi′ are non-isomorphic, and hence there are simple
[ ]k
graphs G1 , . . . , Gk such that the matrix hom(Fi′ , Gj ) i,j=1 is nonsingular. Since
hom(Fi′ , Gj ) = v(Gj )n t(Fi′ , Gj ), we can scale the columns and get that the matrix
[ ]k
t(Fi′ , Gj ) i,j=1 is nonsingular. Since clearly t(Fi , Gj ) = t(Fi′ , Gj ), this proves the
proposition. 
The following corollary of these constructions goes back to Whitney [1932]. We
have seen that the homomorphism functions satisfy the multiplicativity relations
hom(F1 F2 , G) = hom(F1 , G)hom(F2 , G) (where F1 F2 denotes disjoint union). Is
there any other algebraic relation between them? Using multiplicativity, we can
turn any algebraic relation to a linear relation, so the question is: are the graph
parameters hom(F, .) linearly independent (in the sense that any finite number of
them are). Thus (b) above implies:
Corollary 5.45. The simple graph parameters hom(F, .) (where F ranges over
simple graphs) are linearly independent. Equivalently, the simple graph parameters
hom(F, .) (where F ranges over connected simple graphs) are algebraically indepen-
dent.
What about non-algebraic relations? Such relations sound unlikely, and in fact
it can be proved (Erdős, Lovász and Spencer [1979]) that they don’t exist. To be
more precise, for any finite set of(distinct connected graphs ) A = {F1 , . . . , Fk }, if we
construct the set T (A) of points t(F1 , G), . . . , t(Fk , G) ∈ Rk , where G ranges over
all finite graphs, then the closure T (A) has an internal point. We will talk more
about these sets T (A) in Chapter 16.
A
Exercise 5.46. Show by an example that Mhom may be singular.
Exercise 5.47. Prove a version of part (a) of Proposition 5.44 in which the edges
are weighted (instead of the nodes).
Exercise 5.48. Find an upper bound on the number of nodes in the graphs Gi
in part (b) and (c) of Proposition 5.44.
Exercise 5.49. For every m ≥ 1, construct a family A of m simple graphs such
A
that the matrix Mhom is the identity matrix.
Exercise 5.50. For every m ≥ 1 there exist simple graphs F1 , . . . , Fm such that
for every integer vector a ∈ Nm there is a simple graph G such that hom(Fi , G) =
ai for all i ∈ [m].
 5.6. CHARACTERIZING HOMOMORPHISM NUMBERS 75

Exercise 5.51. Let H1 , . . . , Hm be non-isomorphic simple graphs. Prove that

there are no linear relations between the graph parameters hom(., Hi ).
Exercise 5.52. (a) Let H1 , . . . , Hm be non-isomorphic simple connected graphs.
Prove that there are no linear relations between the graph parameters hom(., Hi ),
even when they are restricted to connected graphs. (b) Show that this is no longer
true if we don’t assume the connectivity of the Hi .
Exercise 5.53. (a) Let H1 , . . . , Hm be simple nonisomorphic connected nonbi-
partite graphs. Prove that there is a simple connected graph F such that the
homomorphism numbers hom(F, Gi ) are distinct. (b) Show that for every simple
graph F , at least two of the numbers hom(F, C6 ), hom(F, K2 ) and hom(F, K3 K3 )
are equal.

5.6. Characterizing homomorphism numbers

In the previous sections (e.g. in the proof of Theorem 5.34 and related results)
our key tool was to associate, with every graph G, the graph parameter hom(., G).
What else can be said about these graph parameters? It turns out that they have
an interesting characterization, which will play an important role throughout this
book. There are different versions of this characterization, of which we state a
sample.
Multigraph parameters of the form hom(., H), where H is a weighted graph,
were characterized by Freedman, Lovász and Schrijver [2007].
Theorem 5.54. Let f be a graph parameter defined on multigraphs without loops.
Then f is equal to hom(., H) for some weighted graph H on q nodes if and only if
it is reflection positive, f (K0 ) = 1, and r(f, k) ≤ q k for all k ≥ 0.
Let us note that the condition for k = 0 says that r(f, 0) ≤ 1, which implies
that f is multiplicative (Exercise 4.4).
In terms of statistical physics, this theorem can be viewed as a characteriza-
tion of partition functions of vertex coloring models. Theorem 5.54 implies that
those graph parameters that can be expressed as homomorphism numbers into fixed
weighted graphs are all reflection positive and have exponentially bounded connec-
tion rank. It may be instructive to see directly why this is so for the number of
nowhere-zero flows.
Example 5.55. For two k-labeled graphs G1 and G2 , the value flo(G1 G2 , q) can
be computed by a simple formula provided we know, for all a1 . . . , ak ∈ Zq , the
number flo(Gi ; a1 , . . . , ak ) of nowhere-zero q-flows in Gi with “surplus” ai at each
node i ∈ [k]; then we have a formula similar to (4.6), except that the summation
will range over a ∈ Zkq instead of Zk :
∑
flo(G1 G2 , q) = flo(G1 ; a1 , . . . , ak )flo(G2 ; −a1 , . . . , −ak )
a
∑
(5.51) = flo(G1 ; a1 , . . . , ak )flo(G2 ; a1 , . . . , ak ).
a

From this, we see that M (flo, k) is positive semidefinite and has rank at most q k .

Schrijver [2009] gave the following characterization of graph parameters repre-
sentable as homomorphism functions into weighted graphs with node weights 1 and
complex edgeweights. Recalling the Möbius inverse on the partition lattice (4.3),
we can state the result as follows:
76 5. GRAPH HOMOMORPHISMS

Theorem 5.56. Let f be a complex valued graph parameter defined on looped

multigraphs. Then f = hom(., H) for some edge-weighted graph H on q nodes with
complex edgeweights if and only if f is multiplicative, f (K1 ) = q, and f ⇓ (G) = 0
for every graph G with more than q nodes.

Using this theorem, Schrijver gave a real-valued version, which is more similar
to Theorem 5.54.

Theorem 5.57. Let f be a real valued graph parameter defined on looped multi-
graphs. Then f = hom(., H) for some edge-weighted graph H with real edgeweights
if and only if f is multiplicative and, for every integer k ≥ 0, the multilabeled
connection matrix M mult (f, k) is positive semidefinite.

Every graph parameter f defined on looped-simple graphs can be extended

to looped-multigraphs so that it is invariant under adding parallel edges. Every
homomorphism function hom(., H) where all edge-weights are 0 or 1 defines such a
multigraph parameter. Conversely, if f = hom(., H) (where H is a weighted graph)
is invariant under adding parallel edges, then every edge of H must have weight 0
or 1 (Exercise 5.66). In particular, if all nodeweights of H are 1, then H can be
viewed as a looped-simple graph itself. Hence Theorem 5.57 implies the following
characterization of homomorphism numbers into looped-simple graphs, as noticed
by Lovász and Schrijver [2010, 2009]:

Corollary 5.58. Let f be a graph parameter defined on looped-simple graphs. Then

f = hom(., H) for some looped-simple graph H if and only if f is multiplicative and,
for every integer k ≥ 0, the connection matrix M mult (f, k) is positive semidefinite.

Note that in Theorem 5.57 and Corollary 5.58 no bound on the connection rank
is assumed; in fact (somewhat surprisingly), it follows from the multiplicativity
and reflection positivity conditions that f has finite connection rank, and r(f, k) ≤
f (K1 )k for all k. Furthermore, in Corollary 5.58 it also follows from the conditions
that the values of f are integers.
Next, we state an analogous (dual) characterization of graph parameters of the
form hom(F, .), defined on looped-simple graphs, where F is also a looped-simple
graph (Lovász and Schrijver [2010]). To state the result, we need some definitions.
Recall the notion of H-colored graphs and their products from the proof of Lemma
5.38; we need only the rather trivial version where H = Kq◦ is a fully looped
complete graph. We define dual connection matrices N (f, q) of a graph parameter
f : the rows and columns are indexed by Kq◦ -colored graphs, and the entry in row
G1 and column G2 is f (G1 × G2 ).

Theorem 5.59. Let f be a graph parameter defined on looped-simple graphs. Then

f = hom(F, .) for some looped-simple graph F if and only if f is multiplicative over
direct product, and for each k ≥ 1, the dual connection matrix N (f, k) is positive
semidefinite.

It is interesting to note that “primal” connection matrices of these “dual” ho-

momorphism numbers hom(F, .) also have finite rank (see Exercise 5.67). However,
no characterization of homomorphism numbers in terms of these “primal” connec-
tion matrices is known.
 5.6. CHARACTERIZING HOMOMORPHISM NUMBERS 77

5.6.1. Randomly weighted graphs. The last result to be presented in this

line is a characterization of multiplicative and reflection positive multigraph pa-
rameters with finite connection rank (Lovász and Szegedy [2012c]). To state the
result, we have to generalize the notion of weighted graphs.
A randomly weighted graph is a finite graph H (which we may assume to be a
looped complete graph) in which the nodes are weighted with positive real numbers
(just like in the case of ordinary weighted graphs) and each edge ij is weighted by
a random variable Bij taking values from a finite set of reals. Ordinary weighted
graphs can be regarded as randomly weighted graphs in which the edgeweights are
random variables concentrated on a single value.
To define homomorphism numbers into a randomly weighted graph takes a lit-
tle care. A first idea is to define it as the expectation of hom(F, H) where H is the
weighted graph where the edgeweights are generated randomly and independently
from the corresponding distributions. However, this quantity would not be multi-
plicative. We could start with taking the expectation separately for each edge; this
would then give nothing new relative to the homomorphism numbers into weighted
graphs. We therefore take a middle ground:
∑ ∏ ∏ Fij
(5.52) hom(F, H) = αφ(i) E(Bφ(i)φ(j) ),
φ: V (F )→V (H) i∈V (F ) ij∈E(F simp )

where Fij is the multiplicity of the edge ij in F . This quantity is multiplicative,

and it specializes to the previously defined homomorphism number when the edge
weights are deterministic.
We note two special cases. If F is simple, then we could take the expectation
all the way in; in other words, homomorphisms into randomly weighted graphs
give no new simple graph parameters. On the other hand, if we consider inj(F, H)
(restricting the summation in (5.52)( to injections),
) then we can take the expectation
all the way out, i.e., inj(F, H) = E inj(F, H) .
simp
Example 5.60. Consider the multigraph parameter f (G) = pe(G ) , where
0 < p < 1 is fixed. It is not hard to see that this is reflection positive and its con-
k
nection rank is 2(2) . We can characterize it as hom(G, K1◦ [p]), where K1◦ [p] is the
randomly weighted graph on a single node with a loop, where the loop is decorated
by the probability distribution on {0, 1} in which 1 has probability p. We can also
characterize it as the expectation of tinj (G, G(n, p)), where n ≥ v(G). This param-
k
eter is multiplicative, reflection positive, and its connection rank if r(f, k) = 2(2) ,
which is finite for every k, but has superexponential growth. 
With this generalized notion of homomorphism numbers, we are able to state
the theorem announced above:
Theorem 5.61. A multigraph parameter f is equal to hom(., H) for some randomly
weighted graph H if and only if it is multiplicative, reflection positive and r(f, 2) is
finite.
It would not be enough to assume that r(f, 1) is finite instead of r(f, 2) (see
Exercise 5.65). While in Theorem 5.61 we don’t have to assume anything about
the higher connection ranks, it does follow that they are all finite. In fact, we have
the following “Theorem of Alternatives”:
78 5. GRAPH HOMOMORPHISMS

Supplement 5.62. Let f be a multiplicative and reflection positive parameter de-

fined on multigraphs without loops. Then one of three alternative must occur: (i)
r(f, k) is infinite for all k ≥ 2; (ii) r(f, k) is finite for all k, and log r(f, k) = Θ(k);
(iii) r(f, k) is finite for all k, and log r(f, k) = Θ(k 2 ).

It follows that alternative (ii) obtains iff f = hom(., H) for some weighted graph
H, and alternative (iii) obtains when f = hom(., H) for some randomly weighted
graph H in which at least one edgeweight has a proper distribution.
It is possible to give a more precise description of the asymptotic behavior
of log r(hom(., H), k), but we have to refer to the paper of Lovász and Szegedy
[2012c] for details. Let us note that no such conclusion can be drawn without
assuming reflection positivity. For example, the chromatic polynomial chr(., x)
satisfies log rk(chr(., x), k) = Θ(k log k).

Remark 5.63. Several further improvements, versions and extensions of these

results have been obtained, extending them to directed graphs, hypergraphs, semi-
groups, and indeed, to all categories satisfying reasonable conditions. In this book,
related characterizations will be described for homomorphisms into graphons and
random graphons (Theorem 11.52 and Proposition 14.60), morphisms in categories
(Theorem 23.16), and edge coloring models (Theorem 23.5). One would wish to
derive all of these from a single “Master Theorem”; alas, this has not yet been
found.
The least appealing feature of these theorems is that the necessary and sufficient
condition involves infinite matrices, and in most cases infinitely many of them.
While this is clearly unavoidable in a sense (the condition must involve the value
of the parameter on all graphs), one can formulate conditions that involve only
submatrices with a simpler structure. For example, in Theorem 5.61, it suffices to
consider fully and simply labeled graphs, fully labeled edgeless graphs, and fully
labeled bonds (cf. also the proof of Theorem 5.57 given is Section 6.6).

5.6.2. About the proofs. The proofs of the theorems above follow at least
three different lines. To be more precise, the necessity of the conditions is easy to
prove; below we prove the “easy” direction of Theorem 5.54, and the others follow
by essentially the same argument. The sufficiency parts will be postponed until
some further techniques will be developed:
—The completion of the proof of Theorem 5.54 will be given in Section 6.2.2,
after the development of graph algebras. (These algebras will be useful to study
other related properties of homomorphism functions, and the technique will also be
applied in extremal graph theory.) Corollary 5.58 and its dual, Theorem 5.59, can
be proved by a similar technique. This technique extends to a much more general
setting, to categories, as we will sketch in Section 23.4.
—The proofs of Theorems 5.56 and 5.57 will be described in Section 6.6, where
a general connection to the Nullstellensatz and invariant theory will be developed.
This method extends to edge coloring models (see Section 23.2).
—The proof of Theorem 5.61 will use a lot of the analytic machinery to be
developed in Part 3 of the book, and will be sketched at the end of that part
(Section 17.1.4). For the details of this proof, and for the proof of Supplement 5.62,
we refer to Lovász and Szegedy [2012c].
We conclude this section proving the “easy” direction in Theorem 5.54:
 5.7. THE STRUCTURE OF THE HOMOMORPHISM SET 79

Proposition 5.64. For every weighted graph H, the graph parameter hom(., H) is
reflection positive and r(hom(., H), k) ≤ v(H)k .
Proof. For any two k-labeled graph F1 and F2 and φ : [k] → V (H), we have
(5.53) homφ (F1 F2 , H) = homφ (F1 , H)homφ (F2 , H)
(recall the definition of homφ from (5.10)). Let F = [[F1 F2 ]], then the decomposition
∑
hom(F, H) = αφ homφ (F, H).
φ: [k]→V (H)

writes the matrix M (hom(., H), k) as the sum of v(H)k matrices, one for each
mapping φ : [k] → V (H); (5.53) shows that these matrices are positive semidefinite
and have rank 1. This implies Lemma 5.64. 

Exercise 5.65. For a multigraph G on [n], let X1 , . . . , Xn be random points on

the unit circle, and let f (G) denote the probability that Xi · Xj ≥ 0 for every edge
ij of G. Prove that f is reflection positive, r(f, 0) = r(f, 1) = 1, but r(f, 2) = ∞.
Exercise 5.66. Let H be a weighted graph for which the looped-multigraph
parameter hom(., H) is invariant under adding parallel edges. Prove that all
edgeweights of H are 0 or 1.
Exercise 5.67. Prove that the connection rank r(hom(F, .), k) is bounded by
(k + 2)v(F ) .

5.7. The structure of the homomorphism set

We have discussed the existence of homomorphisms between two graphs F
and G, i.e., the emptiness or non-emptiness of the homomorphism set Hom(F, G).
We considered the size of this set (in fact, much of this book turns around this
number). This set has further structure, which is quite interesting and which can
be exploited to obtain combinatorial results about graphs. We only give a glimpse
of these questions.

5.7.1. The graph of homomorphisms. Let F and G be two simple graphs.

The set Hom(F, G) can be endowed with a graph structure. Brightwell and Win-
kler [2004] define a graph Hom(F, G) by connecting two nodes, meaning homomor-
phisms φ, ψ : F → G, if they differ only on one node of F .
Example 5.68 (Linegraph). If F = K2 , then we get a version of the line-graph of
G: every edge of G will be represented by two nodes in Hom(K2 , G) corresponding
−
→
to the two orientations of the edge, and two nodes (oriented edges ij and −
→ will
uv)
be connected if either i = u or j = v. (The ordinary linegraph is obtained by
merging the two copies of each edge.) 
Example 5.69 (Colorings). Let the target graph G be a complete q-graph. In
this case, nodes are legitimate q-colorings of F , and two of them are adjacent if only
one node of F is recolored to get one coloring from the other. This construction is
important when analyzing the “heat bath” or “Glauber dynamics” Markov chain
in statistical physics, which corresponds to a random walk on this graph. We will
need this Markov chain in Section 20.1.2. 
80 5. GRAPH HOMOMORPHISMS

Brightwell and Winkler relate properties of the Hom graph to a number of

important issues in statistical physics, like long-range actions and phase transitions.
These would be too difficult to state here, but a corollary of their main result is
worth formulating:
Theorem 5.70. Suppose that G is a graph such that the graph Hom(F, G) is con-
nected for every connected graph F with maximum degree d. Then the chromatic
number of G is at least d/2 + 1. 
(They conjecture that d/2 + 1 can be replaced by d.)
5.7.2. The complex of homomorphisms. The set Hom(F, G) can also
be equipped with a topological structure. We say that a set of homomorphisms
φ1 , . . . , φk : F → G is a cluster if for every edge uv ∈ E(F ) and any 1 ≤ i < j ≤ k,
we have φi (u)φj (v) ∈ E(G). It is clear that these clusters form a simplicial complex
Hom(F, G) (i.e., they are closed under taking subsets). It is quite surprising that
topological properties of this complex have graph-theoretic consequences.
What is important about this construction is that it is “functorial”, which
means that every homomorphism ψ : G1 → G2 induces a simplicial (and hence
continuous) map ψb : Hom(F, G1 ) → Hom(F, G2 ) in a canonical way: For every
b
homomorphism
( )φ : F → G1 , we define ψ(φ)(= φψ. )It is trivial that this map from
V Hom(F, G1 ) = Hom(F, G1 ) to V (Hom F, G2 ) = Hom(F, G2 ) maps clusters
onto clusters. We also note that the automorphism group of F acts on Hom(F, G):
if α is an automorphism of F , then α̌ : φ 7→ αφ is an automorphism of Hom(F, G).
We quote two theorems relating properties of these topological spaces to col-
orability of the graph, and they are important tools in determining the chromatic
number of certain graph families. (See Kozlov [2008] and Matoušek [2003] for de-
tailed treatments of this topic.) The first is a re-statement in this language of a
result of Lovász [1978].
Theorem 5.71. If Hom(K2 , G) is k-connected as a topological space, then the
chromatic number of G is at least k + 3.
The second theorem is due to Babson and Kozlov [2003, 2006, 2007].
Theorem 5.72. If Hom(C2r+1 , G) is k-connected as a topological space for some
r ≥ 1, then the chromatic number of G is at least k + 4.
These results suggest the more general assertion that if Hom(F, G) is k-
connected as a topological space, then χ(G) ≥ k + χ(F ) + 1. This is, however,
false, as shown by Hoory and Linial [2005]. But the relationship between the chro-
matic numbers of F and G and the topology of the complex Hom(F, G) is mostly
unexplored.
Exercise 5.73. Prove that if G is a simple graph with maximum degree d, then
for all q ≥ d + 2 the graph Hom(G, Kq ) is connected, i.e., we can transform any
q-coloring of G into any other, changing the color of one node at a time, going
through legitimate q-colorings.
Exercise 5.74. Prove that the graph Hom(F, G) is connected if and only if the
simplicial complex Hom(F, G) is connected.
Exercise 5.75. Prove that Hom(., .) is a contravariant functor in its first
variable: every homomorphism ξ : F1 → F2 induces a simplicial map
φ̌ : Hom(F2 , G) → Hom(F1 , G). Analogously, Hom(., .) is a covariant func-
tor in its second variable.
 5.7. THE STRUCTURE OF THE HOMOMORPHISM SET 81

Exercise 5.76. Let α be an automorphism of F which has an orbit on V (F ) that

is not a stable set. Assume that G has no loops. Then the simplicial map α̌ is
fixed-point-free on the geometric realization of Hom(F, G).
Exercise 5.77. Prove that Hom(K2 , Kp ) is homotopy equivalent to the (p − 2)-
dimensional sphere.
 CHAPTER 6

Graph algebras and homomorphism functions

6.1. Algebras of quantum graphs

A quantum graph is defined as a formal linear combination of a finite number of
multigraphs with real coefficients. To be pedantic, let’s add that these coefficients
can be zero, but terms with zero coefficient can be deleted without changing the
quantum graph. Those graphs that occur with non-zero coefficient are called the
constituents of x. Quantum graphs form an infinite dimensional linear space, which
we denote by Q0 .
∑n Every graph parameter ∑nf can be extended to quantum graphs linearly: if x =
i=1 λ i F i , then f (x) = i=1 λi f (Fi ). In particular, the definition
∑n of hom(F, G)
and
∑m t(F, G) extends to quantum graphs bilinearly: if x = i=1 α i Fi and y =
j=1 β j G j , then we define
∑
n ∑
m
hom(x, y) = αi βj hom(Fi , Gj ),
i=1 j=1

and similarly for t(x, y). (Most of the time we will use linearity in the first argument
only.)
Quantum graphs are useful in expressing various combinatorial situations. For
example, for any signed graph F , we consider the quantum graph
∑ ′
(6.1) x= (−1)e(F )−|E+ | F ′ ,
F′

where the summation extends over all simple graphs F ′ such that V (F ′ ) = V (F )
and E+ ⊆ E(F ′ ) ⊆ E+ ∪ E− . By inclusion-exclusion we see that for any simple
graph G, hom(F, G) = hom(x, G) is the number of maps V (F ) → V (G) that
map positive edges onto edges and negative edges onto non-edges. The equation
hom(F, G) = hom(x, G) remains valid if G is a weighted graph (one way to see it is
to expand the parentheses in definition (5.9)). Due to these nice formulas, we will
denote the quantum graph x by F ; this will not cause any confusion.
The relationships between homomorphism numbers and injective homomor-
phism numbers, equations
∑ (5.16) and (5.18), ∑
can be expressed as follows: For every
graph G, let ZG = P G/P and M G = P µP G/P , where P ranges over all
partitions of V (G). Here the quotient graph G/P is defined by merging every
class into a single node, and adding up the multiplicities of pre-images of an edge
to get its multiplicity in G/P . Then
(6.2) hom(F, G) = inj(ZF, G), and inj(F, G) = hom(M F, G).
More generally, for any graph parameter f , we have f (M G) = f ⇓ (G). The op-
erators Z and M extend to linear operators Z, M : Q0 → Q0 . Clearly, they are
83
84 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

inverses of each other: ZM = M Z = idQ0 . (In Appendix A.1 these operators are
discussed for general lattices.)
We will see that other important facts, like the contraction/deletion relation
of the chromatic polynomial (4.5) can also be conveniently expressed by quantum
graphs (cf. Section 6.3).
For any k ≥ 0, a k-labeled quantum graph is a formal linear combination of
k-labeled graphs. We say that a k-labeled quantum graph is simple [loopless] if all
its constituents are simple [loopless].
6.1.1. The gluing algebra. Let Qk denote the (infinite dimensional) vector
space of k-labeled quantum graphs. We can turn Qk into an algebra by using the
gluing product F1 F2 introduced in Section 4.2 as the product of two generators, and
then extending this multiplication to the other elements of the algebra by linearity.
Clearly Qk is associative and commutative. The fully labeled graph Ok on [k] with
no edges is the multiplicative unit in Qk .
Every graph parameter f can be extended linearly to quantum graphs, and
defines an inner product on Qk by
(6.3) ⟨x, y⟩ = f (xy).
This inner product has nice properties, for example it satisfies the Frobenius identity
(6.4) ⟨x, yz⟩ = ⟨xy, z⟩.
Let Nk (f ) denote the kernel (annihilator) of this inner product, i.e.,
Nk (f ) = {x ∈ Qk : f (xy) = 0 ∀y ∈ Qk }.
Note that it would be equivalent to require this condition for (ordinary) k-labeled
graphs only in place of y. Sometimes we write this condition as x ≡ 0 (mod f ),
and then use x ≡ y (mod f ) if x − y ≡ 0 (mod f ). We define the factor algebra
Qk /f = Qk /Nk (f ).
Formula (6.3) still defines an inner product on Qk /f , and identity (6.4) remains
valid. While the algebra Qk is infinite dimensional, the factor algebra Qk /f is finite
dimensional for many interesting graph parameters f .
Proposition 6.1. The dimension of Qk /f is equal to the rank of the connection
matrix M (f, k). The inner product (6.3) is positive semidefinite on Qk if and only
if M (f, k) is positive semidefinite.
So if the parameter f is reflection positive, then the inner product is positive
semidefinite on every Qk ; equivalently, it is positive definite on Qk /f . It follows
that the examples in section 4.3 provide several graph parameters for which the
algebras Qk /f have finite dimension. This means that in these cases our graph
algebra Qk /f is a Frobenius algebra (see Kock [2003]). For a reflection positive
parameter, the inner product is positive definite on Qk /f , so it turns Qk /f into an
inner product space.
Example 6.2 (Number of perfect matchings). Consider the number pm(G)
of perfect matchings in the graph G. It is a basic property of this value that
subdividing an edge by two nodes does not change it. This can be expressed as
≡ (mod pm),
where the black nodes are labeled. 
 6.1. ALGEBRAS OF QUANTUM GRAPHS 85

Sometimes it will be convenient to put all k-labeled graphs into a single struc-
ture as follows. Recall the notion of partially labeled graphs from Section 4.2, and
also the notion of their gluing product. Let QN denote the (infinite dimensional)
vector space of formal linear combinations (with real coefficients) of partially labeled
graphs. We can turn QN into an algebra by using the product G1 G2 introduced
above (gluing along the labeled nodes) as the product of two generators, and then
extending this multiplication to the other elements linearly. Clearly QN is associa-
tive and commutative, and the empty graph is a unit element.
A graph parameter f defines an inner product on the whole space QN by (6.3),
and we can consider the kernel N (f ) = {x ∈ QN : ⟨x, y⟩ = 0 ∀y ∈ QN } of this
inner product. It is not hard to see that Nk (f ) = Qk ∩ N (f ).
For every finite set S ⊆ N, the set of all formal linear combinations of S-labeled
graphs form a subalgebra QS of QN . We set QS /f = {x/f : x ∈ QS }. Clearly
QS /f is a subalgebra of QN /f , and it is not hard to see that QS /f ∼= Q|S| /f . The
graph with |S| nodes labeled by the elements of S and no edges, which we denote
by OS , is a unit in the algebra QS .

6.1.2. The concatenation algebra. There is another algebra on these vector

spaces. For two 2-multilabeled multigraphs F and G, we define their concatenation
by identifying node 2 of F with node 1 of G, and unlabeling this merged node. We
denote the resulting 2-labeled graph by F ◦G. It is easy to check that this operation
is associative (but not commutative). We extend this operation linearly over Q2 .
This algebra has a ∗ (conjugate) operation: for a 2-labeled graph F , we define
F ∗ by interchanging the two labels. Clearly (F ◦G)∗ = G∗ ◦F ∗ . We can also extend
this linearly over Q2 .
Let f be a graph parameter. It is easy to see that if x ≡ 0 (mod f ) then
x∗ ≡ 0 (mod f ), so the ∗ operator is well defined on elements of Q2 /f . A further
important property of concatenation is that for any three 2-labeled graph F , G and
H,
(F ◦ G)H =∼ F (H ◦ G∗ ),
and hence
( ) ( )
(6.5) f (x ◦ y)z = f x(z ◦ y ∗ ) ,
for any three elements x, y, z ∈ Q2 . It follows that if x ≡ 0 (mod f ) then x ◦ y ≡ 0
(mod f ) for every y ∈ Q2 and thus concatenation is well defined on the elements
of Q2 /f . It is easy to see that (Q2 /f, ◦) is an associative (but not necessarily
commutative) algebra.
We can think of a 2-labeled graph as a graph having one labeled node on its
left side and one on its right side. Then concatenation means that we identify the
right labeled node of one graph with the left labeled node of another. This suggests
a generalization: Instead of a single node, we consider graphs that have k labeled
nodes on each side. Let’s say the labels are 1, . . . , k on both sides, so each label
occurs twice, once on the left and once on the right. It is convenient to allow that
one and the same node gets a left label and a right label. Such a graph will be
called (k, k)-labeled, and we denote their set by Fk,k .
We can define a multiplication on Fk,k , denoted by ◦, in which we identify each
right labeled node of the first graph with the left labeled node of the second graph
with the same label. We can take the space Qk,k of “quantum bi-labeled graphs”,
86 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

i.e., formal linear combinations of graphs in Fk,k . The graph Ok on [k] with no
edges, with its nodes labeled 1, . . . , k from both sides, is a unit in the algebra.
This algebra is associative, but not commutative. It has a “conjugate” opera-
tion, which we denote by ∗, of interchanging “left” and “right”. This is related to
multiplication through the identity (A ◦ B)∗ = B ∗ ◦ A∗ .
Given a graph parameter f defined on looped-multigraphs, we can define the
inner product of two (k, k)-labeled graphs as before: we consider them as multil-
abeled graphs (where left label i is different from right label i), form their gluing
product, and evaluate the parameter on the resulting multigraphs. (We have to
work with multilabeled graphs, since a node is allowed to have two labels. As a
consequence, the gluing product can have loops.)
A further natural generalization involves graphs with possibly different numbers
of labeled nodes on the left and on the right. Let Fk,m denote the set of multigraphs
with k labeled nodes on the left and m labeled nodes on the right. We cannot
form the product of any two graphs, but we can multiply a graph F ∈ Fk,m
with a graph G ∈ Gm,n to get a graph F ◦ G ∈ Fk,n . So bi-labeled graphs form
the morphisms of a category, in which the objects are the natural numbers. The
star operation (interchanging left and right) maps Fk,m onto Fm,k . Any graph
parameter f defines a scalar product on every Fk,m by ⟨F, G⟩ = f (F G), where F G
is defined by identifying nodes with the same left-label as well as nodes with the
same right-label in the disjoint union of F and G.
Just as above, the operations ◦, ∗, and ⟨., .⟩ extend linearly to the linear spaces
Qk,m of formal linear combinations of graphs in Fk,m . This leads us to semisimple
categories and topological quantum field theory (see Witten [1988]), which topics
are beyond the limits of this book.

6.1.3. Unlabeling. Having defined the graph algebras we need, we are going
to describe the relationship between algebras of labeled graphs using different label
sets. There is nothing terribly deep or surprising here; but it might serve as a
warm-up, illustrating how combinatorial and algebraic constructions correspond to
each other.
The unlabeling operator G 7→ [[G]]S extends to Q by linearity. We note that for
= [[[[G]]S [[H]]S ]] ∼
any two partially labeled graphs G and H, [[[[G]]S H ]] ∼ = [[G[[H]]S ]],
and hence we get the identity
(6.6) ⟨[[x]]S , y⟩ = ⟨[[x]]S , [[y]]S ⟩ = ⟨x, [[y]]S ⟩ (x, y ∈ Q).
By a similar argument we get that if S, T ⊂ N are finite sets, then
(6.7) ⟨x, y⟩ = ⟨[[x]]S∩T , [[y]]S∩T ⟩ (x ∈ QS , y ∈ QT )
One consequence of identity (6.6) is that if some x ∈ Q is congruent modulo f
to some S-labeled quantum graph y ∈ QS , then such a y can be obtained by simply
removing the labels outside S:
(6.8) x − y ∈ N (f ) =⇒ x − [[x]]S ∈ N (f ).
Indeed, for any z ∈ Q, we have
⟨x − [[x]]S , z⟩ = ⟨x, z⟩ − ⟨[[x]]S , z⟩ = ⟨y, z⟩ − ⟨[[x]]S , z⟩
= ⟨y, [[z]]S ⟩ − ⟨x, [[z]]S ⟩ = ⟨y − x, [[z]]S ⟩ = 0.
 6.1. ALGEBRAS OF QUANTUM GRAPHS 87

As a special case, we get that [[x]]S ∈ N (f ) for all x ∈ N (f ). This implies that
the operator x 7→ [[x]]S is defined on the factor algebra Q/f , and in fact it gives the
orthogonal projection of Q/f to the subalgebra QS /f . Indeed, by (6.6)
⟨[[x]]S , x − [[x]]S ⟩ = ⟨x, [[x − [[x]]S ]]S ⟩ = ⟨x, [[x]]S − [[x]]S ⟩ = 0.
Another consequence of (6.8) is that for every x ∈ Q there is a unique smallest
set S ⊂ N such that x ≡ [[x]]S (mod f ).
For the rest of this section, we assume that f is multiplicative and normalized
so that f (K1 ) = 1. (This latter condition is usually easily achieved by replacing
f (G) by f (G)/f (K1 )v(G) .)
One important consequence of this assumption is that deleting isolated nodes
(labeled or unlabeled) from a graph G does not change f (G). This implies that it
does not change G/f either. Indeed, let F denote the graph obtained from G by
deleting some isolated nodes, then for every partially labeled graph H, the products
F H and GH differ only in isolated nodes, and hence f (F H) = f (GH), showing
that F/f = G/f . In particular, every graph with no edges has the same image in
Q/f , which is the unit element of Q/f .
Lemma 6.3. Let f be a multiplicative and normalized graph parameter, and let
S ⊆ T be finite subsets of N.
(a) If S ⊆ T , then QS /f has a natural embedding into QT /f .
(b) For any two S, T ⊆ N, we have QS /f ∩ QT /f = QS∩T /f .
(c) If S ∩ T = ∅, then QS QT ∼
= QS ⊗ QT and (QS /f )(QT /f ) ∼
= QS /f ⊗ QS /f .
Proof. (a) Every S-labeled graph G can be turned into a T -labeled graph G′
by adding |T \ S| new isolated nodes, and label them by the elements of T \ S.
(This is equivalent to multiplying it by UT ). As remarked above, G − G′ ∈ N (f ),
and so G/f = G′ /f .
(b) The containment ⊇ follows from (a). To prove the other direction, we
consider any z ∈ QS /f ∩ QT /f . Then we have an x ∈ QS with x/f = z, and a
y ∈ QT with y/f = z. So x−y = x−[[y]]T ∈ N (f ), and so by (6.8), x−[[x]]T ∈ N (f ).
But we can write this as [[x]]T − [[x]]S ∈ N (f ), and then by the same reasoning
[[x]]T − [[[[x]]T ]]S = [[x]]T − [[x]]T ∩S ∈ N (f ), showing that x − [[x]]T ∩S ∈ N (f ), and
so z = x/f ∈ QS∩T /f .
(c) The first relation is trivial, since the partially labeled graphs F G, F ∈ FS• ,
G ∈ FT• are different generators of QS∪T . To prove the second, let a1 , a2 , . . . be any
basis of QS /f and b1 , b2 , . . . , any basis of QT /f . Consider the map ai ⊗ bj 7→ ai bj
(which is defined on a basis of QS /f ⊗ QT /f ), and extend it linearly to a map
Φ : QS /f ⊗ QT /f → (QS /f )(QT /f ). We show that Φ is an isomorphism between
QS /f ⊗ QT /f and (QS /f )(QT /f ).
It is straightforward to check that Φ preserves product in the algebra and also
the unit element. It is also clear that (QS /f )(QT /f ) is generated by the elements
ai bj , so Φ is surjective. To prove that Φ is injective, suppose that ∑ there are real
numbers cij of which a finite but positive number is nonzero such that i,j cij ai bj =
0. Then for every x ∈ QS /f and y ∈ QT /f , we have by multiplicativity
∑ ∑ ∑ ⟨ ∑ ⟩
cij f (xai )f (ybj ) = cij f (xyai bj ) = cij ⟨xy, ai bj ⟩ = xy, cij ai bj = 0.
i,j i,j i,j i,j
88 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

⟨ ∑ ⟩ ∑
Writing this equation as y, i,j cij f (xai )bj = 0, we see that i,j cij f (xai )bj = 0.
Since the bi are linearly independent, this means that for every 1 ≤ j ≤ m,
⟨ ∑ ⟩ ∑
x, cij ai = cij f (xai ) = 0.
i i
∑
This implies that i cij ai = 0, and since the ai are linearly independent, it follows
that cij = 0 for all i and j. 
Corollary 6.4. For every multiplicative graph parameter f with finite rank, r(f, k)
is a supermultiplicative function of k in the sense that
r(f, k + l) ≥ r(f, k)r(f, l).
Proof. It follows from Lemma 6.3(c) that for any two disjoint finite sets S and
T there is an embedding
(6.9) QS /f ⊗ QT /f ,→ QS∪T /f.
Considering the dimensions, the assertion follows. 

Exercise 6.5. Prove that if all nodes of a simple graph F are labeled, then both
F and the quantum graph Fb introduced above are idempotent in the algebra of
simple partially labeled graphs: F 2 = F and Fb2 = Fb.
Exercise 6.6. Let f be a graph parameter for which r(f, 2) = r is finite.
(a) Prove that every path labeled at its endpoints can be expressed, modulo f , as
a linear combination of paths of length at most r.
(b) Prove that a 2-labeled m-bond B m•• can be expressed, modulo f , as a linear
combination of 2-labeled k-bonds with k ≤ r − 1.
(c) A series-parallel graph is a 2-labeled graph obtained from K2•• by repeated
application of the gluing and concatenation operations. Prove that every series-
parallel graph can be expressed, modulo f , as a linear combination of series-
parallel graphs with at most 2r−1 edges.

6.2. Reflection positivity

In this section we assume that f is reflection positive, multiplicative, normalized
and has finite connection rank. We are going to prove that the gluing algebras
(modulo f ) have a very tight structure.
6.2.1. The idempotent basis. Let S be a finite subset of N. If f is reflection
positive and the dimension of QS /f is a finite number, then the factor algebra
QS /f is a finite dimensional commutative Frobenius algebra: it has a commutative
and associative product as well as a positive definite inner product, related by the
Frobenius identity (6.4).
This implies that QS /f has a very simple structure. Every element x ∈ QS /f
defines a linear transformation Ax : QS /f → QS /f by Ax y = xy. Clearly x 7→ Ax
is an algebra homomorphism, and the fact that the inner product is definite on
QS /f implies that x 7→ Ax is injective. Commutativity and the Frobenius identity
imply that Ax is symmetric, and that any two transformations Ax commute. This
implies that there is an orthonormal basis in which all the Ax are simultaneously
diagonal. Counting dimensions shows that every diagonal matrix is of the form Ax ,
and so QS /f is isomorphic with the algebra of diagonal matrices. Another way of
saying this is that QS /f is isomorphic to Rm endowed with the coordinate-wise
product and the usual inner product (where m = dim(QS /f ) = r(f, k)).
 6.2. REFLECTION POSITIVITY 89

The algebra elements corresponding to the standard basis vectors form a

(uniquely determined) basis BS = {pS1 , . . . , pSr } such that (pSi )2 = pSi (the basis
elements are idempotent in the algebra) and pSi pSj = 0 for i ̸= j. We call this the
idempotent basis of QS /f .
If p ∈ Q/f is any nonzero idempotent, then
f (p) = f (pp) = ⟨p, p⟩ > 0.
In particular, > 0. Note, however, that f (pSi ) ̸= 1 in general, so the algebra
f (pSi )
isomorphism between QS /f and Rm does not preserve the inner product.
This purely algebraic construction carries a lot of combinatorial information,
as we shall see. But first, let us work out an example.
Example 6.7 (Eulerian property). Let us compute the algebras associated with
Eul(G), the indicator function of G being eulerian (example 4.21). We know that
this is a homomorphism function into a 2-node weighted graph again, hence it is
reflection positive and r(Eul, k) ≤ 2k , but this last inequality will not hold with
equality for k ≥ 1. The space Q0 /Eul is 1-dimensional. To determine Q1 /Eul, we
note that for two 1-labeled graphs G1 and G2 , the product G1 G2 is eulerian if and
only if both G1 and G2 are eulerian (G1 and G2 cannot have a single node with odd
degree!), and so Eul(G1 G2 ) = Eul(G1 )Eul(G2 ) holds for 1-labeled graphs as well.
This shows that Q1 /Eul is also 1-dimensional.
Next, for general k, let Odd(G) denote the set of nodes of G with odd degree.
Clearly G ≡ 0 (mod Eul) for any graph with Odd(G) ̸⊆ [k]. Since |Odd(G)| is
even, the set Odd(G) is uniquely determined by the intersection Odd′ (G) = [k −
1] ∩ Odd(G). For two k-labeled graphs G1 and G2 , the product G1 G2 is eulerian
if and only if Odd′ (G1 ) = Odd′ (G2 ) (which implies that the unlabeled nodes have
even degree). Furthermore, Odd′ (G1 G2 ) = Odd′ (G1 )△Odd′ (G2 ), and hence G 7→
Odd′ (G) induces an algebra isomorphism between Qk /Eul and the group algebra
of Z2k−1 . Hence r(Eul, k) = 2k−1 for k ≥ 1.
The idempotents of a finite abelian group are determined by its characters,
which in this simple case means that they are indexed by subsets S ⊆ [k − 1], and
the idempotent is the group algebra element
1 ∑
pS = k−1 (−1)|S∩X| X.
2
X⊆[k−1]

The set Odd(G) can be expressed in this basis by discrete Fourier inversion:
1 ∑
Odd(G) = k−1 (−1)|S∩X| pX .
2
X⊆[k−1]

It follows that
( )
G 7→ (−1)|S∩Odd(G)| : S ⊆ [k − 1]
k−1
defines an algebra isomorphism between Qk /Eul and R2 . 
For two idempotents p and q in Q/f , we say that q resolves p, if pq = q. It is
clear that this relation is transitive.
Lemma 6.8. Let r be any idempotent element of QS /f . Then r is the sum of
those idempotents in BS that resolve it.
90 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

∑
Proof. Indeed, we can write r = p∈BS µp p with some scalars µp . Using that
r is idempotent, we get that
∑ ∑
r = r2 = µp µp′ pp′ = µ2p p,
p,p′ ∈BS p∈BS

which shows that = µp for every p, and so µp ∈ {0, 1}. So r is the sum of some
µ2p
subset X ⊆ BS . It is clear that rp = p for p ∈ X and rp = 0 for p ∈ BS \ X, so X
consists of exactly those elements of BS that resolve q. 
As a special case, we see that
∑
(6.10) u= p
p∈BS

is the unit element of QS (this is the image of the edgeless graph US ), and also the
unit element of the whole algebra Q.
Lemma 6.9. Let S ⊂ T be two finite sets. Then every q ∈ BT resolves exactly one
element of BS .
Proof. We have by (6.10) that
∑ ∑ ∑
u= p= q,
p∈BS p∈BS q∈BT
q resolves p

and also ∑
u= q,
q∈BT
so by the uniqueness of the representation we get that every q must resolve exactly
one p. 
f (q)
Lemma 6.10. If p ∈ BS and q resolves p, then [[q]]S = f (p) p.

Proof. Clearly ff (p)

(q)
p ∈ QS /f . Furthermore,
⟨ f (q) ⟩ f (q) f (q)
q− p, p = f (qp) − f (p2 ) = f (q) − f (p) = 0,
f (p) f (p) f (p)
and for every other basic idempotent p′ ∈ BS , we have
⟨ f (q) ⟩ f (q) f (q)
q− p, p′ = f (qp′ ) − f (pp′ ) = f (qpp′ ) − f (pp′ ) = 0.
f (p) f (p) f (p)
This shows that ff (p)
(q)
p is the orthogonal projection of q to QS /f . Since [[q]]S has
the same characterization, the lemma follows. 
Lemma 6.11. Let S, T ⊂ N be finite sets, let p ∈ BS∩T , and let q ∈ BS resolve p.
Then for any x ∈ QT /f we have f (p)f (qx) = f (q)f (px).
Indeed, by Lemma 6.10 and (6.7),
f (q)
f (qx) = f ([[q]]S∩T x) = f (px).
f (p)
Lemma 6.12. If both idempotents q ∈ BS and r ∈ BT resolve the same idempotent
p ∈ BS∩T , then qr ̸= 0.
 6.2. REFLECTION POSITIVITY 91

Indeed, by Lemma 6.11,

f (q) f (q)
f (qr) = f (pr) = f (r) > 0.
f (p) f (p)
Let S ⊂ N and p ∈ BS . For u ∈ N \ S, let q1u , . . . , qD
u
denote the elements
of BS∪{u} resolving p. Note that for u, v ∈ N \ S, there is a natural isomorphism
between QS∪{u} /f and QS∪{v} /f (induced by the map that fixes S and maps u
onto v), and we may choose the labeling so that qiu corresponds to qiv under this
isomorphism.
Let T ⊃ S and V = T \ S. For every map φ : V → [D], let
∏
v
(6.11) qφ = qφ(v) .
v∈V

Lemma 6.13. The algebra elements qφ are nonzero idempotents in QT /f such

that qφ qψ = 0 if φ ̸= ψ.
It is clear that the qφ are idempotents. By Lemma 6.11,
∏ ( ∏ f (qφ(v)
v
))
(6.12) f (qφ ) = f ( v
qφ(v) )= f (p) ̸= 0,
f (p)
v∈V v∈V

and so qφ ̸= 0. Finally, if φ ̸= ψ, then there is a v ∈ V such that φ(v) ̸= ψ(v), and

v v
then qφ(v) qψ(v) = 0, which implies that qφ qψ = 0.
If p ∈ BS and S ⊂ T , |T | = |S| + 1, then the number of elements in BT that
resolve p will be called the degree of p, and denoted by deg(p). Obviously this value
is independent of which (|S| + 1)-element superset T of S we are considering.
Lemma 6.14. If S ⊂ T , and q ∈ BT resolves p ∈ BS , then deg(q) ≥ deg(p).
It suffices to show this in the case when |T | = |S| + 1. Let T = S ∪ {u} and
v = N\T . Let q1u , . . . , qD
u
denote the elements of BT resolving p, where (say) q = q1u ,
v v
and let q1 , . . . , qD be the basic idempotents in BS∪{v} resolving p. Then by Lemma
6.13, the elements q1u qiv , i ∈ [D] are nonzero idempotents in QS∪{u,v} /f resolving
q, such that the product of any two of them is 0. Writing every such q1i as a sum
of basic idempotents we see that deg(q) ≥ D.

6.2.2. Homomorphisms into weighted graphs. After all these prepara-

tions, we are able to complete the proof of Theorem 5.54. Our goal is to construct
a weighted graph H for which f (G) = hom(G, H) for every loopless multigraph G.
The non-degenerate case. We start with sketching the proof in the non-
degenerate case, when dim(Qk /f ) = q k for all k. (This is in fact the generic case,
in the sense that it occurs with probability 1 if f = hom(., H) for a weighted graph
H with randomly chosen edge and nodeweights; see Section 6.4.1.) This implies
that the embedding QS /f ⊗ QT /f ,→ QS∪T /f in (6.9) is an isomorphism. In this
case, the idempotent bases of the algebras Qk /f are easy to construct explicitly: if
p1 , . . . , pq is the idempotent basis of Q1 /f , then the elements pi1 ⊗· · ·⊗pik (ij ∈ [q])
form the idempotent basis of Qk /f .
We can define a weighted complete graph H on [q] as follows: let αi = f (pi ) =
f (p2i ) > 0 and define βij by expressing the graph k2 (a single edge with both nodes
92 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

labeled) in the idempotent basis:

∑
(6.13) k2 = βij (pi ⊗ pj )
i,j∈[q]

To show that the weighted graph H obtained this way satisfies f (G) =
hom(G, H) for any multigraph G, we may assume that V (G) = [k] and all nodes of
G are labeled. Then we can write
∏
(6.14) G= Kuv ,
uv∈E(G)

where Kuv is the graph on k labeled nodes, with a single edge connecting u and v.
Defining pφ = pφ(1) ⊗ · · · ⊗ pφ(k) for φ : [k] → [q], the k-labeled quantum graphs
pφ form a basis of Qk /f consisting of idempotents, and hence
∑
(6.15) Ok = pφ .
φ: [k]→[q]

By the definition of the βij , we have

(6.16) Kuv pφ = βφ(u)φ(v) pφ .
We want to evaluate f (G) = ⟨G, Ok ⟩. Substituting from (6.14) and (6.15), and
using (6.16) and the Frobenius identity (6.4) repeatedly, we get that
∑ ∏ ∏
f (G) = βφ(u)φ(v) αφ(u) = hom(G, H).
φ: [k]→[q] uv∈E(G) u∈V (G)

The general (and degenerate) case takes more work, but we are ready to give
it now.
Proof of Theorem 5.54. The idea is that we find a basic idempotent p ∈ BS
for a sufficiently large finite set S ⊆ N, with the property that the subalgebra pQ/f
behaves like the whole algebra behaved in generic case. So the idempotent bases in
it, and from these the weighted graph H, can be constructed explicitly.
Bounding the expansion. If a basic idempotent p ∈ BS has degree D, then by
Lemma 6.14, there are D basic idempotents in BT with |T | = |S| + 1 with degree
≥ D that resolve p. Hence if |T | ≥ |S|, then the dimension of QT is at least D|T \S| .
It follows that the degrees of basic idempotents are bounded by q. Let us choose
S and p ∈ BS so that D = deg(p) is maximum degree. Then it follows by Lemma
6.14 that all basic idempotents resolving p have degree exactly D.
Describing the idempotents. Let us fix a set S and a basic idempotent p ∈ BS
with maximum degree D. For u ∈ N \ S, let q1u , . . . , qDu
denote the elements of
BS∪{u} resolving p.
We can describe, for a finite set T ⊃ S, all basic idempotents in BT that resolve
p. Let V = T \ S, and for every map φ : V → {1, . . . , D}, let
∏
v
(6.17) qφ = qφ(v) .
v∈V
Note that by Lemma 6.11,
∏ ( ∏ f (qφ(v)
v
))
(6.18) f (qφ ) = f ( v
qφ(v) ) = f (p) ̸= 0,
f (p)
v∈V v∈V
and so qφ ̸= 0.
 6.2. REFLECTION POSITIVITY 93

Claim 6.15. The basic idempotents in QT /f resolving p are exactly the algebra
elements of the form qφ , φ ∈ {1, . . . , D}V .

We prove this by induction on |T \ S|. For |T \ S| = 1 the assertion is trivial.

Suppose that |T \ S| > 1. Let u ∈ T \ S, U = S ∪ {u} and W = T \ {u}; thus
U ∩ W = S. By the induction hypothesis, the basic idempotents in BW resolving
p are elements of the form qψ (ψ ∈ {1, . . . , D}V \{u} ).
Let r be one of these. By Lemma 6.12, rqiu ̸= 0 for any 1 ≤ i ≤ D, and clearly
resolves r. We can write rqiu as the sum of basic idempotents in BT , and it is easy
to see that these also resolve r. Furthermore, the basic idempotents occurring in
the expression of rqiu and rqju (i ̸= j) are different. But r has degree D, so each
rqiu must be a basic idempotent in BT itself.
Since the sum of the basic idempotents rqiu (r ∈ BW,p , 1 ≤ i ≤ D) is p, it
follows that these are all the elements of BT,p . This proves the Claim.
It is immediate from the definition that an idempotent qφ resolves qiv if and
only if φ(v) = i. Hence it also follows that
∑
(6.19) qiv = qφ .
φ: φ(v)=i

Constructing the target graph. Now we can define H as follows. Let H be the
looped complete graph on V (H) = {1, . . . , D}. We have to define the node weights
and edge weights.
Fix any u ∈ N \ S. For every i ∈ V (H), let αi = f (qiu )/f (p) be the weight of
the node j. Clearly αi > 0.
Let u, v ∈ N\S, v ̸= u, and let W = S ∪{u, v}. Let Kuv denote the graph on W
which has only one edge connecting u and v, and let kuv denote the corresponding
element of QW . We can express pkuv as a linear combination of elements of BW,p
(since for any r ∈ BW \ BW,p one has rp = 0 and hence rpku,v = 0):
∑
pkuv = βij qiu qjv .
i,j

This defines the weight βij of the edge ij. Note that βij = βji , since pkuv = pkvu .

Verifying the target graph. We prove that this weighted graph H gives the
right homomorphism function: f (G) = hom(G, H) for every multigraph G. By
(6.19), we have for each pair u, v of distinct elements of V (G)
∑ ∑ ∑ ∑
pkuv = βi,j qiu qjv = βi,j qφ = βφ(u),φ(v) qφ .
i,j∈V (H) i,j∈V (H) φ: φ(u)=i φ∈V (H)V
φ(v)=j

Consider any V -labeled graph G with V (G) = V ⊆ N \ S, and let g be the

corresponding element of Q/f . Then
∏ ∏ ( ∑ )
pg = pkuv = βφ(u),φ(v) qφ
uv∈E(G) uv∈E(G) φ∈V (H)V
∑ ( ∏ )
= βφ(u),φ(v) qφ .
φ:V →V (H) uv∈E(G)
94 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

Since p ∈ QS /f and g ∈ QV /f where S ∩ V = ∅, we have f (p)f (g) = f (pg), and

so by (6.12),
∑ ( ∏ )
f (p)f (g) = f (pg) = βφ(u),φ(v) f (qφ )
φ∈V (H)V uv∈E(G)
∑ ( ∏ )( ∏ )
= βφ(u),φ(v) αφ(v) f (p) = hom(G, H)f (p).
φ:V →V (H) uv∈E(G) v∈V (G)

The factor f (p) > 0 can be cancelled from both sides, completing the proof of the
theorem. 

6.3. Contractors and connectors

We study the existence and properties of two special elements in graph algebras.
These will serve as important building blocks for other constructions, like a more
explicit description of the idempotent basis.

6.3.1. Contractors and connectors for general graph parameters. In

the algebra Q2 of 2-multilabeled graphs, multiplication by the single node with
two labels (denoted by K1•• ) results in identifying nodes labeled 1 and 2. In the
factor algebra modulo a graph parameter, it may be the case that multiplication by
some other graph has essentially the same result. For example, for the chromatic
polynomial chr(., x), the contraction-deletion identity (A.10) can be written like
this:
(6.20) K1•• ≡ O2•• − K2•• (mod chr(., x)).
or in pictograms
12
(6.21) • ≡ − (mod chr(., x)).
We say that the 2-labeled quantum graph on the right side is a contractor for the
graph parameter chr(., x). Starting with any multilabeled quantum graph, we can
apply this identity repeatedly to construct a simply labeled quantum graph which
represents the same element in Q/f .
A consequence of the contraction-deletion relation is the identity

(6.22) ≡ − (mod chr(., x)),

which does not involve multiple labels. One way to read this is that the edge (with
both endnodes labeled) can be replaced by the difference of two simple graphs in
which the two labeled nodes are nonadjacent. We say that the 2-labeled quantum
graph on the right is a connector for the graph parameter chr(., x). One important
consequence of this identity is that applying it repeatedly, we can eliminate all edge
multiplicities.
To state the definition formally, let us say that a quantum graph x is a proper
expansion of a partially labeled graph F modulo a graph parameter f , if x ≡ F
(mod f ), and no constituent of x is of the form F G for some partially labeled graph
G. Then a contractor is a proper expansion of K1•• , and a connector is a proper
expansion of K2•• . Also, the remarks after (6.20) and (6.22) can be formalized like
this:
 6.3. CONTRACTORS AND CONNECTORS 95

Proposition 6.16. A graph parameter f has a contractor if and only if every

multilabeled quantum graph is congruent modulo f to a simply labeled quantum
graph. A graph parameter f has a simple connector if and only if every simply
labeled quantum graph is congruent modulo f to a simply labeled simple quantum
graph with no edge connecting the labeled nodes.
We can put the notion of a contractor in a different context. Let Fkstab denote
the set of k-labeled multigraphs in which the labeled nodes form an independent
(stable) set, and let Qstab
k denote the subalgebra of Qk generated by them. For
a 2-labeled graph F ∈ F2stab , let F ′ denote the graph obtained by identifying the
two labeled nodes and labeling it by 1. (So F ′ is obtained from the product F K1••
by removing the label 2.) The map F 7→ F ′ maps 2-labeled graphs to 1-labeled
graphs. We can extend it linearly to get an algebra homomorphism x 7→ x′ from
Qstab
2 into Q1 .
The map x 7→ x′ does not in general preserve the inner product or even its
kernel; we say that the graph parameter f is contractible, if for every x ∈ Qstab
2 ,
x ≡ 0 (mod f ) implies x′ ≡ 0 (mod f ); in other words, x 7→ x′ factors to a linear
map Qstab
2 /f → Q1 /f . With this notation, z ∈ Q2 is a contractor for f if and only
′
if for every x ∈ Qstab
2 , we have f (xz) = f (x ).
Contractors also relate to the algebra of concatenations:
Proposition 6.17. A contractor for f is the multiplicative identity for the opera-
tion ◦ modulo f .
Proof. We have to verify that if z is a contractor, then
(6.23) z◦x≡x (mod f )
( )
for all x ∈ Q2 . This is equivalent to f (z ◦ x)y = f (xy) for all x, y ∈ Q2 . Using
(6.5) and that in x ◦ y the labeled nodes are nonadjacent, we obtain:
( ) ( ) ( )
f (z ◦ x)y = f z(y ◦ x∗ ) = f (y ◦ x∗ )′ = f (xy),
which proves (6.23). 

Note that in every constituent of x ◦ y the labeled nodes are nonadjacent for
all x, y ∈ Q2 . It follows that if the algebra (Q2 /f, ◦) has a multiplicative identity
(in particular, if it has a contractor), then every y ∈ Q2 /f can be represented by a
2-labeled quantum graph with nonadjacent labeled nodes.
While the existence of a contractor, the existence of a connector, and con-
tractibility are three different properties of a graph parameter, there is some con-
nection, as expressed in the following propositions.
Proposition 6.18. If a graph parameter has a contractor, then it is contractible.
Proof. Let w be a contractor for f . Suppose that x ∈ Q2 satisfies x ≡ 0
(mod f ), and let y ∈ Q1 . Choose a z ∈ Q2 such that z ′ = y. Then
( ) ( ) ( )
f (x′ y) = f (x′ z ′ ) = f (xz)′ = f (xz)w = f x(zw) = 0,
showing that x′ ≡ 0 (mod f ). 

Proposition 6.19. If a graph parameter has a contractor, then it has a connector;

if it has a simple contractor, then it has a simple connector.
96 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

Proof. Let z be a contractor. We claim that z ◦ P2•• is a connector. Indeed,

for any 2-labeled graph
( G, we)have ((z ◦ P2•• )G ∼
= )z(G ◦ P2•• ), and by the definition
of a contractor, f z(G ◦ P2 ) = f K1 (G ◦ P2 ) = f (P2•• G). So z ◦ P2•• ≡ P2••
•• ••

(mod f ), and thus z ◦ P2•• is a connector. The second assertion is trivial by the
same construction. 
Proposition 6.20. If f is contractible, has a connector, and r(f, 2) is finite, then
f has a contractor.
Proof. Since ⟨x, y⟩ = f (xy) is a symmetric (possibly indefinite) bilinear form
that is not singular on Q2 /f , there is a basis p1 , . . . , pN in Q2 /f such that f (pi pj ) =
0 if i ̸= j and f (pi pi ) ̸= 0. By the assumption that f has a connector, we may
represent this basis by quantum graphs with nonadjacent labeled nodes; then the
contracted quantum graphs p′i have no loops. Let
∑N
f (p′i )
z= pi .
i=1
f (p2i )
We claim that z is a contractor. Indeed, let x ∈ Q2 be a quantum graph with
∑N
nonadjacent labeled nodes, and write x ≡ i=1 ai pi (mod f ). Then we have
∑
N
f (p′i ) ∑ N
f (xz) = ai
2 f (p2i ) = ai f (p′i ).
i=1
f (pi ) i=1
∑N
On the other hand, contractibility implies that x ≡ i=1 ai p′i (mod f ), and so
′

∑
N
f (x′ ) = ai f (p′i ) = f (xz). 
i=1

Proposition 6.21. If M (f, 2) is positive semidefinite and has finite rank r, and
f is contractible, then f has a connector whose constituents are paths of length at
most r + 1.
Proof. Since Q2 /f is finite dimensional, there is a linear dependence between
••
P2•• , P3•• , . . . , Pr+2 in Q2 /f . Hence there is a (smallest) k ≥ 2 such that Pk•• can
be expressed as
∑r
(6.24) Pk•• ≡ ••
ai Pk+i (mod f )
i=1
with some real numbers
∑r a1 , . . . ••
, ar . The assertion is equivalent to saying that k = 2.
Let x = P2•• − i=1 ai P2+i . Then (6.24) can be written as x ◦ Pk−1 ••
≡ 0
••
(mod f ). If k = 3, then this implies that x ◦ P2+i ≡ 0 (mod f ) for all( i ≥ 0, and )
hence x ◦ x ≡ 0 (mod f ). Using contractibility we obtain that 0 = f (x ◦ x)′ =
f (x2 ). Now semidefiniteness of M (f, 2) shows that x ≡ 0 (mod f ). So (6.24) holds
with k = 2 as well, a contradiction.
Suppose that k > 3, then
•• 2 •• ••
(x ◦ Pk−2 ) = (x ◦ Pk−1 )(x ◦ Pk−3 ) ≡ 0 (mod f ),
and so by the assumption that M (f, 2) is positive semidefinite, we get that x ◦
••
Pk−2 ≡ 0 (mod f ), which contradicts the minimality of k again. 
The following statement is a corollary of Propositions 6.20 and 6.21.
 6.3. CONTRACTORS AND CONNECTORS 97

Corollary 6.22. If M (f, 2) is positive semidefinite and has finite rank, and f is
contractible, then f has a contractor.
We conclude this section with a number of examples of connectors and con-
tractors.
Example 6.23 (Perfect matchings). Recall that pm(G) denotes the number of
perfect matchings in the graph G. We have seen that rk(pm, k) = 2k is exponentially
bounded, but pm is not reflection-positive, and thus pm(G) cannot be represented
as a homomorphism function.
On the other hand: pm has a contractor: a path of length 2, and also a con-
nector: a path of length 3. 
Example 6.24 (Number of triangles). The graph parameter hom(K3 , .) has no
connector. Indeed, suppose that x ∈ Q2 is a connector, then we must have
hom(K3 , xP3•• ) = hom(K3 , K2•• P3•• ) = hom(K3 , K3 ) = 6,
and also
hom(K3 , xP4•• ) = hom(K3 , K2•• P4•• ) = hom(K3 , C4 ) = 0.
On the other hand,
hom(K3 , xP3•• ) = hom(K3 , xP4•• ),
since x has with nonadjacent labeled nodes, so no homomorphism from K3 touches
the edges of the P3•• factor. This contradiction shows that hom(K3 , .) has no
connector.
A similar argument shows that hom(K3 , .) is not contractible (and so it has no
contractor). 
Example 6.25 (S-Flows). The number flo(G) of flows on G with values from a
given subset of a finite abelian group can be described as a homomorphism function
(Example 5.16). It has a trivial connector, a path of length 2 (which is an algebraic
way of saying that if we subdivide an edge, then the flows don’t change essentially).
In the case of nowhere-0 flows, K2•• + U2•• is a contractor (which amounts to the
contraction-deletion identity for the flow polynomial). In general, it is more difficult
to describe the contractors, but it is possible (Garijo, Goodall and Nešetřil [2011]).

Example 6.26 (Density in a random graph). Recall the multigraph parameter
simp
from Example 5.60: f (G) = pe(G ) (0 < p < 1), which is reflection positive, mul-
tiplicative, and has finite connection rank. This parameter has neither a contractor
nor a connector; it is not even contractible. We have
≡ (mod f ),
but identifying the labeled nodes produces a pair of parallel edges in the first graph
but not in the second, so they don’t remain congruent. 
−
→
Example 6.27 (Eulerian orientations). Recall that eul(G) denotes the number
−
→
of eulerian orientations of the graph G. We have seen that the graph parameter eul
is reflection positive, but has infinite connection rank (so it is not a homomorphism
function). Similarly as in Example 6.25, a path of length 2 is a connector. Fur-
thermore, this graph parameter is contractible, but has no contractor (see Exercise
6.33). 
98 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

6.3.2. Contractors and connectors for homomorphism functions.

Many graph parameters have contractors and/or connectors, as we have seen.
Our next theorem asserts this for homomorphism functions f = hom(., H) for any
weighted graph H. It is easy to check from the definitions that a 2-labeled quantum
graph z is a contractor for hom(., H) if and only if
(6.25) homij (z, H) = (1/αi )1(i = j)
for every i, j ∈ V (H). It is a connector for hom(., H) if and only if z is simple, it
has nonadjacent labeled nodes, and
(6.26) homij (z, H) = βij
for every i, j ∈ V (H).
Theorem 6.28. Let f = hom(., H) for some weighted graph H. Then f has a
simple contractor and a simple connector.
We can in fact construct connectors and contractors of a rather simple form.
To state the result, we define classes of 2-labeled multigraphs as follows. We start
with K = {K2•• }. The (gluing) products of at most a of members of K form the
class K(a) (so these are 2-labeled bonds). Concatenations of at most b bonds form
the class K(a, b). The gluing products of at most c members of K(a, b) form the
class K(a, b, c). Concatenations of at most d members of K(a, b, c) form the class
K(a, b, c, d) etc (Fig. 6.1).

Figure 6.1. Graphs in classes K

Supplement 6.29. Let q = v(H).

(a) The graph parameter f = hom(., H) has a connector whose constituents are
paths Pk•• with 3 ≤ k ≤ q + 3.
(b) The graph parameter f = hom(., H) has a contractor whose constituents
are in the class K(q 2 − 1, 2, q 2 − 1, 2, q). In particular, the number of nodes in this
contractor is at most 2q 3 .
All graphs in the classes K(a, b, . . . ) are series-parallel. So in particular, the
graph parameter f = hom(., H) has a contractor whose constituents are series-
parallel (Lovász and Szegedy [2009]). The Supplement above is a refinement of this
statement. For us, the bound on the number of nodes will be more important than
the structure. If we get rid of the parallel edges in the contractor by replacing each
edge by the simple connector constructed above, every constituent will still be a
series-parallel graph, and the number of nodes in any constituent will be bounded
by 2q 6 .
 6.3. CONTRACTORS AND CONNECTORS 99

Proof of Theorem 6.28 and Supplement 6.29. We may assume that H is

twin-free, since identifying twins does not change the graph parameter hom(., H).
(a) To construct a simple connector, let B = (βij ) be the (weighted) adjacency
√ √
matrix of H, and let D = diag( α1 , . . . , αm ). Let λ1 , . . . , λt be the nonzero
eigenvalues of the matrix DBD (which are real as DBD is symmetric), and consider
∏t
the polynomial ρ(z) = z i=1 (1 − z/λi ). Then ρ(DBD) = 0 (since the eigenvalues
of DBD are roots of ρ). Since the constant term in ρ(z) is 0 and the linear term is
z,
∑tthis expressess DBD as a linear combination of higher powers of DBD: DBD =
s=2 as (DBD) , or

∑
t
(6.27) B= as (BD2 )s−1 B.
s=2

For all i, j ∈ V (H), we have

( )
(6.28) homij (Ps•• , H) = (BD2 )s−2 B ij .
∑t ••
Let y = s=2 as Ps+1 , then (6.27) and (6.28) imply that y is a connector, with the
structure described in the Supplement.
(b) The existence of a contractor, a 2-labeled quantum graph z satisfying (6.25),
follows by Theorem 6.38. We can replace each edge of the contractor by a simple
connector, to make it simple.
The construction of a contractor of the special type described in the Supplement
takes more work. We consider the values βij = homij (K2•• , H). Replacing K2 by the
•• m
m-bond, we get homij (Bm , H) = βij . (Note that this relation also holds for m = 0.)
Hence for every real polynomial p ∈ R[x] we get a 2-labeled quantum graph yp that
is a quantum bond (a linear combination of bonds), such that homij (yp , H) =
p(βij ). Since every real function equals to a polynomial of degree less than q 2 on
the q 2 values βij , we get that for every real function g there is quantum graph yg
with constituents from K(q 2 − 1), such that homij (yg , H) = g(βij ).
Let K denote the field obtained from Q by adjoining the node weights of H.
We choose the function g above so that the values γij = g(βij ) are algebraically
independent over K for different values of βij .
Next, we form the concatenation ug = yg ◦ yg∗ (which is a linear combina-
tion of graphs in K(q 2 − 1, 2)), and consider the values homij (ug , H). We claim
that a “diagonal” value homii (ug , H) can never be equal to an “off-diagonal” value
homjk (ug , H), i ̸= j. Indeed, if they are equal, then we have
∑
q ∑
q
2
αr γir = αr γjr γkr .
r=1 r=1

Since different γ’s are algebraically independent over K (which contains the coef-
ficients αr ), the two sides must be equal formally. In particular, every product
γjr γkr must occur on the left side, which implies that γjr = γkr . By the definition
of γ, this means that βjr = βkr for every r, and so nodes j and k are twins, which
was excluded.
It follows that we can find a polynomial h of degree at most q 2 − 1 such that
h(γii ) = 1 but h(γij ) = 0 if i ̸= j. Hence we get a quantum graph w such that
(6.29) homij (w, H) = 1(i = j).
100 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

Every constituent of w is the (gluing) product of at most q 2 −1 graphs in K(q 2 −1, 2),
so it is in class K(q 2 − 1, 2, q 2 − 1).
Consider the quantum graph w′ = w ◦ w. The constituents of w′ are graphs in
class K(q 2 − 1, 2, q 2 − 1, 2). Using (6.29), we get
∑
(6.30) homij (w′ , H) = αk homik (w, H)homkj (w, H) = αi 1(i = j).
k

Expressing the function i(t) = 1t 1(t ̸= 0) on the values of homij (w′ , H) by a poly-
nomial of degree at most q as above, we can construct a quantum graph z in class
K(q 2 − 1, 2, q 2 − 1, 2, q) satisfying (6.25). 

Using the notion of a contractor, we can give the following characterization of

homomorphism functions, which does not involve the finiteness of the connection
rank.

Theorem 6.30. A graph parameter f can be represented in the form f = hom(., H)

for some weighted graph H if and only if it is multiplicative, reflection positive and
has a contractor.

Proof. The necessity of the conditions follows by Theorem 6.28.

To prove the sufficiency of the conditions, it suffices to prove that there exists a
q > 0 such that r(f, k) ≤ q k for all k ≥ 0, and then invoke Theorem 5.54. Note that
reflection positivity is used twice: the existence of a contractor does not in itself
imply an exponential bound on the connection rank (cf. our introductory example
of the chromatic polynomial).
Let g0 be a contractor for f ; we show that q = f (g02 ) provides the appropriate
upper bound on the connection rank. Since f is multiplicative, we already know
this for k = 0.
We may normalize f so that f (K1 ) = 1. If the conclusion is false (with q =
f (g02 )) then for some integer k > 0 we have (possibly infinite) r(f, k) > q k and
hence for N = ⌊q k + 1⌋ there are mutually orthogonal unit vectors q1 , . . . , qN in the
algebra Qk /f . Let qi ⊗ qi denote the (2k)-labeled quantum graph obtained from
2k labeled nodes by attaching a copy of qi at {1, . . . , k} and another copy of qi at
{k + 1, . . . , 2k}. Let h denote the (2k)-labeled quantum graph obtained from 2k
labeled nodes by attaching a copy of g0 at {i, k + i} for each i = 1, . . . , k. Consider
the quantum graph
∑
N
x= qi ⊗ qi − h.
i=1

By reflection positivity, we have f (x2 ) ≥ 0. But

∑
N ∑
N ∑
N
f (x2 ) = ⟨qi ⊗ qi , qj ⊗ qj ⟩ − 2 ⟨qi ⊗ qi , h⟩ + ⟨h, h⟩.
i=1 j=1 i=1

Here ⟨qi ⊗ qi , qj ⊗ qj ⟩ = ⟨qi , qj ⟩2 = 1(i = j), and so

∑
N ∑
N
⟨qi ⊗ qi , qj ⊗ qj ⟩ = N.
i=1 j=1
 6.4. ALGEBRAS FOR HOMOMORPHISM FUNCTIONS 101

Furthermore, by the definition of g0 and h, we have

∑
N ∑
N
⟨qi ⊗ qi , h⟩ = ⟨qi , qi ⟩ = N.
i=1 i=1

Finally, by the definition of h and the multiplicativity of f , we have ⟨h, h⟩ = f (g02 )k .

Thus f (x2 ) ≥ 0 implies that N ≤ f (g02 )k = q k , a contradiction. 

Exercise 6.31. Let z be a contractor for a graph parameter f , and let F be a

k-labeled graph. Let us delete an edge 1i from F , and add the edge 2i, to obtain
another k-labeled graph F ′ . Prove that zF ≡ zF ′ (mod f ).
Exercise 6.32. Let z be a contractor for a graph parameter f , and let z = O2 −z.
Construct a 3-labeled quantum graph x by gluing a copy of z on nodes 1 and 2,
a copy of z on nodes 2 and 3, and a copy of z on nodes 1 and 3. Prove that x ≡ 0
(mod f ).
−→
Exercise 6.33. Prove that the number of eulerian orientations eul is contractible,
but has no contractor.
Exercise 6.34. For every multigraph G, let Ğ be obtained by repeatedly con-
tracting edges with multiplicity larger than 1 until a simple graph is obtained,
and define f (G) = hom(K3 , Ğ). (a) Prove that Ğ is uniquely determined. (b)
Show that f has a contractor. (c) Prove that f has no connector.

6.4. Algebras for homomorphism functions

We have proved that reflection positive graph parameters with exponentially
bounded connection rank are homomorphism functions into weighted graphs. In
this section we continue the study of such parameters, now using their represen-
tation as homomorphism functions. A key step in the proof was to construct an
idempotent basis in the appropriate graph algebra. What is the exact size of this ba-
sis? How large graphs we need to represent the basis elements as quantum graphs?
Our goal is to prove an exact answer to the first question, and to give good bounds
for the second.
6.4.1. The connection rank of homomorphism functions. One can give
an exact formula for the connection rank r(f, k) of homomorphism functions f =
hom(., H) (Lovász [2006b]); to state it, we need a definition. Two nodes i and j in
the weighted graph H are twins, if βH (i, k) = βH (j, k) for every node k ∈ V (H).
(Note that this applies also to k = i and k = j. On the other hand, twin nodes may
have different nodeweights.) Twin nodes can be merged (adding up their weights)
without changing the homomorphism functions t(., H) and hom(., H).
If H is a weighted graph, then we denote the factor algebra Qk /hom(., H)
simply by Qk /H. The proof of the following characterization of the kernels of the
algebras is left to the reader as an exercise:
Proposition 6.35. If H is a weighted graph and x is a k-labeled quantum graph,
then x ≡ 0 (mod hom(., H)) if and only if homφ (x, H) = 0 for every φ : [k] →
V (H). 
This fact allows us to consider the functions homφ as functions on the factor
algebra Qk /H.
Theorem 6.36. If H is a twin-free weighted graph, then r(hom(., H), k) is the
number of orbits of the automorphism group of H on the ordered k-tuples of nodes.
102 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

Corollary 6.37. Let H be a weighted graph that has no twins and no proper au-
tomorphisms. Then r(hom(., H), k) = v(H)k for every k.
Theorem 6.36 has a number of essentially equivalent formulations, which are
interesting on their own right. One of these characterizes homomorphism functions
of the form homφ (F, H).
Theorem 6.38. Let H be a twin-free weighted graph and h : V (H)k → R. Then
there exists a k-labeled quantum graph z such that homφ (z, H) = h(φ) for every
φ ∈ V (H)k if and only if h is invariant under the automorphisms of H: for every
φ ∈ V (H)k and every automorphism σ of H, h(σ ◦ φ) = h(φ).
Another variant of these theorems gives a combinatorial description of the basic
idempotents p1 , . . . , pn in the algebra Qk /H, which played an important role in the
proof of the characterization theorem. For every φ ∈ V (H)k , we have
homφ (pi , H) = homφ (p2i , H) = homφ (pi , H)2 ,
and hence homφ (pi , H) ∈ {0, 1}. Furthermore, for i ̸= j, we have
homφ (pi , H)homφ (pj , H) = homφ (pi pj , H) = 0,
and hence the sets Φi = {φ ∈ V (H)k : homφ (pi , H) = 1}, which we call idempotent
supports, are disjoint. Since
∑
αφ (H)homφ (pi , H) = hom(pi , H) = hom(p2i , H) > 0,
φ
∑
the idempotent supports are nonempty. We have i pi = Ok , and hence
∑
n
homφ (pi , H) = homφ (Ok , H) = 1,
i=1

and so the idempotent supports form a partition of V (H)k . Since p1 , . . . , pn form

a basis of Qk /H, it follows that functions φ 7→ homφ (x, H) (x ∈ Qk ) are exactly
those that are constant on every idempotent support.
The following characterization of the partition Bk into idempotent supports is
a further equivalent version of Theorem 6.36.
Theorem 6.39. For a twin-free weighted graph H, the idempotent supports are
exactly the orbits of the automorphism group of H on V (H)k .
We call two maps φ, ψ ∈ V (H)k equivalent (in notation φ ∼ ψ) if homφ (F, H) =
homψ (F, H) for every k-labeled graph F . It follows from the discussion above that
this means that they belong to the same idempotent support. With this notion
finally we come to the version of these equivalent theorems that we are going to
prove first.
Theorem 6.40. Two maps φ, ψ ∈ V (H)k are equivalent if and only if there exists
an automorphism σ of H such that ψ = σ ◦ φ.
Proof. The “if” part is trivial, so let do the “only if” part. For any map
φ : [k] → [q], let φ′ denote its restriction to [k − 1]. We start with some easy facts
about equivalence of maps.
Claim 6.41. If two maps φ, ψ are equivalent, then so are φ′ and ψ ′ .
 6.4. ALGEBRAS FOR HOMOMORPHISM FUNCTIONS 103

Indeed, for any (k − 1)-labeled graph F , and the graph F1 obtained from F
by adding a new isolated node labeled k, we have homφ′ (F, H) = homφ (F1 , H) =
homψ (F1 , H) = homψ′ (F, H).
Claim 6.42. Suppose that φ, ψ ∈ [q]k are equivalent. Then for every µ ∈ [q]k+1
such that φ = µ′ there exists a ν ∈ [q]k+1 such that ψ = ν ′ and µ and ν are
equivalent.
Indeed, let µ belong to the support of the basic idempotent p ∈ Qk+1 /H, then
for every ν ∈ V (H)k+1 we have homν (p, H) = 1(ν ∼ µ). Let p′ be obtained by
unlabeling k + 1 in p. Then
∑ ∑
(6.31) homφ (p′ , H) = αη(k+1) (H)homη (p, H) = αη(k+1) (H),
η:η ′ =φ η:η ′ =φ
η∼µ

and similarly
∑
(6.32) homψ (p′ , H) = αη(k+1) (H).
η:η ′ =ψ
η∼µ

These two numbers are equal since φ ∼ ψ. Since the right side of (6.31) is positive,
this implies that the sum in (6.32) is nonempty, and hence there is a map ν such
that ν ′ = ψ and ν ∼ µ.
The next observation makes use of the twin-free assumption.
Claim 6.43. Every map σ : [q] → [q] such that βσ(i)σ(j) = βij for every i, j ∈ [q]
is bijective.
To prove this, note that the mapping σ has some power γ = σ s that is idem-
potent. Then for all i, j ∈ [q], we have βij = βγ(i)γ(j) = βγ 2 (i)γ(j) = βγ(i)j , which
shows that i and γ(i) are twins for all i ∈ [q]. Since H is twin-free, this implies
that γ is the identity, and so σ must be bijective.
After this preparation, we prove the theorem for larger and larger classes of
mappings.
Case 1: φ is bijective. Then k = q. We may assume that the nodes of H are
labeled so that φ is the identity, and then we want to prove that ψ (viewed as a
map of V (H) into itself) is an automorphism of H. First, we show that
(6.33) βij = βψ(i)ψ(j)
for every i, j ∈ [k]. Indeed, let kij be the k-labeled graph consisting of k nodes and
a single edge connecting nodes i and j. Then βij = homφ (kij , H) = homψ (kij , H) =
βψ(i)ψ(j) . It follows by Claim 6.43 that ψ is also bijective.
Second, we show that for every j ∈ [k],
(6.34) αj = αψ(j) .
It suffices to prove this for the case j = k. For the graph Ok−1 consisting of k − 1
isolated labeled nodes,
∏
k−1
homφ′ (Ok−1 , H) = αj ,
j=1
104 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

and since ψ is bijective,

∏
k−1
1 ∏
k
homψ′ (Ok−1 , H) = αψ(j) = αj .
j=1
αψ(k) j=1

Since ψ ′ ∼ φ′ by Claim 6.41, equation (6.34) follows.

Case 2: φ is surjective. By permuting the labels 1, . . . , k if necessary, we may
assume that φ(1) = 1, . . . , φ(q) = q. Claim 6.41 implies that the restriction of
ψ to [q] is equivalent to the restriction of φ to [q], and so by Case 1, there is an
automorphism σ of H such that ψ(i) = σ(i) for i = 1, . . . , q.
Consider any q + 1 ≤ j ≤ k, and let φ(j) = r. We claim that ψ(j) = ψ(r).
Indeed, the restriction of φ to {1, . . . , r−1, r+1, . . . , q, j} is bijective, and equivalent
to the restriction of ψ to this set; hence the restriction of ψ to this set must be
) implies that ψ(j) = ψ(r). This implies that for every 1 ≤ i ≤ k,
bijective,( which
ψ(i) = σ φ(i) as claimed.
Case 3. φ is arbitrary. We can extend φ to a mapping µ : [ℓ] → [q] (ℓ ≥ k) which
is surjective. By Claim 6.42, there is a mapping ν : [ℓ] → [q] extending ψ such that
µ and ν are equivalent. Then by Case 2, there is an automorphism σ of G such
that ν = σ ◦ µ. Restricting this map to [k], the assertion follows. 

The other theorems stated above are easy to derive now.

Proof of Theorems 6.36, 6.38, and 6.39. Theorem 6.39 is trivially equiv-
alent to Theorem 6.40 by the description of idempotent supports. The “only if” part
of Theorem 6.38 is also trivial. To prove the “if” part, notice that every function
h : V (H)k → R invariant under automorphisms can be written as a linear combi-
nation of indicator functions of the orbits of the automorphism group. By Theorem
6.39, this means that it is a linear combination of the functions homφ (pi , H), and
hence it is of the form homφ (z, H) with some z ∈ Qk .
Finally, it follows that the number of orbits of the automorphism group of H on
V (H)k is the number of the idempotents pi , which is r(f, k), which proves Theorem
6.36. 

These results describe an interesting isomorphism between the graph algebras

defined by a homomorphism function hom(., H) and algebras of functions on a
twin-free weighted graph H: Qk /H is isomorphic to the algebra of those functions
V (H)k → R that are invariant under the automorphisms of H. The isomorphism
is defined by the map F 7→ homv1 ...vk (F, H), where homv1 ...vk (F, H) is viewed as a
function of v1 , . . . , vk . This correspondence between quantum graphs and functions
on V (H) is useful in constructing quantum graphs with special properties.
As an application of the tools developed in this section, we are now able to
prove a weaker version of Theorem 5.33, without the bounds on the size of the
graph.
Corollary 6.44. If H1 and H2 are twin-free weighted graphs such that
hom(F, H1 ) = hom(F, H2 ) holds for all simple graphs F , then H1 ∼
= H2 .
Proof. Let H be the graph obtained by taking the disjoint union of H1 and
H2 , creating two new nodes v1 and v2 , and connecting vi to all nodes of Hi . Also
add loops at both vi . The new nodes and new edges have weight 1, except for the
 6.4. ALGEBRAS FOR HOMOMORPHISM FUNCTIONS 105

loops at v1 and v2 , which get some weight β different from all other edgeweights.
This last trick is needed to make sure that the graph H is twin-free.
We claim that for every 1-labeled graph F
(6.35) homv1 (F, H) = homv2 (F, H).
Indeed, if F is not connected, then those components not containing the labeled
node contribute the same factors to both sides. So it suffices to prove (6.35) when
F is connected. Then we have
∑
homv1 (F, H) = β eF (S) hom(F \ S, H1 ).
v1 ∈S⊆V (F )

Indeed, if we fix the set S = φ (v1 ), then the restriction φ′ of φ to V (F ) \ S is

−1

a map into V (H1 ) (else, the contribution of the map to hom(F, H) is 0), and the
contribution of φ to homv1 (F, H) is the product of contributions from the edges
induced by S and the contribution of φ′ to hom(F \ S, H).
Since homv2 (F, H) can be expressed by a similar formula, and the sums on the
right hand sides are equal by hypothesis, this proves (6.35).
Now (6.35) can be phrased as the maps 1 7→ v1 and 1 7→ v2 are equivalent, and
so Theorem 6.40 implies that there is an automorphism of H mapping v1 to v2 .
This automorphism gives an isomorphism between H1 and H2 . 
6.4.2. The size of basis graphs. Every element of the factor algebra Qk /H
has many representations as a quantum graph in Qk . The following theorem asserts
that it has a representation whose constituents are (in a sense) small.
Theorem 6.45. Let H be a weighted graph with V (H) = [q]. The algebra Qk /H
is generated by simple k-labeled graphs with at most 2(k + q 2 )q 6 nodes, in which the
labeled nodes form a stable set.
Proof. Let F = (V, E) be any k-labeled graph; we construct a simple k-labeled
quantum graph x, where each constituent has no more than 2(k + q 2 )q 6 nodes, and
F ≡ x (mod H). ( )
Let z be a 2-labeled quantum graph such that homφ (z, H) = 1 φ(1) = φ(2)
for all φ : {1, 2} → [q]. (So z is very similar to a contractor. We have z 2 = z,
but z ◦ z ̸= z.) We may assume that every constituent of z has at most 2q 6 nodes
(Supplement 6.29 and the Remark after it). Let z = O2 − z. Let w be a simple
connector; we can assume that w is a linear combination of paths of length at least
3 and at most q + 3 by Exercise 6.46.
Let us glue a copy of U2 on every pair of distinct nodes of V ; this does not
change F . But we can expand every O2 as O2 = z + z, and obtain a representation
|V |
of F as a sum of quantum graphs xℓ (ℓ = 1, . . . , 2( 2 ) ), each of which is obtained
from F by gluing either z or z on every pair of nodes in V .
Many of these terms will be 0. For any term xℓ , let Gℓ denote the graph on
S in which two nodes are connected if and only if they have a copy of z glued on.
If (i, j) and (j, k) have z glued on, but (i, k) has z, then the union of these three
is 0 as a quantum graph in Q3 (this is easy to check; cf. Exercise 6.32). Hence if
xi is a nonzero term, then adjacency must be transitive in Gℓ , and so Gℓ consists
of disjoint complete graphs. If Gℓ has more than q components, then any map
V → V (H) will collapse two nodes of V on which a z is glued, and hence xℓ = 0.
So we are left with only those terms in which Gℓ consists of at most q disjoint
106 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

complete graphs. Let V = V1 ∪ · · · ∪ Vr be the partition onto the node sets of these
components (r ≤ q).
Let us select a representative node vi from every Vi . It is easy to see that
deleting the copies of z except those which are attached to a vi , and also the copies
of z except those connecting two nodes vi , does not change xℓ .
If uv ∈ E with u ∈ Vi and v ∈ Vj (i ̸= j), then we can “slide” this edge to
vi vj without changing xℓ (cf. Exercise 6.31). If u, v ∈ Vi , then we replace the edge
uv by a simple connector w in which the labeled nodes are at a distance at least 3
(cf. Exercise 6.46), and then slide both attachment nodes to vi , to get a copy of w′
hanging from vi .
Each constituent of the resulting quantum graph consists of a “core”, the set
of the nodes vi and the set of labeled nodes, at most k + q nodes altogether. What
is not bounded is the sets of edges connecting a vi and a vj , the sets of copies of w′
hanging from a vi , and the copies of z connecting vi to other nodes in Vi .
However, we can get rid of these unbounded multiplicities. First, a set of q 2 or
more parallel edges can be replaced by a linear combination of sets of parallel edges
with multiplicity at most q 2 − 1, by Exercise 6.6(b). By a similar argument, a set of
q or more copies of w′ hanging from the same node vi can be expressed as a linear
combination of sets of at most q − 1 copies. Finally, again by the same argument,
a set of q or more copies of z connecting vi to unlabeled nodes can be expressed as
(aq)linear combination of sets of at most q − 1 copies. So we are left with at most
2 (q 2
− 1) edges that may be parallel to others, at most q(q − 1) hanging copies of
w′ , and at most k + q(q − 1) copies of z.
We get rid of the edge multiplicities by replacing each edge between core nodes
by a simple connector w.
After that, each constituent will be a simple graph. By the choice of z( and ) 2 q,
2 (q −
q
the number of nodes in( each constituent
) will be bounded by k + q + (q + 2)
1) + (q + 2)q(q − 1) + k + q(q − 1) (2q 6 ) < 2(k + q 2 )q 6 . 
As an application of the previous theorem, we prove Theorem 5.33 in its full
strength, including the bounds on the sizes of the graphs needed.
Proof of Theorem 5.33. Following the proof of Corollary 6.44, we have to
show that (6.35) holds. We do know that it holds for every F with at most 2(v(H1 )+
v(H2 ) + 3)8 nodes. Since this includes all basis graphs of Q1 /H by Theorem 6.45,
it follows that (6.35) holds for all simple 1-labeled graphs F . From here, the proof
is unchanged. 

Exercise 6.46. Prove that for every weighted graph H with q nodes and every
t ≥ 2, hom(., H) has a connector whose constituents are Pt•• , Pt+1
•• ••
, . . . , Pt+q .
Exercise 6.47. Prove that for every weighted graph H, hom(., H) has a contrac-
tor whose constituents are series-parallel graphs.

6.5. Computing parameters with finite connection rank

As an application of graph algebras, we prove the theorem announced before,
namely that graph parameters with finite connection rank can be computed in
polynomial time for certain classes of graphs. Of course, the algorithm could be
described without reference to algebras, but I feel that the essence is better shown
through this tool.
 6.5. COMPUTING PARAMETERS WITH FINITE CONNECTION RANK 107

Suppose you want to evaluate a graph parameter on a graph G. There is a cut

of size k in a graph, and while you know everything about one side of the cut, you
have to pay for information about the other side. How much information do you
need about the other side? To avoid the trivial solution “just tell me the value of
the parameter, if my side looks like this”, let us assume that the information about
the other side must be independent from what is on our side, and it is encoded in
the form of an m-tuple of real numbers. Furthermore, the answer must be obtained
by taking an appropriate linear combination of these m numbers, with coefficients
that depend only on the graph on our side.
As an example, let k = 1 (so we have a cutset {v} with one node), and suppose
that we want to compute the number of independent sets in the whole graph. Then
we need to know two data about the other side: the number a0 of independent sets
not containing v, and the number a1 of independent sets containing v. We determine
the analogous numbers b0 , b1 for our side, and then the number of independent sets
in the whole graph is a0 b0 + a1 b1 .
One reason to be interested in the finiteness of connection rank is the fact
that such a graph parameter can be evaluated in polynomial time for graphs with
bounded treewidth, based on the idea explained above. The treewidth of a graph
G is defined as follows. A tree-decomposition of a graph G is determined by a
tree T and a family (Gi )i∈V (T ) of induced subgraphs of G such that G = ∪i Gi
and whenever i is on the path from j to k in T (i, j, k ∈ V (T )), then V (Hi ) ⊃
V (Hj ) ∩ V (Hk ). The tree-width of a graph G is the smallest integer k such that G
has a tree-decomposition into subgraphs of size at most k + 1.

Theorem 6.48. Let f be a graph parameter and k ≥ 0. If r(f, k) is finite, then f

can be computed in polynomial time for graphs with treewidth at most k.

Proof. We describe a dynamic programing algorithm to compute the param-

eter. If the connection matrix M (f, k) has finite rank m, then the graph algebra
Qm /f is finite dimensional for all m ≤ k (Exercise 4.30). We need to do some
(large, but finite) precomputation.
First, we compute a basis Bm , consisting of (ordinary) m-labeled graphs, for
each of the algebras Qm /f . We also express the product of any two basis elements
in this basis (the “Schur constants”).
Second, let H be an l-labeled graph with at most k + 1 nodes (l ≤ k), and
for every ordered subset S ⊆ V (H) with |S| ≤ k, let a basis graph FS ∈ B|S| be
assigned. Let us glue the labeled nodes of FS onto the set S (erasing the labels in
FS at the same time), to get an l-labeled graph H ′ . We compute the representation
of H ′ in the basis Bl . We do this precomputation for every H and every assignment
of basis graphs FS .
Third, we compute the values f (G) for every G ∈ B0 .
Let G be a graph with treewidth at most k, and let (Gi )i∈V (T ) be a tree-
decomposition of the graph with v(Gi ) ≤ k + 1. Designate any leaf r of T as its
root, and for i ∈ V (T ) \ {r}, let i′ denote its parent.
For every node i ∈ V (T )\{r}, the set Si = V (Gi )∩V (Gi′ ) is a cutset in G with
ki ≤ k nodes. Let Fi denote the union of all graphs Gj where j is a descendant of
i (including i), in which the ki nodes of Si are labeled. The algorithm will consist
of expressing every Fi in the basis Bki , starting from the leaves and working our
way up to the root.
108 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

Suppose that such an expression has been computed for every proper descen-
dant of i. The ki -labeled graph Fi is obtained from Gi by attaching different
branches Fj at the sets Sj . We already know how to express each Fj in the basis
Bkj ; let us substitute this expression for Fj , to get a representation of Fi as a linear
combination of graphs, each of which consists of Gi with some number of basis
graphs attached at various subsets S ⊆ V (Gi ) with |S| ≤ k. If two or more basis
graphs are attached at the same set S, we can replace them by one, since we have
precomputed products of basis graphs. But then we have a linear combination of
ki -labeled graphs of the type we have already expressed in the basis Bki .
When we get to the root, we consider it 0-labeled, and get an expression for G
in the basis B0 , which yields the value f (G). 

6.6. The polynomial method

In this section we describe a method of proving representation theorems for
graph parameters, which depends on commutative algebra (properties of multivari-
ate polynomials). This method was developed by B. Szegedy [2007] for the proof of
the characterization of graph parameters that are partition functions of edge col-
oring models (we will describe this result without proof in Chapter 23.2). Here we
use an adaptation of this method to prove Theorem 5.57, due to Schrijver [2009].
The basic idea is to treat the edge weights of the target graph H as variables.
Then homomorphism numbers into H will be polynomials in these variables. One
treats this as a polynomial-valued graph parameter, works out the corresponding
graph algebras, and then proves that one can find a substitution for the variables
that reproduces the given graph parameter.
Let H be a weighted graph on [q], in which the nodeweights are 1, and the
edgeweights are different variables xij (where xij = xji ). It will be convenient to
arrange these variables into a symmetric q × q matrix X. Every substitution of
complex numbers for the xij gives a complex valued homomorphism function into
an edge-weighted graph on [q], and vice versa. The homomorphism number
∑ ∏
hom(G, H) = xφ(i)φ(j)
φ:V (G)→[q] ij∈E(G)
(q )
is a polynomial in the 2 + q variables xij , which we denote by hom(G, X).
We define the polynomial inj(G, X) analogously. Clearly hom(K0 , X) = 1,
hom(K1 , X) = q, and hom(G1 G2 , X) = hom(G1 , X)hom(G2 , X). In particular,
hom(GK1 , X) = qhom(G, X) for every graph G. We extend the definition linearly
to quantum graphs, to get polynomials hom(g, X), inj(g, X) ∈ C[X] associated with
every quantum graph g.
We start with describing the range and kernel of the map g 7→ hom(g, X) (where
q is fixed). Clearly, hom(g, X) is invariant under the permutations of [q]. To be
more precise, if σ ∈ Sq and we define X σ = (xσ(i),σ(j) : i, j ∈ [q]), then trivially
hom(., X) = hom(., X σ ). Let C[X]Sq denote the space of polynomials in C[X] that
are invariant under Sq in this sense.

Lemma 6.49. The polynomials hom(g, X), where g ∈ Q0 , form the space C[X]Sq .
∏ a
Proof. Let X a = i≤j xijij be any monomial, and let G denote the multi-
graph on [q] in which nodes i and j are connected by aij edges. Then inj(G, X) =
 6.6. THE POLYNOMIAL METHOD 109

∑
σ∈Sq (X ) . Since every polynomial in C[X]Sq can be written as a linear combi-
σ a

nation (with constant coefficients) of such special polynomials, it follows that every
polynomial in C[X]Sq can be written as inj(g, X) for some quantum graph g. By
identity (5.18) (which remains valid if the graph G is replaced by the matrix X),
this implies the Lemma. 
Next, we describe quantum graphs g with hom(g, X) = 0 (identically 0 as a
polynomial in the entries of X). Note that if we remove an isolated node to a
constituent of any quantum graph g, and multiply its coefficient by q, then we get
a quantum graph g ′ such that hom(g, X) = hom(g ′ , X). Let us call the repeated
application of this operation an isolate removal.
Lemma 6.50. A quantum graph g satisfies hom(g, X) = 0 if and only if there
is a quantum graph h in which all constituents have more than q nodes such that
removing isolates from g we obtain M h.
Proof. If g = M h, where all constituents of h have more than q nodes, then
hom(g, X) = inj(h, X) = 0. Isolate removal does not change the value of hom(g, X).
Conversely, suppose that hom(g, X) = 0. We may assume that the constituents
of g have no isolated nodes. We have inj(Zg, X) = hom(g, X) = 0. If Zg has a
constituent with at most q nodes, then this produces in inj(Zg, X) a term which
does not cancel (here we use that the constituent has no isolated nodes). So all
constituents of Zg have more than q nodes, and we can take h = Zg. 
Now we are ready to prove Theorems 5.56 and 5.57.
Proof of Theorem 5.56. Multiplicativity implies that f (K0 ) = 1 (since f is
not identically 0), and f (GK1 ) = qf (G).
We want to prove that f = hom(., A) for an appropriate symmetric complex
matrix A; in other words, we want to show that the polynomial equations
(6.36) hom(G, X) − f (G) = 0 (for all looped multigraphs G)
are solvable for the variables xij (1 ≤ i, j ≤ q) over the complex numbers. We are
going to use Hilbert’s Nullstellensatz for this, but we need some preparation. We
begin with relating the kernel of the map hom(., X) to the kernel of f .
Claim 6.51. If hom(g, X) = c (a constant polynomial) for some quantum graph g,
then f (g) = c.
First we consider the case when c = 0. We may assume that g has no isolated
nodes, since isolate removal does not change the values hom(g, X) and f (g). By
Lemma 6.50, g = M h for some quantum graph h in which all constituents have
more than q nodes. But then f (M h) = 0 by the hypothesis of the Theorem.
The case of general constant c follows easily: we have hom(g − cK0 , X) =
hom(g, X) − c = 0, and hence f (g) = f (g − cK0 ) + c = c.
Claim 6.52. The ideal generated by the polynomials hom(g, X) with f (g) = 0 does
not contain the constant polynomial 1.
Suppose that we have a representation
∑
N
1= pi (X)hom(gi , X),
i=1
110 6. GRAPH ALGEBRAS AND HOMOMORPHISM FUNCTIONS

where f (gi ) = 0, and the pi are arbitrary polynomials in C[X]. Let us apply a
permutation σ ∈ Sq to the variables, and sum over all σ. We get:
∑
N ∑ N (∑
∑ )
q! = pi (X σ )hom(gi , X σ ) = pi (X σ ) hom(gi , X).
i=1 σ∈Sq i=1 σ∈Sq

The expression in the large parenthesis is a polynomial in C[X]Sq , and hence by

Lemma 6.49, it is a polynomial of the form hom(hi , X). Hence we get
∑
N (∑
N )
q! = hom(hi , X)hom(gi , X) = hom hi gi , X .
i=1 i=1
By Claim 6.51, this implies that
(∑
N )
f hi gi = q!.
i=1
But we have, using the multiplicativity of f ,
(∑
N ) ∑ N ∑
N
f hi gi = f (hi gi ) = f (hi )f (gi ) = 0,
i=1 i=1 i=1
a contradiction. So Claim 6.52 is proved.
Now it is easy to finish the proof of the theorem. Claim 6.52 and the Nullstel-
lensatz imply that there are complex numbers aij such that aij = aji (1 ≤ i, j ≤ q),
and hom(g, A) = 0 for every quantum graph g for which f (g) = 0 (where A is the
matrix with entries aij ). Applying this to the quantum graph G − f (G)K0 (where
G is an arbitrary multigraph), we get that hom(G − f (G)K0 , A) = 0, and hence
f (G) = hom(G, A). 
Proof of Theorem 5.57. We will apply Theorem 5.56, but first we need to
show a couple of properties of f following from reflection positivity.
Claim 6.53. f (K1 ) is a nonnegative integer.
Let q = f (K1 ). For k ≥ 1, and for a partition P = (S1 , . . . , Sm ) of [k], let UP
denote the k-multilabeled graph on [m] with no edges, where node i is labeled by
the elements of Si . Consider the submatrix M of M mult (f, k) formed by those rows
and columns indexed
∑ by the graphs UP .
Let hk = P µP UP . Then using identity (A.5) for the Möbius function, we
get
∑ ∑
(6.37) ⟨hk , hk ⟩ = µP µQ f (UP ∨Q ) = µP µQ q |P ∨Q| = q(q − 1) · · · (q − k + 1).
P,Q P,Q

Since f is reflection positive, this value must be nonnegative for every k, which
implies that q is a nonnegative integer.
Claim 6.54. If G is a multigraph with k = v(G) > q, then f (M G) = 0.
Let us label the nodes of G by [k], to get a k-labeled graph. Then M G = [[hk G]],
and so f (M G) = ⟨hk , G⟩. Equation (6.37) implies that ⟨hk , hk ⟩ = 0, which (using
reflection positivity again) implies that ⟨hk , G⟩ = 0, which proves the Claim.
So Theorem 5.56 applies, and we get that there exists a symmetric matrix
A ∈ Cq×q for which f = hom(., A). To complete the proof, we have to show:
 6.6. THE POLYNOMIAL METHOD 111

Claim 6.55. The matrix A is real.

This does not follow just from the assumption that f is real (see Exercise 6.56);
we have to use reflection positivity again. Suppose that A has an entry auv which
∑N
is not real. There is a polynomial p = j=0 cj z j ∈ C[z] such that


i if z = auv ,
p(z) = −i if z = auv ,


0 if z ∈ {ast , ast } for some entry ast ̸= auv , auv .
(so p takes pure imaginary values on the entries of A and on their conjugates). This
polynomial may have complex coefficients, but it is easy to see that its complex
conjugate p(x) satisfies the same conditions, and hence, replacing p by (p + p)/2 if
necessary, we may assume that p has real coefficients.
∑
Consider the 2-labeled quantum graph g = j cj Bj•• . We have
∑ ∑ ∑ ∑
⟨g, g⟩ = ck cj hom(Bk+j , A) = ck cj ak+j
uv = p(auv )2 < 0,
k,j k,j u,v u,v

which contradicts the assumption that f is reflection positive. 

Exercise 6.56. Show by an example that hom(G, Z) can be real for every multi-
graph G for a non-real matrix Z.
 Part 3

Limits of dense graph sequences

 CHAPTER 7

Kernels and graphons

The aim of this Chapter is to introduce certain analytic objects, which will
serve as limit objects for graph sequences in the dense case. In the Introduction
(Section 1.5.3) we already gave an informal description of how these graphons enter
the picture as limit objects; however, for the next few chapters we will not talk
about graph sequences, but we treat graphons as generalizations of graphs, to
which many graph-theoretic definitions and results can be extended. Quite often,
the formulation and even the proof of these more general facts are easier in this
analytic setting. We will define the cut norm and cut-distance of these objects,
state and prove regularity lemmas for them, and prove basic properties of sampling
from them. These results will enable us to show that these are just the right objects
to represent the limits of convergent dense graph sequences.

7.1. Kernels, graphons and stepfunctions

Let W denote the space of all bounded symmetric measurable functions
W : [0, 1]2 → R. The elements of W will be called kernels (the name refers
to the fact that they give rise to kernel operators on function spaces on [0, 1]; we
will return to this connection in Section 7.5). Let W0 denote the set of all kernels
W ∈ W such that 0 ≤ W ≤ 1. The elements of W0 will be called graphons (the
name comes from the contraction of graph-function). Sometimes we will also need
to consider the set of all functions W ∈ W such that −1 ≤ W ≤ 1; this will be
denoted by W1 .
As usual, we will not distinguish functions that are almost everywhere equal
(. . . most of the time). Then the space W is just the space of symmetric functions in
L∞ ([0, 1]2 ), which we could identify with the space L∞ (T ), where T is the triangle
{(x, y) ∈ [0, 1]2 : x ≤ y}. We introduced a separate notation for it because we
want to consider a number of different norms on W, of which the L∞ norm will
play a relatively minor role.
A graphon whose values are 0 and 1 can be considered as a graph on node set
[0, 1]. In this case, we can talk about its subgraphs, induced subgraphs, complement,
and so on. Such 0-1 valued graphons will come up in our discussions repeatedly;
however, they would not be sufficient for our main goal, namely, describing limit
objects for convergent graph sequences.
Kernels generalize weighted graphs in the following sense. A function W ∈ W
is called a stepfunction, if there is a partition S1 ∪ · · · ∪ Sk of [0, 1] into measurable
sets such that W is constant on every product set Si × Sj . The sets Si are the
steps of W . For every weighted graph H (on node set V (H) = [n]), we define a
stepfunction WH ∈ W as follows: Split [0, 1] into n intervals J1 , . . . , Jn of length
λ(Ji ) = αi /αH , and for x ∈ Ji and y ∈ Jj , let WH (x, y) = βij (H). Note that the
function WH depends on how the nodes of H are labeled.
115
116 7. KERNELS AND GRAPHONS

Conversely, every stepfunction U corresponds to a weighted graph: if S1 , . . . , Sk

are its steps, then the graph is defined on [k], and the edge ij has weight U (x, y),
where x ∈ Si and y ∈ Sj .
If H is a weighted graph with nodeweights 1 and weighted adjacency matrix
A, then we write WA = WH .
If the edgeweights of H are from the interval [0, 1], then WH is a graphon. In
particular, for every simple (unweighted) graph G, WG is a 0-1 valued graphon
(recall Figure 1.3). In this sense, simple graphs can be considered as special 0-1-
valued graphons.
This correspondence with simple graphs suggests how to extend some basic
quantities associated with graphs to kernels (or at least to graphons). Most impor-
tant of these is the (normalized) degree function

∫1
(7.1) dW (x) = W (x, y) dy.
0

(If the graphon is associated with a simple graph G, this corresponds to the scaled
degree dG (x)/v(G).) We will see more such quantities in the next sections.
Instead of the interval [0, 1], we can consider any probability space (Ω, A, π)
with a symmetric measurable function W : Ω × Ω → [0, 1]. This would not provide
substantially greater generality, but it is sometimes useful to represent graphons by
probability spaces other than [0, 1]. We’ll discuss this in detail in Chapter 13, but
will use this different way of representing a graphon throughout.
Graphons will come up in several quite different forms in our discussions. In
Theorem 11.52 we will collect the many disguises in which they occur.

7.2. Generalizing homomorphisms

Homomorphism densities in graphs extend to homomorphism densities in
graphons and, more generally, in kernels. For every W ∈ W and multigraph
F = (V, E) (without loops), define
∫ ∏ ∏
t(F, W ) = W (xi , xj ) dxi
ij∈E i∈V
[0,1]V

We can think of the interval [0, 1] as the set of nodes, and of the value W (x, y)
as the weight of the edge xy. Then the formula above is an infinite analogue of
weighted homomorphism numbers. We get weighted graph homomorphisms as a
special case when W is a stepfunction: For every unweighted multigraph F and
weighted graph G,

(7.2) t(F, G) = t(F, WG ).

Of the two modified versions of homomorphism densities (5.12) and (5.13), the
notion of the injective density tinj has no significance in this context, since a random
assignment i 7→ xi (i ∈ V (F ), xi ∈ [0, 1]) is injective with probability 1. In other
words, tinj (F, W ) = t(F, W ) for any kernel W and any graph F . But the induced
subgraph density is worth defining, and in fact it can be expressed by a rather
 7.2. GENERALIZING HOMOMORPHISMS 117

simple integral:
∫ ∏ ∏ ( ) ∏
(7.3) tind (F, W ) = W (xi , xj ) 1 − W (xi , xj ) dxi .
ij∈E ij∈(V2 )\E i∈V
[0,1]V

We have an analogue of the inclusion-exclusion formula (5.20), which follows by

expanding the parentheses in the integrand (7.3):
∑ ′
(7.4) tind (F, W ) = (−1)e(F )−e(F ) t(F ′ , W ) = t↑ (F, W ).
F ′ ⊇F
V (F ′ )=V (F )

We should point out that tinj (F, WH ) ̸= tinj (F, H) and tind (F, WH ) ̸= tind (F, H) in
general. We have seen that tinj (F, WH ) = t(F, WH ) = t(F, H). For the induced
density, tind (F, WH ) has a combinatorial meaning if H is a looped-simple graph:
it is the probability that a random map V (F ) → V (H) (not necessarily injective)
preserves both adjacency and nonadjacency.
Many other basic properties of homomorphism numbers extend to graphons,
often to kernels, in a straightforward way, like (5.19) generalizes to
∑
(7.5) t(F, W ) = tind (F ′ , W ),
F ′ ⊇F

and (5.28) generalizes to the identity

(7.6) t(F1 F2 , W ) = t(F1 , W )t(F2 , W ).
We can also generalize homomorphism numbers from partially labeled graphs.
Let F = (V, E) be a k-labeled multigraph. Let V0 = V \ [k] be the set of unlabeled
nodes. For W ∈ W and x1 , . . . , xk ∈ [0, 1], we define
∫ ∏ ∏
tx1 ,...,xk (F, W ) = W (xi , xj ) dxi
ij∈E i∈V0
x∈[0,1]V0

(this is a function of x1 , . . . , xk ). In particular, we have tx (K2• , W ) = dW (x). It is

often convenient to use the notation tx , where x = (x1 , . . . , xk ). The product of
two k-labeled graphs F1 and F2 satisfies
(7.7) tx (F1 F2 , W ) = tx (F1 , W )tx (F2 , W )
′
If F arises from F by unlabeling node k (say), then
∫
(7.8) tx1 ,...,xk−1 (F ′ , W ) = tx1 ,...,xk (F, W ) dxk .
[0,1]

By repeated application of this equation, we get that if F is a k-labeled multigraph,

then
∫
(7.9) t([[F ]], W ) = tx (F, W ) dx.
[0,1]k

Further versions of homomorphism densities treated before can be extended

to homomorphism densities in kernels in a straightforward way. Homomorphism
densities of quantum graphs in kernels are defined simply by linearity. Densities of
signed graphs can be defined generalizing expression 7.3 for the induced subgraph
118 7. KERNELS AND GRAPHONS

densities. Explicitly, let F = (V, E + , E − ) be a signed graph and W ∈ W, then we

define
∫ ∏ ∏ ( ) ∏
(7.10) t(F, W ) = W (xi , xj ) 1 − W (xi , xj ) dxi .
ij∈E+ ij∈E− i∈V
[0,1]V

From this definition, it follows that if W is a graphon, then 0 ≤ t(F, W ) ≤ 1 for

every signed graph F . We can also express t(F, W ) as
∑ ( )
(7.11) t(F, W ) = (−1)|Y | t (V, E+ ∪ Y ), W .
Y ⊆E−

This shows that∑we can still identify a signed graph F = (V, E + , E − ) with the
quantum graph Y ⊆E− (−1)|Y | (V, E+ ∪ Y ).
If all edges are signed “+”, then t(F, W ) is the same as for unsigned graphs.
If Fb is the signed complete graph, obtained from an unsigned simple graph F on
the same node set, in which the edges of F are signed positive and the edges of the
complement are signed negative, then we get the following identity, equivalent to
(7.4):

(7.12) t(Fb, W ) = tind (F, W ).

We define the induced density tind,x (F, W ) of a k-labeled graph F , and the
density of a k-labeled signed graph or quantum graph in the obvious way.
The following proposition states some main properties of subgraph densities in
kernels.

Proposition 7.1. The graph parameter t(., W ) is multiplicative and reflection pos-
itive for every kernel W ∈ W. The corresponding simple graph parameter is also
multiplicative, and it is reflection positive if W ∈ W0 .

Proof. The second assertion is more difficult to prove, and we describe the
proof in this case only. Multiplicativity is trivial. To prove that t(., W ) is reflection
positive, consider any finite set F1 , . . . , Fm of k-labeled graphs, and real numbers
y1 , . . . , ym . We want to prove that
∑
m
t([[Fp Fq ]], W )yp yq ≥ 0.
p,q=1

For every k-labeled graph F with node set [n], let F ′ denote the subgraph of F
induced by the labeled nodes, and F ′′ denote the graph obtained from F by deleting
the edges spanned by the labeled nodes. Then we have
∑
m
(7.13) yp yq t([[Fp Fq ]], W )
p,q=1
∫ ∑
m
= yp yq tx (Fp′′ , W )tx (Fp′′ , W )tx (Fp′ ∪ Fq′ , W ) dx.
[0,1]k p,q=1
∑
We substitute tx (Fp′ ∪ Fq′ , W ) = H tind,x (H, W ), where the summation extends
over all graphs on [k] containing Fp′ ∪ Fq′ as a subgraph. Interchanging summation,
 7.2. GENERALIZING HOMOMORPHISMS 119

we get
∑
m
(7.14) yp yq t([[Fp Fq ]], W )
p,q=1
∫ ∑ ∑
= yp yq tx (Fp′′ , W )tx (Fp′′ , W )tind,x (H, W ) dx
[0,1]k H F ,F ⊆H
p q

For a fixed H, the integrand can be written as

∑ ( ∑ )2
yp yq tx (Fp′′ , W )tx (Fp′′ , W )tind,x (H, W ) = yp tx (Fp′′ , W ) tind,x (H, W ),
Fp ,Fq ⊆H Fp ⊆H

which is nonnegative by the assumption that 0 ≤ W ≤ 1. 

We will see (Theorem 11.52) that multiplicativity and reflection positivity, to-
gether with the trivial condition that t(K1 ) = 1, characterize simple graph param-
eters of the form t(F, W ).
For a general graphon W , the graph parameter t(., W ) cannot be represented
as a homomorphism number into a (finite) weighted graph: it is multiplicative and
reflection positive, but it may have infinite connection rank. We will see (Corol-
lary 13.48) that t(., W ) has finite connection rank if and only if W is equal to a
stepfunction almost everywhere.
The multigraph parameter t(., W ) is contractible, but has no contractor. This
will follow from Theorem 6.30 together with the uniqueness of representation of a
parameter in the form t(., W ) (Theorem 13.10).
Example 7.2 (Eulerian orientations revisited). We have seen that the number
of eulerian orientations eul(G) is not a homomorphism function. However, it can
be expressed as a homomorphism density in a kernel:
−→ ( )
(7.15) eul(F ) = t F, 2 cos(2π(x − y)) .
Indeed, we can write
2 cos(2π(x − y) = e2πi(x−y) + e2πi(y−x) ,
so if we expand the product
∏
(e2πi(xu −xv ) + e2πi(xv −xu ) ),
uv∈E(F )
−
→
then every term corresponds to an orientation F of F , where selecting e2πi(xv −xu )
corresponds to orienting the edge uv from u to v. Thus
∏ ∑ ∏
2 cos(2π(xu − xv )) = e2πi(xv −xu )
uv∈E(F ) →
− →
−
F uv∈E( F )
∑ ∏ +
2πi(d→ −
− (u))xu
− (u)−d→
= e F F .
→
− u∈V (F )
F

If we integrate over all the xu , every term cancels in which the orientation is not
−
− (u) − d→
− (u) ̸= 0. Those terms
+
eulerian, i.e., where any of the nodes u has d→
F F
corresponding to eulerian orientations contribute 1. So the sum counts eulerian
orientations. 
120 7. KERNELS AND GRAPHONS

Remark 7.3. There is probably no good way to define homomorphism numbers

from graphons into graphs or into other graphons. The parameters related to such
homomorphisms that extend naturally to graphons are defined by maximization,
like the normalized maximum cut, and more generally, restricted maximum multi-
cuts. We will discuss these in Chapter 12.

We can generalize the functional t(F, W ) further (believe me, not for the sake
of generality). Let A be a set of kernels. An A-decorated graph is a finite simple
graph F = (V, E) in which every edge e ∈ E is labeled by a function We ∈ A. We
write w = (We : e ∈ E). For every W-decorated graph (F, w) we define
∫ ∏ ∏
(7.16) t(F, w) = Wij (xi , xj ) dxi .
ij∈E i∈V
[0,1]V

For a fixed graph F , the functional t(F, w) is linear in every edge decoration We .
So it may be considered as linear functional on the tensor product W ⊗ · · · ⊗ W
(one factor for every edge of F ), or equivalently, as a tensor on W with e(F ) slots.
This definition contains some of the previous variations on homomorphism
numbers, and it can be used to express homomorphism densities in sums of kernels.

Example 7.4. Let F = (V, E + , E − ) be a signed graph and W ∈ W0 . Let us

decorate each edge in E + by W , and each edge in E − by 1 − W . Let F0 be the
unsigned version of F . Then for the W-decorated graph (F, w) obtained this way,
we have t(F0 , w) = t(F, W ). 

Example 7.5. For W1 , . . . , Wk ∈ W, we have

∑
t(F, W1 + · · · + Wk ) = t(F, w),
w

where w ranges over all {W1 , . . . , Wk }-decorations of F . 

Exercise 7.6. Let F and G be two simple graphs, and let W be a graphon such
that t(F, G) > 0 and t(G, W ) > 0. Prove that t(F, W ) > 0. [Hint: Use the
Lebesgue Density Theorem.]
Exercise 7.7. Prove that for any two simple graphs F and G with v(F ) ≤ v(G)
we have (v(F ))
 
tind (F, G) − tind (F, WG ) ≤ 2 .
v(G)
Exercise 7.8. Let us generalize the construction of graph integrals by adding
“nodeweights”: for every graph F and bounded measurable functions α : [0, 1] →
R and W : [0, 1]2 → R (where W is symmetric), we define
∫ ∏ ∏
t(F, α, W ) = α(xi ) W (xi , xj ) dx.
i∈V (F ) ij∈E(F )
[0,1]V (F )

Show that if we require that α ≥ 0, then t(F, α, W ) can be expressed as

cv(F ) t(F, U ) with some c ≥ 0 and U : [0, 1]2 → R, where c and U depend on
α and W , but not on F .
Exercise 7.9. Prove that the number of perfect matchings in a graph G = (V, E)
can be expressed as t(G, e−2πix , 1 + e2πi(x+y) ).
 7.3. WEAK ISOMORPHISM I 121

7.3. Weak isomorphism I

One complication caused by moving to infinite objects is that isomorphism does
not have an obvious (and unique) definition any more. We can of course talk about
two kernels U, W to be equal as functions, but this is not very useful. More in the
spirit of functional analysis, we will talk about the two kernels to be equal almost
everywhere, i.e., W (x, y) = U (x, y) for almost all (x, y) ∈ [0, 1]2 (with respect to
the Lebesgue measure).
This notion, however, is not what we mean by two kernels being “essentially
the same”: it corresponds to the equality of labeled graphs, not to isomorphism
of unlabeled graphs, which involves finding the right bijection between the node
sets. In terms of graphons (or kernels), we can define this as follows: two kernels
U, W ∈ W are isomorphic up to a null ( set if there
) is an invertible measure preserving
map φ : [0, 1] → [0, 1] such that U φ(x), φ(y) = W (x, y) almost everywhere. (See
Appendix A.3 and the book of Sinai [1976] for the basics of measure preserving
maps.) Since the inverse of an invertible measure preserving map φ : [0, 1] → [0, 1]
is also measure preserving, isomorphism up to a null set is an equivalence relation.
However, there is a weaker notion of isomorphism, which will be more important
for us. The motivation for this notion is the fact that a measure preserving map
need not be invertible. Let W ∈ W and let φ : [0, 1] → [0, 1] be a measure
preserving map. We define a kernel W φ by
( )
W φ (x, y) = W φ(x), φ(y) .
From the point of view of using these functions as continuous analogues of graphs,
the functions W and W φ are not essentially different. For example, we have the
following important fact:
Proposition 7.10. Let W ∈ W and let φ : [0, 1] → [0, 1] be a measure preserving
map. Then for every multigraph F = (V, E), we have t(F, W φ ) = t(F, W ).
( )
Proof. This follows from the fact that (x1 , . . . , xn ) 7→ φ(x1 ), . . . , φ(xn ) is a
measure preserving map [0, 1]n → [0, 1]n , and hence for every integrable function
f : [0, 1]n → R we have
∫ ∫
( )
f φ(x1 ), . . . , φ(xn ) dx1 . . . dxn = f (x1 , . . . , xn ) dx1 . . . dxn
[0,1]n [0,1]n

by
∏ (A.16) in the Appendix. Applying this equation to the function f (x1 , . . . , xn ) =
ij∈E W (xi , xj ), we get the assertion. 
We want to say that W and W φ are “weakly isomorphic”. One has to be a little
careful though, because measure preserving maps are not necessarily invertible,
and so the relationship between W and W φ in Proposition 7.10 is not symmetric
(see Example 7.11). For the time being, we take the easy way out, and call two
kernels U and W weakly isomorphic if t(F, U ) = t(F, W ) for every simple graph F .
We will come back to a characterization of weakly isomorphic kernels in terms of
measure preserving maps (in other words, proving a certain converse of Proposition
7.10) in Sections 10.7 and 13.2. It will also follow that in this case the equation
t(F, U ) = t(F, W ) holds for all multigraphs F (see Exercise 7.18 for a direct proof).
Weak isomorphism of kernels is clearly an equivalence relation, and we can iden-
tify kernels that are weakly isomorphic. This identification will play an important
role in our discussions.
122 7. KERNELS AND GRAPHONS

Example 7.11. The map φ2 : x 7→ 2x (mod 1) is measure preserving. For every

kernel W , the kernel W φ2 consists of four “copies” of W (see Figure 7.1). Similarly,
φ3 : x 7→ 3x (mod 1) is measure preserving, and W φ3 consists of nine “copies” of
W . The kernels W , W φ2 and W φ3 are weakly isomorphic, but there is no measure
preserving map transforming W φ2 to W φ3 (Exercise 7.13). 

ϕ2 ϕ3
W W W
Figure 7.1. Gray-scale images of three graphons that are weakly
isomorphic, but not isomorphic up to a null set. Recall that the
origin is in the upper left corner.

This example illustrates that weak isomorphism is not a very easy notion. We
will return to it and develop more and more information about it when we introduce
distances between graphons, sampling, twin reduction, and other tools in the theory
of graphons.
Exercise 7.12. Suppose that two kernels U and W are weakly isomorphic. Prove
that so are the kernels aU + b and aW + b (a, b ∈ R).
Exercise 7.13. Prove that the kernels W , W φ2 and W φ3 in Example 7.11 are
weakly isomorphic, but not isomorphic up to a null set.

7.4. Sums and products

Perhaps the first tool we use in graph theory is the decomposition into con-
nected components. For kernels, a similar decomposition exists, but one must be a
bit careful with 0-sets. This was worked out by Janson [2008].
Let W1 , W2 , . . . be a finite or countably
∑ infinite family of kernels, and let
a1 , a2 . . . be positive real numbers with i ai = 1. We define the direct sum of
the Wi with weights ai , in notation W = a1 W1 ⊕ a2 W2 ⊕ . . . , as follows. We
split the interval [0, 1] into intervals J1 , J2 , . . . of length a1 , a2 , . . . , consider the
monotone affine maps φi mapping Ji onto [0, 1], and let
{ ( )
Wi φi (x), φi (y) , if x, y ∈ Ji , i = 1, 2, . . . ,
W (x, y) =
0, otherwise.
A kernel will be called connected, if it is not isomorphic up to a null set to
the direct sum of two kernels. This is equivalent to saying that for every subset
S ⊆ [0, 1] with 0 < λ(S) < 1, we have
∫
|W (x, y)| dx dy > 0.
S×([0,1]\S)

Every kernel can be written as the direct sum of connected kernels and perhaps the
0 kernel. (We have to allow the 0 kernel, which cannot be written as the sum of
 7.4. SUMS AND PRODUCTS 123

connected kernels.) This decomposition is unique (up to zero sets); see Bollobás,
Janson and Riordan [2007] and Janson [2008] for more.
Somewhat confusingly, we can introduce three “product” operations on kernels,
and we will need all three of them. Let U, W ∈ W. We denote by U W their
(pointwise) product as functions, i.e.,
(U W )(x, y) = U (x, y)W (x, y).
We denote by U ◦ W their operator product (the name refers to the fact that this is
the product of U and W as kernel operators, see Section 7.5)
∫1
(U ◦ W )(x, y) = U (x, z)W (z, y) dz.
0

We note that U ◦ W is not symmetric in general, but it will be in the cases we use
this operation (for example, when U = W ).
Finally, we denote by U ⊗ W their tensor product; this is defined as a function
[0, 1]2 × [0, 1]2 → [0, 1] by
(U ⊗ W )(x1 , x2 , y1 , y2 ) = U (x1 , y1 )W (x2 , y2 ).
This function is not defined on [0, 1]2 and hence it is not in W; however, we can
consider any measure preserving map φ : [0, 1] → [0, 1]2 , and define the kernel
( )
(U ⊗ W )φ (x, y) = (U ⊗ W ) φ(x), φ(y) .
It does not really matter which particular measure preserving map we use here:
these kernels obtained from different maps φ are weakly isomorphic by the same
computation as used in the proof of Proposition 7.10, and so we can call any of
them the tensor product of U and W .
We note that the tensor product has the nice property that
(7.17) t(F, U ⊗ W ) = t(F, U )t(F, W )
for every multigraph F .
We denote the n-th power of a kernel according to these three multiplications
by W n (pointwise power), W ◦n (operator power), and W ⊗n (tensor power).
There are many other properties and constructions for graphs that can be
generalized to graphons in a natural way. For example, we call a graphon W
bipartite, if there is a partition V (G) = V1 ∪ V2 such that W (x1 , x2 ) = 0 for almost
all (x1 , x2 ) ∈ V1 × V2 . We can define k-colorable kernels similarly. We call a
graphon triangle-free, if t(K3 , W ) = 0. Simple facts like “every bipartite graphon is
triangle-free” can be proved easily. Often one faces minor complications because of
exceptional nullsets; a rather general remedy for this problem, called pure graphons,
will be introduced in Section 13.3.
Exercise 7.14. Show that for every simple graph F , t(F, W ◦n ) = t(F ′ , W ), where
F ′ is obtained from F by subdividing each edge by n − 1 new nodes.
Exercise 7.15. Prove that connectivity of a graphon is invariant under weak
isomorphism.
Exercise 7.16. Prove that a graphon W is bipartite if and only if t(C2k+1 , W ) = 0
for all k ≥ 1.
124 7. KERNELS AND GRAPHONS

7.5. Kernel operators

Every function W ∈ W defines an operator TW : L1 [0, 1] → L∞ [0, 1], by
∫1
(7.18) (TW f )(x) = W (x, y)f (y) dy.
0
Sometimes it will be useful to consider TW as an operator L∞ [0, 1] → L1 [0, 1] or
L2 [0, 1] → L2 [0, 1]; the formula is meaningful in each of these cases.
If we consider TW : L2 [0, 1] → L2 [0, 1], then it is a Hilbert-Schmidt operator,
and the rich theory of such operators can be applied. It is a compact operator,
which has a discrete spectrum, i.e., a countable multiset Spec(W ) of nonzero (real)
eigenvalues {λ1 , λ2 . . . } such that λn → 0. In particular, every nonzero eigenvalue
has finite multiplicity. Furthermore, it has a spectral decomposition
∑
(7.19) W (x, y) ∼ λk fk (x)fk (y),
k
where fk is the eigenfunction belonging to the eigenvalue λk with ∥fk ∥2 = 1. The
series on the right may not be almost everywhere convergent (only in L2 ), but one
has ∫
∑∞
2
λk = W (x, y)2 dx dy = ∥W ∥22 ≤ ∥W ∥2∞ .
k=1 [0,1]2
A useful consequence of this bound is that if we order the λi by decreasing absolute
value: |λ1 | ≥ |λ2 | ≥ . . . , then
∥W ∥2
(7.20) |λk | ≤ √ .
k
It also follows that for every other kernel U on the same probability space, the inner
product can be computed from the spectral decomposition:
∫ ∑ ∫
(7.21) ⟨U, W ⟩ = U (x, y)W (x, y) dx dy = λk U (x, y)fk (x)fk (y) dx dy
[0,1]2 k [0,1]2
∑
= λk ⟨fk , U fk ⟩
k
(where the series on the right is absolute convergent). The spectral decomposition
is particularly useful if we need to express operator powers: The spectral decom-
position of the n-th operator power is
∑
W ◦n (x, y) = λnk fk (x)fk (y),
k
and the series on the right hand side converges to the left hand side almost every-
where if n ≥ 2.
Proposition 7.17. The eigenfunctions fk belonging to a nonzero eigenvalue λk of
any function W ∈ W are bounded.
Proof. Indeed,
∫1
1  
 1
|fk (x)| =  W (x, y)fk (y) dy  ≤ ∥W ∥∞ ∥fk ∥1 . 
|λk | |λk |
0
 7.5. KERNEL OPERATORS 125

Some subgraph densities have nice expressions in terms of this spectrum. Gen-
eralizing (5.31), we have
(7.22) ∫ ∑
t(Cn , W ) = W (x1 , x2 ) · · · W (xn−1 , xn )W (xn , x1 ) dx1 . . . dxn = λnk .
k
[0,1]n

This expression is also valid for n = 2:

∫ ∑
(7.23) t(C2 , W ) = W (x, y)2 dx dy = ∥W ∥22 = λ2k .
k
[0,1]2

Furthermore, for every n ≥ 3,

∑
(7.24) txy (Pn•• , W ) = λn−1
k fk (x)fk (y)
k
almost everywhere. For n = 2 the left side is just W , and so we don’t always get
pointwise equality, only convergence in L2 .
For a general multigraph F = (V, E), we can express its density in a kernel W by
a rather hairy spectral formula (Lovász and Szegedy [2011]), which is nevertheless
useful. Substituting (7.19) in the definition of t(F, W ) and expanding, we get
∑ ∏ ∏
(7.25) t(F, W ) = λχ(e) Mχ (v),
χ:E→N∗ e∈E v∈V

where
∫1 ∏
(7.26) Mχ (v) = fχ(uv) (x) dx.
0 u: uv∈E
(One has to be careful, since (7.19) only converges in L2 , not necessarily almost
everywhere. But using (7.21) we can substitute for the values W (xi , xj ) one by one.)
This representation expresses t(F, W ) in an infinite “edge-coloring model”, which is
analogous to homomorphism numbers with the role of nodes and edges interchanged
(see Section 23.2 for a discussion of finite edge-coloring models): we sum over all
colorings of the edges with N; for every coloring, we take the product of nodeweights
and the product of edgeweights; the edgeweights are just the eigenvalues, and the
weight of a node is computed from the colors of the edges incident with it.
One consequence of (7.22) is that the cycle densities in W determine the spec-
trum of TW and vice versa. In fact, we don’t have to know all cycle densities: any
“tail” (t(Ck , W ) : k ≥ k0 ) is enough. This follows from Proposition A.21 in the
Appendix. In particular, we see that t(C2 , W ) = ∥W ∥22 is determined by the cycle
densities t(Ck , W ), k ≥ 3.
Exercise 7.18. (a) Let F = (V, E) be a multigraph without loops, and let us
subdivide each edge e ∈ E by m(e) ≥ 0 new nodes, to get a multigraph F ′ . Show
that using (7.24) the density of F ′ in W can be expressed by a formula similar
to (7.25). (b) Show that the densities of simple graphs in a kernel determine the
densities of multigraphs.
Exercise 7.19. Let W be a graphon. Prove that (a) all eigenvalues of TW
are contained in the interval [−1, 1]; (b) the largest eigenvalue is also largest
in absolute value; (c) at least one of the eigenvectors belonging to the largest
eigenvalue is nonnegative almost everywhere.
 CHAPTER 8

The cut distance

We have announced in the Introduction that we are going to define the distance
of two arbitrary graphs, so that this distance will reflect structural similarity. The
definition is quite involved, and we will approach the problem in several steps:
starting with two graphs on the same node set, then moving to graphs with the
same number of nodes (but on unrelated sets of nodes), then moving to the general
case. Finally, we extend the definition to kernels, where it will turn out simpler (at
least in words) than in the finite case.
In this section we consider dense graphs. The definitions are of course valid for
all graphs, but they give a distance of o(1) between two graphs with edge density
o(1), so they are not useful in that setting.

8.1. The cut distance of graphs

8.1.1. Norms of a matrix. Let A be an n × n matrix. There are a number
of norms that come up in various studies. We will need the ℓ1 -norm
1 ∑
n
(8.1) ∥A∥1 = 2 |Aij |,
n i,j=1
the ℓ2 or Frobenius norm
(1 ∑ n )1/2
(8.2) ∥A∥2 = 2
A2ij ,
n i,j=1
and the ℓ∞ -norm
(8.3) ∥A∥∞ = max |Aij |.
i,j

(Note the normalization for the ℓ1 and ℓ2 norms: when A an adjacency matrix, all
these norms are between 0 and 1.)
Our main tool will be a less standard norm, called the cut norm, which was
introduced by Frieze and Kannan [1999]. This is defined by
 ∑ 
1  
(8.4) ∥A∥ = 2 max  Aij .
n S,T ⊆[n]
i∈S,j∈T

It is clear that
(8.5) ∥A∥ ≤ ∥A∥1 ≤ ∥A∥2 ≤ ∥A∥∞ .
Example 8.1. Let A be an n × n matrix, whose entries are independent random
±1’s (with expectation 0).
∑ Then ∥A∥1 = ∥A∥2 = ∥A∥∞ = 1. On the other
hand, the expectation of i∈S,j∈T Aij is 0, and the variance is Θ(n2 ), and so the
∑ 
 
expectation of  i∈S,j∈T Aij  is Θ(n). The expectation of the maximum in (8.4)

127
128 8. THE CUT DISTANCE

is more difficult to compute, but using the Chernoff–Hoeffding inequality, one gets
that ∥A∥ < 4n−1/2 with high probability. 
Alon and Naor [2006] relate the cut norm of a symmetric matrix to its
Grothendieck norm (well known in functional analysis). It follows by the results
of Grothendieck that the cut norm is between two absolute constant multiples of
the Grothendieck norm. The Grothendieck norm can be viewed as a semidefinite
relaxation of the cut norm, and it is polynomial time computable to an arbitrary
precision. So we can compute, in polynomial time, an approximation of the cut
norm with a multiplicative error less than 2. We don’t go into the details of these
results here; in our setting it will be more important to approximate the cut norm
by a randomized sampling algorithm, to be described in Section 10.3.
We’ll say more about approximation of the cut norm in the more general setting
of graphons in Section 14.1.
8.1.2. Two graphs on the same set of nodes. Let G and G′ be two graphs
with a common node set [n]. From any of the matrix norms introduced above, the
norm of the difference of their adjacency matrices defines a distance between two
graphs. Two of these distances have special significance.
The ℓ1 distance
|E(G)△E(G′ )|
d1 (G, G′ ) = = ∥AG − AG′ ∥1
n2
is also called the edit distance (usually without the normalization). It can be
thought of as the fraction of pairs of nodes whose adjacency we have to toggle
to get from one graph to the other.
The cut metric derived from the cut norm can be described combinatorially
as follows. For an unweighted graph G = (V, E) and sets S, T ⊆ V , let eG (S, T )
denote the number of edges in G with one endnode in S and the other in T (the
endnodes may also belong to S ∩ T ; so eG (S, S) = 2eG (S) is twice the number of
edges spanned by S). For two graphs G and G′ on the same node set [n], we define
their cut distance (as labeled graphs) by
|eG (S, T ) − eG′ (S, T )|
d (G, G′ ) = max = ∥AG − AG′ ∥ .
S,T ⊆V (G) n2
In this setting dividing by |S|×|T | instead of n2 might look more natural. However,
dividing by |S|×|T | would emphasize small sets too much, and the maximum would
be attained when |S| = |T | = 1. With our definition, the contribution of a pair
S, T is at most |T ||S|/n2 (for simple graphs).
It is easy to see that d (G, G′ ) ≤ d1 (G, G′ ), and in general the two distances
are quite different. For example, if G and G′ are two independent random graphs
on [n] with edge probability
√ 1/2, then with high probability d1 (G, G′ ) ≈ 1/2 but
′
d (G, G ) = O(1/ n).
We will have to define the distance of two weighted graphs G and G′ on the
same node set V , but with possibly different nodeweights. In this case, we have
to add a term accounting for the difference in their node weighting. To simplify
notation, let αi = αi (G)/αG , αi′ = αi (G′ )/αG′ , βij = βij (G) and βij ′
= βij (G′ ).
Then we define
∑ ∑
(8.6) d1 (G, G′ ) = |αi − αi′ | + |αi αj βij − αi′ αj′ βij
′
|
i∈V i,j∈V
 8.1. THE CUT DISTANCE OF GRAPHS 129

and
∑  ∑ 
 ′ 
(8.7) d (G, G′ ) = |αi − αi′ | + max  (αi αj βij − αi′ αj′ βij ).
S,T ⊆V
i∈V i∈S,j∈T

It is easy to check that these formulas define metrics, and they specialize to the “old”
definitions when the nodeweights are 1 and the edgeweights are 0 or 1. Another
special case worth mentioning is when the nodeweights of the two graphs are the
same: in this case, the first term in both definitions disappears, and inside the
′
second term, we get the slightly simpler expression αi αj (βij − βij ). We note,
′
furthermore, that since G and G can be represented as points in the same finite
dimensional space, all usual distance functions on the set of weighted graphs on the
same set of nodes would give the same topology.
Example 8.2. Let Hn denote the complete graph on [n], where all nodes have
weight 1 and all edges have weight 1/2. Then for a random graph G = G(n, 1/2)
on the same node set, we have d (G, Hn ) = o(1) with high probability. 
8.1.3. Two graphs with the same number of nodes. If G and G′ are
unlabeled unweighted graphs on possibly different node sets but of the same cardi-
nality n, then we define their distance by
(8.8) δb (G, G′ ) = min d (G,
b Gb ′ ),
bG
G, b′

where G b and Gb ′ range over all labelings of G and G′ by 1, . . . , n, respectively. (Of

course, it would be enough to fix a labeling for one of the graphs and minimize over
all labelings of the other.)
The hat above the δ indicates that the “ultimate” definition will be somewhat
different. Indeed, handling of this quantity δb (G, G′ ) is quite difficult, due to the
min-max in the definition.
8.1.4. Two arbitrary graphs. Let G = (V, E) and G′ = (V ′ , E ′ ) be two
graphs with (say) V = [n] and V ′ = [n′ ]. To define their distance, recall that
for every graph G and positive integer m, the graph G(m) is obtained from G by
replacing each node of G by m nodes, where two new nodes are connected if and
only if their predecessors were. Using this operation, we can change the graphs
so that they have the same number of nodes, by replacing them with G(n′ ) and
G′ (n), or more generally, by G(kn′ ) and G′ (kn) for any k ∈ N. Now we can use
the distance δb to define the distance
( )
δ (G, G′ ) = lim δb G(kn′ ), G′ (kn) .
k→∞
A more complicated but “finite” definition of the same quantity can be given
as follows (cf. Exercise 8.5). A fractional overlay of G and G′ is a nonnegative
∑n′ ∑n
n × n′ matrix X = (Xiu ) such that u=1 Xiu = n1 and i=1 Xiu = n1′ . If n = n′
and σ : V → V ′ is a bijection, then Xiu = n1 1(σ(i) = u) is a fractional overlay
(which in this case is an honest-to-good overlay). We denote by X (G, G′ ) the set
of all fractional overlays.
Fixing a fractional overlay X, we can define a generalization of the labeled cut
distance:
 ∑ ( )

(8.9) d (G, G′ , X) = max ′  Xiu Xjv 1(ij ∈ E) − 1(uv ∈ E ′ ) .
Q,R⊆V ×V
iu∈Q, jv∈R
130 8. THE CUT DISTANCE

The distance of the two graphs can be described by optimizing over fractional
overlays:
(8.10) δ (G, G′ ) = min d (G, G′ , X)
X∈X (G,G′ )

One can generalize this to weighted graphs. Let G = (V, E) and G′ = (V ′ , E ′ )

be two weighted graphs with normalized nodeweights αi = αi (G) and αu′ = αu (G′ )
′
(so that αG = αG′ = 1), and edgeweights βij = βij (G) and βij = βij (G′ ). A
fractional overlay of G and G′ is defined as a nonnegative n × n′ matrix X such
∑n′ ∑n
that u=1 Xiu = αi (G) and i=1 Xiu = αu (G′ ). We define
 ∑ 
 
(8.11) d (G, G′ , X) = max ′  ′
Xiu Xjv (βij − βuv )
Q,R⊆V ×V
iu∈Q, jv∈R

′
and then δ (G, G ) can be defined by the same formula (8.10). This formula can
be rephrased as follows, using two more V × V ′ matrices Y and Z:
 ∑ 
 
(8.12) δ (G, G′ ) = min ′ max  ′
Yiu Zjv (βij − βuv ).
X∈X (G,G ) 0≤Y,Z≤X
i,j∈V,u,v∈V ′

Indeed, the absolute value on the right is a convex function of the entries of Y
and Z, and so it is maximized when every entry is equal to either 0 or to the
corresponding entry of X.
To illuminate definition (8.10) a little, we can think of a fractional overlay as a
probability distribution χ on V × V ′ whose marginals are uniform. In other words,
it is a coupling of the uniform distribution on V with the uniform distribution on
V ′ . Select two pairs (i, u) and (j, v) from the distribution χ. Then (8.9) expresses
some form of correlation between ij being an edge and uv being an edge.
One word of warning: δ is only a pseudometric, not a true metric, because
δ (G, G′ ) may be zero for different graphs G and G′ . This is the case e.g. if
G′ = G(k) for some k (cf. Exercise 8.6).
We have to discuss a technical problem, for which only partial results are avail-
able (but these will be enough for our purposes). If G and G′ have the same number
of nodes, then the definition of δ may give a value different from their δb distance.
It is trivial that
δ (G, G′ ) ≤ δb (G, G′ ),
but how much larger can the right side be? It may be larger (see Exercise 8.8.
Perhaps the increase is never larger than a factor of 2, but this is open. To prove
anything nontrivial requires tools to be developed later; in Section 9.4 we are going
to prove, among others, the (rather weak) inequality
45
δb (G, G′ ) ≤ √ .
− log δ (G, G′ )
(One important consequence of this weak inequality will be that any Cauchy se-
quence of graphs in the δ distance is also a Cauchy sequence in the δb distance.)

Example 8.3. Let K denote the graph with a single node of weight 1, endowed
with a loop with weight 1/2. Then for a random graph G = G(n, 1/2), we have
δ (G, K) = o(1) with high probability. 
 8.2. CUT NORM AND CUT DISTANCE OF KERNELS 131

Exercise 8.4. Let A be a symmetric matrix. Show that restricting the pairs
(S, T ) in the definition (8.4) of the cut norm in any of the following ways will
decrease it by a small factor only: (a) T = S, by at most 2; (b) T ∩ S = ∅, by at
most 4; (c) T = [n] \ S, by at most 6; (d) |S|, |T | ≥ n/2, by at most 4.
Exercise 8.5. Prove that the definitions of δ (G, G′ ) through blow-ups and
through fractional overlays lead to the same value.
Exercise 8.6. Let G1 and G2 be two simple graphs with δ (G1 , G2 ) = 0. Prove
that there is a simple graph G and n1 , n2 ≥ 1 such that Gi ∼
= G(ni ).
Exercise 8.7. Let A be a symmetric n × n matrix with all entries in [−1, 1]. Let
A′ be obtained from A by deleting a row and the corresponding column. Prove
that  
 ′  2
∥A∥ − ∥A ∥  ≤ .
n
Exercise 8.8. (a) Let H denote the graph on two nonadjacent nodes, with a loop
at each of them. Prove that δb (H, K2 ) = 1/4 but δ (H, K2 ) = 1/8. (b) Prove
b n,n , K n,n ) > δ(Kn,n , K n,n ).
that if n is odd, then δ(K

8.2. Cut norm and cut distance of kernels

After the rather heavy going with the cut distance for graphs, it sounds fright-
ening that we want to extend all this to kernels. But in fact, the definitions become
simpler and more transparent. (This is not the last time when graphons will provide
a more user-friendly environment.)

8.2.1. Cut norm. We define the cut norm on the linear space W of kernels
by
∫ 
 
 
(8.13) ∥W ∥ = sup  W (x, y) dx dy 
S,T ⊆[0,1] 
S×T

where the supremum is taken over all measurable subsets S and T . It is sometimes
convenient to use the corresponding metric d (U, W ) = ∥U − W ∥ .
The cut norm is a norm; this is easy to prove using standard analysis. Simi-
larly as in the case of matrices, we have the trivial inequalities between the most
important norms of a kernel in W1 :
(8.14) ∥W ∥ ≤ ∥W ∥1 ≤ ∥W ∥2 ≤ ∥W ∥∞ ≤ 1.
1/2
In the opposite direction, we have trivially ∥W ∥2 ≤ ∥W ∥1 (showing that ∥.∥1
and ∥.∥2 define the same topology on W1 ), but the other two norms in the formula
above define different topologies. However, for a stepfunction U with k steps we
have the trivial inequality
(8.15) ∥U ∥1 ≤ k 2 ∥U ∥ .
√
It can be shown, in fact, that the coefficient k 2 can be replaced by 2k (see Janson
[2010], Remark 9.8, and also our Exercise 8.18); but the inequality above will be
enough for us.
There is some natural notation that goes with this norm. For every set R ⊆ W0 ,
we define its ε-neighborhood in the cut-norm
B (R, ε) = {W ∈ W0 : d (W, R) < ε} = {W ∈ W0 : (∃U ∈ R) d (W, U ) < ε}.
132 8. THE CUT DISTANCE

We define the ε-neighborhood B1 (R, ε) in the L1 -norm analogously. (We defined

all this in the graphon space W0 , where we need this notation. One could of course
take other sets of kernels as the universe.)

8.2.2. Cut distance of unlabeled kernels. Kernels, defined on the fixed set
[0, 1], correspond to labeled graphs. Just as for graphs, we introduce an “unlabeled”
version of the cut norm, by finding the best overlay of the underlying sets. Let S [0,1]
denote the set of measure preserving maps [0, 1] → [0, 1], and let S[0,1] denote the
set of all invertible measure preserving maps [0, 1] → [0, 1] (the inverse of such a
map is known to be measure preserving as well, so S[0,1] is a group; see Appendix
A.3.2). We define the cut distance of two kernels by
(8.16) δ (U, W ) = inf d (U, W φ ),
φ∈S[0,1]
( )
(where W φ (x, y) = W φ(x), φ(y) ). It is easy to see that either one of the following
expressions could be used to define the cut distance:
(8.17) δ (U, W ) = inf d (U φ , W ) = inf d (U, W φ )
φ∈S[0,1] φ∈S [0,1]

= inf d (U ψ , W φ ).
φ,ψ∈S [0,1]

We will prove the much less trivial fact that in the last expression the infimum is
attained: Theorem 8.13 below establishes this in larger generality, for all norms
satisfying some natural conditions.
The distance δ of kernels is only a pseudometric, since different kernels can
have distance zero. (Such pairs of kernels will turn out exactly the weakly isomor-
phic pairs, but this will take more work to prove.) We can identify two kernels
whose cut distance is 0, to get the set Wf of unlabeled kernels. We define the sets
f0 and W
W f1 analogously.
Going into all the complications with using the cut norm and then minimizing
over measure preserving transformations is justified by the important fact that the
metric δ defines a compact metric space on graphons. We will state and prove
this fact in Section 9.3.
One main advantage in using graphons instead of graphs is that many formu-
las and proofs become much simpler and more transparent. (Just compare the
definition (8.16) of the distance of two graphons with the definition (8.12) of the
analogous quantity for two weighted graphs!) When going from graphs to graphons
via the correspondence G 7→ WG , we may pay a prize by having to estimate how
much error we make by this. This will indeed require extra work in some cases, but
in other cases we will be lucky, and no error will be made. For example, equation
(7.2) shows that homomorphism numbers “from the left” don’t change when we
replace G by WG . The next lemma shows that the situation is similar with the
δ distance. (We will not always be so lucky; Section 12.4.4 will be devoted to
estimating this kind of error for multicuts.)
Lemma 8.9. For any two weighted graphs H and H ′
δ (H, H ′ ) = δ (WH , WH ′ ).
(
Proof.
) (Let φ : [0, 1] →) [0, 1] be a measure preserving map. Let Si : i ∈
V (H) and Tu : u ∈ V (H ′ ) be the partitions of [0, 1] into the steps of WH and
 8.2. CUT NORM AND CUT DISTANCE OF KERNELS 133

( )
WH ′ . Define Xiu = λ Si ∩ φ(Tu ) , then the matrix (Xiu ) is a fractional overlay
of H and H ′ . Conversely, every fractional overlay can be obtained from a measure
preserving map this way.
We claim that for this measure preserving map and the corresponding fractional
overlay we have
 ∑   ∫ 
 ′   φ 
(8.18) max ′  Xiu Xjv (βij − βuv ) = sup  (WH − WH ′ ).
Q,R⊆V ×V Y,Z⊆[0,1]
iu∈Q, jv∈R Y ×Z

For every Q ⊆ V ×V , let SQ = ∪(i,u)∈Q Si ∩φ(Tu ). Then for a fixed Q, R ⊆ V ×V ′ ,

′

it is easy to check that

∑ ∫
′ φ
Xiu Xjv (βij − βuv )= (WH − WH ′ ).

iu∈Q, jv∈R SQ ×SR

On the other hand, if Ziu = λ(Z ∩ Si ∩ φ(Tu )) and Yiu = λ(Y ∩ Si ∩ φ(Tu )), then
0 ≤ Yiu , Ziu ≤ Xiu , and
∫ ∑
φ ′
(WH − WH ′) = Yiu Zjv (βij − βuv ).
Y ×Z i,j∈V,u,v∈V ′

So the definition (8.10) of δ (H, H ′ ) implies the direction ≤ in (8.18), while formula
(8.12) implies reverse direction. This proves (8.18), from which the Lemma follows.

8.2.3. Maxima versus suprema: cut norm. One price we have to pay for
working with infinite objects like graphons is that when maximizing a function over
an infinite set of objects (e.g. subsets), we don’t necessarily have a maximum, only
a supremum; hence we have to work with approximate optima. With two impor-
tant definitions, the cut norm and the cut distance, we don’t have this difficulty.
(The Compactness Theorem 9.23 will provide another powerful tool to avoid such
problems in many cases.) Next we prove this for the cut norm, and at the end of
this chapter, for the cut distance. This would not be absolutely necessary: in most
cases, we could just carry along an arbitrarily small error term. Nevertheless, it
makes sense to include these facts in this book: if you want to work with these
notions, you might as well work with them as conveniently as possible. The next
lemma also provides a useful expression for the cut norm.
Lemma 8.10. For any kernel W ∈ W, the optima
∫ 
 
(8.19) sup  W (x, y) dx dy 
S,T ⊆[0,1]
S×T

and
 ∫ 
 
(8.20) sup  f (x)g(y)W (x, y) dx dy 
f,g: [0,1]→[0,1]
[0,1]2

are attained, and they are both equal to ∥W ∥ .

The sets S, T and the functions f, g are tacitly assumed to be measurable. We
can write the expression to be maximized in (8.19) as ⟨1S , TW 1T ⟩, and in (8.20), as
⟨f, TW g⟩ (where TW is the operator defined by (7.18)). The assertion of the lemma
is equivalent to saying that the optimum in (8.20) is attained, and it is attained
134 8. THE CUT DISTANCE

by 0-1 valued functions f and g. I am grateful to Svante Janson for suggesting a

simplification of the proof that follows.

Proof. Let D = supf,g ⟨f, TW g⟩. We start with proving that this supremum is
attained by appropriate functions f and g. Let fn , gn : [0, 1] → [0, 1] (n = 1, 2, . . . )
be functions such that ⟨fn , TW gn ⟩ → D. The set of functions [0, 1] → [0, 1] are
weak*-compact, which means that by selecting a subsequence, we may assume that
it tends to a limit f : [0, 1] → [0, 1] in the sense that ⟨fn , h⟩ → ⟨f, h⟩ for every
h ∈ L1 [0, 1]. Similarly, we can go to a further subsequence to assume that gn
converges to a function g in the same sense. It is easy to see that f and g are
bounded (perhaps after changing them on a null set). Now we claim that
∫ ∫
fn (x)gn (y)W (x, y) dx dy −→ f (x)g(y)W (x, y) dx dy.
[0,1]2 [0,1]2

This convergence is trivial when W = 1S×T for two measurable sets S, T ⊆ [0, 1].
Hence it follows when W is stepfunction, since stepfunctions are linear combinations
of a finite number of functions of the type 1S×T . Hence it follows for every kernel,
since every kernel can be approximated by stepfunctions in L1 ([0, 1]2 ), and the
factors fn , gn , f, g are bounded. This implies that ⟨f, TW g⟩ = D.
Next we show that the maximizing functions f and g can be chosen to be 0-1
valued. Let S = {x : 0 < f (x) < 1}, and suppose that λ(S) > 0. Define
( )
fs (x) = f (x) + s min f (x), 1 − f (x) .
Then for −1 ≤ s ≤ 1, the function fs satisfies 0 ≤ fs ≤ 1, and hence, by the
maximality property of f , we have ⟨fs , Tw g⟩ ≤ ⟨f, Tw g⟩. Since ⟨fs , Tw g⟩ is a linear
function of s and equality holds for s = 0, we must have equality for ( all values
) of
s, in particular for s = 1, and so we can replace f by f1 (x) = min 1, 2f (x) . Re-
peating this construction, we get a sequence of optimizing functions that monotone
converges to the 0-1 valued function f = 1(f (x) > 0). So we can replace f by f ,
and similarly we can replace g by a 0-1 valued function g. 

8.2.4. Operator norms and cut norm. While the cut norm is best suited
for combinatorial purposes, it is equivalent to more traditional norms, such as the
operator norm of TW as an operator L∞ → L1 , as the following simple lemma
shows:
Lemma 8.11. For every kernel W , we have
∥W ∥ ≤ ∥TW ∥∞→1 ≤ 4∥W ∥ .
Proof. By definition,
 
∥TW ∥∞→1 = sup ∥TW g∥1 = sup ⟨f, TW g⟩ = sup ⟨f, TW g⟩.
−1≤g≤1 −1≤f,g≤1 −1≤f,g≤1

Comparing this expression with (8.20), we get the first inequality. For the second,
we write
∥TW ∥∞→1 = sup ⟨f − f ′ , TW (g − g ′ )⟩.
0≤f,f ′ ,g,g ′ ≤1
Here
⟨f − f ′ , TW (g − g ′ )⟩ = ⟨f, TW g⟩ − ⟨f ′ , TW g⟩ − ⟨f, TW g ′ ⟩ + ⟨f ′ , TW g ′ ⟩ ≤ 4∥T ∥ . 
 8.2. CUT NORM AND CUT DISTANCE OF KERNELS 135

There are many other variations on the definition which give norms that are
some constant factor away from the cut norm; these are useful since in some proofs
they come up more directly than the cut norm. Some of these are stated as exercises
at the end of this section.
There are other well-studied operator norms that are topologically equivalent
to the cut norm (even though they are not equivalent up to a constant factor). The
Schatten p-norm Sp (TW ) of a kernel operator TW is defined as the ℓp -norm of the
sequence of its eigenvalues. For an even integer p, these can be expressed in terms
of homomorphism densities:
Sp (TW ) = t(Cp , W )1/p .
(It is not trivial that t(C2r , U )1/(2r) is a norm, i.e., it is subadditive (the other
defining properties of a norm are easy). In Proposition 14.2 we’ll describe a method
to prove that Schatten norms are indeed norms, along with certain more general
norms defined by graphs.)
These norms define the same topology on W1 as the cut norm. We prove the
explicit relationship for the case p = 4, which we need.
Lemma 8.12. For every graphon U ∈ W1 , ∥U ∥4 ≤ t(C4 , U ) ≤ 4∥U ∥ .
Proof. The second inequality is a special case of Lemma 10.23. To prove the
first inequality, we use
∥U ∥ = sup ⟨f, TU g⟩,
0≤f,g≤1
where
⟨f, TU g⟩ ≤ ∥f ∥2 ∥TU g∥2 ≤ ∥TU g∥2 = ⟨TU g, TU g⟩1/2 = ⟨g, TU2 g⟩1/2
1/2 1/2 1/2
= ⟨g, TU ◦U g⟩1/2 ≤ ∥g∥2 ∥TU ◦U ∥2→2 ≤ ∥TU ◦U ∥2→2 ≤ ∥U ◦ U ∥2
= t(C4 , U )1/4 . 
8.2.5. Minima versus infima: cut distance. The last result in this section
is of a similar nature as Lemma 8.10: we prove that the “inf” in the last quantity
in formula 8.17 above is in fact a “min”. This was proved by Bollobás and Riordan
[2009]. An analogous result for the L1 -norm was proved by Pikhurko [2010]. With
later applications in mind, we prove it in greater generality.
The construction that gives the cut distance δ from the cut norm can be
applied to any other norm on W that is invariant under maps W 7→ W φ for all
φ ∈ S[0,1] . We will call such a norm invariant. For an invariant norm N on the
linear space W, we define
δN (U, W ) = inf N (U − W φ ).
φ∈S[0,1]

We call this function the distance derived from N . The distances δN will be inter-
esting for us mainly in the cases when N = ∥.∥ , N = ∥.∥1 and N = ∥.∥2 . The
corresponding unlabeled distances are δ , δ1 and δ2 .
Since the norm is invariant under measure preserving bijections, we have N (U −
−1
W φ ) = N (U φ − W ), implying that δN (U, W ) = δN (W, U ). It is trivial that the
triangle inequality holds for δN , so it is a semimetric (and clearly it is not a true
metric, since δN (U, U φ ) = 0 for every measure preserving map φ ∈ S[0,1] ).
We call a norm N smooth, if it is continuous in the topology of pointwise
convergence in W. In other words, for every sequence of kernels (Wn ) such that
136 8. THE CUT DISTANCE

Wn ∈ W1 and Wn → 0 almost everywhere, we have N (Wn ) → 0. This implies that

if Wn → W almost everywhere, then N (Wn ) → N (W ). The L1 , L2 and cut norms
are smooth, but L∞ is not.
We have defined invariance of a norm using measure preserving bijections, but
(at least for smooth norms) this implies invariance under all measure preserving
maps φ : [0, 1] → [0, 1]. This is easy to see for stepfunctions W , since for any
measure preserving map W φ is a stepfunction with the same number of steps,
same size of steps, and same function value on these steps as W , and hence there
is a bijective measure preserving map ψ such that W φ = W ψ . For a general kernel
W ∈ W, we have a sequence of stepfunctions Wn such that Wn → W almost
everywhere, and then also Wnφ → W φ almost everywhere. By the smoothness of
N this implies that N (Wn ) → N (W ) and N (Wnφ ) → N (W φ ). Since we know that
N (Wnφ ) = N (Wn ), it follows that N (W φ ) = N (W ).
Let us note that an invariant norm N on W defines a norm on bounded sym-
metric measurable functions on any standard probability space. Indeed, if (Ω, A, π)
is such a space, then there is a measure preserving map ψ : [0, 1] → Ω, and then
we can define N (W ) = N (W ψ ) for every bounded symmetric measurable function
W : Ω × Ω → R. This value will not depend on the choice of ψ, which follows
easily from the invariance of N .
One can also give a more probabilistic description of the distance δN , using
coupling measures (see Appendix A.3). For every coupling measure µ between two
copies of [0, 1], the two projection maps π, ρ : [0, 1]2 → [0, 1] (where [0, 1]2 is
equipped with the measure µ and [0, 1], with the Lebesgue measure) are measure
preserving. So for every kernel U , the function U π is a kernel on the probability
space ([0, 1]2 , B, µ), and similarly for the projection ρ. As remarked above, N defines
a norm on kernels on ([0, 1]2 , B, µ); we denote this norm by Nµ . It is easy to see
that for every kernel U on [0, 1], we have
(8.21) Nµ (U π ) = N (U ).
After this explanation, we can state the theorem:
Theorem 8.13. Let N be a smooth invariant norm on W. Then we have the
following alternate expressions for the unlabeled distance derived from N :
(8.22) δN (U, W ) = inf N (U − W φ ) = inf N (U − W φ )
φ∈S[0,1] φ∈S [0,1]

= inf N (U ψ − W ) = inf N (U ψ − W )
ψ∈S[0,1] ψ∈S [0,1]

= inf N (U ψ − W φ ) = min N (U ψ − W φ ),
φ,ψ∈S[0,1] φ,ψ∈S [0,1]

and
(8.23) δN (U, W ) = min Nµ (U π − W ρ ),
µ

where µ ranges over all coupling measures on [0, 1]2 .

Proof. The equality of the first expressions in each line of (8.22) follows from
the fact that invertible measure preserving maps form a group.
First, let U and W be stepfunctions. As used before, the kernel W φ for any
measure preserving map φ can be realized by an invertible measure preserving map,
 8.2. CUT NORM AND CUT DISTANCE OF KERNELS 137

which implies that in each line of 8.22, the two expressions are equal. Equation
(8.23) follows similarly easily in this case.
Second, we consider arbitrary functions U, W ∈ W, and prove the formulas with
the two occurrences of “min” replaced by “inf”. Let (Un ) and (Wn ) be sequences
of stepfunctions converging almost everywhere to U and W , respectively. Then
N (Un − U ) → 0 by the smoothness of N , and similarly for W . Since N (Unφ − U φ ) =
N (Un − U ) for every measure preserving map φ, this implies that
inf N (Un − Wnφ ) = inf N (Un − Wnφ ) → inf N (U − W φ ) = δN (U, W ),
φ∈S [0,1] φ∈S[0,1] φ∈S[0,1]

and also that

inf N (Un − Wnφ ) → inf N (U − W φ ),
φ∈S [0,1] φ∈S [0,1]

which proves the equality in the first line of (8.22). The other equations follow
similarly.
However, this argument only gives an “inf” in the last two expressions for δN .
To prove that it is in fact a minimum, we begin with (8.23). The space of coupling
measures is compact in the weak topology, so it suffices to show that Nµ (U π − W ρ ),
as a function of µ, is lower semicontinuous. This means that if µn → µ weakly
(where µ and µn are coupling measures), then for every two kernels U and W , we
have
(8.24) lim inf Nµn (U π − W ρ ) ≥ Nµ (U π − W ρ ).
n

As a first step, we prove that Nµn (V ) → Nµ (V ) for every continuous function V .

Let fn and f be the functions representing the measures µn and µ as in Proposition
fn f
A.6(iv). Then Nµn (V ( ) = N (V )), and (Nµ (V ) = N) (V ).f Since V is continuous, we
have V (x, y) = V fn (x), fn (y) → V f (x), f (y) = V (x, y) for almost all x, y ∈
fn

[0, 1]2 . By our assumption on the norm N , this implies that Nµn (V ) → Nµ (V ). As
a special case, we get (8.24) for continuous kernels U and W .
Let U, W : [0, 1] × [0, 1] → R be arbitrary kernels, and fix any ε > 0. There
are continuous kernels Uk and Wk (k = 1, 2, . . . ) such that Uk → U and Wk → W
almost everywhere. By the smoothness of N , we can fix k large enough so that
N (Uk − U ) ≤ ε and N (W k − W ) ≤ ε.
By the special case proved above, we know that
Nµn (Ukπ − Wkρ ) → Nµ (Ukπ − Wkρ ) (n → ∞),
and we can fix n so that |Nµn (Ukπ − Wkρ ) − Nµ (Ukπ − Wkρ )| ≤ ε. Then, using (8.21),
Nµ (U π − W ρ ) ≤ Nµ (Ukπ − Wkρ ) + Nµ (Ukπ − U π ) + Nµ (Wkρ − W ρ )
= Nµ (Ukπ − Wkρ ) + N (Uk − U ) + N (Wk − W )
≤ Nµ (Ukπ − Wkρ ) + 2ε.
Here, by the choice of n,
Nµ (Ukπ − Wkρ ) ≤ Nµn (Ukπ − Wkρ ) + ε
≤ Nµn (U π − W ρ ) + Nµn (Ukπ − U π ) + Nµn (Wkρ − W ρ ) + ε
= Nµn (U π − W ρ ) + N (Uk − U ) + N (Wk − W ) + ε
≤ Nµn (U π − W ρ ) + 3ε.
138 8. THE CUT DISTANCE

Combining these inequalities, we get that Nµ (U π − W ρ ) ≤ Nµn (U π − W ρ ) + 5ε if

n is large enough. This proves (8.24) and thereby the existence of the minimum in
(8.23).
The existence of the minimum in (8.22) follows easily now. Let µ be a coupling
measure such that δN (U, W ) = Nµ (U π − W ρ ). Let σ be a measure preserving
bijection from [0, 1] with the Lebesgue measure into [0, 1]2 with the measure µ, and
let π and ρ be the projections of [0, 1]2 to the two coordinates. The fact that µ is
a coupling measure implies that the compositions φ = σπ and ψ = σρ are measure
preserving, and
N (U φ − W ψ ) = Nµ (U π , W ρ ) = δN (U, W ). 
This theorem has an important corollary:
Corollary 8.14. For any smooth and invariant norm N on W, we have
δN (U, W ) = 0 if and only if there exist maps φ, ψ ∈ S [0,1] such that U ψ = W φ
almost everywhere.
f
This corollary allows us to consider the distances δ1 and δ2 as defined on W
(just as δ ). In other words, the condition δN (U, W ) = 0 is independent of N ,
and identifying such pairs of kernels gives the same space for every smooth and
invariant norm N .
Exercise 8.15. Let H and H ′ be two weighted graphs on the same set with
αH = αH ′ = 1, with the same edgeweights, but different nodeweights. Prove that
δ1 (H, H ′ ) ≤ ∥α(H) − α(H ′ )∥1 .
Exercise 8.16. Prove that for every kernel W ∈ W1 and k ≥ 2, we have
t(C4 , W ) ≤ t(C2k , W )1/(2k) ≤ t(C4 , W )1/4 .
Exercise 8.17. Let σ : [0, 1] → [0, 1] range over maps that can be obtained as
follows: we split (0, 1] into the intervals Ik = ( k−1
n
, nk ] (k = 1, . . . , n) and permute
these intervals arbitrarily. Prove that for every smooth and invariant norm N we
have
δN (U, W ) = inf N (U − W σ ).
σ
√
Exercise 8.18. Improve the coefficient in (8.15) to (a) 2k, (b) 2 k (this is not
easy!).
∫
Exercise 8.19. Show that if we use the formula supS S×S W to define a norm
(which is only a constant factor off the cut norm), then the supremum is not
always attained (Laczkovich [1995]).
Exercise 8.20. Show by examples that one could not replace any of the “inf”-s
by “min” in Theorem 8.13.
Exercise 8.21. Show that if N is the L∞ norm on W, then even the “easy” part
of Theorem 8.13 fails: there are two kernels U and W such that inf φ∈S [0,1] ∥U −
W φ ∥∞ = 0 but inf φ∈S[0,1] ∥U − W φ ∥∞ > 0.

8.3. Weak and L1 -topologies

We end this discussion of graphon distances with a further somewhat technical
issue. The topology on W defined by the cut norm is certainly different from the
topology defined by the L1 -norm; there are, however, some nontrivial relationships
between them. We will discuss these in larger generality and detail in Section 14.2,
but a few simple facts can be proved here easily, and we will need some of them
soon.
 8.3. WEAK AND L1 -TOPOLOGIES 139

The key to relating the cut norm to other topologies is the following lemma.
Lemma 8.22. Suppose that ∥Wn ∥ → 0 as n → ∞ (Wn ∈ W1 ). Then∫ for every
function Z ∈ L1 ([0, 1]2 ), ∥ZWn ∥ → 0. In particular, ⟨Z, Wn ⟩ → 0 and S Wn → 0
for every measurable set S ⊆ [0, 1]2 .
Proof. If Z is the indicator function of a rectangle, these conclusions follow
from the definition of the ∥.∥ norm. Hence the conclusion follows for stepfunc-
tions, since they are linear combinations of a finite number of indicator functions of
rectangles. Then it follows for all integrable functions, since they are approximable
in L1 ([0, 1]2 ) by stepfunctions. 

A uniformly bounded sequence of kernels Wn ∈ W is called weak* convergent

to a kernel W if ⟨Wn , U ⟩ → ⟨W, U ⟩ ∫for every integrable
∫ function U : [0, 1]2 → R.
This is equivalent to requiring that S×T Wn → S×T W for all measurable sets S
and T . This sound almost like convergence in the cut norm, but it is not the same!
Lemma 8.22 implies that convergence in the ∥.∥ norm implies weak* convergence.
However, weak* convergence does not imply convergence in the cut norm (Exercise
8.26; an interesting counterexample follows from Example 11.41).
Since ∥.∥ ≤ ∥.∥1 , the cut norm is continuous with respect to the L1 -norm. The
converse is not true (recall the example of random graphs from the Introduction,
Figure 1.5), but the following fact, proved and used by Lovász and Szegedy [2010a],
shows that it is at least lower semicontinuous:
Proposition 8.23. Let Wn → W in the cut norm (Wn , W ∈ W1 ). Then
lim inf ∥Wn ∥1 ≥ ∥W ∥1 .
n→∞

Proof. Let Y = sgn(W ). Then by Lemma 8.22,

∥Wn ∥1 ≥ ⟨Wn , Y ⟩ → ⟨W, Y ⟩ = ∥W ∥1 . 

As noted above, we cannot claim in Proposition 8.23 that ∥Wn ∥1 → ∥W ∥1 .

However, as further applications of Lemma 8.22, we can state two weaker facts in
this direction (Lovász and Szegedy [2010a]); the first of these was also proved (in a
slightly different form) by Pikhurko [2010].
Proposition 8.24. Let W be a 0-1 valued graphon and let (Wn ) be a sequence of
graphons such that ∥Wn − W ∥ → 0. Then ∥Wn − W ∥1 → 0.
Proof. By Lemma 8.22, we have
∫ ∫ ∫ ∫
∥Wn − W ∥1 = Wn + (1 − Wn ) → W+ (1 − W ) = 0.
{W =0} {W =1} {W =0} {W =1}

Proposition 8.25. Suppose that Un → U in the cut norm as n → ∞ (U, Un ∈ W0 ).

Then for every W ∈ W0 there is a sequence of graphons Wn ∈ W0 such that
Wn → W in the cut norm, and ∥Un − Wn ∥1 → ∥U − W ∥1 .
It is important that we want Wn ∈ W0 ; if we only wanted kernels, we could
take simply Wn = W + Un − U .
140 8. THE CUT DISTANCE

Proof. First we consider the case when U ≥ W . Let

{
W (x, y)/U (x, y) if U (x, y) > 0,
Z(x, y) =
0 otherwise.
and define Wn = ZUn . Trivially Wn ∈ W0 , W = ZU , and
∥W − Wn ∥ = ∥Z(U − Un )∥ → 0
by Lemma 8.22. Furthermore, using that Un ≥ Wn and U ≥ W , we get
   
∥Un −Wn ∥1 −∥U −W ∥1  = ∥Un −Wn ∥ −∥U −W ∥  ≤ ∥U −Un ∥ +∥W −Wn ∥ .
This implies that ∥Un − Wn ∥1 → ∥U − W ∥1 as n → ∞.
The case when U ≤ W follows by a similar argument, replacing U, W, Un by
1 − U, 1 − W, 1 − Un .
Finally, in the general case, consider the graphon V = max(U, W ). Then
clearly ∥U − V ∥1 + ∥V − W ∥1 = ∥U − W ∥1 . Since U ≤ V , there exists a sequence
(Vn ) of graphons such that ∥Vn − V ∥ → 0 and ∥Vn − Un ∥1 → ∥V − U ∥1 . Since
V ≥ W , there is a sequence (Wn ) of graphons such that ∥Wn − W ∥ → 0 and
∥Wn − Vn ∥1 → ∥W − V ∥1 . Hence
lim sup ∥Un − Wn ∥1 ≤ lim sup ∥Un − Vn ∥1 + lim sup ∥Vn − Wn ∥1
n→∞ n→∞ n→∞
= ∥U − V ∥1 + ∥V − W ∥1 = ∥U − W ∥1 .
Using Proposition 8.23, the lemma follows. 
Exercise 8.26. Show that weak* convergence of a sequence of graphons does not
imply convergence in the cut norm.
Exercise 8.27. Show that ∥Wn ∥ → 0 (Wn ∈ W1 ) does not imply that ∥Wn ∥1 →
0.
 CHAPTER 9

Szemerédi partitions

One of the most important tools in understanding large dense graphs is the
Regularity Lemma of Szemerédi [1975, 1978] and its extensions. This lemma has
many interesting connections to other areas of mathematics, including analysis and
information theory (see Lovász and Szegedy [2007], Bollobás and Nikiforov [2008],
Tao [2006a]). It also has weaker (but more effective) and stronger versions. Here
we survey as much as we need from this rich theory, extend it to graphons (as it
happens quite often, this leads to simpler, more elegant formulations), and prove a
very general version of it using the space of graphons.

9.1. Regularity Lemma for graphs

9.1.1. Homogeneous bipartite graphs and the original lemma. For a
graph G = (V, E) and for X, Y ⊆ V , let eG (X, Y ) denote the number of edges
with one endnode in X and another in Y ; edges with both endnodes in X ∩ Y are
counted twice. We denote by
eG (X, Y )
dG (X, Y ) =
|X||Y |
the density of edges between X and Y . If X and Y are disjoint, we denote by
G[X, Y ] the bipartite graph on X ∪ Y obtained by keeping just those edges of G
that connect X and Y .
Let P = {V1 , . . . , Vk } be a partition of V . We define the weighted graph GP on
V by taking the complete graph and weighting its edge uv by dG (Vi , Vj ) if u ∈ Vi
and v ∈ Vj . A related, but different construction is that of the template graph of
the partition P. This weighted quotient graph G/P is defined on [k]: node i gets
nodeweight |Si |/|V |, and the edge ij gets edgeweight eG (Si , Sj )/(|Si ||Sj |) (we allow
loops here).
The Regularity Lemma says, roughly speaking, that the node set of every graph
has an equitable partition P into a “small” number of classes such that GP is “close”
to G. Various (non-equivalent) forms of this lemma can be proved, depending on
what we mean by “close”.
Let G be a bipartite graph G with bipartition {U, W }. On the average, we
expect that for X ⊆ U and Y ⊆ W ,
eG (X, Y ) ≈ dG (U, V )|X||Y |.
For two arbitrary subsets of the nodes, eG (X, Y ) may be very far from this “ex-
pected value”, but if G is a random graph, or at least “random-like”, then it will
be close; random graphs are very “homogeneous” in this respect. We say that G is
ε-homogeneous, if
 
(9.1) eG (X, Y ) − dG (U, V )|X||Y | ≤ ε|U ||W |

141
142 9. SZEMERÉDI PARTITIONS

holds for all subsets X ⊆ U and Y ⊆ W .

Remark 9.1. Here we diverge from the usual statement of the Regularity Lemma:
one usually considers ε-regular bipartite graphs, where the stronger condition
 
(9.2) eG (X, Y ) − dG (U, V )|X||Y | ≤ ε|X||Y |
is required for all Y ⊆ U and Y ⊆ W such that |X| > ε|U | and |Y | > ε|W |. (Clearly
we could not require condition (9.2) to hold for small X and Y : for example, if
both have one element, then the quotient eG (X, Y )/(|X||Y |) is either 0 or 1.) The
properties of ε-homogeneity and ε-regularity are essentially equivalent (see Exercise
9.6). In either version, this property can be viewed as a quantitative version of
quasirandomness discussed in the Introduction.
With these definitions, the Regularity Lemma can be stated as follows:
Lemma 9.2 (Regularity Lemma, almost original form). For every ε > 0
there is an S(ε) ∈ N such that every graph G = (V, E) has an equitable partition
{V1 , . . . , Vk } (1/ε ≤ k ≤ S(ε)) such that for all but εk 2 pairs of indices 1 ≤ i < j ≤
k, the bipartite graph G[Vi , Vj ] is ε-homogeneous.
The important point is that the bound S(ε) on the number of classes is inde-
pendent of the graph G. Note that the Regularity Lemma does not say anything
about the internal structure of the classes Vi . The lower bound k ≥ 1/ε guarantees
that the number of edges inside the classes is bounded by k(n/k)2 ≤ εn2 . The
exceptional pairs of classes contain at most εk 2 (n/k)2 = εn2 edges, so all these
edges can be considered as “error terms”. If we need information about the inter-
nal structure of the classes, we have to appeal to the Strong Regularity Lemma to
be discussed below.
One feature of the Regularity Lemma, which unfortunately forbids practical
applications, is that the upper bound S(ε) it...provides on the number of classes is
2
very large: standard proofs give a tower 22 of height about 1/ε2 , and unfortu-
nately this is not far from the truth, as was shown by Gowers [2006] (for a simpler
recent construction, see Conlon and Fox [2011]).
9.1.2. Weak Regularity Lemma. A version of the Regularity Lemma with
a weaker conclusion but with a more reasonable error bound was proved by Frieze
and Kannan [1999]. This is the form that we use most of the time in this book.
Lemma 9.3 (Weak Regularity Lemma). For every k ≥ 1 and every graph
G = (V, E), V has a partition P into k classes such that
2
d (G, GP ) ≤ √ .
log k
Note that we do not require here that P be an equitable partition; it is not
hard to see that this version implies that there is also an equitable
√ partition with
similar property, just we have to increase the error bound to 4/ log k (see Exercise
9.7).
To see the connection with the original lemma, we note that if G0 is an ε-
homogeneous bipartite graph, and H is the weighted complete bipartite graph with
the same bipartition {U, W } and with edge weights d = e(G0 )/(|U ||W |), then (9.1)
says that d (G0 , H) ≤ ε. Hence if P is a Szemerédi partition in the sense of Lemma
9.2, then the distance between the bipartite subgraph of G induced by Vi and Vj ,
 9.1. REGULARITY LEMMA FOR GRAPHS 143

and the corresponding weighted bipartite subgraph of GP , is at most ε for all but
εk 2 pairs (i, j), and at most 1 for the remaining εk 2 pairs. This implies that the
cut distance between G and GP is at most 2ε. So the partition in Lemma 9.3 has
indeed weaker properties than the partition in Lemma 9.2. This is compensated
for by the relatively decent number of partition classes.
The Weak Regularity Lemma implies that there is a partition P such that the
template graph satisfies
2
(9.3) δ (G, G/P) ≤ d (G, GP ) ≤ √ .
log k
9.1.3. Strong Regularity Lemma. Other versions of the Regularity Lemma
strengthen, rather than weaken, the conclusion (of course, at the cost of replacing
the tower function by an even more formidable value). Such a “super-strong”
Regularity Lemma was proved by Alon, Fischer, Krivelevich and Szegedy [2000].
To state this lemma, we need a further definition. Let P be an equitable partition
of V (G), and let Q be an equitable refinement of it. Following Conlon and Fox
[2011], we say that Q is ε-close to P, if for almost every pair S ̸= T ∈ P (with at
most ε|P|2 exceptions), for almost every pair X, Y ∈ Q (with at most (|Q|/|P|)2
exceptions), we have  
 eG (X, Y ) eG (S, T ) 
 −  ≤ ε.
 |X||Y | |S||T | 
Lemma 9.4 (Very Strong Regularity Lemma). For every sequence ε =
(ε0 , ε1 , ...) of positive numbers there is a positive integer S(ε) such that for ev-
ery graph G = (V, E), the node set V has an equitable partition P and an equitable
refinement Q of P such that |Q| ≤ S(ε), P is ε0 -regular, Q is ε|P| -regular, and Q
is ε0 -close to P.
While this Very Strong Regularity Lemma has many important applications,
it is not easy to explain its significance at this point. One important feature is
that through the second partition Q, it carries information about the inside of the
partition classes of P.
A somewhat weaker (but essentially equivalent) version, which is simpler to
state but more difficult to apply, was proved by Tao [2006b] and by Lovász and
Szegedy [2007].
Lemma 9.5 (Strong Regularity Lemma). For every sequence ε = (ε0 , ε1 , ...)
of positive numbers there is a positive integer S(ε) such that for every graph G =
(V, E), there is a graph G′ on V , and V has a partition P into k ≤ S(ε) classes
such that
(9.4) d1 (G, G′ ) ≤ ε0 and d (G′ , (G′ )P ) ≤ εk .
Note that the first inequality involves the normalized edit distance, and so it is
stronger than a similar condition with the cut distance would be. The second error
bound εk in (9.4) can be thought of as being very small. If we choose εk = ε/2
for all k, we get the Weak Regularity Lemma 9.3 (without an explicit bound on
the number of classes). Choosing εk = ε20 /k 2 , the partition obtained satisfies the
requirements of the Original Regularity Lemma 9.2.
We can replace εk by the much smaller number εk /(k 2 S(εk )2 ), where S is
the bound in the Original Regularity Lemma. Then we can apply the Original
Regularity Lemma to each of the partition classes obtained in Lemma 9.5, to get
144 9. SZEMERÉDI PARTITIONS

the Very Strong Regularity Lemma 9.4. (The details of this derivation are left to
the reader as an exercise.)
We will formulate the Strong Regularity Lemma for kernels, and prove it in
that version, in Section 9.3.
Exercise 9.6. Show that (a) if a bipartite graph is ε-regular, then it is ε-
homogeneous; (b) if a bipartite graph is ε3 -homogeneous, then it is ε-regular.
Exercise 9.7. Prove that for every k ≥ 1 and every graph G √ = (V, E), V has an
equitable partition P into k classes such that d (G, GP ) ≤ 4/ log k.

9.2. Regularity Lemma for kernels

The Weak Regularity Lemma extends to kernels, and this is the form we are
going to prove first. While stating and proving the Lemma directly would be quite
easy, we make a detour by introducing the “stepping operator” formally and stating
some basic properties. These will be useful later on.

9.2.1. The stepping operator. Let W ∈ W and let P = (S1 , . . . , Sq ) be a

partition of [0, 1] into a finite number of measurable sets. (When we speak of a
partition of [0, 1], we always mean such a partition.) We define the function WP by
∫
1
WP (x, y) = W (x, y) dx dy (x ∈ Si , y ∈ Sj ).
λ(Si )λ(Sj )
Si ×Sj

So WP is obtained by averaging W over the “steps” Si × Sj ; it is a stepfunction

with steps in P. If λ(Si ) = 0 or λ(Sj ) = 0, then we define WP (x, y) = 0 (this is
just to have a complete definition; sets of measure zero in the partition can usually
be ignored). We call this construction a stepping of W ; it will be used throughout
this book.
Analytically, the stepping operator is the orthogonal projection of the Hilbert
space L2 ([0, 1]2 ) onto the subspace of stepfunctions with P-steps. In probability
language, it is the conditional expectation relative to the (finite) sigma-algebra
generated by the sets in P. These remarks may make some of the basic properties
below easier to understand, but we will use the more elementary direct formulation.
Most of the information about the stepfunction WP is contained in a finite
weighted graph, the quotient graph W/P. This is a weighted graph on [q], with
nodeweights αi (W/P) = λ(Si ) and edgeweights βij (W/P) = WP (x, y) for any
x ∈ Si , y ∈ Sj .
On the space W, the stepping operator W 7→ WP is a linear operator which is
idempotent and symmetric:
(9.5) ⟨UP , WP ⟩ = ⟨UP , W ⟩ = ⟨U, WP ⟩.
Stepfunctions with steps in a fixed partition P form a finite dimensional linear
space, and the stepping operator is the orthogonal projection onto this space, which
is shown by the simple identity
(9.6) ⟨WP , W − WP ⟩ = 0.
This implies that stepping operator is contractive with respect to the L2 norm:
(9.7) ∥WP ∥22 = ∥W ∥22 − ∥W − WP ∥22 ≤ ∥W ∥22 .
 9.2. REGULARITY LEMMA FOR KERNELS 145

It is not hard to see that the stepping operator is also contractive with respect to
the cut norm (Exercise 9.17). In fact, we will see in Section 14.2.1 that stepping is
contractive with respect to any other reasonable norm on W.
9.2.2. Weak Regularity Lemma. It is a basic fact from analysis that every
kernel W can be approximated arbitrarily well by stepfunctions in the L1 norm.
The approximating stepfunctions can be obtained by averaging over “steps”:
Proposition 9.8. Let (Pn ) be a sequence of measurable partitions of [0, 1] such
that every pair of points is separated by all but a finite number of partitions Pn .
Then WPn → W almost everywhere for every W ∈ W. 
The Weak Regularity Lemma for kernels, proved by Frieze and Kannan [1999]
(and in particular its Corollary 9.13 below), is a related statement about approxima-
tion by stepfunctions in the cut norm (instead of in the sense of almost everywhere
convergence).
Lemma 9.9 (Weak Regularity Lemma for Kernels). For every function W ∈
W and k ≥ 1 there is stepfunction U with k steps such that
2
∥W − U ∥ < √ ∥W ∥2 .
log k
Roughly speaking, this Lemma says that every kernel can be approximated well
in the cut norm by stepfunctions (in fact, by its steppings). Proposition 9.8 asserts
something similar about approximating in the L1 -norm. Since ∥W ∥ ≤ ∥W ∥1 ,
approximating in the L1 norm seems to be a stronger result. However, the error in
the L1 -norm approximation depends not only on the number of steps, but on W
as well. The crucial fact about Lemma 9.9 is that the error tends to 0 as k → ∞,
uniformly in W .
The error bound in Lemma 9.9 is only attractive when compared with the error
bound in the stronger versions; for a prescribed error ε, the number of partition
classes we need is still exponential in 1/ε2 . Frieze and Kannan give a stronger
form of this result that provides a polynomial size description of the approximating
stepfunction.
Lemma 9.10. For every kernel U ∈ W1 and k ≥ 1 there are k pairs of subsets
Si , Ti ⊆ [0, 1] and k real numbers ai such that
∑
k
1
∥U − ai 1Si ×Ti ∥ < √ .
i=1
k
∑
It is clear that the function 1Si ×Ti is a stepfunction; we can make it
i ai∑
symmetric by taking the average with i ai 1Ti ×Si , getting 2k terms. This sym-
metric stepfunction has at most 22k steps, so Lemma 9.9 follows from Lemma 9.10
(replacing k by 22k ).
We have mentioned the significance of the interplay between the cut norm and
other kernel norms. The proof of the Regularity Lemma is the first point where
this is apparent. For later reference, we state the key observation in the proof of
the Weak Regularity Lemma separately, in two versions.
Lemma 9.11. (a) For every U ∈ W there are two sets S, T ⊆ [0, 1] and a real
number 0 ≤ a ≤ ∥U ∥∞ such that
∥U − a1S×T ∥22 ≤ ∥U ∥22 − ∥U ∥2 .
146 9. SZEMERÉDI PARTITIONS

(b) Let U ∈ W and let P be a measurable k-partition of [0, 1]. Then there is a
partition Q refining P with at most 4k classes such that
∥U − UP ∥ = ∥UQ − UP ∥ .
Proof. Let S and T be measurable subsets of [0, 1] such that
∫ 
 
∥U ∥ =  U  = |⟨U, 1S×T ⟩|,
S×T

where we may assume that ⟨U, 1S×T ⟩ ≥ 0. Let a = λ(S)λ(T ) ∥U ∥ .

1
Then
1
(9.8) ∥U − a1S×T ∥22 = ∥U ∥22 − ∥U ∥2 ≤ ∥U ∥22 − ∥U ∥2 .
λ(S)λ(T )
This proves (a).
The proof of (b) is similar. The inequality ∥U − UP ∥ ≥ ∥UQ − UP ∥ follows
by the contractivity of the stepping operator (Exercise 9.17). To prove the other
direction, let S and T be measurable subsets of [0, 1] such that ⟨U − UP , 1S×T ⟩ =
∥U − UP ∥ , and let Q denote the partition generated by P, S and T . Clearly Q
has at most 4k classes. Using (9.5), we get ⟨U, 1S×T ⟩ = ⟨UQ , 1S×T ⟩, and hence
   
∥U − UP ∥ = ⟨U − UP , 1S×T ⟩ = ⟨UQ − UP , 1S×T ⟩ ≤ ∥UQ − UP ∥ .
This completes the proof. 

Proof of Lemma 9.10. We apply Lemma 9.11(a) repeatedly, to get pairs of

∑j
sets Si , Ti and real numbers ai such that the “remainders” Wj = U − i=1 ai 1Si ×Ti
satisfy
∑
j−1
∥Wj ∥22 ≤ ∥U ∥22 − ∥Wi ∥2 .
i=0
Since the right hand side remains nonnegative, it follows that for every k there is a
0 ≤ i < k with ∥Wi ∥2 ≤ 1/k. Changing ai+1 , . . . , ak to 0, we get the lemma. 

The stepfunction approximating a given graphon W in Lemma 9.9 is usually not

a stepping of W . Is the optimally approximating stepfunction necessarily a stepping
of W ? While this looks plausible, the answer is negative (see Exercise 9.18). But,
as noted by Frieze and Kannan [1999], such steppings are almost optimal:
Lemma 9.12. Let W ∈ W1 , let U be a stepfunction, and let P denote the partition
of [0, 1] into the steps of U . Then
∥W − WP ∥ ≤ 2∥W − U ∥ .
Proof. Using that U = UP and the contractivity of the stepping operator with
respect to the cut norm, we get
∥W − WP ∥ ≤ ∥W − U ∥ + ∥U − WP ∥ = ∥W − U ∥ + ∥UP − WP ∥
= ∥W − U ∥ + ∥(U − W )P )∥ ≤ 2∥W − U ∥ .


Lemmas 9.9 and 9.12 imply:

 9.2. REGULARITY LEMMA FOR KERNELS 147

Corollary 9.13. For every function W ∈ W1 and k ≥ 1 there is a partition P of

[0, 1] into at most k sets with positive measure for which
2
∥W − WP ∥ ≤ √ . 
log k
A partition P of [0, 1] such that ∥W −WP ∥ ≤ ε will be called a weak regularity
partition of w with error ε.
Lemma 9.11(b) provides an alternative way of getting this corollary. It is easy
to check that ⟨UQ − UP , UP ⟩ = ⟨U − UP , UP ⟩ = 0 if Q is a refinement of P. Hence
∥U − UP ∥2 = ∥UQ − UP ∥2 ≤ ∥UQ − UP ∥22 = ∥UQ ∥22 − ∥UP ∥22
= ∥U − UP ∥22 − ∥U − UQ ∥22 .
So UQ is a better approximation of U in L2 than UP , and the gain is at least as
large as ∥U − UP ∥2 . From here we can conclude just as in the proof above.
Coming back to the approximation described in Lemma 9.10, it is often useful
to have a bound on the numbers ai . With the notation of its proof, looking at the
proof carefully, we see that
1
ai = ∥Wi ∥ ,
λ(Si )λ(Ti )
and
1
∥Wi ∥2 = ∥Wi ∥22 − ∥Wi+1 ∥22 ,
λ(Si )λ(Ti )
whence
∑ k ∑k
λ(Si )λ(Ti )a2i = (∥Wi ∥22 − ∥Wi+1 ∥22 ) = ∥U ∥22 − ∥Wk ∥22 ≤ ∥U ∥22 .
i=1 i=1
This bound allows ai to be large when λ(Si )λ(Ti ) is small, but it is easy to fix the
argument to get a more useful bound. Instead of choosing the optimal sets Si and
Ti , we choose a pair Si , Ti such that λ(Si ), λ(Ti ) ≥ 1/2, and
1
⟨Wi , 1Si ×Ti ⟩ ≥ ∥Wi ∥ .
4
It is easy to see that such a pair exists (cf. Exercise 8.4). Then we get the following:
Lemma 9.14. For every kernel W ∈ W1 and∑k ≥ 1 there are k pairs of subsets
Si , Ti ⊆ [0, 1] and k real numbers ai such that i a2i ≤ 4 and
∑
k
4
∥W − ai 1Si ×Ti ∥ < √ .
i=1
k
We can easily add other requirements in Lemma 9.9.
Lemma 9.15. Let W ∈ W1 and 1 ≤ m < k.
(a) For every m-partition Q of [0, 1] there is k-partition P refining Q such that
2
∥W − WP ∥ ≤ √ .
log k/m
(b) For every m-partition Q of [0, 1] there is an equipartition P with k classes
such that
2m
∥W − WP ∥ ≤ 2∥W − WQ ∥ + .
k
148 9. SZEMERÉDI PARTITIONS

Proof. Statement (a) follows by the same argument as Lemma 9.9, just start-
ing with Q instead of the indiscrete partition. To prove (b), we partition each class
of Q into classes of measure 1/k, with at most one exceptional class of size less then
1/k. Keeping all classes of size 1/k, let us take the union of exceptional classes,
and repartition it into classes of size 1/k, to get a partition P.
To analyze this construction, let us also consider the common refinement R =
P ∧ Q. Then WR and WP differ on a set of measure less than 2(m/k), and so
2m
∥W − WP ∥ ≤ ∥W − WR ∥ + .
k
Lemma 9.12 implies that ∥W − WR ∥ ≤ 2∥W − WQ ∥ , which completes the
proof. 

9.2.3. Strong Regularity Lemma. The Strong Regularity Lemma too has
a “continuous” version:

Lemma 9.16 (Strong Regularity Lemma for Kernels). For every sequence
ε = (ε0 , ε1 , ...) of positive numbers there is a positive integer S(ε) such that for
every graphon W , there is another graphon W ′ , and a stepfunction U ∈ W0 with
k ≤ S(ε) steps such that

(9.9) ∥W − W ′ ∥1 ≤ ε0 and ∥W ′ − U ∥ ≤ εk .

We will give a proof of this Lemma, deriving it from an even more general
theorem, in the next section. Here we sketch how to derive the graph version 9.5
from the kernel version. Let (ε0 , ε1 , ...) be a sequence of positive numbers, which
we may assume is monotone decreasing. Let G be a simple graph on [n]. We apply
Lemma 9.16 with εk /2 to WG , to get a threshold S ′ (depending only on (ε0 , ε1 , . . . ),
a kernel W ′ and a partition P of [0, 1] such that |P| ≤ S ′ , ∥WG − W ′ ∥1 ≤ ε0 /2 and
∥W ′ − WP′ ∥ ≤ εk /2.
′ ′
First,(we have to turn
] W into a graph G . This can be done by randomization.
Let∫ Ii = (i − 1)/n, i/n and Rij = Ii × Ij . We connect i and j with probability
n2 Rij W ′ . The probability that this edge will be in the symmetric difference of
∫
E(G) and E(G′ ) is at most n2 |WG − W ′ |, and hence the expected (normalized)
Rij
edit distance between G and G′ is at most ∥WG −W ′ ∥1 ≤ ε0 /2. Markov’s inequality
gives that with probability at least 1/2, the distance d1 (G, G′ ) ≤ ε0 .
Next, we have to turn the partition P of [0, 1] into a partition Q of [n]. We do
this randomly again, by selecting a uniform random point Xi ∈ Ii (i = 1, . . . , n),
and putting i into the m-th class of Q if Xi is in the m-th class of P. A bit
trickier computation with second moments (which is similar to the proof of Propo-
sition 12.19, but simpler, and √ is not given here) shows that with high probability,
d (G′ , (G′ )Q ) ≤ εk /2 + 10/ n.
Now we choose k0 = max(k0′ , 400/ε2k′ ). If n ≤ k0 , then we can take G = G′
0
and partition [n] into singletons. If n > k0 , then with positive probability the
′ ′
partition Q constructed above √ satisfies |Q| = k ≤ k0 ≤ k0 , d1 (G, G ) ≤ ε0 and
d (G′ , (G′ )Q ) ≤ εk /2 + 10/ n ≤ εk .

Exercise 9.17. Prove that the stepping operator is contractive with respect to
the L1 norm and the cut norm.
 9.3. COMPACTNESS OF THE GRAPHON SPACE 149

Exercise 9.18. Show by an example that the best approximation in the cut
norm of a function W ∈ W1 by a stepfunction with a given number of steps is not
necessarily a stepping of W . Is stepping the best approximation in the L2 or in
the L1 norm?
Exercise 9.19. Are analogues of Lemma 9.12 valid for the L1 L2 and L∞ norms?
Exercise 9.20. Formulate and prove the original Regularity Lemma for kernels.
Exercise 9.21. Give a proof of the Strong Regularity Lemma 9.16 along the lines
of the proof of Lemma 9.9.
Exercise 9.22. Let K1 , K2 , . . . be arbitrary nonempty subsets of a Hilbert space
H. Prove for every ε > 0 and f ∈ H there is an integer 1 ≤ m ≤ ⌈1/ε2 ⌉ and a
vector f0 = α1 f1 + · · · + αm fm (αi ∈ R, fi ∈ Ki such that for every g ∈ Km+1
we have |⟨g, f − f0 ⟩| ≤ ε∥g∥∥f ∥. Derive the weak, original, and strong lemmas by
choosing the sets Ki appropriately.

9.3. Compactness of the graphon space

In this section we prove a theorem of Lovász and Szegedy [2007] that is equiv-
alent (at least in a non-effective sense) to all versions of the Regularity Lemma.
To be more precise, we will derive the theorem from the Weak Regularity Lemma,
and then we will show that the Strong Regularity Lemma can be derived from it
quite easily. In a sense, this theorem can be considered as the strongest form of
regularity.
f0 , δ ) is compact.
Theorem 9.23. The space (W
Proof. In a metric space, it suffices to prove that every sequence W1 , W2 , . . .
of graphons has a convergent subsequence.
For every n ≥ 1, we can construct the partitions Pn,k of [0, 1] (k = 1, 2, . . . ),
using Lemma 9.15, such that these partitions and the corresponding stepfunctions
Wn,k = (Wn )Pn,k ∈ W0 satisfy the following conditions:
(i) ∥Wn − Wn,k ∥ ≤ 1/k,
(ii) The partition Pn,k+1 refines Pn,k ,
(iii) |Pn,k | = mk depends only on k.
Once we have such partitions, we can rearrange the points of [0, 1] for every
fixed n by a measure preserving bijection so that every partition class in every Pn,k
is an interval.
Claim 9.24. We can replace the sequence (Wn ) by a subsequence so that for every
k, the sequence Wn,k converges almost everywhere to a stepfunction Uk with mk
steps as n → ∞.
Indeed, we can select a subsequence of the Wn for which the length of the i-th
interval of Wn,1 converges for every i, and also the value of Wn,1 on the product of
the i-th and j-th intervals converges for every i and j (as n → ∞). It follows then
that the sequence Wn,1 converges to a limit U1 almost everywhere, which itself is a
stepfunction with m1 steps that are intervals.
We repeat this for k = 2, 3, . . . , to get subsequences for which Wk,n → Uk
almost everywhere, where Uk is a stepfunction with mk steps that are intervals.
As usual, we always keep the k-th function after the k-th step. This yields the
subsequence with the properties in the Claim.
150 9. SZEMERÉDI PARTITIONS

Let Pk denote the partition of [0, 1] into the steps of Uk . For every k < l, the
partition Pn,l is a refinement of the partition Pn,k , and hence Wn,k = (Wn,l )Pn,k .
It is easy to see that this kind of relation is inherited by the limiting stepfunctions:
(9.10) Uk = (Ul )Pk .
Let (X, Y ) be a random point in [0, 1]2 chosen uniformly, then (9.10) implies
that the sequence (U1 (X, Y ), U2 (X, Y ), . . . ) is a martingale. Since the random
variables Ui (X, Y ) remain bounded, the Martingale Convergence Theorem A.12
implies that this sequence is convergent with probability 1. In other words, the
sequence of functions (U1 , U2 , . . . ) is convergent almost everywhere. Let U be its
limit; we show that ∥U − Wn ∥ → 0.
Fix any ε > 0. Then there is a k > 3/ε such that ∥U − Uk ∥1 < ε/3. Fixing this
k, there is an n0 such that ∥Uk − Wn,k ∥1 < ε/3 for all n ≥ n0 . Then
δ (U, Wn ) ≤ δ (U, Uk ) + δ (Uk , Wn,k ) + δ (Wn,k , Wn )
ε ε ε
≤ ∥U − Uk ∥1 + ∥Uk − Wn,k ∥1 + δ (Wn,k , Wn ) ≤ + + = ε.
3 3 3
This completes the proof of Theorem 9.23. 

The theorem remains valid if we replace W f0 by any uniformly bounded subset

f
of W, closed in the δ distance. For example, the space (W f1 , δ ) is also compact.
This can be proved by the same argument, or by noticing that W 7→ 2W − 1 is a
mapping (Wf0 , δ ) → (W
f1 , δ ) that is continuous and surjective, and so it preserves
compactness.
An easy consequence of Theorem 9.23 is the following:
Corollary 9.25. For every ε > 0 there is an integer k(ε) ≥ 1 such that simple
graphs with kε nodes form an ε-net in (W0 , δ ).
We conclude this section with showing that Theorem 9.23 implies the Strong
Regularity Lemma quite easily.

Proof of Lemma 9.16. Every graphon W is the limit of stepfunctions in the

∥.∥1 norm, hence there is a stepfunction U ∈ W0 with ∥W − U ∥1 ≤ ε0 . This
means that the sets B1 (U, ε0 ), where U is a stepfunction, cover the whole space
W0 . Unfortunately, these sets are not open in the d metric. Therefore, we take
a little larger sets. Let k(U ) denote the number of steps of a stepfunction U , and
define
AU = {W ∈ W0 : (∃V ∈ W0 ) ∥U − V ∥ < εk(U ) , ∥V − W ∥1 < ε0 }.
Claim 9.26. The set AU is open in the cut norm.
Indeed, let W ∈ AU and Wn ∈ W0 such that ∥W − Wn ∥ → 0. By the
definition of AU , there is a graphon V such that ∥U −V ∥ < εk(U ) and ∥V −W ∥1 <
ε0 . By Proposition 8.25, there are graphons Vn such that ∥V − Vn ∥ → 0 and
∥Wn − Vn ∥1 → ∥V − W ∥1 < ε0 . So if n is large enough, we have ∥U − Vn ∥ < εk(U )
and ∥Vn − Wn ∥1 < ε0 , showing that Wn ∈ AU .
Now we have to go to the factor space W f0 . For every stepfunction U , we
consider the sets
eU = {W ∈ W
A f0 : (∃V ∈ W
f0 ) δ (U, V ) < εk(U ) , δ1 (V, W ) < ε0 }.
 9.4. FRACTIONAL AND INTEGRAL OVERLAYS 151

It is easy to see, using Claim 9.26, that A eU is open in (W f0 , δ ). The sets A

eU cover
the whole space, so by the compactness of the space (Theorem 9.23), we obtain a
eU = W
finite set of stepfunctions U1 , . . . , Ut such that ∪ti=1 A f0 .
i
We claim that we can choose S(ε) = maxi≤t k(Ui ) to satisfy the requirements of
the lemma. Indeed, for every graphon W there is a stepfunction Ui (1 ≤ i ≤ t) such
that W ∈ A eU , which means that there is a graphon V such that δ (Ui , V ) < εk(U )
i
and δ1 (V, W ) < ε0 . We can apply a measure preserving bijections φ, ψ to get
d1 (V φ , W ) < ε0 and δ (Uiψ◦φ , V φ ) < εk(Ui ) . Since Uiψ◦φ is a stepfunction with
k(Ui ) steps, we can take U = Uiψ◦φ and W ′ = V φ to complete the proof. 
Exercise 9.27. Prove that for every ε > 0 there is a positive integer S(ε) such
that for every graphon W there is another graphon W ′ and a stepfunction U ∈ W0
with k ≤ S(ε) steps such that ∥W − W ′ ∥ ≤ ε/k! and ∥W ′ − U ∥1 ≤ ε.

9.4. Fractional and integral overlays

Using the Regularity Lemma, we are now ready to discuss the problem of
comparing the two distances δ and δb , as raised in Section 8.1.4. If two graphs
G1 and G2 have the same number of nodes, then the inequality
δ (G1 , G2 ) ≤ δb (G1 , G2 ),
is easy, but what can we say in the other direction? (This is admittedly a very
technical issue, but it comes up in all sorts of arguments.) Perhaps the following
very close connection is true:
Conjecture 9.28. For any two simple graphs G and G′ on n nodes, δb (G, G′ ) ≤
2δ (G, G′ ).
Unfortunately, I can only offer some weaker bounds (but these will be sufficient
for the applications later). We will need these results for weighted graphs too. The
following theorem is a combination of results of Borgs, Chayes, Lovász, Sós and
Vesztergombi [2008] and Alon [unpublished].
Theorem 9.29. For any two edge-weighted graphs H1 and H2 with the same num-
ber n of nodes, with edgeweights in [0, 1], we have the following inequalities:
(9.11) δb (H1 , H2 ) ≤ n6 δ (H1 , H2 ),
17
(9.12) δb (H1 , H2 ) ≤ δ (H1 , H2 ) + √
log n
45
(9.13) δb (H1 , H2 ) ≤ √ .
− log δ (H1 , H2 )
Proof. The first inequality is quite easy. Let (Xui ) be an optimal fractional
overlay of H1 and H2 . We claim that there is a bijection π : V1 → V2 such that
Xu,π(u) ≥ 1/n3 for all u ∈ V (H1 ). This follows from the Marriage Theorem: if
there is no such bijection, then there are two sets S ⊆ V1 and T ⊆ V2 such that
|S| + |T | > n and Xst < 1/n3 for all s ∈ S and t ∈ T . Then X(S, T ) ≤ |S||T |/n3 <
1/n. On the other hand,
|S| |V2 \ T | |S| + |T | − n 1
X(S, T ) = X(S, V2 ) − X(S, V2 \ T ) ≥ − = ≥ ,
n n n n
a contradiction.
152 9. SZEMERÉDI PARTITIONS

Let Y be the fractional overlay corresponding to the bijection π. Then Y ≤

n3 X, and hence
δb (H1 , H2 ) ≤ d (H1 , H2 , Y ) ≤ n6 d (H1 , H2 , X) = n6 δ (H1 , H2 ).
This proves (9.11).
To prove the second inequality (9.12), let k = ⌊n1/3 ⌋. By the Weak Regularity
Lemma 9.3, there are partitions P = {V1 , . . . , Vk } of V (H1 ) and Q = {U1 , . . . , Uk }
of V (H2 ) into k almost equal classes so that
4
d (H1 , (H1 )P ), d (H2 , (H2 )Q ) ≤ √ .
log k
For the weighted k-node “template” graphs H1 /P and H2 /Q we have
δ (H1 /P, H2 /Q) = δ ((H1 )P , (H2 )Q )
≤ δ (H1 , H2 ) + δ (H1 , (H1 )P ) + δ (H2 , (H2 )Q )
8
≤ δ (H1 , H2 ) + √ .
log k
Let (Xij )ki,j=1 be an optimal fractional overlay of H1 /P and H2 /Q. We define
a bijection φ : V (H1 ) → V (H2 ) by mapping ⌊Xij n⌋ nodes of Vi to Uj arbitrarily.
This is possible, since
∑
k ∑
k ∑
k ∑
k
⌊Xij n⌋ ≤ Xij n = |Vi | and ⌊Xij n⌋ ≤ Xij n = |Uj |.
j=1 j=1 i=1 i=1

The nodes left in the two graphs are matched with each other arbitrarily.
The bijection φ between V (H1 ) = V ((H1 )P ) and V (H2 ) = V ((H2 )Q ) defines a
fractional overlay Y between H1 /P and H2 /Q, such that
d (φ((H1 )P ), (H2 )Q ) = d (H1 /P, H2 /Q, Y ).
The fractional overlays X and Y are very close: |Xij − Yij | ≤ 1/n for every 1 ≤
i, j ≤ k. Hence it follows that
k2 k2
d (H1 /P, H2 /Q, Y ) ≤ d (H1 /P, H2 /Q, X) + = δ (H1 /P, H2 /Q) + .
n n
Combining, we get
b 1 , H2 ) ≤ d (φ(H1 ), H2 ) ≤ d (φ((H1 )P ), (H2 )Q ) + √ 8
δ(H
log k
8 k2 8
= d (H1 /P, H2 /Q, Y ) + √ ≤ δ (H1 /P, H2 /Q) + +√
log k n log k
k2 16
≤ δ (H1 , H2 ) + +√ .
n log k
Recalling the choice of k, we get (9.12).
Finally, the third inequality (9.13) follows easily from the first two. If n <
δ (H1 , H2 )−1/7 , then (9.11) implies that

b 1 , H2 ) ≤ δ (H1 , H2 )1/7 < √ 55

δ(H ,
− log δ (H1 , H2 )
 9.4. FRACTIONAL AND INTEGRAL OVERLAYS 153

while for n ≥ δ (H1 , H2 )−1/7 , (9.12) gives that

17 45
δb (H1 , H2 ) ≤ δ (H1 , H2 ) + √ <√ . 
log(δ (H1 , H2 )−1/7 ) − log δ (H1 , H2 )

The first and third inequalities in the theorem are very weak (at least, in
comparison with the conjectured bound of δb ≤ 2δ ). Nevertheless, (9.13) will be
important for us, since it implies that δ and δb define the same Cauchy sequences
of graphs. Borgs, Chayes, Lovász, Sós and Vesztergombi [2008] prove a stronger
inequality of this nature:
(9.14) δb (H1 , H2 ) ≤ 32δ (H1 , H2 )1/67 .
Since this is still far from Conjecture 9.28, we don’t reproduce the proof here.
We can ask the same question about any of the unlabeled distances: if two
graphs have the same number of nodes, is the distance between them defined
through optimal overlay essentially the same as the distance defined by going to
the associated graphons and considering their distance? For the edit distance, an
affirmative answer was proved by Pikhurko [2010]; this bound is much stronger
than 9.14: it is optimal except for the constant 3. We prove it in a bit more general
form, for weighted graphs, since we will need it.
Theorem 9.30. For any two edge-weighted graphs H1 and H2 on [n] we have the
following inequalities:
δ1 (H1 , H2 ) ≤ δb1 (H1 , H2 ) ≤ 3δ1 (H1 , H2 ).
Proof. The first inequality is trivial. Let A and B be the adjacency matrices
of H1 and H2 , respectively. Then we have
δb1 (H1 , H2 ) = min ∥A − P BP ∥1 ,
P

where P ranges over all permutation matrices. To express the distance δ1 is a bit
more complicated:
δ1 (H1 , H2 ) = min d1 (H1 , H2 , X),
X
where ∑
d1 (H1 , H2 , X) = Xiu Xjv |Aij − Buv |,
i,j,u,v∈[n]

and X ranges over all n × n fractional overlays. We want to prove that

(9.15) δb1 (H1 , H2 ) ≤ 3d1 (H1 , H2 , X)
for every fractional overlay X. It suffices to prove this for overlays X of the form
1 ∑
m
X= Pk ,
mn
k=1

where the Pk are n × n permutation matrices, because matrices of this form are
dense among all fractional overlays by the Birkhoff–von Neumann Theorem. Then
1 ∑ ∑ 1 ∑
m n m
d1 (H1 , H2 , X) = |Aij − BPk (u),Pl (v) | = ∥A − Pk BPl ∥1 .
m 2 n2 i,j=1
m2
k,l=1 k,l=1
154 9. SZEMERÉDI PARTITIONS

We need the following simple inequality for three permutation matrices P , Q

and R:
(9.16) ∥A − P BP ∥1 ≤ ∥A − P BQ∥1 + ∥A − QBR∥1 + ∥A − RBP ∥1 .
Indeed, by the triangle inequality and the invariance of the ∥.∥1 -norm under per-
mutations of rows and columns, we have
∥A − P BP ∥1 ≤ ∥A − P BQ∥1 + ∥P R−1 A − P BQ∥1 + ∥P R−1 A − P BP ∥1
= ∥A − P BQ∥1 + ∥A − RBQ∥1 + ∥A − RBP ∥1 .
Transposing the middle term, we get (9.16).
Averaging this inequality with P, Q, R ranging over {P1 , . . . , Pm }, we get
1 ∑ 3 ∑
m m
∥A − Pk BPk ∥1 ≤ 2 ∥A − Pk BPl ∥1 = 3d1 (H1 , H2 , X).
m m
k=1 k,l=1
∑ b 1 , H2 ), this completes the proof.
Since (1/m) k ∥A − Pk BPk ∥1 ≥ δ(H 

Exercise 9.31. Show by an example that δ1 ̸= δb1 in general.

9.5. Uniqueness of regularity partitions

Is the regularity partition of a graph (weak, original or strong) uniquely deter-
mined? Perhaps uniquely determined up to some error?
The answer to this question is negative. The image of any kind of regularity
partition under any automorphism of a graph G is another regularity partition
with the same properties. Typically, we get this way many different partitions.
For example, consider the graph G on [n] in which every node i is connected to
i ± 1, i ± 2, . . . i ± ⌈n/4⌉ (modulo m). A regularity partition (in the original sense,
and also in the strong sense) is obtained by splitting [n] into k blocks of consecutive
numbers of approximately equal size, and rotations give many different partitions
of this kind.
Giving up the uniqueness of the partition itself, we can ask about the approx-
imate uniqueness of the template graph G/P . For partitions in the weak or in the
original sense, even this does not hold. Let (Fn ) be a quasirandom graph sequence
with v(Fn ) = n. Consider the direct product G = Kn ×Fn , and the two n-partitions
P1 and P2 induced by the projections onto the two factors. The bipartite graph
between any two classes of P1 is K2 × Fn , which is ε-homogeneous for an arbitrarily
small ε if n is large enough. So P1 is a regularity partition with template graph Kn ,
where every edge has weight 1/2. On the other hand, the subgraph of G between
two classes of P2 is either edgeless, or it is a complete bipartite graph with a perfect
matching removed. Both of these are ε-homogeneous if n is large enough, and the
template graph is Fn .
But for strong regularity partitions, Alon, Shapira and Stav [2009] did prove
a uniqueness result. We state it for graphons, and for our version of the Strong
Regularity lemma; for other versions, we refer to the paper.
Theorem 9.32. Let ε > 0, let W, W1 , W2 be graphons and let P1 , P2 be equipar-
titions of [0, 1] into k parts so that ∥W − Wi ∥1 ≤ ε and ∥Wi − (Wi )Pi ∥ ≤ ε/k 4 .
Then δ1 (W1 /P1 , W2 /P2 ) ≤ 8ε.
 9.5. UNIQUENESS OF REGULARITY PARTITIONS 155

Figure 9.1. Proving uniqueness of strong regularity partitions.

Heavy lines mean small cut distance, broken lines mean small L1 -
distance.

Proof. Let Q = P1 ∨ P2 denote the common refinement of P1 and P2 ; clearly

Q has at most k 2 classes, and ∥W1 −W2 ∥ ≤ 2ε. By the contractivity of the stepping
operator (Exercise 9.17 of Proposition 14.13), we have
∥(W1 )P1 − (W2 )P1 ∥1 ≤ 2ε, ∥(W1 )P2 − (W2 )P2 ∥1 ≤ 2ε, ∥(W1 )Q − (W2 )Q ∥1 ≤ 2ε,
and
∥(W1 )P2 − (W1 )Q ∥ ≤ ∥W1 − (W1 )P1 ∥ ≤ ε/k 4 ,
∥(W2 )P1 − (W1 )Q ∥ ≤ ∥W2 − (W2 )P2 ∥ ≤ ε/k 4
(see Figure 9.1 for the chain of small distances followed by the proof). Hence
δ1 (W1 /P1 , W2 /P2 ) = δ1 ((W1 )P1 , (W2 )P2 ) ≤ ∥(W1 )P1 − (W2 )P2 ∥1
≤ ∥(W1 )P1 − (W2 )P1 ∥1 + ∥(W2 )P1 − (W1 )Q ∥1
+ ∥(W1 )Q − (W2 )Q ∥1 + ∥(W2 )P1 − (W1 )Q ∥1
+ ∥(W1 )P2 − (W2 )P2 ∥1 .
By the trivial inequality (8.15),
∥(W2 )P1 − (W1 )Q ∥1 ≤ k 4 ∥(W2 )P1 − (W1 )Q ∥ ≤ ε,
and similarly
∥(W1 )P2 − (W2 )Q ∥1 ≤ ε.
Substituting these bounds, the theorem follows. 
The proof above bounds the fractional edit distance of the two template graphs
(W1 )/P1 and (W2 )/P2 . Using Pikhurko’s Theorem 9.30, we could replace it by the
integral version of the edit distance δb1 , at the cost of another factor of 3. Using
Exercise 8.18, we could replace the bound ∥Wi − (Wi )Pi ∥ ≤ ε/k 4 by ∥Wi −
(Wi )Pi ∥ ≤ ε/(2k) Also, we could derive similar bounds for the weighted graphs
W/P1 and W/P2 .
 CHAPTER 10

Sampling

We turn to the analysis of sampling from a graph, our basic method of gathering
information about very large dense graphs. In fact, most of the time we prove our
results in the framework of sampling from a graphon. We start with describing
what it means to sample from a graphon.

10.1. W -random graphs

A graphon W gives rise to a way of generating random graphs that are more
general than the Erdős–Rényi graphs. This construction was introduced inde-
pendently by Diaconis and Freedman [1981], Boguñá and Pastor-Satorras [2003],
Lovász and Szegedy [2006], and Bollobás, Janson and Riordan [2007], and quite
probably implicitly by others.
Given a graphon W and an ordered set S = (x1 , . . . , xn ), where xi ∈ [0, 1], we
define a weighted graph H(S, W ) on node set [n] by assigning weight W (xi , xj ) to
edge ij (i, j ∈ [n], i ̸= j). We give weight 0 to the loops.
Every weighted graph H with edgeweights βij (H) ∈ [0, 1] gives rise to a random
simple graph G(H) on V (H): we connect nodes i and j with probability βij (H),
making an independent decision for distinct pairs (i, j) (i, j ∈ ([n], i ̸= )j). In
particular, we can construct a random simple graph G(S, W ) = G H(S, W ) .
For an integer n > 0, we define the random weighted graph H(n, W ) = H(S, W ),
and the random simple graph G(n, W ) = G(S, W ), where S is an ordered n-tuple
of independent uniform random points from [0, 1].
To mention some special cases, if W is the identically p function, we get “or-
dinary” random graphs G(k, p). If W = WG for some simple graph G, then
G(k, WG ) = H(k, WG ) is “almost” the same as the random induced subgraph
G(k, G) of G. To be more precise, if we condition on x1 , . . . , xk belonging to dif-
ferent steps of WG , then G(k, WG ) is a random k-node induced subgraph. The set
this condition excludes,
( ) namely sequences x1 , . . . , xk containing repetitions, has a
measure at most k2 /v(G). Hence
( )
( ) k 1
(10.1) dvar G(k, G), G(k, WG ) ≤ .
2 v(G)
It is straightforward to extend this construction to generating a countable ran-
dom graph G(W ) on N: We generate an infinite sequence X1 , X2 , . . . of independent
uniformly distributed random points from [0, 1], and (as before) connect nodes i
and j with probability W (Xi , Xj ).
Remark 10.1. There are two ways of thinking about a graphon as a generalized
graph. First, we can consider it as a weighted graph with node set [0, 1]. Second, we
may think of each element x ∈ [0, 1] as an infinite set Sx of nodes with infinitesimally
157
158 10. SAMPLING

small measure, where there is a random bipartite graph Gx,y between Sx and Sy
with density W (x, y). These random bipartite graphs must be independent as
random variables, which makes this impossible to construct in standard measure
theory (one can construct such an object in non-standard analysis, cf. Section
11.3.2). But often this is a useful informal way of thinking of a graphon. The two
random samples H(n, W ) and G(n, W ) correspond to these two ways of looking at
graphons.
The definition of the sampling distance can also be extended from simple graphs
to graphons (recall (1.2) for graphs):
∞
∑ 1 ( )
(10.2) δsamp (U, W ) = k
dvar G(k, U ), G(k, W ) .
2
k=1

Using the fact that for any graphon U and simple graph F on node set [k], the
probability that G(k, U ) = F is just tind (F, U ), we have for all U, W ∈ W0
( ) 1 ∑
(10.3) dvar G(k, U ), G(k, W ) = |tind (F, U ) − tind (F, W )|.
2 simp
F ∈Fk

Hence
∑
(10.4) δsamp (U, W ) = 2−v(F )−1 |tind (F, U ) − tind (F, W )|,
F

where F ranges through all finite graphs with V (F ) = {1, . . . , v(F )}. By (10.1) the
distributions of G(k, G) and G(k, WG ) are almost the same if v(G) is large, and
hence
 
(10.5) δsamp (F, G) − δsamp (WF , WG ) ≤ 4 .
v(G)
While the sampling procedure described above is the most natural and most
often used, we sometimes need to sample in other ways. In Lemma 10.18 we
will describe a sampling method where the random selection of the nodes is more
restricted, but which is still good enough to get the same information about W
(however, we need much larger samples).
There are other uses of graphons and kernels in generating random graphs.
Bollobás, Borgs, Chayes and Riordan [2010] and Bollobás, Janson and Riordan
[2007] study sparse random graphs generated from a nonnegative kernel W by con-
structing a (W/n)-random graph on n nodes. Bollobás, Janson and Riordan [2010]
and Bollobás and Riordan [2009] study random trees generated from a graphon.
Palla, Lovász and Vicsek [2010] construct sparse random graphs as (W ⊗n )-random
graphs with n′ nodes, where n and n′ are chosen so as to keep the average degree
constant. We will not go into the details of these constructions.

10.2. Sample concentration

If we take a bounded size sample from a graph, we can see very different graphs.
For a sufficiently large random graph, for example, we can see anything. The natural
way to use the sample G[S] is to compute some graph parameter f (G[S]). But this
parameter can vary wildly with the choice of the sample, so what information do
we get?
 10.2. SAMPLE CONCENTRATION 159

The following theorem asserts that every reasonably smooth parameter of a

sample is highly concentrated. (Note: we don’t say anything here about the con-
nection between the value of the parameter on the whole graph and on the sample!
We return to this question in Chapter 15.) Let us define a reasonably smooth graph
parameter as a parameter f satisfying |f (G) − f (G′ )| ≤ 1 for any two graphs G and
G′ on the same node set whose edge sets differ only in edges incident with a single
node. More generally, we define parameter of edge-weighted graphs as reasonably
smooth if |f (H) − f (H ′ )| ≤ 1 for any two edge-weighted graphs on the same node
set that differ only in the weights of edges incident with a single node.
Theorem 10.2 (Sample Concentration for Graphs). Let f be a reasonably
smooth
( graph) parameter, let G be a graph, and let 1 ≤ k ≤ v(G). Let f0 =
E f (G(k, G)) , then for every t ≥ 0,
( ( ) √ )
P f G(k, G) ≥ f0 + 2tk ≤ e−t .
The result extends to graphons. We formulate two versions, corresponding to
the two sampling methods defined above.
Theorem 10.3 (Sample Concentration for Graphons). (a) Let f be a rea-
sonably( smooth simple
) graph parameter, let W ∈ W0 , and let k ≥ 1. Let
f0 = E f (G(k, W )) , then for every t ≥ 0,
( ( ) √ )
P f G(k, W ) ≥ f0 + 2tk ≤ e−t .
(b) Let f be a reasonably( smooth parameter
) of edge-weighted graphs. Let W ∈
W, let k ≥ 1, and let f0 = E f (H(k, W )) , then for every t > 0,
( ( ) √ )
P f H(k, W ) ≥ f0 + 2tk ≤ e−t .
In both theorems, we can apply the same inequality to the function −f , to
obtain a bound on the probability of a large deviation from the mean in the other
direction.
( )
Proof. The function f G({x1 , . . . , xk }, W ) (as a function of x1 , . . . , xk ∈
[0, 1]) satisfies the conditions of Corollary A.15 of Azuma’s Inequality, and hence
applying the inequality with n = k and ε = (2t/k)1/2 , the inequality in (a) follows.
The proof of (b) is essentially the same. 
( )
Applying this theorem with f (G) = v(G)/v(F ) tinj (F, G) (which is reason-
ably smooth), and combining it with (5.21), we get the following concentration
inequalities for subgraph densities:
Corollary 10.4. Let W ∈ W0 , n ≥ 1, 0 < ε < 1, and let F be a simple graph,
then the W -random graph G = G(n, W ) satisfies
( )
( ) ε2 n
P |tinj (F, G) − t(F, W )| > ε ≤ 2 exp −
2v(F )2
and ( )
( ) ε2 n
P |t(F, G) − t(F, W )| > ε ≤ 2 exp − . 
8v(F )2
We will see in Section 10.4 that not only numerical parameters of subgraph
samples are concentrated, but the samples themselves are concentrated in the cut
distance.
160 10. SAMPLING

10.3. Estimating the distance by sampling

10.3.1. The main sampling lemma. Among the main technical tools used
in this book are a couple of probabilistic theorems, which relate sampling to cut
distance. The first of these theorems is due to Alon, Fernandez de la Vega, Kan-
nan and Karpinski [2003], with an improvement by Borgs, Chayes, Lovász, Sós
and Vesztergombi [2008]. Its proof will be quite involved. Its main implication is
that the d -distance of two graphs on the same set of nodes can be estimated by
sampling.
Lemma 10.5 (First Sampling Lemma for Graphs). Let G and H be weighted
graphs with V (G) = V (H), with the same node weights, and with edge weights in
[0, 1]. Let k ≤ v(G) be a positive integer, and let S be chosen uniformly
√
from all
subsets of V (G) of size k. Then with probability at least 1 − 4e− k/10 ,
 
  8
d (G[S], H[S]) − d (G, H) ≤ 1/4 .
k
This Lemma extends to kernels, and this is the form which we prove. For
U ∈ W and X = (X1 , . . . , Xk ) ⊆ [0, 1], let U [X] denote the symmetric k × k matrix
defined by (U [X])ij = U (Xi , Xj ).
Lemma 10.6 (First Sampling Lemma for Kernels). Let U ∈ W1 and let√X be
a random ordered of k-subset of [0, 1]. Then with probability at least 1 − 4e− k/10 ,
3 8
− ≤ ∥U [X]∥ − ∥U ∥ ≤ 1/4 .
k k
Not only are the lower and upper bounds in this lemma different, they are also
quite different in difficulty. To prove the lower bound is rather straightforward, but
the proof of the upper bound will need a couple of lemmas about a tricky sampling
procedure estimating the sum of entries of a matrix.
It will be more convenient to work with the following one-sided version of the
cut norm: ∑
1
∥A∥+ = n2 max Aij
S,T ⊆[n]
i∈S,j∈T
for an n × n matrix A, and
∫
∥W ∥+
 = sup W (x, y) dx dy
S,T ⊆[0,1]
S×T

for a kernel W . We note that ∥A∥ = max{∥A∥+  , ∥−A∥ }, and similarly for the
+

cut norm of kernels. In terms of this norm, we are going to prove the following
similar bounds:
Lemma 10.7. Let U ∈ W1 and let √ X be a random ordered of k-subset of [0, 1].
− k/10
Then with probability at least 1 − 2e ,
3 8
− ≤ ∥U [X]∥ − ∥U ∥ ≤ 1/4 .
k k
Let B = U ∑[X]. For any set Q 1 of rows and any set Q2 of columns, we set
B(Q1 , Q2 ) = B
i∈Q1 ,j∈Q2 ij . We denote by Q+
1 the set of columns j ∈ [k] for
−
which B(Q1 , {j}) > 0. We define the set of columns Q1 and the sets of rows
−
2 , Q2 analogously. Note that B(Q1 , Q1 ), B(Q2 , Q2 ) ≥ 0 by this definition.
Q+ + +
 10.3. ESTIMATING THE DISTANCE BY SAMPLING 161

We start with proving an inequality for the case when only a random subset Q
of columns is selected.
Lemma 10.8. Let S1 , S2 ⊆ [k], and let Q be random q-subset of [k] (1 ≤ q ≤ k).
Then
( ) k2
B(S1 , S2 ) ≤ EQ B((Q ∩ S2 )+ , S2 ) + √ .
q
Proof. The inequality is clearly equivalent to the following:
( ) k2
(10.6) EQ B((Q ∩ S2 )− , S2 ) ≤ √ .
q
Note that there is no absolute value on the left side: the expectation of B(Q ∩
S2 )− , S2 ) can be very negative, but not very positive. The lemma says that the set
Q ∩ S2 )− tends to pick out those rows whose sum∑is small. ∑
Consider row i of B. Let m = |S2 |, bi = j∈S2 Bij , ci =
2
j∈S2 Bij and
∑
Ai = j∈Q∩S2 Bij . The contribution of row i to the left side is bi if Ai ≤ 0
(i.e., i ∈ (Q ∩ S2 )− ), and 0 otherwise. So the expected contribution of row i is
P(Ai ≤ 0)bi .
If bi ≤ 0, then this contribution is nonpositive. Else, we use Chebyshev’s
inequality to estimate the probability of Ai ≤ 0. We have E(Ai ) = qbi /k and
Var(Ai ) < qci /k. Hence
(
 qbi  qbi ) k 2 Var(Ai ) kci
P(Ai ≤ 0) ≤ P Ai − ≥ ≤ < 2.
k k q 2 b2i qbi
The probability on the left is at most 1, and so we can bound it from above by its
square root:
√
√ kci
P(Ai ≤ 0) ≤ P(Ai ≤ 0) ≤ √ .
qbi
( ) √
So the contribution of row i to EQ B((Q ∩ S2 )− , S2 ) is P(Ai ≤ 0)bi ≤ kci /q ≤
√
k/ q. Summing over all i ∈ S1 , inequality (10.6) follows. 

The following lemma gives an upper bound on the one-sided cut norm, using
the sampling procedure from the previous lemma.
Lemma 10.9. Let S1 , S2 ⊆ [k], and let Q1 and Q2 be random q-subsets of [k],
(1 ≤ q ≤ k). Then
1 ( ) 2
∥B∥+
 ≤ 2
E Q 1 ,Q2 max B(R2
+
, R +
1 ) +√ .
k Ri ⊆Qi q
The Lemma estimates the (one-sided) cut norm by maximizing only over cer-
tain rectangles (at the cost of averaging these estimates). the main point for our
purposes will be that (for a fixed Q1 and Q2 ), the number of rectangles to consider
is only 4q , as opposed to 4k in the definition of the cut norm.

Proof. Fix any two sets S1 , S2 ⊆ [k]. By Lemma 10.8,

( ) k2
(10.7) B(S1 , S2 ) ≤ EQ2 B((Q2 ∩ S2 )+ , S2 ) + √ .
q
162 10. SAMPLING

We apply Lemma 10.8 again, interchanging the roles of rows and columns:
( k2
B((Q2 ∩ S2 )+ , S2 ) ≤ EQ1 B((Q2 ∩ S2 )+ , (Q1 ∩ (Q2 ∩ S2 )+ )+ ) + √
q
( ) k2
≤ EQ1 max B(R1+ , R2+ ) + √ .
Ri ⊆Qi q
Substituting in (10.7), the Lemma follows. 

Now we can turn to the main part of the proof.

Proof of Lemma 10.7. To bound the difference ∥B∥+  −∥U ∥ , we first bound
+

its expectation. For any two measurable subsets S1 , S2 ⊂ [0, 1], we have
1
∥B∥+
 ≥ U (S1 ∩ X, S2 ∩ X),
k2
∑
(where U (Z1 , Z2 ) = x∈Z1 ,y∈Z2 U (x, y) for finite subsets Z1 , Z2 ⊂ [0, 1]). Choosing
the set X randomly, we get
( ) 1 ( )
EX ∥B∥+  ≥ 2 EX U (S1 ∩ X, S2 ∩ X)
k ∫ ∫
k−1 1
= U (x, y) dx dy + U (x, x) dx
k k
S1 ×S2 S1 ∩S2
∫
2
≥ U (x, y) dx dy −
k
S1 ×S2

Taking the supremum of the right side over all measurable sets S1 , S2 we get
( ) 2
EX ∥B∥+  ≥ ∥U ∥ − .
+
k
From here, the bound follows by sample concentration (Theorem 10.3).
To prove an upper bound on the difference ∥B∥+  − ∥U ∥ , let Q1 and Q2 be
+
√
random q-subsets of [k], where q = ⌊ k/4⌋. Lemma 10.9 say that for every X,
1 ( ) 2
∥B∥+  ≤ 2
E Q 1 ,Q2 max B(R2
+
, R +
1 ) +√ .
k Ri ⊆Qi q
Next we take expectation over the choice of X. More precisely, we fix the sets
Ri ⊆ Qi ⊆ [k], and also those∑ points Xi ∈ [0, 1] for which i ∈ Q = Q1 ∪ Q2 .
Define Y1 = {y ∈ [0, 1] : i∈R1 U (Xi , y) > 0}, and define Y2 analogously. Let
X ′ = (Xi : i ∈ [k] \ Q), then for every i ∈ S∫1 \ Q and j ∈ S2 \ Q, the contribution
of the term U (Xi , Xj ) to EX ′ B(R2+ , R1+ ) is Y1 ×Y2 U ≤ ∥U ∥+
 . The contribution of
the remaining terms U (Xi , Xj ) with either i ∈ Q or j ∈ Q is at most 2k|Q| ≤ 4kq
in absolute value. Hence
(10.8) EX ′ B(R2+ , R1+ ) ≤ k 2 ∥U ∥+
 + 4kq.

Next we show that the value of B(R2+ , R1+ ) is highly concentrated around its
expectation. This is a function of the independent random variables Xi , i ∈ [k] \ Q,
and if we change the value of one of these Xi , the sum B(R2+ , R1+ ) changes by at
most 4k (there are fewer than 2k entries that may change, and each of them by at
 10.3. ESTIMATING THE DISTANCE BY SAMPLING 163

most 2). We can apply Corollary A.15 of Azuma’s Inequality, and conclude that
with probability at least 1 − e−1.9q , we have
√ √
B(R2+ , R1+ ) ≤ EX ′ B(R2+ , R1+ ) + 7.9k kq ≤ k 2 ∥U ∥+
 + 4kq + 7.9k kq.
The number of possible pairs of sets R1 and R2 is 4q , and hence with probability
at least 1 − 4q e−1.9q > 1 − e−q/2 , this holds for all R1 ⊆ Q1 and R2 ⊆ Q2 , and so it
holds for the maximum. Taking expectation over Q1 and Q2 does not change this,
so we get that with probability (over X) at least 1 − e−q/2 , we have
√
2 4q 7.9 q
∥B∥+  ≤ ∥U ∥+
 + √ + + √ .
q k k
This implies the upper bound in the lemma by simple computation (if k large
enough). 

Proof of Lemma 10.6. Applying √

Lemma 10.7 to both kernels U and −U ,
− k/10
with probability at least 1 − 4e all four inequalities will hold, and in this case
so do the inequalities in the Lemma. 

10.3.2. First applications. We can apply the First Sampling lemma when
U = W1 − W2 is a difference of two graphons. Considering Wi [X] as the edge-
weighted graph H(X, Wi ), Lemma 10.6 implies the following:
Corollary 10.10. Let W1 , W2 ∈ W0 and let X be a sequence of k ≥ 1 random
points of [0, 1] chosen independently
√
from the uniform distribution. Then with prob-
− k/10
ability at least 1 − 4e ,
 
  8
d (H(X, W1 ), H(X, W2 )) − ∥W1 − W2 ∥  ≤ 1/4 .
k
In terms of the random weighted ( graphs H(k, W1 ) )and H(k, W2 ) this means
that they can be coupled so that d H(k, W1 ), H(k, W2 ) ≈ δ (W1 , W2 ) with high
probability. We will see that more is true: H(k, W ) will be close to W in the cut
distance with high probability. (However, quantitatively “closeness” will be much
weaker.)
We have seen that the cut distance of two samples H(k, W1 ) and H(k, W2 ) is
close to the distance of W1 and W2 (if coupled appropriately). How about the
simple graphs G(k, W1 ) and G(k, W2 )? The following simple lemma shows that if
k is large enough, then G(k, W ) is close to H(k, W ), so similar conclusions hold.
Lemma 10.11. For every edge-weighted graph H with edgeweights in [0, 1], and
√
for every ε ≥ 10/ q,

P(d (G(H), H) > ε) ≤ e−ε q /100 .

2 2

√
Applying this inequality with ε = 10/ q and bounding the distance by 1 in
the exceptional cases, we get the inequality
( ) 11
(10.9) E(d G(H), H) ≤ √ .
q
Note that no similar assertion would hold for the distances d1 or d2 . For example,
if all edgeweights of H are 1/2, then d1 (G(H), H) = d2 (G(H), H) = 1/2 for any
instance of G(H).
164 10. SAMPLING

( )
Proof. For i, j ∈ [q], define the random variable Xij = 1 ij ∈ E(G(H)) . Let
S and T be two disjoint subsets of [q]. Then the Xij (i ∈ S, j ∈ T ) are independent,
and E(Xij ) = βij (H), which gives that
∑ ( )
eG(H) (S, T ) − eH (S, T ) = Xij − E(Xij ) .
i∈S, j∈T

Let us call the pair (S, T ) bad, if |eG(H) (S, T ) − eH (S, T )| > εq 2 /4. The probability
of this can be estimated by the Chernoff–Hoeffding Inequality:
( ∑ (
 ) 1 ) ( ε2 q 4 ) ( −ε2 q 2 )
P  Xij − E(Xij )  > εq 2 ≤ 2 exp − ≤ 2 exp .
4 32|S||T | 32
i∈S, j∈T

The number of disjoint pairs (S, T ) is 3q , and so the probability that there is a bad
pair is bounded by 2 · 3q e−ε q /32 < e−ε q /100 . If there is no bad pair, then it is
2 2 2 2

easy to see that d (G(H), H) ≤ ε (cf. Exercise 8.4). This completes the proof. 
This lemma implies that the weighted sample in the First Sampling Lemma
can be replaced by a simple graph at little cost. We state one corollary:
Corollary 10.12. Let W1 , W2 ∈ W0 and k ≥ 1. Then the random graphs √
G(k, W1 )
and G(k, W2 ) can be coupled so that with probability at least 1 − 5e− k/10 ,
 ( ) 
  10
d G(k, W1 ), G(k, W2 ) − ∥W1 − W2 ∥  ≤ 1/4 . 
k

Exercise 10.13. Derive the First Sampling Lemma for graphs (Lemma 10.5)
from the graphon version (Lemma 10.6). Attention: sampling from a graph G
and sampling from WG does not quite give the same distribution!
Exercise 10.14. Prove the (much easier) analogue of the First Sampling Lemma
for the edit distance: Let G and H be simple graphs V (G) = V (H). Let k ≤ v(G)
be a positive integer, and let S be chosen uniformly from all ordered subsets of
V (G) of size k. Then
( ) (k − 1)n
E d1 (G[S], H[S]) = d1 (G, H),
k(n − 1)
2
and for every ε > 0, with probability at least 1 − 2e−kε /2 ,
 
 
d1 (G[S], H[S]) − d1 (G, H) ≤ ε.

10.4. The distance of a sample from the original

A second lemma about sampling that will be used very often, due to Borgs,
Chayes, Lovász, Sós and Vesztergombi [2008], shows that a sample is close to the
original graph (or graphon) with high probability. Note that here we have to use the
δ distance, rather than the d distance, since the graphs have different number of
nodes, and no overlaying is given a priori. Also note that the bound on the distance
is much weaker than in the previous lemma (but it does tend to 0 with the sample
size).
Lemma 10.15 (Second Sampling Lemma for Graphs). Let k ≥ 1, and
let G be
( a simple graph
) on at least k nodes. Then with probability at least
1 − exp −k/(2 log k) ,
( ) 20
δ G, G(k, G) ≤ √ .
log k
 10.4. THE DISTANCE OF A SAMPLE FROM THE ORIGINAL 165

The Second Sampling Lemma also extends to graphons, which can be stated
in terms of the W -random graphs H(k, W ) and G(k, W ).
Lemma 10.16 (Second Sampling Lemma for Graphons). ( Let k ≥ 1,) and let
W ∈ W0 be a graphon. Then with probability at least 1 − exp −k/(2 log k) ,
20
δ (H(k, W ), W ) ≤ √ ,
log k
and
22
δ (G(k, W ), W ) ≤ √ .
log k
Proof. First we prove that these inequalities hold in expectation. Let m =
⌈k 1/4 ⌉. By Lemma 9.15, there is an equipartition P = {V1 , . . . , Vm } of [0, 1] into m
classes such that
8
d (W, WP ) ≤ √ .
log k
Let S be a random k-subset of [0, 1], then by the First Sampling Lemma 10.6, we
have
 
d (W [S], WP [S]) − d (W, WP ) ≤ 8
k 1/4
with high probability. This implies that
( ) 10
E d (W [S], WP [S]) − d (W, WP ) ≤ 1/4
k
(k is large enough for this, else the bound in the lemma is trivial), and so
( ) ( )
E d (W [S], WP [S]) ≤ E d (W [S], WP [S]) − d (W, WP ) + d (W, WP )
9
≤√ .
log k
So it suffices to prove that δ (WP , WP [S]) is small on the average.
Let H = WP [S]. The graphons WP and WH are almost the same: both are
stepfunctions with m steps, with the same function values on corresponding steps.
The only difference is that the measure of the i-th step Vi in WP is 1/m, while the
measure of the i-th step in WH is |Vi ∩ S|/k, which is expected to be close to 1/m
if k is large enough. ∑
Write |Vi ∩ S|/k = 1/m + ri , then it is easy to see that δ (WP , WH ) ≤ i |ri |.
Hence it is easy to estimate the expectation of this distance, using elementary
probability theory:
√ √
( ) ∑ ( ) ( ) m−1 1
E δ (WP , WH ) ≤ E |ri | = mE |r1 | ≤ m E(r1 ) = 2 < 3/8 .
i
k k
Hence
( ) ( )
E(δ (W, W [S]) ≤ δ (W, WP ) + E δ (WP , WP [S]) + E δ (WP [S], W [S])
8 1 9 18
≤√ + +√ ≤√ .
log k k 3/8 log k log k
( )
A similar estimate for δ W, G(k, W ) follows if we invoke inequality (10.9):
( ) ( ) ( )
E δ (W, G(k, W )) ≤ E δ (W, H(k, W )) + E δ (H(k, W ), G(k, W ))
18 11 20
≤√ +√ <√ .
log k k log k
166 10. SAMPLING

Now the Lemma follows by the Sample Concentration Theorem 10.3 applied
to the graph parameter f (G) = v(G)δ (G, W ). 

We have seen that the (weak) Regularity Lemma implies that we can approxi-
√
mate every simple graph G by a weighted graph on k nodes with error O(1/ log k).
As an application of the Second Sampling Lemma, we get that we can also approx-
imate every simple graph G by an (unweighted) simple graph H on k nodes, at the
cost of a constant factor in the error:
Corollary 10.17. For every k ≥ 1 and simple graph G, there is a simple graph H
with k nodes such that
10
δ (G, H) ≤ √ .
log k
We need a version of the Second Sampling Lemma for the modified sampling
procedure mentioned above in Remark 10.1. Let W be a graphon and n ≥ 1.
Let S = (s1 , . . . , sn ), where si is a random uniform point from the interval [(i −
1)/n, i/n]. We denote the random graph G(S, W ) by G′ (n, W ). The following
bound was proved by Lovász and Szegedy [2010a].
Lemma 10.18. For√every graphon W and positive integer k, we have with proba-
bility at least 1 − 5/ k,
176
δ (G′ (k, W ), W ) < √ .
log k
Proof. The trick is to do a second sampling: we choose a random r-tuple T of
nodes of G′ = G′ (k, W ), where r = ⌈k 1/4 ⌉. We may assume that k > 25, else there
is nothing to prove.
( The Second
) Sampling Lemma implies that with probability at
least 1 − 2 exp −r/(2 log r) , we have
22
δ (G′ [T ], G′ ) ≤ √ .
log r
Now we can generate G′ [T ] = G(r, G′ ) in the following way: we choose a random
sequence X of r independent uniform points in [0, 1]; if they belong to different
intervals Ji = [(i − 1)/k, i/k], then we return G(X, W ); else, we try again. This
gives us a coupling between G(r, W ) and G(r, G′ ) such that
( ) r(r − 1)
P G(r, W ) ̸= G(r, G′ ) ≤ P(∃i : |X ∩ Ji | ≥ 2) ≤ .
k
Invoking
( the Second
) Sampling Lemma again, with probability at least 1 −
2 exp −r/(2 log r) we have
22
δ (G(r, W ), W ) ≤ √ ,
log r
and hence with probability at least
r r(r − 1) 5
1 − 4 exp(− )− ≥1− √
2 log r k k
we have
44 176
δ (G′ , W ) ≤ δ (G′ , G′ [T ]) + δ (G(r, W ), W ) ≤ √ ≤√ . 
log r log k
 10.5. COUNTING LEMMA 167

Exercise 10.19. Consider the template graph H of a weak regularity partition,

with k almost equal classes, of a large graph G, and turn it to a simple graph
by the method of Lemma 10.11. Prove that the (random)( simple
) graph
√ G(H)
obtained this way satisfies, with high probability, δ G, G(H) ≤ 10/ log k.
Exercise 10.20. Let k ≥ 1, let W be a graphon, and let S1 and S2 be two
independent random k-subsets of [0, 1]. Then with probability at least 1 − 21−k ,
( ) 22
δb G(S1 , W ), G(S2 , W ) ≤ √
log k
(note the “hat” over the δ).
Exercise 10.21. Let f be a graph parameter and assume that |f (G) − f (G′ )| ≤
d (G, G′ ) for any two graphs on the same node set. Then for every graph G and
1 ≤ k ≤ v(G) there is a value f0 such that if S ⊆ V (G) is a random k-subset, then
22
|f (G[S]) − f0 | < √
log k
with probability at least 1 − o(1).

10.5. Counting Lemma

It is time to relate the two main quantities we introduced to study large dense
graphs and graphons: homomorphism densities (which are equivalent to sample
distributions) and the cut distance. The following simple but fundamental relation
between them, due to Lovász and Szegedy [2006], is a generalization of the “Count-
ing Lemma” in the theory of Szemerédi partitions. (Lemma 10.32, which will be
more difficult to prove, will state a certain converse of this fact.)
We start with a combinatorial formulation.
Lemma 10.22 (Counting Lemma for Graphs). For any three simple graphs
F , G and G′
|t(F, G) − t(F, G′ )| ≤ e(F ) δ (G, G′ ).
The Lemma extends to graphons:
Lemma 10.23 (Counting Lemma for Graphons). Let F be a simple graph and
let W, W ′ ∈ W0 . Then
|t(F, W ) − t(F, W ′ )| ≤ e(F )δ (W, W ′ ).
This lemma shows that for any simple graph F , the function W 7→ t(F, W )
is Lipschitz-continuous on W0 in the metric δ . At the end of this section, we
state several further versions of the Counting Lemma as exercises. The proof of
the lemma will be given in the more general setting of W0 -decorated graphs (which
actually makes the proof simpler!). Recall that a W0 -decorated graph is a simple
graph in which a graphon We is assigned to each edge e. Also recall the definition
(7.16) of homomorphism density of such a decorated graph.
Lemma 10.24 (Counting Lemma for decorated graphs). Let (F, w) and
(F, w′ ) be two W0 -decorated graphs with the same underlying simple graph, where
w = (We : e ∈ E) and w′ = (We′ : e ∈ E). Then
∑
|t(F, w) − t(F, w′ )| ≤ ∥We − We′ ∥ .
e∈E(F )
168 10. SAMPLING

Proof. It suffices to prove this bound for the case when We = We′ for all edges
′
but one. Let F = (V, E), and let uv be the edge with Wuv ̸= Wuv . Then
∫ ∏ ( )
t(F, w) − t(F, w′ ) = Wij (xi , xj ) Wuv (xu , xv ) − Wuv′
(xu , xv ) dx
[0,1]V ij∈E(F )\{uv}
∫
( ′
)
= f (x)g(x) Wuv (xu , xv ) − Wuv (xu , xv ) dx,
[0,1]V

where ∏
f (x) = Wij (xi , xj )
ij∈∇(u)\uv

does not depend on xv , and satisfies 0 ≤ f ≤ 1. Similarly,

∏
g(x) = Wij (xi , xj )
ij∈E\∇(u)

does not depend on xu , and satisfies 0 ≤ g ≤ 1. Fixing all variables except xu and
xv , we get the following estimate by Lemma 8.10:
 ∫ ( ) 
 ′  ′
 f (x)g(x) Wuv (xu , xv ) − Wuv (xu , xv ) dxu dxv  ≤ ∥Wuv − Wuv ∥ .
[0,1]2

Integrating over the remaining variables, we get that

|t(F, w) − t(F, w′ )| ≤ ∥Wuv − Wuv
′
∥ . 

From this lemma, along with (10.3) and (7.4), it is easy to derive a relationship
between the variation distance of the distributions of the random graphs G(k, U )
and G(k, W ), and the cut distance of U and W .
Corollary 10.25. Let U and W be two graphons, then for every k ≥ 2, we have
( ) 2
dvar G(k, U ), G(k, W ) ≤ 2k δ (U, W ).

Exercise 10.26. Show that the Counting Lemma does not hold for multigraphs,
not even for F = C2 .
Exercise 10.27. Let F be a simple graph with m edges and let W, W ′ ∈ W1 .
Then
|t(F, W ) − t(F, W ′ )| ≤ 4mδ (W, W ′ ).
Exercise 10.28. Let F be a simple graph with m edges and let W ∈ W1 . Then
|t(F, W )| ≤ 4m∥W ∥ .
Exercise 10.29. For every W1 -decorated graph (F, w),
t(F, w) ≤ 4 min ∥We ∥ .
e∈E(F )

Exercise 10.30. Prove the following “induced” version of the Counting Lemma:
If F is a simple graph and U, W ∈ W0 , then
( )
k
|tind (F, U ) − tind (F, W )| ≤ 4 ∥U − W ∥ .
2
2
Use this to improve the coefficient 2k in Corollary 10.25.
 10.6. INVERSE COUNTING LEMMA 169

10.6. Inverse Counting Lemma

Our goal is to establish a converse to the Counting Lemma: if two “large”
graphs are locally close (in the sense of sampling or homomorphism densities) then
they are globally close (in the sense of cut distance). This treatment is based on
Borgs, Chayes, Lovász, Sós and Vesztergombi [2008]. We prove two versions, both
of which will play an important role later on.
Lemma 10.31. Let U and W be two graphons and suppose that for some k ≥ 2,
we have
( ) ( k )
dvar G(k, U ), G(k, W ) < 1 − 2 exp − .
2 log k
Then
50
δ (U, W ) ≤ √ .
log k
Note that the bound on the variation distance of the distributions of the random
subgraphs G(k, U ) and G(k, W ) is very weak: a tiny overlap between them already
implies that the graphons U and W are close. Applying the lemma to WG1 and
WG2 gives a similar result for two large graphs.

Proof. The assumption implies that we can couple G(k, U )( and G(k, W))
so that G(k, U ) = G(k, W ) with probability larger than 2 exp −k/(2 log k) .
The( Second Sampling
) Lemma 10.16 implies that with probability at least 1 −
exp −k/(2 log k) , we have
( ) 22
δ U, G(k, U ) ≤ √ ,
log k
and similar assertion holds for W . It follows that with positive probability all three
happen, and then we get
( ) ( ) 50
δ (U, W ) ≤ δ U, G(k, U ) + δ W, G(k, W ) ≤ √ . 
log k

Lemma 10.32 (Inverse Counting Lemma). Let k be a positive integer, let

U, W ∈ W0 , and assume that for every simple graph F on k nodes, we have

|t(F, U ) − t(F, W )| ≤ 2−k .

2

Then
50
δ (U, W ) ≤ √ .
log k
Proof. Assume that U, W ∈ W0 satisfy

|t(F, U ) − t(F, W )| ≤ 2−k

2

for every graph F with k nodes. This implies (by inclusion-exclusion) that
k k+1
|tind (F, U ) − tind (F, W )| ≤ 2(2) 2−k = 2−( 2 ) .
2

In terms of the W -random graphs G(k, U ) and G(k, W ),

 ( ) ( )
P G(k, U ) = F − P G(k, W ) = F  ≤ 2−( 2 ) .
k+1
170 10. SAMPLING

Hence
( ) ∑ ( ) ( )
P G(k, U ) = F − P G(k, W ) = F  ≤ 2(2) 2−( 2 )
k k+1
dvar G(k, U ), G(k, W ) =
F
(k )
= 2−k < 1 − 2 exp − .
2 log k
An application of Lemma 10.31 completes the proof. 

Exercise 10.33. Prove that for any two graphons U and W ,

√
1 1 ( 400 )
log ≤ log ≤ exp 2
.
δ (U, W ) δsamp (U, W ) δ (U, W )

10.7. Weak isomorphism II

An important consequence of (10.4) is that two weakly isomorphic graphons
have sampling distance 0, i.e., they are indistinguishable by sampling. The converse
of this assertion follows by the same kind of argument.
The significance of this easy remark is that it allows us to relate weak iso-
morphism to the cut distance via the sampling distance. We start with a more
general fact, showing the topological equivalence of the sampling distance and the
cut distance on the space of graphons.
We have noted that two graphons U and W are weakly isomorphic (i.e.,
t(F, U ) = t(F, W ) for every simple graph) if and only if their sampling distance
is 0. The Counting Lemma and the Inverse Counting Lemma imply that two
graphons are weakly isomorphic if and only if their cut distance is 0. It is easy to
see that this implies the same conclusion for general kernels:
Corollary 10.34. Two kernels U and W are weakly isomorphic if and only if
δ (U, W ) = 0.
Since δ (U, W ) = 0 expresses the existence of a correspondence between the
points of the two graphons, this theorem can be considered as a generalization of
Theorem 5.29 (which can be derived from it with some effort).
The proof of Corollary 10.34, if we include the proofs of the Counting Lemma
and the Inverse Counting Lemma, is quite long, and in particular the proof of the
Inverse Counting Lemma, which builds on the First Sampling Lemma, is quite
involved. One can get a more direct proof using only rather standard analysis; see
Exercise 11.27.
Theorem 8.13 and its Corollary 8.14, for the special case of the cut norm and
cut-distance 0, imply the following further characterizations of weak isomorphism:
Corollary 10.35. (a) Two kernels U and W are weakly isomorphic if and only
if there exist measure preserving maps φ, ψ : [0, 1] → [0, 1] such that U φ = W ψ
almost everywhere.
(b) Two kernels U and W are weakly isomorphic if and only if there exists a
coupling measure µ on [0, 1]2 such that for two random samples (x1 , y1 ) and (x2 , y2 )
from µ, we have U (x1 , x2 ) = W (y1 , y2 ) with probability 1.
As a further corollary we get the following fact (stated before as Exercise 7.18):
Corollary 10.36. If two kernels U, W ∈ W are weakly isomorphic, then t(F, U ) =
t(F, W ) holds for all multigraphs F .
 10.7. WEAK ISOMORPHISM II 171

Exercise 10.37. Construct the coupling measures in Theorem 8.13 for the cut
distance of the three weakly isomorphic graphons in Example 7.11.
Exercise 10.38. Show by an example that the sampling distance and the cut
distance do not define the same topology on the set of finite graphs.
 CHAPTER 11

Convergence of dense graph sequences

Finally we have come to the central topic of this book: convergent graph se-
quences and their limits. The two key elements, namely sampling and graphons,
have been introduced in the Introduction. Here we take our time to look at them
from various aspects.

11.1. Sampling, homomorphism densities and cut distance

Recall from the introduction that we can define a notion of convergence if
we fix a sampling method. For dense graphs, we use subgraph sampling: We select
uniformly a random k-element subset of V (G), and return the subgraph induced by
it. The probability that we see a given graph F is the quantity tind (F, G) introduced
in (5.13). A sequence of graphs (Gn ) with v(Gn ) → ∞ is convergent if the induced
subgraph densities tind (F, Gn ) converge for every finite graph F .
It is often more convenient to define convergence using the homomorphism
densities t(F, Gn ) or the subgraph densities tinj (F, Gn ). This does not change the
notion of convergence as introduced above in terms of sampling. Indeed, subgraph
densities can be expressed as linear combinations of induced subgraph densities and
vice versa (we have discussed such relations in Section 5.2.3), and hence tinj (F, Gn )
tends to a limit as n → ∞ if and only if tind (F, Gn ) does. For the homomorphism
densities
( ) argument is a bit more involved: we know that t(F, G) − tinj (F, G) =
the
O 1/v(G) , and so this difference tends to 0 if v(G) → ∞. Hence t(F, Gn ) tends to
a limit as n → ∞ if and only if tinj (F, Gn ) does.
This notion of convergence of dense graphs is often called left-convergence, since
it is based on homomorphisms “from the left”. In the case of dense graphs, this no-
tion is rather robust (it seems to be the only reasonable way to define convergence),
and hence we call it simply “convergence”. The parallel notion of right-convergence
(which will turn out to be equivalent, at least if defined properly) will be discussed
in Chapter 12.
Many examples of convergent graph sequences will be shown in Section 11.4.2,
but let us describe a couple of very simple ones here.
Example 11.1. The sequence of complete graphs is convergent, since a random
induced k-node subgraph is always a complete graph itself. 
Example 11.2. Fix any 0 ≤ p ≤ 0, and generate a random graph Gn = G(n, p)
for every n. This sequence will be convergent with probability 1. Indeed, for large
n, a random induced k-subgraph Gn [S] of Gn will be very close in distribution to
G(k, p), for most choices of Gn . (Note that what we mean here is that Gn is fixed,
the randomness comes from the choice of the k-subset S). This is not hard to verify
directly, using elementary probability theory; it also follows from the much stronger
results in Chapter 10. 
173
174 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

The definition of convergence can be reformulated using the notion of sampling

distance (1.2): a sequence (Gn ) of simple graphs with v(Gn ) → ∞ is convergent if
for every graph F , (tind (F, Gn ) : n = 1, 2, . . . ) is a Cauchy sequence (equivalently,
(t(F, Gn ) : n = 1, 2, . . . ) is a Cauchy sequence). This is equivalent to saying that
the graph sequence is Cauchy in the δsamp metric (1.2). The following theorem of
Borgs, Chayes, Lovász, Sós and Vesztergombi [2006, 2008], which is one of the main
results in this theory, justifies the use of the cut metric δ .
Theorem 11.3. A sequence (Gn ) of simple graphs with v(Gn ) → ∞ is convergent
if and only if it is a Cauchy sequence in the metric δ .
Proof. The Counting Lemma 10.22 implies that every Cauchy sequence in
the metric δ is convergent. The Inverse Counting Lemma 10.32 (applied to the
graphons WGn ) implies the converse. 

Remark 11.4. This proof builds on a fairly long chain of previous results, some of
which, like the First Sampling Lemma 10.6, were quite involved. The advantage of
this proof is that it gives a quantitative form of the equivalence of two convergence
notions. As pointed out by Schrijver, a weaker qualitative form is easier to prove,
inasmuch we can replace the use of the Inverse Counting Lemma by the characteri-
zation of weak isomorphism (for which a simple direct proof is sketched in Exercise
11.27). Indeed, consider the two metric spaces (W f0 , δ ) (the graphon space) and
F
[0, 1] (the space of graph parameters with values in [0, 1]). Both of these are com-
pact (one by the Compactness Theorem 9.23, the other by Tychonoff’s Theorem).
The map W 7→ t(., W ) is continuous by the Counting Lemma, and injective by
Corollary 10.34, and hence its inverse is also continuous. For a convergent sequence
of graphs, this means precisely that the graphons WGn form a convergent sequence
in (Wf0 , δ ).

Theorem 11.3 can be generalized to characterize convergence in the space W.

The proof is the same, except that the graphon versions of the Counting Lemmas
must be used.
Theorem 11.5. Let (Wn ) be a sequence of graphons in W0 and let W ∈ W0 . Then
t(F, Wn ) converges for all finite simple graphs F if and only if Wn is a Cauchy
sequence in the δ distance. Furthermore, t(F, Wn ) → t(F, W ) for all finite simple
graphs F if and only if δ (Wn , W ) → 0. 

11.2. Random graphs as limit objects

Once we have defined convergence of a graph sequence, we would like to an-
swer the question: what does it converge to? I have told you (and justified by some
pictures) that the answer is “graphons”, but let us dwell on this question in a more
abstract setting for a while. In an abstract sense, we can assign a “limit object”
to every convergent sequence: we say that two convergent sequences are “equiva-
lent” if interlacing them we get a convergent sequence, and the limit objects can
be defined as equivalence classes of convergent sequences. This abstract definition
is not of much help; we are going to describe much more explicit representations
for the limit objects. Our favorite one is the graphon, but we’ll see other, equiva-
lent representations in the form of random graph models, reflection positive graph
parameters, and more.
 11.2. RANDOM GRAPHS AS LIMIT OBJECTS 175

11.2.1. Finite random graph models. The first construction of a limit

object, which we call the weak limit, is actually quite general: it can be constructed
for convergence of any reasonable sequence of structures for which we have any
reasonable sampling process. (We will see later, for example, how it works for
graphs with bounded degree.)
Given a simple graph G and k ∈ [v(G)], the random sample G(k, G) is a
random graph on k labeled nodes; we denote its distribution by σG,k . Clearly
σG,k (F ) = tind (F, G). If (G1 , G2 , . . . ) is a convergent graph sequence, then the dis-
tributions σGn ,k tend to some distribution σk on k-node labeled graphs. Conversely,
if the distributions σGn ,k tend to a limit for every k, then the graph sequence is
convergent. (The distribution σGn ,k may be undefined for a finite number of indices
n for every fixed k.) So the sequence of limit distributions (σ1 , σ2 , . . . ) encodes the
“limit” of the convergent graph sequence. Which sequences of distributions arise
this way?
A random graph model is a probability distribution σk on simple graphs on
node set [k], for every k ≥ 1, which is invariant under the reordering of the nodes.
In other words, it is a sequence of random variables Gk , whose values are simple
graphs on [k], and isomorphic graphs have the same probability. We say that a
random graph model is consistent if deleting node k from Gk , the distribution of
the resulting graph is the same as the distribution of Gk−1 . In formulas, this means
that for every graph H on k − 1 nodes,
∑
(11.1) σk−1 (H) = σk (F ),
F : F ′ =H

where F ′ denotes the graph obtained by deleting node k from F . We say that the
model is local, if for two disjoint subsets S, T ⊆ [k], the subgraphs of Gk induced
by S and T are independent as random variables.
We note that consistency, together with the invariance under reordering
the nodes,
( implies
) that for every simple graph F on k nodes, the expectation
E(tind F, Gn ) = σk (F ) is independent of n once n ≥ k.
Example 11.6. For every graphon W , the random graph model Gk = G(k, W )
is both consistent and local, which is trivial to check. (It will turn out that this
example represents all such models.) 
Theorem 11.7. If a graph sequence (G1 , G2 , . . . ) is convergent, then the distribu-
tions σk = limn→∞ σk,Gn form a consistent and local random graph model. Con-
versely, every consistent and local random graph model arises this way.
Before proving this theorem, we need some preparation. Let G be a graph and
k ≤ v(G). The sequence of distributions (σG,1 , σG,2 , . . . ) is not quite consistent,
because it breaks down for k > v(G); but it is consistent for the values of k for
which it is defined. There is a more serious problem with locality: selecting i
distinct random nodes of G will bias the selection of the remaining k − i, if we insist
on selecting distinct points. So locality will be only approximately true.
We can fix both problems if we consider the slightly modified distributions
′ ′ ′
σG,k (F ) = tind (F, WG ). The sequence (σG,1 , σG,2 , . . . ) is consistent. The random
graphs corresponding to this sequence of distributions are G(k, WG ). (We could
also generate this from G by selecting the k random nodes with replacement.) The
′
difference between σG,k and σG,k is very small if G is large: if we sample G(k, WG )
176 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

and keep it iff the sampled points correspond to different nodes of G, and otherwise
resample, then we get a sample from the distribution G(k, G). This shows that
( )
′ n(n − 1) . . . (n − k + 1) 1 k
(11.2) dvar (σG,k , σG,k ) ≤ 1 − < .
nk n 2
As discussed in the introduction, random graphs satisfy quite strong laws of
large numbers in the sense that two large random graphs are very much alike; this
translates to the fact that a sequence of independently generated random graphs
G(n, p) is convergent with probability 1. The next lemma shows that all local and
consistent random graph models have a similar property.

Lemma 11.8. Let (σ1 , σ2 , . . . ) be a local consistent random graph model, and gen-
erate a graph Gn from every σn , independently for different values of n. Then the
sequence (G1 , G2 , . . . ) is convergent with probability 1.

Proof. First we note that for every simple graph F on [k] and n ≥ k, we have
( )
(11.3) E tind (F, Gn ) = σk (F ).
Indeed, consider any injective map φ : V (F ) → V (Gn ). It follows from the
isomorphism invariance of σn that the probability that φ is an induced embedding
is the same for every map φ, so it suffices to compute this probability when φ
is the identity map on [k]. By the consistency of the model, this probability is
P(Gk = F ) = σk (F ).
Next we show that tind (F, Gn ) is concentrated around its expectation σk (F ).
We could compute second moments, but this would not give a sufficiently good
bound. So (sigh!) we compute the fourth moment.
Let S1 , S2 , S3 , S4 be independent random ordered k-subsets of [n] (we assume
that n > k 2 ). Define Xi = 1(Gn [Si ] = F ) − σk (F ). Note that E(Xi ) = 0 by (11.3),
even if we condition on the choice of the Si , since the distribution of Gn [S] is the
same for every ordered k-set S ⊆ [n]. Furthermore,
( )
(11.4) E(X1 X2 X3 X4 ) = E (tind (F, Gn ) − σk (F ))4 ,
since for a fixed Gn the variables Xi are independent, and E(Xi | Gn ) =
tind (F, Gn ) − σk (F ).
Let A denote the event that every Si meets at least one other Sj . The key
observation is that
E(X1 X2 X3 X4 | A) = 0.
This follows since if the Si are fixed so that (say) S4 does not meet the others, then
X4 is independent of {X1 , X2 , X3 }, and its expectation is 0. (This is where we use
the assumption that our random graph model is local!) Thus

E(X1 X2 X3 X4 ) ≤ E(X1 X2 X3 X4 | A)P(A) + E(X1 X2 X3 X4 | A)P(A)

7k 4
≤ P(A) ≤ .
n2
(The last inequality follows by elementary combinatorics.) Thus we get that
( ) 7k 4
E (tind (F, Gn ) − σk (F ))4 ≤ 2 ,
n
 11.2. RANDOM GRAPHS AS LIMIT OBJECTS 177

and hence by Markov’s Inequality

( )
(11.5) P(|tind (F, Gn ) − σk (F )| > ε) = P (tind (F, Gn ) − σk (F ))4 > ε4
1 ( ) 7k 4
≤ 4 E (tind (F, Gn ) − σk (F ))4 ≤ 4 2 .
ε ε n
If we sum (11.5) for n ≥ 1 with a fixed ε > 0, then the sum of the right hand sides
is convergent, so it follows by the Borel–Cantelli Lemma that with probability 1,
|tind (F, Gn ) − σk (F )| > ε holds for a finite number of values of n only, and so
tind (F, Gn ) → σk (F ) with probability 1. Hence the graph sequence Gn converges
with probability 1. 

With this lemma at hand, our theorem is easy to prove.

Proof of Theorem 11.7. First, consider a convergent graph sequence

(G1 , G2 , . . . ) and the probability distributions σk defined by it. Consistency and
′
locality follow by the consistency and locality of the distributions σk,G n
.
Second, consider consistent and local random graph model (σ1 , σ2 , . . . ), and a
sequence of random graphs Gn (n = 1, 2, . . . ), which are independently generated
from distribution σn for different indices n. It follows from Lemma 11.8 that with
probability 1, this graph sequence is convergent. Equation 11.3 implies that it
reproduces the right random graph model. 

11.2.2. Countable random graph models. We can arrange all labeled sim-
ple graphs in a locally finite rooted tree, where the empty graph is the root, and
F ′ is the parent of F . If (σ1 , σ2 , . . . ) is a consistent sequence of distributions, then
σk is a probability distribution on the k-th level of the tree, and the probability of
each node is the sum of probabilities of its children.
From this setup, we can combine all the distributions σk into a single probability
distribution on all infinite paths starting at the root. To be more precise, let Ω
denote the set of such paths, and let ΩF denote the set of paths passing through
the node F . Then the sets ΩF generate a sigma-algebra A on Ω. The Kolmogorov
Extension Theorem implies that there is a (unique) probability measure σ on (Ω, A)
such that σ(ΩF ) = σk (F ) for every F .
This is so far an abstract construction. We can, however, make explicit sense of
the elements of Ω. A path in the tree starting at the root is a sequence (F0 , F1 , . . . )
′
of graphs such that Fk = Fk+1 . Hence the path gives rise to the countable graph
F = ∪n Fn on the set of positive integers N∗ . Conversely, every graph on N∗
corresponds to a path in the tree starting at the origin.
Thus the points of Ω can be identified with the graphs on N∗ . The sets ΩF
are obtained by fixing adjacency between a finite number of nodes. Thus σ can be
thought of as a probability distribution on graphs on N∗ .
A countable random graph model is a probability distribution σ on (Ω, A),
invariant under permutations of N∗ . Such a random graph can also be considered
as a symmetric exchangeable array of 0-1 valued random variables (we will come
back to this way of looking at them in Section 11.3.3). The countable random
graph model is local if for any two finite disjoint subsets S1 , S2 ⊆ N∗ , the subgraphs
induced by S1 and S2 are independent (as random variables). The discussion above
shows that every consistent random graph model defines a countable random graph
model.
178 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

There is a way to go back from a countable random graph model σ to finite

consistent random graph models, which is even simpler: to get a random graph on
[n], generate a random graph G from σ, and take the induced subgraph G[n]. It is
easy to see that this random graph model is consistent. Furthermore, a consistent
random graph model is local if and only if the corresponding countable random
graph model is local. To sum up,
Proposition 11.9. There is a bijection between consistent random graph models
and countable random graph models. This bijection preserves locality.
It follows that local countable random graph models can serve as representa-
tions of the limit objects for convergent graph sequences.
Example 11.10 (Cliques and stable sets). Let G be either the complete graph
or the edgeless graph on N∗ , each with probability 1/2. Clearly G does not depend
on the ordering of N∗ , so it is a countable random graph model. The subgraph G[n]
is also complete with probability 1/2 and edgeless with probability 1/2 (at least
for n > 1). This defines a consistent random graph model. However, this model is
not local: the subgraph induced by {0, 1} is not independent (in the probabilistic
sense) from the subgraph induced by {2, 3}; to the contrary, they are always the
same. 
Example 11.11 (The Rado graph). We construct a random graph on N∗ by
connecting a pair of distinct integers i, j ∈ N∗ with probability 1/2, independently
for different pairs. The resulting random graph G(N∗ , 1/2), called the Rado graph,
has many interesting properties (some of these are stated in the exercises at the
end of the section), but right now, what is important for us is that it is local by
construction. The corresponding local and consistent random graph model is the
ordinary random graph G(n, 1/2). 
Example 11.12 (Infinite W -random graph). We can extend the definition of
W -random graphs (Section 10.1) to get a countable random graph G(N∗ , W ). Given
a graphon W , we select a sequence of independent random points (X1 , X2 , . . . )
from [0, 1], and connect i and j (i, j ∈ N∗ ) with probability W (Xi , Xj ). This
construction generalizes the Rado graph. It is immediate that the distribution of
the infinite W -random graph is invariant under permutations of N∗ , and it is also
local: for two disjoint sets S, T ⊆ N∗ , the graph we construct on S is independent
(in the probability sense) from the graph we construct on T . We will see that all
local countable random graph models can be constructed this way. 
Example 11.13 (Triangle-free random graph model). It is easy to construct
a countable triangle-free graph that contains every finite triangle-free graph as an
induced subgraph (Exercise 11.17. In a theory developed independently from ours,
Petrov and Vershik [2010] prove that there exist a local countable random graph
model that is triangle-free with probability 1. They also construct an appropriate
graphon, which turns out 0-1 valued. 
We conclude this section with a construction of a convergent graph sequence
from a countable random graph model (without the locality assumption) given
independently by Lovász and Szegedy [2012a] and Diaconis and Janson ([2008],
Theorem 5.3).
 11.2. RANDOM GRAPHS AS LIMIT OBJECTS 179

Proposition 11.14. Let G be a random graph on N∗ drawn from a countable ran-

dom graph model. Let G[n] denote the subgraph induced by [n]. Then the sequence
(G[1], G[2], . . . ) is convergent with probability 1.

Proof. For every fixed simple graph F , the sequence (tinj (F, G[n])( : n =
1, 2, . . . ) )is a reverse martingale for n ≥ v(F ) in the sense that E tinj (F, G[n −
1]) | G[n] = tinj (F, G[n]) (this follows by the simple averaging principle (5.27)). By
the Reverse Martingale Convergence Theorem A.17, it follows that this sequence
is convergent with probability 1. Hence with probability 1, (tinj (F, G[n]) : n =
1, 2, . . . ) is convergent for every F . 

This last proposition may sound similar to the construction in Lemma 11.8, but
there is a significant difference: in this construction, locality is not needed. Unlike
in Lemma 11.8, G[n] and G[m] are not independently generated. If we apply the
construction in Lemma 11.8 twice, and then pick the even-indexed graphs from
one sequence and interlace them with the odd-indexed graphs from the other, we
get a sequence that is constructed in the same way, and so it is convergent with
probability 1. This means that almost all sequences generated by Lemma 11.8
(for a fixed consistent and local random graph model) have the same limit. In
contrast to this, running the construction in Proposition 11.14 twice we could not
necessarily interlace the resulting sequences into a single convergent sequence: in
Example 11.10, we get a sequence of growing cliques with probability 1/2 and a
sequence of growing edgeless graphs with probability 1/2. Both of these sequences
are convergent, but they don’t have the same limit. Sequences constructed from
one and the same countable random graph model are almost always convergent,
but they may converge to different limits.
In view of Examples 11.6 and 11.12, we also get:

Corollary 11.15. For every graphon W , generating a W -random graph G(n, W )

for n = 1, 2, . . . we get a convergent sequence with probability 1, whose limiting
countable random graph model is G(N∗ , W ).

Exercise 11.16. (a) Prove that the Rado graph almost surely has the extension
property: for any two disjoint finite subsets S, T ⊆ N∗ there is a node connected
to all nodes in S but to no node in T .
(b) Prove that every countable graph with the extension property is isomorphic
to the Rado graph.
(c) Prove that if you generate two Rado graphs independently, they will be iso-
morphic with probability 1.
Exercise 11.17. Construct a universal triangle-free graph: a countable graph
containing every finite triangle-free graph as an induced subgraph.
Exercise 11.18. (a) We can define a random graph G(N∗ , p) for all 0 < p < 1.
Prove that with probability 1, this random graph will be isomorphic to the Rado
graph for any p.
(b) More generally, if W is a graphon with 0 < W (x, y) < 1 for all x, y ∈ [0, 1],
then G(n, W ) is almost always isomorphic to the Rado graph.
(c) Construct a graphon W such that two independent countable W -random
graphs are almost surely non-isomorphic [Gábor Kun].
Exercise 11.19. Show that without the assumption of locality, Lemma 11.8 does
not remain valid.
180 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Exercise 11.20. Prove that if we generate two sequences as in Proposition 11.14

from a local countable random graph model, then interlacing the two sequences,
we get a convergent sequence.

11.3. The limit graphon

In this section we give a more explicit description of limit objects for convergent
graph sequences: we show that graphons (up to weak isomorphism) are precisely
the structures that are needed.
11.3.1. Existence. Let (Gn ) be a convergent graph sequence, so that the
densities t(F, Gn ) tend to a limit t(F ) for every finite simple graph F . We know
that a limit object can be described as a consistent and local random graph model
(finite or countable). The main motivation behind introducing graphons is that they
provide a much more explicit representation for this limit object, as the following
theorem shows (Lovász and Szegedy [2006]).
Theorem 11.21. For any convergent sequence (Gn ) of simple graphs there exists
a graphon W such that t(F, Gn ) → t(F, W ) for every simple graph F .
We say that this graphon W is the limit of the graph sequence, and write
Gn → W .
The reader might wonder if one really needs complicated objects like integrable
functions to describe limits of graph sequences; would perhaps piecewise linear, or
monotone, or continuous functions suffice? It turns out that (up to weak isomor-
phism) all measurable functions are needed: Every graphon W can be obtained as
the limit of a convergent sequence of simple graphs; this follows by Corollary 11.15.
There are three quite different ways to prove Theorem 11.21. The original
one by Lovász and Szegedy [2006] uses Szemerédi partitions and the Martingale
Convergence Theorem. This is very closely related to the proof of the compactness
of the graphon space (Theorem 9.23). Below we will use compactness to prove
Theorem 11.21, but we could as well go the other way, since Theorem 11.21 easily
implies the compactness of the graphon space (Exercise 11.28).
A more recent proof by Elek and Szegedy [2012] constructs a different limit
object first, in the form of a graph on a very large sigma-algebra, by taking an
ultraproduct; then obtains the graphon as an appropriate projection of this. This
proof technique is quite general, it extends to hypergraphs and many other struc-
tures.
As a third route to prove Theorem 11.21, it was shown by Diaconis and Jan-
son [2008] that it can be derived from results of Aldous [1981] and Hoover [1979]
on exchangeable random variables, pointing out a basic connection to probability
theory. We will sketch these alternative proofs in the next two sections.
Proof of Theorem 11.21. By Theorem 9.23, the metric space (W f0 , δ ) is
compact, and hence the sequence (Wn = WGn : n = 1, 2, . . . ) has a convergent
f0 . By the Counting Lemma
subsequence (Wnj : j = 1, 2, . . . ) with limit W ∈ W
10.23, we have for every simple graph F
|t(F, Wnj ) − t(F, W )| ≤ e(F ) δ (Wnj , W ) −→ 0 (j → ∞),
and so t(F, Wnj ) = t(F, Gnj ) → t(F, W ). Since (t(F, Gn ) : n = 1, 2, . . . ) is a
Cauchy sequence, this implies that t(F, Gn ) → t(F, W ) for every simple graph
F. 
 11.3. THE LIMIT GRAPHON 181

There may be several graphons W representing the limit of a convergent graph

sequence. Of course, we may change the value of W on a set of measure 0; but
more generally, we can replace W by any other kernel weakly isomorphic with W .
Conversely, of W and W ′ both represent the limit of a convergent graph sequence,
then they are weakly isomorphic by the definition of weak isomorphism. (Weak
isomorphism has been discussed in Sections 7.3 and 10.7, and we will say more
about it in Section 13.2.)
Convergence to the limit object can also be characterized by the distance func-
tion:
Theorem 11.22. For a sequence (Gn ) of graphs with v(Gn ) → ∞ and graphon
W , we have Gn → W if and only if δ (WGn , W ) → 0.
Proof. If δ (WGn , W ) → 0, then Gn → W follows by the Counting Lemma
just like in the proof of Theorem 11.21 above. Conversely, suppose that Gn → W ,
so t(F, Gn ) → t(F, W ) for every simple graph F , and hence for every fixed k, the
Inverse Counting Lemma 10.32 implies that
20
δ (WGn , W ) ≤ √
log k
if n is large enough. Hence δ (WGn , W ) → 0 as claimed.
(This proof gives an explicit connection between the rates of convergence in
Gn → W and δ (WGn , W ) → 0. If we don’t care about this, the theorem follows by
abstract arguments: the space (W f0 , δ ) is compact, the map W 7→ (t(F, W ) : F ∈
F) is continuous and injective, hence its inverse is also continuous.) 

11.3.2. Ultralimit and limit. The theory of ultraproducts and ultralimits

provides a general way to construct limit objects. This proof technique has some
drawbacks: it is non-constructive, and requires advanced special techniques from
model theory (outlined in Appendix A.5). On the other hand, its advantage is
that it is very flexible: one can define the limit of virtually any kind of sequence
of structures this way, and only have to wonder later whether this limit object can
be “brought down to earth” to have combinatorial (or algebraic, or arithmetic)
significance.
Let ω be an ultrafilter on N∗ . Recall that we call the sets in ω “Large”, the
other subsets of N∗ , “Small”. As a first ∏ try, let us define the limit of a graph
sequence (G1 , G2 , . . . ) as the ultraproduct ω Gn .
This assigns a limit object to every graph sequence, not just to those that
are convergent. Unfortunately, this ultraproduct will depend on the ultrafilter we
use, even for convergent graph sequences (so the situation is not as simple as for
numerical sequences, see Exercise 11.31). For example, let Hn be the edgeless graph √
on n nodes if n is even,
√ and let it be a graph consisting of a clique of size ⌊ n⌋
together with n − ⌊ n⌋ isolated nodes ∏ if n is odd. This graph sequence is clearly
convergent. On the other hand, ω Gn has no edge if the set of odd integers is
Small, and it consists of a clique of continuum size and continuum many isolated
nodes if the set of odd integers is Large (and either can happen).
While it is of the same cardinality as the whole ultraproduct, this clique should
occupy a negligible part of this huge product graph. To make this precise, we have
to introduce a measure on the product. Let (G1 , G2 , . . . ) be a sequence of finite
graphs. For every graph Gn = (Vn , En ), we can consider the (finite) sigma-algebra
182 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

An = 2Vn of all subsets of Vn , and

∏ the uniform probability measure πn on Vn . Then
the ultraproduct G = (V, E)∏ = ω Gn is a graph whose node set V∏ is also equipped
with a sigma-algebra A = ω An and a probability measure π = ω πn on ∏ it.
This answers our concern with the example above: the ultraproduct ω Hn
becomes edgeless if we delete an appropriate set of nodes of measure zero.
However, the construction is still not satisfactory. First, the sigma-algebra A
is an ugly one: for example, it is not separable in general. Second, the set E of
edges of the product, as a subset of V × V , may not be measurable with respect to
the product sigma-algebra A × A. This is really serious: we would like to be able
to assert, for example, that if the graphs Gn have edge density 1/2, then the limit
has edge density 1/2, which means that (π × π)(E) = 1/2. However, this assertion
does not even make sense!
The way out of this trouble is to do∏some fiddling with the ultraproduct. Let
us start with taking the ultraproduct ω (Vn × Vn ). This ultraproduct can be
identified with V × V in a natural way: if an , bn ∈ Vn , then
(11.6) [(a1 , b1 ), (a2 , b2 ), . . . ] ↔ ([a1 , a2 , . . . ], [b1 , b2 , . . . ]).
It is easy to see that if the bracket on the left side is represented by another
equivalent sequence of pairs, then the brackets on the right don’t change, and vice
versa. ∏
As before, sets of the form ω Sn (Sn ⊆ Vn × Vn ) form a set algebra on V × V ,
which generates a sigma-algebra A′ . In particular, ∏ we can take the ultraproduct
of the edge sets of the graphs Gn , to get E = ω En ∈ A′ . Furthermore, the
ultraproduct η of the uniform measures on Vn × Vn is a probability measure on
(V × V, A′ ).
∏ It is easy to see that the sigma-algebra A × A is generated
′
by sets of the form
ω (S n × Tn ), where Sn , T n ⊆ Vn . Hence A × A ⊆ A . In general, we don’t have
equality (this is why E is not measurable in A×A). But the following lemma shows
that in an important way, A′ is not too much larger than A × A:
Lemma 11.23. For every set B ∈ A′ and every x ∈ V , the neighborhood B(x) =
{y ∈ V : (x, y) ∈ B} belongs to A.
∏
Proof. First, we consider
∏ the case when B = ω Bn , and let x = [x1 , x2 , . . . ].
We claim that B(x)∏= ω Bn (xn ) (which of course implies that B(x) ∈ A). Indeed:
y = [y1 , y2 , . . . ] ∈ ω Bn (xn ) if and only if yn ∈ Bn (xn ) for a large set of indices
n, if and only if (xn , yn ) ∈ Bn for a large set of indices n, if and only if [x, y] ∈ B.
Now taking complements and countable intersections of sets B ∈ A′ corre-
sponds to carrying out the same operations on the sets B(x) (where x ∈ V is
fixed), so these sets stay in A. 

The measure µ, when restricted to A × A, gives the probability measure π × π.

Hence we can take the conditional expectation W = E(1E | A × A), which is a
function on V × V , measurable with respect to A × A, and has the property that
∫
(11.7) W (x, y) dπ(x) dπ(y) = µ(X ∩ E)
X

for every set X ∈ A × A. This results in a probability space (V, A, π) with a

symmetric measurable function W : V × V → [0, 1].
 11.3. THE LIMIT GRAPHON 183

We show that W can serve as the limit graphon (at least as long as we ignore
the ugliness of the underlying sigma-algebra):
Proposition 11.24. For every simple graph F and any sequence (Gn : n =
1, 2, . . . ) of graphs,
lim t(F, Gn ) = t(F, W ).
ω

If the graph sequence (G1 , G2 , . . . ) is convergent, then the ultralimit on the left
side is equal to limn→∞ t(F, Gn ), independently from ω.
Proof. Let V (F ) = [k]. As a first step, we express the left hand side in
the ultraproduct space. We have to introduce more sigma-algebras for ∏ this. For
every set U ⊆ [k], we take the set VnU of all maps∏U → Vn . The set ω VnU can
be identified with V U , just like (11.6) identifies ω Vn × Vn with V × V . The
ultraproduct of the Boolean algebra of all subsets of VnU gives a sigma-algebra AU
on V U . The ultraproduct of the uniform measures on the sets VnU gives a measure
τU on the product sigma-algebra on V U . We abbreviate τ[k] by τk .
The∏set Hom(F, Gn ) of homomorphisms ∏ from F into Gn is ∏ a subset of Vnk ,
k k
and so ω Hom(F, Gn ) is a subset of ω Vn = V . The set ω Hom(F, Gn )
can be identified with the set Hom(F, G) ⊆ V k of homomorphisms of F into G.
Furthermore,
(∏ ) ( )
lim t(F, Gn ) = τk Hom(F, Gn ) = τk Hom(F, G)
ω
ω
by the definition of the ultraproduct of measures. So it suffices to prove that
( )
(11.8) τk Hom(F, G) = t(F, W ).
Let (X1 , . . . , Xk ) ∈ V k be a random node chosen from the distribution τk . Then
we can rephrase the equality to be proved as
( ∏ ) ( ∏ )
(11.9) E 1E (Xi , Xj ) = E W (Xi , Xj ) .
ij∈E(F ) ij∈E(F )

(It is easy to check that the functions 1E (Xi , Xj ) are measurable with respect to
the sigma-algebra A[k] .)
If the random variables 1E (Xi , Xj ) were independent, we could take the ex-
pectation factor-by-factor, and we would be done. But of course they are not. The
trick in the proof is to replace the factors 1E (Xi , Xj ) by W (Xi , Xj ) one by one.
Consider any edge uv of F ; we show that
( ∏ ) ( ∏ )
(11.10) E 1E (Xi , Xj ) = E 1E (Xi , Xj )W (Xu , Xv )) .
ij∈E(F ) ij̸=uv

This will show that we can replace 1E (Xu , Xv ) by W (Xu , Xv ) without changing
the expectation, and repeating a similar argument for all edges of F , we get (11.9).
For notational convenience, assume that u = 1 and v = 2. The main difficulty
in the rest of the argument is to be careful about measurability, because we have
several sigma-algebras floating around. Using Lemma 11.23, it is not hard to argue
that fixing X1 and X2 , the functions 1E (Xi , Xj ) are measurable with respect to
A[k] , and so the expectation
∏
f (X1 , X2 ) = EX3 ,...,Xk 1E (Xi , Xj )
ij∈E(F ){i,j}̸={1,2}
184 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

is well defined. Furthermore, again by Lemma 11.23, if we fix X3 , . . . , Xk , then every

function 1E (Xi , Xj ) ({i, j} ̸= {1, 2}) becomes either constant, or A{1} -measurable
(if the edge ij is incident with 1) or A{2} -measurable (if ij is incident with 2).
Hence it follows that f (x1 , x2 ) is A{1} × A{2} -measurable.
By the definition of W , this implies that
∫ ∫
f (x1 , x2 )1E (x1 , x2 ) dτ{1,2} (x1 , x2 ) = f (x1 , x2 )W (x1 , x2 ) dπ(x1 ) dπ(x2 ),
V ×V V ×V

which proves (11.10). 

To finish the construction of the limit graphon, one has to map this big sigma-
algebra (V, A) onto [0, 1], define the appropriate image of W , and show that it
represents the same subgraph densities. We refer to the paper of Elek and Szegedy
[2012] for these details.
Remark 11.25. If you compare this construction of the limit object with the
construction given before, some parallelism between their elements is apparent.
For example, (11.7), applied to a generator set X = S × T , asserts that the density
of edges of G between S and T is the same as the integral of W on S × T , so W and
1E have distance 0 in the cut norm. Perhaps further exploration of this parallelism
could shed some light on the nature of the use of non-constructive infinite methods
like ultraproducts in the theory of (very large, but) finite graphs.
In the last proof, one needs to handle various sigma-algebras; indeed, a little
“calculus of sigma-algebras” was used. Elek and Szegedy push this much further,
and develop more from this combinatorial theory of sigma-algebras, which enables
them to extend this construction to hypergraphs; cf. Section 23.3.
11.3.3. Exchangeable random variables. Let (G1 , G2 , . . . ) be a conver-
gent graph sequence. We start with the weak limit of the sequence in the form of
a local countable random graph model σ. Let G be a graph from this distribution, ( )
and let (Xij )∞i,j=1 be its adjacency matrix; in other words, Xij = 1(ij ∈ E G) .
It follows from the invariance of σ under permutations that the random vari-
ables Xij have the same distribution: P(Xij = 1) = limn→∞ t(K2 , Gn ). They are
not independent, but have the following property, which is called symmetrically ex-
changeable: if α is a permutation of N∗ , then for every k ≥ 1, the joint distribution of
(Xij : 1 ≤ i, j ≤ n) is the same as the joint distribution of (Xα(i)α(j) : 1 ≤ i, j ≤ n).
This is of course just a reformulation of the condition that σ is invariant.
Now symmetrically exchangeable random variables have a representation theo-
rem due to Aldous [1981] and Hoover [1979]: every such system can be represented
as a mixture of random variables of the form Xij = W (Yi , Yj ), where Yi (i ∈ N∗ )
are independent random variables with values in [0, 1] and W : [0, 1]2 → [0, 1] is
a symmetric measurable function in two variables. Furthermore, as Diaconis and
Janson [2008] show, if the random graph G is local, then you don’t get a real mix-
ture: local exchangeable distributions are precisely the extreme points of the space
of symmetrically exchangeable random variables, in other words, random variables
of the form Xij = W (Yi , Yj ). So the representation theorem of Aldous and Hoover
provides the limit graphon W directly!
For the details of this theory and its application here, we refer to the monograph
of Kallenberg [2005] and the paper [2008]. See also Austin [2008] for a survey of
the many other applications of this approach.
 11.4. PROVING CONVERGENCE 185

Exercise 11.26. Prove that if a graph sequence (Gn ) satisfies v(Gn ) → ∞, then
it is Cauchy in the cut distance if and only if it is Cauchy in the sampling distance.
Exercise 11.27. Prove the following facts: (a) For every stepfunction W ,
δ1 (W, H(n, W )) → 0 as n → ∞ with probability 1. (b) For every graphon W ,
δ1 (W, H(n, W )) → 0 as n → ∞ with probability 1. (c) For every graphon W ,
δ (G(n, W ), H(n, W )) → 0 as n → ∞ with probability 1. (d) If U and W are
weakly isomorphic graphons, then G(n, U ) and G(n, W ) have the same distri-
bution. (e) If U and W are weakly isomorphic graphons, then δ (U, W ) = 0
(A. Schrijver).
Exercise 11.28. Show that Theorems 11.21, 11.22 and Corollary 11.15 imply
that the space (Wf0 , δ ) is compact.
Exercise 11.29. Prove that the following properties of graphs are inherited to
their ultraproduct: (a) 3-regular; (b) all degrees bounded by 10; (c) triangle-free;
(d) containing a triangle; (e) bipartite; (f) disconnected.
Exercise 11.30. Prove that the following properties of graphs are not inherited
to their ultraproduct: (a) connected; (b) all degrees are even; (c) non-bipartite.
Exercise 11.31. (a) Prove that every bounded sequence of real numbers has a
unique ultralimit. (b) Prove that the limω (ai + bi ) = limω ai + limω bi . (c) Prove
that the ultralimit limω ai is independent of the choice of the ultrafilter ω if and
only if the sequence is convergent in the classical sense.

11.4. Proving convergence

It is not always easy to show that a certain graph sequence is convergent. We
have several characterizations (in terms of subgraph densities, sample distribution,
cut distance) and sometimes one, sometimes the other condition is easier to apply.
First we develop some useful sufficient conditions for convergence, and then apply
these to give a number of examples of interesting convergent graph sequences.

11.4.1. Convergence of sampling methods. We start with a supplement

to Corollary 11.15:
Proposition 11.32. For every graphon W , generating a W -random graph G(n, W )
for n = 1, 2, . . . we get a graph sequence such that G(n, W ) → W with probability
1.
Proof. It is straightforward to check that for every simple graph F and n ≥
v(F ),
E(tinj (F, G(n, W )) = t(f, W ).
Since tinj (F, G(n, W ) is highly concentrated around its expectation (by Theorem
10.2 or by the proof of Lemma 11.8), it follows that tinj (F, G(n, W ) → t(f, W ) with
probability 1. 

Sometimes we need that other sequences constructed by similar, but different

sampling from a graphon are convergent. We describe one lemma of this type by
Borgs, Chayes, Lovász, Sós and Vesztergombi [2008] [2011].
We start with a definition. For every n ≥ 1, let Sn ⊆ [0, 1] be a finite set such
that |Sn | → ∞. We say that the sequence (Sn ) is well distributed , if |Sn ∩J|/|Sn | →
λ(J) for every interval J as n → ∞. Equivalently, the uniform measure on Sn
converges weakly to the uniform measure on [0, 1] (see Billingsley [1999] for more
on related notions like uniform distribution of sequences).
186 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Lemma 11.33. Let W ∈ W0 be almost everywhere continuous, and let Sn be a

well distributed sequence of sets. Then G(Sn , W ) → W with probability 1.
It is clear that such a conclusion cannot hold without some assumption on W ,
since a general measurable function could be changed on the sets Sn ×Sn arbitrarily
without changing its subgraph densities. For an extension to graphons on general
metric spaces, see Exercise 13.8. Note that there need not be any randomness in
the sequence Sn .
Proof. Consider a partition {Z1 , . . . , Zm } of [0, 1] into m intervals of equal
length. Since (Sn ) is well distributed, we have 1/(m+1) ≤ |Sn ∩Zj |/|Sn | ≤ 1/(m−1)
for every j if n ≥ n0 (m). Given any n that is large enough, we choose the largest
m for which n ≥ n0 (m), and partition each set Zj into |Sn ∩ Zj | sets of equal
measure, each containing exactly one point of Sn ∩ Zj , to get the partition Qn .
This partition has the properties that |Qn | = |Sn |, every partition class contains
exactly
{ one point of Sn , the maximum diameter of partition classes tends to 0, and
max λ(Q)|Sn | − 1 : Q ∈ Qn } → 0.
For s ∈ Sn , let Qns be the partition class of Qn containing s. Define the function
Wn as follows: for s, s′ ∈ Sn and (x, y) ∈ Qns × Qns′ , let Wn (x, y) = W (s, s′ ). Then
Wn (x, y) → W (x, y) in every point (x, y) where W is continuous, in particular
Wn → W almost everywhere. This implies that
(11.11) ∥Wn − W ∥1 −→ 0 (n → ∞).
We can view Wn as the graphon WHn associated with a weighted graph Hn
with V (Hn ) = Sn , where the weight of node s ∈ Sn is λ(Qns ), and the weight of
edge ss′ (s, s′ ∈ S) is W (s, s′ ). Note that Hn is almost the same weighted graph as
Hn = H(Sn , W ): they are defined on the same set of nodes, the edges have the same
weights, and the nodeweight λ(Qns ) is asymptotically 1/|Sn | by the construction of
Qn . Given ε > 0, we have |λ(Qns ) − 1/|Sn || < ε/|Sn | if n is large enough. Hence
there is a measure preserving bijection φ : [0, 1] → [0, 1] and a set R ⊆ [0, 1] of
measure ε such that
WHn (x, y) = WHφn (x, y) (x, y ∈
/ R).
This implies that
(11.12) δ1 (Hn , Hn ) −→ 0 (n → ∞).
Formulas (11.11) and (11.12) imply that H(Sn , W ) → W , which in turn implies
that G(Sn , W ) → W with probability 1 (cf. Lemma 10.11). 
Corollary 11.34. Let W be a graphon that is almost everywhere continuous. Then
G(Sn , W ) → W with probability 1, where
(a) Sn ⊆ [0, 1] is obtained by selecting a uniform random point from every
interval [j/n, (j + 1)/n], n = 0, . . . , n − 1;
(b) Sn = {1/n, 2/n, . . . , n/n}.
11.4.2. Examples: Convergent graph sequences. We discuss a variety of
examples of convergent graph sequences. It turns out that it is not always easy to
prove that these are convergent and determine the limit graphon. In some examples,
the limit can be guessed and the proof of convergence in the cut distance is easy.
In other examples, it is quite tricky to guess the limit graphon. In other cases,
the convergence of subgraph densities can be proved. We will also see a randomly
 11.4. PROVING CONVERGENCE 187

growing graph sequence which is convergent with probability 1, but if we run it

again, it converges to a different limit!
We start with two easy examples.
Example 11.35 (Complete bipartite graphs). It is natural to guess, and easy
to prove, that complete bipartite graphs Kn,n converge to the graphon W (x, y) =
1(0 ≤ x ≤ 1/2 ≤ y ≤ 1) + 1(0 ≤ y ≤ 1/2 ≤ x ≤ 1). 
Example 11.36 (Simple threshold graphs). These graphs are defined on the
set {1, . . . , n} by connecting i and j if and only if i + j ≤ n. These graphs converge
to the graphon defined by 1(x + y ≤ 1), which we call the simple threshold graphon.


Figure 11.1. Simple threshold graphs and their limits

A more interesting example is the following.

Example 11.37 (Quasirandom graphs). A sequence of graphs tending to the
identically-p function is exactly what we called a “quasirandom sequence” with
density p (by the second property in Section 1.4.2). In particular, the Paley graphs
(Example 1.1) converge to the graphon W ≡ 1/2. 
Example 11.38 (Multitype quasirandom graphs). Generalizing the previous
example, we consider a multitype quasirandom graph sequence (Gn ) with a template
graph H. This means that (assuming that V (H) = [q] and V (Gn ) = [n]) V (Gn ) has
a partition (V1 , . . . , Vq ) such that |Vi | = αi (H)n + o(n) and for every fixed i, j ∈ [q],
the bipartite graphs Gn [Vi , Vj ] form a quasirandom bipartite graph sequence with
edge density βij (H) (in the case when i = j, the induced subgraphs Gn [Vi ] form a
quasirandom graph sequence).
A multitype quasirandom graph sequence with template graph H tends to
the graphon WH , and vice versa. The first statement follows easily, since
δ (WGn , WH ) → 0 by the definition of a multitype quasirandom sequence. The
converse is less trivial; if δ (WGn , WH ) → 0, then there is a way to label the nodes
of Gn so that ∥WGn − WH ∥ → 0 (this follows from Theorem 11.59 below), and
then we can partition the nodes of Gn by putting node u in class i (u ∈ [n], i ∈ [q])
iff α1 (H) + · · · + αi−1 (H) ≤ u/n < α1 (H) + · · · + αi (H). It is not hard to verify
that this partition has the right properties to guarantee that the sequence (Gn ) is
multitype quasirandom with template H. 
We continue with several examples, given by Borgs, Chayes, Lovász, Sós and
Vesztergombi [2011], of convergent graph sequences obtained by random growing
processes.
188 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Example 11.39 (Growing uniform attachment graphs). We generate a ran-

domly growing graph sequence Gua n as follows. We start with a single node. At
the n-th iteration, a new node is born, and then every pair of nonadjacent nodes is
connected with probability 1/n. We call this graph sequence a uniform attachment
graph sequence; see Figure 1.8.
Let us do some simple calculations. After n steps, let {0, 1, . . . , n − 1} be the
nodes (born in this order). The probability that nodes i < j are not connected
j
is j+1 · j+1
j+2 · · · n
n−1
= nj . These events are independent for all pairs (i, j). The
expected degree of j is
∑
j−1
n−j ∑
n−1
n−i n − 1 j(j − 1)
+ = − .
i=0
n i=j+1
n 2 2n

The expected number of edges is

n−1 ( )
1 ∑ n − 1 j(j − 1) n2 − 1
− = .
2 j=0 2 2n 6

To figure out the limit graphon, note that the probability that nodes i and j
are connected is 1 − max(i, j)/n. If i = xn and j = yn, then this is 1 − max(x, y).
This motivates the following:

n tends to the limit function 1 − max(x, y)

Proposition 11.40. The sequence Gua
with probability 1.
Proof. For a fixed n, the events that nodes i and j are connected are in-
dependent for different
( i, j, and so) by the computation above, Gua n has the same
distribution as G Sn , 1 − max(x, y) , where Sn = {0, 1/n, . . . , (n − 1)/n}. It is easy
to see that this sequence is well distributed in [0, 1], and so the Proposition follows
by Lemma 11.33. 

One can get a good explicit bound on the convergence rate by estimating the
cut-distance of WGua
n
and 1 − max(x, y), using the Chernoff-Hoeffding bound. 
Example 11.41 (Prefix attachment graphs). In this construction, it will be
more convenient to label the nodes starting with 1. At the n-th iteration, a new
node n is born, a node z is selected at random, and node n is connected to nodes
1, . . . , z − 1. We denote the n-th graph in the sequence by Gpfx
n , and call this graph
sequence a prefix attachment graph sequence (Figure 11.2).
Again we start with some simple calculations. The probability that nodes
i < j are connected is j−i j (but these events are not independent in this case!). The
expected degree of j is therefore
∑
j−1
j−i ∑n
i−j j n
+ = n − + j ln + o(n).
i=1
j i=j+1
i 2 j

The expected number of edges is n(n − 1)/4.

Looking at the picture, it seems that it tends to some function, which we can try
to figure out similarly as in the case of uniform attachment graphs. The probability
that i and j are connected can be written in a symmetric form as |j − i|/ max(i, j).
If i = xn and j = yn, then this is |x − y|/ max(x, y).
 11.4. PROVING CONVERGENCE 189

Figure 11.2. A randomly grown prefix attachment graph with

100 nodes, and the same graph with nodes ordered by their degrees.

Does this mean that the graphon U (x, y) = |x − y|/ max(x, y) is the limit?
Somewhat surprisingly, the answer is negative, which we can see by computing
triangle
( )(densities.
) The probability that three nodes i < j < k form a triangle is
1 − kj 1 − ji (since if k is connected to j, then it is also connected to i). Hence
the expected number of triangles is
∑ ( j
)(
i
) ( )
1 n
1− 1− = .
k j 6 3
i<j<k

Hence ( )
1 n 1
t(K3 , Gn ) = −→ .
n3 3 6
On the other hand,
∫
|x − y| |x − z| |y − z|
t(K3 , U ) = · · dx dy dz.
max(x, y) max(x, z) max(y, z)
[0,1]3

Since the integrand is independent of the order of the variables, we can compute
this easily:
∫ ( )
x ( x) ( y) 5
t(K3 , U ) = 6 1− 1− 1− dx dy dz = .
y z z 36
0≤x<y<z≤1

So U is not the limit of the sequence Gpfxn .

Perhaps ordering the nodes by degrees helps? The second pixel picture in Figure
11.2 suggests that after this reordering, the functions WGpfx
n
converge to some other
continuous function. But again this convergence is only in the weak* topology, not
in the δ distance. We’ll see that no continuous function can represent the “right”
limit: the limit graphon is 0-1 valued, and it is uniquely determined up to measure
preserving transformations by Theorem 13.10, which do not change this property.
Is this graph sequence convergent at all? Our computation of the triangle
densities above can be extended to computing the density of any subgraph, and it
follows that the sequence of densities t(F, Gpfx
n ) is convergent for every n. How to
figure out the limit?
190 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Let us label a node born in step k, connected to {1, . . . , m}, by (k/n, m/k) ∈
[0, 1] × [0, 1]. Then we can observe that nodes with label (x1 , y1 ) and (x2 , y2 ) are
connected if and only if either x1 < x2 y2 or x2 < x1 y1 .
This suggests a description of the limit in the following form: Consider the
function W pfx : [0, 1]2 × [0, 1]2 → [0, 1], given by
( )
W pfx (x1 , y1 ), (x2 , y2 ) = 1(x1 < x2 y2 or x2 < x1 y1 ).
(As remarked before, we can consider 2-variable functions on other probability
spaces, not just [0, 1]; in this case, [0, 1]2 is a more convenient representation. In
the proof below, we use an analogue of Lemma 11.33, adopted to this case. For a
general statement containing both, see Exercise 13.8.)
Proposition 11.42. The prefix attachment graphs Gpfx
n tend to W
pfx
almost surely.
Proof sketch. Let Sn be the (random) set of points in [0, 1]2 of the form
(i/n, zi /i) where i = 1, . . . , n and zi is a uniformly chosen random integer in [i].
Then Gpfx n = G(Sn , W ) = H(Sn , W pfx ).
pfx

Furthermore, with probability 1, the sets Sn are well distributed in [0, 1]2 in
the sense that |Sn ∩ A|/|Sn | → λ(A) for every open set A. It suffices to verify
this for the case when A = J1 × J2 , where J1 , J2 are open intervals, and it will
be also convenient to assume that J1 does not start at 0. The assertion is then
easily verified, based on the fact that the first coordinates (i/n) are well distributed
in [0, 1], and the second coordinates are uniformly distributed random points in
{1/i, . . . , i/i}. Thus the generalized version of Lemma 11.33 applies and proves the
Proposition. 

Figure 11.3. The limit of randomly grown prefix attachment

graphs as a function on [0, 1]2 .

Proposition 11.42 gives a nice and simple representation of the limit object with
the underlying probability space [0, 1]2 (with the uniform measure). If we want a
2
representation by a graphon on [0, 1], we can ( map [0, 1]
) into [0, 1] by a measure
φ pfx
preserving map φ; then Wpfx (x, y) = W φ(x), φ(y) gives a representation of
the same graphon as a 2-variable function. For example, using the map φ that
separates even and odd bits of x, we get the fractal-like picture in Figure 11.3.
It is interesting to note that the graphs G(n, W ) form another (different) se-
quence of random graphs tending to the same limit W with probability 1.
 11.4. PROVING CONVERGENCE 191

A final remark on this graph sequence. It is not hard to verify that for the
graphon U (x, y) = |x − y|/ max(x, y), we have
∫
(11.13) (WGpfx
n
− U ) −→ 0
S×T

for every S, T ⊆ [0, 1]. (Indeed, it is enough to prove this for sets S, T from a
generating set of the sigma-algebra of Borel sets, e.g. rational intervals. Since
there is only a countable number of these intervals, it suffices to prove that (11.13)
holds with probability 1 for any two rational intervals S and T . This is a rather
straightforward computation in probability.)
So WGpfx
n
→ U in the weak* topology of L∞ ([0, 1]2 ), but not in our sense. We
will see (Lemma 8.22) that our convergence implies weak* convergence, but not
the other way around. This example also shows that had we defined convergence
of a graph sequence by weak* convergence (after appropriate relabeling), the limit
would not be unique. The uniqueness of the limit graphon is a nontrivial fact! 
So far, our randomly grown sequences tended to well-defined limit graphons
with probability 1. Now we turn to examples of randomly grown sequences that are
convergent with probability 1, but if we run the process again, they may converge
to a different limit. There is in fact a very simple sequence with this property.
Example 11.43 (Cloning). Given a simple graph G0 , we select a uniform random
node v, and create a twin of v (a new node v ′ connected to the same nodes as v; v and
v ′ are not connected). Repeating this we get a sequence of graphs G0 , G1 , G2 . . . .
We claim that this sequence is convergent with(probability 1. Let ) v(G0 ) = k.
Note that each Gn is determined by the sequence ni : i ∈ V (G0∑ ) , where ni is
the number of clones of i we created, including i itself. Clearly i ni = n + k,
and the probability that node i will ( be cloned in)the next step is ni /(n + k). So
the development of the sequence ni : i ∈ V (G0 ) follows a Pólya urn model (see
e.g. Grimmett and Stirzaker [1982]), which implies that with probability ∑ 1, every
ratio ni /(n + k) tends to some real number xi . Clearly xi ≥ 0 and i xi = 1. So
Gn → WH , where H is obtained from G0 by weighting node i with weight xi (the
edges remain unweighted).
What are these values xi ? They can be anything (as long as they are nonnega-
tive and sum to 1). In fact, it follows from the theory of the Pólya urn that
∑ the vector
(xi ) is uniformly distributed over the simplex {x ∈ RV (G0 ) : xi ≥ 0, i xi = 1}. 
Let us close this section with a more interesting example with similar property.
Example 11.44 (Growing preferential attachment graphs). This randomly
growing graph sequence Gpa n is generated as follows. We start with a single node.
At the n-th step (when we already have a graph with n nodes), a new node labeled
n + 1 is created. This new node is connected to each old node i with probability
(dn (i) + 1)/(n + 1), independently for different nodes i, where dn (i) is the current
degree of node i. (Adding 1 in the numerator and denominator is needed in order
to generate anything other than empty graphs.)
The behavior of the graph sequence Gpan is somewhat unexpected: it is conver-
gent with probability 1, but the limit is not determined. More precisely:
Proposition 11.45. With probability 1, the sequence Gpa
n is quasirandom, i.e., it
converges to a constant function.
192 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

It is perhaps surprising that if we build a new preferential attachment graph

sequence by this method, it may tend to a different constant function. The distri-
bution of limit density is not known.
Proof. Set Xn = e(Gpa
n ). Then

∑
n−1
dn−1 (i) + 1 2 n−1
E(Xn | Gpa
n−1 ) = Xn−1 + = Xn−1 + Xn−1 + .
i=1
n n n
Hence
1 1
E(2Xn + 2n + 1 | Xn−1 ) = (2Xn−1 + 2n − 1),
(n + 2)(n + 1) (n + 1)n
which shows that the values Yn = (2Xn + 2n + 1)/((n + 2)(n + 1)) form a mar-
tingale. Since they are obviously bounded, the Martingale Convergence Theorem
implies that with probability 1 there is a value a such that Yn → a. Clearly,
Yn ∼ t(K2 , Gpa
n ), and so t(K2 , Gn ) → a.
pa
pa ∑n−1
Given Gn−1 , the degree of node n when it is born is i=1 Xi , where the Xi
are independent 0-1 random variables with E(Xi ) = (dn−1 (i) + 1)/(n + 1). Hence
∑
n−1
dn−1 (i) + 1 2 n−1
E(dn (n) | Gpa
n−1 ) = = e( Gpa
n−1 ) + ,
i=1
n n n
and hence (dn (n) + 1)/(n + 1) will be heavily concentrated around a. In particular,
(dn (n) + 1)/(n + 1) → a as n → ∞.
Next, observe that the development of dn (i), for a fixed i, follows a Pólya Urn
model with di (i) + 1 red and i − di (i) green balls, whence (dn (i) + 1)/(n + 1) is
a martingale converging to the beta distribution with parameters di (i) + 1 and
i − di (i). So for large i, (dn (i) + 1)/(n + 1) will be heavily concentrated around its
expectation (di (i) + 1)/(i + 1), which in turn is heavily concentrated around a. So
for large n, most nodes will have degree around an.
It follows that the process is almost the same as G(n, a), where we can also think
of the nodes created one-by-one and joined to each previous node with probability
a. We can couple the two processes to show that with probability 1, they converge
to the same limit, which is clearly the identically-a function.
Note that by the Martingale Stopping Theorem A.11,
2e(Gpa
n ) + 2n + 1
E(a | Gpa
n ) = Yn = .
(n + 2)(n + 1)
Since Gpan can be any simple graph on n nodes with positive probability, it follows
that a is not determined, and with a more careful computation one can see that a
falls into any interval with positive probability. 

Remark 11.46. In several examples above (11.35, 11.36, 11.41, 11.43), the limit
graphing is 0-1 valued. Consider, for example, the case of prefix attachment graphs.
It follows by Proposition 8.24 that WGpfx n
→ W pfx with probability 1 in the edit
distance, not just in the cut distance. This means that while the graphs Gpfx n
are random, they are very highly concentrated: two instances of Gpfx n differ in
o(n2 ) edges only, if overlayed properly (not in the original ordering of the nodes!).
Informally, they have a relatively small amount of randomness in them, which
 11.5. MANY DISGUISES OF GRAPH LIMITS 193

disappears as n → ∞. Indeed, Gpfx n is generated using

( ) only O(n log n) bits, as
opposed to, say, G(n, 1/2), which is generated using n2 bits. We’ll further explore
this phenomenon in Section 14.3.2.

Exercise 11.47. We define a randomly growing graph sequence Gn as follows.

We start with a single node. At the n-th iteration, a new node is born, it is
connected to node i with probability 1 − i/n, and every pair of nonadjacent nodes
is connected with probability 2/n. Prove that the sequence Gn tends to the
graphon 1 − xy with probability 1.
Exercise 11.48. Let us modify the cloning sequence as follows: at each step, we
select one of the original nodes (uniformly) and clone it. What is the limit of the
resulting graph sequence?
Exercise 11.49. Find limn→∞ t(F, Gpfx n ) for a general simple graph F .
Exercise 11.50. A random threshold graph is obtained by the following simple
procedure: starting with a single node, at each step we create a new node, flip
a coin, and connect the new node either to all previous nodes, or to none of
them, depending on the outcome of the coin flip. Prove that with probability 1,
a random threshold graph sequence converges to the simple threshold graphon
(Diaconis, Holmes and Janson [2008]).
Exercise 11.51. Verify that the uniform attachment graphon 1 − max(x, y) in
Example 11.39 is the operator square of the simple threshold graphon 1(x+y ≤ 1)
in Example 11.36. Find a combinatorial interpretation of this fact.

11.5. Many disguises of graph limits

We have seen that graphons, homomorphism numbers into graphons, certain
finite random graph models, as well as certain infinite random graph models can
all be considered as limit objects for convergent graph sequences. All of these
carry the same information, so it is a matter of taste which of these is called “the”
limit object. The following theorem summarizes a number of structures that are
equivalent to graphons. All these are familiar from our discussions.
Theorem 11.52. The following are equivalent (cryptomorphic):
(a) a graphon, up to weak isomorphism;
(b) a multiplicative, normalized simple graph parameter that is nonnegative on
signed graphs;
(c) a consistent and local random graph model;
(d) a local random countable graph model;
(e) a point in the completion of the space of finite graphs with the cut distance.
What we mean by equivalent (or cryptomorphic) is that given one of these
objects, one can construct an object from each of the other types (a)–(e), so that
these constructions are consistent (going around a cycle we get back to the original
object). Below we describe these constructions. This is a summary, most of the
work has been done in earlier parts of this book.
The conditions on these objects have useful alternative forms; for example, the
condition in (b) that the graph parameter is nonnegative on signed graphs could
be replaced by reflection positivity. We will discuss these equivalences in the more
general context of random graphons in Section 14.5.2.
194 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Proof. (a)→(b): Every graphon W gives rise to the simple graph parameter
t(., W ), which is, as we have seen, multiplicative, normalized, and nonnegative on
signed graphs.
(b)→(c): Let f be a multiplicative, normalized simple graph parameter that
is nonnegative on signed graphs. The conditions imply that the value of f does
not change if isolated nodes are added or deleted. We consider the signed graph Fb
obtained from a simple graph F by signing its edges with +, and the edges of its
∑
complement with −. It is easy to check that F ∈F simp Fb = Ok (the graph with no
∑ k
b) = f (Ok ) = f (O1 )k = 1. So the values f (Fb) form
edges), and hence simp f (F
F ∈Fk
a probability distribution σk on Fksimp . It is clear that this distribution is invariant
under isomorphism, so we get a random graph model (σk ).
∑ b = F K1 , and so
It is also easy to check that H: H ′ =F H
∑
f (Fb) = f (F
[ K1 ) = b
f (H).
H: H ′ =F
This means that generating a random graph from σk+1 , and deleting its last node,
we get a random graph from σk . So the model is consistent.
To show that the model is local, let S and T be disjoint subsets of [k], and let
FS and FT be two simple graphs on S and T , respectively. Let G be a random
graph from σk . Then
( ) ∑ ( )
P G[S] = FS , G[T ] = FT = P G[S ∪ T ] = H
V (H)=S∪T
H[S]=FS ,H[T ]=FT
∑
= b = f (F
f (H) cS F
cT ) = f (F
cS )f (F
cT )
V (H)=S∪T
H[S]=FS ,H[T ]=FT
( ) ( )
= P G[S] = FS P G[T ] = FT .
Thus the model is local.
(c)↔(d): This follows by Proposition 11.9 and the discussion before it.
(c)→(a): Generate a random graph Gn from the consistent local random graph
model. By Lemma 11.8, we get a convergent graph sequence with probability 1.
(d)↔(e): We have seen that a graph sequence is convergent if and only if it is
Cauchy in the cut distance. Every point in the completion is defined by a Cauchy
sequence, which tends to a graphon W . Two Cauchy sequences define the same
point of the completion if and only if merging them we get a Cauchy sequence,
which implies that they have the same limit graphon (up to weak isomorphism).
Conversely, every graphon is the limit of a Cauchy sequence (for example, the
sequence of W -random graphs), and so it corresponds to a point in the completion.


11.6. Convergence of spectra

We have seen in Section 7.5 that graphons can be considered as kernel operators
and they have a discrete spectrum such that there are only a finite number of
eigenvalues outside any neighborhood of 0; in particular, every nonzero eigenvalue
has finite multiplicity. All these eigenvalues are contained in the interval [−1, 1].
Magnitude does not order these eigenvalues into a single sequence: there may be an
infinite number of positive ones and an infinite number of negative ones, forming two
 11.6. CONVERGENCE OF SPECTRA 195

sequences λ1 (W ) ≥ λ2 (W ) ≥ · · · ≥ 0 and λ′1 (W ) ≤ λ′2 (W ) ≤ · · · ≤ 0 converging to

0. One or the other of these sequences may be finite; in this case, we still define
these infinite sequences, padding them with 0’s at the end.
It was proved by Borgs, Chayes, Lovász, Sós and Vesztergombi [2012] that if a
sequence of graphs converges to a graphon W , then their spectra converge to the
spectrum of W in an appropriate sense. (The more general Theorem 11.54 below
also implies a similar result for weighted graphs.)
Let G be a simple graph with n nodes, and let µ1 ≥ µ2 ≥ · · · ≥ µn be
the eigenvalues of its adjacency matrix. We normalize these eigenvalues to get
λi = λi (G) = µi /n, (i = 1, . . . , n). It will be useful to consider the sequence
ordered increasingly too, and denote it by λ′1 ≤ λ′2 ≤ . . . .
Theorem 11.53. Let (G1 , G2 , , . . . ) be a sequence of simple graphs converging to
a graphon W . Then for every fixed i ≥ 1 and n → ∞,
λi (Gn ) → λi (W ) and λ′i (Gn ) → λ′i (W ).
One may be puzzled by the fact that λi (W ) is nonnegative, while for a finite
graph Gn , we have no control over the sign of λi (Gn ). For example, if Gn = Kn ,
then every eigenvalue except the largest is negative. The theorem implies that
λi (Gn ) must have a nonnegative limit for every fixed i. This is of course not hard
to see directly (see Exercise 11.58).
This observation allows us to re-formulate this theorem in terms of the graphons
Wn = WGn . It is easy to see that the spectrum of Wn consists of the normalized
spectrum of Gn , together with infinitely many 0’s. So λ1 (Wn ) ≥ λ2 (Wn ) ≥ . . .
starts with the positive (or nonnegative) part of the spectrum of Gn , followed
by 0-s, and similar description can be given of the negative eigenvalues. Hence
λi (Gn ) ≤ λi (Wn ), but by the remark above, λi (Wn ) − λi (Gn ) → 0 for every fixed
i. So the following theorem implies Theorem 11.53.
Theorem 11.54. Let (W1 , W2 , . . . ) be a sequence of graphons, converging to a
graphon W in the δ distance. Then for every fixed i ≥ 1 and n → ∞,
λi (Wn ) → λi (W ) and λ′i (Wn ) → λ′i (W ).
Proof. By choosing a subsequence, we may assume that the limits
αi = lim λi (Wn ) and αi′ = lim λ′i (Wn )
n→∞ n→∞
exist for every j ≥ 1. We claim that for every k ≥ 3,
∞
∑ ∑∞
(11.14) λi (Wn )k −→ αik .
i=1 i=1
Indeed, every term on the left converges to the corresponding∑term on the right,
and the series are uniformly majorized by the convergent series i 1/ik/2 by (7.20).
Similar arguments can be applied to the negative eigenvalues, and we get that
∑ ∑∞ ∞
∑
λk −→ αik + (αi′ )k (n → ∞).
λ∈Spec(Wn ) i=1 i=1

For the sum of powers of the eigenvalues we have the graph-theoretic expression
∑ ∑
λk = t(Ck , Wn ) and λk = t(Ck , W ).
λ∈Spec(Wn ) λ∈Spec(W )
196 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

Now we know that Wn → W and hence t(Ck , Wn ) → t(Ck , W ). Thus we get

∑ ∞
∑ ∞
∑
λk = αik + (αi′ )k
λ∈Spec(W ) i=1 i=1

for every k ≥ 3. By Proposition A.21 in the Appendix, this implies that

{α1 , α2 , . . . } and {λ1 (W ), λ2 (W ), . . . } are equal as multisets, and similar conclusion
holds for {α1′ , α2′ , . . . } and {λ′1 (W ), λ′2 (W ), . . . }. 

The functionals λi (W ) and λ′i (W ) are invariant under measure preserving trans-
formations, and so they can be considered as a functional on the space (W, f δ ).
The theorem shows that these functionals are continuous. Of course, similar con-
clusion holds for the eigenvalues of kernels in Wf1 . By compactness, these maps are
uniformly continuous, which can be stated as follows:
Corollary 11.55. For ε > 0 and every i ≥ 1, there is a δi > 0 such that if
U, W ∈ W1 and δ (U, W ) ≤ δi , then
|λi (U ) − λi (W )| ≤ ε and |λ′i (U ) − λ′i (W )| ≤ ε.
Example 11.56. If (Gn ) is a quasirandom sequence with density p, then the largest
normalized eigenvalue of Gn tends to p, while the others tend to 0. The limiting
graphon, the identically-p function, has one nonzero eigenvalue (namely p). 
The last example suggests that perhaps convergent graph sequences can be
characterized through the convergence of their spectra, since if (Gn ) is a sequence
of graphs such that the edge density on Gn tends to p, the largest normalized
eigenvalue of Gn tends to p, and all the other eigenvalues tend to 0, then (Gn ) is
quasirandom. There is no real hope for this, however:
Example 11.57. Consider two non-isomorphic graphs G1 and G2 with the same
spectrum (for example, the incidence graphs of two non-isomorphic finite projective
planes of the same order). Consider the blow ups G1 (n) and G2 (n), n = 1, 2, . . . ,
and merge them into a single sequence. This sequence is not convergent, but all
graphs in it have the same spectra except for the multiplicity of 0. 

Exercise 11.58. Prove that for any finite simple graph G, λi (G) ≥ −(i −
1)/(v(G) − i + 1).

11.7. Convergence in norm

We discuss a somewhat technical issue that is important and also often conve-
nient when working with convergence of graph and graphon sequences.
Let Gn → W , so δ (WGn , W ) → 0. Of course, this does not mean that ∥WGn −
W ∥ → 0: the function WGn depends on the labeling of the nodes of Gn (the
distance δ (WGn , W ) does not, since relabeling Gn results in weak isomorphism of
WGn ). By the definition of the δ distance, we have measure preserving bijections
φn : [0, 1] → [0, 1] such that ∥WGφnn − W ∥ → 0. The functions WGφnn can be quite
complicated for general measure preserving bijections φn , so it is nice to know
that the maps φn can be chosen to be just permutations of the steps of WGn . In
other words, choosing the labeling of the graphs Gn appropriately, we can achieve
convergence in norm.
 11.8. FIRST APPLICATIONS 197

Theorem 11.59. Let (Gn ) be a sequence of graphs such that Gn → W . Then the
graphs Gn can be labeled so that ∥WGn − W ∥ → 0.
Proof. Let Pn be a partition of [0, 1] into consecutive intervals of length
1/v(Gn ). By Proposition 9.8, we have that ∥W − WPn ∥ → 0, so combined
with the assumption that δ (W, WGn ) → 0 we see that δ (WPn , WGn ) → 0.
Here δ (WPn , WGn ) = δ (W/Pn , Gn ) can be thought of as the distance of two
weighted graphs on the same number of nodes, so by Theorem 9.29, we get that
δb (W/Pn , Gn ) → 0. This means that the graphs in the sequence (Gn ) can be
relabeled to get a graph sequence (G′n ) such that
∥WPn − WG′n ∥ = d (W/Pn , G′n ) → 0.
Since ∥W − WPn ∥ → 0, this proves the Theorem. 

Exercise 11.60. Let us extend the definition of the distance δb to the case when
one of the arguments is a graphon U : δb (U, G) = minG′ ∥U − WG′ ∥ , where
G′ ranges over all relabeled versions of G. (a) Prove that if Gn → U , then
δb (U, Gn ) → 0. (b) Show by an appropriate construction that the following
stronger statement is not true: there exists a function f : [0, 1] × N∗ → [0, 1] such
that f (x, n) → 0 if x → 0 and n → ∞, and δb (U, G) ≤ f (δ (U, G), v(G)).
Exercise 11.61. Prove that a sequence of graphs is convergent if and only if they
have weak regularity partitions with convergent templates. More precisely, (Gn )
is convergent if and only if for every k ∈ N∗ V (Gn ) has
√ a k-partition Pk,n such
that (a) for every n, we have d (Gn , (Gn )Pn,k ) ≤ 10/ log k, and (b) for every k,
the template graphs Gn /Pn,k converge to some weighted graph Hk on k nodes as
n → ∞.
11.8. First applications
As a first illustration how our graph limits help proving theorems about finite
graphs, we describe two proofs working with graph limits of two important results: a
characterization of quasirandom graphs (see Section 1.4.2) and the Removal Lemma
(see Section 11.8.2). This should illustrate not only that graph limits are useful, but
also the method of obtaining a limit graphon from a sequence of counterexamples.
11.8.1. Quasirandom graphs. We start with quasirandom graphs. Let (Gn )
be a quasirandom sequence with density p (see Examples 11.37 and 11.56). This
means that t(F, Gn ) → pe(F ) for every simple graph F , or even simpler, Gn → p
(the identically-p graphon). We mentioned in the introduction the surprising fact,
due to Chung, Graham and Wilson [1989], that it is enough to require this relation
for F = K2 and F = C4 :
Theorem 11.62. If (Gn ) is a sequence of simple graphs such that v(Gn ) → ∞,
t(K2 , Gn ) → p, and t(C4 , Gn ) → p4 , then (Gn ) is quasirandom with density p.
Proof. Suppose that (Gn ) is not quasirandom, i.e., there is a simple graph F
such that t(F, Gn ) ̸→ pe(F ) . We can select a subsequence for which t(F, Gn ) →
c ̸= pe(F ) , and then we can select a convergent subsequence. Let W be its limit
graphon; then t(K2 , W ) = p, t(C4 , W ) = p4 , and W ̸= pJ as t(F, W ) = c ̸= pe(F ) .
To get a contradiction, it suffices to prove:
Claim 11.63. If W is a kernel such that t(K2 , W ) = p and t(C4 , W ) = p4 for
some real number p, then W = pJ.
198 11. CONVERGENCE OF DENSE GRAPH SEQUENCES

We start with two applications of the Cauchy–Schwarz inequality:

t( , W ) = ⟨txy ( , W ), txy ( , W )⟩ ≥ ⟨txy ( , W ), J⟩2 = t( , W )2

= ⟨tx ( , W ), tx ( , W )⟩2 ≥ ⟨tx ( , W ), J⟩4 = t( , W )4 .

Since we have equality, the function txy ( , W ) must be constant, and by integra-
tion we see that its value is p . This means that W ◦ W = txy ( , W ) = p2 J.
2
2
This means that the operator TW ◦W = TW has a single nonzero eigenvalue p2 with
eigenfunction ≡ 1. But then trivially TW has a single eigenvalue ±p with the same
eigenfunction, i.e., W ≡ p or W ≡ −p. The condition t(K2 , W ) = p rules out the
second alternative. 
11.8.2. Removal Lemma. One of the first, most famous, and in a sense in-
famous consequences of the Regularity Lemma was proved by Ruzsa and Szemerédi
[1976].
Lemma 11.64 (Removal Lemma). For every ε > 0 there is an ε′ > 0 such that
if a simple graph G with n nodes has at most ε′ n3 triangles, then we can delete εn2
edges from G so that the remaining graph has no triangles.
This lemma sounds innocent, almost like a trivial average computation. This
is far from the truth! No simple proof is known, and (worse) all the known proofs
give a terrible dependence of ε′ on ε. The best bound, due to Fox [2011] gives an
ε′ such that 1/ε′ is a tower 22 of height about log(1/ε). The original proof gives
2···

a tower √of height about 1/ε2 . Perhaps this looks friendlier (?) if we write it as
∗ ′
ε ≈ 1/ log (1/ε ). The proof given below does not give any explicit bound, but it
illustrates the way graph limit theory can be used.
Proof. Suppose that the lemma is false. This means that there is an ε > 0
and a sequence of graphs (Gn ) such that t(K3 , Gn ) → 0 but deleting any set of εn2
edges, the remaining graph will contain a triangle. By selecting a subsequence, we
may assume that t(F, Gn ) is convergent for every simple graph F , and then there is a
graphon W such that Gn → W . We have then t(K3 , W ) = limn→∞ t(K3 , Gn ) = 0.
The condition on the deletion of edges is harder to deal with, because it does
not translate directly to any property of the limit graphon W . What we can do
is to “pull back” information from W to the graphs Gn . By Theorem 11.59, we
may assume that ∥WGn − W ∥ → 0. (This step is not absolutely necessary, but
convenient.)
Let S = {(x, y) ∈ [0, 1]2 : W (x, y) > 0}. By Lemma 8.22, we have
∫ ∫
(1 − 1S )WGn → (1 − 1S )W = 0,
[0,1]2 [0,1]2
∫
so we can choose n large enough so that (1 − 1S )WGn < ε/4. Let V (Gn ) = [N ],
Ji = [(i − 1)/N, i/N ], and Rij = Ji × Jj .
We modify Gn by deleting the edge ij if λ(S ∩ Rij ) < 3/(4N 2 ).
Claim 11.65. The remaining graph G′n is triangle-free.
Indeed, suppose that i, j, k are three nodes such that λ(S ∩ Rij ) ≥ 3/(4N 2 ),
λ(S ∩ Rjk ) ≥ 3/(4N 2 ) and λ(S ∩ Rik ) ≥ 3/(4N 2 ). Observe that t(K3 , W ) = 0
 11.8. FIRST APPLICATIONS 199

implies that t(K3 , 1S ) = 0. But we have

∫
t(K3 , 1S ) ≥ 1S (x, y)1S (y, z)1S (x, z) dx dy dz
Ji ×Jj ×Jk
1 1 1 1
≥ 3
− λ(Rij \ S) − λ(Rjk \ S) − λ(Rik \ S)
N N N N
1 1 1
≥ 3 −3 = > 0,
N 4N 3 4N 3
which is a contradiction.
What is left is to bound the number m of edges deleted. This is easy: if edge
ij is deleted, then WGn = 1 on Rij , and so
∫ ∫
1
(1 − 1S )WGn = (1 − 1S ) = λ(Rij \ S) ≥ ,
4N 2
Rij Rij

We know that the number of deleted edges must be at least εN 2 , and so

∫
εN 2 ε
(1 − 1S )WGn ≥ = ,
4N 2 4
[0,1]2

which contradicts the choice of n. 

If you know the usual derivation of the Removal Lemma from the Regularity
Lemma, or have worked it out yourself by solving Exercise 11.66, you may feel
that the two proofs are analogous; and you are quite right. One may even say
that it is the same proof told in a different language, or at least, that it points
out a nontrivial connection between the Regularity Lemma and measure theory.
Indeed, other much deeper versions and applications of the Regularity Lemma
(like the Regularity Lemma for hypergraphs or Szemerédi’s Theorem on arithmetic
progressions) can be proved using mainly measure theoretic arguments; see Elek
and Szegedy [2012] [2012] for details.
Exercise 11.66. (a) Let G be a graph on n nodes, and consider a Szemerédi
partition of G with k classes and error bound ε as in the Original Regularity
Lemma. Let us delete all edges within the classes, between exceptional pairs of
classes, and between classes where the edge-density is less than 100ε1/3 . Prove
that if the remaining graph contains any triangle, then it contains at least ε(n/k)3
triangles. (b) Prove the Removal Lemma, based on (a).
Exercise 11.67. Extend the proof of the Removal Lemma above to the following
more general theorem: For every simple graph F and every ε > 0 there is an
ε′ > 0 such that if a simple graph G with n nodes has t(F, G) ≤ ε′ , then we
can delete εn2 edges from G so that F has no homomorphism into the remaining
graph.
 CHAPTER 12

Convergence from the right

Recall from the Introduction the formula

F −→ G −→ H,
which referred to the framework in which we study large graphs from “both sides”:
by mapping small graphs into them, and mapping them into small graphs. So far,
we have used homomorphisms into the large graph to define convergence. We have
seen many examples, however, of homomorphism numbers from the large graph
into fixed target graphs H which were very interesting: the number of k-colorings
of a graph; Ising and Potts models in statistical physics; approximating maximum
cuts. We have seen that homomorphism functions hom(G, .) have a characteriza-
tion (Theorem 5.59) perfectly analogous to the characterization of homomorphism
functions hom(., G) (Corollary 5.58). Does this duality extend to convergence?
In this chapter we set out to characterize convergence of a graph sequence
(G1 , G2 , . . . ) in terms of mappings “to the right”, using homomorphisms from the
graphs Gn into some fixed graph H, based on the results of Borgs, Chayes, Lovász,
Sós and Vesztergombi [2012].
The first and most natural approach is not going to work. It is clear that
considering simple graphs H would not give sufficient information: if the chro-
matic number of the graphs in the sequence tends to infinity, then hom(Gn , H) is
eventually 0 for every fixed simple graph H. We will see that even counting homo-
morphisms into weighted graphs H would not suffice to characterize convergence.
On the other hand, we show that a modification of the notion of homomorphism
numbers, or replacing counting by maximization (in other words, considering re-
stricted multicuts), does lead to a characterizations of convergence. It will turn out
that the convergence conditions hold for general weighted target graphs H, but it
is enough to require them for simple graphs (in the maximization version) of for
weighted graphs with just two positive edgeweights (in the counting case).
We will talk informally about left-convergence and right-convergence. Left-
convergence of a sequence (Gn ) means our notion of convergence as defined and
studied in the previous chapter. Right-convergence means a number of possible
convergence notions defined in terms of homomorphisms from the graph Gn into
fixed smaller graphs.

12.1. Homomorphisms to the right and multicuts

12.1.1. Naive right-convergence. Consider homomorphisms G → H,
where we think of G as a very large simple graph and H is a small weighted
graph. It will be convenient to scale the nodeweights of H so that αH = 1; this
only scales the values hom(G, H) by an easily computable factor. We will assume
that V (G) = [n] and V (H) = [q]. (We refer to Borgs, Chayes, Lovász, Sós and
201
202 12. CONVERGENCE FROM THE RIGHT

Vesztergombi [2012] for a treatment of the case when G is also weighted, and also
for consequences of these results in statistical physics.)
Recall that for a fixed weighted graph H, the value hom(G, H) grows expo-
nentially with n2 , and so a reasonable normalization is to consider the (dense)
homomorphism entropy
log hom(G, H)
ent(G, H) = .
v(G)2
The first, “naive” notion of right-convergence would be to postulate that these
homomorphism entropies converge for all weighted graphs H with (say) positive
edgeweights. This is at least a necessary condition for convergence:
Proposition 12.1. Let (Gn ) be a convergent graph sequence. Then for every
weighted graph H with positive edgeweights, the sequence ent(Gn , H) is convergent.
To prove this proposition, let us recall from Example 5.19 that the homomor-
phism entropy can be approximated by the maximum weighted multicut density:
Let G be a simple graph on [n], and H, a weighted graph on [q] with positive
edgeweights and αH = 1. Define Bij = log βij (H), then
log hom(G, H) log q
(12.1) cut(G, B) ≤ ≤ cut(G, B) + ,
n2 n
where cut(G, B) is the maximum weighted multicut density
1 ∑
cut(G, B) = max Bij eG (Si , Sj ).
(S1 ,...,Sq )∈Πn n2
i,j∈[q]

We need a couple of facts about weighted multicut densities. First, they are
invariant under blow-ups:
Lemma 12.2. For a simple graph G, symmetric matrix B ∈ Rq×q and integer
k ≥ 1, we have
cut(G(k), B) = cut(G, B).
Proof. The inequality cut(G(k), B) ≥ cut(G, B) is clear, since every q-
partition of V (G) can be lifted to a q-partition of V (G(k)), contributing the same
value to the maximization in the definition of cut(G(k), B). To prove the reverse
inequality, let (S1 , . . . , Sq ) be the q-partition of V (G(k)) attaining the maximum in
the definition of cut(G(k), B). For every node v ∈ V (G), we pick a random element
v ′ ∈ V (G(k)) uniformly from the set of twins of v created when blowing it up, and
let T = {v ′ : v ∈ V (G)}. Let G′ = G(k)[T ]. Then G′ ∼ = G, and
(1 ∑ ) 1 ∑
E 2 Bij eG′ (Si ∩ T, Sj ∩ T ) = 2
Bij eG(k) (Si , Sj )
n (nk)
i,j∈[q] i,j∈[q]

= cut(G(k), B).
It follows that for at least one choice of the nodes v ′ , we have
1 ∑
cut(G′ , B) ≥ 2 Bij eG′ (Si ∩ T, Sj ∩ T ) ≥ cut(G(k), B). 
n
i,j∈[q]

The following lemma is superficially similar to the Counting Lemma 10.22; it

is in fact quite a bit simpler.
 12.1. HOMOMORPHISMS TO THE RIGHT AND MULTICUTS 203

Lemma 12.3. For two simple graphs G and G′ and symmetric matrix B ∈ Rq×q ,
we have
|cut(G, B) − cut(G′ , B)| ≤ q 2 δ (G, G′ ).
Proof. We start with proving the weaker inequality
(12.2) |cut(G, B) − cut(G′ , B)| ≤ q 2 δb (G, G′ )
in the case when v(G) = v(G′ ) = n. We may assume that G and G′ are optimally
overlayed, so that V (G) = V (G′ ) = [n] and δb (G, G′ ) = d (G, G′ ). Then for every
partition (S1 , . . . , Sq ) ∈ Πn , we have
1 ∑ 1 ∑ 
 
 2 Bij eG (Si , Sj ) − 2 Bij eG′ (Si , Sj )
n n
i,j∈[q] i,j∈[q]
1 ∑
≤ 2 Bij |eG (Si , Sj ) − eG′ (Si , Sj )| ≤ q 2 d (G, G′ ).
n
i,j∈[q]

This proves (12.2). To get the more general inequality in the lemma, we apply
(12.2) to the graphs G(n′ k) and G′ (nk), where n = v(G), n′ = v(G′ ), and k is a
positive integer. The left side equals |cut(G, B) − cut(G′ , B)| for any k by Lemma
12.2, while the right side tends to q 2 δ (G, G′ ) if k → ∞ by the definition of δ . 
Proof of Proposition 12.1. By the Theorem 11.3, we have δ (Gn , Gm ) → 0
as n, m → ∞; by Lemma 12.3, this implies that the sequence of numbers cut(Gn , B)
is a Cauchy sequence; by (12.1), it follows that the values ent(Gn , H) form a Cauchy
sequence. 
It would be a natural idea here to define convergence of a graph sequence
in terms of the convergence of the homomorphism entropies ent(Gn , H). How-
ever, this notion of convergence would not be equivalent to left-convergence, and it
would allow sequences that we would not like to consider “convergent”, as Example
12.4 below shows. (Some suspicion could have been raised by (12.1) already: the
nodeweights of H disappeared, which indicated loss of information.)
Example 12.4. Let (Fn ) be a quasirandom graph sequence with edge density p,
and let (Gn ) be a quasirandom graph sequence of density 2p, where (to keep the
notation simple), we assume that v(Fn ) = v(Gn ) = n. Then we have, for every
weighted graph H with positive edgeweights,
1 ∑ 1
ent(Fn , H) = 2 max Bij eFn (Si , Sj ) + O( )
n (S1 ,...,Sq )∈Pn n
i,j∈[q]
∑ ( |S | |S | ) 1
i j
= max Bij p + o(1) + O( )
(S1 ,...,Sq )∈Pn n n n
i,j∈[q]

= p max{x Bx : x ∈ Rq+ , xT 1 = 1} + o(1).

T

Applying the same computation to Gn , we get that for the graphs G2n (disjoint
union of two copies of Gn ), we have
log hom(G2n , H) log(hom(Gn , H)2 ) log hom(Gn , H)
ent(G2n , H) = = =
(2n)2 4n2 2n2
1
= ent(Gn , H) = p max{xT Bx : x ∈ Rq+ , xT 1 = 1} + o(1).
2
204 12. CONVERGENCE FROM THE RIGHT

So merging the sequences (Fn ) and (G2n ) we get a graph sequence for which the
quantities ent(Gn , H) converge for every H, but which is clearly not convergent
(check the triangle density!). 
12.1.2. Typical homomorphisms. ∏ Let us try to take the nodeweights of H
into account. The values αφ = v∈V (G) αφ(v) form a probability distribution on the
maps φ : [n] → [q], where (by the Law of Large Numbers) we have |φ−1 (i)| ≈ αi n
with high probability, if n is large. However, this information becomes irrelevant
as n → ∞, and only the largest term will count rather than the “typical”. It turns
out that it is often advantageous to restrict ourselves to maps that are ”typical”,
by forcing φ to divide the nodes in the given proportions. Let Π(n, α) denote the
set of partitions (V1 , . . . , Vq ) of [n] into q parts with ⌊αi n⌋ ≤ |Vi | ≤ ⌈αi n⌉, , and
consider the set of maps
{ }
Φ(n, α) = φ ∈ [q]n : ⌊αi n⌋ ≤ |φ−1 (i)| ≤ ⌈αi n⌉ for all i ∈ [q] .
[ √ √ ]
(We could be less restrictive and allow, say φ−1 (i) ∈ αi n − n, αi n + n . This
would not change the considerations below in any significant way.)
We define a modified homomorphism number, by summing only over the “typ-
ical” homomorphisms:
∑ ∏
hom∗ (G, H) = αφ βφ(u)φ(v) .
φ∈Φ(n,α) uv∈E(G)

From this, we get the typical homomorphism entropy

log hom∗ (G, H)
ent∗ (G, H) = .
n2
12.1.3. Restricted multicuts. In maximum multicut problems, it is quite
natural to fix the proportion into which a multicut separates the node set. For
example, the “maximum bisection problem” asks for the maximum size of a cut that
separates the nodes into two almost equal parts. We can formulate the restricted
multicut problem as follows. We specify (in addition to the coefficients Bij ), q
further numbers α1 , . . . , αq > 0 with α1 + · · · + αq = 1.
It will convenient to consider the parameters αi and Bij as the nodeweights
and edge weights of a weighted graph H ′ with V (H ′ ) = [q]. Then we are interested
in the value
1 ∑
(12.3) cut(G, H ′ ) = max 2 Bij eG (Si , Sj ),
n i,j

where {S1 , . . . , Sq } ranges over all partitions of V (G) such that

(12.4) ⌊αi n⌋ ≤ |Si | ≤ ⌈αi n⌉ (i = 1, . . . , q).
This quantity is called a maximum restricted multicut, or in terms of statistical
physics, a microcanonical ground state energy. This quantity can be defined for all
graphs H with positive nodeweights, by scaling the nodeweights so that they sum
to 1.
The same simple computation as in Example 5.19 gives the following formula:
1
(12.5) ent∗ (G, H) = cut(G, H ′ ) + O( ).
n
Here H ′ is obtained from H by replacing all edgeweights by their binary logarithms
(while keeping the same nodeweights).
 12.2. THE OVERLAY FUNCTIONAL 205

In the study of subgraph densities, the identity t(F, G) = t(F, WG ) guaranteed

an easy transition between
( graphs
) and graphons. A finite consequence of this is
the identity t(F, G) = t F, G(m) (where G(m) is the m-blowup of the graph G).
Lemma 12.2 above shows that a similar identity holds for multicuts. However, for
restricted multicuts or homomorphisms to the right such a simple identity does not
hold any more. We will have to estimate the error we are making when we replace
a graph G by the associated graphon WG (luckily, this error will be small if G is
large enough).

Exercise 12.5. Show that if αi = 1/q for all i and n < q, then hom∗ (G, H) =
tinj (G, H).
Exercise 12.6. Let (Gn ) be a quasirandom sequence with edge density p, and
let F be a simple graph such that 2e(F [S]) ≤ q|S|2 for every subset S ⊆ V (F ).
Prove that cut(Gn , F ) ≤ pq + o(1) (n → ∞).

12.2. The overlay functional

The main advantage of the maximum-cut type functions introduced in the
previous section is that it is easy to extend their definitions to the case when the
graph G is replaced by a graphon. (Let me repeat that there does not seem to
be any reasonable extension of the hom function to graphons.) For a probability
distribution α on [q], let Π(α) denote the set of partitions (S1 , . . . , Sq ) of [0, 1] into
q measurable sets with λ(Si ) = αi . For every graphon U and weighted graph H on
node set [q], we define
∑ ∫
C(U, H) = sup Bij U (x, y) dx dy
(S1 ,...,Sq )∈Π(α)
i,j∈[q] Si ×Sj

This notion does not quite extend the maximum restricted weighted multicut of
graphs G; the reason is that in a graph, we cannot partition the set of nodes in
exactly the desired proportions. But the difference is small; we will come back to
this question in Section 12.4.1.
We can generalize even further and define, for two kernels U and W ,
∫
( )
C(U, W ) = sup ⟨U, W φ ⟩ = sup U (x, y)W φ(x), φ(y) dx dy.
φ∈S[0,1] φ∈S[0,1]
[0,1]2

It is easy to see that this extends the definition of maximum restricted weighted
multicuts in the sense that if U is any graphon and H is a weighted graph, then
(12.6) C(U, H) = C(U, WH ).
The functional C(U, W ), which we call the overlay functional, has many good
properties. It follows just like the similar statement for norms in Theorem 8.13 that
(12.7) C(U, W ) = sup ⟨U, W φ ⟩ = sup ⟨U φ , W ⟩ = sup ⟨U φ , W ψ ⟩
φ∈S[0,1] φ∈S[0,1] φ,ψ∈S [0,1]
{ }
= sup ⟨U0 , W0 ⟩ : (∃φ, ψ ∈ S [0,1] ) U = U0φ , W = W0ψ .
Hence it follows that the overlay functional is invariant under measure preserving
f0 × W
transformations of the kernels, i.e., it is a functional on the space W f0 . It is also
206 12. CONVERGENCE FROM THE RIGHT

immediate from the definition that this quantity has the (somewhat unexpected)
symmetry property C(U, W ) = C(W, U ), and satisfies the inequalities
(12.8) ⟨U, W ⟩ ≤ C(U, W ) ≤ ∥U ∥2 ∥W ∥2 , C(U, W ) ≤ ∥U ∥∞ ∥W ∥1 .
This suggests that C(., .) behaves like some kind of inner product. This analogy
is further supported by the following identity, reminiscent of the cosine theorem,
relating it to the distance δ2 derived from the L2 -norm:
1( )
(12.9) C(U, W ) = ∥U ∥22 + ∥W ∥22 − δ2 (U, W )2
2
1( )
= δ2 (U, 0)2 + δ2 (W, 0)2 − δ2 (U, W )2 .
2
Indeed,
δ2 (U, W )2 = inf ∥U − W φ ∥22 = ∥U ∥22 + ∥W ∥22 − 2 sup ⟨U, W φ ⟩
φ∈S[0,1] φ∈S[0,1]

= ∥U ∥22 + ∥W ∥22 − 2C(U, W ).

We have to be a bit careful: the functional C(U, W ) is not bilinear, only subadditive
in each variable:
(12.10) C(U + V, W ) ≤ C(U, W ) + C(V, W ).
It is homogeneous for positive scalars: if λ > 0, then
(12.11) C(λU, W ) = C(U, λW ) = λC(U, W ).
We have C(U, W ) = C(−U, −W ), but C(U, W ) and C(−U, W ) are not related in
general.
A less trivial property of the overlay functional is that it is continuous in each
variable (with respect to the δ distance). This does not follow from (12.9), since
the distance δ2 (U, W ) is not continuous with respect to δ (only lower semicontin-
uous; see Section 14.2.1).
Lemma 12.7. If δ (Un , U ) → 0 as n → ∞ (U, Un ∈ W1 ), then for every W ∈ W1
we have C(Un , W ) → C(U, W ).
Proof. By subadditivity (12.10), we have
−C(U − Un , W ) ≤ C(Un , W ) − C(U, W ) ≤ C(Un − U, W ),
and hence it is enough to prove that C(Un − U, W ), C(U − Un , W ) → 0. In other
words, it suffices to prove the lemma in the case when U = 0.
By definition, we have C(Un , W ) ≥ ⟨Un , W ⟩, and the right side tends to 0 by
Lemma 8.22. Hence lim inf n C(Un , W ) ≥ 0.
To prove the opposite
∑m inequality, we start with the case when W is a stepfunc-
tion. Write W = i=1 ai 1Si ×Ti , then using (12.10) and (12.11), we get
∑
m ∑
m
C(Un , W ) ≤ C(Un , ai 1S×T ) = C(ai Un , 1S×T )
i=1 i=1
∑m ∑
m
≤ ∥ai Un ∥ = |ai |∥Un ∥ → 0.
i=1 i=1
 12.3. RIGHT-CONVERGENT GRAPHON SEQUENCES 207

Now if W is an arbitrary kernel, then for every ε > 0 we can find a stepfunction W ′
such that ∥W −W ′ ∥1 ≤ ε/2. We know that C(Un , W ′ ) → 0, and hence C(Un , W ′ ) ≤
ε/2 if n is large enough. But then
C(Un , W ) ≤ C(Un , W − W ′ ) + C(Un , W ′ ) ≤ ∥Un ∥∞ ∥W − W ′ ∥1 + ε/2 ≤ ε.
This shows that lim supn C(Un , W ) ≤ 0, and completes the proof. 

While the functional C(U, W ) is continuous in each variable, it is not contin-

uous as a functional on the product space (W f1 , δ ) × (W
f1 , δ ). Let (Gn ) be any
quasirandom graph sequence and let Wn = Un = 2WGn − 1. Then Un , Wn → 0 in
the cut norm (and so also in δ ), but C(Un , Wn ) = 1 for all n.
Exercise 12.8. Define the following functional on W0 × W0 :
C ∗ (U, W ) = sup ⟨U, W φ ⟩
φ: [0,1]→[0,1]

where φ is measurable, but not necessarily measure preserving. Prove the formulas
maxcut(G) = C ∗ (WG , WK2 ),
and  ( )
∥U ∥ ≈ C ∗ U, 1(x, y ≤ 1/2) ,
where the ≈ sign means equality up to a factor of 2.

12.3. Right-convergent graphon sequences

As we have experienced before, many results are easier and cleaner when formu-
lated for graphons. This is particularly true for results about right-convergence. the
goal of this section is to formulate various characterizations of convergent graphon
sequences in terms of quantities defined by maps from the underlying set of a
graphon. Then in the next section we give a characterization of convergent graph
sequences, which will sound almost identical, but whose prof will be much more
tedious.

12.3.1. Quotient sets of graphons. If W is a kernel and P = (S1 , . . . , Sq )

is a measurable q-partition of [0, 1], we have defined the template (quotient) graph
W/P: it is a weighted graph on [q], with node weights αi (W/P) = λ(Si ) and edge
weights ∫
1
βij (W/P) = W.
λ(Si )λ(Sj ) Si ×Sj
The Regularity Lemma (Lemma 9.13 and its versions) said that there is always
a template that is close to the original graphon (or graph in the finite case) in the
cut norm. In order to get right-convergence criteria, we will study all templates of
a given graphon. For a kernel W and probability distribution a on [q], we denote
by Qa (W ) the set of templates L = W/P with α(L) = a. We denote by Qq (W )
the set of all q-partitions. ()
We will consider these quotient sets as subsets of the 2q + q dimensional real
space. The quotient set Qq (W ) is not always closed, but it is closed if W is a step-
function (see Exercise 12.13). The closure can be described in terms of “fractional
partitions”, also discussed in exercises.
Quotient sets can be used to express the overlay functional, at least if one of the
kernels involved is a weighted graph. For every weighted graph H with α(H) = a
208 12. CONVERGENCE FROM THE RIGHT

and every kernel W , we have

∑
(12.12) C(W, H) = max αi (L)αj (L)βij (H)βij (L).
L∈Qa (W )
i,j∈[q]

In this case, we have αi (L) = αi (H) for all i.

We need a classical definition: if (X, d) is a metric space, then the Hausdorff
metric is defined on the set of subsets of X by the formula
(12.13) dHaus (A, B) = inf{c : d(a, B) ≤ c ∀a ∈ A and d(b, A) ≤ c ∀b ∈ B}
(here d(a, B) denotes the distance of point a from set B). The special case we need
is when A, B are sets of weighted graphs on [q]. we can use the edit distance or the
cut distance, and denote the corresponding Hausdorff distance of two sets A and B
by dHaus
1 (A, B) or dHaus
 (A, B). Most of the time it does not make much difference
which one we use, because of the trivial inequalities
(12.14) dHaus
 (A, B) ≤ dHaus
1 (A, B) ≤ q 2 dHaus
 (A, B).
We start with a lemma relating quotient sets with different node weights.
Lemma 12.9.( For any W ∈ W)1 and any two probability distributions a, a′ on [q],
we have dHaus
1 Qa (W ), Qa′ (W ) ≤ 3∥a − a′ ∥1 .
Proof. Let L = W/P ∈ Qa (W ), where P = {S1 , . . . , Sq } ∈ Π(a). It is easy to
construct a partition P ′ = {S1′ , . . . , Sq′ } ∈ Π(a′ ) such that either Si ⊆ Si′ or Si′ ⊆ Si
for every i. Let L′ = W/P ′ ∈ Qa′ (W ). By definition,
∑ ∫ ∫ 

d1 (L, L′ ) = ∥a − a′ ∥1 +  W − W .
i,j∈[q] Si ×Sj Si′ ×Sj′

Here
∫ ∫  ( )
 
 W− W  ≤ λ (Si × Sj )△(Si′ × Sj′ )
Si ×Sj Si′ ×Sj′

≤ |ai − a′i | max(aj , a′j ) + |aj − a′j | min(ai , a′i ).

Summing over all i and j, we get
( ∑ ∑ )
d1 (L, L′ ) ≤ ∥a − a′ ∥1 1 + max(aj , a′j ) + min(ai , a′i )
j i
( ∑ )
= ∥a − a′ ∥1 1+ (aj + a′j ) = 3∥a − a′ ∥1 .
j

This proves the Lemma. 

We need a couple of lemmas about the Hausdorff distance of quotient sets of

different graphons.
Lemma 12.10. For any two graphons U and W and any integer q ≥ 1, we have
( )
dHaus
 Qq (U ), Qq (W ) ≤ δ (U, W ).
Proof. The quotient sets are invariant under weak isomorphisms, and hence
we may assume that U and W are optimally overlayed, so that δ (U, W ) =
∥U − W ∥ . The contractivity of the stepping operator (Exercise 9.17) asserts
 12.3. RIGHT-CONVERGENT GRAPHON SEQUENCES 209

that d (U/P, W/P) ≤ ∥U − W ∥ for any q-partition P of [0, 1]. By the definition
of Hausdorff distance, this implies that
( )
dHaus
 Qq (U ), Qq (W ) ≤ ∥U − W ∥ = δ (U, W ). 
Lemma 12.11. For any two graphons U and W and any integer q ≥ 1, we have
( ) ( ) ( )
dHaus
 Qq (U ), Qq (W ) ≤ sup dHaus
 Qa (U ), Qa (W ) ≤ 4dHaus
 Qq (U ), Qq (W )
a
(where a ranges over all probability distributions on [q]).
Proof. The first inequality is easy: let H ∈ Qq (U ), then H ∈ Qb (U ) for the
distribution b = α(H). Hence
( ) ( ) ( )
d H, Qq (W ) ≤ d H, Qb (W ) ≤ dHaus  Qb (U ), Qb (W )
( )
≤ sup a dHaus
 Qa (U ), Qa (W ) .
Since this holds for every H ∈ Qq (U ), and analogously for every graph in Qq (W ),
the inequality follows by the definition of the Hausdorff distance.
To prove the second inequality, let a be any probability distribution on [q]
and H ∈ Qa (U ).( For every ε > ) 0, there is a quotient L ∈ Qq (W ) such that
d (H, L) ≤ dHaus
 Q q (U ), Q q (W ) + ε. Let L ∈ Qb (W ), then ∥a − b∥1 ≤ d (H, L)
by the definition of d (H, L). By Lemma 12.9, there is a quotient L′ ∈ Qa (W )
such that d (L, L′ ) ≤ d1 (L, L′ ) ≤ 3|a − b| + ε ≤ 3d (H, L) + ε. Thus
( )
d (H, L′ ) ≤ d1 (H, L) + d (L, L′ ) ≤ 4d (H, L) + ε ≤ 4dHaus
 Qq (U ), Qq (W ) + 5ε.
Since ε was arbitrary, this proves the lemma. 
12.3.2. Graphon convergence from the right. After this preparation, we
are ready to characterize convergence of a graphon sequence in terms of homomor-
phisms into fixed weighted graphs.
Theorem 12.12. For any sequence (Wn ) of graphons, the following are equivalent:
(i) the sequence (Wn ) is convergent in the cut distance δ ;
(ii) the overlay functional values C(Wn , U ) are convergent for every kernel U ;
(iii) the restricted multicut densities C(Wn , H) are convergent for every simple
graph H;
(iv) the quotient sets Qq (Wn ) form a Cauchy sequence in the dHaus Hausdorff
metric for every q ≥ 1.
It follows from conditions (ii) and (iii) that it would be equivalent to assume
the convergence of the sequence C(Wn , H) for every weighted graph H. Lemma
12.11 implies that we could require in (iv) the convergence of Qa (Wn ) for every
q ≥ 1 and probability distribution a on [q]. In fact, it would be enough to require
this for the uniform distribution (see Exercise 12.24). In (iv), we could use the
dHaus
1 Hausdorff metric as well.
Proof. (i)⇒(ii) by Lemma 12.7. (ii)⇒(iii) is trivial. (i)⇒(iv) by Lemma 12.10.
(iii)⇒(i): Let (Wn ) be a sequence of graphons that is not convergent in the
cut distance. By the compactness of the graphon space, it has two subsequences
(Wni ) and (Wmi ) converging to different unlabeled graphons W and W ′ . There is
a graphon U such that C(W, U ) ̸= C(W ′ , U ); in fact, (12.9) implies
( ) ( )
C(W ′ , W ′ ) − C(W ′ , W ) + C(W, W ) − C(W ′ , W ) = δ2 (W ′ , W )2 > 0,
210 12. CONVERGENCE FROM THE RIGHT

and so either C(W ′ , W ′ ) ̸= C(W, W ′ ) or C(W, W ′ ) ̸= C(W, W ), and we can take

either U = W or U = W ′ . Furthermore, we can choose U of the form U = WH ,
where H is a simple graph. This follows using the fact that simple graphs are dense
in the graphon space, and the continuity of the overlay functional (Lemma 12.7).
Since C(Wni , H) → C(W, H) and C(Wmi , H) → C(W ′ , H) by Lemma 12.7, it follows
that the values C(Wn , H) cannot form a convergent sequence, contradicting (ii).
(iv)⇒(iii): Fix any simple graph H on [q], and let a be the uniform distribution
on [q]. Let n, m ≥ 1, then we have
1 ∑ 1 ∑
C(Wn , H) = sup 2
βij (L) = max 2
βij (L).
L∈Qa (Wn ) q i,j L∈Qa (Wn ) q i,j
ij∈E(H) ij∈E(H)

Let Ln ∈ Qa (Wn ) attain the maximum. By the definition ( of Hausdorff distance,

)
there is an L′ ∈ Qa (Wm ) such that d (Ln , L′ ) ≤ dHaus Qa (Wn ), Qa (Wn ) . The
 ∑
definition of the C functional implies that C(Wm , H) ≥ (1/q 2 ) ij∈E(H) βij (L′ ).
Hence
1 ∑ 1 ∑
C(Wn , H) − C(Wm , H) ≤ 2 βij (Ln ) − 2 βij (L′ )
q q i,j
ij∈E(H)
ij∈E(H)

1 ∑
≤ 2 |βij (Ln ) − βij (L′ )| = d1 (Ln , L′ ) ≤ q 2 d (Ln , L′ )
q i,j
( )
≤ q 2 dHaus
 Qa (Wn ), Qa (Wn ) .
By Lemma 12.11, we have
( ) ( )
dHaus
 Qa (Wn ), Qa (Wn ) ≤ 4dHaus
 Qq (Wn ), Qq (Wn ) ,
which tends to 0 as n, m → ∞ by hypothesis. This implies that
( )
lim sup C(Wn , H) − C(Wm , H) ≤ 0.
n

Since
( a similar conclusion ) holds with n and m interchanged, we get that
C(Wn , H) : n = 1, 2, . . . is a Cauchy sequence. 

Some of the arguments in the proof of Theorem 12.12, most notably the proof
of (iii)⇒(i), were not effective. One can in fact prove explicit inequalities between
the different distance measures that occur. We refer to Borgs, Chayes, Lovász, Sós
and Vesztergombi [2012] for the details.
Exercise 12.13. Show by an example that the set Qq (W ) is not closed in general,
but it is closed if W is a stepfunction.
Exercise 12.14. Show by an example that Qa (W ) is not convex in general, even
if W is a stepfunction.
Exercise 12.15. A fractional partition of [0, 1] into q parts is an ordered q-tuple
of measurable functions ρ1 , . . . , ρq : S → [0, 1] such that for all x ∈ [0, 1], we
have ρ1 (x) + · · · + ρq (x) = 1. For a fractional partition ρ of [0, 1] and a kernel
W ∈ W, we define the fractional quotient graph W/ρ as a weighted graph on [q]
with αi (W/ρ) = ∥ρi ∥1 and
∫
1
βij (W/ρ) = ρi (x)ρj (y)W (x, y) dx dy.
∥ρi ∥1 ∥ρj ∥1
[0,1]2

Prove that the operation W 7→ W/ρ is contractive for the L1 and L2 .

 12.4. RIGHT-CONVERGENT GRAPH SEQUENCES 211

Exercise 12.16. Let ρ be a fractional q-partition of [0, 1]. Prove that W/ρ ∈
Qq (W ). Also proved that every weighted graph in Qq (W ) can be represented this
way.

12.4. Right-convergent graph sequences

Our basic plan is to apply theorem 12.12 to graphons WGn to get characteriza-
tions of convergent graph sequences in terms of homomorphisms to the right. There
are two difficulties in the way. First, in the restricted multicut density condition
(iii), thee is no way in general to partition the nodes of a graph in given proportions;
we have to allow some rounding of the prescribed sizes for the partition classes. Sec-
ond, a partition of [0, 1] obtained when we overlay WGn and H optimally may not
correspond to any partition of V (Gn ). (It corresponds to a “fractional partition”,
which we will have to define because of this.) This problem is more serious, and we
have to work harder to obtain true partitions of V (Gn ).

12.4.1. Restricted quotients. Let G be a simple graph with nodeset [n] and
let P = (S1 , . . . , Sq ) be a partition of [n]. We consider the quotient graph G/P as
a weighted graph on [q], with node weights αi (G/P) = |Si |/n (i ∈ [q]), and edge
weights βij (G/P) = eG (Si , Sj )/|Si ||Sj | (i, j ∈ [q]).
The set of all weighted graphs G/P, where P ranges over all q-partitions of [n],
will be called the quotient set of G (of size q), and will be denoted by Qq (G). For
a graph G and a probability distribution a on [q], to define the restricted quotient
set Qa (G), we have to allow the relative sizes of the partition classes to deviate a
little from the prescribed values a: we consider the set of quotients G/P, where
P ∈ Π(n, a).
Quotient sets can be used to express multicut functions. For every weighted
graph H,
∑
(12.15) cut(G, H) = max αi (L)αj (L)βij (L)βij (H).
L∈Qa (G)
i,j∈[q]

Note that the nodeweights of H and L are not the same in general, but almost:
|αi (L) − αi (H)| ≤ n1 .

Remark 12.17. The quotient sets are in a sense dual to the (multi)sets of induced
subgraphs of a given size, which was one of the equivalent ways of describing what
we could see by sampling. Instead of gaining information about a large graph by
taking a small subgraph, we take a small quotient.
However, there are substantial differences. On the set of induced subgraphs of a
given size, we had a probability distribution, which carried the relevant information.
We can also introduce a probability distribution on quotients of a given size of a
graph G, by taking a random partition. This would be quite relevant to statistical
physics, but we would run into difficulties when tending to infinity with the size
of G. The probability distributions would concentrate more and more on boring
average quotients, while the real information would be contained in the outliers.
To be more specific, a random induced subgraph (of a fixed, but sufficiently large
size) approximates the original graph well, but a random quotient does not carry
this information. In other words, it is the set of quotients that characterizes the
convergence of a graph sequence, and not the distribution on it.
212 12. CONVERGENCE FROM THE RIGHT

12.4.2. Fractional partitions. A fractional partition of a set S is an ordered

q-tuple of functions ρ1 , . . . , ρq : S → [0, 1] such that for all x ∈ [0, 1], we have
ρ1 (x) + · · · + ρq (x) = 1. An ordinary partition corresponds to the special case when
every ρi is 0-1 valued. For every vector α ∈ [0, 1]q , we denote by Π∗ (n, α) the set
of fractional partitions ρ of [n] with ∥ρi ∥1 = αi n.
We extend the notion of quotients to fractional partitions. For every fractional
partition ρ of [n], we consider the fractional quotient graph G/ρ, which is a weighted
graph on [q], with node weights
1 1 ∑
αi (G/ρ) = |ρi | = ρi (u) (i ∈ [q])
n n
u∈[n]

and edge weights

∑ /
βij (G/ρ) = ρi (u)ρj (v) |ρi ||ρj | (i, j ∈ [q])
u,v∈[n]

In the special case when every value ρi (u) is 0 or 1, then the supports of the
functions ρi form a partition P, and G/P = G/ρ.
We also introduce fractional quotient sets, replacing partitions by fractional
partitions. The set of all fractional q-quotients of a graph G is denoted by Q∗q (G),
and the set of all fractional q-quotients G/ρ for which α(G/ρ) is a fixed distribution
a on [q], by Q∗a (G).

12.4.3. Relations between quotient sets. Our goal is to use quotient sets
to characterize convergence of a graph sequence. But before doing so, we have to
formulate and prove a number of rather technical relationships between different
quotient sets.
For any simple graph G and positive integer q, we have two quotient sets: the
set Qq (G) of quotients G/P, and the set Q∗q (G) of fractional quotients G/ρ. In
addition, we have the restricted versions Qa of both of these. The quotient sets
Qq (WG ) and Qa (WG ) will also come up; but it is easy to see that these are just
the same as Q∗q (G) and Q∗a (G).
Turning to the quotient set Qq (G) (which is of course the most relevant from
the combinatorial point of view), it follows immediately from the definition that
Qq (G) ⊆ Q∗q (G). Note, however, that Qa (G) and Q∗a (G) are in general not com-
parable. The first set is finite, the second is typically infinite. On the other hand,
Qa (G) contains graphs whose nodeweight vector is only approximately equal to a,
and so it is not contained in Q∗a (G).
In the rest of this section we are going to prove that the “true” quotient sets
and their fractional versions are not too different, at least if the graph is large. We
will need the following version of Lemma 12.9, which can be proved along the same
lines.

Lemma 12.18. For (any simple graph) G and any two probability distributions a, a′
on [q], we have dHaus
1 Qa (G), Qa′ (G) ≤ 3∥a − a′ ∥1 .

The two kinds of quotient sets of the same graph are related by the following
proposition.
 12.4. RIGHT-CONVERGENT GRAPH SEQUENCES 213

Proposition 12.19. For every simple graph G on [n], integer q ≥ 1, and probability
distribution a on [q],
( ∗ ) 4q ( ∗ ) 16q
dHaus
1 Qq (G), Qq (G) ≤ √ and dHaus
1 Qa (G), Qa (G) ≤ √ .
n n
Proof. We start with the first inequality, whose proof gets somewhat technical.
Since Qq (G) ⊆ Q∗q (G), it suffices to prove that if H is a fractional q-quotient of G,
√
then there exists a q-quotient L of G such that d1 (H, L) ≤ 4q/ n. We may assume
that q ≥ 2 and n > 9q 2 ≥ 36 (else, the assertion is trivial).
Let ρ ∈ Π∗ (n, α) be a fractional partition such that G/ρ = H. We want to
“round” the values ρi (u) to ri (u) ∈ {0, 1} so that we get an integer partition in
Π(n, α) with “almost” ∑ the same quotient. Let A denote the adjacency matrix of G,
and define Fij (r) = u,v∈[n] Auv ri (u)rj (v), then we want
∑ ∑
(12.16) ri (u) = 1, ⌊αi n⌋ ≤ ri (u) ≤ ⌈αi n⌉, Fij (r) ≈ Fij (ρ)
i u

for all i, j ∈ [q].

We do random ( rounding: for) each u ∈ [n], let Zu be chosen randomly from
the distribution ρ1 (u), . . . , ρq (u) , and let Ri (u) = 1(Zu = i). So P(Ri (u) = 1) =
ρi (u), and for∑ different nodes u of G the random variables Ri (u) are independent.
We have i Ri (u) = 1 for every u ∈ [n], and so the Ri define a partition P.
Let L = G/P.( For the other ) two conditions we get the right value at least in
∑ ( )
expectation: E u Ri (u) = αi n and E Fij (R) = Fij (ρ) (in the last equation we
use that Ri (u) and Rj (v) ∑ are independent if u ̸= v, and Auv = 0 if u = v). We
consider the errors Xi = u Ri (u) − αi n and Yij = Fij (R) − Fij (ρ), and use a
second moment argument to show that these are small. We have
∑ ( ) ∑
Var(Xi ) = Var Ri (u) = (ρi (u) − ρi (u)2 ) < αi n,
u u

and hence ∑ ∑
E( Xi2 ) = Var(Xi ) < n.
i i
Furthermore,
(12.17)
( ) ∑ ( )
Var(Yij ) = Var Fij (R) = Auv Au′ v′ cov Ri (u)Rj (v), Ri (u′ )Rj (v ′ ) .
u,v,u′ ,v ′ ∈[n]

Each covariance in this sum depends on which of u, v, u′ , v ′ and also which of i and
j are equal, but each case is easy to treat. The covariance term is 0 if the edges uv
and u′ v ′ are disjoint. If i ̸= j, we get:


 ρi (u)ρj (v) − ρi (u)2 ρj (v)2 < ρi (u)ρj (v), if u = u′ and v = v ′ ,

−ρ (u)ρ (v)ρ (u)ρ (v) < 0,
i i j j if u = v ′ , v = u′ ,

 ρj (v)ρj (v ′ )(ρi (u) − ρi (u)2 ) < ρi (u)ρj (v)ρj (v ′ ), if u = u′ and v ̸= v ′ ,


−ρi (u)ρj (v)ρi (v)ρj (v ′ ) < 0, if u = v ′ , u′ ̸= v.
(The other possibilities are covered by symmetry.) Summing over u, v, u′ , v ′ , we get
that the sum in (12.17) is at most (αi n)(αj n) + 0 + (αi n)(αj n)2 + (αi n)2 (αj n) + 0.
214 12. CONVERGENCE FROM THE RIGHT

The case when i = j can be treated similarly, and we get that the sum in (12.17)
is at most 2(αi n)2 + 4(αi n)3 . Hence, summing over all i and j, we get
∑ ∑ ∑ ∑
E(Yij2 ) = Var(Yij ) ≤ n2 + (n2 + 2n3 ) αi2 + 4n3 αi3 ≤ 6n3 + 2n2 .
i,j i,j i i

By Cauchy–Schwarz,
(∑ |X | ∑ |Y | )2 (1 ∑ 1 ∑ 2)
i ij
d1 (H, L)2 = + ≤ (q 2
+ q) Xi
2
+ Y ,
i
n i,j
n2 n2 i n4 i,j ij
and so
(1
6 2 ) 16q 2
E(d1 (H, L)2 ) = (q 2 + q) +
+ 2 < .
n n √ n n
Hence with positive probability, d1 (H, L) ≤ 4q/ n.
The second inequality in the proposition is quite easy to prove now, except
that we cannot use containment in either direction, and so we have to prove two
“almost containments”. Let H ∈ Qa (G) and let b = α(H), ∥a − b∥1 ≤ q/n,
and H ∈ Q∗b (G). By Lemma ∗
√ 12.18, there is an L ∈ Qa (G) such that d1 (H, L) ≤
3∥a − b∥1 ≤ 3q/n < 16q/ n.
Conversely, let H ∈ Q∗a (G), then ′
√ by part (a), there exists a q-quotient H ∈
′
Qq (G) such that d1 (H, H ) ≤ 4q/ n. Lemma 12.18 implies that there exists an
L ∈ Q∗a (G) such that
d1 (L, H ′ ) ≤ 3|a − α(H ′ )| = 3|α(H) − α(H ′ )| ≤ 3d1 (H, H ′ ),
√
and so d1 (L, H) ≤ d1 (L, H ′ ) + d1 (H ′ , H) ≤ 4d1 (H, H ′ ) ≤ 16q/ n. 
12.4.4. Right-convergent graph sequences. In a sense, right-convergence
of a graph sequence is a special case of right-convergence of a graphon sequence.
However, quantities like multicuts associated with a graphon of the form WG are
only approximations of the analogous combinatorial quantities associated with the
corresponding graph G. (This is in contrast with the homomorphism densities
from the left, recall e.g. (7.2).) In this section we prove that these approximations
are good enough for the equivalent characterizations of convergence of graphon
sequences to carry over to graph sequences.
We prove the following characterization of convergence of a dense graph se-
quence, analogous to the characterization of convergence of a graphon sequence
given in Theorem 12.12.
Theorem 12.20. Let (Gn ) be a sequence of simple graphs such that v(Gn ) → ∞
as n → ∞. Then the following are equivalent:
(i) the sequence (Gn ) is convergent;
(ii) the overlay functional values C(WGn , U ) are convergent for every kernel U ;
(iii) the restricted multicut densities cut(Gn , H) are convergent for every simple
graph H;
(iv) the quotient sets Qq (Gn ) are Cauchy in the Hausdorff metric for every
q ≥ 1.
Clearly, conditions (ii) and (iii) are also equivalent to the convergence of
cut(Gn , H) for every weighted graph H. By (12.5), this is equivalent to the con-
vergence of typical homomorphism entropies ent∗ (G, J) for every weighted graph
J with positive edgeweights. By our discussion in Section 2.2, we could talk about
 12.4. RIGHT-CONVERGENT GRAPH SEQUENCES 215

microcanonical ground state energies instead of restricted multicuts. By the results

of the previous section, we could replace Qq (Gn ) by Q∗q (Gn ), or we could require
the convergence of Qa (Gn ) for every q ≥ 1 and probability distribution a on [q].
On the other hand, it would be enough to require this for the uniform distribution
(see Exercise 12.24).
Proof. If we replace Gn by WGn , then (i) and (ii) do not change. In (iii), we
have ∑
C(Gn , H) = max αi (L)αj (L)βij (L)βij (H),
L∈Qa (G)
i,j∈[q]
and ∑
C(WGn , H) = max
∗
αi (L)αj (L)βij (L)βij (H).
L∈Qa (G)
i,j∈[q]
( ∗ ) √
Since dHaus
1 Qa (G), Qa (G) ≤ 9q/ n by Proposition 12.19, it follows that C(Gn , H)
is convergent as n → ∞ if and only if C(WGn , H) is. By a similar argument, the
validity of (iv) does not change if we replace Gn by WGn . So the theorem follows
by Theorem 12.12. 
The last theorem in this chapter describes the limiting values in Theorem 12.20.
The proof is contained in our previous considerations.
Theorem 12.21. Let Gn be a convergent sequence of simple graphs such that
Gn → W (W ∈ W0 ). Let F be a simple graph, let H be a weighted graph with
positive edgeweights, and let J be obtained from H by replacing every edgeweight by
its binary logarithm. Then cut(Gn , J) → C(W, J) and ent∗ (Gn , H) → C(W, J).
Remark 12.22. How far can we push convergence results for right-homomorphism
parameters of a convergent graph sequence? Suppose that (Gn ) is a convergent
graph sequence. Does it follow that log t(Gn , W )/v(Gn )2 tends to a limit for every
graphon W ? Or at least, for every graphon W > 0? or at least for every graphon W
with 1/2 ≤ W ≤ 1? Theorem 12.20(ii) suggests an affirmative answer, but this is
false: the example in Exercise 12.25 (which is hard!) shows that right-convergence
in this more general sense does not follow from left-convergence.

Exercise 12.23. Prove that maxH |C(U, H) − C(W, H)|, where the maximum is
taken over all weighted graphs H on [q] with nodeweight
( vector a) and edgeweights
( )
in [−1, 1], is equal to the Hausdorff distance dHaus
1 conv(Qa (U ) , conv Qa (U ) ).
Exercise 12.24. Prove that a sequence (Wn ) of graphons is convergent if and
only if the quotient sets Qu (Wn ) are convergent in the Hausdorff metric for every
q ≥ 1, where u is the uniform distribution on [q].
Exercise 12.25. Let (Gn ) be a quasirandom graph sequence with edge density
1/2, such that v(Gn ) ∑ = kn is a sufficiently fast increasing sequence of integers.
Define W (x, y) = 1 + ∞ n=1 WG2n (k2n x, k2n y) (where WGn (x, y) = 0 if x ∈
/ [0, 1]
or y ∈
/ [0, 1]). Prove that the kernel W is 1-2 valued, and
log t(G2n , W ) 1 log t(G2n+1 , W ) 1
2
= + o(1), 2
= + o(1).
k2n 2 k2n+1 4
 CHAPTER 13

On the structure of graphons

13.1. The general form of a graphon

A probability space J = (Ω, A, π) together with a symmetric function W : J ×
J → R, measurable with respect to the completion of the sigma-algebra A × A, will
be called a kernel, and if the range of W is contained in [0, 1], a graphon.
So far, we have assumed that J is the unit interval [0, 1] with the Lebesgue
measure, and this was good enough to get a limit object for every convergent graph
sequence. But recall Example 11.41, where it took an artificial step of applying a
measure preserving map to carry the graphon structure over to the unit interval.
A similar step was needed in the definition of the tensor product of two kernels
(Section 7.4). In both cases, allowing more general underlying sets leads to a
simpler and cleaner situation. If we have to make this distinction, a graphon on
[0, 1] with the Lebesgue measure will be called a graphon on [0, 1].
Subgraph densities can be defined in any graphon by the same formula as in
the special case of [0, 1]: For a multigraph F = (V, E), we define
∫ ∏ ∏
t(F, W ) = W (xi , xj ) dπ(xi ).
ij∈E i∈V
ΩV

Hence we can define weak isomorphism of kernels as before.

We can sample H(n, W ) and G(n, W ) from any graphon. If we fix the under-
lying space, we can talk about kernel norms as before. If (Ω, A, π) and (Ω′ , A′ , π ′ )
are two probability spaces, (Ω, A, π, W ) is a kernel, and φ : Ω′ → Ω is a measure
preserving map, then the pullback W φ can be defined as before:
( )
W φ (x, y) = W φ(x), φ(y) .
It is clear that (Ω′ , A′ , π ′ , W φ ) will be a kernel, which we call the pullback of (J ′ , W ′ ).
Simple computation shows that for every graph F ,
(13.1) t(F, W ) = t(F, W φ ),
so (Ω′ , A′ , π ′ , W φ ) is weakly isomorphic to (Ω, A, π, W ).
To define the δ distance of two kernels, we have to be a little careful, since at
this point we allow the underlying spaces to have atoms, and then the difficulties
in the definition of the cut-distance for graphs (Sections 8.1.2–8.1.4) resurface. Let
A1 = (J1 , W1 ) and A2 = (J2 , W2 ) be two kernels. To avoid the difficulty with
atoms, we consider all kernels (J, W ) and all pairs of measure preserving maps
φi : J → Ji . We define
δ (A1 , A2 ) = inf ∥W1φ1 − W2φ2 ∥ ,
where the infimum ranges over all choices of J, W, φ1 and φ2 .
217
218 13. ON THE STRUCTURE OF GRAPHONS

At this point, there are some rather technical questions to address. Do we want
to assume that J is a standard probability space (see Appendix A.3)? Do we want
to consider A as a complete sigma-algebra with respect to the probability measure
(like Lebesgue measurable sets in [0, 1]), or to the contrary, do we want to assume
that it is countably generated (like Borel sets in [0, 1])?
It was shown by Borgs, Chayes and Lovász [2010] that a kernel on an arbitrary
probability space can be transformed, by very simple steps, into a kernel on a stan-
dard probability space, which is equivalent for all practical purposes (in particular,
weakly isomorphic). The steps of such a transformation are described in Exer-
cises 13.11–13.13 below. This implies that we can work with standard probability
spaces whenever necessary (or just convenient). We call a kernel standard, if the
underlying probability space is standard.
From the point of view of subgraph densities, multicuts, etc. the underlying
space does not matter much, as we shall see; but choosing the underlying probability
space appropriately may lead to a simpler form for the function W and to simpler
computations. If the space J is finite, we get just a weighted graph with normalized
nodeweights. Let us see a number of further examples where allowing this more
general form is very useful (and so Example 11.41 was not an isolated occurrence).

Example 13.1 (Limits of interval graphs). An interval graph is a graph ob-

tained from a finite set of intervals on a line, where the intervals are the nodes,
and two intervals are connected by an edge if and only if they have a point in com-
mon. Diaconis, Holmes and Janson [2011] give the following description of limits
of interval graphs. ( ) (
Let J =) {(x, y) ∈ [0, 1]2 : x ≤ y}, and define W (x1 , y1 ), (x2 , y2 ) = 1 [x1 , y1 ]∩
[x2 , y2 ] ̸= ∅ . So (J, W ) can be considered as an infinite interval graph, whose nodes
are all sub-intervals of [0, 1]. We can take any probability measure on the Borel
sets in J: we always obtain a graphon that is the limit of interval graphs, and all
interval graph limits arise this way. 

Example 13.2. Fix some d ≥ 2, and let Vn be a set of n unit vectors in Rd , chosen
independently from the uniform distribution on the unit sphere. Connect two
elements x, y ∈ Vn by an edge if and only if xT y ≥ 0, to get a graph Gn = (Vn , En ).
The sequence (Gn : n = 1, 2, . . . ) is convergent, and its limit is the graphon whose
underlying set is S d−1 , with the uniform distribution, and W (x, y) = 1(xT y ≥ 0).


13.1.1. Atomfree and twin-free kernels. There are still ways to further
simplify a kernel (J, W ) on a standard probability space (Ω, A, π). One possibility
is to get rid of the atoms by a procedure generalizing the construction of the kernel
WH from a weighted graph H, by assigning to each atom a an interval Ia of length
π(a), and an interval I to the atom-free part of Ω, so that these intervals partition
[0, 1]. This defines a measure preserving map ψ : [0, 1] → Ω, and the pullback W φ
will define a kernel on [0, 1] that is weakly isomorphic to (J, W ).
This procedure takes us to a very familiar domain (two-variable real functions),
but the kernel on [0, 1] is still not uniquely determined by its weak isomorphism
class, as we have seen in Example 7.11. To really standardize a kernel, we go the
opposite way, by creating and merging atoms as much as we can. To be more
precise, we need some definitions.
 13.1. THE GENERAL FORM OF A GRAPHON 219

Let (J, W ) be a kernel. Two points x, x′ ∈ J are called twins if W (x, y) =

W (x′ , y) for almost all y ∈ J. This defines an equivalence relation on J. We call
the kernel H twin-free if no two distinct points in J are twins.
Proposition 13.3. For every kernel (J, W ) there is a twin-free kernel (J1 , W1 )
and a measure preserving map φ : J → J1 such that W = W1φ almost everywhere.
If J is standard, then we can require that J1 be standard as well.
Proof. The twin-free kernel and the measure preserving map φ will be easy to
define, but we have to work on verifying their properties. First, we only modify the
sigma-algebra of the probability space J = (Ω, A, π). Let us define a new sigma-
algebra A′ consisting of those sets in A that do not separate any twin points. Let
J ′ = (Ω, A′ , π) (this is a little abuse of notation, since we should restrict π to A′ ).
Define W ′ = E(W | A′ × A′ ).
Claim 13.4. W = W ′ almost everywhere.
It suffices to show that
∫ ∫
(13.2) W dπ × dπ = W ′ dπ × dπ
A×B A×B

for all A, B ∈ A. (Note that this holds for A, B ∈ A′ by the definition of conditional
probability.) Consider the functions
∫ ∫
UA = W (., y) dπ(y), gA = E(1A | A′ ), VA = W (., y)gA (y) dπ(y).
A Ω
′
These functions are A -measurable by the definition of twins. Furthermore, gA
is the orthogonal projection of 1A into the space of A′ -measurable functions,
gAB (x, y) = gA (x)gB (y) is the orthogonal projection of 1A×B into the space of
A′ × A′ -measurable functions, and by definition, W ′ is the orthogonal projection
of W into this space. Using these observations, we have
∫
W dπ × dπ = ⟨1B , UA ⟩ = ⟨gB , UA ⟩ = ⟨VB , 1A ⟩ = ⟨VB , gA ⟩
A×B
∫
= ⟨W, gAB ⟩ = ⟨W , gAB ⟩ = ⟨W , 1A×B ⟩ =
′ ′
W ′ dπ × dπ.
A×B

This proves (13.2) and the Claim.

It seems that instead of eliminating twin points, we made them even more
“twin-like”: they are now not separated by any set from A′ . But this is good,
because then identifying them does not change anything: Let Ω1 denote the set
of equivalence classes of “being twins” on Ω, let φ(x) be the equivalence ( class)
containing x ∈ Ω, let A1 = {φ(X) : X ∈ A′ }, and define π1 (X) = π φ−1 (X)
for X ∈ A1 . Then J1 = (Ω1 , A1 , π1 ) is a probability space. Furthermore, W ′
is constant on S × T for two equivalence classes, and hence W1 (S, T ) = W ′ (x, y)
(x ∈ S, y ∈ T ) is well defined. Trivially, W1φ = W ′ , so by Claim 13.4, we have
W1φ = W almost everywhere.
For the proof of the second statement of the Proposition, we need the following.
Claim 13.5. If A is countably generated, then A1 is countably separated.
220 13. ON THE STRUCTURE OF GRAPHONS

Let R be a countable generating set in A. For every R ∈ R and rational

number r, consider the set
{ ∫ }
SR,r = x ∈ Ω : W (x, y) dπ(y) ≥ r .
R

Clearly SR,r ∈ A′ . Furthermore, if x and x′ are not twins, then W (x, .) and
′
W
∫ (x , .) differ on a ∫set of positive
′
∫ a set R ∈ R
measure, and so there is ∫ such that
R
W (x, y) dπ(y) ̸
= R
W (x , y) dπ(y).
∫ Assume that (say)
∫ R
W (x, .) > R
W (x′ , .),
′
then for any rational number r with R W (x, .) > r > R W (x , .) we have x ∈ SR,r
but x′ ∈ / SR,r . So the countable family of sets SR,r separates any two points of Ω
that are not twins. It follows that the sets φ(SR,r ) ∈ A1 separate any two points
of Ω1 .
Claim 13.5 and Proposition A.4 in the Appendix complete the proof. 

Exercise 13.6. Show that for the set J and function W in Example 13.1, several
different measures on J can yield the same—isomorphic—graphons.
Exercise 13.7. Consider two graphons U = (Ω, A, π, W ) and U ′ = (Ω, A, π ′ , W )
which only differ in their probability measures. Prove that δ1 (U, U ′ ) ≤
2dvar (π, π ′ ). [Hint: use Exercise 8.15.]
Exercise 13.8. Suppose that a graphon (Ω, A, π, W ) is defined on a metric space
(Ω, d), where A is the set of Borel sets, π is atom-free, and W is almost everywhere
continuous. Suppose that the sequence Sn ⊆ J is well distributed in the sense
that |Sn ∩ U |/|Sn | → π(U ) for every open set U . Then t(F, G(Sn , W )) → t(F, W )
for every simple graph F with probability 1.

13.2. Weak isomorphism III

We give a characterization of weakly isomorphic kernels first in the twin-free
case, and then use this to complete our arsenal of characterizations in general.
Theorem 13.9. If two standard twin-free kernels are weakly isomorphic, then they
are isomorphic up to a nullset.
Proof. Let (J1 , W1 ) and (J2 , W2 ) be two weakly isomorphic twin-free kernels,
where Ji = (Ωi , Ai , πi ) (i = 1, 2) are standard probability spaces. By Corollary
10.35, there is a third kernel (J, W ) (J = (Ω, A, φ)) and measure preserving maps
φi : Ji → J such that Wi = W φi almost everywhere.
Let Ω′i be the set of all elements u ∈ Ωi for which Wi (u, v) = W φi (u, v) for
almost all v ∈ Ωi . By the definition of W and φi , we must have πi (Ωi \ Ω′i ) = 0.
So we can delete the elements of Ωi \ Ω′i from Ωi for i = 1, 2. In other words, we
may assume that for every u ∈ Ωi , Wi (u, v) = W φi (u, v) for almost all v ∈ Ωi .
This implies that φi is injective. Indeed, the set φ−1 i (x) consists of twins in
the kernel (Ji , W φi ), which remain twins in (Ji , Wi ). Since (Ji , Wi ) is twin-free, it
follows that φ−1i (x) has only one element.
An injective measure preserving map from a standard probability space into
another one is almost bijective: the set of points with no inverse image has measure
0 (Proposition A.4 in the Appendix). Hence Ω∗ = φ1 (Ω1 ) ∩ φ2 (Ω2 ), Ω∗1 = φ−1 ∗
1 (Ω )
∗ −1 ∗
and Ω2 = φ2 (Ω ) have measure 1 in the corresponding graphons, and we can
restrict these kernels to them. But then φ1 and φ2 are isomorphisms between these
three kernels. 
 13.2. WEAK ISOMORPHISM III 221

This theorem adds a further characterization of weak isomorphism of standard

kernels, formulated before. The following theorem summarizes these characteriza-
tions: Corollaries 10.34, 10.35 and 10.36 established the equivalence of (a) with (c),
(e), (f) and (b); Proposition 13.3 and Theorem 13.9 imply the nontrivial part of
(d).
Theorem 13.10. For two standard kernels (J1 , W1 ) and (J2 , W2 ) the following are
equivalent:
(a) t(F, W1 ) = t(F, W2 ) for every simple graph F (i.e., (J1 , W1 ) and (J2 , W2 )
are weakly isomorphic);
(b) t(F, W1 ) = t(F, W2 ) for every loopless multigraph F ;
(c) δN (W1 , W2 ) = 0 for every (or just for one) smooth invariant norm;
(d) there exist a standard kernel (J0 , W0 ) and measure preserving maps
φi : J0 → Ji such that Wi = W0φi almost everywhere;
(e) there exist a standard kernel (J0 , W0 ) and measure preserving maps
φi : Ji → J0 such that Wiφi = W0 almost everywhere;
(f) there exists a coupling measure µ between J1 and J2 such that W1 (X1 , Y1 ) =
W2 (X2 , Y2 ) for almost all pairs (X1 , X2 ) and (Y1 , Y2 ) selected independently from
the distribution µ. 
There are (at least) four quite different ways to prove Theorem 13.10 (not
counting trivial differences like the order in which various conditions are proved).
• The main steps in the route followed here was to establish the equivalence of (a)
and (c) (which was proved by Borgs, Chayes, Lovász, Sós and Vesztergombi [2008])
and then use Theorem 8.13 (due to Bollobás and Riordan [2009]).
• The original proof by Borgs, Chayes and Lovász [2010] is more direct but a
lot longer; it is built on the natural idea to bring every kernel to a “canonical
form”, so that weakly isomorphic kernels would have identical canonical forms. In
the case of functions in a single variable, a canonical form that works in many
situations is “monotonization” (see the Monotone Reordering Theorem A.19 in
the Appendix). For kernels there does not seem to exist such a canonical form,
but one can construct, for every kernel W , a “canonical ensemble”: a probability
distribution πW on the set of kernels such that two kernels U and W are weakly
isomorphic if and only if the distributions πU and πW are identical.
• Diaconis and Janson [2008] showed that theorem 13.10 also follows from results
of Kallenberg [2005] in the theory of exchangeable random variables [2005].
• The proof of Janson [2010] is based on the idea of pure kernels (see Section 13.3)
and Theorem 13.9.
Exercise 13.11. Let J = (Ω, A, π) be a probability space and W : Ω × Ω → R,
a symmetric function measurable with respect to the completion of A × A. Show
that W can be changed on a set of measure 0 so that it becomes measurable with
respect to A × A.
Exercise 13.12. Let (J, W ) be a kernel on a (non-standard) probability space
J = (Ω, A, π). Show that there is a countably generated σ-algebra A0 ⊆ A on Ω
such that W is measurable with respect to the completion of A0 × A0 .
Exercise 13.13. Let (J, W ) be a kernel on a (non-standard) countably gener-
ated separating probability space J = (Ω, A, π). Show that there is a standard
probability space J0 = (Ω0 , A0 , π0 ), a kernel (J0 , W0 ), and an injective measure
preserving map φ : J → J0 such that W = W0φ almost everywhere.
222 13. ON THE STRUCTURE OF GRAPHONS

Exercise 13.14. Consider two probability spaces (Ω, A, π) and (Ω′ , A′ , π ′ ), a

kernel (Ω, A, π, W ), and a measure preserving map φ : Ω → Ω′ (note: it goes in
the opposite direction than in the definition of the pullback!). (a) Show that one
can construct a “push-forward” kernel (Ω′ , A′ , π ′ , Wφ such that
( )
(Wφ )φ = E W | φ−1 (A′ ) × φ−1 (A′ ) .
(b) show by an example that t(F, Wφ ) = t(F, W ) does not hold in general.

13.3. Pure kernels

In the spirit of classical analysis, there is no room to further standardize a
kernel: We have seen that every kernel is equivalent to a twin-free kernel, and two
weakly isomorphic twin-free kernels are isomorphic up to a nullset, and usually we
don’t care about nullsets. But it turns out that cleaning up these nullsets is worth
the trouble.

13.3.1. Purifying kernels. We introduce a distance notion on the points of

a kernel. Let (J, W ) be a kernel. We can endow the space J with the distance
function
∫
rW (x, y) = ∥W (x, .) − W (y, .)∥1 = |W (x, z) − W (y, z)| dz.
J

This function is defined for almost all pairs x, y; we can delete those points from
J where W (x, .) ∈/ L1 (J) (a set of measure 0), to have rW defined on all pairs.
It is clear that rW is a pseudometric (it is symmetric and satisfies the triangle
inequality). We call rW the neighborhood distance on W .
Example 13.15 (Stepfunctions). For stepfunctions, the underlying metric space
is finite. 
Example 13.16 (Spherical distance). Let S d denote the unit sphere in Rd+1 ,
consider the uniform probability measure on it, and let W (x, y) = 1 if x · y ≥ 0
and W (x, y) = 0 otherwise. Then (S d , W ) is a graphon, in which the neighborhood
distance of two points a, b ∈ S d is just their spherical distance (normalized by
dividing by π). 
Example 13.17. Let (M, d) be a metric space, and let π be a Borel probability
measure on M . Then d can be viewed as a kernel on (M, d). For x, y ∈ M , we have
∫ ∫
rd (x, y) = |d(x, z) − d(y, z)| dπ(z) ≤ d(x, y) dπ(z) = d(x, y),
M M

so the identity map (M, d) → (M, rd ) is contractive. This implies that if (M, d)
is compact, and/or finite dimensional (in many senses of dimension), then so is
(M, rd ). For most ”everyday” metric (spaces )(like segments, spheres, or balls)
rd (x, y) can be bounded from below by Ω d(x, y) , in which case (M, d) and (M, rd )
are homeomorphic.
( More)generally, if F : [0, 1] → R is a continuous function, then W (x, y) =
F d(x, y) defines a kernel, and the identity map (M, d) → (M, rW ) is continuous.

A kernel (J, W ) is pure if (J, rW ) is a complete separable metric space and the
probability measure has full support (i.e., every open set has positive measure).
 13.3. PURE KERNELS 223

This definition includes that rW (x, y) is defined for all x, y ∈ J and rW (x, y) > 0 if
x ̸= y, i.e., the kernel has no twins.
Theorem 13.18. Every twin-free kernel is isomorphic, up to a null set, to a pure
kernel.
Proof. Let (J, W ) be a twin-free kernel. Let T be the set of functions f ∈
L1 (J) such that for every L1 -neighborhood U of f , the set {x ∈ J : W (x, .) ∈ U }
has positive measure. Clearly T is a closed subset of L1 (J), and it is complete and
separable in the L1 -metric. Let J ′ be the set of points in J for which W (x, .) ∈ T ,
and let T ′ = {W (x, .) : x ∈ J ′ }. The map φ : J ′ → T ′ defined by x 7→ W (x, .)
is bijective, since (J, W ) is twin-free. The set T inherits a probability measure
π ′ = π ◦ φ−1 from J. It is easy to see from the construction that J ′ and T ′ are
measurable.
We claim that
(13.3) π(J \ J ′ ) = 0.
It is clear that for almost all x ∈ J, W (x, .) ∈ L1 (J). Every function g ∈ L1 (J) \ T
has an open∪ neighborhood Ug in L1 (J) such that π{x ∈ J : W (x, .) ∈ Ug } = 0.
Let U = g∈T / Ug . Since L1 (J) is separable, U equals the union of some countable
subfamily {Ugi : i ∈ N} and thus π{x ∈ J : W (x, .) ∈ U } = 0. Since J \ J ′ ⊆ U ,
this proves (13.3).
The functions W (x, .) (x ∈ J ′ ) are everywhere dense in T and have measure
1. So T is a complete separable metric space with a probability measure on its
Borel sets. It also follows from the definition of T that every open set has positive
measure, and (13.3) implies that π ′ (T \ T ′ ) = 0.
We define a kernel W ′ : T × T → [0, 1] as follows. Let f, g ∈ T . If f ∈ T ′ , then
f = W (x, .) for some x ∈ J, and we define W ′ (f, g) = g(x). Similarly, if g ∈ T ′ ,
then g = W (., y) and we define W ′ (f, g) = f (y). Note that if both f, g ∈ T ′ , then
this definition is consistent: W ′ (f, g) = f (y) = g(x) = W (x, y). If f, g ∈ / T ′ , then
′
we define W (f, g) = 0. We note that f and g are determined up to a zero set only;
we can choose any function representing them, and since we are changing W on a
set of measure 0 only, it remains measurable.
The kernel (T, W ′ ) is pure; indeed, we just have to check that rW ′ coincides
with the L1 metric on T ; then T will have all the right properties. For f, g ∈ T , we
have
∫ ∫
rW ′ (f, g) = |W ′ (f, y) − W ′ (g, y)| dπ ′ (y) = |W ′ (f, y) − W ′ (g, y)| dπ ′ (y)
′
∫T ∫ T

= |f (y) − g(y)| dπ(y) = |f (y) − g(y)| dπ(y) = ∥f − g∥1 .

J′ J
′
The kernels (J, W ) and (T, W ) are isomorphic up to a 0-set; indeed, we can get the
kernel (J ′ , W |J ′ ) from (J, W ) and the kernel (T ′ , W ′ |T ′ ) from the kernel (T, W ′ )
by deleting appropriate 0-sets, and φ is an isomorphism between (J ′ , W |J ′ ) and
(T ′ , W ′ |T ′ ). This proves the theorem. 
There is still some freedom left: given a pure kernel (J, W ), we can change the
value of W on a symmetric subset of J × J that intersects every fiber J × {v} in a
set of measure 0. We can take the integral of W (which is a measure ω on J × J),
and then the derivative of ω with respect to π × π wherever this exists. This way
we get back W almost everywhere, and a well defined value for some further points.
224 13. ON THE STRUCTURE OF GRAPHONS

After all these changes, where W is left undefined is the set of “essential dis-
continuities” of W (of measure 0). It would be interesting to relate this set to
combinatorial properties of W .

13.3.2. Density functions on pure kernels. Now we come to the utiliza-

tion of our work with purifying kernels. The following technical lemma will be very
useful in the study of homomorphism densities on pure graphons.
Lemma 13.19. Let (J, W ) be a pure graphon and let F = (V, E) be a k-labeled
multigraph with nonadjacent labeled nodes. Then
|tx (F, W ) − tx (F, W )| ≤ e(F ) max rW (xu , x′u )
u∈[k]

′
for all x, x ∈ J . k

It follows that the functions t are Lipschitz (and hence continuous).

Proof. Let E = {u1 v1 , . . . um vm }, where we may assume that vi is unlabeled.

For each v ∈ V \ [k], let yv = xv = x′v be a variable. Then, using a telescoping sum,
∫ ∏
m ∫ ∏
m
tx (F, W ) − tx′ (F, W ) = W (xui , xvi ) dy − W (x′ui , x′vi ) dy
i=1 i=1
J V \[k] J V \[k]
∑ ∫ ∏
m
( )∏
= W (xui , xvi ) W (xuj , xvj ) − W (x′uj , x′vj ) W (x′ui , x′vi ) dy
j=1 V \[k] i<j j>i
J

and hence
∑
m ∫
|tx (F, W ) − tx′ (F, W )| ≤ |W (xuj , xvj ) − W (x′uj , x′vj )| dy.
j=1 V \[k]
J

By the assumption that vi is unlabeled, we have xvj = x′vj for every j, and so
∑
m ∫
|tx (F, W ) − tx′ (F, W )| ≤ |W (xuj , xvj ) − W (x′uj , xvj )| dy
j=1 V \[k]
J
∑m
≤ rW (xuj , x′uj ) ≤ e(F ) max rW (xu , x′u ),
u∈[k]
j=1

which proves the assertion. 

Corollary 13.20. Let (J, W ) be a pure kernel, and let F = (V, E) be a k-labeled
graph with nonadjacent labeled nodes. Then tx (F, W ) is a continuous function of
x ∈ J S with respect to the metric rW .
In the case when F is a path of length 2, we get a corollary that will be
important in the next section.
Corollary 13.21. For every pure kernel (J, W ), W ◦ W is a continuous function
(in two variables) on the metric space (J, rW ).
Most applications of Corollary 13.20 use the following consequence:
 13.4. THE TOPOLOGY OF A GRAPHON 225

Corollary 13.22. Let (J, W ) be a pure kernel, and let F1 , . . . , Fm be (k + n)-

labeled multigraphs with nonadjacent labeled nodes. Let a1 , . . . , am be real numbers
and x ∈ J k , such that the equation
∑
m
(13.4) ai tx,y (Fi , W ) = 0
i=1

holds for almost all y ∈ J k . Then it holds for all y ∈ J k .

Proof. By Corollary 13.20, the left side of (13.4) is a continuous function of
(x, y), and so it remains a continuous function of y if we fix x. Hence the set where
it is not 0 is an open subset of J k . Since the graphon is pure, it follows that this
set is either empty or has positive measure. 

Going to pure graphons is a good proof method, which can lead to nontrivial
results. We illustrate this by discussing properties of t(., W ) (W ∈ W) from the
point of view of Section 6.3.2. This parameter is multiplicative and reflection pos-
itive. If W is not a stepfunction, then t(., W ) has no contractor (else, 6.30 would
imply that it is a homomorphism function). On the other hand, it is contractible.
We prove this in a more general form.
Let F be a k-labeled multigraph, and let P = {S1 , . . . , Sm } be a partition of
[k]. We say that P is legitimate for F , if each set Si is stable in F . If this is the
case, then the m-labeled multigraph F/P (obtained by identifying the nodes in each
Si , and labeling the obtained node with i) has no loops. For a k-labeled quantum
graph g, we say that the partition P of [k] is legitimate for g if it is legitimate for
every constituent. Then we can define g/P by linear extension.
Proposition 13.23. Let g be a k-labeled quantum graph and P, a legitimate par-
tition for g. Let W ∈ W, and suppose that tx (g, W ) = 0 almost everywhere on
[0, 1]k . Then ty (g/P, W ) = 0 for almost all y ∈ [0, 1]|P| .
Proof. We may assume that W is pure. If tx (g, W ) = 0 almost everywhere,
then it holds everywhere by Corollary 13.20. In particular, it holds for every sub-
stitution where the variables corresponding to the same class of P are identified,
which means that ty (g/P, W ) is identically 0 on [0, 1]|P| . 

Corollary 13.24. The multigraph parameter t(., W ) is contractible for every kernel
W.
Exercise 13.25. Show that using properties of pure kernels, one gets a very short
proof of the statement of Exercise 7.6.

13.4. The topology of a graphon

We have seen that every graphon is weakly isomorphic to a pure graphon (which
is unique up to changing the function on special nullsets), which has an underlying
complete metric space J. So we could ask: what topological properties does J have
for special graphons, and are they related to the combinatorial properties of graph
sequences converging to W ? It turns out that these questions are more interesting
if we ask them for a different, but related topology on the underlying set, and this
is what we are going to introduce now.
226 13. ON THE STRUCTURE OF GRAPHONS

13.4.1. The similarity distance. It was noted by Lovász and Szegedy [2007,
2010b] that for a pure graphon (J, W ), the distance function rW = rW ◦W defined
by the operator square of W is also closely related to combinatorial properties of a
graphon. We call this the similarity distance. In the special case of finite graphs,
this notion was defined in the Introduction, where the motivation for its name was
also explained. We will use the graph version to design algorithms in Section 15.4.1.
In explicit terms, we have
∫ ∫ ∫ 
 
rW (a, b) = rW ◦W (a, b) =  W (a, y)W (y, x) dy − W (b, y)W (y, x) dy  dx
J J J
∫ ∫ 
 ( ) 
(13.5) =  W (x, y) W (a, y) − W (b, y) dy  dx .
J J

(We write here and in the sequel dx instead of dπ(x), where π is the probability
measure of the graphon.)
Lemma 13.26. If (J, W ) is a pure graphon, then the similarity distance rW is a
metric.
Proof. The only nontrivial part of this lemma is that rW (a, b) = 0 implies
that a = b. The condition rW (a, b) = 0 implies that for almost all x ∈ J we have
∫
( )
W (x, y) W (a, y) − W (b, y) dy = 0.
J

Using that (J, W ) is pure, Corollary 13.22 implies that this holds for every x ∈ J.
In particular, it holds for x = a and x = b. Substituting these values and taking
the difference, we get that
∫
( )2
W (a, y) − W (b, y) dy = 0,
J

and hence W (a, y) = W (b, y) for almost all z. Using again that (J, W ) is pure, we
conclude that a = b. 

So (J, rW ) is a metric space, and hence Hausdorff. We have to be careful

though, since the metric space (J, rW ) is not necessarily complete. We will work
with its completion (J, rW ), but first we define it differently.
Let us say that a sequence of points xn ∈ J is weakly convergent if
∫ ∫
W (xn , y) dy → W (x, y) dy
A A

for every measurable set A ⊆ J. We call this the weak topology on J. (We need this
name only temporarily, since we are going to show that rW gives a metrization of
the weak topology.) It is well known that this topology is metrizable.
Let J denote the completion of J in the weak topology. The map x 7→ W (x, .)
embeds J into L1 (J), and weak convergence corresponds to weak* convergence of
functions in L1 (J). Hence J corresponds to the weak* closure of J. It follows
in particular that J is a compact separable metric space (compactness follows by
Aleoglu’s Theorem, since J is a closed subset of the unit ball of L1 (J)).
 13.4. THE TOPOLOGY OF A GRAPHON 227

We can extend the definition of W to J × J: for x ∈ J and y ∈ J \ J, we define

W (x, y) = x(y) (recall that the elements of J can be identified with functions on
J), and we define W (x, y) = 0 if x, y ∈ J \ J. We also extend the measure π by
π(J \ J) = 0.
Theorem 13.27. For any pure graphon, the metric rW defines exactly the weak
topology on J.
Proof. First we show that the weak topology is finer than the topology of
(J, rW ). Suppose that xn → x in the weak topology, this means that the functions
W (xn , .) converge weakly to W (x, .). Consider
∫ ∫ 
 ( ) 
rW (xn , x) =  W (xn , y) − W (x, y) W (y, z) dy  dz.
J J

Here the inner integral tends to 0 for every z, by the weak convergence xn → x.
Since it also remains bounded, it follows that the outer integral tends to 0. This
implies that xn → x in (J, rW ). (Let us note that since π(J \ J) = 0, it does not
matter whether we integrate over J or over J.)
From here, the equality of the two topologies follows by general arguments: the
weak topology on J is compact, and the coarser topology of rW is Hausdorff, which
implies that they are the same. 
Corollary 13.28. For every pure graphon (J, W ), the space (J, rW ) is compact.
Another useful corollary of these considerations concerns continuity in the sim-
ilarity metric. We have seen that W ◦ W is continuous in the metric rW ; one might
hope that the function W is continuous as a function on (J, rW ), but this would be
too much to ask for (the half-graphon is an easy example). However, integrating
out one of the variables we get a continuous function. To be more precise (and
more general):
Corollary 13.29. For every pure graphon (J, W ), and every function g ∈ L1 (J),
the function ∫
(TW g)(.) = W (., y)g(y) dy
J
is continuous on (J, rW ).
In particular, it follows that every eigenfunction of TW is continuous on (J, rW ).
Proof. Let xn → x in the rW metric. Then W (xn , .) → W (x, .) in the weak
topology by Theorem 13.27, and hence
∫ ∫
W (Xn , y)g(y) dy → W (x, y)g(y) dy. 
J J

We conclude this section with an example in which the topologies defined by

rW and rW are different. Note that for any two points x, y ∈ J, we have
(13.6) rW (x, y) ≤ rW (x, y),
which implies that the topology of (J, rW ) is finer than the topology of (J, rW ).
The two topologies may be different. Graphons for which the finer space (J, rW ) is
also compact seem to have special importance in combinatorics.
228 13. ON THE STRUCTURE OF GRAPHONS

Example 13.30. For y ∈ [0, 1), let y = 0.y1 y2 . . . be the binary expansion of y. Let
us decompose [0, 1) into the intervals Ik = [1 − 2−k , 1 − 2−k−1 ). Define U (x, y) = yk
for 0 ≤ y ≤ 1 and x ∈ Ik . Define W (1, y) = 1/2 for all y. This function is not
symmetric, so we put it together with a reflected copy to get a graphon:


U (2x, 2y − 1), if x ≤ 1/2 and y ≥ 1/2,
W (x, y) = U (2y, 2x − 1), if x ≥ 1/2 and y ≤ 1/2,


0, otherwise.
(This is rather difficult to parse, but it is an important example. Perhaps Figure
13.1 helps.) Selecting one point from each interval [1 − 2−k , 1 − 2−k−1 ), we get an
infinite number of points in [0, 1) mutually at rW -distance 1/4; so this sequence is
not convergent even in the completion of this graphon. (In particular, (J, rW ) is
not compact.) On the other hand, this same sequence converges in (J, rW ). So the
two topologies are different. 

Figure 13.1. A graphon defined on a space whose completion in

rW is not compact. The picture on the left shows just one half.

13.4.2. Similarity distance and regularity partitions. Now we come to

the results that first motivated the introduction of a second distance notion defined
by a graphon.
Let (J, d) be a metric space and let π be a probability
∫ measure on its Borel
sets. We say that a set S ⊆ J is an average ε-net, if J d(x, S) dπ(x) ≤ ε.
Let S ⊆ J be a finite set and s ∈ S. The Voronoi cell of S with center s is the
set of all points x ∈ J for which d(x, s) ≤ d(x, y) for all y ∈ S. Clearly, the Voronoi
cells of S cover J. (We can break ties arbitrarily to get a partition.)
Theorem 13.31. Let (J, W ) be a pure graphon, and ε > 0.
(a) Let S be an average ε-net in the metric space (S, rW ). Then √ the Voronoi
cells of S form a weak regularity partition P with error at most 8 ε.
(b) Let P = {J1 , . . . , Jk } be a weak regularity partition with error ε. Then there
are points vi ∈ Ji such that the set S = {v1 , . . . , vk } is an average (4ε)-net in the
metric space (S, rW ).
Proof. (a) Let P be the partition into the√Voronoi cells of S. Let us write
R = W − WP . We want to show that ∥R∥ ≤ 8 ε. It suffices to show that for any
0-1 valued function f ,
√
(13.7) ⟨f, Rf ⟩ ≤ 2 ε.
 13.4. THE TOPOLOGY OF A GRAPHON 229

Let us write g = f − fP , where fP (x) is obtained by replacing f (x) by the average

of f over the class of P containing x. Clearly ⟨fP , RfP ⟩ = 0, and so
(13.8) ⟨f, Rf ⟩ = ⟨g, Rf ⟩ + ⟨fP , Rf ⟩ = ⟨f, Rg⟩ + ⟨fP , Rg⟩ ≤ 2∥Rg∥1 ≤ 2∥Rg∥2 .
For each x ∈ J, let (φ(x) ∈ )S be the center of the Voronoi
( cell )containing x, and
define W ′ (x, y) = W x, φ(y) and similarly R′ (x, y) = R x, φ(y) . Then using that
(W − R)g = WP g = 0, W − W ′ = R − R′ and R′ g = 0, we get
∥Rg∥22 = ⟨Rg, Rg⟩ = ⟨W g, (R − R′ )g⟩ = ⟨W g, (W − W ′ )g⟩ = ⟨g, W (W − W ′ )g⟩
∫ ∫
′  ( ) 
≤ ∥W (W − W )∥1 =  W (x, y) W (y, z) − W (y, φ(z)) dy  dx dz
J2 J
∫
( ) ( )
= rW z, φ(z) = Ex (rW x, S) ≤ ε.
J

This proves (13.7).

(b) Suppose that P is a weak regularity partition with error ε. Let R = W −WP ,
then we know that ∥R∥ ≤ ε. For every x ∈ [0, 1], define
∫ ∫  ∫
 
F (x) =  R(x, y)W (y, z) dy  dz = s(x, z)R(x, y)W (y, z) dy dz,
J J J2
∫
where s(x, z) is the sign of R(x, y)W (y, z) dy. Lemma 8.10 implies that for every
z ∈ J,
∫
s(x, z)R(x, y)W (y, z) dx dy ≤ 2∥R∥ ≤ 2ε,
J2

and hence
∫
(13.9) F (x) dx ≤ 2ε.
J

Let x, y ∈ J be two points in the same partition class of P. Then WP (x, s) =

WP (y, s) for every s ∈ J, and hence
∫ ∫ 
 ( ) 
(13.10) rW (x, y) =  W (x, s) − W (y, s) W (s, z) ds dz
J J
∫ ∫ 
 ( ) 
=  R(x, s) − R(y, s) W (s, z) ds dz

J J
∫ ∫  ∫ ∫ 
   
≤  R(x, s)W (s, z) ds dz +  R(y, s)W (s, z) ds dz
 
J J J J
= F (x) + F (y).
For every set T ∈ P, let vT ∈ T be a point “below average” in the sense that
∫
1
F (vT ) ≤ F (x) dx,
π(T )
T
230 13. ON THE STRUCTURE OF GRAPHONS

and let S = {vT : T ∈ P}. Then using (13.9),

∑∫ ∑∫ ( )
Ex rW (x, S) ≤ rW (x, vT ) dx ≤ F (x) + F (vT ) dx
T ∈P T T ∈P T
∫ ∑ ∫
≤ F (x) dx + λ(T )F (vT ) ≤ 2 F (x) dx ≤ 4ε. 
J T ∈P J

By the Weak Regularity Lemma, we get that every graphon has an average
2
ε-net of size 2O(1/ε ) . How about a “true” ε-net? By a standard trick, if we take a
maximum set R of points such that any two are at a distance at least ε, then every
point is at a distance of at most ε from R. But is such a set necessarily finite? And
if so, how can we bound its size?
It turns out that one can give a bound that is similar to the bound on the size
of an average ε-net derived from Theorem 13.31. The following result is due to
Alon [unpublished].
Proposition 13.32. Let (J, W ) be a graphon and let R ⊆ J be a set such that
2
rW (s, t) ≥ ε for all s, t ∈ R (s ̸= t). Then |R| ≤ (16/ε2 )257/ε .
The bound on the size of R is somewhat worse than for the average ε-net, but
the main point is that it depends on ε only. There are examples showing that an
exponential dependence on 1/ε is unavoidable (Exercise 13.41).

Proof. Consider the Frieze–Kannan decomposition of W provided by Lemma

9.14:
∑
k
(13.11) W = ai 1Si ×Ti + U,
i=1
∑ 2
√ k = ⌈256/ε ⌉, the sets Si , Ti ⊆ J are measurable, i ai ≤ 4, and ∥U ∥ ≤
2
where
4/ k ≤ ε/4. For s, t ∈ R, we have
∫ ∫ 
 ( ) 
rW (s, t) =  W (s, y) − W (t, y) W (y, z) dy  dz
J J
∫
( )
= σst (z) W (s, y) − W (t, y) W (y, z) dy dz,
J×J
∫( )
where σst (z) is the sign of W (s, y) − W (t, y) W (y, z) dy. Substituting from
(13.11) for the last occurrence of W , we get
∑ ∫
k
( )
rW (s, t) = ai σst (z) W (s, y) − W (t, y) 1Si ×Ti (y, z) dy dz
i=1 J×J
∫
( )
(13.12) + σst (z) W (s, y) − W (t, y) U (y, z) dy dz.
J×J

Here the last term is small by Lemma 8.10 and the choice of U :
∫ ( ) 
  ε
 σst (z) W (s, y) − W (t, y) U (y, z) dy dz  ≤ 2∥U ∥ ≤ .
2
J×J
 13.4. THE TOPOLOGY OF A GRAPHON 231

To bound a term from the first sum, we do a little computation:

∫ ∫ (∫ )(∫ )
( )
σst (z) W (s, y) − W (t, y) dy dz = σst (z) dz W (s, y) − W (t, y) dy ,
Ti Si Ti Si
∫
and hence, with fi (s) = Si
W (s, y) dy, we have
∫ ∫ ( ) 
 
 σst (z) W (s, y) − W (t, y) dy dz 
Ti Si
∫ 
 
≤  W (s, y) − W (t, y) dy  = |fi (s) − fi (t)|.
Si
2 √
Now if |R| > (16/ε2 )257/ε √> (4 k/ε)k , then there is a pair of points s, t ∈ R such
that |fi (s) − fi (t)| ≤ ε/(4 k) for all i, and for this choice of s and t we have

∑ ∫ ∑
k
( ) k
ε ε
ai σst (z) W (s, y) − W (t, y) 1Si ×Ti (y, z) dy dz < |ai | √ ≤ .
i=1 i=1
4 k 2
J×J
∑
(In the last step we used that i a2i ≤ 4 and the inequality between arithmetic and
quadratic means.) By (13.12) this implies that rW (s, t) < ε, a contradiction. 

13.4.3. Finite dimensional graphons. The main reason to be interested in

this topology is the following consequence of Theorem 13.31. We define the (upper)
Minkowski dimension of a metric space (M, d) as
log N (ε)
lim sup ,
ε→0 log(1/ε)
where N (ε) is the maximum number of points in M mutually at distance at least ε.
This dimension is finite if and only if there is a d ≥ 0 such that every set of points
mutually at distance at least ε has at most ε−d elements.
Corollary 13.33. If (J, W ) is a graphon for which the space (J, rW ) has finite
Minkowski dimension
( d,
) then for every ε > 0 the graphon has a weak regularity
partition with O (1/ε)d classes. 
Which graphons are finite dimensional in this sense? Almost all of those we
have met so far are 1 or at most 2-dimensional (see Exercise 13.44). We will formu-
late some conjectures later, in Section 16.7.1; right now we describe an interesting
combinatorially defined class with this property.
We say that a graphon W misses a signed graph F if t(F, W ) = 0. Trivially, if
W misses F , then the complementary graphon 1 − W misses the signed graph F −
obtained from F by negating the signs of the edges. We will only consider bipartite
graphs F ; extension of these results to non-bipartite graphs is open.
Graphons missing a signed bipartite graph can be characterized in terms of the
Vapnik–Chervonenkis dimension (see Appendix A.6 for basic information about the
VC-dimension). This characterization is not much more than a reformulation of
the definition, but useful nonetheless.
232 13. ON THE STRUCTURE OF GRAPHONS

Proposition 13.34. A pure graphon (J, W ) misses some signed bipartite graph
with k nodes in the smaller bipartition class if and only if W is 0-1 valued al-
most everywhere
{ ( } VC-dimension of the family of neighborhoods RW =
) and the
supp W (x, .) : x ∈ J is less than k.
Proof. First, suppose that (J, W ) misses a signed bipartite graph F with bi-
partition V1 ∪ V2 , where V1 = [k] and V2 = {1′ , . . . , m′ }. We start with showing
that W is 0-1 valued almost everywhere. Let F • be obtained by labeling all nodes
of V1 . Then for almost all x ∈ J k , we have tx (F • , W ) = 0. By Corollary 13.22, it
follows that tx (F • , W ) = 0 for every x ∈ J k . In particular, tz...z (F • , W ) = 0 for all
z ∈ J. But for this substitution,
∫ ∏m
( )d− (j)
tz...z (F • , W ) =
+
W (z, yj )d (j) 1 − W (z, yj ) dy1 . . . dym
J m j=1
+ −
(where d (j) and d (j) are the numbers of positive and negative edges of F incident
with j, respectively). If there is a z ∈ J such that 0 < W (y, z) < 1 for all y ∈ Y ,
where Y has positive measure, then the part of the integral over Y m is already
positive, so tz...z (F • , W ) > 0, a contradiction.
Next, we show that the VC-dimension of RW is less than k. Suppose not,
then
{ there
( is a) set S = {x } 1 , . . . , xk } ⊆ J with |S| = k such that the family H =
supp W (x, .) : x ∈ S is qualitatively independent (this means that for every
H′ ⊆ H there is a point contained in all sets of H′ but in no set of H \ H′ ). This
implies that tx1 ...xk (F • , W ) > 0. By the purity of (J, W ) and Corollary 13.22,
the set of points (y1 , . . . , yk ) ∈ [0, 1]k for which ty1 ...yk (F • , W ) > 0 has positive
measure. Hence t(F, W ) > 0.
Conversely, suppose that W is 0-1 valued (we may assume everywhere), and
dimV C (RW ) < k. Let F denote the signed complete bipartite graph with k nodes
in one class U and 2k nodes in the other class U ′ , in which each node in U ′ is
connected to a different set of nodes in U by positive edges. Let F • be obtained by
labeling the nodes in U . Then any choice of x1 , . . . , xk for which tx1 ...xk (F • , W ) > 0
gives k points with qualitatively independent neighborhoods, which is impossible.
So we must have t(F, W ) = 0. 
Our main goal is to connect the VC-dimension of neighborhoods to the dimen-
sion of J. The following theorem was proved (in a slightly more general form) by
Lovász and Szegedy [2010b].
Theorem 13.35. If a pure graphon (J, W ) misses some signed bipartite graph F ,
then
(a) W is 0-1 valued almost everywhere,
(b) (J, rW ) is compact, and
(c) it has Minkowski dimension at most 10v(F ).
Proof. (a) is just repeated from Proposition 13.34. To prove (b), we start with
studying weakly convergent sequences of functions W (x, .). Let (x1 , x2 , . . . ) be a
sequence of points in J and suppose that there is a function f ∈ L1 (J) such that
∫ ∫
W (xn , y) dy −→ f (y) dy
S S
for every measurable set S ⊆ J.
 13.4. THE TOPOLOGY OF A GRAPHON 233

Claim 13.36. The weak limit function f is almost everywhere 0-1 valued.
Suppose not, then there is an ε > 0 and a set( Y ⊆ J2 with
) positive measure such
that ε ≤ f (x) ≤ 1 − ε for x ∈ Y . Let Sn = supp W (xn , .) ∩ Y . We select, for every
k ≥ 1, k indices n1 , . . . nk so that the Boolean algebra generated by Sn1 , . . . Snk (as
subsets of Y ) has 2k atoms of positive measure. If we have this for some k, then
for every atom A of the Boolean algebra
∫ ∫
λ(A ∩ Sn ) = W (x, yn ) dx −→ f (x) dx (n → ∞),
A A
and so if n is large enough, then
ε ( ε)
λ(A) ≤ λ(A ∩ Sn ) ≤ 1 − λ(A).
2 2
If n is large enough, then this holds for all atoms A, and so Sn cuts every previous
atom into two sets with positive measure, and we can choose nk+1 = n.
But this means that the VC-dimension of the supports of the W (x, .) is infinite,
contradicting Proposition 13.34. This proves Claim 13.36.
Claim 13.37. The convergence W (xn , .) → f also holds in L1 .
Indeed, we know that f (x) ∈ {0, 1} for almost all x, and hence
∫ ∫
( )
∥f − W (xn , .)∥1 = 1 − W (xn , y) dy + W (xn , y) dy −→ 0.
{f =1} {f =0}

Now it is easy to prove that (J, rW ) is compact. Consider any infinite sequence
(x1 , x2 , . . . ) of points of J. By Alaoglu’s Theorem, this has a subsequence for which
the functions W (xn , .) converge weakly to a function f ∈ L1 (J). By Claim 13.37,
they converge to f in L1 . This implies that they form a Cauchy sequence in L1 ,
and so (x1 , x2 , . . . ) is a Cauchy sequence in (J, rW ). Since (J, rW ) is a complete
metric space, this sequence has a limit in J.
To prove (c), let F be a signed bipartite graph such that t(F, W ) = 0, and let
(V1 , V2 ) be a bipartition of F with |V1 | = k, where we may assume that k ≤ v(F )/2.
We may assume that F is complete bipartite, since adding edges (with any signs)
does not change the condition that t(F, W ) = 0. Let F • be obtained from F by
labeling the nodes in V1 .
We want to show that the Minkowski dimension of (J, rW ) is at most 20k. It
suffices to show that every finite set Z ⊆ J such that the rW{-distance ( of any
) two
−20k
elements
} is at least ε, is bounded by |Z| ≤ c(k)ε . Let H = supp W (x, .) : x∈
Z . Since W is 0-1 valued, the condition on Z means that
(13.13) π(X△Y ) ≥ ε
for any two distinct sets X, Y ∈ H.
We do a little clean-up: Let A be the union of all atoms of the set algebra
generated by H that have measure 0. Clearly A itself has measure 0, and hence the
family H′ = {X \ A : X ∈ H} still has property (13.13).
We claim that H′ has VC-dimension less than k. Indeed, suppose that J \ A
contains a shattered k-set S. To each j ∈ V1 , we assign a point
( qj ∈) S bijectively.
To each i ∈ V2 , we assign a point pi ∈ Z such that qj ∈ supp W (pi , .) if and only if
ij ∈ E + . (This is possible since S is shattered.) Now fixing the pi , for each j there
234 13. ON THE STRUCTURE OF GRAPHONS

is a subset of J of positive measure whose points are contained in exactly the same
members of H′ as qj , since qj ∈ / A. This means that the function tx1 ...xk (F • , W ) is
positive for xi = pi . Corollary 13.22 implies that tx1 ...xk (F • , W ) > 0 for a positive
fraction of the choices of x1 , . . . xk ∈ J, and hence t(F, W ) > 0, a contradiction.
Applying Proposition A.30 we conclude that |Z| = |H| ≤ (80k)10k ε−20k . This
proves that the Minkowski dimension of (J, rW ) is bounded by 20k. 
The results in this section do not remain true if the signed graph we exclude
is nonbipartite. For example, if we exclude any non-bipartite graph, then any
bipartite graph satisfies the condition, but some bipartite graphs are known to
need an exponential (in 1/ε) number of classes in their weak regularity partitions.
Exercise 13.38. Two metrics d1 and d2 on the same set are called uniformly
equivalent, if( there is) a function f : R+( → R+ )such that f (x) ↘ 0 if x ↘ 0,
d1 (x, y) ≤ f d2 (x, y) and d2 (x, y) ≤ f d1 (x, y) . Prove that for a pure kernel
(J, W ), the space (J, rW ) is compact if and only if the metrics rW are rW are
uniformly equivalent.
Exercise 13.39. Figure out the completions of the spaces ([0, 1], rW ) and
([0, 1], rW ) for the graphon W in Example 13.30.
Exercise 13.40. For the graphon (S d , W ) defined in Example √ 13.16, show that
the similarity distance of two points a, b ∈ S d is Ω(](a, b)/ d).
Exercise 13.41. Show that the graphon in the previous exercise, with an ap-
2
propriate choice of d, contains 2Ω(1/ε ) points mutually at least ε apart in the
similarity distance.
Exercise 13.42. Let W be a graphon such that (J, rW ) can be covered by m
balls of radius ε. Prove that there exists a stepfunction U with m(1/ε)m steps
such that ∥W − U ∥1 ≤ 2ε.
Exercise 13.43. Let M (ε) denote the minimum number of sets of diameter at
most ε covering( a metric
) space
( ) and define the covering dimension of (S, d)
(S, d),
by lim supε→0 log M (ε) / log(1/ε) . Prove that this is the same as the Minkowski
dimension.
Exercise 13.44. (a) Check that all graphons constructed in Section 11.4.2 are
at most 2-dimensional. (b) Prove that a graphon W on [0, 1] that is a continuous
function is at most 1-dimensional. (c) Find the dimension of the graphon in
Example 13.16. (d) Construct an infinite dimensional graphon.
Exercise 13.45. Let W be a graphon such that t(F, W ) = 0 for a signed bipartite
graph F = (V, E). Prove that for every 0 < ε < 1, there exists a 0-1 valued
2
stepfunction U with O(ε−10v(F ) ) steps such that ∥W − U ∥1 ≤ ε.

13.5. Symmetries of graphons

An automorphism of a graphon W on [0, 1] is an invertible measure preserving
map σ : [0, 1] → [0, 1] such that W σ = W almost everywhere. Clearly, the
automorphisms of W form a group Aut(W ). An example with many automorphisms
is a setfunction W : here Aut(W ) contains the group of all invertible measure
preserving transformations that leave the steps invariant, and it contains all the
automorphisms of the corresponding weighted graph. Note, however, that if we
purify a stepfunction, then we get a finite weighted graph, so the large and ugly
subgroups consisting of measure preserving transformations of the steps disappear.
We can endow Aut(W ) with the topology of pointwise convergence in the rW
metric. Szegedy observed that if W is pure, then Aut(W ) is compact in this topol-
ogy. This follows from the facts that every automorphism of a graphon is an
 13.5. SYMMETRIES OF GRAPHONS 235

isometry of the compact metric space (J, rW ) (this is trivial), and those isometries
that correspond to automorphisms form a closed subgroup (this takes some work
to prove; see Lovász [Notes]).
We will not go into the detailed study of Aut(W ) in this book, even though it
has interesting and nontrivial properties. We restrict our treatment to generalizing
the easy direction of Theorem 6.36, and to an application of the results of this
chapter to characterizing when t(., W ) has finite connection rank.
The group Aut(W ) acts on J k for any k. The number of orbits of this action
can be estimated from below as follows.
Proposition 13.46. The number of orbits of the automorphism group of W on
[0, 1]k is at least r(t(., W ), k).
Proof. Suppose that Aut(W ) has a finite set of orbits O1 , . . . Om on [0, 1]k .
Let F and F ′ be two k-labeled graphs. Then
∫
t([[F F ′ ]], W ) = tx1 ...xk (F, W )tx1 ...xk (F ′ , W ) dx1 . . . dxn .
[0,1]k

The functions tx1 ...xk (F, W ) and tx1 ...xk (F ′ , W ) are constant on every orbit, and
hence
∑
m
t([[F F ′ ]], W ) = λ(Oj )txj,1 ...xj,k (F, W )txj,1 ...xj,k (F ′ , W ),
j=1

where (xj,1 . . . xj,k ) is any representative point of Oj . This shows that M (t(., W ), k)
is the sum of m matrices of rank 1, and so it has rank at most m. 

To be able to say something about the finiteness of the number of orbits of

the automorphism group on k-tuples of points of a graphing, we need the following
theorem.
Theorem 13.47. Let W be a graphon such that r(t(., W ), 2) is finite. Then W is
a stepfunction.
Proof. First we show that TW has finite rank. It is clear that a kernel W
has at most m different nonzero eigenvalues if and only if there are real numbers
a0 , . . . , am , not all 0, such that
∑
m
(13.14) ak W ◦(k+2) = 0
k=0

almost
∑ everywhere (so that all eigenvalues of W will be roots of the polynomial
k+2
k a k x ). We claim that this is equivalent to requiring that
∑
m
(13.15) ak ⟨W ◦(k+2) , W ◦(l+2) ⟩ = 0 (l = 0, . . . , m).
k=0

Indeed, (13.14) clearly implies (13.15) for every l; on the other hand, (13.15) implies
that
⟨∑ m ∑
m ⟩
(13.16) ak W ◦(k+2) , ak W ◦(k+2) = 0,
k=0 k=0
236 13. ON THE STRUCTURE OF GRAPHONS

which implies (13.14). Using (7.22), equation (13.15) can be rewritten as

∑
m
(13.17) ak t(Ck+l+4 , W ) = 0 (l = 0, . . . , m).
k=0

This is a system of m+1 homogeneous linear equations in m+1 variables ak , which

is solvable if and only if its determinant vanishes:
 
 t(C4 , W ) t(C5 , W ) . . . t(Cm+4 , W ) 

 t(C5 , W ) t(C6 , W ) . . . t(Cm+5 , W ) 

(13.18)  .. ..  = 0.
 . . 
 
t(Cm+4 , W ) t(Cm+5 , W ) . . . t(C2m+4 , W )

Since this matrix is a submatrix of M (f, 2), this determinant will certainly vanish
if m ≥ r(f, 2). It follows that the number of distinct nonzero eigenvalues of the
operator TW is at most r(f, 2). Since every eigenvalue has finite multiplicity, it
follows that TW has finite rank.
Next, we show that the range of W ◦ W is finite (up to a set of measure
0). Consider its moments ∫ as a single variable function on the probability space
[0, 1]2 : Mk (W ◦ W ) = [0,1]2 (W ◦ W )k , and the corresponding moment matrix
( )∞
M (W ◦ W ) = Mk+l (W ◦ W ) k,l=0 . Note that Mk (W ◦ W ) = t(K2,k , W ), and
•• ••
so Mk+l (W ◦ W ) = t(K2,k K2,l , W ). It follows that M (W ◦ W ) is a submatrix of
M (f, 2), and hence its rank is finite. By Theorem A.22, the range of W ◦ W is finite
(up to a set of measure 0).
2
The fact that TW has finite rank implies that TW ◦W = TW has finite rank.
This, together with the fact that the range of W is finite, implies that W is a
stepfunction. Indeed, the row space of W is finite dimensional, so we can select
a finite set of points x1 , . . . , xr so that every row W (x, .) is a linear combination
of the functions W (xi , .). Since W has finite range, the functions W (xi , .) are
stepfunctions. There is a finite partition [0, 1] = S1 ∪ · · · ∪ Sp such that every
function W (xi , .) is constant on every Si , and hence every row is constant on every
Si . By symmetry, this implies that W is constant on every rectangle Si × Sj , i.e.,
it is a stepfunction. 

It is easy to derive from this the following analogue of Theorem 5.54.

Corollary 13.48. Let f be a reflection positive, multiplicative, and normalized
simple graph parameter. Then either r(f, k) is infinite for all k ≥ 2, or there is a
twinfree weighted graph H such that f = t(., H), r(f, k) is finite for all k ≥ 0, and
r(f, k)1/k → v(H).
This result is not a strengthening of Theorem 5.54, because it concerns simple
graph parameters. The extension to multigraph parameters is more complicated,
and we will return to it later, in Chapter 17.

Proof. By Theorem 11.52 and Proposition 14.61, there is a graphon W such

that f = t(., W ). If r(f, 2) = ∞, then trivially r(f, k) = ∞ for all k ≥ 2. Suppose
that r(f, 2) < ∞, then by Theorem 13.47, W is a stepfunction, and so there is
a weighted graph H such that f = t(., H). The conclusion follows by theorem
6.36. 
 13.5. SYMMETRIES OF GRAPHONS 237

For the graph parameter f = t(., W ), we have r(f, 0) = 1 by multiplicativity,

so this is always finite. The rank r(f, 1) may be finite; this happens if W has
an automorphism group that has a finite number of orbits with positive measure.
Proposition 13.46 and Theorem 13.47 imply that if Aut(W ) has a finite number
of orbits on pairs, then W is a stepfunction. It follows in particular that Aut(W )
has a finite number of orbits on k-tuples of points for every k. This is somewhat
surprising in view of the fact that there are arbitrarily large finite graphs whose
automorphism group has only three orbits on pairs, but an unbounded number of
orbits on k-tuples for k ≥ 3 (for example, the Paley graphs in Example 1.1).
 CHAPTER 14

The space of graphons

The space of graphons is the stage where many acts of interaction between
graph theory and analysis take place. This chapter collects a number of questions
about the structure of this space that arise naturally and that have at least partial
answers.

14.1. Norms defined by graphs

We mentioned, and through the results in the last chapters also illustrated, that
the cut norm and the cut-distance are best suited for measuring structural simi-
larity of two graphons. However, other norms are also important; the connection
between norms on the graphon space and our theory is twofold: first, homomor-
phism densities give rise to interesting norms, and second, norms with some natural
properties are closely related to the cut norm. We start with a discussion of norms
defined by homomorphism densities, based on the work of Hatami [2010].
We have seen that many Schatten norms of a kernel operator can be expressed
by the homomorphism densities of even cycles. Here we consider a more general
question: for which graphs F is |t(F, W )|1/e(F ) a norm? We call such a graph
F norming. This condition can be relaxed in two directions: (1) We can ask
whether the functional W 7→ |t(F, W )|1/e(F ) is a seminorm on W (i.e., whether it is
subadditive, but could be 0 even if W is not identically 0; homogeneity is trivial).
We call F seminorming if this holds. (2) We can ask whether W 7→ t(F, |W |)1/e(F )
is a norm (we moved the absolute value signs in); we call F weakly norming if
this holds. This is equivalent to asking: which graphs F have the property that
the subadditivity inequality t(F, W1 + W2 )1/e(F ) ≤ t(F, W1 )1/e(F ) + t(F, W2 )1/e(F )
holds for all W1 , W2 ∈ W0 ? (There is no fourth version: if the functional W 7→
t(F, |W |)1/e(F ) is a seminorm, then it is a norm; see Exercise 14.9).
A related property is that the graph F satisfies t(F, W ) ≥ 0 for every kernel
W . Such graphs are called positive. If F = [[F12 ]] for some k-labeled graph F1
with nonadjacent labeled nodes, then F is positive, but the converse is not known.
Exercises 14.3 and 14.4 state some of the known properties of positive graphs.
Returning to graphs that are norming in one sense or the other, we collect some
(easy) facts about them, due to Hatami [2010] and Kunszenti-Kovács [unpublished].
Every seminorming or weakly norming graph is bipartite (Exercise 14.7). Being
seminorming is almost equivalent to being norming: every seminorming graph that
is not norming is a star with an even number of edges. Every seminorming graph
is positive (and hence we don’t need the absolute value in the definition; Exercise
14.8), but not every positive graph is seminorming. Graphs with an odd number
of edges cannot be seminorming, but they can be weakly norming, as the example
of K2 shows.
239
240 14. THE SPACE OF GRAPHONS

There are several classes of graphs F with norming properties. Besides cycles,
complete bipartite graphs with an even number of nodes in each bipartition class are
norming, and all complete bipartite graphs are weakly norming. For more properties
and examples of graphs with norming properties, see Exercises 14.5–14.8.
Norming properties are closely related to Hölder-type inequalities for homo-
morphism densities, which can be stated using the notion of W-decorated graphs
introduced in Section 7.2. A(simple graph F) = (V, E) has the Hölder property, if
for every W-decoration w = we : e ∈ E(F ) of F ,
∏
(14.1) t(F, w)e(F ) ≤ t(F, we ).
e∈E

It has the weak Hölder property, if this inequality holds for every W0 -decoration of
F (equivalently, for every W-decoration with nonnegative functions).
Hatami [2010] gives the following characterizations of seminorming and weakly
norming graphs in terms of Hölder properties.
Theorem 14.1. A simple graph is seminorming if and only if it has the Hölder
property. It is weakly norming if and only if it has the weak Hölder property.
Proof. We prove the second assertion; the proof of the first is similar. In the
“if” direction, suppose that a simple graph F = (V, E) with m edges has the weak
Hölder property. Let W1 , W2 ∈ W0 . We have
∑
t(F, W1 + W2 ) = t(F, w),
w

where the summation extends to all {W1 , W2 }-decorations w of F . So by the weak

Hölder property,
∑∏ ∑m ( )
m
t(F, W1 + W2 ) ≤ t(F, we )1/m = t(F, W1 )k/m t(F, W2 )(m−k)/m
w e∈E
k
k=0
( )m
= t(F, W1 )1/m + t(F, W2 )1/m ,

which shows that the functional t(F, W )1/m is subadditive on W0 .

The proof of the “only if” direction is trickier. Suppose that F is weakly
norming, and let (F, w) be a W0 -decoration of F . Inequality (14.1) is homogeneous
of degree m in each graphon we , so we may scale those and assume that t(F, we ) = 1
for every edge. We want to prove that t(F, w) ≤ 1, but first we prove the weaker
inequality
(14.2) t(F, w) ≤ mm .
∑
Indeed, if W = e we , then using subadditivity, we get
(∑ )m
t(F, w) ≤ t(F, W ) ≤ t(F, we )1/m = mm .
e

To conclude, we use a method called tensoring. Let n ≥ 1, and let us decorate

every edge e by the tensor product we⊗n . Then one has
t(F, w⊗n ) = t(F, w)n , t(F, we⊗n ) = t(F, we )n = 1,
and hence 14.2 implies that t(F, w) ≤ mm/n . Since this holds for every n, it follows
that t(F, w) ≤ 1. 
 14.1. NORMS DEFINED BY GRAPHS 241

Using this Theorem, one can prove about some graphs that they are norming
(Hatami [2010]). A characterization of such graphs is open.
Proposition 14.2. (a) Even cycles are norming.
(b) Hypercubes are weakly norming.
(c) Deleting a perfect matching from a complete bipartite graph Kn,n , we get a
weakly norming graph.
Proof. We describe the proof of (b); the proofs of (a) and (c) are similar
(in fact, simpler). Consider the d-dimensional hypercube graph Qd . We consider
its node set as V = {0, 1}d , and its edge set as E = {xy : x, y ∈ V, xi =
yi for all but one i}.
By Theorem 14.1, it is enough to prove that (14.1) holds for F = Qd and
any decoration with graphons. Let A be the set of graphons that occur. We may
assume that A does not contain the graphon that is almost everywhere 0 (else,
the inequality/is trivial). We say that an A-decoration W of Qd is pessimal, if
d ∏
t(Qd , W )e(Q ) e∈E t(Q , We ) is maximal among all A-decorations. Since there
d

are only a finite number of such decorations, at least one of them is pessimal. In
these terms, inequality (14.1) means that there is a pessimal A-decoration with all
decorating graphons equal.
Let S1 denote the set of nodes x of Qd with x1 = 1, x2 = 0; let S2 be the set
of nodes x with x1 = 0, x2 = 1; and let T = V \ S1 \ S2 . Note that T separates S1
and S2 . Let Ei be the set of edges incident with any node in Si , and let E0 be the
set of edges spanned by T .
We can write
∫ ∏ ∫ ∏ ∏ ∏
t(F, W ) = Wij (xi , xj ) dx = .
ij∈E ij∈E0 ij∈E1 ij∈E2
[0,1]V [0,1]V

Considering the first factor as a weight function (here we use that W ≥ 0), we can
apply the Cauchy–Schwarz Inequality to get
( ∫ )1/2 ( ∫ )1/2
∏ ( ∏ )2 ∏ ( ∏ )2
t(Q , W ) ≤
d
.
ij∈E0 ij∈E1 ij∈E0 ij∈E2
[0,1]V [0,1]V

Interchanging x1 and x2 in every x ∈ V (in other words, reflecting in the hyperplane

x1 = x2 ) is an automorphism σ of Qd which maps E1 onto E2 , and therefore
∫ ∏ ( ∏ )2
= t(Qd , W ′ ),
ij∈E0 ij∈E1
[0,1]V

where We′ = We if e ∈ E1 ∪ E0 , and We′ = Wσ(e) if e ∈ E2 . Similarly,

∫ ∏ ( ∏ )2
= t(Qd , W ′′ ),
ij∈E0 ij∈E2
[0,1]V

where We′′ = We if e ∈ E2 ∪ E0 , and We′ = Wσ(e) if e ∈ E1 . Thus we get

t(Qd , W ) ≤ t(Qd , W ′ )1/2 t(Qd , W ′′ )1/2 .
By the definition of pessimal decoration, we must have equality here, and the deco-
rations W ′ and W ′′ must also be pessimal. Here (say) W ′ is a pessimal decoration
242 14. THE SPACE OF GRAPHONS

which is invariant under interchanging the first two entries in every x ∈ V . We call
this symmetrization with respect to the hyperplane x1 = x2 . We can symmetrize
similarly with respect to the hyperplane x1 + x2 = 1.
Now consider a pessimal decoration W and a face Z of the cube such that all
edges of Z are decorated by the same graphon W . Suppose that Z is not the whole
cube. We may assume that Z is the face defined by x1 = x2 = · · · = xk = 0,
where 0 < k < d. Let Z ′ be the face obtained by reflecting Z in the hyperplane
xk = xk+1 . The intersection of Z and Z ′ is the face defined by x1 = x2 = · · · =
xk = xk+1 = 0. The smallest face Z ′′ containing both Z and Z ′ is defined by
x1 = x2 = · · · = xk−1 = 0.
Let us symmetrize with respect to the hyperplane xk = xk+1 . The decoration of
the edges of Z does not change, but the decoration of the edges of Z ′ also becomes
U . Symmetrizing with respect to xk + xk+1 = 1, we get a pessimal decoration in
which all edges of Z ′′ have the same decoration. Repeating this procedure, we get
a pessimal decoration with all edges decorated by the same graphon and we are
done. 

Exercise 14.3. A graph F is positive if and only if every connected component

of F that is not positive occurs with even multiplicity.
Exercise 14.4. Let F be a positive simple graph. (a) F × G is positive for every
simple graph G. (b) There is a homomorphism F → F such that every edge
has an even number of pre-images (c) [Harder] If F is positive, then there is a
homomorphism F → G into a simple graph G with v(G) ≥ v(F )/2 such that
every edge has an even number of pre-images (Camarena, Csóka, Hubai, Lippner
and Lovász [2012]).
Exercise 14.5. Prove that (a) complete bipartite graphs with an even number of
nodes in both color classes are norming, (b) complete bipartite graphs are weakly
norming, (c) stars with an even number of edges are seminorming, (c) K2,3 is not
seminorming (but weakly norming).
Exercise 14.6. Let T be a tree that is seminorming. (a) Prove that if U, W ∈ W
∫1 ∫1
and 0 U (x, y) dx = 0 W (x, y) dx for every y, then t(T, U ) = t(T, W ). (b) Prove
that T is a star.
Exercise 14.7. (a) Every seminorming or weakly norming graph is bipartite.
(b) Every seminorming graph is either a star, or eulerian. (c) Every seminorming
graph that is not norming is a star.
Exercise 14.8. Let F be a seminorming graph. Prove that (a) kernels with
t(F, W ) = 0 form a linear space; (b) e(F ) is even; (c) F is positive.
Exercise 14.9. Prove that if the functional W 7→ t(F, |W |)1/e(F ) is a seminorm,
then it is a norm.

14.2. Other norms on the kernel space

The topologies on W1 defined by the cut norm, L1 -norm, weak convergence etc.
are different, but there are some subtle, nonobvious relationships between them.
This turns out to be quite important for graph-theoretic applications: the interplay
between the cut norm and L2 -norm is crucial in the proof of the Regularity Lemma
(Section 9.1.2), and the relationship between the cut norm and L1 -norm is the key
to the analytic theory of property testing (Section 15.3) and to the stability theory
of extremal graphs (Section 16.4). We are not going to explore all the connections
between these norms, just those that have graph theoretical significance.
 14.2. OTHER NORMS ON THE KERNEL SPACE 243

Almost all norms on W (in short, norms for this section) that we need have
some natural properties. Recall from Section 8.2 that a norm N is called invariant,
if N (W φ ) = N (W ) for every measure preserving transformation φ ∈ S[0,1] , and
smooth, if for every sequence Wn ∈ W1 of kernels such that Wn → 0 almost
everywhere, we have N (Wn ) → 0. The norms L1 , L2 , the cut norm, and the graph
norms from the previous section share these properties. (But the L∞ -norm is not
smooth!)
Recall the obvious inequalities
(14.3) ∥W ∥ ≤ ∥W ∥1 ≤ ∥W ∥2 .
1/2
For W ∈ W1 , we have ∥W ∥2 ≤ ∥W ∥1 , and hence these two norms define the
same topology on W1 . Trivially, the cut norm is continuous in this topology. How
about the other way around? There are easy examples showing that ∥Wn ∥ → 0
does not imply that ∥Wn ∥1 → 0 or ∥Wn ∥2 → 0: let (Gn ) be a quasirandom graph
sequence with edge density 1/2, and Wn = 2WGn − 1. Then ∥Wn ∥ → 0, but
∥Wn ∥1 = ∥Wn ∥2 = 1.
The main goal in this section is to establish the following picture about smooth
invariant norms.
Theorem 14.10. (a) Every smooth invariant norm (as a function on W f1 ) is con-
tinuous with respect to the L1 norm, and the cut norm is continuous with respect
to any smooth invariant norm.
(b) Any smooth invariant norm is lower semicontinuous with respect to any
other smooth invariant norm.
We also prove an analogous (but not equivalent!) theorem about the distances
δN on Wf defined by smooth invariant norms N . Let us call these, for brevity,
delta-metrics.
Theorem 14.11. (a) Every delta-metric is continuous (as a function on W f1 × W
f1 )
with respect to δ1 , and δ is continuous with respect to any delta-metric.
(b) Any delta-metric is lower semicontinuous with respect to any other delta-
metric.
The fact that we prove continuity (or lower semicontinuity) as a function in two
variables, and not just separately in each variable, is significant. As an example of
a different nature, recall that the overlay functional C(U, W ) is continuous in each
variable, but not as a 2-variable function (Section 12.2).
Some of the above statements are trivial, and some follow easily from each
other. Along the lines, we are going to prove a couple of facts that will be useful
in other contexts too.
14.2.1. Smooth and invariant norms. As a technical preparation, we have
to prove some simple facts about smooth and invariant norms.
Lemma 14.12. Every smooth norm N is uniformly continuous with respect to the
L1 norm on W1 .
Proof. Suppose not, then there exists an ε > 0 and a sequence of kernels
Wn ∈ W1 such that ∥Wn ∥1 → 0 but N (Wn ) > 0. By selecting a subsequence, we
may assume that Wn → 0 almost everywhere, contradicting the assumption that
N is smooth. 
244 14. THE SPACE OF GRAPHONS

Next we prove a basic property of the stepping operator.

Proposition 14.13. The stepping operator is contractive with respect to any
smooth invariant norm.
Proof. Let P be any finite measurable partition of [0, 1], and let N be an
invariant norm. We want to prove that N (WP ) ≤ N (W ) for every W ∈ W. By the
invariance of N , we may assume that the partition classes of P are intervals. We
may also assume that W ∈ W1 .
For every interval I = [a, b) ∈ P, let φI : I → I denote a measure
( preserving
)
map x 7→ a + 2(x − a) (mod b − a). Then every map (x, y) 7→ φI (x), φJ (y) is
ergodic on I × J. Let φ : [0, 1] → [0, 1] denote the map that acts on I ∈ P as φI .
For n ≥ 1, define
1 ∑ φk 1∑ ( k )
n−1 n−1
Un (x, y) = W (x, y) = W φ (x), φk (y) .
n n
k=0 k=0

Using the subadditivity and invariance of N , we get

1∑
n−1
k
N (Un ) ≤ N (W φ ) = N (W ).
n
k=0

On the other hand, the Ergodic Theorem implies that Un → WP almost everywhere
as n → ∞. Since, trivially, Un ∈ W1 and N is smooth, this implies that N (WP ) =
limn→∞ N (Un ) ≤ N (W ). 
Next we give a useful representation of smooth invariant norms. By the Hahn–
Banach Theorem, we can represent any norm on W that is continuous in the L∞
norm as
(14.4) N (W ) = sup ℓ(W ),
ℓ∈L

where L is an appropriate set of linear functionals on W, continuous in the L∞

norm. We show that for our norms, the linear functionals in L can be represented
as inner products with functions in W.
Proposition 14.14. For every smooth and invariant norm N there is a set K ⊆ W
such that
N (W ) = sup ⟨U, W ⟩
U ∈K
for every W ∈ W.
Proof. Define K = {Y ∈ W : ⟨Y, U ⟩ ≤ N (U ) ∀U ∈ W}. Let W ∈ W, then
we want to prove that N (W ) = supY ∈K ⟨Y, W ⟩. Suppose not, then we may assume
that N (W ) > 1 > supY ∈K ⟨Y, W ⟩.
First, we assume that W is a stepfunction. Let P be the partition of [0, 1] into
the steps of W . The linear space WP of kernels with steps in P is finite dimensional,
and B = {U ∈ WP : N (U ) ≤ 1} is a convex set in it. Since W ∈ / B, and we are
working in a finite dimensional space, there is a hyperplane of the form ⟨Y, .⟩ = 1
(Y ∈ WP ) through the point W such that ⟨Y, X⟩ ≤ 1 for all X ∈ B. Then for any
U ∈ W, using Proposition 14.13, we get
⟨ 1 ⟩
⟨Y, U ⟩ = ⟨Y, UP ⟩ = N (UP ) Y, UP ≤ N (UP ) ≤ N (U ),
N (UP )
 14.2. OTHER NORMS ON THE KERNEL SPACE 245

which shows that Y ∈ K. Since ⟨Y, W ⟩ = 1, this is a contradiction.

Second, let W be an arbitrary kernel. Proposition 9.8 implies that there is a
stepfunction W ′ such that N (W −W ′ ) < N (W )−1. Then N (W −W ′ ) < N (W )−1,
and hence N (W ′ ) > 1. We know already that there is a stepfunction Y ∈ K
with the same steps as W ′ such that ⟨Y, W ′ ⟩ = 1. Since ⟨Y, W ⟩ = ⟨Y, W ′ ⟩, this
contradicts the choice of W , and completes the proof. 
The following lemma is a special case of the theorem, but it is best to formulate
and prove it separately.
Lemma 14.15. Let N be any smooth invariant norm, and let Wn → W in the cut
norm (Wn , W ∈ W1 ). Then
lim inf N (Wn ) ≥ N (W ).
n→∞

Proof. By Proposition 14.14, the norm N can be represented as N (X) =

supY ∈K ⟨X, Y ⟩ for some K ⊆ W. Let ε > 0 and choose a function Y ∈ K such
that ⟨Y, W ⟩ ≥ N (W ) − ε. Then by Lemma 8.22, N (Wn ) ≥ ⟨Y, Wn ⟩ → ⟨Y, W ⟩ ≥
N (W ) − ε. Since ε > 0 was arbitrary, this proves the Proposition. 
We can now give the proof of the first main theorem in this section.
Proof of Theorem 14.10. (a) Lemma 14.12 proves the first statement. To
prove the second, let N be a smooth invariant norm. Suppose that the cut norm is
not continuous with respect to N , then there is a sequence of kernels Wn ∈ W1 such
that N (Wn ) → 0 but ∥Wn ∥ ≥ c > 0 for all n. By the compactness of the graphon
space, we may also assume that δ (Wn , U ) → 0 for some nonzero graphon U . This
means that there are invertible measure preserving transformations φn such that
∥Wnφn −U ∥ → 0. Lemma 14.15 implies that lim inf n N (Wnφn ) = lim inf n N (Wn ) ≥
N (U ) > 0. Since N is invariant, this contradicts the assumption N (Wn ) → 0.
(b) Let N1 and N2 be two smooth invariant norms on W, and let
W1 , W2 , . . . , W ∈ W1 such that N1 (Wn − W ) → 0. Then ∥Wn − W ∥ → 0 by
(a). Hence liminfn N2 (Wn ) ≥ N (W ) by Lemma 14.15. 
14.2.2. Delta-distances. We start with a fact similar to Lemma 14.15, but
more difficult to prove, with the distances δN and δ replacing the norm N and
the cut norm. For the case when N is the L1 -norm, this was proved by Lovász and
Szegedy [2010a].
Lemma 14.16. Let N be a smooth invariant norm on W. Let δ (Un , U ) → 0 and
δ (Wn , W ) → 0 as n → ∞ (U, W, Un , Wn ∈ W1 ). Then
lim inf δN (Wn , Un ) ≥ δN (W, U ).
n→∞

In other words, the distance δN is lower semicontinuous on the compact metric

space (Wf1 , δ ).

Proof. Applying appropriate measure preserving transformations to the ker-

nels Un and Wn , we may assume that ∥Un − U ∥ → 0 and ∥Wn − W ∥ → 0
when n → ∞. Fix an ε > 0. Let P and Q denote finite partitions of [0, 1]
such that N (W − WP ) ≤ ε and N (U − UQ ) ≤ ε. For any positive integer n,
there are measure preserving transformations φn , ψn : [0, 1] 7→ [0, 1] such that
δN (Wn , Un ) = N (Wnφn − Unψn ). The difficulty (why we cannot apply Lemma 14.15)
is that these transformations φn and ψn may depend on n.
246 14. THE SPACE OF GRAPHONS

But as a next step, we fix n as follows. Let R = {R1 , . . . , Rm } denote the

common( refinement of )the partitions P and Q. We claim that if n is large enough
then N (Wn )R − WR ≤ ε. By Lemma 14.12, there is an ε′ > 0 such that it
is enough to guarantee that ∥(Wn )R − WR ∥1 ≤ ε′ . By (8.15), this follows if we
have ∥(Wn )R − WR ∥ ≤ ε′ /m2 . By Proposition 14.13, this will hold if ∥Wn −
W ∥ ≤ ε′ /m2 , which holds for every n that is(large enough)by the assumption that
∥Wn − W ∥ → 0 when n → ∞. Similarly, N (Un )R − UR ≤ ε holds if we choose
n large enough.
We consider n fixed from now on, and so we can replace Wn , W and P by
Wnφn , W φn and φn (P), and replace Un , U and Q by Unψn , U ψn and ψ(Q). With
this new notation, we have δN (Wn , Un ) = N (Wn − Un ). Then
δN (W, U ) ≤ N (W − U ) ≤ N (W − WR ) + N (WR − UR ) + N (UR − U ).
By the choice of P and by Proposition 14.13
( )
N (W −WR ) ≤ N (W −WP )+N (WP −WR ) = N (W −WP )+N (WP −W )R ≤ 2ε,
and using the analogous estimate for U , we get that
δN (W, U ) ≤ N (WR − UR ) + 4ε.
Using Proposition 14.13 again and the fact that the kernels WP and UQ are both
constant on the rectangles Ri × Rj , we get
( )
δN (Wn , Un ) = N (Wn − Un ) ≥ N (Wn )R − (Un )R
( ) ( )
≥ N (WR − UR ) − N (Wn )R − WR − N UR − (Un )R
≥ N (WR − UR ) − 2ε ≥ δN (W, U ) − 6ε.
Since this holds for every ε > 0 if n is large enough, the assertion follows. 
Proof of Theorem 14.11. The theorem follows from Lemma 14.16 similarly
as Theorem 14.10 followed from Lemma 14.15. The details are not repeated. 
A consequence of Lemma 14.16 (or Theorem 14.11) is worth formulating.
Corollary 14.17. Let N be a smooth invariant norm. Let R ⊆ W f1 be compact with
respect to the δ distance. Then the functional δN (., R) is lower semicontinuous
f1 , δ ).
on (W
Proof. Suppose that Un → U in the δ distance; we want to prove that
lim inf δN (Un , R) ≥ δN (U, R).
n→∞
By selecting an appropriate subsequence, we may assume that the limes inferior is
actually a limit. For every n, let Wn ∈ R be such that δN (Un , Wn ) ≤ δN (Un , R) +
1/n. Again by going to a subsequence, using the compactness of R, we may assume
that Wn → W for some W ∈ R. By Lemma 14.16,
( 1)
lim inf δN (Un , R) ≥ lim inf δN (Un , Wn ) − ≥ δN (U, W ) ≥ δN (U, R). 
n→∞ n→∞ n
Exercise 14.18. Prove that the statement of Lemma 14.12 remains valid if the
assumption of smoothness is replaced by the assumption that N has a represen-
tation as in Proposition 14.14.
Exercise 14.19. Construct a norm N such that the stepping operator is not
contractive with respect to N .
 14.3. CLOSURES OF GRAPH PROPERTIES 247

Exercise 14.20. Let W1 and W2 be graphons that are monotone increasing in

both variables. Prove that (a) δ1 (W1 , W2 ) = ∥W1 − W2 ∥1 ; (b) δ (W1 , W2 ) =
2/3
∥W1 − W2 ∥ ; (c) ∥W1 − W2 ∥1 ≤ 10∥W1 − W2 ∥ (Bollobás, Janson and Riordan
[2012]).

14.3. Closures of graph properties

A class of graphons closed under weak isomorphism is called a graphon property.
Since every graphon is weakly isomorphic to a graphon on [0, 1], we can usually
restrict our attention to graphons on [0, 1] when studying graphon properties. In
this section, we study graphon properties obtained as closures of graph properties,
and graphon properties defined by equations.
A graph property P is a class of finite graphs closed under isomorphism. Let P
is the set of graphons (J, W ) that arise as limits of graph sequences in P. Geometric
and topological properties of the closure reveal important information about the
graph property, as we shall see.
14.3.1. Hereditary properties. Recall that a graph property is called hered-
itary, if whenever G ∈ P, then every induced subgraph is also in P. For every
graphon W , let I(W ) denote the set of its “induced subgraphs”, i.e., the set of
those graphs F for which tind (F, W ) > 0. Clearly, I(W ) is a hereditary graph
property.
Let P be a hereditary property of graphs. Then
(14.5) ∪W ∈P I(W ) ⊆ P.
Indeed, if F ∈
/ P, then tind (F, G) = 0 for every G ∈ P, since P is hereditary. This
implies that tind (F, W ) = 0 for all W ∈ P, and so F ∈ / I(W ).
Equality does not always hold in (14.5). For example, we can always add
a graph G and all its induced subgraphs to P without changing P. As a less
trivial example, consider all graphs with degrees bounded by 10. This property is
hereditary, and P consists of a single graphon (the identically 0 function), so the
left hand side of (14.5) consists of edgeless graphs only. Equality in (14.5) can be
characterized by assuming that P is closed not only under induced subgraphs, but
also under a certain version of multiplying points (Exercise 14.28).
The closure of the set of triangle-free graphs is the set of triangle-free graphons,
which can be characterized by the property t(K3 , W ) = 0. More generally:
Proposition 14.21. Let P be a hereditary graph property. Then its closure is
characterized by the (infinitely many) equations tind (F, W ) = 0 for all F ∈
/ P.
Proof. By the definition of hereditary properties, the equations tind (F, G) = 0
hold for all G ∈ P and F ∈ / P, which implies that tind (F, W ) = 0 for every W ∈ P.
Conversely, suppose that W has the property that tind (F, W ) = 0 for all F ∈
/ P.
This means that P(G(k, W ) ∈ P) = 1 for every k. Since G(k, W ) → W with
probability 1, this implies that W ∈ P. 
14.3.2. Random-free properties. A graph property P is random-free , if
every W ∈ P is 0-1 valued almost everywhere. (For an explanation of the name, see
Exercise 14.29.) By Proposition 8.24, if (Gn ) is a convergent random-free sequence
of graphs with limit graphon W , then WGn → W in the L1 norm. Example 11.41
illustrated how random-freeness was related to a small amount of randomness in a
randomly generated random-free graph sequence. We can also point out that if W
248 14. THE SPACE OF GRAPHONS

is 0-1 valued, then G(n, W ) = H(n, W ), and so to generate G(n, W ), we don’t need
randomness to get the edges (of course, we still need randomness to generate the
nodes).
Among hereditary properties, it is quite easy to characterize random-free prop-
erties.
Lemma 14.22. A hereditary graph property P is random-free if and only if there
is a signed bipartite graph F such that t(F, W ) = 0 for all W ∈ P.
The proof will show that it would be enough to assume that for every graphon
W ∈ P there is a signed bipartite graph with t(F, W ) = 0.

Proof. Suppose that for every signed bipartite graph F there is a graphon
W ∈ P such that t(F, W ) > 0. Let (Fn ) be a quasirandom sequence of bipartite
graphs with bipartition V (Fn ) = Vn′ ∪ Vn′′ , with edge density 1/2, and with |Vn′ | =
|Vn′′ |. Consider the signed bipartite graphs Fbn , obtained from Kn,n by signing
the edges of Fn with +, the other edges with −. Let Wn ∈ P be a graphon
such that t(F cn , Wn ) > 0. It is easy to see that this means that there is a simple
graph Gn obtained from Fn by adding edges within the color classes such that
tind (Gn , Wn ) > 0. By Proposition 14.21, this implies that Gn ∈ P.
By selecting a subsequence we may assume that the graph sequences G′n =
(Gn [Vn′ ]) and G′′n = (Gn [Vn′′ ]) are convergent. By Theorem 11.59, we can order the
nodes in Vn′ and in Vn′′ so that WG′n converges to a graphon W ′ on [0, 1] in the
cut norm, and similarly WG′′n converges to a graphon W ′′ on [0, 1]. If we order the
nodes of Gn so that the nodes in Vn′ precede the nodes in Vn′′ , and keep the above
ordering inside Vn′ and Vn′′ , then WGn converges to the graphon

 ′
W (2x, 2y) if x, y < 1/2,
′′
U (x, y) = W (2x − 1, 2y − 1) if x, y > 1/2,


1/2 otherwise.

So U ∈ P is not 0-1 valued, and hence P is not random-free.

Conversely, suppose that P is not random-free, and let W ∈ P be a graphon
that is not 0-1 valued almost everywhere. Then by Theorem 13.35(a), t(F, W ) > 0
for every signed bipartite graph F . 

Corollary 14.23. If a hereditary property of bipartite graphs does not contain all
bipartite graphs, then it is random-free.
Using Theorem 13.35(c), we can associate a finite dimension with every nontriv-
ial hereditary property of bipartite graphs. It would be interesting to find further
combinatorial properties of this dimension.
The natural analogue of this corollary for properties of nonbipartite graphs fails
to hold.
Example 14.24. Let P be the property of a graph that it is triangle-free. Then
every bipartite graphon is in its closure, but such graphons need not be 0-1 valued.

For more characterizations of hereditary and random-free properties, and for
more on their connection, see Janson [2011c].
 14.3. CLOSURES OF GRAPH PROPERTIES 249

14.3.3. Flexible properties. A graphon U is called a flexing of a graphon

W , if U (x, y) = W (x, y) for all x, y with W (x, y) ∈ {0, 1} (so we may change the
values of W that are strictly between 0 and 1; we may change them to 0 or to 1,
so the relation is not symmetric). We say that a graphon property is flexible if it is
preserved under flexing.
Every random-free graphon property is trivially flexible. It is also clear that
the intersection and union of any set of flexible graphon properties is flexible. If R
is a flexible graphon property, then so are its “complement” {1 − W : W ∈ R},
its “downward closure” {U ∈ W0 : (∃W ∈ R) U ≤ W }, and its “upward closure”
defined analogously.
For every signed graph F , the graphon property {W ∈ W0 (: tind (F, W ) = )0} is
flexible, since the condition means that for almost all vectors
( xi : i )∈ V (F ) , at
∏ ∏
least one of the factors in ij∈E(F ) W (x1 , xj ) ij∈E(F ) 1 − W (x1 , xj ) is 0, which
is preserved if values strictly between 0 and 1 are changed. By Proposition 14.21,
the closure of any hereditary property can be defined by conditions of the form
tind (F, W ) = 0. This implies:
Proposition 14.25. The closure of a hereditary property is flexible. 
There are other, non-hereditary graph properties whose closure is flexible. Some
of these are described in Exercise 14.31
Every flexible property has the following interesting geometric feature:
Proposition 14.26. If R ⊆ W0 is flexible, then W0 \ R is convex.
Proof. Indeed, let W1 , W2 ∈ W0 \ R, and suppose that a convex combination
W = α1 W1 + α2 W2 ∈ R. Then for every x, y ∈ [0, 1] with W (x, y) ∈ {0, 1} we have
W1 (x, y) = W2 (x, y) = W (x, y), and so by the definition of flexibility, we must have
W1 , W2 ∈ R, a contradiction. 

Corollary 14.27. If P is a hereditary graph property, then W0 \ P is convex.

We will discuss an application of this fact in extremal graph theory in Section
16.5.1.
Exercise 14.28. Prove that for a hereditary property P of graphs equality holds
in (14.5) if and only if for every graph G ∈ P and v ∈ V (G), if we add a new
node v ′ and connect it to all neighbors of v, then at least one of the two graphs
obtained by joining or not joining v and v ′ has property P.
Exercise 14.29. Prove that a graph property is not random-free if it contains
large quasirandom bipartite graphs in the following sense: for every ε > 0 there
is δ > 0, a sequence of graphs G1 , G2 , · · · ∈ P, and disjoint sets Sn , Tn ⊆ V (Gn )
with |Sn | = |Tn | ≥ δv(Gn ) such that the bipartite graphs Gn [Sn , Tn ] form a
quasirandom sequence with error ε.
Exercise 14.30. Prove that a graph property P is random-free if and only if for
every ε > 0 there is an n ∈ N such that for every graph G ∈ P with v(G) ≥ n
and every ε-regular k-partition of G, all but εk2 pairs of partition classes span
bipartite graphs whose edge density is at most ε or at least 1 − ε.
Exercise 14.31. Prove that the closure of the following graph properties is flex-
ible: (a) G is clique of size ⌈|V (G)|/2⌉ together with ⌊|V (G)|/2⌋ isolated nodes;
(b) ω(G) ≥ |V (G)|/2; (c) α(G) ≥ |V (G)|/2; (c) there is a labeling of the nodes
by {1, . . . , n} such that all (i, j) with i + j ≤ n are connected by an edge.
250 14. THE SPACE OF GRAPHONS

14.4. Graphon varieties

The topic of this section is reminiscent of the setup of algebraic geometry: we
study subsets of W, called varieties, defined by equations specifying linear (equiv-
alently, algebraic) dependence between subgraph densities. These equations are
invariant under weak isomorphism, so we may work in W f if we wish. Conditions
like this play a role in extremal graph theory, and a better understanding of these
varieties seems to be an important direction in the study of graphons.
A set of kernels satisfying a condition of the form t(g, W ) = 0, where g is a
quantum graph, will be called a kernel variety. We get, of course, different versions
by putting restrictions on g and on W . A simple variety is defined by a simple
quantum graph g. A graphon variety is the intersection of a kernel variety with
W0 . It is often convenient to restrict our attention to pure graphons; since every
graphon is weakly isomorphic to a pure graphon, this is not an essential restriction.
It is clear that every kernel/graphon variety is closed under weak isomorphism.
Every simple graphon variety is closed in the cut distance, and hence it can be
considered as a closed (and hence compact) subset of the graphon space (W f0 , δ ).
However, this does not hold for non-simple varieties (see Example 14.36). The union
and intersection of two [simple] graphon varieties are [simple] graphon varieties
(Exercise 14.50).
We could try to be more general and consider the common solutions (in W ) of
a system of constraints t(f1 , W ) = 0, . . . , t(fm , W ) = 0. However, this could always
be replaced by the single condition t(f12 + · · · + fm 2
, W ) = 0.
While a general theory of graphon varieties is not at hand, there are some
interesting examples, which will be needed later on. We will see some less trivial
varieties in Section 14.4.2, but this will need some preparation in Section 14.4.1.

Example 14.32 (Constants). As an immediate application of Claim 11.63, we

get that every constant function W = Jp forms a simple kernel variety. Indeed, this
kernel can be defined by the equations t(K2 , W ) = p and t(C4 , W ) = p4 . 

Example 14.33 (Complete graphs). We have mentioned in the introduction

that the densities of triangles and edges in a graph G satisfy the inequality
t(K3 , G) ≥ 2t(K2 , G)2 − t(K2 , G), and equality holds if and only if G is a blow-up
of the complete graph. This is equivalent to saying that (up to weak isomorphism)
the graphon variety defined by the equation t(K3 − 2K2 K2 + K2 , W ) = 0 consists
of the countably many graphons WKn (n = 1, 2, . . . ) and the identically-1 graphon.

∫1
Example 14.34 (Regularity). We call a kernel d-regular, if 0 W (x, y) dy = d for
almost all 0 ≤ x ≤ 1. This kernel variety can be defined by two subgraph density
constraints: t(K2 , W ) = d and t(P3 , W ) = d2 . (This can be shown by a simpler
version of the argument in the proof of Claim 11.63.) Regular kernels (without
specifying the degree d) can be defined by the constraint t(P3 − K2 K2 , W ) = 0. 

Example 14.35 (Hadamard kernels). Our next examples show that graphon
varieties can encode quite substantial combinatorial complications. A symmetric
n × n Hadamard matrix B gives rise to a kernel WB , which we alter a little to get
a graphon UB = (WB + 1)/2. We call UB an Hadamard graphon.
 14.4. GRAPHON VARIETIES 251

Hadamard graphons, together with the graphon J1/2 , form a simple graphon
variety. Indeed, the condition
( )
(14.6) t K3 − 2K2 K2 + K2 , 1 − (2U − 1) ◦ (2U − 1) = 0
implies that 1 − (2U − 1) ◦ (2U − 1) is a kernel that is either identically 1 or it
corresponds to a complete graph (Example 14.33). Let W = 2U − 1, then either
W ◦ W = 0 or W ◦ W = WI where I is an identity matrix of some size n. In the first
case, W = 0 and so U = J1/2 . In the second case, we note that every eigenvector
of TW ◦W = TW 2
is a stepfunction with steps [0, 1/n), . . . , [(n − 1)/n, 1), and hence
so are the eigenvectors of TW . It follows that W is a stepfunction with these steps,
and so W = WB for some n × n matrix B. Furthermore, W ◦ W = WI implies
that B 2 = nI. Since U is a graphon, we have −1 ≤ ∑ W ≤ 1, and so every entry of
B is in [−1, 1]. The condition B 2 = nI implies that i Bij 2
= n for every j, which
implies that every entry of B is either 1 or −1, and so B is an Hadamard matrix.
We must add that (14.6) can be expanded into a subgraph density condition
on U , using the fact that t(F, U ◦ U ) = t(F ′ , U ) (where F ′ is the subdivision of F ).

Example 14.36 (Zero-one valued graphons). It is not hard to see that W ∈ W
is 0-1 valued almost everywhere if and only if t(B4 −2B3 +B2 , W ) = 0 (one approach
is to note that W is 0-1 valued iff txy (B2•• , W ) = txy (B1•• , W ), and use Lemma 14.37
below). Hence 0-1 valued graphons form a kernel variety. However, the variety of
0-1 valued graphons is not simple, because it is not closed in the cut distance: for
a quasirandom graph sequence (Gn ) the associated graphons WGn are 0-1 valued,
but WGn → J1/2 in the ∥.∥ norm. 
14.4.1. Unlabeling. Before describing more complicated graphon varieties,
we introduce a tool that is very useful in constructing varieties. Instead of pre-
scribing subgraph densities, we can try to define graphon or kernel varieties by a
(seemingly) more general condition on the density function of a k-labeled graph or
quantum graph: such conditions can be written as tx (g, W ) = 0 (for all x ∈ [0, 1]k )
for some k-labeled quantum graph g. However, there is a way to translate labeled
constraints to unlabeled constraints. This fact will be convenient in constructions,
since it is often easier to describe a property by the density of a labeled quantum
graph.
Lemma 14.37. For every k-labeled quantum graph f there is an unlabeled quantum
graph g such that for any W ∈ W, t(g, W ) = 0 if and only if tx1 ...xk (f, W ) = 0
almost everywhere. If f is simple, then we can require that g is simple, and the
labeled nodes form a stable set in every constituent of g.
Proof. The first assertion is trivial: tx1 ...xk (f, W ) = 0 almost everywhere if
and only if t([[f 2 ]], W ) = 0. This construction works for the second statement as
well, provided the labeled nodes form a stable set in every constituent of f .
To prove the second statement for every simple k-labeled quantum graph g,
define Lb(g) as the disjoint union of the subgraphs of the constituents induced
by the labeled nodes (note that these subgraphs ( all )have the
( same) node set [k]).
We use induction on the chromatic number χ Lb(f ) . If χ Lb(f ) = 1, then the
labeled nodes are nonadjacent in every constituent, and the trivial construction
above works.
252 14. THE SPACE OF GRAPHONS

( )
Suppose that χ Lb(f ) = r > 1, let [k] = S1 ∪· · ·∪Sr be an r-coloring of Lb(f ),
and let q = |Sr |. We may suppose that Sr = {k − q + 1, . . . , k}. We glue together
two copies of f along Sr . Formally, let f1 be obtained from f by increasing the
labels in Sr by k − q (the labels not in Sr are not changed). Let f2 be obtained
from f by increasing all labels by k − q. So the product f1 f2 is a (2k − q)-labeled
quantum graph, in which the nodes of Sr are labeled 2k − 2q + 1, . . . , 2k − q. Let
h be obtained from f1 f2 by unlabeling the nodes in Sr .
Claim 14.38. For every W ∈ W, tx1 ...xk (f, W ) = 0 almost everywhere if and only
if tx1 ...x2k−2q (h, W ) = 0 almost everywhere.
The “only if” part is obvious, since
tx1 ...xk (f, W ) = 0 ⇒ tx1 ...x2k−q (f1 , W ) = tx1 ...x2k−q (f2 , W ) = 0
⇒ tx1 ...x2k−q (f1 f2 , W ) = 0 ⇒ tx1 ...x2k−2q (h, W ) = 0.
To prove the “if” part, note that two labeled nodes whose labels correspond to the
same label in f are never adjacent, so we can identify these labels in h to get f 2
(with the labels in Sr removed). So tx1 ...x2k−2q (h, W ) = 0 almost everywhere implies
by Proposition 13.23 that t([[f 2 ]], W ) = 0, and hence we get that tx1 ...xk (f, W ) = 0
almost everywhere. This proves the Claim.
Thus it suffices to express the constraint tx1 ...x2k−2q (h, W ) =( 0 by )an appropri-
ate unlabeled constraint. This can be done by induction, since χ Lb(h) ≤ r−1. 
In some cases, the following simple observation suffices to go between labeled
and unlabeled conditions.
Lemma 14.39. Let F be a k-labeled signed graph. Then in W0 , the constraints
tx1 ...xk (F, W ) = 0 and t([[F ]], W ) = 0 define the same graphon variety.
Proof. Clearly tx1 ...xk (F, W ) = 0 implies that t([[F ]], W ) = 0. Conversely, in
the constraint
∫ ∏ ∏ ( )
t([[F ]], W ) = W (xi , xj ) 1 − W (xi , xj ) dx = 0
ij∈E+ ij∈E−
[0,1]V (F )

the integrand is nonnegative, so it must be 0 almost everywhere. Thus integrating

only over the unlabeled nodes, we also get 0 almost everywhere. 
Example 14.40. The identically-p graphon Jp is defined by the constraint
txy (K2•• , U ) = p. Following the above construction, we get that it can be defined
by the constraint t(C4 (p), U ) = 0, where C4 (p) is obtained from C4 by replacing
each edge by the quantum graph K2•• − pO2 . It is a good exercise to verify that
this is equivalent to the conditions t(K2 , U ) = p and t(C4 , U ) = p4 . 
We have seen that 0-1 valued graphons do not form a simple variety. On the
other hand, there are many interesting varieties whose elements are 0-1 valued. As
a first application of Lemma 14.37, we describe a rather general sufficient condition,
generalizing Theorem 13.35(a) from graphons to kernels.
Lemma 14.41. Let F be a signed bipartite graph on n nodes, all labeled. Suppose
that for some W ∈ W we have
(14.7) tx1 x2 ...xn (F, W ) = 0
almost everywhere. Then W (x, y) ∈ {0, 1} almost everywhere.
 14.4. GRAPHON VARIETIES 253

Proof. By Proposition 13.23, (14.7) implies that for the 2-labeled signed bond
B •• obtained by identifying each color class of F , we have txy (B •• , W ) = 0. This
clearly implies that W is 0-1 valued almost everywhere. 

It follows by Lemma 14.39 that if W ∈ W0 , then it is enough to assume that

t([[F ]], W ) = 0, and we get another proof of Theorem 13.35(a).
Example 14.42 (Threshold graphons). Let α : [0, 1] → [0, 1] be a “weight
function”, and consider the graphon Uα (x, y) = 1(α(x) + α(y) < 1). We call these
graphons threshold graphons. They have been studied by Diaconis, Holmes and
Janson [2008] as limits of threshold graphs.
Every such weight function α has a “monotone reordering” in the form of a mea-
sure preserving function φ : [0, 1] → [0, 1] such that α ◦ φ is monotone increasing
(see Proposition A.19). Then Uα◦φ = (Uα )φ is a graphon that is weakly isomorphic
to Uα . Furthermore, Uα◦φ is monotone decreasing, and clearly 0-1 valued. Con-
versely, every monotone decreasing 0-1 valued graphon is almost everywhere equal
to a threshold graphon (see Exercise 14.53).
Threshold graphons form a simple variety. Let C b4 denote a signed 4-labeled
4-cycle, with two opposite edges signed “+”, the other two signed “−”. Then, a
kernel W ∈ W is a threshold graphon if and only if
(14.8) b4 , W ) = 0
tx1 x2 x2 x4 (C
almost everywhere. The necessity of this condition is trivial, for the (elementary)
proof of the sufficiency, see Exercise 14.55. It follows by Lemma 14.39 that if
b4 ]], W ) = 0.
W ∈ W0 , then it is enough to assume that t([[C 
Example 14.43 (Excluded induced subgraphs). From any class of graphs that
is characterized by a finite number of excluded induced subgraphs, we get a graphon
variety by taking the closure (this is immediate by Proposition 14.21). It seems
that the study of this closure often leads to quite interesting questions, which are
mostly unexplored.
As an interesting special case, we can consider graphs not containing an induced
path on 4 nodes. These graphs are called complement reducible, or cographs, and
have many interesting properties and characterizations into which we don’t go here.
The closure consists of graphons W satisfying the equation tind (P4 , W ) = 0. While
a complete characterization of such graphons is awkward, it turns out that the
regular ones among them (those satisfying t(P3 , W ) = t(K2 K2 , W )) have a quite
pretty characterization (we refer to Lovász and Szegedy [2011] for details). 
14.4.2. Stepfunctions and finite rank kernels. Let Sk denote the set of
kernels that are almost everywhere equal to a stepfunction with k steps.
Proposition 14.44. The set Sk is a simple kernel variety, defined by an equation
t(f, .) = 0, where f is a simple quantum graph whose constituents have at most
(k + 1)(k + 2) nodes.
Proof. It is clear that the set Sk is closed under weak isomorphism. Every
function U ∈ Sk satisfies the following equation:
∏ ( )
(14.9) U (xij , xi ) − U (xij , xj ) = 0
1≤i<j≤k+1
254 14. THE SPACE OF GRAPHONS

for almost all choices of the variables xi (1 ≤ i ≤ k + 1) and xij (1 ≤ i < j ≤ k + 1).
Indeed, there are always two variables with one index, say xi and xj , which belong
to the same step, and then the corresponding factor in (14.9) is 0 for any choice of
xij .
Conversely, suppose that U ∈ W is not a stepfunction with k steps. Then there
is a set of (k + 1)-tuples (x1 , . . . , xk+1 ) with positive measure such that no two of
the xi are twins. For every (k + 1)-tuple in this set, there is a positive measure of
choices for xij such that U (xij , xi ) ̸= U (xij , xj ). So (14.9) fails to hold on a set of
positive measure.
Now (14.9) can be written as tx (g,( U ) )= 0 (for an ) appropriate simple m-labeled
quantum graph g, where m = k + 1 + k+1 2 = k+2
2 . By Lemma 14.37, this can be
expressed as t(g, U ) = 0 with a simple unlabeled quantum graph g. This implies
that Sk is a simple kernel variety. Let us note in addition that every constituent of
g is bipartite and fully labeled, hence the construction in the proof of Lemma 14.37
is finished in two steps, and it only doubles the number of nodes. 
We define the rank of a kernel W as the rank of the corresponding kernel
operator TW . This is usually infinite, but we will be interested in the cases when it
is finite. Since every nonzero eigenvalue of TW has finite multiplicity, we know that
a kernel has finite rank if and only if it has a finite number of distinct eigenvalues.
Every stepfunction has finite rank. It is easy to see that the sum and product of
two kernels with finite rank have finite rank (see Exercise 14.56).
If the rank r of W is finite, then some of the formulas in Section 7.5 become
simpler: the spectral decomposition (7.19) of W will be finite (and therefore, it will
hold almost everywhere, not just in L2 ):
∑
r
(14.10) W (x, y) = λk fk (x)fk (y).
k=1

The expression (7.25) for the density of a graph F in W also becomes finite:
∑ ∏ ∏
(14.11) t(F, W ) = λχ(e) Mχ (f ).
χ:E→[r] e∈E v∈V

Example 14.45 (Stepfunctions). Every stepfunction has finite rank. Indeed,

if W is a stepfunction with m steps, then every function in the range of TW is a
stepfunction with the same m steps, and so the dimension of this range is at most
m. 
Example 14.46. If W (x, y) is a (symmetric) polynomial in x and y, then W has
finite rank. It suffices to prove this for the elementary symmetric polynomials x + y
and xy, in which case the corresponding operators have rank 2 and 1, respectively.

Proposition 14.47. (a) Kernels with at most m different nonzero eigenvalues
form a simple variety. (b) Kernels of rank at most m form a simple variety.
Proof. (a) We have seen in the proof of Theorem 13.47 that a kernel has at
most m different nonzero eigenvalues if and only if the determinant (13.18) vanishes.
This determinant can be expressed as t(f, W ), where f is a quantum graph (in which
every constituent is a disjoint union of cycles). Thus W has at most m nonzero
eigenvalues if and only if t(f, W ) = 0. (b) Let x1 , . . . xm , y1 , . . . , ym ∈ [0, 1], and
 14.4. GRAPHON VARIETIES 255

consider the matrix W [x, y] = (W (xi , yj ))m

i,j=1 . A kernel has rank at most m if
and only if for det W [x, y] = 0 for almost all x, y ∈ [0, 1]m . This condition can
be transformed into the form of t(f, W ) = 0 similarly to the proof of Proposition
14.44. 

Example 14.46 shows that not every kernel with finite rank is a stepfunction.
The following theorem asserts that, at least from the point of view of varieties, the
two classes are not very far.
Theorem 14.48. If a kernel variety contains a kernel with finite rank, then it
contains a stepfunction.
Proof. Let the variety V be defined by the equations
t(gi , X) = 0 (i = 1, . . . , m),
and let W ∈ V have finite rank r. Let H be the set of all constituents of the gi .
The equations
t(F, X) = t(F, W ) (F ∈ H)
define a kernel variety V ′ ⊆ V.
Let f1 , . . . , fr be the eigenfunctions of TW . By (14.11), if (u1 , . . . , ur ) is another
set of bounded measurable functions that satisfy
(14.12) Ma (u) = Ma (f )
for every vector ∑of exponents a ∈ [M ]r (where M = max{e(F ) : F ∈ H}), then
r
the kernel U = t=1 λt ut (x)ut (y) satisfies t(F, U ) = t(F, W ) for all F ∈ H and
′
so U ∈ V . By Proposition A.25 in the Appendix there is a system of functions u
satisfying (14.12) which are stepfunctions, and then U is also a stepfunction. 

We will see that every stepfunction in W forms a simple variety in itself (Corol-
lary 16.47). Hence the family of stepfunctions in Theorem 14.48 could not be
replaced by any other family of finite rank kernels (e.g., polynomials).
Exercise 14.49. Show how the facts that regular graphons form simple varieties
and 0-1 valued kernels form a variety (Examples 14.32, 14.34 and 14.36) follow
immediately from the unlabeling method.
Exercise 14.50. Prove that the union and intersection of two graphon varieties
are graphon varieties.
Exercise 14.51. Graphons with values in a fixed finite set S ⊆ R form a variety,
but this variety is not simple unless |S| = 1.
Exercise 14.52. Show that W0 is not a variety in W.
Exercise 14.53. Prove that every monotone decreasing 0-1 valued graphon is
almost everywhere equal to a threshold graphon.
Exercise 14.54. We say that ∪ a kernel W on [0, 1] is an equivalence graphon if
there is a partition [0, 1] = i∈I S1 into measurable parts such that for almost all
pairs x, y ∈ [0, 1], W (x, y) = 1(i = j) if x ∈ Si and y ∈ Sj . Prove that equivalence
graphons (up to isomorphism modulo a 0 set) form a simple variety in W0 .
Exercise 14.55. Assume that a graphon W satisfies (14.8). Prove the following
consequences: (a) W is 0-1 valued( almost) everywhere.
( (b) Defining ) N (x) = {y ∈
[0, 1] : W (x, y) = 1}, either λ N (x)\N (y) = 0 or λ N (y)\N (x) = 0 for all x, y ∈
( { ( ) } )
[0, 1]. (c) Defining α(x) = 12 1 + λ z ∈ [0, 1] : λ N (z) \ N (x) = 0 − λ(N (x)) ,
we have α ≥ 0; (d) W (x, y) = 1(α(x) + α(y) ≤ 1) almost everywhere.
256 14. THE SPACE OF GRAPHONS

Exercise 14.56. Let U and W be two kernels with finite rank. Prove that the
kernels U + W , U W , and U ⊗ W have finite rank. If U has finite rank and W is
arbitrary, then the kernel U ◦ W + W ◦ U has finite rank.

14.5. Random graphons

Going on with the study of the graphon space, we would like to introduce
probability measures on it.
It is not clear that random graphons can be defined at all: in general, one
cannot define the notion of a random measurable function on an uncountable set in
a useful way. However, the situation is better with unlabeled graphons (equivalence
classes of graphons under weak isomorphism): they form a compact metric space,
and the σ-algebra B of Borel subsets in this metric space has good properties (see
e.g. Exercise 14.63). A probability distribution on B will be called a random graphon
model.
Example 14.57 (Limits of growing graph sequences). Let (G1 , G2 , . . . ) be a
sequence of graphs growing under some randomized rule, such that it is convergent
with probability 1. Sometimes the limit graphon is uniquely determined: Among
our examples of convergent graph sequences, growing uniform attachment graphs
(Example 11.39), growing ranked attachment graphs (Exercise 11.47) and growing
prefix attachment graphs (Example 11.41) had this property. On the other hand,
a growing preferential attachment graph sequence (Example 11.44) converges to a
proper distribution on graphons (this distribution is supported on rather simple
graphons in this case, namely constant functions where the constant is chosen from
an appropriate distribution). 
Example 14.58 (The ultimate random graph). When one learns about ran-
dom graphs, it is a very counterintuitive fact that they look so much alike: you
expect to see a great variety in their edge-densities, triangle-densities, etc., to call
them “random” in the everyday sense of word. Well, here is a construction that
satisfies this (it is not clear though whether it can be used in actual constructions
or proofs, due to the non-constructive steps in the definition).
We start with constructing a random graphon model. Let (F1 , F2 , . . . ) be any
ordering of all connected simple graphs, up to isomorphism. (For example, we could
order them by increasing number of nodes, and within this, randomly). We are
going to specify the values of the densities t(Fi , W ) one by one as follows. Suppose
that we have fixed a1 = t(F1 , W ), . . . , am = t(Fm , W ), so that the set Wa1 ...am of
graphons satisfying all these conditions is nonempty. Since the functionals t(F, .)
are continuous on (W f0 , δ ), this set is closed, and hence compact. Consider the set

Ia1 ...am = {t(Fm+1 , W ) : W ∈ Wa1 ...am }.

This is a compact subset of [0, 1], and hence measurable. If it has positive measure,
we choose from it uniformly a random real number for am+1 . If it is of measure 0,
then we choose its minimal element as am+1 .
When we are done with our infinitely many choices, we have a system of equa-
tions t(Fi , W ) = ai , i = 1, 2, . . . . By construction, any finite number of these
are satisfiable, so by the compactness of the graphon space, all of them are si-
multaneously satisfiable. Furthermore, since the subgraph densities determine the
graphon up to weak isomorphism, we get a unique unlabeled random graphon W.
 14.5. RANDOM GRAPHONS 257

This defines a probability distribution on graphons. Now if we want a large simple

“ultimately random” graph, we construct G(n, W).
This is of course a toy example at the moment. It would be interesting to see if
this or perhaps other, similarly constructed “ultimately random” graphs have any
interesting applications. 
14.5.1. Equivalent descriptions of a random graphon. The following
theorem (combining results from Diaconis and Janson [2008] and Lovász and
Szegedy [2012a]) characterizes random graphon models. Another way of looking
at it is that it shows that for each of the objects in Theorem 11.52 we can relax
the conditions to get another important family of cryptomorphic structures. For
example, to relax the condition of multiplicativity (for a normalized parameter), we
consider isolate-indifferent paremeters (invariant under deletion of isolated nodes).
A motivation for the introduction of random graphon models is that they give a
representation of graph parameters that are more general than those represented
by graphons.
Theorem 14.59. The following are equivalent (cryptomorphic):
(a) an isolate-indifferent, normalized simple graph parameter with nonnegative
upper Möbius inverse;
(b) a consistent random graph model;
(c) a countable random graph model;
(d) a random graphon model.
Proof. The proof is similar to the proof of Theorem 11.52. The construction
(a)→(b)→(c) is essentially the same as (b)→(c)→(d) there. The remaining two
steps need some modification.
(c)→(d). Let G be a random countable graph from a consistent countable
random graph model. Let Gn be the finite graph spanned by the first n nodes of
G. By Proposition 11.14, the graph sequence (Gn ) is almost certainly convergent.
Thus by Theorem 11.52, it tends to a limit graphon W with probability 1.
This way we have described a method to generate a random graphon W. It is
not hard to check that the distribution of W is a probability measure on the Borel
f0 , δ ). For every graph F , this random graphon satisfies
sets of (W
t(F, W) = lim t(F, Gn ) = lim tinj (F, Gn ).
n→∞ n→∞

By the consistency of G, the expectation of tinj (F, Gn ) is independent of n for

n ≥ k = |V (F )|, and so
( ) ( ) ( )
E t(F, W) = lim E tinj (F, Gn ) = E tinj (F, Gk ) = P(F ⊆ Gk ).
n→∞

(d)→(a). Let W be a random graphon from any random graphon model. This
defines a graph parameter f by
( )
f (F ) = E t(F, W) .
For every fixed graphon W , the graph parameter f (·) = t(·, W ) is normalized,
isolate-indifferent (since it is multiplicative), and has nonnegative Möbius inverse
(by Theorem 11.52). Trivially, these properties are inherited by the expectation. 
258 14. THE SPACE OF GRAPHONS

14.5.2. Equivalent properties of the descriptions. Graph parameters as

well as random graph models in Theorems 11.52 and 14.59 have a number of alter-
native characterizations. The next Propositions give some of these.
Proposition 14.60. Let f be an isolate-indifferent, normalized simple graph pa-
rameter. Then the following are equivalent:
(a) f is reflection positive;
(b) f is flatly reflection positive (i.e., connection matrices for fully labeled
graphs are semidefinite);
(c) f is nonnegative on signed graphs;
(d) f has nonnegative Möbius inverse;
(e) f = E(t(., W)), where W is a random graphon from a random graphon
model;
(f) f is in the convex hull of the functions t(., W ), where W is a graphon.
Proof. (a) The implications (a) ⇒(b), (c)⇒(d) and (e)⇒(f) are trivial.
(b)⇒(c): The Lindström–Wilf Formula (see (A.1) in the Appendix) gives the
following diagonalization of M flat (f, k): Let Fk′ denote the set of fully labeled simple
graphs on [k]. Let Z denote the Fk′ × Fk′ matrix defined by ZF1 ,F2 = 1F1 ⊆F2 . Let
D be the diagonal matrix with DF,F = f ↑ (F ). Then M flat (f, k) = Z T DZ. This
implies that M flat (f, k) is positive semidefinite if and only if f ↑ (F ) ≥ 0 for all
graphs with k nodes.
(d)⇒(c): Let F = (V, E1 , E2 ) be a signed graph. Then an easy computation
shows that
∑ ′ ∑
f (F ) = (−1)e(F )−|E1 | f (F ′ ) = f ↑ (F ′ ) ≥ 0.
E1 ⊆E(F ′ )⊆E1 ∪E2 E1 ⊆E(F ′ ),E2 ∩E(F ′ )=∅
V (F ′ )=V V (F ′ )=V

(c)⇒(e): Let f be an isolate-indifferent, normalized graph parameter with

nonnegative Möbius
( inverse.
) By Theorem 14.59, it defines a random graphon W
such that f = E t(·, W) .
(f)⇒(a): Every function t(., W ) is reflection positive, and this is clearly inher-
ited by their convex hull. 

Among graph parameters described above, the multiplicative ones have differ-
ent characterizations as well.
Proposition 14.61. Let f be a graph parameter satisfying the conditions of Propo-
sition 14.60. Then the following are equivalent:
(c) f is multiplicative;
(b) f = t(., W ), where W is a graphon;
(c) f is the limit of homomorphism density functions t(., G), where G is a
simple graph.
Proof. Clearly both (b) and (c) imply (a). Conversely, ( if f is)multiplicative,
then by Theorem 11.52, there is graphon W such that f = E t(., W ) . By Corollary
11.15, the function t(., W ) is the limit of homomorphism density functions for every
W. 
 14.6. EXPONENTIAL RANDOM GRAPH MODELS 259

The following proposition describes a connection between graph-theoretic and

group-theoretic properties of countable random graph models, and indicate a con-
nection with ergodic theory. Recall from Section 11.2.2 that a countable random
graph model is given by a probability measure on the sigma algebra A on the set
of graphs on N∗ , invariant under permutations of N∗ (for short, an invariant mea-
sure). We say that an invariant measure µ is ergodic, if no subset T ∈ A with
0 < µ(T ) < 1 is invariant under permutations of the nodes.
We note that invariant measures on A form a convex set K, and µ is ergodic
if and only if it is an extreme point of K. Indeed, if we can write µ = 12 (µ1 + µ2 ),
where µ1 , µ2 are different invariant measures, then the positive support of the signed
measure µ1 − µ2 is an invariant set in A with positive measure. Conversely, if
T ∈ A is an invariant set with 0 < µ(T ) < 1, then µ is a convex combination
( ) of the
invariant measures µ1 (X) = µ(X ∩ T )/µ(T ) and µ2 = µ(X \ T )/ 1 − µ(T ) .
Proposition 14.62. A countable random graph model is local if and only if it is
ergodic.
Proof. Let µ be an invariant measure on A. First, suppose that µ is ergodic,
then µ is an extreme point of K. By Theorem 14.59, there is a probability distribu-
tion π on graphons representing µ in the following sense: if W is a random graphon
drawn from this distribution, then G(N∗ , W) has distribution µ. It follows that µ
is a convex combination of distributions G(N∗ , W ) where W is in the support of
π. Since µ is an extreme point of K, the distribution π must be concentrated on
a single graphon W0 . Thus µ is the distribution of G(N∗ , W0 ), and Theorem 11.52
implies that µ is local.
Conversely, if µ is not ergodic, then we can write it as µ = 12 (µ1 + µ2 ), where
µ1 , µ2 ∈ K and µ1 ̸= µ2 . Let G1 and G2 be random countable graphs from the
distributions µ1 and µ2 , respectively, and let G be G1 with probability 1/2 and G2
with probability 1/2. Let S, T ⊆ N∗ be finite sets with |T | = |S| and T ∩ S = ∅, and
consider a labeled graph F on |S| nodes such that P(G1 [S] = F ) ̸= P(G2 [S] = F ).
Set a1 = P(G1 [S] = F ) = P(G1 [T ] = F ) (by invariance, these two probabilities are
equal), and define a2 analogously.
Thus we have
( ) ( ) ( )
P G[S] = F, G[T ] = F − P G[S] = F P G[T ] = F
1( ( ) ( ))
= P G1 [S] = F, G1 [T ] = F + P G2 [S] = F, G2 [T ] = F
2
1( ( ) ( ))( ( ) ( ))
− P G1 [S] = F + P G2 [S] = F P G1 [T ] = F + P G2 [T ] = F
4
1 2 1 1
= (a1 + a22 ) − (a1 + a2 )2 = (a1 − a2 )2 > 0.
2 4 4
This shows that the events G[S] = F and G[T ] = F are not independent, and hence
µ is not local. 
Exercise 14.63. Prove that the sigma-algebra of Borel sets in the metric space
f0 , δ ) is generated by the “semivarieties” S(F, a) = {W ∈ W
(W f0 : t(F, W ) ≥ a},
where F is a simple graph and a is a rational number.

14.6. Exponential random graph models

In this last section of our study of the graphon space, we sketch some results in
the theory of random graphs where viewing the graphon space as a single compact
260 14. THE SPACE OF GRAPHONS

metric space is crucial. Chatterjee and Varadhan [2011] applied the theory of
graph limits to the theory of large deviations for Erdős–Rényi random graphs.
This was extended by Chatterjee and Diaconis [2012] to more general distributions
on graphs, which they call exponential random graph models. We summarize their
ideas without going into the details of the proofs.
Let f be a bounded graph parameter such that for every convergent graph
sequence (Gn ), the numerical sequence f (Gn ) is convergent. Such parameters are
called estimable. The canonical examples of such parameters are subgraph densities
t(F., ), but there are many others. We will return to them in Section 15.1 to study
their estimation through sampling and other characterizations. Right now we only
need the fact that every such parameter can be extended to the graphon space
f0 so that if Gn → W then f (Gn ) → f (W ), and the extension is continuous in
W
the distance δ (in particular, the extension is invariant under weak isomorphism).
These facts are immediate consequences of the definition.
Suppose that we want to understand the structure of a random graph, but
under the condition that f (G) is small. For example, Chatterjee and Varadhan
were interested in random graphs G(n, 1/2) in which the triangle density is much
less than 1/8 (the expectation). To this end, we consider a weighting of all simple
graphs on n nodes by e−f (G)n ; this will emphasize those graphs for which f (G)
2

2
is small. The factor n in the exponent is needed to make the logarithms of the
weights to have the same order of magnitude
( ) as the logarithm of the total number of
simple graphs on [n] (which is just n2 , if you take binary logarithm). We introduce
the probability distribution φn on Fnsimp by
e−f (G)n
2

φn (G) = ∑ −f (G′ )n2

.
simp e
G′ ∈Fn

Let ψn denote the normalizing factor in the denominator. It looks quite hairy, but
Chatterjee and Diaconis derived an asymptotic formula for it. To state their result,
we need some notation. For W ∈ W0 , consider the entropy-like functional
∫
1 ( ) ( )
I(W ) = W (x, y) log W (x, y) + 1 − W (x, y) log 1 − W (x, y) .
2 [0,1]2
Chatterjee and Varadhan proved that this functional is invariant under weak iso-
f0 , δ ) (this fact is quite similar
morphism and lower semicontinuous on the space (W
to Lemma 14.16).
The formula of Chatterjee and Diaconis can be stated as follows:
Theorem 14.64. If f is an estimable graph parameter, then
( )
lim ψn = sup f (W ) − I(W ) . 
n→∞ W ∈W0

Using this, they prove the following result about the behaviour of a random
graph drawn from the distribution φn . Since f (W )−I(W ) is upper semicontinuous
on the compact space (W f0 , δ ), the supremum in the above formula is in fact a
maximum, and it is attained on a compact set Kf ⊆ W0 .
Theorem 14.65. Let f be an estimable graph parameter, and let Gn be a random
graph from the distribution φn . Then for every η > 0 there are C, ε > 0 such that
P(δ (WGn , Kf ) > η) ≤ Ce−εn .
2

 14.6. EXPONENTIAL RANDOM GRAPH MODELS 261

This implies that if we choose a random Gn from φn for every n, then

δ (WGn , Kf ) → 0 with probability 1. If Kf consists of a single graphon W0 (which
is in a sense the “generic” case), then Gn → W0 with probability 1.
Theorems 14.64 and 14.65 provide a framework for analyzing the behavior of
exponential random graph models. This is by no means easy, and the results are
interesting. Most work has been done for the parameters of the form f (G) =
β1 t(K2 , G) + β2 t(K3 , G) (extending, in a sense, our discussions in Sections 2.1.1
and 16.3.2). We refer to Chatterjee and Diaconis [2012], Aristoff and Radin [2012],
and Radin and Yin [2012] for recent results.
 CHAPTER 15

Algorithms for large graphs and graphons

We have seen in the Introduction that different kinds of algorithmic questions

can be asked for a very large graph: we may want to estimate a parameter, test
a property, or compute (in some sense) an additional structure for the graph. We
sneak in a fourth one, making a distinction between “property distinction” and
“property testing”. We will see that the theory of graph limits and other methods
developed in this book provide valuable tools for the theoretical understanding of
all these types of algorithmic problems.

15.1. Parameter estimation

We want to determine some parameter of a very large graph G. Of course, we’ll
not be able to determine the exact value of this parameter; the best we can hope
for is that if we take a sufficiently large sample, we can find the approximate value
of the parameter with high probability.
To be precise, a graph parameter f is estimable, if for every ε > 0 there is a
positive integer k such that if G is a graph with at least k nodes and we select a
random k-set X of nodes of G, then from the subgraph G[X] induced by them we
can compute an estimate g(G[X]) of f such that

(15.1) P(|f (G) − g(G[X])| > ε) < ε.

We call the parameter g a test parameter for f . However, we don’t really need
this notion: we can always use g = f (cf. Goldreich and Trevisan [2003]). Indeed,
(15.1) implies that P(|f (G[X]) − g(G[X])| > ε) < ε, and so

P(|f (G) − f (G[X])| > 2ε) ≤ P(|f (G) − g(G[X])| > ε)

+ P(|g(G[X]) − f (G[X])| > ε) < ε + ε = 2ε,

so we can choose the threshold k belonging to ε/2 in the original definition to get
the condition obtained by replacing g by f .
It is easy to see that estimability is equivalent to (saying )that for every con-
vergent graph sequence (Gn ), the sequence of numbers f (Gn ) is convergent. (So
graph parameters of the form t(F, .) are estimable by the definition of convergence.)
Using this, for any estimable parameter f we can define a functional fb on W0 , where
fb(W ) is the limit of f (Gn ) for any sequence of simple graphs Gn → W . It is also
immediate that this functional fb is continuous on (W f0 , δ ). The functional fb does
not determine the graph parameter f : defining f0 (G) = fb(WG ) we get a graph
parameter with fb0 = fb, but f could be any parameter of the form f0 + h, where
h(G) → 0 if v(G) → ∞.
263
264 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

All this is, however, more-or-less just a reformulation of the definition. Borgs,
Chayes, Lovász, Sós and Vesztergombi [2008] gave a number of more useful condi-
tions characterizing testability of a graph parameter. We formulate one, which is
perhaps easiest to verify for concrete parameters.
Theorem 15.1. A graph parameter f is estimable if and only if the following three
conditions hold:
(i) If Gn and G′n are simple graphs on the same node set (n = 1, 2, . . . ) and
d (Gn , G′n ) → 0, then f (Gn ) − f (G′n ) → 0.
( )
(ii) For every simple graph G, f G(m) has a limit as m → ∞ (recall that
G(m) denotes the graph obtained from G by blowing up each node into m twins).
(iii) f (GK1 ) − f (G) → 0 if v(G) → ∞ (recall that GK1 is obtained from G by
adding a single isolated node).
Note that all three conditions are special cases of the statement that
(iv) if |V (Gn )|, |V (G′n )| → ∞ and δ (Gn , G′n ) → 0, then f (Gn ) − f (G′n ) → 0.
This condition is also necessary, so it is equivalent to its own three special cases
(i)–(iii) in the Theorem.
Proof. The necessity of condition (iv) (which implies (i)–(iii)) is easy: Sup-
pose that there are two sequences of graphs Gn such that |V (Gn )|, |V (G′n )| → ∞
and δ (Gn , G′n ) → 0, but f (Gn ) − f (G′n ) ̸→ 0. By selecting a subsequence, we
may assume that |f (Gn ) − f (G′n )| > ε for all n for some ε > 0. Going to a further
subsequence, we may assume that the sequences (G1 , G2 , . . . ) and (G′1 , G′2 , . . . )
are convergent. But then δ (Gn , G′n ) → 0 implies that the interlaced graph se-
quence (G1 , G′1 , G2 , G′2 , . . . ) is convergent as well. However, the numerical sequence
(f (G1 ), f (G′1 ), f (G2 ), f (G′2 ), . . . ) is not convergent, a contradiction.
To prove the sufficiency of (i)–(iii), we start with proving the following stronger
form of (i):
(i’) If Gn and G′n are simple graphs with the same number of nodes (n =
1, 2, . . . ) and δ (Gn , G′n ) → 0, then f (Gn ) − f (G′n ) → 0.
This follows by Theorem 9.29, which implies that one can overlay the graphs
Gn and G′n so that d (Gn , G′n ) → 0.
Consider a convergent graph sequence (G1 , G2 , . . . ), we prove that the sequence
(f (G1 ), f (G2 ), . . . ) is convergent. Let ε > 0. Using (i’), we can choose an ε1 > 0
so that if δ (G, G′ ) ≤ ε1 , then |f (G) − f (G′ ) ≤ ε. Since the graph sequence is
convergent, we can choose and( fix an integer n ≥ 1 so )that δ (Gn , Gm ) ≤ ε1 /2 for
m ≥ n. By (ii), the sequence f (Gn (p)) : p = 1, 2, . . . (is convergent. )  Let a be its
limit, then we can choose a threshold p0 ≥ 1 so that f Gn (p) − a ≤ ε for every
integer p ≥ p0 . We may assume that p0 ≥ 4/ε1 . Finally, based on (iii), we can
chose a threshold q ≥ 1 such that |f (GK1 ) − f (G)| ≤ ε/v(Gn ) whenever v(G) ≥ q.
Now consider a member Gm of the sequence for which m ≥ n and v(Gm ) ≥
max(q, p0 v(Gn ), 4v(Gn )/ε1 ). We can write v(Gm ) = pv(Gn ) + r, where p ≥ p0 and
0 ≤ r < v(Gn ). Then Gm and G′ = Gn (p)K1r have the same number of nodes.
Furthermore,
( )
δ (Gm , G′ ) ≤ δ Gm , Gn (p) + δ (Gn (p), G′ )
2r
≤ δ (Gm , Gn ) + ≤ ε1 .
pv(Gn )
 15.1. PARAMETER ESTIMATION 265

Hence |f (Gm ) − f (G′ )| ≤ ε, and so

   ( ) ( ) 
f (Gm ) − a| ≤ |f (Gm ) − f (G′ ) + f (G′ ) − f Gn (p) | + |f Gn (p) − a
ε
≤ε+r + ε < 3ε.
v(Gn )
So for any two indices m1 and m2 that are large enough, we have |f (Gm1 ) −
f (Gm2 )| < 6ε, which proves that f is estimable. 
Example 15.2 (Maximum cut). As a basic example, consider the density of
maximum cuts (recall Example 5.18). One of the first substantial results on prop-
erty testing (Goldreich, Goldwasser and Ron [1998], Arora, Karger and Karpinski
[1995]) is that this parameter is estimable. In the introduction we gave an argu-
ment (which can be made precise using high concentration results like Azuma’s
inequality) that if S is a sufficiently large random subset of nodes of G, then
maxcut(G[S]) ≥ maxcut(G) − ε: a large cut in G, when restricted to S, gives a
large cut in G[S]. It is harder, and in fact quite surprising, that if most subgraphs
G[S] have a large cut, then so does G. This follows from Theorem 15.1 above, since
conditions (i)–(iii) are easily verified for f = maxcut. 
Example 15.3 (Free energy). The “free energy” is a statistical physical quantity.
Recall the definition of the energy of a map σ : V (G) → [q] (a configuration) from
the introduction:
∑
(15.2) H(σ) = − Jφ(u),φ(v) ,
uv∈E(G)

and also the partition function

∑
(15.3) Z= e−H(σ)/T ,
σ:V (G)→[q]

where T is the temperature (for simplicity, we don’t consider an external field).

The mean field partition function of G can be obtained (formally) by considering a
very high temperature:
∑
(15.4) Zmean = e−H(σ)/v(G) .
σ:V (G)→[q]

The free energy is defined by

ln Z(G, H)
(15.5) F(G, H) = − .
v(G)
It would exceed the framework of this book to explain the physics behind these
names; let us just treat them as graph parameters related to homomorphism num-
bers. Note that the normalization is different from (2.15) in the exponent and
therefore we only divide by v(G) (as opposed to (5.33), for example). For more
about this connection, we refer to Borgs, Chayes, Lovász, Sós and Vesztergombi
[2012].
The free energy (for a fixed weighted graph H) is a more complicated example
of a estimable parameter, which illustrates the power of Theorem 15.1. It is difficult
to verify directly either the definition, or say condition (iv). The theorem splits this
task into three: condition (i) is easy by the definition of d (G, G′ ); (ii) is an exercise
in classical combinatorics, in which we have to count mappings that split the twin
classes in given proportions; finally, (iii) is trivial. 
266 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

Exercise 15.4. Show that neither one of the three conditions in Theorem 15.1
can be dropped.
Exercise 15.5. Use Theorem 15.1 to prove that the i-th largest eigenvalue of
a graph is an estimable parameter for every fixed i ≥ 1. Give a new proof of
Theorem 11.54 based on this argument.
Exercise 15.6. Fix a graphon W , then cut(W, F ), as a function of F , defines a
simple graph parameter. Prove that it is estimable.

15.2. Distinguishing graph properties

An algorithmic task of different nature is to test whether a given graph G has a
certain property (e.g., is it connected, bipartite, or perfect). Before (e.g. in Section
4.3) we considered graph properties as 0-1 valued graph parameters. In this case,
this would not be a useful approach, at least not if we wanted to reduce the question
to parameter estimation: getting a 0-1 valued parameter with less than 50 percent
error is tantamount to getting it exactly, which is clearly too much to require for
very large graphs. Therefore we have to modify the definition to allow an error
(which is then equal to 1, so very large in our setting), but with small probability.
We start with a version of the problem that turns out simpler: how to distin-
guish two properties? Deciding whether a graph has a given property P will be
then treated as distinguishing graph property P from the set of graphs that are far
from having this property.
Let P1 and P2 be two graph properties that are exclusive (i.e., P1 ∩ P2 = ∅).
We want to design a method that, given an arbitrarily large graph G, looks at a
sample of some fixed size k, and guesses whether the graph has property P1 or
P2 . If the graph has property P1 or property P2 , we would like to guess right with
rather high probability, say at least 2/3. If the graph does not have either one of
the properties, then we don’t care what the guess was.
The guess will be based on a third graph property Q, which we call the test
property. If the sample G(k, G) has property Q, we guess that G has property P1 ;
else, we guess that G has property P2 . So the precise definition is the following:
we call properties P1 and P2 distinguishable by sampling, if there exists a positive
integer k and a test property Q such that for every graph G with at least k nodes
 2

( ) ≥ 3 , if G ∈ P1 ,
P G(k, G) ∈ Q
 1
≤ , if G ∈ P2 .
3
The following lemma shows that the numbers 1/3 and 2/3 in this definition are
arbitrary; we could replace them with any two numbers a and b with 0 < a < b < 1.
Furthermore, we can replace the number k by every sufficiently large integer.

Lemma 15.7. Let P1 and P2 be two graph properties with P1 ∩ P2 = ∅. Let

0 < a < b < 1 and 0 < c < d < 1. Suppose that there is a positive integer k and a
test property Q such that for every graph G with at least k nodes
{
( ) ≥ b if G ∈ P1 ,
P G(k, G) ∈ Q
≤ a if G ∈ P2 .
 15.2. DISTINGUISHING GRAPH PROPERTIES 267

Then for every positive integer k ′ that is large enough there is a test property Q′
such that for every graph G with at least k ′ nodes
{
( ) ≥ d if G ∈ P1 ,
P G(k ′ , G) ∈ Q′
≤ c if G ∈ P2 .

Proof. Let k ′ > k. For every graph F on k ′ nodes, let

f (F ) = P(G(k, F ) ∈ Q),

and define the property Q′ as the set of graphs F on k ′ nodes such that f (F ) ≥
(a + b)/2. ( )
Let G ∈ P1 , v(G) ≥ k ′ . Since G k, G(k ′ , G) is a random k-node subgraph of
G, we have
( )
f0 = E f (G(k ′ , G)) = P(G(k, G) ∈ Q) ≥ b.

Furthermore, the graph parameter f has the property that if we change edges in F
incident with a given node v, then the validity of the event G(k, F ) ∈ Q changes
only if the random k-subset contains v, which happens with probability k/k ′ . So
the value f (F ) changes by at most k/k ′ . We can apply the Sample Concentration
Theorem 10.2 to the parameter (k ′ /k)f , and get that
( ) ( a + b) ( b − a)
P G(k ′ , G) ∈
/ Q′ = P f (G(k ′ , G) ≤ ≤ P f (G(k ′ , G) ≤ f0 − ≤ e−t ,
2 2
where t = (b − a)2 k ′ /(8k 2 ). Choosing k ′ large enough, this will be less than 1 − d,
proving that Q′ and k ′ satisfy the first condition in the lemma. The second condition
follows similarly. 

Using our Sampling Lemmas, we can give the following characterization of

distinguishable properties.

Theorem 15.8. For two graph properties P1 and P2 , the following are equivalent:
(a) P1 and P2 are distinguishable by sampling;
(b) there exists a positive integer k such that for any Gi ∈ Pi with v(Gi ) ≥ k,
we have δ (G1 , G2 ) ≥ 1/k;
(c) there exists a positive integer k such that for any Gi ∈ Pi with v(Gi ) ≥ k,
we have
( ) 1
dvar G(k, G1 ), G(k, G2 ) ≥ .
3
Note that (b) could be phrased as P1 ∩ P2 = ∅.

Proof. (a)⇒(c): Let P1 and P2 be distinguishable with sample size k and test
property Q. Then for any two graphs G1 , G2 ∈ Pi , we have
( ) ( ) 1
P G(k, G1 ) ∈ Q − P G(k, G2 ) ∈ Q ≥ .
3
268 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

On the other hand,

( ) ( )
P G(k, G1 ) ∈ Q − P G(k, G2 ) ∈ Q
∑( ( ) ( ))
= P G(k, G1 ) = F − P G(k, G2 ) = F
F ∈Q
∑ ( ) ( )
≤ P G(k, G1 ) = F − P G(k, G2 ) = F 
+
F
( )
= dvar G(k, G1 ), G(k, G2 ) .
This proves (c).
(c)⇒(b): This follows immediately from the Counting Lemma 10.22.
(b)⇒(a): Let δ (P1 , P2 ) = c > 0. Let k be large enough, and define the test
property Q by
{ c}
(15.6) Q = F : v(F ) = k, δ (F, P1 ) ≤ .
2
To see that this is a valid test property, consider any graph G ∈ P1 with v(G) ≥ k.
By the Second Sampling Lemma 10.15, we have with probability at least 2/3 that
( ) 20 c
δ G, G(k, G) ≤ √ < ,
log k 2
and if this happens, then G(k, G) ∈ Q. The condition on graphs in P2 follows
similarly. 

As a consequence of this proof, we can make the following observation: While

the test property Q looks like the key to this testing method, it can in fact be
chosen in a very specific way, in the form (15.6).
We can also talk about distinguishing two graphon properties by testing. We
assume that we can get information about a given graphon W by generating a
W -random graph G(k, W ). All the above can be repeated in this model, and we
will not go through the details.

15.3. Property testing

Testing for a single property is a more complicated business than distinguishing
two properties. There is a large literature on this subject; here we restrict our
attention to aspects that are related to graph limit theory, based on Lovász and
Szegedy [2010a]. We start with a discussion of testing for a graphon property (which
is easily handled using the results about distinguishing two properties), and then
add the necessary work to apply this to testing graph properties.

15.3.1. Testable graphon properties. We call a graphon property R

testable if it is closed in the δ metric, and there is a graph property R′ (called the
test property for R), such that
(a) for every graphon W ∈ R and every k ≥ 1, we have G(k, W ) ∈ R′ , with
probability at least 2/3, and
(b) for every ε > 0 there is a kε ≥ 1 such that for every graphon W with
d1 (W, R) > ε and every k ≥ kε we have G(k, W ) ∈ / R′ with probability at least
2/3.
 15.3. PROPERTY TESTING 269

The definition above is clearly one-sided, there is no symmetry between the

property and its complement. Several other versions could be defined, but we’ll
restrict our attention to this one. We could require (a) only for k ≥ k0 , but we can
put all smaller graphs into R′ for free, so this would not give anything different.
On the other hand, we need a threshold in (b) to depend on ε: for a fixed k, if
d1 (W, R) is very small, then there is a graphon U ∈ R such that d1 (W, U ) is very
small, and the distributions of G(k, U ) and G(k, W ) are almost the same, so no test
property R′ can distinguish them.
Example 15.9 (Complete or edgeless). Let R be the graphon property satisfied
by the identically-0 and identically-1 graphons. Then the graph property “complete
or edgeless” is a good test property for R, so R is testable. 
Example 15.10 (Constant graphon). Let R be the graphon property satisfied
by the identically-1/2 graphon U . We show that this property is not testable.
Consider a random graph Gn = G(n, 1/2); then ∥WGn − U ∥ → 0 with probability
1. Fix a sequence for which this happens. The distribution of G(k, WGn ) tends to
the distribution of G(k, 1/2) for every fixed k, so for every possible test property
R′ , if G(k, 1/2) ∈ R′ with probability at least 2/3, then G(k, WGn ) ∈ R′ with
probability at least 1/2 (for every k, if n is large enough). On the other hand,
d1 (WGn , R) = 1/2 for every n, and hence we should have G(k, WGn ) ∈ / R′ with
probability at least 2/3 if k is large enough.
We note that the complementary property W0 \R is testable: since d1 (W, Rc ) =
0 for every graphon W , the identically true property as a good test property. 
Similarly as for property distinguishing, one may feel that the tricky choice of
the test property R′ is crucial. but in fact, once a property is testable, we can use
a very simple test property:
Proposition 15.11. Let R be a testable graphon property. Then
{ 20 }
R′ = F : v(F ) = 1 or δ (WF , R) ≤ √
log v(F )
is a valid test property for R.
Proof. First, suppose that W ∈ R, and let k ≥ 2. By the Second Sampling
Lemma 10.16, we have
( ) 20
δ W, G(k, W ) ≤ √
log k
with probability at least 2/3. Thus G(k, W ) ∈ R′ with probability larger than 2/3.
Second, let R′′ be any valid test property for R. By definition, for every ε > 0
there is a k ≥ 1 (depending on ε, R and R′′ ) such that whenever d1 (W, R) > ε
for some graphon W , then P(G(k, W ) ∈ R′′ ) ≤ 1/3. Let U ∈ R, then P(G(k, U ) ∈
R′′ ) ≥ 2/3. This implies that the variation distance of the distributions of G(k, W )
and G(k, U ) is at least 1/3. By Corollary 10.25, this implies that δ (W, U ) ≥ 13 2−k .
2

This holds for all U ∈ R, so δ (W, R) ≥ 13 2−k . Let n be large enough (depending
2

on k), and consider the W -random graph G(n, W ), and the corresponding graphon
Wn = WG(n,W ) . Then with high probability
( ) 1 ( ) 1 20 20
δ Wn , R ≥ 2−k − δ Wn , W ≥ 2−k − √
2 2
>√ .
3 3 log n log n
270 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

/ R′ with high probability, and this shows that R′ is a valid test

So G(n, W ) ∈
property. 

It is not always easy to decide about a graphon property whether it is testable,

and we will have to develop some theory to prove properties of further, more inter-
esting examples. We start with showing that testability and distinguishability are
closely related. Let Rcε denote the set of graphons W such that d1 (W, R) ≥ ε.
Proposition 15.12. A graphon property R is testable if and only if it is distin-
guishable by sampling from the property Rcε for every ε > 0. 
Proof. The “only if” part is straightforward to check. To verify the “if” part,
suppose that R and Rcε are distinguishable by sampling for every ε > 0. Using
Lemma 15.7, this means that for every ε > 0 there is a positive integer kε and a
test property Qε such that for every k ≥ kε , we have
 2

( ) ≥ 3 , if G ∈ R,
P G(k, W ) ∈ Qε

≤ 1 , if G ∈ Rc .
ε
3
We may assume that kε increases if ε decreases.
One difficulty is that we need to define a single test property Q, not the family
{Qε }. But this is easy. Let F ∈ Q if and only if k1/m ≤ v(F ) < k1/(m+1) and
F ∈ Q1/m for some positive integer m. Then Q works for every ε > 0.
The second difficulty is that we want P(G(k, W ) ∈ Q) for all k, not just for
k ≥ kε . But this property holds for k ≥ k1 by definition, so all we have to do is to
include all graphs with fewer that k1 nodes in Q. 

The following corollary to Theorem 15.8 and Proposition 15.12 provides an

analytic characterization of testable graphon properties. Recall that the distances
d1 and d are related trivially by d ≤ d1 . Testability of a property means, in a
sense, an inverse relation:
Corollary 15.13. A closed graphon property R is testable if and only if
δ (R, Rcε ) > 0 for every ε > 0.
The condition given in this corollary can be rephrased in various ways, for
example: For every sequence of graphons (Wn ) such that d (Wn , R) → 0, we have
d1 (Wn , R) → 0. We show that this condition is equivalent to a seemingly weaker
condition. (Recall the definition of flexing from Section 14.3.3.)
Lemma 15.14. A closed graphon property R is testable if and only if for every
U ∈ R and every sequence of graphons (Wn ) such that Wn → U and every Wn is
a flexing of U , we have d1 (Wn , R) → 0.
Proof. Let (Wn ) be a sequence of graphons such that d (Wn , R) → 0. If
d1 (Wn , R) 9 0, then we may take a subsequence for which lim inf d1 (Wn , R) > 0,
and then a further subsequence such that δ (Wn , U ) → 0 for some U ∈ R. Define
{
′ U (x, y) if U (x, y) ∈ {0, 1},
Wn (x, y) =
Wn (x, y) otherwise.
 15.3. PROPERTY TESTING 271

Then Wn′ is a flexing of U . Furthermore, we have

∫ ∫
∥Wn − Wn′ ∥1 = Wn + (1 − Wn ) → 0
U =0 U =1

by Lemma 8.22. This implies that δ (Wn , Wn′ ) → 0, and hence δ (Wn′ , U ) → 0.
By the hypothesis of the lemma implies that d1 (Wn′ , R) → 0. Hence
d1 (Wn , R) ≤ d1 (Wn′ , R) + ∥Wn − Wn′ ∥1 → 0. 
Since the condition of the last lemma is trivially fulfilled if R is flexible, we get
a useful corollary:
Corollary 15.15. Every closed flexible graphon property is testable. In particular,
the closure of every hereditary property is testable.
The following result can be viewed as the graphon analogue of the theorem of
Fischer and Newman [2005] (from which the finite theorem can be derived).
Theorem 15.16. A closed graphon property R is testable if and only if the func-
tional d1 (., R) is continuous in the cut norm.
Proof. If d1 (., R) is continuous, then Corollary 15.13 implies that R is testable.
Suppose that R is testable. The functional d1 (., R) is lower semicontinuous in
the cut norm by Lemma 14.15. To prove upper semicontinuity, let W, Wn ∈ W0
and let ∥Wn − W ∥ → 0. We claim that lim supn d1 (Wn , R) ≤ d1 (W, R).
Let ε > 0, and let U ∈ R be such that ∥W − U ∥1 ≤ d1 (W, R) + ε. By
Proposition 8.25, there is a sequence of graphons Un such that ∥Un − U ∥ → 0 and
∥Un − Wn ∥1 → ∥U − W ∥1 . By Corollary 15.13 (in the form in the remark after its
statement), it follows that d1 (Un , R) → 0, and so
d1 (Wn , R) ≤ ∥Wn − Un ∥1 + d1 (Un , R) → ∥U − W ∥1 .
Hence
lim sup d1 (Wn , R) ≤ ∥U − W ∥1 ≤ d1 (W, R) + ε.
n→∞
Since ε > 0 is arbitrary, this implies that d1 (., R) is upper semicontinuous. 
Example 15.17 (Neighborhood of a property). Let S ⊆ W0 be an arbitrary
graphon property and let a > 0 be an arbitrary number. Then the property R =
{U ∈ W0 : δ (U, S) ≤ a} is testable.
To show this, we use Corollary 15.13. For ε > 0 define ε′ = aε/2. Let W ∈
B (R, ε′ ). Then W ∈ B (S, a+ε′ ), and so there is a U ∈ S such that ∥U −W ∥ ≤
a + 2ε′ . Consider Y = (1 − ε)W + εU . Then ∥Y − U ∥ = ∥(1 − ε)(U − W )∥ ≤
(1 − ε)(a + 2ε′ ) < a, so Y ∈ R. Furthermore, ∥W − Y ∥1 = ∥ε(W − U )∥1 ≤ ε, and
so W ∈ B1 (R, ε). Since W was an arbitrary element of B (R, ε′ ), this implies that
δ (R, Rcε ) > ε′ . 
Example 15.18 (Subgraph density). For every fixed graph F and 0 < c < 1,
the property R of a graphon W that t(F, W ) = c is testable. Let us verify that for
every ε > 0 there is an ε′ > 0 such that d1 (W, R) ≥ ε implies that d (W, R) ≥ ε′ .
Assume that d1 (W, R) ≥ ε, then t(F, W ) ̸= c; let (say) t(F, W ) > c. The graphons
Us = (1 − s)W , 0 ≤ s ≤ ε, are all in B1 (W, ε), and hence not in R. It follows that
t(F, Us ) > c for all 0 ≤ s ≤ ε. Since t(F, Uε ) = (1 − ε)e(F ) t(F, W ), this implies
that t(F, W ) > (1 − ε)−e(F ) c. Thus for every U ∈ R we have t(F, W ) − t(F, U ) ≥
272 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

( )
(1 − ε)−e(F ) − 1 c. By the Counting Lemma 10.23 this implies that δ (U, W ) ≥
( )
(1 − ε)−e(F ) − 1 c/e(F ). Choosing the right hand side in this inequality as ε′ , we
get that δ (W, R) ≥ ε′ , which we wanted to verify.
Fixing two subgraph densities, however, may yield a non-testable property: for
example, t(K2 , W ) = 1/2 and t(C4 , W ) = 1/16 imply that W ≡ 1/2 (see Section
1.4.2), and we have seen that this graphon property is not testable. 
15.3.2. Testable graph properties. We call a graph property P testable, if
for every ε > 0, graphs in P can be distinguished from graphs farther than ε from
P in the edit distance. To be more precise, recall that d1 (F, G) is defined for two
graphs on the same node set, and it denotes their normalized edit distance. So for
a graph G on n nodes, d1 (G, P) is the minimum number of edges to be changed in
order to get a graph with property P, divided by n2 . If there is no graph in P on
n nodes, then we define d1 (F, P) = 1. Let Pεc denote the set of simple graphs F
such that d1 (F, P) ≥ ε; then we want to distinguish P from Pεc by sampling.
This notion of testability is usually called oblivious testing, which refers to the
fact that no information about the size of G is assumed.
Using our analytic language, we can give several reformulations of the definition
of testability of a graph property, which are often more convenient to use.
(T1) if (Gn ) is a sequence of graphs such that v(Gn ) → ∞ and δ (Gn , P) → 0,
then d1 (Gn , P) → 0;
(T2) for every ε > 0 there is an ε′ > 0 such that if G and G′ are simple
graphs such that v(G′ ), v(G′ ) ≥ 1/ε′ , G ∈ P and δ (G, G′ ) < ε′ , then
d1 (G′ , P) < ε;
(T3) P ∩ Pεc = ∅ for every ε > 0.
The equivalence of these with testability follows by Theorem 15.8.
From this characterization of testability it follows that if P is a testable property
such that infinitely many graphs have property P, then for every n that is large
enough, it contains a graph on n nodes. Indeed, suppose that for infinitely many
n, P contains a graph Gn on n nodes but none on n + 1 nodes. We may assume
that Gn → W ∈ W0 . Then Gn K1 ∈ P1/2 c
and Gn K1 → W , so W ∈ P ∩ Pεc , a
contradiction.
It is surprising that this rather restrictive definition allows many testable graph
properties: for example, bipartiteness, triangle-freeness, every property definable
by a first order formula (Alon, Fischer, Krivelevich and Szegedy [2000]). Let us
begin with some simple examples.
Example 15.19 (Nonempty). Let P be the graph property that “G has at least
one edge”. This is testable with the identically true test property. 
This example sounds like playing in a trivial way with the definition. The
following examples are more substantial.
Example 15.20 (Large clique). Let P be the graph property ω(G) ≥ v(G)/2.
Then P is testable. This can be verified using (T2): we show that for every ε > 0
there is an ε′ > 0 such that δ (G, P) ≤ ε′ implies that d1 (G, P) ≤ ε. We show that
ε′ = exp(−10000/ε6 ) does the job.
Indeed, if δ (G, P) ≤ ε′ , then there is (a graph H ∈) P such that δ (G, H) ′
( ≤ )2ε .
′
Let V (H) = [q] and V (G) = [p], then δ G(q), H(p) ≤ 2ε , and since v G(q) =
( ) ( ) √
v H(p) = pq, Theorem 9.29 implies that δb G(q), H(p) ≤ 45/ − log(2ε′ ) <
 15.3. PROPERTY TESTING 273

( )
ε3 /2, which means that G(q) and H(p) can be overlaid so that d G(q), H(p) ≤
ε3 /2. Since H ∈ P, it contains a complete graph of size at least v(H)/2, and so
H(p) contains a complete graph K of size at least pq/2. Now the definition of
the cut distance implies that G(q)[V (K)] is almost complete: it misses at most
ε3 (pq)2 /2 edges. Let qi denote
∑ the number of nodes in K that come from node i of
G, so that 0 ≤ qi ≤ q and i qi = pq. For simplicity of presentation, assume that
p = 2r is even and ε ≤ 1/2. Also assume that q1 ≥ q2 ≥ · · · ≥ qp .
Let k denote the number of qi with i ≤ r and qi ≤ εq. We claim that k ≤ 2rε.
We
∑ may assume that k > 0, then qr+1 , . . . , qp < εq, and so (r − k)q + (r + k)εq ≥
i qi = rq. This implies the bound on k. Let G1 = G[1, . . . , r], then
1 3 2 2 ∑
ε p q ≥ qi qj ≥ (e(G1 ) − kr)(εq)2 ,
2
ij∈E(G1 )

whence
ε ε ε
e(G1 ) ≤ kr + p2 ≤ p2 + p2 < εp2 .
2 2 2
So adding at most εp2 edges to G, we can create a complete subgraph with p/2
nodes, showing that d1 (G, P) ≤ ε. 
Example 15.21 (Triangle-free). Let P be the property of being triangle-free.
Then P is a valid test property for itself. It is trivial that if G is triangle-free, then
any sample G(k, G) is also triangle-free. The other condition is, however, far from
being trivial. If G(k, G) is triangle-free with probability at least 2/3, and k is large
enough, then G has very few triangles. Hence by the Removal Lemma 11.64, we
get that we can change (in this case, delete) a small number of edges so that we
get rid of all triangles. In fact, the Removal Lemma is equivalent to the testability
of triangle-freeness.
Theorem 15.24 below will give a general sufficient condition for testability,
which will imply the Removal Lemma. 
There is a tight connection between testability of graph properties and the
testability of their closures. To formulate it, we need a further definition. A graph
property P is robust, if for every ε > 0 there is an ε0 > 0 such that if G is a
graph with v(G) ≥ 1/ε0 and d1 (WG , P) ≤ ε0 , then d1 (G, P) ≤ ε. Another way
of stating this is that if (Gn ) is a sequence of graphs such that d1 (WGn , P) → 0,
then d1 (Gn , P) → 0. (For a more combinatorial formulation of this property, see
Exercise 15.30.)
Theorem 15.22. (a) A graphon property is testable if and only if it is the closure
of a testable graph property.
(b) A graph property is testable if and only if it is robust and its closure is
testable.
It is not the first time in this book that results about graphons are nice and
easy, but to describe the connection between the notions for graphons and the
corresponding notions for graphs is the hard part. This is true in this case too, and
we will omit some details of the rather long proof of this theorem; see Lovász and
Szegedy [2010a] for these details.
Before going into the proof, let us look at an example.
274 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

Example 15.23 (Alternating). A graph property with a testable closure need

not be itself testable. Let P be the graph property that the graph is complete if
the number of nodes is even, but edgeless if the number of nodes is odd. Clearly,
this is not testable. The closure of P is the graphon property valid for W ≡ 0 and
W ≡ 1, which is testable (Example 15.9). 
Proof. We start with proving that the closure of a testable graph property
is testable. It suffices to prove that if (Wn ) is a sequence of graphons such that
d (Wn , P) → 0, then d1 (Wn , P) → 0. We may assume that the sequence Wn is
convergent, so Wn → U for some U ∈ W0 (in the δ distance). Clearly U ∈ P, so
by the definition of closure, there are graphs Hn ∈ P such that Hn → U .
Fix any ε > 0. By Theorem 15.16, there is an ε′ > 0 such that if |V (G)|, |V (H)|
are large enough, H ∈ P, and δ (G, H) < ε′ , then d1 (G, P) < ε. Furthermore,
there is an nε ≥ 1 such that if n ≥ nε , then δ (WHn , U ), δ (Wn , U ) ≤ ε′ /3.
Fix any n ≥ nε , and let Gn,m (m = 1, 2, . . . ) be a sequence of graphs such that
Gn,m → Wn as m → ∞. Then, provided m is large enough,
δ (Hn , Gn,m ) ≤ δ (WHn , U ) + δ (U, Wn ) + δ (Wn , WGn,m ) < ε′ .
Hence by the choice of ε′ , we have d1 (Gn,m , P) ≤ ε. This means that there are
graphs Jn,m ∈ P with V (Jn,m ) = V (Gn,m ) such that d1 (Gn,m , Jn,m ) ≤ ε. By
choosing a subsequence, we can assume that Jn,m → Un as m → ∞ for some
Un ∈ P. Applying Lemma 14.16 we obtain that
d1 (Wn , P) ≤ δ1 (Wn , Un ) ≤ lim inf δ1 (WGn,m , WJn,m ) ≤ lim inf d1 (Gn,m , Jn,m ) ≤ ε.
m→∞ m→∞

The proof of the converse in (a) (which we will not use in this book) is omitted.
Next we show that every testable graph property P is robust. Let (Gn ) be a
sequence of graphs such that d1 (WGn , P) → 0, then d (WGn , P) → 0. This implies
that there are graphons Un ∈ P such that d (WGn , Un ) → 0. By the definition of P,
this implies that there are simple graphs Hn ∈ P such that δ (Gn , Hn ) → 0. This
means that δ (Gn , P) → 0. Since P is testable, this implies that d1 (Gn , P) → 0.
Finally, we show that if a graph property P is robust and its closure is testable,
then P is testable. Let (Gn ) be a sequence of graphs such that d (Gn , P) → 0.
Then d (WGn , P) → 0. Since P is testable, this implies that d1 (WGn , P) → 0. By
robustness, we get that d1 (Gn , P) → 0, which proves that P is testable. 

15.3.3. Hereditary and testable properties. A surprisingly general suffi-

cient property for testability was found by Alon and Shapira [2008].
Theorem 15.24 (Alon–Shapira). Every hereditary graph property is testable.
We have seen that the Removal Lemma is perhaps the simplest nontrivial spe-
cial case of this theorem. The analytic proof of the Removal Lemma we described
in Section 11.8 can be extended; we don’t go into the details of this, instead we
state and prove a more general result, which gives a full characterization of testable
graph properties in terms of being “almost hereditary”.
Testable properties are not necessarily hereditary, as Examples 15.19 and 15.20
show, so this condition is not necessary and sufficient. But there is a weaker version
of heredity that can be used to characterize testability: we consider only those
induced subgraphs that are close to the original graph in the cut-distance. (Recall
that by the Second Sampling Lemma 10.15, almost all sufficiently large subgraphs
 15.3. PROPERTY TESTING 275

have this property.) Clearly this theorem implies Theorem 15.24. The proof will
be quite involved, using much of the material developed earlier in this chapter.
Theorem 15.25. A graph property P is testable if and only if for every ε > 0
there is an ε′ > 0 such that if H ∈ P and G is an induced subgraph of H with
v(G) ≥ 1/ε′ and δ (G, H) < ε′ , then d1 (G, P) < ε.
Another way to state the condition is that if Hn ∈ P (n = 1, 2, . . . ) is a
sequence of simple graphs, Gn is an induced subgraph of Hn , and δ (Gn , Hn ) → 0,
then d1 (Gn , P) → 0. Informally, induced subgraphs inherit the property of the
big guy, but they have to pay an inheritance tax; the tax is however small if the
descendants are also big and they are close to the big guy.

Proof. The “only if” part is trivial, since the condition is a special case of the
reformulation (T2) of testability. By theorem 15.22, it suffices to prove that P is
testable and P is robust.
We start with proving a graphon version of the condition in the theorem.
Claim 15.26. Let U ∈ P and let (Gn ) be a sequence of simple graphs with Gn → U .
Also assume that tind (Gn , U ) > 0. Then d1 (Gn , P) → 0.
Since U ∈ P, there is a sequence of simple graphs Hm ∈ P (m = 1, 2, . . . )
such that Hm → U . Condition tind (Gn , U ) > 0 implies that tind (Gn , Hm ) > 0 for
every n if m is large enough. Furthermore, both Gn → U and Hm → U , and hence
δ (Gn , Hm ) → 0 if n, m → ∞. So if n is large enough,
( we can) select an m(n) such
that Gn is an induced subgraph of Hm (n) and δ Gn , Hm (n) → 0. The condition
in the Theorem implies the claim.
To prove that P is testable, we use Lemma 15.14. So let us consider a graphon
U ∈ P and a sequence of graphons Wn → U where every Wn is a flexing of U . We
want to prove that d1 (Wn , P) → 0. For each n, we choose a simple graph Gn such
that
(15.7) v(Gn ) ≥ n,
(15.8) tind (Gn , Wn ) > 0,
1
(15.9) δ (Gn , U ) ≤ δ (Wn , U ) + ,
n
1
(15.10) d1 (Gn , P) ≥ d1 (Wn , P) − .
n
This is not difficult: Gn = Gnk = G(k, Wn ) will satisfy these conditions with high
probability if k is sufficiently large. Indeed, (15.7) and (15.8) are essentially trivial,
and (15.9) follows by Lemma 10.16. To verify (15.10), we select a graph Hnk ∈ P
with V (Hnk ) = [k] such that
d1 (Gnk , P) = d1 (Gnk , Hnk ) ≥ δ1 (WGnk , WHnk ).
Let k → ∞, then WGnk → Wn in the δ -distance with probability 1, and (by
selecting an appropriate subsequence) WHnk → Un ∈ P. By Lemma 14.16, we get
that
lim inf δ1 (WGnk , WHnk ) ≥ δ1 (Wn , Un ) ≥ δ1 (Wn , P),
n→∞

which proves that (15.10) is satisfied if k is large enough.

276 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

Claim 15.26 implies that d1 (Gn , P) → 0. Indeed, condition (15.8) implies that
tind (Gn , U ) > 0 (here we use that Wn is a flexing of U ). Furthermore, Gn → U by
(15.9), so Claim 15.26 applies.
From here, the testability of P follows easily:
1
d1 (Wn , P) ≤ δ1 (Wn , P) ≤ δ1 (Gn , P) + → 0.
n
Our second task is to prove that P is robust: if (Gn ) is a sequence of simple
graphs such that d1 (WGn , P) → 0, then d1 (Gn , P) → 0. Let Wn ∈ P be such that
∥WGn − Wn ∥1 → 0. By selecting an appropriate subsequence, we may assume that
δ (Wn , U ) → 0, for some graphon U . Clearly U ∈ P and Gn → U .
Consider the random graph G′n = G′ (v(Gn )), U ). We have tind (G′n , U ) > 0
with probability 1, and by Lemma 10.18, with probability tending(to 1, G′n →)U .
Furthermore, an easy computation (cf. Exercise 10.14) gives that E d1 (Gn , G′n ) =
E(∥WGn − WG′n ∥1 ) = ∥WGn − Wn ∥1 → 0, and so with high probability, we have
d1 (Gn , G′n ) → 0. By Claim 15.26, this implies that d1 (G′n , P) → 0. Hence
d1 (Gn , P) ≤ d1 (Gn , G′n ) + d1 (G′n , P) → 0, which proves that P is robust. 
Other characterizations of testable graph properties are known. Alon, Fischer,
Newman and Shapira [2006] characterized testable graph properties in terms of
Szemerédi partitions (we refer to their paper for the formulation). Fischer and
Newman [2005] connected testability to estimability. We already stated a version
of this result for graphons (Theorem 15.16), from which it can be derived (we don’t
go into the details):
Theorem 15.27. A graph property is testable if and only if the normalized edit
distance from the property is an estimable parameter. 

Exercise 15.28. Prove that the graph property ω(G) ≥ v(G)/2 satisfies the
condition given in Theorem 15.25.
Exercise 15.29. Prove the following analogue of Proposition 15.11 for finite
graphs: If P is a testable graph property, then
{ 20 }
P ′ = F : v(F ) = 1 or δ (F, P) ≤ √
log v(F )
is a valid test property for P.
Exercise 15.30. Prove that graph property P is robust if and only if for every
ε > 0 there is an ε0 > 0 such that if G is a graph with v(G) ≥ 1/ε0 and G has
infinitely many near-blowups G′ with d1 (G′ , P) ≤ ε0 , then d1 (G, P) ≤ ε.

15.4. Computable structures

15.4.1. Similarity distance and representative sets. Let us recall from
the Introduction that if we want to run algorithms on a very large graph G, it is
very useful to define and also compute a “representative set” of nodes, a (fairly
large, but bounded size) subset R ⊆ V (G) such that every node is “similar” to one
of the nodes in R. We want to define a distance that reflects for two nodes how
“similar” their positions in the graph are.
We define the similarity distance of two nodes s, t ∈ V (G) as follows: for any
node w, compute
1 
a(s, t; w) = |N (w) ∩ N (s)| − |N (w) ∩ N (t)|,
n
 15.4. COMPUTABLE STRUCTURES 277

and let
1 ∑
(15.11) dsim (s, t) = a(s, t; w).
n
w∈V (G)

We can think of a(s, t; w) as a measure of how different s and t are from the point of
view of w; then dsim (s, t) is an average measure of this difference. Of course, w could
be more myopic and not look for neighbors of s and t among its own neighbors,
but look only for s and t; then dsim (s, t) would measure the size of the symmetric
difference of the neighborhoods of s and t. As explained in the Introduction, this
is a perfectly reasonable definition, but it would not measure what we want. The
node w could also look for second or third neighbors of s and t, but this would not
give anything more useful than this definition, at least for dense graphs.
There are many ways to rephrase this definition. We can pick three random
nodes w, v, u ∈ V (G), and define
 
(15.12) dsim (s, t) = Ew Ev (asv avw ) − Eu (atu auw ),
where (aij ) is the adjacency matrix of G. We could use v = u here, I just used
different variables to make the correspondence with the definition clearer. We can
also notice that dsim (s, t) is the L1 distance of rows s and t of the square of the
adjacency matrix, normalized by n2 . Finally, the similarity distance is quite closely
related to the distance rWG , discussed in Section 13.4: dsim (s, t) = rWG (x, y), where
x and y are arbitrary points of the intervals representing s and t in WG .
There is an easy algorithm to compute (approximately) the similarity distance
of two nodes.
Algorithm 15.31.
Input: A graph G given by a sampling oracle, two nodes s, t ∈ V , and an error
bound ε > 0.
Output: A number D(s, t) ≥ 0 such that with probability at least 1 − ε,
D(s, t) − ε ≤ dsim (s, t) ≤ D(s, t) + ε.
The algorithm is based on (15.12). Select a random node w and fix it tem-
porarily. Select O(1/ε2 ) random nodes v and compute the average of asv avw , to
get a number that is within an additive error of (ε/4 to Ev (asv avw ) ). Estimate
Ev (atv avw ) similarly. This gives an estimate for |Ev auw (asv − atv ) | with error at
most ε/2 with high probability. Repeat this O(1/ε2 ) times and take the average to
get D(s, t).
Next, we specialize Theorem 13.31 to graphs.
Theorem 15.32. Let G = (V, E) be a graph.
(a) If P = {S1 , . . . , Sk } is a partition of V (G) such that d (G, GP ) = ε, then
we can select a node vi ∈ Si from each partition class such that the average dsim -
distance from S = {v1 , . . . , vk } is at most 4ε.
(b) If S ⊆ V is a subset such that the average dsim -distance from S =
{v1 , . . . , vk } is √ε, then the Voronoi cells of S form a partition P such that
d (G, GP ) ≤ 8 ε. 
We define a representative set with error ε > 0 as a subset R ⊆ V (G) such that
any two elements of R are at a (similarity) distance at least ε/2, and the average
distance of nodes from R is at most 2ε. (The first condition is not crucial for the
278 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

applications we want to give, but it guarantees that the set is chosen economically.)
2
Theorem 13.31 implies that such a set R exists with |R| ≤ 232/ε . Furthermore,
such a set can be constructed in our model.

Algorithm 15.33.
Input: A graph G given by a sampling oracle, and an error bound ε.
2
Output: A random set R ⊆ V (G) such that |R| ≤ (64/ε2 )1028/ε , and with
probability at least 1 − ε, R is a representative set with error ε.

The set R is grown step by step, starting with the empty set. At each step,
a new uniform random node w of G is generated, and the approximate distances
D(w, v) are computed for all v ∈ R with error less than ε/4 with high probability.
If all of these are larger than 3ε/4, then w is added to R. Else, w is discarded and
a new random node is generated. If R is not increased in k = ⌈2000 1ε log 1ε ⌉ steps,
the algorithm halts.
We have to make sure that we don’t make the mistake of stopping too early.
It is clear that as long as the average distance from R is larger than 2ε, then the
probability that a sample has distance at least ε is at least ε, and so the probability
that in k iterations we don’t pick a node whose distance from R is less than ε is less
than e−kε . If we find a good node u, then with high probability the approximate
distance satisfies D(u, R) > 3ε/4, and so we add u to R. Hence the probability
that we stop prematurely is less than e−kε E(|R|) ≤ ε.
The size of R can be bounded using Proposition 13.32, which gives the bound
on the output size. We can say more. Suppose √ that there exists a representative
set R with error ε. Then only a fraction √ of 2 ε nodes of G (call these nodes
“remote”) are at a distance
√ more than ε from R. Let us run the above
√ algorithm
with ε replaced by 2 ε, to get a representative set R′ with error q ε. The set R′
will contain at most one non-remote node from every Voronoi cell of R. We have
little control√over how many remote nodes we selected, but we can post-process the
′
result. The ε-balls around the non-remote √ nodes in R cover all the non-remote
nodes, so only leave out a fraction of 2 ε of all nodes. By sampling and brute
force, we can select the smallest subset R′′ ⊆ R′ with this property. This way √ we
have constructed a representative set R′′ with |R′′ | ≤ |R| and error at most 3 ε.
As a special case, if there is a representative set whose size is polynomially
bounded in the error ε, our algorithm will find one with a somewhat worse polyno-
mial bound.

Remark 15.34. One could try to work with a stronger notion: define a strong
representative set with distance ε > 0 as a subset R ⊆ V (G) such that any two
elements of R are at a (similarity) distance at least ε, and any other node of G is
at a distance at most ε from R.
It is trivial that every graph contains a strong representative set: just take a
maximal set of nodes any two of which are at least ε apart. Furthermore, Propo-
sition 13.32 shows that the size of such a set can be bounded by a function of
ε. There are, however, several problems with the idea of computing and using it.
First, in our very large graph model, the similarity distance cannot be computed
exactly; second (and more importantly) the graph can have a tiny remote part
which no sampling will discover but a representative of which should be included
in the strong representative set.
 15.4. COMPUTABLE STRUCTURES 279

15.4.2. Computing regularity partitions. As an easy application of The-

orem 15.32, we give an algorithm to compute a weak Szemerédi partition in a huge
graph. Our goal is to illustrate how an algorithm works in the pure sampling model,
as well as in what form the result can be returned. (This way of presenting the out-
put of an algorithm for a large graph was proposed by Frieze and Kannan [1999]).
A polynomial time algorithm to compute a regularity partition in the traditional
setting of graph algorithms was given by Alon, Duke, Lefmann, Rödl and Yuster
[1994].
Algorithm 15.33 enables us to encode a partition of V (G) as a subset R ⊆ V (G):
for each r ∈ R, we can define the partition class Vr as the Voronoi cell of r. Ties
will be broken arbitrarily, and nodes to which there are several “almost closest”
nodes may be misclassified, but this is the best one can hope for. To formalize,
Algorithm 15.35.
Input: A graph G given by a sampling oracle, a subset R ⊆ V (G), a node
u ∈ V , and an error bound ε > 0.
Output: A (random) node v ∈ R almost closest to v in the sense that with
probability at least 1 − ε, dsim (u, v) ≤ (1 + ε)dsim (u, R).
This algorithm uses Algorithm 15.31 to compute (approximately) the distances
dsim (u, r), r ∈ R, and returns the node r ∈ R that it finds closest to u. In other
words, we compute the Voronoi cells of the set R.
Theorem 15.32 says in this context that the partition determined by Algorithms
15.33 and 15.35 satisfies d (G, GP ) ≤ ε with high probability. We omit the details
of the error analysis. So we get a weak regularity partition. It is not known whether
stronger regularity partitions can be computed in the sampling model.
15.4.3. Computing a maximum cut. The algorithm to approximately
compute the maximum cut is similar.
Algorithm 15.36.
Input: A graph G given by a sampling oracle, a subset R ⊆ V (G), and an
error bound ε > 0.
Output: A partition R = R1 ∪ R2 .
For the partition P implicitly determined above, we can also compute the
edge densities between the partition classes, which we use to weight the edges
of the complete graph on R, so that we get a weighted graph H. We find the
maximum cut in H by brute force, to get a partition R = R1 ∪ R2 . This gives an
implicit definition of a cut in G, where a node u is put on the left side of the cut
iff D(u, R1 ) < Db(u, R2 ) for the approximate distances computed by Algorithm
15.31.
Remark 15.37. 1. A notion closely related to testability and to computing a
structure, called “local reparability”, was introduced by Austin and Tao [2010].
We only give an informal definition. We start with a graph property P and a graph
for which almost all induced subgraphs of size N have property P; we say that G
has the property N -locally. Now we want to repair G to have property P itself,
by changing a small fraction of the edges. So far, this is essentially the same as
testability, but we have to do the repair by a local algorithm as follows. We select
a random sample A ⊆ V (G) of size k (where k ≤ N is chosen appropriately), and
280 15. ALGORITHMS FOR LARGE GRAPHS AND GRAPHONS

for every pair of nodes u, v ∈ V (G), we compute whether they should be adjacent,
knowing only the induced subgraph G[A ∪ {u, v}]. We could decide simply the
adjacency of them in G, but then we would not do any repair. Taking the subgraph
G[A] and its connections to u and v also into account, our algorithm will define in
a modified graph G′ . This graph G′ should have property P, and its edit distance
from G should be arbitrarily small if N and k are large enough.
It may or may not be possible to do so. We say that P is locally reparable if
it is always possible. Austin and Tao prove, among others, that every hereditary
property is reparable. For the exact definitions, formulation, proofs, generalizations
to hypergraphs and other results we refer to the paper.
2. There is a natural nondeterministic version of testability, introduced by
Lovász and Vesztergombi [2012] . A property of finite graphs is called nondeter-
ministically testable if it has a “certificate” in the form of a coloring of the nodes
and edges with a bounded number of colors, adding new edges with other colors,
and orienting the edges, such that once the certificate is specified, its correctness
can be verified by random local testing. Here are a few examples of properties that
are nondeterministically testable in a natural way: “the graph is 3-colorable;” “the
graph contains a clique on half of its nodes;” “the graph is transitively orientable”;
“one can add at most v(G)2 /100 new edges to make the graph perfect.”
Using the theory of graph limits, it is proved that every nondeterministically
testable property is deterministically testable. In a way, this means that P = N P in
the world of property testing for dense graphs. (Many, but not all, of the properties
described above are also covered by Theorem 15.24.)
We will see that for bounded-degree graphs, the analogous statement does not
hold. In fact, the study of nondeterministic certificates will lead to a new interesting
notion of convergence (Section 19.2).
 CHAPTER 16

Extremal theory of dense graphs

Extremal graph theory was one of the motivating fields for graph limit theory, as
described in the Introduction. It is also one of the most fertile fields of applications
of graph limits. In this chapter we give an exposition of some of the main directions.
We start with two sections developing some technical tools, reflection positivity
and variational calculus. Then we discuss extremal problems for complete graphs
and some other specific problems. We re-prove some classical general results in
extremal graph theory, and finally, we treat some very general questions (formulated
in the introduction) about decidability of extremal graph problems and the possible
structure of extremal graphs.

16.1. Nonnegativity of quantum graphs and reflection positivity

We are interested in linear inequalities between the densities of some subgraphs.
(Why just linear? We have seen in the Introduction that, using the multiplicativ-
ity of subgraph densities, algebraic inequalities between them can be replaced by
equivalent linear inequalities. To be sure, there are nontrivial non-algebraic inequal-
ities that hold between subgraph densities; see Exercise 16.21. But their theory is
virtually completely unexplored.)
As we have seen in the Introduction, many results in extremal graph theory
can be stated in this form. We have also seen an example of a proof that relied on
simple computations with quantum graphs. To make the problem and the methods
precise, for a set U ⊆ W and for a quantum graph x we write x ≥ 0 (for U) if
t(x, W ) ≥ 0 for every graphon W ∈ U. Most of the time (but not always) we will
be concerned with U = W0 , and then we will suppress the “(for U)” part of the
notation. The condition x ≥ 0 is equivalent to requiring that t(x, G) ≥ 0 for every
simple graph G.
Reflection positivity of the subgraph densities (Proposition 7.1) implies many
inequalities of this type. (In a sense it implies all, as we will see in Theorem
16.41.) For every k ≥ 0, every graphon W , every set of simple k-labeled graphs
{F1 , . . . , Fm }, and every vector a ∈ Rm , we have

∑
m
ai aj t([[Fi Fj ]], W ) ≥ 0,
i,j=1

or, using our notation,

∑
m
ai aj [[Fi Fj ]] ≥ 0.
i,j=1

281
282 16. EXTREMAL THEORY OF DENSE GRAPHS

This is equivalent to
(∑
m )2
(16.1) [[ ai Fi ]] ≥ 0.
i,j=1

So we get the fact, used implicitly in the Introduction (Section 2.1.3), that unlabel-
ing the square of a k-labeled quantum graph, we get a nonnegative quantum graph.
Let us also recall the trivial fact that adding or deleting isolated nodes to a graph
F does not change the homomorphism densities t(F, .). We call a quantum graph
g a square-sum if there are k-labeled
∑ quantum graphs y1 , . . . , ym for some k such
that g can be obtained from i yi2 by unlabeling and adding or deleting isolated
nodes. We have just shown that every square-sum satisfies g ≥ 0.
Another important property of semidefinite matrices is that their determinant
is nonnegative: for any set {F1 , . . . , Fm } of graphs and any graphon W , we have
 
 t([[F1 F1 ]], W ) . . . t([[F1 Fm ]], W ) 
 
 .. .. 
 . .  ≥ 0.
 
t([[Fm F1 ]], W ) . . . t([[Fm Fm ]], W )

Expanding this determinant, we get a polynomial inequality, which can be turned

into a linear inequality involving subgraph densities, using multiplicativity of the
parameter t(., W ). This can be expressed as nonnegativity of a quantum graph:
 
 [[F1 F1 ]] . . . [[F1 Fm ]] 
 
 .. .. 
(16.2)  . .  ≥ 0.
 
[[Fm F1 ]] . . . [[Fm Fm ]]

This is still a rather complicated inequality, but one special case will be useful:
(16.3) [[F1 F1 ]][[F2 F2 ]] ≥ [[F1 F2 ]]2 .
Another consequence of reflection positivity is the following: let (aij )m
i,j=1 be a
symmetric positive semidefinite matrix, then
∑
m
(16.4) aij [[Fi Fj ]] ≥ 0.
i,j=1

We can add two relations related to subgraphs. First, adding an isolated node
to a graph F does not change its density in any graph or graphon, and so
F K1 − F ≥ 0, but also F − F K1 ≥ 0.
Furthermore, if F ′ is a subgraph of F , then t(F ′ , W ) ≥ t(F, W ) for every graphon
W , and hence
(16.5) F ′ − F ≥ 0.

Exercise 16.1. Show that inequalities (16.2) (16.4) and (16.5) can be derived
from the inequalities (16.1).
Exercise 16.2. Prove the following “supermodularity” inequality: if F1 and F2
are two simple graphs on the same node set, then F1 ∪ F2 + F1 ∩ F2 ≥ F1 + F2 .
 16.2. VARIATIONAL CALCULUS OF GRAPHONS 283

16.2. Variational calculus of graphons

One advantage of working in the large space of graphons instead of the much
simpler (discrete) space of graphs is that we can do continuous deformations of a
graphon. This is useful when trying to optimize some functional, or when studying
conditions that uniquely determine graphons. The following considerations are
conceptually not too difficult, but technically more involved.
To express the variation of graphon functionals, we need some notation. For
every multigraph F = (V, E) and node i ∈ V , let F i denote the 1-labeled quantum
graph obtained by labeling i by 1. For every edge ij ∈ E, let F ij denote the 2-
labeled quantum graph obtained from F by deleting the edge ij, and labeling i by 1
and j by 2. The 2-labeled quantum graph Fij is∑ constructed similarly,∑ but the edge
ij is not deleted. So Fij = K2•• F ij . Let F † = i∈V F i and F ‡ = 21 i,j: ij∈E F ij
(each edge contributes two terms, since its endpoints can be labeled∑ in two ways;
this is why it will be convenient to divide by 2). Similarly, let F ♮ = 12 i,j: ij∈E Fij .
We extend the operators F 7→ F † , F 7→ F ‡ and F 7→ F ♮ linearly to all quantum
graphs. We have x♮ = K2•• x‡ .

Example 16.3. Clearly Cn‡ = nPn•• , where Pn•• denotes the path on n nodes with
its endpoints labeled. So txy (Cn‡ , W ) = ntxy (Pn•• , W ) = nW ◦(n−1) . 

The Edge Reconstruction Conjecture 5.31 says in this language that if F and
G are simple graphs that are large enough (both have at least four non-isolated
nodes), then [[F ‡ ]] = [[G‡ ]] implies that F ∼
= G.
We study two kinds of variations of a kernel W : in the more general version,
we change the values of W at every node. However, it is often easier to construct
variations in which the measure on [0, 1] is rescaled. This simpler kind variation will
have the advantage that if we start with a graphon, then we don’t have to worry
about the values of W running out of the interval [0, 1]. We start with describing
the variation of the measure.
Consider a family αs : [0, 1] → R+ (s ∈ [0, 1]) of weight functions such that
∫1
0
α s (x) dx = 1 for every s. Every such function defines a probability measure µs
on [0, 1] by
∫
µs (A) = αs (x) dx.
A

Every kernel W on [0, 1]2 gives rise to a family of kernels Ws , where Ws =

([0, 1], µs , W ). For every finite graph F , we have
∫ ∏ ∏
t(F, Ws ) = W (xi , xj ) dµs (xi )
[0,1]V (F ) ij∈E(F ) i∈V (F )
∫ ∏ ∏ ∏
= W (xi , xj ) αs (xi ) dxi .
[0,1]V (F ) ij∈E(F ) i∈V (F ) i∈V (F )

We say that the family (αs ) has uniformly bounded derivative, if for every x ∈ [0, 1]
d
the derivative α̇s (x) = ds αs (x) exists, and there is a constant M > 0 such that
|α̇s (x)| ≤ M for all x and s. If αs is a family of weight functions with uni-
formly bounded derivative, then by elementary analysis it follows that the function
284 16. EXTREMAL THEORY OF DENSE GRAPHS

t(F, Ws ) is differentiable as a function of s, and

d ⟨ ⟩
(16.6) t(F, Ws ) = α̇s , tx (F † , Ws ) .
ds
In the second type of variation, we consider a family Us ∈ W (0 ≤ s ≤ 1) of
kernels. We say that the family Us has uniformly bounded derivative, if for every
x, y ∈ [0, 1] the derivative U̇s (x, y) = dsd
Us (x, y) exists, and there is a constant
M > 0 such that |U̇s (x)| ≤ M for all x and s. If Us is a uniformly bounded
family of kernels with uniformly bounded derivative, then the function t(F, Ws ) is
differentiable as a function of s, and
d ⟨ ⟩
(16.7) t(F, Us ) = U̇s , txy (F ‡ , Us ) .
ds
As an application of these formulas, we derive versions of the Kuhn–Tucker
conditions in optimization for a graphon minimizing a smooth function of a finite
number of homomorphism densities (rephrasing conditions of Razborov [2007, 2008]
in our language). For a functional ω on W0 , we say that a graphon W is a local
minimizer, if there is an ε > 0 such that ω(U ) ≥ ω(W ) for every graphon U with
∥U − W ∥1 < ε. (We could define this notion with respect to other norms, but this
will be the version we use.)
Lemma 16.4. Let g be a simple quantum graph, and suppose that W is a local
minimizer of t(g, W ) over W ∈ W0 .
(a) For almost all x ∈ [0, 1], tx (g † , W ) = t([[g † ]], W ).
(b) For almost all x, y ∈ [0, 1],


= 0 if 0 < W (x, y) < 1,
‡
txy (g , W ) ≥ 0, if W (x, y) = 0,


≤ 0, if W (x, y) = 1.
(c) For almost all x, y ∈ [0, 1], txy (g ♮ , W ) ≤ 0.
We will prove a more general lemma.
Lemma 16.5. Let Φ : Rm → R be a differentiable function, and let Φi =
∂
∂xi Φ.
(
Let F1 , . . . , Fm , simple graphs, and suppose that W is a local minimizer
) ( )
of Φ t(F1 , W ), . . . , t(Fm , W ) over W ∈ W0 . Set ai = Φi t(F1 , W ), . . . , t(Fm , W ) .
(a) For almost all x ∈ [0, 1],
∑
m
( )
ai tx (Fi† , W ) − v(Fi )t(Fi , W ) = 0.
i=1

(b) For almost all x, y ∈ [0, 1],


∑
m 
= 0 if 0 < W (x, y) < 1,
ai txy (Fi‡ , W ) ≥ 0, if W (x, y) = 0,


i=1 ≤ 0, if W (x, y) = 1.
(c) For almost all x, y ∈ [0, 1],
∑
m
ai txy (Fi♮ , W ) ≤ 0.
i=1
 16.3. DENSITIES OF COMPLETE GRAPHS 285

∫
Proof. (a) Let φ : [0, 1] → [−1, 1] be a measurable function such that φ = 0.
For s ∈ [−1, 1], we re-weight the points of [0, 1] by αs (x) = 1 + sφ(x) to get the
graphon Ws . Using (16.6), we get

∫1
d ( ) ∑m
Φ t(F1 , Ws ), . . . , t(Fm , Ws )  = ai φ(x)tx (Fi† , W ) dx
ds s=0
i=1 0
∫1 ∑
m
= φ(x) ai tx (Fi† , W ) dx.
0 i=1

Since W is a local minimizer, and ∥Ws −W ∥1 is arbitrarily small if s is small (see

∑
Exercise 13.7), this derivative must be 0 for all φ. This implies that i ai tx (Fi† , W )
is a constant function of x almost everywhere. Integrating over x, we recover the
value of the constant, which proves (a).

(b) Let U ∈ W1 be a function such that U (x, y) ≥ 0 if W (x, y) = 0 and

U (x, y) ≤ 0 if W (x, y) = 1. Then for every s ≥ 0 in a small neighborhood of 0,
W + sU ∈ W0 . Using (16.7),

d ( )
Φ t(F1 , W + sU ), . . . , t(Fm , W + sU ) 
ds s=0
∫ ∑
m
= U (x, y) a′i txy (Fi‡ , W ) dx dy.
i=1
[0,1]2

Since W is a local minimizer, we must have

∫ ∑
m
U (x, y) a′i txy (Fi‡ , W ) dx dy ≥ 0.
i=1
[0,1]2

This must hold for all functions U ∈ W1 such that U (x, y) ≥ 0 if W (x, y) = 0 and
U (x, y) ≤ 0 if W (x, y) = 1, which implies (b).

(c) This follows from (b) by multiplying by W (x, y). 

Exercise 16.6. For W ∈ W0 , define δ(W ) = minx∈[0,1] tx (K2• , W ) and ∆(W ) =

maxx∈[0,1] tx (K2• , W ) (minimum degree and maximum degree). Prove that for any
tree T , δ(W )v(T )−1 ≤ t(T, W ) ≤ ∆(W )v(T )−1 .
Exercise 16.7. For a simple graph F , prove that
sup{t(K2 , W ) : W ∈ W0 , t(F, W ) = 0} = sup{δ(W ) : W ∈ W0 , t(F, W ) = 0}.

16.3. Densities of complete graphs

The problem of describing relationships between densities of complete graphs
in a graph G has received special attention. In this section we survey results about
these questions in the framework of our book.
286 16. EXTREMAL THEORY OF DENSE GRAPHS

16.3.1. Linear inequalities. We start with a result that characterizes linear

inequalities between complete subgraph densities. This was proved in a special
case by Bollobás [1976], but, as observed by Schelp and Thomason [1998], his proof
extends to the proof of the general theorem. Schelp and Thomason give further
applications of the method (see Exercise 16.16).
Theorem 16.8. Let g be a quantum graph whose constituents are complete graphs.
Then g ≥ 0 if and only if t(g, Kn ) ≥ 0 for every n ≥ 1.
Note that the case n = 1 is included, which means that t(g, 0) ≥ 0; in other
words, the coefficient of K0 in g is nonnegative. Since Kn tends to the all-one
function as n → ∞, the condition implies that t(g, 1) ≥ 0. Stating the result more
directly: an inequality of the form
∑
m
(16.8) ai t(Ki , G) ≥ 0
i=1

holds for every graph G if and only if

∑
m
(n)i
(16.9) ai ≥0
i=1
ni

holds for every integer n ≥ 1.

Proof. The “only if” direction is trivial. To prove the “if” direction, suppose
that t(g, Kn ) ≥ 0 for all n; we want to prove that t(g, W ) ≥ 0 for every W ≥ 0.
It suffices to prove this for any dense set of graphons W , and we choose the set
graphons WH , where H is a node-weighted simple graph (all edgeweights are 0 or 1).
Let V (H) = [q], and let α1 , . . . , αq ≥ 0 be the
∑ nodeweights (we allow 0 nodeweights
for this argument). We may assume that i αi = 1. Supposing that there is an
H with t(g, H) < 0, choose one with minimum number of nodes, and choose the
nodeweights so as to minimize t(g, H). Then all the nodeweights must be positive,
since a node with weight 0 could be deleted without changing any subgraph density,
contradicting the minimality of q. Clearly t(g, H) is a polynomial in the nodeweights
αi . Furthermore, the assumptions that the constituents of g are complete and H
has no loops imply that every homomorphism contributing to t(g, H) is injective,
and so t(g, H) is multilinear.
Next we prove that H must be complete. Indeed, if (say) nodes 1, 2 ∈ V (H) are
nonadjacent, then t(g, H) has no term containing the product α1 α2 , i.e., fixing the
remaining variables, t(g, H) is a linear function of α1 and α2 . Since only the sum
of α1 and α2 is fixed, we can shift them keeping the sum fixed and not increasing
the value of t(g, H) until one of them becomes 0. This is a contradiction, since we
know that all weights must be positive.
To show that all weights are equal, let us push the argument above a bit further.
Fixing all variables but α1 and α2 , we can write t(g, H) = a + b1 α1 + b2 α2 + cα1 α2 .
Since H is complete, we know that t(g, H) is a symmetric multilinear polynomial
in α1 , . . . , αq , and so b1 = b2 . Since α1 + α2 is fixed, we get t(g, H) = a′ + cα1 α2 ,
where a′ does not depend on α1 or α2 . If c ≥ 0, then this is minimized when α1 = 0
or α2 = 0, which is a contradiction as above. Hence we must have c < 0, and in
this case t(g, H) is minimized when α1 = α2 . Since this holds for any two variables,
all the αi are equal.
 16.3. DENSITIES OF COMPLETE GRAPHS 287

But this means that t(g, H) = t(g, Kq ), which is impossible since t(g, Kq ) ≥ 0
by hypothesis. This completes the proof. 
As a corollary (which is in fact equivalent to the theorem) we get the following.
( an integer m ≥ 1, )and associate with every graphon W the vector tW =
Fix
t(K2 , W ), . . . , t(Km , W ) . Let Tm denote the set of the vectors tW . It follows
from Theorem 11.21 and Corollary 11.15 that Tm is the closure of the points tG ,
where G is a simple graph (we write tG for tWG ).
Corollary 16.9. The extreme points of the convex hull of Tm are the vectors tKn
(n = 1, 2, . . . ) and (1, . . . , 1).
The following corollary is interesting to state in view of the undecidability result
of Hatami and Norine [2011] already mentioned in the Introduction (which will be
proved as Theorem 16.34 a little later). It is easy to design an algorithm to check
whether (16.9) holds for every n, and hence:
Corollary 16.10. For quantum graphs g with rational coefficients whose con-
stituents are complete graphs, the property g ≥ 0 is algorithmically decidable.
As a further corollary, we derive Turán’s Theorem for graphons.
Corollary 16.11. For every r ≥ 2, we have
f0 , t(Kr , W ) = 0} = 1 − 1
max{t(K2 , W ) : W ∈ W ,
r−1
and the unique optimizer is W = WKr−1 .
One could prove this result along the lines of several well-known proofs of
Turán’s Theorem; we could also prove a generalization of Goodman’s inequality
2.2. Specializing the proof of Theorem 16.8 above just to this case, we get the
proof by “symmetrization”, due to Zykov [1949].
Proof. Let us prove the inequality
(16.10) rr t(Kr , W ) − (r − 1)t(K2 , W ) + r − 2 ≥ 0.
By Theorem 16.8, it suffices to verify this inequality when W = WKn for some
n ≥ 1. This is straightforward, and we also see that equality holds for n = r − 1
only. Corollary 16.9 implies that equality holds in (16.10) only if W = WKr−1 . In
the special case with t(Kr , W ) = 0, we get Corollary 16.11. 
16.3.2. Edges vs. triangles. In the introduction (Section 2.1.1) we men-
tioned several results about the number of triangles in a graph, if the number of
edges is known: Goodman’s bound and its improvements, and the Kruskal–Katona
Theorem. In this Section we describe the exact relationship between the edge den-
sity and triangle density in a graph, i.e., we describe the set D2,3 ; for convenience,
we recall how it looks (Figure 16.1; also recall that the figure is distorted to be able
to see some features better).
As a special case of Corollary 16.9, we get the result of Bollobás [1976] men-
tioned in the introduction:
Corollary 16.12. The set D2,3 is contained in the convex hull of the points (1, 1)
and
( ) ( n − 1 (n − 1)(n − 2) )
tn = t(K2 , Kn ), t(K3 , Kn ) = , (n = 1, 2, . . . ).
n n2
288 16. EXTREMAL THEORY OF DENSE GRAPHS

Figure 16.1.

However, a quick look at Figure 2.1 shows that Corollary 16.12 does not tell the
whole story: between any two special points (including the endpoints of the upper
boundary), the domain D2,3 is bounded by a curve that appears to be concave. It
turns out that these curves are indeed concave, which can be proved by the same
kind of argument as used in the proof of Theorem 16.8 (see Exercise 16.18). The
formula for these curves (cubic equations) is more difficult to obtain, and this will
be our main concern in the rest of this section.
The Kruskal–Katona bound. Let us start with the curve bounding the
domain T from above, which is not hard to determine: its equation is y = x2/3 .
As we mentioned in the introduction, this follows from (a very special case of) the
Kruskal–Katona Theorem in extremal hypergraph theory. Here we give a short
direct proof using the formalism of graph algebras. Applying (16.3) with F1 = P3••
and F2 = P2•• , we get
( )( ) 2 2 2
=[[ ]] ≤ [[ ]][[ ]] = ≤
(the last step uses the trivial monotonicity (16.5)). This shows that t(K3 , W ) ≤
t(K2 , W )3/2 for every graphon W , what we wanted to prove.
We also want to prove that this upper bound on the triangle density is sharp.
For n ≥ 1, let G consist of a complete graph on k nodes and n − k isolated nodes.
Then
( t(K2 , G) =) (k)2 /n2 and t(K3 , G) = (k)3 /n3 . Clearly, points of the form
(k)2 /n2 , (k)3 /n3 get arbitrarily close to any point on the curve y = x2/3 .
Razborov’s Theorem. To determine the lower bounding curve of D2,3 is
much harder (Razborov [2008]); even the result is somewhat lengthy to state. Per-
haps the best way to remember it is to describe a family of extremal graphons
(which are all node-weighted complete graphs).
Theorem 16.13. For all 0 ≤ d ≤ 1, the minimum of t(K3 , W ) subject to W ∈ W0
and t(K2 , W ) = d, is attained by the stepfunction W = WH , where H is a weighted
complete graph on k = ⌈ 1−d1
⌉ nodes with edgeweights 1 and appropriate nodeweights:
k − 1 of the nodeweights are equal and the last one is at most as large as these.
One indication of the difficulty of the proof is that the extremal graphon is not
unique, except for the special values d = 1 − 1/k. Let us consider the interval I
representing the smallest weighted node and the interval J representing any other
node. Restricted to I ∪ J, the graphon is bipartite and hence triangle-free. If we
 16.3. DENSITIES OF COMPLETE GRAPHS 289

replace the function WH on (I ∪J)×(I ∪J) by any other triangle-free function with
the same integral, then neither the edge density nor the triangle density changes,
but we get a different extremal graphon.
The nodeweights can be determined by simple computation. With a con-
venient parametrization suggested by Nikiforov [2011], they can be written as
(1 + u)/k, . . . , (1 + u)/k, (1 − (k − 1)u)/k. The edge density in the extremal graph
is
k−1 k−1
(16.11) t(K2 , H) = d = (1 − u2 ) = (1 + u)(1 − u),
k k
and the triangle density is
(k − 1)(k − 2) (k − 1)(k − 2)
(16.12) t(K3 , H) = 2
(1−3u2 −2u3 ) = (1+u)2 (1−2u).
k k2
This gives a parametric
[ k−2 equation
] for the cubic curve bordering the domain in Figure
2.1 in the interval k−1 , k−1
k . We can solve (16.11) for u as a function of d, and then
substitute this in (16.12) to get an explicit expression for t(K3 , H) as a function of
d (this hairy formula is not the way to understand or remember the result; but we
need it in the proof):
(k − 1)(k − 2)
(16.13) t(K3 , H) = f (d) = (1 + u)2 (1 − 2u)
k2 √ √
(k − 1)(k − 2) ( kd )2 ( kd )
= 1 + 1 − 1 − 2 1 − .
k2 k−1 k−1
(where k = ⌈ 1−d
1
⌉). This function is rather complicated, and I would not even
bother to write it out, except that we need its explicit form in the proof below.
Perhaps the following form says more:
( 3kd k 2 f (d) )2 ( kd )3
(16.14) 1− + = 1− .
2(k − 1) 2(k − 1)(k − 2) k−1
This shows that after an appropriate affine transformation, every concave piece
of the boundary of the region D2,3 looks alike (including the curve bounding the
region from above). Perhaps this is trying to tell us something—I don’t know.
These considerations allow us to reformulate Theorem 16.13 in a more direct
form:
Theorem 16.14. If G is a graph with t(K2 , G) = d, then t(K3 , G) ≥ f (d).
The original proof of this theorem uses Razborov’s flag algebra technique, which
is basically equivalent to the methods developed in this book. Since then, the result
has been extended by Nikiforov [2011] to the number of K4 ’s and by Reiher [2012] to
all complete graphs. We describe the proof of Razborov’s Theorem in our language.
(Reiher’s proof can be viewed as a generalization of this argument to all complete
graphs; the generalization is highly nontrivial.)
( )
Proof. Let W ∈ W0 minimize t(K3 , W ) − f t(K2 , W ) subject to k−2 k−1 ≤
t(K2 , W ) ≤ k , and suppose (by way of contradiction) that the minimum value
k−1

is negative. Set d = t(K2 , W ) and

√
′ 3(k − 2) ( kd ) 3(k − 2)
(16.15) λ = f (d) = 1+ 1− = (1 + u).
k k−1 k
290 16. EXTREMAL THEORY OF DENSE GRAPHS

(The representation in terms of the parameter u will be useful if you want to follow
some of the computations below). Since the objective function is 0 at the endpoints
k−1 < d < k , and so W is a local minimizer in W0 .
of the interval, we must have k−2 k−1

Let us simplify notation by writing g1 ≡ g2 (for W ) and g1 ≤ g2 (for W ) for

two quantum graphs g1 and g2 if t(g1 , W ) = t(g2 , W ) and t(g1 , W ) ≤ t(g2 , W ),
respectively. We invoke the formulas obtained by variational calculus on graphons,
and get by Lemma 16.5(a) that
(16.16) 3 − 2λ =3 − 2λ (for W )
(if a graph pictogram has a black node, this means that the equation or inequality
holds for every choice of the image of the black node). Multiplying (16.16) by the
edge with one node labeled, and then unlabeling, we get
(16.17) 3 − 2λ =3 − 2λ (for W ).
By Lemma 16.5(c),
(16.18) 3 −λ ≤ 0 (for W ).
Multiplying with the signed 3-node path with both edges negative and with both
endpoints labeled, we get
(16.19) 3 ≤λ (for W ),
which can be written as
3 −6 +3 ≤λ − 2λ +λ (for W ),
or, simplifying and leaving just the 4-node graphs on the left,
(16.20) 3 −6 ≤ (λ − 3) +λ − 2λ (for W ).
We get a third inequality from inclusion-exclusion:
= −3 +3 − ,
whence
(16.21) 3 −3 = − − ≤ − (for W ).
Adding (16.17), (16.20) and (16.21), we get
0 ≤ (λ − 2) +λ +3 − 2λ − (for W ).
We can replace every by d, to get
(16.22) (λ + 3d − 2) ≥ λ(2d2 − d) + (for W ).
1 2
For k = 3 (i.e., < d <
2 3 ),
we can just ignore the K4 term (it is nonnegative), and
hence (upon verifying that λ + 3d − 2 > 0)
λ(2d2 − d) f ′ (d)(2d2 − d)
t(K3 , W ) ≥ = ′ = f (d)
λ + 3d − 2 f (d) + 3d − 2
(where the last equality is easy to check).
However, if k ≥ 4, then we need a nontrivial lower bound for t(K4 , W ) (note
that the extremal graph H contains many K4 -s). Let
∫ 1
•
(16.23) µ = 2λd − 3t(K3 , W ) and w(z) = tz (K2 , W ) = W (z, y) dy.
0
 16.3. DENSITIES OF COMPLETE GRAPHS 291

Then we can write (16.16) as

(16.24) 3tz (K3• , W ) = 2λw(z) − µ.
The numbers λ and µ will play an important role in our computations, so let us
have a closer look at them. Recall that λ = f ′ (d) is given by (16.15), which yields
the bounds
3(k − 2) 3(k − 2)
(16.25) ≤λ≤ .
k k−1
For µ we don’t have an explicit formula in terms of d alone, but we can get rather
tight bounds: from Goodman’s Theorem and the indirect hypothesis we get d(2d −
1) ≤ t(K3 , W ) < f (d), which in turn implies
2df ′ (d) − 3f (d) < µ ≤ 2df ′ (d) − 3d(2d − 1).
From here, it takes straightforward computation to verify that
k−1 3µ k−2
(16.26) < 2 ≤
k−2 λ k−3
We see that µ > 0. Furthermore, inequality 16.18 implies that
(16.27) 3tz (K3• , W ) ≤ λw(z),
for almost all z ∈ [0, 1], and then (16.27) and (16.24) imply that
µ µ
≤ w(z) ≤ .
2λ λ
For any point z ∈ [0, 1], the density of K4 -s containing it is just the density of
triangles in its neighborhood. To be precise, for any point z ∈ [0, 1], we define a new
graphon Wz on [0, 1], by keeping the same W but adding a nodeweight function
W (z, .)/w(z). Then
tz (K3• , W ) 2λw(z) − µ
t(K2 , Wz ) = 2
= .
w(z) 3w(z)2
The right hand side is a monotone increasing function of w(z) for w(z) ∈ [ 2λ
µ µ
, λ ],
and hence
2λw(z) − µ λ2 k−2
t(K2 , Wz ) = ≤ ≤ .
3w(z)2 3µ k−1
( )
So by induction on k, we know that t(K3 , Wz ) ≥ f t(K2 , Wz ) , and
( 2λw(z) − µ )
tz (K4• , W ) = t(K3 , Wz )w(z)3 ≥ f w(z)3 .
3w(z)2
Hence
∫1 (
2λw(z) − µ )
t(K4 , W ) ≥ f w(z)3 dz.
3w(z)2
0
The ugly integrand is hopeless to evaluate directly; the trick is to find a (lower bound
)
that is linear in w(z) and approximates it well enough. Set g(w) = f 2λw−µ 3w 2 w3 .
Let w0 ∈ [ 2λ
µ µ
, λ ] be chosen so that
2λw0 − µ k−3
(16.28) = .
3w02 k−2
292 16. EXTREMAL THEORY OF DENSE GRAPHS

This is possible, since the left side, as a function of w0 , ranges from 0 to λ2 /(3µ) on
this interval, and the target value (k − 3)/(k − 2) is in this range by (16.26). Then
(k − 3) 3 (k − 3)(k − 4) 3
g(w0 ) = f w = w0 .
k−2 0 (k − 2)2
( )
The linear function we will use goes through the point w0 , g(w0 ) and has appro-
priate slope:
Claim 16.15. Let λ and µ be real numbers satisfying (16.25) and (16.26), and let
w0 satisfy (16.28). Then for every w ∈ [ 2λ
µ µ
, λ ], we have
1
g(w) − g(w0 ) ≥(2λ2 − 3µ)(w − w0 ).
3
All functions in this claim are explicit, which makes it a (hard and tedious)
exercise in first year calculus; we do not reproduce the details of its proof. Using
this claim, we have
∫1 ∫1
( ) 1 1
t(K4 , W ) ≥ g w(z) dz ≥ g(w0 ) − (2λ − 3µ)w0 + (2λ − 3µ) w(z) dz
2 2
3 3
0 0
(k − 3)(k − 4) 3 1
= w0 + (2λ2 − 3µ)(d − w0 ).
(k − 2)2 3
Hence, returning to (16.22),
(k − 3)(k − 4) 3 1
(16.29) (λ+3d−2)t(K3 , W ) ≥ λ(2d2 −d)+ w0 + (2λ2 −3µ)(d−w0 ).
(k − 2)2 3
This is another messy formula, but we can express the variables in terms of u and
y = w0 /(1 + u) (the latter is, of course, chosen with hindsight): we already have
expressions for λ and d; we have w0 = y(1 + u), and then µ can be expressed using
(16.28). With these substitutions, the difference of the two sides looks like
( k − 2)
(16.30) (1 + u)2 y −
( k )
k−1 (k − 3 3k − 7 ) 2(k − 1)(k − 3)
× (1 − u) − u+ y+ 2
(1 + u)y .
k k−2 k−2 k(k − 2)
2
(k−2)
k ≤ y ≤ k(k−3) .
1
We know that 0 < u < k−1 , and (16.28) and (16.26) imply that k−2
Then we face another exercise in calculus, to show that in this range (16.30) is
negative (we don’t describe the details). This contradicts (16.29), and completes
the proof. 
It would of course be important to find a more “conceptual” proof of Theorem
16.13. As a couple of examples of the kind of general question that arises, is an
algebraic inequality between densities of complete graphs decidable? Does such an
inequality hold true if it holds true for all node-weighted complete graphs?
Exercise 16.16. Let g be a quantum graph such that every constituent with
negative coefficient is complete. Prove that tinj (g, W ) ≥ 0 for every graphon W if
and only if tinj (g, Kn ) ≥ 0 for all n ≥ 1 (Schelp and Thomason [1998]).
Exercise 16.17. Prove that for quantum graphs g with rational coefficients whose
constituents are complete graphs, the property g ≥ 0 is in P . (The input length is
the total number of digits in the numerators and denominators of the coefficients,
and complete k-graphs (0 ≤ k ≤ m) contribute 1 even it their coefficient is 0.)
 16.4. THE CLASSICAL THEORY OF EXTREMAL GRAPHS 293

Exercise 16.18. Adopt the proof of Theorem 16.8 to prove that the boundary
of the domain D2,3 (Fig. 2.1) is concave between any two special points tn and
tn+1 .
Exercise 16.19. (a) Let Kr′ denote the graph obtained by deleting an edge from
Kr . Prove that
′ t(Kr , G)2
t(Kr+1 , G) ≥ .
t(Kr−1 , G)
(b) Prove that
( ′ )
t(Kr+1 , G) − t(Kr , G) ≤ r t(Kr+1 , G) − t(Kr+1 , G) .
(c) Prove that
t(Kr , G) t(Kr+1 , G)
r ≤ (r − 1) + 1.
t(Kr−1 , G) t(Kr , G)
(d) Prove the following result of Moon and Moser [1962]: If Nr denotes the number
of complete r-graphs in G, then
( )
Nr+1 1 Nr
≥ 2 r2 − N1 .
Nr r −1 Nr−1
Exercise 16.20. Prove the following generalization of Goodman’s Theorem (2.2):
if d = t(K2 , W ) is the edge density of a graph G, then
( )
t(Kr , G) ≥ d(2d − 1)(3d − 2) · · · (r − 1)d − (r − 2) .
Exercise 16.21. Prove that
( )
t(K4′ , G) ≥ t(K3 , G)2 log∗ 1/t(K3 , G) .
[T. Tao; hint: use the Removal Lemma.]

16.4. The classical theory of extremal graphs

Following the exposition of this topic in the introduction, let us state now in
more general and more precise terms the extremal graph results of Erdős and Stone
[1946], Erdős and Simonovits [1966], and Simonovits [1968]. In this more general
setting, we exclude several graphs L1 , . . . , Lk as subgraphs of a simple graph G,
and we want to determine the maximum number of edges of G, given the number
of nodes n. In our formalism, we want to solve
(16.31) max{t(K2 , G) : tinj (L1 , G) = · · · = tinj (Lk , G) = 0}.
Turán’s Theorem (Corollary 16.11) is a special case when k = 1 and L1 = Kr .
The key results are summed up in the following theorem.
Theorem 16.22. Let L1 , . . . , Lk be simple graphs and let r = mini χ(Li ). Suppose
that a simple graph G does not contain any Li as a subgraph. Then
1
(16.32) t(K2 , G) ≤ 1 − + o(1) (v(G) → ∞).
r−1
Asymptotic equality holds G = T (n, r − 1) is the Turán graph on n = v(G) nodes
with r − 1 color classes. Furthermore, this extremal graph is stable in the following
sense: For every ε > 0 there is an ε′ > 0 (depending on L1 , . . . , Lk and ε, but not
on G) such that if t(K2 , G) ≥ 1 − 1/(r − 1) − ε′ , then δb1 (Gn , T (n, r − 1)) ≤ ε.
Theorem 16.22 can be translated to the language of graphons, and proved quite
easily using our general results on graph limits. As in almost all of our applications
of graph limit theory, the original treatment has the advantage that it provides
explicit bounds for the o(1) term as well as the dependence of ε′ on ε in the theorem
above.
294 16. EXTREMAL THEORY OF DENSE GRAPHS

Theorem 16.23. Let L1 , . . . , Lk be simple graphs and let r = mini χ(Li ). Then
1
max{t(K2 , W ) : W ∈ W0 , t(L1 , W ) = · · · = t(Lk , W ) = 0} = 1 − ,
r−1
and the unique optimizer (up to weak isomorphism) is W = WKr−1 .
Proof. Let, say χ(L1 ) = r. Then t(L1 , Kr ) > 0, and hence it follows easily
that t(Kr , W ) = 0 (Exercises 7.6, 13.25). Application of Turán’s Theorem for
graphons (Corollary 16.11) completes the proof. 

It is not quite trivial, but not very hard either, to derive the classical results
mentioned above from Theorem 16.23. Let us illustrate this by the derivation
of stability. If stability fails, then there exists a sequence Gn of simple graphs
such that tinj (L1 , Gn ) = · · · = tinj (Lk , Gn ) = 0 and t(K2 , Gn ) → 1 − 1/(r − 1), but
δb1 (Gn , T (n, r−1)) ̸→ 0. By Theorem 9.30, this implies that δ1 (Gn , T (n, r−1)) ̸→ 0.
In graphon language, this means that δ1 (WGn , WKr−1 ) ̸→ 0. By choosing a subse-
quence and than a subsequence of that, we may assume that δ1 (WGn , WKr−1 ) > a
for some a > 0 for all n, and Gn → U for some graphon U . Then t(L1 , U ) =
· · · = t(Lk , U ) = 0 and t(K2 , U ) = 1 − 1/(r − 1), so by Theorem 16.23, U is weakly
isomorphic to WKr−1 . So Gn → WKr−1 . Since WKr−1 is 0-1 valued, it follows by
Proposition 8.24 that δ1 (WGn , WKr−1 ) → 0, a contradiction.

16.5. Local vs. global optima

One advantage of embedding the set of graphs to the large space of graphons
is that for every optimization problem, we can define local optima, and study then
with the methods of analysis. We describe three problems where this treatment
gives interesting results.
In our first example, every local optimum is also global by convexity, and this
fact can be exploited to get a short proof. In the second example, we don’t know
whether local and global optima are the same; we can determine the local ones, and
perhaps these are also global. In the third example, local and global optima are
quite different, and the consideration of local optima leads to interesting results.

16.5.1. The distance from a hereditary graph property. A surprisingly

general result in extremal graph theory is the theorem of Alon and Stav [2008],
proving that for every hereditary property, a random graph with appropriate density
is asymptotically the farthest from the property in edit distance.
Theorem 16.24 (Alon and Stav). For every hereditary graph property P there
is a number p, 0 ≤ p ≤ 1, such that for every graph G with v(G) = n,
( )
d1 (G, P) ≤ E d1 (G(n, p), P) + o(1) (n → ∞).
The following theorem of Lovász and Szegedy [2010a] states a graphon version
of this fact. (Recall that, by Proposition 14.25, the closure of a hereditary graph
property is flexible. We omit the details of the derivation of Theorem 16.24 from
Theorem 16.25.)
Theorem 16.25. If R is a flexible graphon property, then the maximum d1 -distance
from R is attained by a constant function.
 16.5. LOCAL VS. GLOBAL OPTIMA 295

Proof. This proof illustrates the power of extending graph problems to a con-
tinuum. By Proposition 14.26, the set W0 \ R is convex. Hence it follows that the
d1 distance from R is a concave function on W0 \ R. We also know by Proposition
15.15 and Theorem 15.16(b) that d1 (., R) is a continuous function on (W f0 , δ ), and
hence it assumes its maximum. Let M be the set of maximizing graphons in W0 ;
this is a convex, closed subset of W0 . Since W0 \ R is invariant under the group
of invertible measure preserving transformations of [0, 1], so is M , and hence M is
also compact in the pseudometric δ .
This implies in many ways that M contains a constant function; here is a fast
argument. Let W ∈ M have minimum L2 -norm (such a graphon exists, since by
Lemma 14.15 the L2 -norm is lower semicontinuous with respect to the cut norm).
For every measure preserving transformation φ, we have (W + W φ )/2 ∈ M . Fur-
thermore,
W + Wφ
≤ 1 (∥W ∥2 + ∥Wφ ∥2 ) = ∥W ∥2 .
2 2
2
By the choice of W , we must have equality here, which implies that W and W φ
are proportional, but since they have the same integral, they must be equal almost
everywhere. Since this holds for any φ, W must be constant almost everywhere. 
16.5.2. The Sidorenko Conjecture. Sidorenko [1991, 1993] conjectured
that the inequality
(16.33) t(F, W ) ≥ t(K2 , W )e(F ) .
holds for all bipartite graphs F and all W ∈ W, W ≥ 0. Several special cases of
this inequality were mentioned in the Introduction, Section 2.1.2. Sidorenko in fact
formulated this not only for graphs but for graphons, being perhaps the first to
use the integral expression for t(., .) as a generalization of subgraph counting. (The
conjecture extends to non-symmetric functions W , but we restrict our attention to
the symmetric case here.) A closely related conjecture in extremal graph theory
was raised earlier by Simonovits [1984]. In spite of its very simple form and a lot
of effort, this conjecture is unproven in general.
It is easy to see that every graph satisfying Sidorenko’s Conjecture must be
bipartite. Indeed, if W = WK2 , then the right side of (16.33) is positive, but the
left side is positive only if F is bipartite.
We can view this as an extremal problem in two ways: (1) for every nonnegative
W ∈ W, matchings minimize t(F, W ) among all bipartite graphs with a given
number of edges; (2) for every bipartite graph F , constant functions W minimize
t(F, W ) among all nonnegative kernels W with a given integral. Since both sides
of (16.33) are homogeneous in W of the same degree, we can scale W and assume
that t(K2 , W ) = 1. Then we want to conclude that t(F, W ) ≥ 1 for every bipartite
graph F .
There are partial results in the direction of the conjecture. Sidorenko proved
it for a fairly large class of graphs, including trees, complete bipartite graphs, and
all bipartite graphs with at most 4 nodes in one of the color classes. After a long
period of little progress, several new (but unfortunately still partial) results were
obtained recently. Each of these is in one way or other related to the material in
this book, so we discuss them in some detail.
We have defined weakly norming graphs in Section 14.1. Hatami [2010] gives a
proof of the following (easy) fact, attributing it to B. Szegedy : If a bipartite graph
296 16. EXTREMAL THEORY OF DENSE GRAPHS

F is weakly norming, then it satisfies the Sidorenko conjecture. Combined with the
result of Hatami (Proposition 14.2) that all cubes are weakly norming, it follows
that all cubes satisfy Sidorenko’s conjecture.
In another direction, Conlon, Fox and Sudakov [2010] proved that the con-
jecture is satisfied by every bipartite graph that contains a node connected to all
nodes on the other side. Their proof uses a sophisticated probabilistic argument.
Li and Szegedy [2012] give a shorter analytic proof, which extends to a larger class
of graphs. Szegedy [unpublished] uses entropy arguments to prove the conjecture
for an even larger class, which includes of all previously settled special cases. The
smallest graph for which the conjecture is not known is the Möbius ladder of length
5 (equivalently, a 10-cycle with the longest diagonals added).

16.5.3. Local Sidorenko Conjecture. We can ask for conditions on W ,

rather than on F , that suffice to prove the conjectured inequality (16.33). Lovász
[2011] proved that for every bipartite graph F , the constant 1 kernel minimizes
t(F, W ) at least locally:
Theorem
∫ 16.26. Let F be a bipartite graph with m edges. Let W be a kernel with
W = 1 and ∥W − 1∥∞ ≤ 1/(4m). Then t(F, W ) ≥ 1. 
The proof of this theorem is not given here; instead, we prove the following
easier related result. Let us say that a graph F has the local Sidorenko property if
( kernel W ≥ 0) there is an εW > 0 such that for every 0 ≤ ε ≤ εW , we
for every
have t F, 1 + ε(W − 1) ≥ 1. (For a graph satisfying the Sidorenko conjecture, we
have εW = 1 for every W ≥ 0.) With this weaker notion, a graph does not have
to be bipartite to satisfy it. In fact, we have the following characterization of these
graphs:
Proposition 16.27. A graph has the local Sidorenko property if and only if either
it is a forest or its girth is even.
In particular, every bipartite graph has the local Sidorenko property.

Proof. Let us start with the “if” part. Replacing W by 1+(W −1)/∥W −1∥∞ ,
we may assume that 0 ≤ W ≤ 2. Let U = W −1, then U ∈ W1 . The homomorphism
density t(F, W ) = t(F, 1 + U ) can be expanded in terms of the subgraphs of F , and
so what we want to prove is
∑
t(F ′ , εU ) ≥ 1
F ′ ⊆F

for a sufficiently small ε > 0. (Let us agree that, for the rest of this section, F ′ ⊆ F
means that F ′ is a subgraph of F without isolated nodes.) The term with F ′ = ∅
is 1, and so (pulling out the ε factors) we want to prove that
∑ ′
(16.34) t(F ′ , U )εe(F ) ≥ 0.
∅̸=F ′ ⊆F

It follows from the definition of U that t(K2 , U ) = 0, and so every term in (16.34)
where F ′ is a matching cancels. If F itself is a matching, we have nothing to
prove. Otherwise, the next smallest term is t(P3 , U ), which is nonnegative, since
P3 = [[(K2• )2 ]] is a square. If t(P3 , U ) > 0, then for every sufficiently small ε > 0 it
dominates the sum (16.34), and we are done. So suppose that t(P3 , U ) = 0, then
 16.5. LOCAL VS. GLOBAL OPTIMA 297

tx (K2• , U ) = 0 for almost all x. This implies that t(F ′ , U ) = 0 whenever U has a
node of degree 1. In particular, if F is a forest, we are done.
Suppose that F is not a forest, then the nonzero term in 16.34 with the small-
est number of edges is t(C2r , U ), where C2r is the shortest cycle in F . Since
t(C2r , U )1/(2r) is a norm by Proposition 14.2, this term is nonzero if U ̸= 0, and so
for a sufficiently small ε > 0, it dominates the remaining terms.
To prove the “only if” part, suppose
( 1) that the girth of F is odd. Let U be
the graphon defined by the matrix −1 1 −1 . Then t x (K2
•
, U ) = 0 for every x, and
hence all those terms in (16.34) are 0 in which F has a node with degree 1. So
the nonzero terms with the smallest exponent of ε correspond to the shortest (odd)
cycles. Trivially t(F ′ , W ) = −1 for such a term, and so for a sufficiently small ε
the whole expression (16.34) will be negative. 
The proof of Theorem 16.26 expands the idea of the proof above, but one has
to do much more careful estimations. Many steps in the proof can be viewed as
using a version of the calculus “for W0 ” developed before, but this time “for W1 ”.
Some steps are described in Exercises 16.29-16.31 to illustrate this potentially useful
technique.
16.5.4. Common graphs. The following inequality is closely related to
Goodman’s Theorem (2.2), and it can be proved along the same lines:
1
(16.35) t(K3 , G) + t(K3 , G) ≥ ,
4
and equality holds asymptotically if G is a random graph with edge density 1/2.
Erdős conjectured that a similar inequality will hold for K4 in place of K3 , but this
was disproved by Thomason [1998]. More generally, one can ask which graphs F
satisfy ( )
tinj (F, G) + tinj (F, G) ≥ 1 + o(1) 21−e(F ) ,
where the o(1) refers to v(G) → ∞. Going to the limit, we get a formulation free
of remainder terms: Which simple graphs F satisfy
1
(16.36) t(F, W ) + t(F, 1 − W ) ≥ 21−e(F ) = 2t(F, )
2
for every graphon W ? Such graphs F are called common graphs. So the triangle is
common, but K4 is not. Are there any other common graphs?
Sidorenko [1996] studied graphs with this and other “convexity” properties.
Let F be a graph satisfying Sidorenko’s conjecture. Then
t(F, W ) + t(F, 1 − W ) ≥ t(K2 , W )e(F ) + t(K2 , 1 − W )e(F )
( t(K , W ) + t(K , 1 − W ) )e(F )
2 2
≥2 = 21−e(F ) ,
2
so F is common. Sidorenko’s conjecture would imply that all bipartite graphs are
common, and all bipartite graphs mentioned above for which Sidorenko’s conjecture
is verified are common. Among non-bipartite graphs, not many common graphs are
known. Jagger, Štovı́ček and Thomason [1996] showed that no graph containing
K4 is common.
Franek and Rödl [1992] showed that if we delete an edge from K4 , the obtained
graph K4′ is common. Recently Hatami, Hladky, Král, Norine and Razborov [2011]
proved that the 5-wheel is common, using computers to find appropriate nonnega-
tive expressions in the flag algebra. We cannot reproduce their proof here; instead,
298 16. EXTREMAL THEORY OF DENSE GRAPHS

let us give the proof of the fact that K4′ is common, which should give a feeling for
this technique. We do our computations in the graph algebra, instead of the flag
algebra.
We start with rewriting (16.36) as follows. Let U = 2W − 1, then substituting
U in (16.36) and multiplying by 2e(F ) , we get
(16.37) t(F, 1 + U ) + t(F, 1 − U ) ≥ 2,
which should hold for every U ∈ W1 . In other words, F is common iff the left side
is minimized by U = 0.
The subgraph densities on the left side of (16.37) can be expanded as before,
and we get ∑
t(F, 1 + U ) + t(F, 1 − U ) = 2 t(F ′ , U ).
F ′ ⊆F
e(F ′ ) even
The term with e(F ′ ) = 0 gives 2, the value on the right side of (16.37), so F is
common if and only if
∑
(16.38) t(F ′ , U ) ≥ 0
F ′ ⊆F
e(F ′ )>0, even

for every U ∈ W1 . Note that inequality (16.38) has to be true for all U ∈ W1 , not
just for U ∈ W0 , so the fact that all terms on the left have nonnegative coefficient
does not make this relation trivial.
As an example, in the case F = K4′ we get from (16.38) that its commonness
follows if we can show that
(16.39) 2 +8 + +4 ≥ 0 (for W1 ).
Here we can write the left side as
( )2
]] + 4( )
2 2
[[
2 +4 + +2 − .
It is easy to see (see Exercise 16.30) that the last term is nonnegative. The other
terms are squares, which proves (16.39).
Locally common graphs. We say that a graph F is locally common, if for every
U ∈ W1 there is a 0 < εU ≤ 1 such that if 0 < ε < εU , then t(F, 1 + εU ) + t(F, 1 −
εU ) ≥ 2.
Franek and Rödl [1992] proved that K4 is locally common. In fact, the follow-
ing more general result holds, and can be proved along the lines of the proof of
Proposition 16.27, using formula 16.38.
Proposition 16.28. Let G be a graph in which the subgraph with the minimum
number of edges such that all degrees are at least 2 and the number of edges is even
is an even cycle. Then G is locally common. 
In particular, every bipartite graph is locally common (this follows by Propo-
sition 16.27 as well), and so is every simple graph containing a 4-cycle. Combining
with the theorem of Jagger, Štovı́ček and Thomason [1996] mentioned above, it
follows that every graph that contains a K4 is not common but locally common.
Not all graphs are locally common (see Exercise 16.33).
Exercise 16.29. Prove that C2 ≥ C4 ≥ C6 ≥ . . . (for W1 ).
 16.6. DECIDING INEQUALITIES BETWEEN SUBGRAPH DENSITIES 299

Exercise 16.30. Prove that [[F12 F2 ]] ≤ [[F12 ]] (for W1 ) for any two k-labeled
multigraphs F1 and F2 .
Exercise 16.31. Suppose that a bipartite graph F contains a 4-cycle. Prove that
F ≤ C4 (for W1 ). More generally, if F is not a forest, then F ≤ C2r (for W1 ),
where C2r is the shortest cycle in F .
Exercise 16.32. Prove that triangles are common (16.35).
Exercise 16.33. Prove that [[C7• C11
•
]] is not locally common.

16.6. Deciding inequalities between subgraph densities

Now we turn to more general questions in extremal graph theory, as we already
indicated in the Introduction.

16.6.1. Undecidability of density inequalities. In analogy with Artin’s

Theorem for real polynomials (see Appendix A.7), we may try to represent quantum
graphs g with g ≥ 0 as sums of squares or, more generally, as quotients of sums of
squares: if y and z are square-sums and y = zg, then g ≥ 0. The following theorem
by Hatami and Norine [2011] tells us once and forever that no such “certificate” of
nonnegativity can be given.
Theorem 16.34. It is algorithmically undecidable whether an inequality g ≥ 0
holds (where g is a quantum graph with rational coefficients).
It is a related, but easier, fact that the following is algorithmically undecidable:
given
∑ simple graphs F1 , . . . , Fm and rational coefficients a1 . . . , am , decide whether
i ai hom(F, Gi ) ≥ 0 holds for every simple graph G. This follows from a result
of Ioannidis and Ramakrishnan [1995], proved in a completely different setting of
databases; for the adaptation to the graph case, see Exercise 16.43, based on the
lecture notes of Kopparty [2011].
The proof will consist of a reduction of the problem of deciding whether a given
polynomial p ∈ Z[x1 , . . . , xk ] is nonnegative for every x1 , . . . , xk ∈ N, to the problem
of deciding whether x ≥ 0 for a quantum graph g with rational coefficients. Since
the latter problem is undecidable by the Theorem of Matiyasevich (see Section A.7
in the Appendix), this will imply that so is the former. To this end, we need a
version of Problem A.33, which asks for nonnegativity on another set of numbers,
namely on the set
{ 1 }
A = 1 − : n = 1, 2, . . . .
n
The variables will be represented by edge densities and triangle densities in ap-
propriate graphs, and the fact that Goodman’s bound is only attained for edge
densities in A (Corollary 16.12) will then be used to force the edge-densities to be
of this form.
Proof of Theorem 16.34. We start with some algebraic reductions.
Claim 16.35. It is algorithmically undecidable whether an inequality p ≥ 0 (p ∈
Z[x1 , . . . , xk ]) holds on Ak .
Indeed, one can reduce deciding if p ≥ 0 on Nk to deciding whether p ≥ 0 on
k
A by a straightforward change of variables.
The connection to graph theory
( is established )by the following reduction. Con-
sider the set D2,3 of all pairs t(K2 , W ), t(K3 , W ) where W is a graphon (Figures
300 16. EXTREMAL THEORY OF DENSE GRAPHS

( Let D ⊆ D2,3 consist ) of the points (x, y) where x ∈ A and y = 2x −x

2
2.1 and 16.1).
(i.e., pairs t(K2 , Kn ), t(K3 , Kn ) together with (0, 0) and (1, 1)).
For a polynomial p ∈ Z[x1 , . . . , xk ], construct the polynomial in 2k variables:
∏
k ∑
k
p∗ (x1 , y1 , . . . , xk , yk ) = (1 − xi )2 p(x1 , . . . , xk ) + M (yi − 2x2i + xi ),
i=1 i=1

where M = 2 max{∥(gradp)(x)∥∞ : x ∈ [0, 1] }. We also construct a polynomial

k

in 3k variables:
( v1 w1 vk wk )
p∗∗ (u1 , v1 , w1 , . . . , uk , vk , wk ) = (u1 . . . uk )N p∗ 2 , 3 , . . . , 2 , 3 ,
u1 u1 uk uk
(where N = 5 deg(p∗ ) is large enough to cancel all denominators).
Claim 16.36. The following are equivalent: (i) p ≥ 0 on Ak ; (ii) p∗ ≥ 0 on D2,3
k
;
∗
(iii) p ≥ 0 on D .
k

Trivially, (ii)⇒(iii)⇒(i). To prove that (i)⇒(ii), assume that p ≥ 0 on Ak , and

let (xi , yi ) ∈ D2,3 . We want to prove that p∗ (x1 , y1 , . . . , xk , yk ) ≥ 0. We start with
bounding the first term in the definition of p∗ . Let zi ∈ A be closest to xi . Then
p(z1 , . . . , zk ) ≥ 0, and so
p(x1 , . . . , xk ) ≥ p(x1 , . . . , xk ) − p(z1 , . . . , zk ) = (x − z) · (gradp)(ξ),
where ξ ∈ [0, 1]k . By the definition of M , we get that
∏
k ∏
k
M∑
k
(16.40) (1 − xi )2 p(x1 , . . . , xn ) ≥ − (1 − xi )2 |xi − zi |
i=1 i=1
2 i=1

M∑
k
≥− (1 − xi )2 |xi − zi |.
2 i=1
We show that each term is compensated for by the corresponding term in the other
part of p∗ , i.e.,
1
(16.41) (1 − xi )2 |xi − zi | ≤ yi − 2x2i + xi .
2
Let us assume e.g. that xi ≤ zi (the other case is similar). Let wi ∈ A be the
closest point to xi with wi < xi .
By Corollary 16.12, yi is above the chord between zi and wi of the parabola
2x2 + x. On the other hand, 2x2i + xi is below the chord between zi and (zi + wi )/2.
The slope of the first chord is 2zi + 2wi − 1; the slope of the second, 3zi + wi − 1.
The difference in slopes is zi − wi , and so yi − 2x2i + xi ≥ (zi − wi )(zi − xi ). Simple
computation shows that
(1 − wi )2 1
zi − wi = ≥ (1 − xi )2 .
2 − wi 2
This proves (16.41) and thereby also Claim 16.36.
As a preparation for the rest of the proof, we need to construct some special
graphs. We fix a simple graph F with node set [k] that has no automorphisms.
For any set of positive integers n1 , . . . , nk , let F (n1 , . . . , nk ) denote the (unlabeled)
graph obtained from F by replacing every node i by a set of ni twins. We will call
these ni nodes the clones of i.
 16.6. DECIDING INEQUALITIES BETWEEN SUBGRAPH DENSITIES 301

As a related construction, we define Fir (i ∈ [k], r ≥ 0) by adding r new twins of

node i, making them mutually adjacent and adjacent to i, and labeling the original
nodes 1, . . . , k. The node i and the new nodes will be called the clones of i. So F1,r
is a k-labeled version of F (r + 1, 1, . . . , 1).

Figure 16.2. Auxiliary graphs for the proof of theorem 16.34.

As a further step, we add all missing edges to Fir with negative sign, to get the
k-labeled signed graph Fbir (as we have seen, this can be considered as a k-labeled
quantum graph). Recall that Fb is defined analogously.
Claim 16.37. Every homomorphism of Fb into any graph G is an induced embed-
ding.
Indeed, every homomorphism preserves both edges and non-edges by the def-
inition of signed graphs. Suppose that two nodes u, v ∈ V (F ) are mapped onto
the same node of G. Then every further node of F must be connected to u and
v in the same way, and so interchanging u and v is an automorphism of F , which
contradicts the choice of F .
Claim 16.38. For every homomorphism of Fb into any of the special graphs
F (n1 , . . . , nk ), each node i ∈ V (F ) is mapped onto a clone of i.
The proof is similar to the previous one. We already know that the map is
injective. For u ∈ V (F ), let σ(u) be defined as the node of F whose clones in
F (n1 , . . . , nk ) contain the image of u. No two nodes u, v ∈ V (F ) have σ(u) = σ(v):
similarly as before, interchanging two such nodes would be an automorphism of F .
Hence σ is an automorphism of F , and hence σ must be the identity. This proves
the Claim.
Our next observation is that homomorphism densities from the signed graphs
Fir into any simple graph G can be expressed quite simply. Let G be any simple
graph, and let φ : [k] → V (G), and let S = φ([k]) be its range. Let Uφ,i be the set
of nodes in V (G) \ S which are connected to φ(i) and all the neighbors of φ(i) in
S, but to no other node in S.
We claim that
{
hom(Kr , G[Uφ,i ]) if φ is an induced embedding of F ,
(16.42) homφ (Fbir , G) =
0 otherwise.
Assume first that φ is an induced embedding of F into G. It is clear that if ψ is any
homomorphism of Fbir into G extending φ, then all the clones of i in Fbir must be
mapped onto nodes in Uφ,i . Since these twins form a complete graph Kr , the num-
ber of ways to map these twins into G[Uφ,i ] homomorphically is hom(Kr , G[Uφ,i ]),
and every such map, together with φ, forms a homomorphism of Fbir into G. Claim
302 16. EXTREMAL THEORY OF DENSE GRAPHS

16.37 implies that homφ (Fbir ) = 0 if φ is not an induced embedding of F . Hence

(16.42) follows.
Normalizing the homomorphism numbers in (16.42), we get that if φ is an
induced embedding, then
|Uφ,i |r
(16.43) tφ (Fbir , G) = t(Kr , G[Uφ,i ]).
v(G)r
We want to reduce the problem of deciding whether p ≥ 0 on Ak (for a given
polynomial p ∈ Z[x1 , . . . , xk ]) to deciding whether g ≥ 0 for a quantum graph
g. Given p, we construct the polynomials p∗ ∈ Z[x1 , y1 , . . . , xk , yk ] and p∗∗ ∈
Z[u1 , v1 , w1 , . . . , uk , vk , wk ] as above, and define the k-labeled quantum graph
g = p∗∗ (Fb11 , Fb12 , Fb13 , . . . , Fbk1 , Fbk2 , Fbk3 ).
The key step in the proof is the following:
Claim 16.39. We have [[g]] ≥ 0 if any only if p ≥ 0 on Ak .
To start with the “if” direction, assume that p ≥ 0 on Ak . Then p∗ ≥ 0 on
k
D2,3 by Claim 16.36. We want to prove that t([[g]], G) ≥ 0 for every graph G. This
follows if we show that tφ (g, G) ≥ 0 for every map φ : [k] → V (G). To simplify
notation, set tir = tφ (Fbir , G). Then
tφ (g, G) = p∗∗ (t11 , t12 , t13 , . . . , tk1 , tk2 , tk3 )
(t tk2 tk3 )
12 t13
= (t11 . . . tk1 )N p∗ 2 , 3 , . . . , 2 , 3 .
t11 t11 tk1 tk1
( )
By (16.43), we have (t12 /t11 , t13 /t11 ) = t(K2 , G[Uφ,i ]), t(K3 , G[Uφ,i ]) ∈ D2,3 , and
2 3

hence it follows that tφ (g, G) ≥ 0.

To prove the converse, assume that [[g]] ≥ 0. By Claim 16.36, it suffices to prove
that p∗ ≥ 0 on Dk . Let xi = (ni − 1)/ni ∈ A and yi = 2x2i − xi = (n2i − 3ni + 2)/n2i ,
and consider the graph G = F (n1 +1, . . . , nk +1). By our assumption, t([[g]], G) ≥ 0.
We can write ∑ 1
t([[g]], G) = tφ (g, G).
v(G)k
φ:[k]→V (G)

In every constituent of g, the labeled nodes induce a copy of Fb , which implies by

(16.42) that only those terms where φ is an induced embedding are nonzero. By
Claim 16.38, such mappings φ map every node of F onto a clone of it, and hence
tφ (g, G) is the same for every induced embedding. This implies that tφ (g, G) ≥ 0
for any such map φ. Let us fix an induced embedding φ of F into G.
Now we can apply (16.43): the set Uφ,i induces a complete graph with ni nodes,
and hence
|ni |r (ni )r
tφ (Fbir , G) = r t(Kr , Kni ) = .
n nr
This implies that
( )
n1 (n1 )2 (n1 )3 nk (nk )2 (nk )3
tφ (g, G) = p∗∗ , 2 , 3 ,..., , ,
n n n n n2 n3
( n . . . n )N
p∗ (x1 , y1 , . . . , xk , yk ).
1 k
=
nk
Since the left side is nonnegative, this implies that p∗ (x1 , y1 , . . . , xk , yk ) ≥ 0 as
claimed.
 16.6. DECIDING INEQUALITIES BETWEEN SUBGRAPH DENSITIES 303

This proves Claim 16.39, and together with Claim 16.35, it completes the proof
of the theorem. 
As mentioned in the introduction, an inequality g ≥ 0, where g is a quantum
graph, is decidable with an arbitrarily small error:
Proposition 16.40. There is an algorithm that, given a quantum graph g with
rational coefficients and an error bound ε > 0, decides either that g  0 or that
g + εK1 ≥ 0 (if both inequalities are true, then it may return either answer).
Proof. This will follow from Theorem 16.41 in the next section, ∑ but let us
describe a simple direct proof suggested
∑ by Pikhurko. Let g = F a F F be the
given quantum graph, let a = F |aF |e(F ), ε1 = ε/a. By Corollary 9.25, there
is an integer k ≥ 1 such that all simple ∑ graphs with k nodes form an ε1 -net in
(W0 , δ ). Let us check the inequality F aF t(F, G) ≥ 0 for all simple graphs G
with at most k nodes. If we find a graph that violates it, we know that g  0. Else,
let G be any simple graph. By the definition of k, there is a simple graph G′ on k
nodes such that δ (G, G′ ) ≤ ε1 , and hence by the Counting Lemma 10.22, we have
∑ ∑ ∑
aF t(F, G) ≥ aF (t(F, G′ )−e(F )ε1 ) ≥ aF t(F, G′ )−ε ≥ −ε = −εt(K1 , G′ ),
F F F
so we can conclude that g + εK1 ≥ 0. 
16.6.2. Positivstellensatz for graphs. Is there a quantum graph g ≥ 0
which is not a square sum? Hatami and Norine [2011] constructed such a quantum
graph. In fact, the existence of such a quantum graph follows from Theorem 16.34,
stating that it is algorithmically undecidable whether a quantum graph with ra-
tional coefficients is nonnegative. To see this, consider two Turing machines, both
working on an input which is a quantum graph g with rational coefficients. We may
assume that the constituents of g have no isolated nodes. One of them will look for
a graph G with t(g, G) < 0; the other, for a representation of g as a square-sum.
If for every input one of them halts, then we know whether or not g ≥ 0. So there
must be an input g on which both Turing machines run forever; then we have g ≥ 0,
and g is not a square sum.
(To be precise, we must add that if g is a square-sum, then it is a square-sum
where the coefficients in the quantum graphs yi in the definition are algebraic real
numbers; then there are only a countable number of possibilities, and the second
Turing machine can check them in an appropriate order. One needs a method to
check that given k-labeled quantum
∑ graphs yi , . . . , yk with algebraic coefficients,
whether g is obtained from i yi2 by unlabeling and deleting isolated nodes. Such
an algorithm follows from Tarski’s Theorem on the decidability of the first order
theory of real numbers.)
But not all is lost: the following weaker result was proved by Lovász and
Szegedy [2012a].
Theorem 16.41. Let f be a quantum graph. Then f ≥ 0 if and only if for every
ε > 0 there is a square-sum g such that ∥f − g∥1 < ε.
An analogous theorem for nonnegative polynomials was proved by Lasserre
[2007].
Proof. The “if” part is trivial. The idea of the proof of
∑ the “only if” part is
the following. Consider the (unlabeled) quantum graph g = F aF F (where only a
304 16. EXTREMAL THEORY OF DENSE GRAPHS

finite number of the aF are nonzero). We may assume that no graph F with aF > 0
contains an isolated node, since removing isolated nodes does not change t(g, W ).
The condition g ≥ 0 means that h(g) ≥ 0 for every graph parameter h of the form
h = t(., W ) with W ∈ W0 . This constraint is linear, so we can(equivalently
) require
the inequality for every graph parameter of the form h = E t(., W) , where the
expectation is over some probability distribution on graphons (see Section 14.5).
By Proposition 14.60, this is equivalent to requiring that h is normalized, isolate-
indifferent∑ and reflection positive. We can forget about the normalization, since the
condition F aF h(F ) ≥ 0 is homogeneous. So the question is: Does the inequality
∑
(16.44) aF h(F ) ≥ 0
F
hold for every isolate-indifferent and reflection positive graph parameter h?
This problem can be rephrased in terms of the connection matrix X = M (h, N),
whose entries we consider as unknowns. These unknowns are not all different: if
(for this proof only) F ′ denotes the graph obtained from F by removing its isolated
nodes, then we have XF1 ,F2 = XG1 ,G2 whenever [[F1 F2 ]]′ ∼
= [[G1 G2 ]]′ . The reflection
positivity conditions mean that X ≽ 0. The question is: Do these constraints imply
the inequality
∑
(16.45) aF XF,K0 ≥ 0
F

(In this last sum, all the graphs F are unlabeled.)

This is just a feasibility problem in semidefinite programming —apart from the
“minor” problem that the unknowns form an infinite matrix, and have to satisfy
an infinite number of constraints. We will have to cut the program to finite size,
which will bring in the error in the theorem.
But let us ignore the problems with infinities, and apply the Semidefinite Farkas
Lemma: We have to assign, as a Lagrange multiplier, a matrix Y ≽ 0, which has
to satisfy
∑
(16.46) YF1 ,F2 = aF
F1 ,F2 : [[F1 F2 ]]′ ∼
=F

for every F ∈ F simp (where the summation extends over all partially labeled simple
graphs F1 and F2 ). We can rewrite this as
∑ ∑
g= aF F = YF1 ,F2 [[F1 F2 ]]′
F F1 ,F2
T
Let us write Y = ZZ with some matrix Z; this takes care of the semidefiniteness
condition (remember, we are ignoring the problem that these matrices are infinite).
Then
∑ ∑ ∑ (∑ )2
g= ZF1 ,m ZF2 ,m [[F1 F2 ]]′ = [[
Zm,F F . ]]
F1 ,F2 m m F
showing that g is a square-sum.
Now we have to make this argument precise. Let Fk′ denote the set of fully
labeled graphs on [k]. Let M denote the linear space of all symmetric matrices
indexed by partially labeled simple graphs, let P be the subset of M consisting of
positive semidefinite matrices, and let L denote the subspace of matrices satisfying
XF1 ,F2 = XG1 ,G2 whenever [[F1 F2 ]]′ ∼
= [[G1 G2 ]]′ . Clearly, P is a convex cone. Let
 16.6. DECIDING INEQUALITIES BETWEEN SUBGRAPH DENSITIES 305

Φk denote the operator mapping a matrix in M to its restriction to Fk′ × Fk′ (this
is a finite matrix!). Then Mk = Φk M is the space of all symmetric Fk′ × Fk′
matrices, and Pk = Φk P is the positive semidefinite cone in Mk . It is also clear
that Lk = Φk L consists of those matrices X ∈ Mk for which XF1 ,F2 = XG1 ,G2
whenever [[F1 F2 ]]′ ∼
= [[G1 G2 ]]′ . Clearly,
(16.47) Φk (P ∩ L) ⊆ Pk ∩ Lk ,
but equality may not hold in general.

'k
%k 'm,k %m %
'm

(k $k (m $m ( $

(k‚$k (m‚$m (‚$

Figure 16.3. Spaces and cones in the proof of Theorem 16.41.

We note that the entries of every matrix X ∈ Pk ∩ Lk are in the interval

[0, X∅,∅ ]. Indeed, for any F ∈ Fk′ and the fully labeled edgeless graph Uk ∈ Fk′ ,
the condition X ∈ Lk implies that XUk ,F = XF,F = XF,Uk , and so X ∈ Pk implies
that 0 ≤ XF,F ≤ X∅,∅ . Since XF,G = XF G,F G , the claim follows.
For k ≤ m, we embed Fk′ into Fm ′
, by adding m − k isolated nodes labeled
′ ′
k + 1, . . . , m. The corresponding operator restricting Fm × Fm matrices to Fk′ × Fk′
will be denoted by Φm,k .
We claim that the following weak converse of (16.47) holds:
∩
(16.48) Φk (P ∩ L) = Φm,k (Pm ∩ Lm ).
m≥k

Indeed, let A be a matrix that is contained in the right hand side. Then for every
m ≥ k we have a matrix Bm ∈ Pm ∩ Lm such that A is a restriction of Bm .
Now let m → ∞; by selecting a subsequence, we may assume that all entries of
Bm tend to a limit. This limit defines a graph parameter f , which is normalized,
isolate-indifferent and flatly reflection positive. By Proposition 14.60, f is reflection
positive, and so the matrix M (f ) is in P ∩ L and Φk M (f ) = A.
We may assume that |V (F )| = k whenever aF ̸= 0. Let A ∈ Mk denote the
matrix {
aF , if F = G,
AF G =
0, otherwise.
Then g ≥ 0 means that A · Z ≥ 0∑for all Z ∈ Φk (P ∩ L) (where the inner product
A·Z of two matrices is defined as i,j Aij Zij ). In other words, A is in the dual cone
of Φk (P ∩ L). From (16.48) it follows that there are diagonal matrices Am ∈ Mk
such that Am → A and Am · Y ≥ 0 for all Y ∈ Φm,k (Pm ∩ Lm ). In other words,
Am · Φm,k Z ≥ 0 for all Z ∈ Pm ∩ Lm , which can also be written as Φ∗m,k Am · Z ≥ 0,
where Φ∗m,k : Mk → Mm is the adjoint of the linear map Φm,k : Mm → Mk .
(This adjoint acts by adding 0-s.) So Φ∗m,k Am is in the polar cone of Pm ∩ Lm ,
306 16. EXTREMAL THEORY OF DENSE GRAPHS

∗
which is Pm + L∗m . The positive semidefinite cone ∑ is self-polar. The linear space
∗
Lm consists of those matrices B ∈ Mm for which F1 ,F2 BF1 ,F2 = 0, where the
summation extends over all pairs F1 , F2 ∈ F ′ M for which F1 F2 ≃ F0 , for every
fixed graph F0 . Thus we have Φ∗m,k Am = P + L, where P is positive semidefinite
∑N
and L ∈ L∗m . Since P is positive semidefinite, we can write it as P = k=1 vk vkT ,
′
where vk ∈ RFm . We can write this as
{
∑ ∑ N
(Am )F0 ,F0 , if F1 F2 ≃ F0 ∈ Fk′ ,
vk,F1 vk,F2 =
F ,F k=0
0, otherwise.
1 2
F1 F2 ≃F0

In other words,
( )2
∑
N ∑ ∑
vk,F F = (Am )F0 ,F0 F0 .
k=1 F F0
So the quantum graph on the right side is a sum of squares. Furthermore, if m → ∞,
then Am → A and so
∑ ∑
(Am )F0 ,F0 F0 → AF0 ,F0 F0 = f. 
F0 F0

In view of the usefulness of extending graphs to graphons, it seems natural to

define graph algebras of infinite formal linear combinations of graphs with appro-
priate convergence properties; in other words, of graph parameters. It has not been
worked out, however, what the structure of the resulting algebra is, and how it is
related to graphons.
Proposition 14.61 suggests that it should be enough to use fully labeled quan-
tum graphs when approximating a nonnegative quantum graph as a square sum.
This is indeed true (see Exercise 16.44), but the approximation is very inefficient,
as shown by the following example.
Example 16.42. Consider the quantum graph
g= − ≥ 0.
If we allow unlabeled nodes, then g can be represented (up to labels and isolated
nodes) as a square sum (in fact, as a single square):
( )2
− .
But g cannot be represented as a square sum of fully labeled graphs. To see this, let
Sk denote the
∑set of quantum graphs obtained by deleting isolates and unlabeling
square sums i yi2 , where every constituent of yi is a fully labeled graph on k nodes.
Consider the graph parameter
 (k)

 3 4 if e(G) = 0,
(k−2) if e(G) = 1,

f (G) = 2

 1 if E(G) consists of two disjoint edges,


0 otherwise.
Then f (g) < 0. On the other hand, we claim that f (y 2 ) ≥ 0 if every constituent
of y is a fully labeled graph on k nodes. This means that the flat connection ma-
trix M flat (f, k) is positive semidefinite. By the Lindström–Wilf Formula (A.1), this
 16.7. WHICH GRAPHS ARE EXTREMAL? 307

is equivalent to saying that the upper Möbius inverse f ↑ (see (4.1)) is nonnega-
tive, which is easy to check. Since f is isolate-indifferent, this proves that f is
nonnegative on every quantum graph in Sk . It follows that g ∈/ Sk for any k.
But g can be approximated by members of Sk for large k. To construct such
an approximation, let Hij (1 ≤ i < j ≤ k) consist of V (G) = [k] and a single edge
connecting i to j. Expanding and unlabeling the quantum graph
 2
1 ∑ 1 ∑
h2k =  H1i − (k−1) Hij  ,
k−1 2
2≤i≤k 2≤i<j≤k

we get
( ) 1 ( )
− + k − (k − 6) + (4k − 10) .
(k − 1)(k − 2)
The last term tends to 0 as k → ∞, so g is arbitrarily well approximated by h2k . 
Razborov and his collaborators found several concrete applications of the
method of semidefinite programming in extremal graph theory. Let us mention
one. Paul Erdős [1984] conjectured in 1984 that the number of pentagons in a
triangle-free graph with a given number of nodes is maximized by blow-ups of a
pentagon. In our language, if G is a triangle-free simple graph, then
2
(16.49) t(C5 , G) ≤ t(C5 , C5 ) = .
625
In spite of its naturality and simple form, this conjecture remained unproven un-
til recently, when Hatami, Hladky, Kral, Norine and Razborov [2011] and inde-
pendently Grzesik [2012] found a proof, using flag algebras and computer-assisted
solutions to semidefinite programs.
∑
Exercise 16.43. Prove that the problem whether i ai hom(F, Gi ) ≥ 0 holds for
every simple graph G (for given simple graphs F1 , . . . , Fm and integer coefficients
a1 . . . , am ) is undecidable. [Hint: Use the result of Exercise 5.50 and Matiyase-
vich’s Theorem.]
Exercise 16.44. Use the method of Example 16.42 to show that if a quantum
graph g can be written as a square sum (unlabeled and isolates removed), then it
can be approximated up to an arbitrarily small error by square-sums of quantum
graphs with fully labeled constituents.

16.7. Which graphs are extremal?

Is there a “template” describing the structure of extremal graphs for an ex-
tremal graph problem? For “classical” extremal problems, discussed in Section
16.4, this has been answered by the Erdős–Simonovits–Stone theory of extremal
graphs (at least in the dense case). As we have seen, in an asymptotic sense (for
large graphs) the only extremal graphs are the Turán graphs. In the language
of graphons, every optimization problem in the sense of Theorem 16.23 has an
optimum solution of the form WKq for some q ≥ 2, and the solution is unique.
More general stepfunctions (with positive integral values) have been found to
serve as optimum solutions of generalizations of these problems to directed graphs
and multigraphs, in the work of Brown, Erdős and Simonovits [1973, 1978].
308 16. EXTREMAL THEORY OF DENSE GRAPHS

In this discussion we stay with simple graphs, but we consider a much more
general type of graph theoretic extremal problem:
(16.50) maximize t(f, W )
subject to t(g1 , W ) = a1
..
.
t(gk , W ) = ak
where f, g1 , . . . , gk are given simple quantum graphs. Most of the graphon versions
of extremal problems discussed so far fit this scheme. Is there a special family of
graphons such that every extremal graph problem has a solution from this family?
We define an interesting class of graphons that are all needed, and we conjecture
that they are also sufficient.

16.7.1. Finitely forcible graphons. To motivate this definition, recall the

result of Chung, Graham and Wilson from Section 1.4.2 that a sequence of graphs
(Gn , n = 1, 2, . . . ) with the property that t(K2 , Gn ) → p and t(C4 , Gn ) → p4
for some 0 < p < 1 is quasirandom, so it is convergent and its limit is the all-p
function. In other words, t(F, Gn ) → t(F, Lp ) for every simple graph F , where Lp
is the weighted graph with one node with a loop of value p.
Are there other convergent graph sequences that can be characterized by the
convergence of a finite number of homomorphism densities t(F, Gn )? Are perhaps
all convergent graph sequences like this?
We can translate this question to the language of graphons, which allows a more
precise statement: Which graphons W have a finite ”forcing” set of simple graphs
F1 , . . . , Fm such that if any other graphon U has the property that t(Fi , U ) =
t(Fi , W ) (i = 1, . . . , m), then U and W are weakly isomorphic, i.e., t(F, U ) =
t(F, W ) for every simple graph F ? Such a graphon will be called finitely forcible.
In other words, a finitely forcible graphon forms a single-element variety in the
graphon space. The theorem of Chung, Graham and Wilson quoted above shows
that every constant function is finitely forcible.
Coming down to the earth, this notion has the following version for finite
graphs: a convergent simple graph sequence (Gn ) is finitely forcible, if there
is a finite set of simple graphs F1 , . . . Fm such that whenever Hn is another
graph sequence such that limn t(Fi , Gn ) = limn t(Fi , Hn ) for i = 1, . . . , m, then
limn t(F, Gn ) = limn t(F, Hn ) for every simple graph F . So it follows from the
characterization of quasirandom graph sequences by Chung, Graham and Wilson
[1989] in Section 1.4.2 that every quasirandom graph sequence is finitely forcible.
We will see that “most” graphons (in a Baire category sense) are not finitely
forcible. On the other hand, there are interesting families of finitely forcible
graphons. In the next section, we describe some such families. See Lovász and
Szegedy [2011] for more constructions of finitely forcible graphons.
Trivially, every finitely forcible graphon is the unique solution of an extremal
problem of type (16.50). Perhaps the converse is also true:
Conjecture 16.45. Every extremal problem has a finitely forcible optimum. In
other words, if a finite set of constraints of the form t(Fi , W ) = ai is satisfied by
some graphon, then we can add a finite number of further constraints of this type
to make the solution unique (up to weak isomorphism).
 16.7. WHICH GRAPHS ARE EXTREMAL? 309

16.7.2. Many finitely forcible graphons. In this section, we describe some

reasonably general constructions for finitely forcible graphons.
Stepfunctions. Almost all classical extremal problems have a solution whose
“template” is a stepfunction. It was shown by Lovász and Sós [2008] that every
stepfunction is finitely forcible, which means that every stepfunction is the template
of an appropriate extremal graph problem. (It was conjectured that these are the
only ones, but as we shall see there are many more.) We give a somewhat stronger
result with a better bound on the size of the forcing graphs F .
Theorem 16.46. Let H be a weighted graph with nonnegative nodeweights and
with edgeweights from [0, 1], and let U be a kernel. If t(F, U ) = t(F, WH ) for all
graphs F with v(F ) ≤ 4(2v(H) + 3)8 , then U is weakly isomorphic to WH .
Proof. By Proposition 14.44 we know that stepfunctions with q = v(H) steps
form a variety: there is a simple quantum graph g such that t(g, W ) = 0 if and only
if W ∈ Sq , where every constituent of g has (q + 2)(q + 1) nodes. So t(g, WH ) =
t(g, H) = 0, and hence t(g, U ) = 0, which implies that U is a stepfunction with
q steps. So we may assume that there is a weighted graph H ′ with q nodes such
that U = WH ′ almost everywhere. It follows that for every simple graph F with
at most 4(2q + 3)8 nodes, we have t(F, H ′ ) = t(F, U ) = t(F, H). By Theorem 5.33,
this implies that H ∼
= H ′ , and so U = WH almost everywhere. 

Corollary 16.47. Every stepfunction is finitely forcible.

Theorem 16.46 implies for finite graphs:
Corollary 16.48. If H is weighted graph with edgeweights from [0, 1], and (Gn )
is a graph sequence such that t(F, Gn ) → t(F, H) for every simple graph F with
v(F ) ≤ 4(2v(H) + 3)8 , then t(F, Gn ) → t(F, H) for every simple graph F (in other
words, Gn → WH ).
Threshold graphons. The simple threshold graphon 1(x + y ≤ 1) introduced
in Example 11.36 is perhaps the simplest finitely forcible graphon that is not a
stepfunction. We in fact have a more general class. Let p(x, y) be a symmetric real
polynomial. We call the graphon U (x, y) = 1(p(x, y) > 0) the p-threshold graphon.
Proposition 16.49. If p is monotone decreasing on [0, 1]2 , then the p-threshold
graphon U is finitely forcible among kernels.
The proof will show that U is determined by the equations
(16.51) b4 , W ) = 0
tx1 x2 x3 x4 (C
and
(16.52) t(Ka,b , W ) = t(Ka,b , U ) (1 ≤ a, b ≤ 2 deg(p) + 4)
b4 is the signed 4-cycle defined in Example 14.42). It can be conjectured that
(here C
the monotonicity condition is not needed.

Proof. The condition on the monotonicity of p implies that W = U satisfies

(16.51). It is trivial that equations (16.52) are satisfied by W = U .
Let W ∈ W be any graphon satisfying (16.51)-(16.52). As discussed in Example
14.42, we may assume that W is a threshold graphon, i.e., 0-1 valued and monotone
310 16. EXTREMAL THEORY OF DENSE GRAPHS

decreasing. Let SW = {(x, y) : W (x, y) = 1}. We have

∫ ∫ a ∏
∏ b
t(Ka,b , W ) = W (xi , yj ) dy dx.
i=1 j=1
[0,1]a [0,1]b

Split this integral according to which xi and which yj is the largest. Restricting the
integral to, say, the domain where x1 and y1 are the largest, we have that whenever
W (x1 , y1 ) = 1 then also W (xi , yj ) = 1 for all i and j, and hence
∫ ∫ ∫ ∫ ∏
a ∏
b
W (xi , yj ) dy dx
i=1 j=1
x1 ∈[0,1] x2 ,...,xa ≤x1 y1 ∈[0,1] y2 ,...,yb ≤y1
∫ ∫ ∫
= W (x1 , y1 )xa−1
1 y1b−1 dy1 dx1 = xa−1 y b−1 dy dx .
x1 ∈[0,1] y1 ∈[0,1] (x,y)∈SW

Since there are a choices for the largest xi and b choices for the largest yj , this
implies that
∫
(16.53) t(Ka,b , W ) = ab xa−1 y b−1 dy dx .
(x,y)∈SW

Let us approximate W by a threshold graphon V for which the boundary curve

∂SV is smooth (except for the endpoints of the intersection of SV with boundary
of the square), and ε = λ(SW △SV ) is very small. Let ds be the arc length of
the boundary curve ∂SV , and let n(x, y) = (n1 (x, y), n2 (x, y)) denote the outward
normal of ∂SV at point (x, y). By the Gauss–Ostrogradsky Theorem, we can rewrite
(16.53) as an integral along the boundary:
∫
t(Ka,b , V ) = b xa y b−1 n1 (x, y) ds .
∂SV

Interchanging the roles of x and y, and adding, we get

∫
( ) 1 1
(16.54) xa y b n1 (x, y) + n2 (x, y) ds = t(Ka+1,b , V ) + t(Ka,b+1 , V ).
a+1 b+1
∂SV

Consider the following integral:

∫
( )
I(V ) = x(1 − x)y(1 − y) p(x, y)2 n1 (x, y) + n2 (x, y) ds.
∂SV

By (16.54), this can be expressed as a linear combination

∑
2 deg(p)+4
I(V ) = cab t(Ka,b , V ),
a,b=1
 16.7. WHICH GRAPHS ARE EXTREMAL? 311

where the coefficients depend only on a, b and p. Furthermore, we have

∑
2 deg(p)+4
∑
2 deg(p)+4
I(V ) = cab t(Ka,b , V ) ≈ cab t(Ka,b , W )
a,b=1 a,b=1

∑
2 deg(p)+4
= cab t(Ka,b , U ) = I(U ) = 0,
a,b=1

where the error at the ≈ sign is bounded by Cε for some C that depends only on
a, b and p.
On the other hand, the integrand in I(V ) is nonnegative everywhere. Letting
ε → 0, we see that the x(1 − x)y(1 − y) p(x, y)2 = 0 on ∂(SW ). Since p is monotone
decreasing, this implies that ∂(SW ) = ∂(SU ), and hence W = U except perhaps on
the boundary (which is of measure 0). 

Complement reducible graphs. We have seen that Complement reducible

graphs form a variety (Example 14.43). Many of them are finitely forcible, and
they are of a rather different nature, sort of fractal-like. Here is a specific example.
Proposition 16.50. For x, y ∈ [0, 1], let U (x, y) = 1, if the first bit where the
binary expansions of x and y differ is at an odd position, and let U (x, y) = 0
otherwise. Then U is finitely forcible. 
Our examples of finitely forcible graphons have had finite range (the function
W assumed only a finite set of values). We refer to the paper of Lovász and Szegedy
[2011] for the proof of Proposition 16.50 and for further constructions of finitely
forcible graphons, where the range of the function W contains an interval.

16.7.3. Not too many finitely forcible graphons. Are there any graphons
that are not finitely forcible? A natural extension of the class of stepfunctions is
the class of kernels with finite rank. However, we don’t get any new finitely forcible
graphons in this class. In fact, Theorem 14.48 implies:
Corollary 16.51. Every finitely forcible kernel with finite rank is a stepfunction.
In view of Proposition 16.49, the following further corollary may be surprising:
Corollary 16.52. Assume that W ∈ W0 can be expressed as a non-constant poly-
nomial in x and y. Then W is not finitely forcible.
We want to derive more general necessary conditions for being finitely forcible.
We start with a rather strong property of finitely forcible functions. Recalling the
definition of the 2-labeled graph F ‡ from Section 16.2, let L(W ) be the linear space
generated by all 2-variable functions txy (F ‡ , W ), where F ranges over all simple
graphs.
Lemma 16.53. Suppose that W ∈ W is forced (in W) by the simple graphs
F1 , . . . , Fm . Then either the functions txy (F1‡ , W ), . . . , txy (Fm
‡
, W ) are linearly de-
pendent, or they generate L(W ) (or both).
Proof. Suppose not, then there is a simple graph Fm+1 such that the func-
tions txy (Fi‡ , W ) (i = 1, . . . , m + 1) are linearly independent. For U ∈ W, set
312 16. EXTREMAL THEORY OF DENSE GRAPHS

hk (U ; x, y) = txy (Fk‡ , U ). So hk (U ) ∈ W, and hk : W → W is a (non-linear) opera-

tor. Let Φ(U ) ∈ W denote the component of hm+1 (U ) orthogonal to the subspace
of W generated by h1 (U ), . . . , hm (U ).
We need the following technical claim.
Claim 16.54. There is an open L∞ -ball U in W centered at the function W such
that Φ : W → W is Lipschitz (in the L∞ norm) on U.
The operators hk are L∞ -Lipschitz on W. This follows from the inequality
∥txy (F, U ) − txy (F, U ′ )∥∞ ≤ e(F )∥U − U ′ ∥∞ ,
which is a simple exercise. Define the (numerical) matrix
⟨ ⟩ ⟨ ⟩
h1 (U ), h1 (U ) ... h1 (U ), hm (U )
 .. .. 
B(U ) =  . . 
⟨ ⟩ ⟨ ⟩
hm (U ), h1 (U ) . . . hm (U ), hm (U )
and the matrices Bi (U ) obtained from B by replacing the i-th column by
⟨h1 (U ), hm+1 (U )⟩, . . . , ⟨hm (U ), hm+1 (U )⟩. By elementary linear algebra, we have
( )
∑m
det Bi (U )
Φ(U ) = hm+1 (U ) + ( ) hi (U ).
i=1
det B(U )
The denominator is bounded away from 0 in a neighborhood U of W , and all the
other functions are Lipschitz in this neighborhood, proving that Φ is Lipschitz.
This completes the proof of the Claim.
By classical results on differential equations in Banach spaces (see e.g. Zeidler
[1985]), there exists a b > 0 and a differentiable family {Us : s ∈ [−b, b]} of
functions in U satisfying the differential equation
U̇s = Φ(Us ), U0 = W.
By (16.7), we have
d
t(Fi , Us ) = ⟨Φ(Us ), t2 (Fi‡ , Us )⟩ = 0
ds
for i = 1, . . . , m, and hence t(Fi , Us ) = t(Fi , U0 ) = t(Fi , W ) for all s ∈ [0, c]. Since
the graphs Fi force W , it follows that the Us are weakly isomorphic to W , and
so t(Fm+1 , Us ) = t(Fm+1 , W ). But then ⟨Φ(W ), t2 (F ‡ , W )⟩ = ds d
t(F, Us )s=0 = 0,
‡
and so ⟨Φ(W ), Φ(W )⟩ = ⟨Φ(W ), t(Fm+1 , W )⟩ = 0, which is a contradiction, since
Φ(W ) ̸= 0. 
A corollary of the previous theorem is that every finitely forcible kernel satisfies
a “nontrivial” relation of the form txy (f, W ) = 0. To specify what we mean by
“nontrivial”, let us say that a connected component of a partially labeled graph F
is a floating component, if it contains no labeled nodes.
Corollary 16.55. If the kernel W ∈ W is finitely forcible, then there is a nonzero
simple 2-labeled quantum graph g with nonadjacent labeled nodes and no floating
components such that txy (g, W ) = 0 almost everywhere.
Proof. The linear dependence of the functions txy (F ‡ , W ) gives a simple 2-
∑
labeled quantum graph with nonadjacent labeled nodes of the form f = i αi Fi‡
that satisfies txy (f, W ) = 0 almost everywhere. To see that f ̸= 0, it suffices to
 16.7. WHICH GRAPHS ARE EXTREMAL? 313

note that from any constituent of Fi‡ we can recover Fi by connecting its labeled
nodes, so no cancellation will occur. 
As an application of Corollary 16.55, we prove:
Theorem 16.56. The set of finitely forcible graphons is of first category in
f0 , δ ).
(W
Proof. For a fixed set {F1 , . . . , Fk } of simple 2-labeled graphs with no floating
components, let T (F1 , . . . , Fk ) denote the set of graphons W for which there is a
∑k
nonzero quantum graph of the form f = i=1 ai Fi satisfying txy (f, W ) = 0 for
all x, y ∈ [0, 1]. Corollary 16.55 implies that every finitely forcible graphon belongs
to one of the sets T (F1 , . . . , Fk ), so it suffices to prove that these sets are nowhere
dense. We do so in two steps.
Claim 16.57. Let F1 , . . . , Fk be simple 2-labeled graphs with no floating compo-
nents, and let W ∈ W0 . Then every neighborhood of W contains a graphon W ′
such that txy (F1 , W ′ ), . . . , txy (Fk , W ′ ) are linearly independent.
It is easy to see that there is a simple 2-labeled graph G such that
[[GF1 ]], . . . , [[GFk ]] are mutually non-isomorphic (a large complete graph Kn•• , with
an edge incident with the node labeled 1 removed, suffices). Proposition 5.44 im-
( )k
plies that there are graphons U1 , . . . , Uk such that the matrix t([[GFi ]], Uj ) i,j=1 is
nonsingular. For 0 < ε < 1/k, define
W ε = (1 − kε)W ⊕ (ε)U1 ⊕ · · · ⊕ (ε)Uk
(so the components of W ε are W, U1 , . . . , Uk , scaled by 1 − kε, ε, . . . , ε).
First we show that W ε → W in (W f0 , δ ) if ε → 0. Indeed, for every connected
simple graph F , we have
( )
t(F, W ε ) = (1 − kε)v(F ) t(F, W ) + εv(F ) t(F, U1 ) + · · · + t(F, Uk ) ,
and hence t(F, W ε ) → t(F, W ) as ε → 0.
Next, we show that txy (F1 , W ε ), . . . , txy (Fk , W ε ) are linearly independent for
all ε > 0. If not, then there are real numbers ai such that
∑
k
ai txy (Fi , W ε ) = 0
i=1

for all x, y ∈ [0, 1]. Suppose 1 − kε + (j − 1)ε ≤ x, y ≤ 1 − kε + jε, then every choice
of the variables for which one of the unlabeled nodes has value outside the interval
[1 − kε + (j − 1)ε, 1 − kε + jε] contributes 0 to txy (Fi , W ε ). Hence
∑
k
ai εv(Fi ) txy (Fi , Uj ) = 0 (j = 1, . . . , k)
i=1

for all x, y ∈ [0, 1]. Multiplying by txy (G, Uj ) and integrating, we get
∑
k
ai εv(Fi ) t([[GFi ]], Uj ) = 0 (j = 1, . . . , k).
i=1
( )
But this contradicts the nonsingularity of the matrix t([[GFi ]], Uj ) , and proves
the claim.
314 16. EXTREMAL THEORY OF DENSE GRAPHS

Claim 16.58. If F1 , . . . , Fk are simple 2-labeled graphs with no floating compo-

f0 , δ ).
nents, then T (F1 , . . . , Fk ) is nowhere dense in (W

Indeed, Claim 16.57 implies that every nonempty open set in (W f0 , δ ) contains
a graphon U such that txy (F1 , U ), . . . , txy (Fk , U ) are linearly independent. Then
 k
their Gram determinant t([[Fi Fj ]], U )i,j=1 is positive. But this determinant is a
continuous in U , and so there is a neighborhood of U in which it does not vanish, and
hence txy (F1 , U ′ ), . . . , txy (Fk , U ′ ) are linearly independent in this neighborhood.
This proves the claim and thereby the theorem. 

16.7.4. Infinitesimally finitely forcible graphons. Lemma 16.53 sug-

gests the following notion. We say that W is infinitesimally finitely forcible if
L(W ) has finite dimension. To explain the name, suppose that the functions
txy (F1‡ , W ), . . . , txy (Fk‡ , W ) generate L(W ). Informally, this means that every infin-
itesimal change in W that preserves t(F1 , W ), . . . , t(Fk , W ), also preserves t(F, W )
for every F .
The following observation, contrasted with Corollary 16.51, shows that infini-
tesimal finite forcibility and finite forcibility behave quite differently:
Lemma 16.59. Every infinitesimally finitely forcible kernel has finite rank.
Proof. If L(W ) has finite dimension d, then consider the functions
txy (Ck‡ , W ) = ktxy (Pk•• , W ) = kW ◦k ∈ L(W ), k = 3, . . . , d + 3.
These are linearly dependent, and so W satisfies a polynomial equation as an oper-
ator. This means that it has a finite number of different nonzero eigenvalues. Since
every nonzero eigenvalue has finite multiplicity, W has finite rank. 

The following corollary shows that the (false) conjecture mentioned above that
only stepfunctions are finitely forcible is true in a weaker sense.
Corollary 16.60. Graphons that are both finitely forcible and infinitesimally
finitely forcible are exactly the stepfunctions.
This corollary implies that our examples of finitely forcible non-step-functions
(e.g., the simple threshold graphon) are finitely forcible but not infinitesimally
finitely forcible. We don’t know any examples for the converse.

Proof. If a graphon is both finitely forcible and infinitesimally finitely forcible,

then it has finite rank by Lemma 16.59, and so it is a stepfunction by Corollary
16.51.
Conversely, we know by Theorem 16.46 that every stepfunction W is finitely
forcible. Since every function txy (F ‡ , W ) is itself a stepfunction with the same
steps, it follows that L(W ) is finite dimensional, so W is infinitesimally finitely
forcible. 

Summarizing Lemma 16.53 and Corollary 16.60, we get the following.

Corollary 16.61. Suppose that W ∈ W is forced (in W) by the simple graphs
F1 , . . . , Fm . Then either W is a stepfunction, or the functions txy (Fi‡ , W ) (i =
1, . . . , m) are linearly dependent. 
 16.7. WHICH GRAPHS ARE EXTREMAL? 315

Remark 16.62. 1. All the examples of finitely forcible graphons discussed above,
indeed all the examples we know of, have dimension at most 1 (in the sense of
the topology of the graphon discussed in Section 13.4). Most likely this is just
due to the lack of more involved constructions; but it is not too far fetched to
ask: Does every finitely forcible graphon have finite dimension? Together with
Proposition 13.34 this would imply that finitely forcible graphons have polynomial-
size weak regularity partitions. Together with Conjecture 16.45 and the properties
of finitely forcible graphons proved above, this would provide nontrivial “templates”
for extremal graphs, and possibly provide some help in finding the extremal graphs
for specific extremal graph problems by imposing limitations on them.
2. We have seen a number of “finiteness” conditions on a graphon W : (a) W is a
stepfunction; (b) W has finite rank; (c) W is finitely forcible; (d) W is infinitesimally
finitely forcible; (e) the graph parameter t(., W ) has finite connection rank, or
equivalently, the corresponding gluing algebras Qk /W have finite dimension; (f)
the spaces (J, rW ) and/or (J, rW ) are finite dimensional. We could add further
such conditions, like (g) the algebras Qk /W are finitely generated (this is true
not only for stepfunctions, but also for a simple threshold function, for example).
Several implications between these finiteness properties have been proved in this
book, but several others are only conjectured.

Exercise 16.63. Prove that the simple threshold graphon (Example 11.36) is
forced by the conditions (16.51) and t(P3 , W ) − t(K2 , W ) + 1/6 = 0.
Exercise 16.64. Show that for every kernel W there is a nonzero simple 2-
labeled quantum graph g with nonadjacent labeled nodes (which may have floating
components) such that txy (g, W ) = 0.
Exercise 16.65. Which implications among the finiteness conditions (a)-(g) in
Remark 16.62 are proved in this book? Which others are trivial/easy/possible?
 CHAPTER 17

Multigraphs and decorated graphs

Limit objects can be defined for multigraphs, directed graphs, colored graphs,
hypergraphs etc. In many cases, like directed graphs without parallel edges, or
graphs with nodes colored with a fixed number of colors, this can be done along
the same lines as for simple graphs.
Turning to multigraphs, even the definition of homomorphisms is not unique,
as we have discussed in Chapter 5. In one version, a homomorphism F → G is a
map V (F ) → V (G) where the image of any edge has at least as large multiplicity
as the edge itself (node-homomorphism); in another version, to specify a homomor-
phism between multigraphs, we have to tell the image of every node as well as the
image of every edge (node-and-edge homomorphism). We also mentioned homo-
morphisms that preserve edge-multiplicities (induced homomorphisms). But this is
not the main complication. To illuminate the content of this chapter, let us discuss
informally convergence of multigraphs. We get to the most general question in sev-
eral steps. We want to define convergence of a multigraph sequence (G1 , G2 , . . . )
in terms of the convergence of the homomorphism densities t(F, Gn ) for every F ,
and want to construct a limit object that appropriately reflects the limiting values.
(1) In the previous chapters, this program was carried out in detail (maybe
even in more detail than you wished to see) in the case when the graphs Gn as well
as the graphs F were simple.
(2) Suppose that the graphs Gn are multigraphs, but we care about the densities
of simple graphs F only. In this case, node-homomorphisms mean nothing new,
but node-and-edge homomorphisms do. Let us assume for the time being that the
edge multiplicities in the graphs Gn remain uniformly bounded by a fixed constant
d. This case is quite easy, and it has been settled (even in greater generality)
by Borgs, Chayes, Lovász, Sós and Vesztergombi [2008]: the limit object can be
described by a kernel with values in [0, d], and the proofs are rather straightforward
generalizations of the proofs from case (1).
(3) Let the graphs Gn be multigraphs with bounded edge multiplicities as
before, but we want the limit object to correctly reflect densities of multigraphs
F . This case is more interesting. It turns out that whether we consider node-
homomorphisms or node-and-edge homomorphisms does not matter much (this is
not obvious at the first sight). Nor do the numerical values of the edge multiplicities:
we can think of them just as decorations of the edges from the set K = {0, 1, . . . , d},
and the only relevant property of this set is that it is finite. Here comes the first
surprise: the limit object can again be defined as a function on [0, 1]2 , but its values
are not numbers, but probability distributions on K (in other words, d-tuples of
numbers). The second surprise is that one can generalize the results to decorations
from a set K that is any compact Hausdorff space. Once the right statement of
317
318 17. MULTIGRAPHS AND DECORATED GRAPHS

the results is found, the proofs can be obtained by essentially the same techniques
as before. These results of B. Szegedy and the author will be discussed in Section
17.1.
(4) Let us backtrack and generalize in another direction: we allow unlimited
edge multiplicities for the graphs Gn , but are only interested in densities of simple
graphs F . The limit object is, not too surprisingly, an unbounded kernel. But the
treatment becomes more technical; one needs appropriate bounds on the growth
of edge multiplicities, and even then, one has to modify the definition of the cut
norm and strengthen the Regularity Lemma to get the proofs. Some preliminary
results of L. Szakács and the author [unpublished] are described in the internet
notes [Notes].
(5) Finally, if we have sequences of graphs with unbounded edge-multiplicities
and we want a limit object that correctly reflects densities of multigraphs, then
we have to combine the ideas of questions (3) and (4). Here the cases of
node-homomorphism densities and node-and-edge homomorphism densities diverge:
there will be graph sequences that are convergent in the node-homomorphism sense
but not in the node-and-edge homomorphism sense. Kolossváry and Ráth [2011]
showed how to assign limit objects if we work with node-homomorphisms; these
results can also be derived from the results mentioned in point (3) above, by com-
pactifying the set of integers. The limit object is a function defined on [0, 1]2 ,
whose values are probability distributions on N. One expects that under appro-
priate bounds on the growth of edge multiplicities, these limit objects will also be
valid for the node-and-edge homomorphism densities. However, as far as I know,
no details have been worked out here.

17.1. Compact decorated graphs

17.1.1. Sampling and homomorphism numbers. We often encounter
graphs with a special decoration: most often we color nodes or edges with a fi-
nite number of colors, but in some cases the objects used for decoration are more
complicated, like kernels in W. Multiplicities of edges can be thought of as deco-
rations from N, and nodeweights and edgeweights can be thought of as decorations
from R. In this section we sketch how to extend the results about convergence and
limit of simple graphs to the more general setting when we decorate every edge ij
G
of a simple graph G by an element βij of an arbitrary, but fixed compact Hausdorff
space K. Most of this is based on the work of Lovász and Szegedy [2012b].
It will be convenient to assume that K contains a special element called 0,
where an edge decorated with 0 means that it is missing (one can always add an
element to K to play this role). This way we may assume that the underlying
simple graph is Kn◦ , the complete graph on [n] with a loop edge on every node. We
denote by Fn (K) the set of all K-decorated graphs on [n], and by F(K), the set of
all K-decorated graphs.
If K is finite, then so is Fn (K). If K is endowed with a topology, then Fn (K) is
a topological space, endowed with the product topology of a finite number of copies
of K. Compactness of K implies that Fn (K) is compact. We can identify graphs
in Fn (K) that are isomorphic (in the obvious sense: the nodes can be permuted
so that we get the same decoration for every edge); we get a topological space
Fn (K)/Sn , which is also compact.
 17.1. COMPACT DECORATED GRAPHS 319

To define subgraph sampling for K-decorated graphs is straightforward: For

G ∈ F(K) and k ∈ [v(G)], let G(k, G) denote the K-decorated graph obtained by
selecting a random ordered subset (v1 , v2 , . . . , vk ) of V (G) uniformly, and decorating
the edge ij of Kn◦ by βvGi ,vj . While G(k, G) comes with labeled nodes, it is clear that
this graph with any other labeling of its nodes arises with the same probability.
Here comes the first (little) surprise: to define homomorphism numbers and
homomorphism densities for K-decorated graphs is not straightforward. There is
no natural way to define hom(F, G) for two K-decorated graphs F and G. What we
can do is the following. Let C denote the space of continuous real valued functions
on K. For every map φ : V (F ) → V (G), where F is a C-decorated graph and G is
a K-decorated graph, we define a real value
∏
F G
homφ (F, G) = βij (βφ(i)φ(j) ).
1≤i<j≤k

F G F G
(Recall that βijis a function on K while βφ(i)φ(j) is an element of K, so βij (βφ(i)φ(j) )
is well defined.) The homomorphism number hom(F, G) is defined, as earlier, by
∑
hom(F, G) = homφ (F, G),
φ: V (F )→V (G)

and we define inj(F, G), as before, by summing over injective maps. We also define
the homomorphism density by
hom(F, G)
t(F, G) = .
v(G)v(F )
These subgraph densities have some new features relative to those used so far.
First of all, they are not necessarily in [0, 1]. Second, while for simple graphs
homomorphism numbers and sample distributions were easily expressed in terms
of each other (recall Proposition 5.5), in this more general setting the situation
is different. For a fixed (large) K-decorated graph G, sampling from G assigns
probabilities to K-decorated graphs, while the homomorphism numbers into G
assign real numbers to C-decorated graphs.
We are going to characterize convergence of a graph sequence in terms of homo-
morphism numbers from C-decorated graphs. This seems to be quite wasteful, since
in the case of simple graphs, only a countable number of convergence conditions
had to be assumed. But we can restrict ourselves to graphs decorated by elements
from an appropriate subset of C. We say that a set B ⊆ C is a generating system if
the linear space generated by the elements of B is dense in C in the L∞ norm. If
C is finite dimensional, then it is the most economical to choose a basis of C for B.
It turns out that the choice of the family B has combinatorial significance, as the
following examples show.
Example 17.1 (Simple graphs). Let K be the discrete space with two elements
called “edge” and “non-edge”(or shortly) 1 and 0. The set C consists of all maps
{0, 1} → R, i.e., of all pairs f (0), f (1) of real numbers. A natural generating
subset (in fact, a basis) in C consists of the pairs f0 = (1, 1) and f1 = (0, 1).
Sampling, convergence, and homomorphism densities correspond to these notions
introduced for simple graphs.
One may, however, take another basis in C, namely the pair g0 = (0, 1) and
g1 = (1, 0). Then again B-decorated graphs can be thought of as simple graphs,
320 17. MULTIGRAPHS AND DECORATED GRAPHS

and hom(F, G) counts the number of maps that preserve both adjacency and non-
adjacency. 
Example 17.2 (Colored graphs). Let K be a finite set of “colors” with the
discrete topology. Continuous functions on K can be thought of as vectors in
RK . The standard basis B in this space corresponds to elements of K, and so
B-decorated graphs are just the same as K-decorated graphs. The homomorphism
density t(F, G) is the probability that a random map V (F ) → V (G) preserves edge
colors. 
Example 17.3 (Multigraphs). Let G be a multigraph with edge multiplicities at
most d. Then G can be thought of as a K-decorated graph, where K = {0, 1, . . . , d}.
However, there are several meaningful ways of picking a basis in C, giving rise
different notions of homomorphisms.
• Taking the standard basis in C = RK means that we think of the edge multi-
plicities just as different labels (colors). The graph F will be decorated with edge
multiplicities too, and then a homomorphism must preserve edge multiplicities.
This is equivalent to Example 17.2; it can be thought of as the induced version of
homomorphisms between multigraphs.
• Take the basis B = {(1, 0, 0, . . . , 0), (1, 1, 0, . . . , 0), . . . , (1, 1, 1, . . . , 1)} in C. Again,
we can think of a B-decorated graph as a multigraph with edge multiplicities at
most d. Then a map φ : V (F ) → V (G) counts as a homomorphism if and only
if the multiplicity of each target edge φ(i)φ(j) ∈ E(G) is at least as large as the
multiplicity of ij ∈ E(F ); in other words, it counts node-homomorphisms.
• Take the functions B = {1, x, . . . , xd } in C. Again, we can think of a B-decorated
graph as a multigraph with edge multiplicities at most d, where an edge decorated
by xi is represented by i parallel edges. With this choice, hom(F, G) is the number
of node-and-edge homomorphisms of F into G as multigraphs. 
Example 17.4 (Weighted graphs). Let K ⊆ R be a bounded closed interval.
Let B be the collection of functions x 7→ xj for j = 0, 1, 2, . . . on K; then B is a
generating system. It is natural to consider a B-decorated graph F as a multigraph,
and then hom(F, G) is just our usual homomorphism into a weighted graph. Note
that Example 17.3 for the third choice of the basis of C is a special case. 
Much of the theory of graph homomorphisms can be built up for compact
decorated graphs without much difficulty, but with some care. For example, the
relationships between hom and inj (equations (5.16) and (5.24)) can be extended
for any C-labeled graph F and K-labeled graph G:
∑
(17.1) hom(F, G) = inj(F/P, G),
P
and
∑
(17.2) inj(F, G) = µP hom(F/P, G),
P
where we have to re-define F/P for a partition P = V1 , . . . , Vq∏of V (F ): the nodes
F
are the partition classes, and an edge (Vi , Vj ) is decorated by u∈Vi , v∈Vj βuv .
(It may seem strange at the first sight that we decorate the edges of F/P by
the product of their inverse images, and not by the sum, say. Looking at Example
17.1 gives an explanation: a missing edge is decorated by the function (1, 1), an
 17.1. COMPACT DECORATED GRAPHS 321

edge present, by the function (0, 1), so the product is (0, 1) if and only if at least
one edge is present.)
From the results in Part 2, we state a generalization of one result (the first
statement of Theorem 5.29), which we will need later.

Theorem 17.5. Two K-decorated graphs G1 , G2 ∈ F(K) are isomorphic if and

only (if hom(F, G1)) = hom(F, G2 ) for every B-decorated graph F with v(F ) ≤
max v(G1 ), v(G2 ) .

Proof. The “only if” part is obvious. For the converse, suppose that
hom(F, G1 ) = hom(F, G2 ) for every B-decorated graph F . In particular, this holds
for F = K1 , which implies that v(G1 ) = v(G2 ) = n (say). Furthermore,
• hom(F, G1 ) = hom(F, G2 ) for every F with v(F ) ≤ n whose edges are deco-
rated by linear combinations of functions from B. Indeed, expanding the product in
the definition of the homomorphism number, we see that hom(F, G1 ) can be written
as a linear combination of values hom(F ′ , G1 ), where every F ′ is B-decorated. We
get a similar expression for hom(F, G2 ) in terms of the values hom(F ′ , G2 ). Since
hom(F ′ , G1 ) = hom(F ′ , G2 ) for all these graphs F ′ by hypothesis, it follows that
hom(F, G1 ) = hom(F, G2 ).
• hom(F, G1 ) = hom(F, G2 ) for every C-decorated F with v(F ) ≤ n. This
follows since linear combinations of functions in B are dense in C.
• inj(F, G1 ) = inj(F, G2 ) for every C-decorated F with v(F ) ≤ n. This follows
from (17.2).
Now let S be the set of all elements of K occurring as edge-decorations in G1 or
G2 . Let F = Kn◦ , and let us decorate the edges of F with functions fe ∈ C such that
the values fe (s) (e ∈ E(F ), s ∈ S) are algebraically independent transcendentals
(such functions clearly exist). In the equation inj(F, G1 ) = inj(F, G2 ), every term
is a product of these transcendentals, so for the equation to hold, we need that
every term on the left cancels a term on the right, and vice versa. But if the
term corresponding to an (injective) map φ : V (F ) → V (G1 ) cancels the term
corresponding to ψ : V (F ) → V (G2 ), then φ−1 ◦ ψ is an isomorphism between G1
and G2 . 

17.1.2. Convergence. The definition of convergence in terms of sampling is

rather straightforward, after we remark that the samples G(k, G) define a distribu-
tion on a compact space Fk (K) (and not on a finite space as before), and probability
distributions on a compact space behave nicely (see Appendix A.3.3). We say that
a sequence (G1 , G2 , . . . ) of K-decorated graphs is convergent, if v(Gn ) → ∞, and
for every k ≥ 1, the samples (G(k, Gn ) : n = 1, 2, . . . ) are weakly convergent in
distribution. In( other words,) for every continuous function f : Fk (K) → R, the
limit limn→∞ E f (G(k, Gn )) exists. We note that it is enough to require this for
continuous functions f : Fk (K)/Sk → R (where Sk is the symmetric group on [k]),
or in other words, for continuous functions f : Fk (K) → R that are invariant un-
der relabeling of the nodes. Indeed, the distribution of G(k, Gn ) is invariant under
relabeling, so if we define
1 ∑
f (G) = f (Gπ )
k!
π∈Sk
322 17. MULTIGRAPHS AND DECORATED GRAPHS

(where Gπ (denotes the graph

) (obtained by )permuting the labels in G according to
π), then E f (G(k, Gn )) = E f (G(k, Gn )) . If the values on the right converge,
then so do the values on the left.
The next theorem shows that our somewhat complicated definition of homo-
morphism functions is right for the notion of convergence.
Theorem 17.6 (Equivalence of convergence notions). Let (G1 , G2 , . . . ) be a
sequence of K-decorated graphs with v(Gn ) → ∞, and let B ⊆ C be a generating
system. Then the following are equivalent:
(a) (G1 , G2 , . . . ) is convergent;
(b) for every C-decorated graph F , the numerical sequence t(F, Gn ) is conver-
gent;
(c) for every B-decorated graph F , the numerical sequence t(F, Gn ) is conver-
gent.
An important consequence of this theorem is that if we define convergence of
the sequence (Gn ) in terms of the convergence of t(F, Gn ) for all B-decorated graph,
then the definition is independent of the choice of B. In particular, convergence
of a multigraph sequence is the same whether we use induced homomorphisms,
node-homomorphisms, or node-and-edge homomorphisms in the definition.
Proof. (a)⇒(b)⇒(c): From the identity
( ( ))
(17.3) tinj (F, Gn ) = E tinj F, G(v(F ), Gn )
we see that tinj (F, Gn ) is convergent. Just as for simple graphs, the convergence of
the sequence tinj (F, Gn ) is equivalent to convergence of t(F, Gn ). Trivially, (c) also
follows.
(c)⇒(a): Let f be any continuous function on Fk (K)/Sk . We claim that F
can be approximated uniformly in the L∞ norm by functions of the type hom(h, .),
where h is a B-decorated quantum graph. (The value hom(h, G) is invariant under
relabeling of the nodes of G, and hence hom(h, .) can be considered as a function
on Fk (K)/Sk .) We will use the Stone–Weierstrass Theorem. The set of functions
hom(h, .) is closed under linear combination and multiplication, and contains the
constant functions. Furthermore, it separates elements of Fk (K)/Sk , i.e., for two
non-isomorphic decorated graphs G1 , G2 ∈ Fk (K) there is a B-decorated quantum
graph h such that hom(h, G1 ) ̸= hom(h, G2 ). This is just the content of Theorem
17.5.
It follows that if (Gn ) is a sequence of decorated graphs that satisfies (c), then
for every continuous function Fk (K)/Sk → R, the sequence f (Gn ) is convergent.
Hence (Gn ) is convergent. 
17.1.3. Limit objects. Somewhat surprisingly, limit objects of K-decorated
graph sequences are not 2-variable functions with values in K, but 2-variable func-
tions with values that are probability measures on K. (This may be the precept of
this whole section!) Let us denote by W(K) the set of functions ω : [0, 1]2 → P(K)
that are measurable and symmetric (ω(x, y) = ω(y, x) for every (x, y) ∈ [0, 1]2 ). El-
ements of W(K) will be called K-graphons.
Note that these K-graphons are different from the random graphons introduced
in Section 14.5. In a random graphon W, the values W(x, y) are also random
variables, but for different points (x, y) they are correlated in specific ways (like
 17.1. COMPACT DECORATED GRAPHS 323

in Example 11.44, where we obtained a distribution on constant functions). On

the other hand, in a K-graphon we think of the probability distribution ω(x, y) as
independent for different x and y. Technically, this cannot be defined (we don’t
have uncountably many independent random variables), but this will be how we
treat them as soon as we can restrict our attention to a finite or countable set of
pairs (x, y).
The above definition of K-graphons through measure-valued functions is rather
abstract; we can give a more down-to-earth definition as follows. ∫ Let ω be a K-
graphon and let f ∈ C. Define Wf : [0, 1]2 7→ R by Wf (x, y) = K f dω(x, y) (recall
that ω(x, y) is a probability measure on the Borel sets of K for every x and y). For
every fixed f ∈ C, the function Wf is a kernel.
Conversely, if we specify a measurable function W : C × [0, 1]2 → R such that
Wf (x, y) is linear in f , Wf ∼
=1 (x, y) = 1 and Wf (x, y) > 0 for every function f > 0
for all (x, y) ∈ [0, 1], then there is a K-graphon represented by Wf (x, y). Indeed,
for every fixed (x, y) ∈ [0, 1]2 , the functional f 7→ Wf (x, y) is a linear functional on
C that is positive on positive functions. The Riesz representation ∫ theorem implies
that there is a probability measure ω(x, y) such that Wf (x, y) = K f dω(x, y), and
it is not hard to check that ω(x, y) defines a K-graphon.
It is enough to know the values Wf (x, y) for f ∈ B, where B is a generating
system; this determines the values Wf (x, y) for all f ∈ C, and through this, the
probability distributions ω(x, y). We call the system of functions (Wf : f ∈ B)
the B-moment representation of ω. (The name refers to the fact that for various
natural choices of K and B, the numbers t(F, Wf ) (F ∈ B) behave similarly to the
moments of a single-variable function. See Example 17.3).
The construction assigning a graphon to every simple graph extends in a
straightforward manner: every K-decorated graph G gives rise to a K-graphon
ωG as follows. Let V (G) = [n]. We split the unit interval into n intervals J1 , . . . , Jn
of length 1/n, and let ωG (x, y) = βij G
for x ∈ Ji , y ∈ Jj (here we identify the
element βij ∈ K with the probability distribution concentrated on βij
G G
).
It is also straightforward to extend homomorphism densities. For every K-
graphon ω and C-decorated graph F on V (F ) = [k], we introduce the homomor-
phism density t(F, ω) by
∫ ∏
t(F, ω) = Wβi,j
F (xi , xj ) dx1 . . . dxk .

1≤i<j≤k
[0,1]k

It is easy to see that for every K-decorated graph G and C-decorated graph F ,

t(F, ωG ) = t(F, G).

Note that if F is B-decorated for some generating system B ⊆ C, then t(F, ω) is

expressed in terms of the B-moment representation of ω.

Example 17.7. It may be worthwhile to revisit our examples from Section 17.1.1.
Simple graphs could be thought of as K-decorated graphs with K = {0, 1}.
Every probability distribution on K can be represented by a number between 0 and
1, which is the probability of being adjacent (i.e., the probability of the element 1 ∈
K). So a K-graphon is described by a symmetric measurable function W : [0, 1]2 7→
[0, 1], i.e., by a graphon.
324 17. MULTIGRAPHS AND DECORATED GRAPHS

For edge-colored graphs, K is a finite set of “colors”. Probability distributions

on K can be described by the probabilities of its points. So a K-graphon
∑ is repre-
sented by k = |K| measurable functions Wi : [0, 1]2 7→ [0, 1] with i Wi (x, y) = 1.
For multigraphs with bounded edge multiplicities, K-graphons have the same
description as above, only the “colors” correspond to “multiplicities” now. It is re-
markable that this notion (and the notion of convergent sequence) does not depend
on which definition of homomorphisms we adopt.
Finally, in the case of weighted graphs with edge weights from K = [0, 1] (for
example), a K-graphon has values that are probability distributions on [0, 1]. These
are uniquely determined by their moments, so we can describe a K-graphon as a
sequence of measurable functions Wn : [0, 1]2 7→ [0, 1] (n = 0, 1, . . . ) such that
(Wn (x, y) : n = 0, 1, . . . ) is the sequence of moments of a random variable with
values in [0, 1] for all x, y ∈ [0, 1]. 
The limit of a convergent sequence of K-decorated graphs can be represented
by a K-graphon:
Theorem 17.8. Let (G1 , G2 , . . . ) be a convergent sequence of K-decorated graphs.
Then there is a K-graphon ω such that t(F, Gn ) → t(F, ω) as n → ∞ for every
C-decorated graph F .
In Chapter 11, we based the proof of the analogous theorem (Theorem 11.21)
on the compactness of the graphon space. One can extend that proof; the difficulty
is that it is awkward to define the cut distance and prove its basic properties. One
can bypass the cut distance and work directly with homomorphism densities. For
the details of the proof, we refer to the paper of Lovász and Szegedy [2012b].
It is worth pointing out that while we can select different generating sets in
C, and we get seemingly very different sequences of homomorphism densities, the
limit object is independent of which generating system we use.
We do get compactness of the K-graphon space, without the metrization given
by the cut distance, by an easy observation. Just as for simple graphs and graphons,
not only sequences of graphs but also sequences of K-graphons have convergent
subsequences, which can be proved along the same lines:
Theorem 17.9. Let B be a countable generating set and let W1 , W2 , . . . be a
sequence of K-graphons. Then we can select an infinite subsequence for which
(t(F, Wn )) is a convergent sequence for every B-decorated graph F . 
17.1.4. An old debt. It is time to sketch the proof of Theorem 5.61 character-
izing multiplicative and reflection positive graph parameters with finite connection
rank r(f, 2). (We do use a fair bit of the material in previous chapters.)

Proof. The necessity of the conditions is easy, along the lines of the proofs of
Propositions 5.64 and 7.1.
The sufficiency takes more work, but the proof can be put together from
arguments that we have established before. Let f be a multiplicative, reflec-
tion positive multigraph parameter for which r(f, 2) is finite. Replacing f by
f (G)/(f (K1 )v(G) f (K2 )e(G) ), we may assume that f (K1 ) = f (K2 ) = 1. We es-
tablish several representations of f .
1. First, we represent f as an expectation of homomorphism-like quantities.
For every n ≥ 1, there is a distribution on [−1, 1]-weighted graphs on [n] such
 17.2. MULTIGRAPHS WITH UNBOUNDED EDGE MULTIPLICITIES 325

( )
that if Hn is chosen from this distribution, then f (G) = E inj(G, Hn ) for every
multigraph G on [n]. This can be proved using Proposition A.24 in the Appendix.
2. If the weighted graph Zn on [n] is chosen from the distribution in Step 1,
then f (G) = limn→∞ t(G, Zn2 ) with probability 1 for every multigraph G. This can
be proved by modifying the proof of Lemma 11.8 appropriately (in fact, it suffices
to compute the second moments only).
3. There is a [−1, 1]-graphon W (in the sense of Section 17.1.3) such that
f = t(., W ). This can be deduced from Theorem 17.8.
4. If W = (W0 , W1 , . . . ) be the moment sequence representation of W , then
every Wn is a stepfunction with the same steps. This is proved using the assumption
that r(f, 2) is finite in a way similar to the proof of Theorem 13.47.
Knowing this, it is easy to construct the randomly weighted graph H. Its nodes
correspond to the steps S1 , . . . , Sq of the Wn , and the measures of the steps also
give their nodeweights. The random variable W (x, y) (x ∈ Si , y ∈ Sj ) gives the
decoration of the edge ij. It is not hard to check that the construction gives the
right graph parameter. 

17.2. Multigraphs with unbounded edge multiplicities

An obvious drawback of our compactly decorated graph model is that it does
not include multigraphs with unbounded multiplicities. We can think of these as
N-decorated graphs, but this is not a compact space in the discrete topology. One
way out is to compactify this space; the simplest compactification consists of adding
a single new element ∞, whose neighborhoods are complements of finite subsets of
N. Let N denote this compactification of N.
Continuous functions N → R are convergent sequences of real numbers, where
the value on ∞ is the limit of the sequence. Let C = C(N) be the set of these func-
tions. Similarly as in Example 17.3 above, we can choose B0 = {f0 , f1 , . . . , f∞ },
where fi = 1{i} for finite i, and f∞ ≡ 1; or we can choose B1 = {g0 , g1 , . . . , g∞ },
where gi = 1{1,...,i} for finite i, and g∞ ≡ 1. In the first case, hom(F, G) counts in-
duced homomorphisms, while in the second, it counts node-homomorphisms. How-
ever, taking the functions xi for x > 1 is not allowed here, since these values don’t
form a convergent sequence. So counting node-and-edge homomorphisms between
multigraphs with unbounded edge multiplicities does not fit in this model.
We know that convergence of a sequence does not depend on the choice of
the generating system, so we can characterize it either through induced homomor-
phisms or node-homomorphisms. It is easy to see that sequence of multigraphs
(Gn : n = 1, 2, . . . ) is convergent if and only if for every k ≥ 0, truncating the
edge multiplicities at k gives a convergent sequence (now graphs with bounded edge
multiplicities).
This is a reasonable definition, but if we want to describe convergence of node-
and-edge homomorphism densities, or convergence of a sequence of weighted graphs
with unbounded edge-weights, then we have to work more, as the following example
shows.
Example 17.10. Let Gn be the multigraph on [n] where the multiplicity of the
edge connecting 1 to 2k is 4k for 1 ≤ k ≤ log n, and all other edge multiplicities are 1.
This graph sequence is convergent in the compactification sense. However, the edge
326 17. MULTIGRAPHS AND DECORATED GRAPHS

densities t(K2 , Gn ) do not form a convergent sequence: we have t(K2 , Gn ) ∼ 11/3

if n is a power of 2, but t(K2 , Gn ) ∼ 5/3 if n is one less. 
Lovász and Szakács [unpublished] gave a different definition of convergence
for multigraph sequences with unbounded edge multiplicities and constructed an
appropriate limit object such that the densities of simple graphs converge to a limit
for every convergent multigraph sequence; see Lovász [Notes].
 Part 4

Limits of bounded degree graphs

 CHAPTER 18

Graphings

This next part of the book treats convergence and limit objects of bounded
degree graphs. We fix a positive integer D, and consider graphs with all degrees
bounded by D. Unless explicitly said otherwise, this degree bound will be tacitly
assumed.
In this chapter we introduce infinite graphs that generalize finite bounded de-
gree graphs. Their main role will be to serve as limit objects for sequences of
bounded degree graphs, analogous to the role of graphons in the previous part.
Graphons (symmetric functions in two variables) are very common objects and of
course they have been studied for many reasons since the dawn of analysis. Graph-
ings are less common; however, they are interesting on their own right, and in fact,
they too have been studied in other contexts, mainly in connection with group
theory.
The situation will be more complex than in the dense case, and there will
be no single “true” limit object. But the connection between these objects is
quite interesting. As a further warning, it is not known whether the objects to be
discussed in this chapter are all limit objects of sequences of finite graphs. This
makes it even more justified to treat them separately from convergent finite graph
sequences.
In this part we will consider finite graphs, countably infinite graphs, and even
larger graphs (typically of continuum cardinality). To keep notation in check, we
will denote finite graphs by F, F ′ , G, G′ , G1 . . . , and families of finite graphs by
calligraphic letters. In particular, we denote by G the family of all finite graphs
(with all degrees bounded by D). We denote countable graphs by H, H ′ , H1 , . . . ,
and their families by Gothic letters. In particular, G denotes the family of all
countable graphs (with all degrees bounded by D). Graphs with larger cardinality
will be denoted by boldface letters like G, G′ , . . . ; we will not talk about families
of them.

18.1. Borel graphs

We start with a quite general notion. Let (Ω, B) be a Borel sigma-algebra;
then (Ω, B) is separating and generated by a countable family J = {J1 , J2 , . . . } of
subsets of Ω. It will be convenient to assume that the generator set J is closed
under complementation and finite intersections, which implies that it is a Boolean
algebra (not a sigma-algebra!). We call the sets in B and also in B × B etc. Borel
sets. (The reader who likes more concrete structures can think of this as the sigma-
algebra of Borel sets in [0, 1], with J consisting of finite unions of open intervals
with rational endpoints.)
Let G be a graph with node set V (G) = Ω. We call G a Borel graph, if its
edge-set is a Borel set in B × B. For example, the complete graph on Ω is Borel,
329
330 18. GRAPHINGS

but we will be interested in graphs with all degrees bounded by D, and will tacitly
assume this condition for these infinite graphs too.
Example 18.1. For a fixed a ∈ (0, 1), we define a graph Pa on [0, 1] by connecting
two points x and y if |x − y| = a. This defines a Borel graph. The graph structure
of this Borel graph is quite simple: it is the union of finite paths. If a > 1/2, then
it is just a matching together with isolated nodes. Of course, the Borel structure
adds additional structure.
We can make the example more interesting, if we wrap the interval [0, 1] around,
and consider the graph Ca on [0, 1) in which a node x is connected to x+a (mod 1)
and x−a (mod 1). If a is irrational, we get a graph that consists of two-way infinite
paths; if a is rational the graph will consist of cycles. 
The following lemma is very useful and it also motivates some of the definitions
in the sequel.
Lemma 18.2. A graph G on a Borel space (Ω, B) is a Borel graph if and only if
for every Borel set B ∈ B, the neighborhood NG (B) is Borel.
Proof. Suppose that G is a Borel graph, and let B ∈ B. Then B ′ = E(G) ∩
(B × Ω) is a Borel set. Furthermore, if we project B ′ to the second coordinate, then
the inverse image of any point is finite. A classical theorem of Lusin [1930] implies
that the projection is also Borel; but this projection is just NG (B).
Conversely, assume that G has the property that the neighborhood of any Borel
set is also Borel. Let Pi (i = 1, 2, . . . ) range over all partitions of Ω into a finite
number of sets in J . We claim that
∩ ∪ ( )
(18.1) E(G) = J × NG (J) ;
i J∈Pi

this will prove that G is Borel.

First, let (x, y) ∈ E(G). If x ∈ J ∈ Pi , then y ∈ NG (J) and so (x, y) ∈
J × NG (J). Hence it follows that (x, y) is contained in the right hand side of
(18.1).
Second, let (x, y) be a pair contained in the right hand side of (18.1), then for
every i ≥ 1 there is a set J ∈ Pi for which (x, y) ∈ J × NG (J). Hence y ∈ NG (J),
and so there is a point zi ∈ J such that y ∈ NG (zi ), which means that zi ∈ NG (y).
But NG (y) is a finite set, and so this can hold for all sets J ∈ Pi , J ∋ x only if
x ∈ NG (y). This shows that (x, y) ∈ E(G). 
There is a rather rich theory of Borel graphs (see e.g. Kechris and Miller [2004]).
We state and prove only a few results that we need. The following theorem, which
extends Brooks’ Theorem from finite graphs to Borel graphs, was proved by Kechris,
Solecki and Todorcevic [1999]. We say that a coloring of the nodes of a Borel graph
is a Borel coloring, if nodes with any given color form a Borel set.
Theorem 18.3. Every Borel graph has a Borel coloring with D + 1 colors.
Proof. We start with constructing a countable Borel coloring. Consider the
countable Boolean algebra J = {J1 , J2 , . . . } generating Borel sets. For every node
v ∈ Ω there is a set Ji such that N (v) ⊆ Ji but v ∈ / Ji . Let ψ(v) denote the
smallest index i for which Ji has this property. Then trivially the sets ψ −1 (i)
(i = 1, 2, . . . ) are disjoint. Two adjacent nodes u and v cannot have ψ(u) = ψ(v),
 18.1. BOREL GRAPHS 331

so the sets ψ −1 (i) (i = 1, 2, . . . ) consist of nonadjacent points. It is easy to check

(using Lemma 18.2) that these sets are Borel sets.
Next, for each i = 1, 2, . . . , we recolor each node in ψ −1 (i) with the least color
that does not occur among its neighbors. We do this in order, so the nodes in
ψ −1 (2) get recolored before the nodes in ψ −1 (3) etc. Using Lemma 18.2, it is easy
to prove that those points that switch their color to a given j form a stable Borel
set. Hence the final coloring is Borel. It is trivial that every new color is bounded
by D + 1. 
Now we turn to coloring the edges. Shannon’s Theorem asserts that the edges
of a multigraph with maximum degree D can be colored by at most 3D/2 colors.
For simple graphs, Vizing’s Theorem gives the better bound of D + 1. For Borel
graphs, only a weaker result is known (which is quite trivial in the finite case):
Theorem 18.4. Every Borel graph has a Borel edge coloring with 2D − 1 colors.
Proof. We define a new Borel graph, the “line-graph” L(G) of G, defined on
the set E(G) (which, as a subset of Ω × Ω, is equipped with a sigma-algebra of
Borel sets), where two edges of G are adjacent if and only if they share a common
endpoint. It is straightforward to see that this graph L(G) is Borel, and has degrees
at most 2D − 2. Applying Theorem 18.3 to L(G), we get the Theorem. 
If we have a Borel graph, then various basic constructions lead to Borel sets
and functions. We state and prove this fact for the degree function, but many other
similar assertions can be proved (see the Exercises below). For every set A ⊆ Ω
and x ∈ Ω, let degGA (x) denote the number of neighbors of x in A. We suppress the
superscript G if there is only one graph around.
Lemma 18.5. Let G be a Borel graph. Then for every Borel set A ⊆ V (G),
degA (x) is a Borel function of x.
Proof. Let Pi (i = 1, 2, . . . ) range over all partitions of V (G) into D sets from
the generator set J . Then
∑
degA (x) = max 1NG (A∩J) (x),
i∈N
J∈Pi

showing that this function is Borel. 

Let G be a graph (of any cardinality). We define the local distance of two nodes
u, v ∈ V (G) by
d◦ (u, v) = inf{2−r : BG,r (u) ∼
= BG,r (v)}.
This turns V (G) into a semimetric space. Unfortunately, two different nodes may
be at distance 0: this happens exactly when there is an automorphism of G moving
one onto the other. (This is the first occurrence of the “curse of symmetry” in this
part; it has caused difficulties in Chapter 6, and it will haunt us when constructing
graphings or designing local algorithms for large graphs.) We call the topology
defined by this semimetric the local topology. This defines a local topology on V (G)×
V (G).
Assuming that there are no points at distance 0, local distance defines an
ultrametric space (i.e., the triangle inequality holds in a very strong sense: d◦ (x, y) ≤
max{d◦ (x, z), d◦ (z, y)}). This implies (or it is easy to see directly) that the set of
nodes whose r-neighborhood has a fixed isomorphism type is both closed and open,
and such sets form an open basis. The space is totally disconnected.
332 18. GRAPHINGS

Proposition 18.6. Let G be a bounded degree graph (of any cardinality), and
suppose that G has no automorphism. Then E(G) is closed in the local topology,
and hence G is a Borel graph with respect to the Borel space defined by the local
topology.
Proof. Let xy ∈ / E(G), and let y1 , . . . , yd be the neighbors of x. Since G has
no automorphism, there is an r ≥ 1 such that BG,r (y) ̸∼ = BG,r (y1 ), . . . , BG,r (yd ).
We claim that if u, v ∈ V (G) such that d◦ (u, x) < 2−r and d◦ (v, y) < 2−r , then
uv ∈
/ E(G).
Assume that uv ∈ E(G). By the definition of the distance function, d◦ (u, x) <
2−r implies that BG,r+1 (x) ∼ = BG,r+1 (u). Let, say, y1 correspond to v under this
isomorphism, then BG,r (y1 ) ∼= BG,r (v). But d◦ (u, x) < 2−r implies that BG,r (v) ∼ =
BG,r (y), so BG,r (y1 ) ∼ B
= G,r (y), a contradiction. 
What to do if G has automorphisms? One possibility is to decorate the nodes
from some set K of “colors”, in order to break all automorphisms. A similar
construction will be described in Section 18.3.4, and here we don’t go into the
details.
Exercise 18.7. Let G be a Borel graph, and let us add all edges that connect
nodes at distance 2. Prove that the resulting graph G2 is Borel.
Exercise 18.8. Let G be a Borel graph. Prove that for every 1-labeled simple
graph F , the quantity homu (F, G) is well-defined, and it is a Borel function of
u ∈ V (G).
Exercise 18.9. Let G be a Borel graph and let Vk denote the set of nodes with
degree k. Prove that Vk is a Borel set.
Exercise 18.10. Let G be a Borel graph and let Vk denote the union of its finite
components with k nodes. Prove that Vk is a Borel set.
Exercise 18.11. Prove that every Borel graph has a maximal stable set of nodes
that is Borel.
Exercise 18.12. Prove that if a graph with bounded degree has no automor-
phism, then its cardinality is at most continuum.

18.2. Measure preserving graphs

Now we come to the definition of graphs that will serve as limit objects for
convergent sequences of bounded degree graphs. We endow the sigma-algebra (Ω, B)
with a probability measure λ. We say that a graph G with node set Ω is measure
preserving, or a graphing, if it is Borel and for any two measurable sets A and B,
we have
∫ ∫
(18.2) degB (x) dλ(x) = degA (x) dλ(x).
A B
In other words, “counting” the edges between A and B from A, we get the same as
counting them from B. To be precise, a graphing is a quadruple G = (Ω, B, λ, E),
where Ω = V (G) is a set, B = B(G) is a Borel sigma-algebra on Ω, λ = λG is a
probability measure on B, and E = E(G) ∈ B × B is a Borel set satisfying the
measure preserving condition (18.2).
Remark 18.13 (On terminology). The name “graphing” was introduced by
Adams [1990], and it refers to the representation of the classes of an equivalence
relation as the connected components of a Borel graph. It seems, however, that
the usage is shifting to the one above. Since these objects are analogous to our
 18.2. MEASURE PRESERVING GRAPHS 333

“graphons” in the dense case (whose name comes from the contraction of graph-
function), I like the parallel “graphon–graphing”, and will adopt the above meaning.
Besides the probability measure λ on the points, there are two (related) mea-
sures that often play a role. The integral measure of the degree function is often
called the volume:
∫
(18.3) Vol(A) = deg(x) dλ(x).
A
The volume of the whole underlying set is the average degree:
∫
(18.4) d0 = Vol(Ω) = deg(x) dλ(x).
Ω

We can normalize the volume to get a probability measure λ∗ (A) = Vol(A)/Vol(Ω).

We call the distribution λ∗ the stationary distribution of G; the name refers to the
random walk on G.
We can define a finite measure η = ηG on (Ω × Ω, B × B) by
∫
(18.5) η(A × B) = degB (x) dλ(x)
A

for product sets (A, B ∈ B). It is not hard to see that Caratheodory’s Theorem
applies and we can extend η to the sigma-algebra B × B. If we want a probability
measure on the edges, we can normalize by the average degree: the measure η/d0
can be considered as the uniform probability measure on E(G). Equation (18.2)
implies that η is invariant under interchanging the coordinates. Both marginals of
η give the volume measure.
Lemma 18.14. The measure η is concentrated on E(G).
Proof. Let J = {J1 , J2 , . . . }. We claim that
∞
∪ ( )
(18.6) E(G) = (Ω × Ω) \ Ji × (Ω \ NG (Ji )) .
i=1

It
∪ is ( clear that E(G) )is contained in the right hand side. Conversely, if (x, y) ∈ /
i Ji × (Ω \ N G (J i )) , then for each i for which Ji ∋ x, we have y ∈ NG (Ji ). So
there is a zi ∈ Ji adjacent to y. Since y has finite degree, this can hold for each Ji
only if x is (adjacent to y. This) proves (18.6).
Since η Ji × (Ω \ NG (Ji )) = 0 by the definition of η, equation (18.6) implies
that η(Ω × Ω \ E(G)) = 0. 

Assuming that the average degree is positive, one way to generate a random
edge from the distribution η/d0 is to select a point x from the distribution λ∗ , and
then select an edge incident with x uniformly at random. Conversely, selecting a
random edge from the distribution η/d0 , and then selecting randomly one of its
endpoints, we get a point from the distribution λ∗ . To describe the connection
between λ and λ∗ in this language, we can generate a point from λ∗ by generating
a random point x according to λ, and keeping it with probability deg(x)/D (else,
rejecting it and generating a new one). If there are no isolated nodes, we can
generate a point from λ by generating a random point x according to λ∗ , and
keeping it with probability 1/ deg(x).
334 18. GRAPHINGS

Example 18.15. If D = 1, then every graphing G is the graph of an involution

φ : S → S for some set S ⊆ V (G). Since S = NG (V (G)), it is measurable.
Furthermore, for any measurable A ⊆ S we have
∫ ∫
( )
λ φ−1 (A) = degA (x) dλ(x) = degφ−1 (A) (x) dλ(x) = λ(A),
φ−1 (A) A

and so φ is a measure preserving map. 

Example 18.16 (Graphings from graphs). For any finite graph F , we define
a graphing GF as follows. Let V (F ) = [n], and let us split the unit interval [0, 1)
into n intervals Ji = [(i − 1)/n, i/n). For every edge ij ∈ E(F ) with i < j, let us
connect every point x ∈ Ji to x + (j − i)/n ∈ Jj . It is not hard to verify that the
resulting graph GF is measure preserving. Every connected component of GF is
isomorphic to F . See Figure 18.1(a) for the graphing on [0, 1] representing of the
pentagon. The picture is similar to the pixel picture of the graphon associated with
a simple graph, except that instead of a black square, we have a white square with
a diagonal. 
Example 18.17 (Cyclic graphing). Consider the graph Ca introduced in Exam-
ple 18.1. Endowing it with the uniform measure on [0, 1] turns it into a graphing. If
a is rational, then every connected component of Ca is a cycle; else, every connected
component is a 2-way infinite path. In this latter case, we call Ca an irrational
cyclic graphing (Figure 18.1(b)). 

Figure 18.1. Three graphings: (a) the pentagon as a graphing,

(b) a cyclic graphing, (c) the union of a symmetric finite set of
segments with slopes ±1; every such picture defines a graphing.
The slopes of ±1 correspond the measure preservation.

Remark 18.18. It may be useful to allow graphs measurable with respect to the
completion of B instead of B with respect to the probability measure λ (in the case
of the interval [0, 1], this means allowing Lebesgue measurable sets instead of Borel
sets). We call a graph Lebesgue measurable if for every set A ∈ A, its neighborhood
NG (A) ∈ A. The correspondence between graphings and Lebesgue measurable
graphs is described in Exercises 18.30–18.32.
18.2.1. Verifying measure preservation. Suppose that we have a measur-
able graph G with a probability measure λ on the node set. How to verify that
this graph is measure preserving? Let us describe some methods to do so.
 18.2. MEASURE PRESERVING GRAPHS 335

Edge measure. The simplest method, which often works well, is to specify a
measure η on the edge set satisfying (18.5). To be more precise, we consider a
probability space (Ω, B, λ) and a Borel graph G on it. Suppose that there exists a
finite measure η on the Borel sets in Ω × Ω, which is invariant under interchanging
the coordinates, concentrated on E, and its marginal is the volume measure Vol.
This trivially implies that (18.2) holds.
Borel subgraphs. Every subgraph of a graphing that is in a sense explicitly
definable is itself a graphing: there is no constructive way to violate (18.2). The
following lemma makes this precise.
Lemma 18.19. Let G = (Ω, B, λ, E) be a graphing, and let L ⊆ E be a symmetric
Borel set. Then G′ = (Ω, B, λ, L) is a graphing.
Proof. Let A, B ⊆ Ω be Borel sets. We want to show that
∫ ∫
′ ′
(18.7) dG
B (x) dλ(x) = dG
A (x) dλ(x).
A B
[ ]
First we prove that this equation holds when L = E ∩ (S × T ) ∪ (T × S) with
two disjoint Borel sets S, T . Indeed, for any two Borel sets A and B,
∫ ∫ ∫
′
dG
B (x) dλ(x) = d G
B∩T (x) dλ(x) + dG
B∩S (x) dλ(x)
A A∩S A∩T
∫ ∫ ∫
′
= dG
A∩S (x) dλ(x) + dG
A∩T (x) dλ(x) = dG
A (x) dλ(x).
B∩T B∩S B

A similar computation shows that (18.7) holds if L = E ∩ (S × S) for any Borel set
S.
To prove the lemma in general, we use induction on the degree bound D. For
D = 1 the assertion is trivial.
Let ε > 0, let J = {J1 , J2 , . . . } be a countable generator set of B, and let Pn
be a partition of V (G) into the atoms generated by J1 , . . . , Jn . Let Xn be the set
of points with degree D all whose neighbors belong to the same class of Pn . Since
any two points are separated by Pn if n is large enough, we have ∩n Xn = ∅, and
hence λ(Xn ) ≤ ε if n is large enough. Let us fix such an n.
For S ∈ Pn , let S ′ = S \ Xn . For every S, T ∈ Pn , the graph G(S, T ) obtained
by restricting the edge set of G to (S ′ × T ′ ) ∪ (T ′ × S ′ ) is measure preserving by
the special case proved above. In G(S, T ), each point has degree at most D − 1,
by the definition of S ′ and T ′ . Hence by the induction hypothesis, restricting the
edge set to E(G(S, T )) ∩ L we get a graphing G′ (S, T ), which means that
∫ ∫
G′ (S,T ) G′ (S,T )
dB dλ(x) = dA (x) dλ(x).
A B

Since the graphings G (S, T ) are edge-disjoint, it follows that G1 = ∪S,T G′ (S, T )
′

is measure preserving. We get G1 from G′ by deleting all edges incident with Xn ,

and hence for any two measurable sets A and B, we have
∫ ∫ ∫ ∫
G′ G1 G′ ′
dB (x) dλ(x) = dB (x) dλ(x) + dB∩Xn (x) dλ(x) + dG
B\Xn (x) dλ(x).
A A A A∩Xn
336 18. GRAPHINGS

Here
∫ ∫ ∫
′
dG
B∩Xn (x) dλ(x) ≤ dG
B∩Xn (x) dλ(x) = A (x) dλ(x) ≤ Dλ(Xn ) ≤ Dε,
dG
A A B∩Xn

and ∫
′
B\Xn (x) dλ(x) ≤ Dλ(Xn ) ≤ Dε.
dG
A∩Xn

Hence
   
∫ ∫  ∫ ∫ 
   
 dG ′
(x) dλ(x) − d G′
(x) dλ(x) ≤  dG1
(x) dλ(x) − dG1
(x) dλ(x) + 2Dε
 B A   B A 
   
A B A B
= 2Dε.
Since ε was arbitrarily small, this proves (18.7). 

Corollary 18.20. The intersection and union of two graphings on the same prob-
ability space are graphings.
Proof. Let G1 and G2 be the two graphings, then consider G1 ∩ G2 (we keep
the underlying point set and do the set operation on the edge set). This is a Borel
subgraph of G1 , and hence it is a graphing.
The assertion about the union is trivial if the graphings are edge-disjoint. In
the general case, consider the graphs G1 \ G2 , G2 \ G1 and G1 ∩ G2 . These three
graphs are Borel subgraphs of one of the graphings G1 and G2 , and hence they are
graphings. But then so is their union, which is just G1 ∪ G2 . 

Measure preserving families. Another way to “certify” the measure preserva-

tion condition is to use the simpler notion of invertible measure preserving maps.
Let A1 , . . . , Ak , B1 , . . . , Bk be measurable subsets of a Borel space (Ω, B),
and let φi : Ai → Bi be invertible measure preserving maps. The tuple
H = (φ1 , . . . , φk ) will be called a measure preserving family (see Gaboriau [2002],
Kechris and Miller [2004]). From every measure preserving family H we get a di-
→
−
rected multigraph G on Ω by connecting x, y ∈ Ω for every i such that y = φi (x).
The edges of this digraph are colored with k colors in such a way that each color-
class defines a measure preserving bijection between two measurable subsets of Ω.
Forgetting the orientation and the edge-colors of this digraph, we get a graph
G with degrees bounded by 2k, which we call the support graph of the measure
preserving family.
It is more natural perhaps to assume that the maps φ1 , . . . , φk are involutions,
in which case we get an undirected graph right away. We say that the measure
preserving family is involutive. A little advantage of working with involutions is
that we could extend the maps φi to measure preserving involutions Ω → Ω, and
would not have to worry about domains Ai and ranges Bi .
A graphing with its edges colored and oriented so that each color defines an
invertible measure preserving map is equivalent to a measure preserving family.
Conversely, every graphing can be “certified” by an appropriate measure preserving
family.
 18.2. MEASURE PRESERVING GRAPHS 337

Theorem 18.21. A graph supporting a measure preserving family is a graphing.

Conversely, for every graphing G there is an involutive measure preserving family
with at most 2D − 1 parts supported by G.

This measure preserving family is not unique in general.

Proof. If the family consists of a single measure preserving map φ, then it is

easy to check that its support graph is measure preserving. For a general measure
preserving family, this follows by Corollary 18.20.
To prove the converse, Theorem 18.4 implies that the edges of G can be split
into at most 2D −1 Borel sets that are matchings. By Lemma 18.19, the involutions
that these matchings define are measure preserving. 

While it is nicer to work with measure preserving involutions, the generality

allowed in the first statement of the previous theorem has some merits. It is often
easier to construct a family of non-involutory measure preserving maps to support
the graph. Furthermore, while the minimum number of maps in a measure pre-
serving family with given support graph, and the minimum number of maps in
an involutive family can be mutually bounded by factors of 2, the former may be
smaller. Both of these merits are illustrated by the following example.

Example 18.22. Consider an irrational cyclic graphing Ca as in Example 18.17;

we may assume without loss of generality that 0 < a < 1/2. This graphing is
2-regular, and as a graph it consists of disjoint 2-way infinite paths.
We claim that Ca cannot be represented by two involutions. Suppose that the
involutions φ1 and φ2 define Ca . Then each point x ∈ V (Ca ) is matched with
x − a (mod 1) by φ1 and to x + a (mod 1) by φ2 , or the other way around. Let
A1 denote the set of points matched the first way, and A2 , the rest. Then trivially
A1 and A2 are Borel sets, A1 ∪ A2 = V (Ca ), and A2 = A1 + a (mod 1). But this
is impossible by basic results in ergodic theory, since the map x 7→ x + a is ergodic.
On the other hand, we can represent Ca by three involutions: One matches
points x and x + a if 0 < x ≤ 1 − a and x ∈ (2ka, (2k + 1)a]( for some k ∈ N; the]
other matches points x and x + a if 0 < x ≤ 1 − a and x ∈ (2k + 1)a, (2k + 2)a
for some k ∈ N; the third matches points x and x + a − 1 if 1 − a < x ≤ 1. 

Example 18.23 (Squaring the circle). Answering a problem of Tarski,

Laczkovich [1990] proved that a circular disc D can be partitioned into a fi-
nite number of sets, and these can be translated so that they form a parti-
tion of a square S with the same area. (It is not known whether this can be
achieved by measurable pieces.) This result gives rise to interesting graphings. Let
X1 , X2 , . . . , Xm be the pieces of D, and let v1 , . . . , vm be the translation vectors
(so that X1 + v1 , . . . , Xm + vm form a partition of S). We may assume that D and
S are disjoint, and λ(D) = λ(S) = 1/2. We define a bipartite graph G on D ∪ S
by connecting x ∈ D to y ∈ S iff y − x ∈ {v1 , . . . , vm }. Clearly G is a Borel graph,
and every point has degree at most m. Furthermore, those edges that are defined
by the same vector vi define a measure preserving map between D ∩ (S − v − i) and
(D + vi ) ∩ S. Hence by Theorem 18.21, G is a graphing.
The theorem of Laczkovich is equivalent to saying that the vectors v1 , . . . , vm
can be chosen so that the resulting graphing G has a perfect matching. It is not
338 18. GRAPHINGS

known whether (for an appropriately rich family of translations) it has a Borel per-
fect matching. (Exercise 18.29 shows that a graphing can have a perfect matching,
but no Borel perfect matching.) 

Exercise 18.24. Let G be a graphing, and let us add all edges that connect
nodes at distance 2. Prove that the resulting graph G2 is a graphing.
Exercise 18.25. Let G be a graphing in which every connected component has at
most k nodes. Let S ⊆ V (G) be a measurable set that intersects every connected
component. Prove that λ(S) ≥ 1/k.
Exercise 18.26. Let G be a graphing on [0, 1], let E ′ ⊆ E(G) be a symmetric
Borel set, and E ′′ = E(G)\E ′ . Consider the graphings G′ and G′′ on [0, 1] defined
by the edge sets E ′ and E ′′ (cf. Lemma 18.19). Prove that ηG = ηG′ + ηG′′ .
Exercise 18.27. Let G be a graphing, and let S ⊆ E(G) be a (not necessarily
symmetric) Borel set. For x ∈ V (G), let d+ S (x) denote the number of pairs
−
(x,
∫ y) ∈ S, and
∫ let dS (x) denote the number of pairs (y, x) ∈ S. Prove that
d+ dλ = V (G) d−
V (G) S S dλ.

Exercise 18.28. Let Gi (i = 1, 2) be a graphing on [0, 1]. Define the categorical

product G1 × G2 (as a graph on [0, 1] × [0, 1]), and prove that it is a graphing
(cf. Aldous and Lyons [2007]).
Exercise 18.29. Let Ca be an irrational cyclic graphing. (a) Show that it con-
tains a perfect matching, but no Borel measurable perfect matching (Laczkovich
[1988]). (b) Show that if M is any Borel measurable matching in Ca , then there is
an augmenting path: a path of odd length such that its endpoints are not covered
by M , but every second edge on the path belongs to M (Elek and Lippner [2010]).
Exercise 18.30. Prove that for every Lebesgue measurable graph G there is a set
T ⊆ V (G) with measure 1 such that the subgraph G[T ] obtained by restricting
the edge set to T × T is Borel.
Exercise 18.31. Show by an example that there is a Borel graph on [0, 1] that
is not Lebesgue measurable.
( )
Exercise 18.32. (a) Let G be a Borel graph such that λ NG (A) = 0 for all
A ⊆ V (G) with λ(A) = 0. Prove that G is Lebesgue measurable. (b) Prove that
every graphing is Lebesgue measurable.

18.3. Random rooted graphs

The construction of Benjamini and Schramm [2001] for limit objects of bounded
degree graph sequences is different from graphings, but closely related to them. (It
is interesting to note that the relationship of these objects with convergent graph
sequences is not completely known, but their relationship with each other is quite
well understood.) They will be helpful throughout, in particular in extending the
notion of weak isomorphism to graphings and characterizing it.
Let us pick a random point x of a graphing G. (When talking about a random
point of a graphing G = (Ω, B, E, λ), we mean that it is selected according to the
probability measure λ.) Consider the connected component Gx containing it: This
is a countable graph with bounded degree, and it has a special “root”, namely the
node x. So we have generated a random rooted connected graph with bounded
degrees. We start with making precise sense of this, by constructing an interesting
Borel graph, the “graph of graphs”.
 18.3. RANDOM ROOTED GRAPHS 339

18.3.1. The graph of graphs. Let G• denote the set of connected countable
graphs (with all degrees bounded by D) that also have a specified node called their
root. We denote the root of a graph H ∈ G• by root(H). (We could consider these
graphs as 1-labeled, but in this context calling the single labeled node the “root” is
common.) Sometimes we will write we also write H = (H ′ , v), where v = root(H),
and H ′ = [[H]] is the unrooted graph underlying H. For every rooted graph H,
we denote by deg(H) the degree of its root. The set of finite graphs in G• will be
denoted by G • .
We consider two graphs in G• the same if there is an isomorphism between
them that preserves the root. Let Br ⊆ G • denote the set of r-balls, i.e., the set
of finite rooted graphs in which every node is at a distance at most r from the
root. (Since we keep the degree bound D fixed, the set Br is finite.) For a rooted
countable graph (H, v) ∈ G• , let BH,r = BH,r (v) ∈ Br denote the neighborhood of
the root with radius r. For every r-ball F , let G•F denote the set of “extensions”
of F , i.e., the set of those graphs H ∈ G• for which BH,r ∼ = F (as rooted graphs).
With all this notation, we can define something more interesting. First we
define a graph H on the set G• . Let (H, v) ∈ G• . For every edge e = vv ′ ∈ E(H),
connect (H, v) by an edge to the rooted graph (H, v ′ ) ∈ G• . So every edge of H
incident with v gives rise to an edge of H incident with (H, v). In particular, all
degrees in H are bounded by D. We call H the “Graph of Graphs”.
The r-neighborhood of a rooted graph H in H is almost the same as the r-
neighborhood of the root in H. To be precise, if [[H]] has no automorphism, then
BH,r (H) ∼ = BH,r (root(H)). The image of v ∈ V (H) under this isomorphism is
obtained by moving the root of H to v. However, if there is an automorphism of
H moving root(H) to v, then the “curse of symmetry” strikes again, and this map
is not one-to-one.
We endow the set G• with a metric: For two graphs H1 , H2 ∈ G• , define their
ball distance by
d• (H1 , H2 ) = inf{2−r : BH1 ,r ∼
= BH2 ,r }.
(This is reminiscent of the semimetric defining the local topology of a graph, but
it is defined on a different set.) This turns G• into a metric space. It is easy to see
(Exercise 18.43) that the sets G•F are both closed and open, they form an open basis,
and the space (G• , d• ) is compact and totally disconnected. The sigma-algebra of
Borel sets of (G• , d• ) will be denoted by A.
As usual, every subset of G• you will ever need, and every function G• → R
you will ever define, will be Borel. The graph H is Borel with respect to the
sigma-algebra A. This follows by the same argument as Proposition 18.6.

18.3.2. Invariant measures. You may have noticed that we have not defined
any measure on the set G• . We will in fact consider many probability measures on
it; these measures will carry the real information. Let σ be any probability measure
on (G• , A). It is easy to see that the degree deg of the root is a measurable function
on G. Nodes with different degrees cause some complication here, and it will be
best to introduce right away another probability measure on G• : we define
∫ /∫
σ ∗ (A) = deg dσ deg dσ.
A G•
340 18. GRAPHINGS

Clearly these integrals are finite (at most D). If the denominator is 0, then σ is
concentrated on the graph consisting of a single node (the only connected graph
with average degree 0). In this trivial case, we set σ ∗ = σ.
Next, we introduce a very important condition on the distribution, which ex-
presses that all possible roots of a graph are taken into account judiciously. (The
meaning of this condition will be clearer when we get to limits of graph sequences.)
Select a rooted graph H according to the distribution σ ∗ and then select a uniform
random edge e from the root. We consider e as oriented away from the root. This
way we get a probability distribution σ → on the set G→ of graphs in G• with an
oriented edge (the “root edge”) from the root also specified. We say that σ is involu-
tion invariant (another name commonly used is unimodular) if the map G→ → G→
obtained by reversing the orientation of the root edge is measure preserving with
respect to σ → . By an involution invariant random graph we mean a random rooted
connected graph drawn from an involution invariant probability measure on G• .
Example 18.33. Let G ∈ G be a connected finite graph. Selecting a root from
V (G) uniformly at random defines a probability distribution σG on G• (concen-
trated on rooted copies of G). If we select the root v with probability proportional
to the degree of v, and a root edge e incident with v uniformly, then simple compu-
tation shows that the edge is uniformly distributed among all oriented edges, and
so the distribution σG is involution invariant. 
Example 18.34 (Path). Let P denote the two-way infinite path with any node
chosen as a root. The distribution on G• concentrated on P is involution invariant,
since selecting any root edge we still get a distribution concentrated on a single
graph, so reversing the edge preserves this distribution. 
Example 18.35 (Triangular Ribbon). Let P be the 2-way infinite path and let
R be the “ribbon” obtained from P by connecting every pair of nodes at distance
2 (Figure 18.2(a)). If we specify any node as its root, we get a connected countable
4-regular rooted graph R• . The distribution on G• (where D = 4) concentrated on
R• is involution invariant. To see this, note that if we select an oriented edge as a
root, we get only two edge-rooted graphs H ′ and H ′′ (up to isomorphism): either
an edge of P is selected, or an edge not on P . Furthermore, reversing the edge
yields an isomorphic edge-rooted graph, so the distribution on {H ′ , H ′′ } remains
involution invariant. 

(a) (b)

Figure 18.2. (a) The triangular ribbon. (b) The grandmother

graph. Note that the neighborhood of a node reveals the orienta-
tion of the tree.
 18.3. RANDOM ROOTED GRAPHS 341

Example 18.36 (Grandmother graph). Let T be a two-way infinite binary tree,

and let us connect every node to its grandparent (Figure 18.2(b)). The resulting
8-regular connected graph H has a node-transitive automorphism group, and hence
if specify any node as its root, we get isomorphic rooted graphs. The distribution
on G• (where D = 8) concentrated on H is, however, not involution invariant.
To see this, we again determine the possible edge-rooted graphs that we obtain
from H. We get 4 types: an edge of T oriented “up”, an edge of T oriented
“down”, an edge not in T oriented “up”, and an edge not in T oriented “down”. It
is not hard to check that these are non-isomorphic, and the probabilities they are
obtained with are (in the above order) 2/8, 1/8, 4/8, 1/8. Reversing the root edge
interchanges the first two and the last two probabilities, so the distribution is not
involution invariant. 
18.3.3. Graphings and random rooted graphs. Let us recall the simple
construction at the beginning of this section, providing a link between graphings
and involution invariant distributions. Let G be a graphing and choose a random
point x ∈ V (G). The connected component Gx of G containing x, with a root x,
is a graph in G• , which we call a random rooted component of G. The map x 7→
Gx , which we will call the component map, is measurable as a map (V (G), A) →
(G• , A), and thus every graphing G defines a probability distribution σ = σG on
(G• , A). Selecting x from the distribution λ∗ , the graph Gx will be a random rooted
connected graph from the distribution σ ∗ . Selecting an edge of Gx incident with
x, we get an edge of G from the probability distribution ηG /d0 , together with an
orientation. Since ηG is symmetric, shifting the root to the other endpoint does
not change the distribution. Hence σ is involution invariant.
So every graphing gives rise to a (well-defined) involution invariant random
graph; we also say that the graphing represents this distribution. The following
converse to this statement was known in various related contexts for some time; for
written versions, see Aldous and Lyons [2007] and Elek [2007a].
Theorem 18.37. Every involution invariant probability distribution on G• can be
represented by a graphing.
Here we don’t claim uniqueness any more. This will be quite relevant a little
later! Before proving this theorem, let us consider a couple of examples.
Example 18.38 (Grid). Consider the involution invariant random graph concen-
trated on the infinite planar grid (with any root). We can construct a graphing
representing this by taking two irrational reals α and β that are independent over
the rationals, and connecting every x ∈ [0, 1) to x + α (mod 1), x − α (mod 1),
x + β (mod 1) and x − β (mod 1). There are many other constructions, for exam-
ple, we could take the unit square [0, 1)2 as the underlying probability space, and
connect (x, y) to (x ± α (mod 1), y ± β (mod 1)). Every connected component of
this graphing will be an infinite grid. 
Example 18.39. Consider the involution invariant random graph that is concen-
trated on the D-regular tree with a root (which is unique up to isomorphism). How
to represent this by a graphing?
For D = 2, an irrational cyclic graphing is a graphing representation. Here is
another one: let us randomly 2-color the path with colors 0 and 1. The sequence of
colors to the right from the root (including the root) can be thought of as a number
342 18. GRAPHINGS

x ∈ [0, 1]. (Let us ignore the ambiguity that one can write rational numbers whose
denominator is a power of two in two different ways; this involves a set of measure 0
anyway.) Similarly, the sequence of colors to the left of the root (this time excluding
the root) gives a number y ∈ [0, 1]. So every point of the unit square corresponds to
a 2-colored 2-way infinite path (with a root), and this correspondence is bijective.
To shift the root to the right by one step corresponds to replacing x by 2x (mod 1)
and y by y/2 if x < 1/2 and by y/2+1/2 if x ≥ 1/2. (This map, as a transformation
of {0, 1}Z , is called a Bernoulli shift. In its other incarnation as a transformation
of the unit square, it is sometimes called the dough folding map.) The graphing
will be defined on [0, 1]2 (with the Lebesgue measure), and every point (x, y) will
be connected to its image and to its inverse image under the dough folding map.
For D > 2 a geometric construction of a representing graphing is more com-
plicated. We can start from the fact that a D-regular tree is the Cayley graph
of a group freely generated by D involutions. This group can be represented, for
example, by reflections in D generic hyperplanes through the origin in D-space.
If we take the surface of the unit sphere in RD with the uniform probability dis-
tribution, and connect every point to its images and inverse images under these
reflections, we get a graphing representing the infinite D-regular tree. (Points on
the D hyperplanes in which we reflect will have lower degree than D; but we can
delete these points and all their images under the group, which is a set of measure
0, and then we get a graphing in which every connected component is a D-regular
tree.) 

As a preparation for the proof of Theorem 18.37, we describe a rather simple

construction of a measurable graph from an involution invariant random graph
(which, unfortunately, does not always represent the right measure). Consider the
“graph of graphs” H constructed in Section 18.3.1. We have seen that H is a Borel
graph. Unfortunately, this Borel graph with the measure σ does not represent the
involution invariant distribution σ in general. For example, in Example 18.34 the
graph H will have a single node (the rest is of measure 0), and this is too simple
to represent anything nontrivial. Exercise 18.47 describes an even worse example,
showing that an involution invariant measure does not necessarily turn the graph
H into a graphing.
The problem is clearly caused by the symmetries in the graph σ is concentrated
on. This motivates the following lemma.

Lemma 18.40. Let σ be an involution invariant distribution on G• such that

almost all graphs from σ have no automorphism (as unrooted graphs). Then (H, σ)
is a graphing that represents σ.

Proof. First, we prove that (H, σ) is a graphing. Choose a rooted graph

(H, v) ∈ G• from the distribution σ ∗ , and a random neighbor u of v (uniformly
from the neighbors). By the assumption that almost surely H has no nontrivial
automorphism, the graph (H, u) is almost surely different from (H, v). The pair
(H, u)(H, v) is an edge of H, and selecting another neighbor of v, we would get a
different edge of H. This describes a way to generate a random edge of H (with
an orientation). It follows from the involution invariance of σ that this distribution
on edges of H is invariant under flipping the orientation. Since the marginal of η
is σ ∗ , this implies that (H, σ) is measure preserving.
 18.3. RANDOM ROOTED GRAPHS 343

Let (H, v) be any rooted graph from σ with no automorphism. We want to

argue that the connected component of H containing (H, v) as a root is isomorphic
to (H, v). Indeed, assigning the role of the root to different nodes of H gives
non-isomorphic rooted graphs, and so we get an injection of V (H) → G• . From
the definition of adjacency in H, this embedding preserves adjacency and non-
adjacency, and the range is a connected component of H. This proves that (H, σ)
represents σ. 

18.3.4. The Bernoulli Graphing. To prove Theorem 18.37, we have to

break the symmetries of the graphs from σ. For this, we generalize the “graph of
graphs” construction. Let G+ denote the set of triples (H, v, α), where (H, v) ∈ G• ,
and α : V (H) → [0, 1] is a weighting of the nodes of H. Two such rooted, weighted
graphs are considered the same, if there is an isomorphism between the graphs that
preserves the root and preserves the weights. Let A+ be the sigma-algebra on G+
generated by the following cylinder sets: for an r ≥ 0, we fix the isomorphism type
of the ball B ∈ Br with radius r about the root, and also for every node in B,
we specify a Borel set in [0, 1] from which the weight is to be chosen. (The choice
of the interval [0, 1] to use for weighting is arbitrary; we could have decorated the
nodes by the points of any other Borel probability space. In fact, other decoration
will be needed later.) It is easy to see that (G+ , A+ ) is a Borel sigma-algebra.
We can define a graph on G+ , the Graph of Weighted Graphs H+ , as follows:
we connect two nodes (G, α) and (G′ , α′ ) by an edge if G′ arises from G by shifting
the root to one of its neighbors (while keeping all the nodeweights); in other( words,
)
if there is an isomorphism ι from( G to G′)(as unrooted graphs) such that α′ ι(u) =
α(u) for every u ∈ V (G), and ι root(G) is a neighbor of root(G′ ).
Given a probability distribution σ on (G• , A), we can define a probability dis-
tribution σ + on (G+ , A+ ) as follows: Select a random graph H ∈ G• from the
distribution σ, and assign independent, uniform random weights from [0, 1] to the
nodes.
Lemma 18.41. If σ is an involution invariant probability distribution on (G• , A),
then Bσ = (G+ , A+ , σ + ) is a graphing, and it represents σ.
This construction associates a graphing with every involution-invariant distri-
bution, which we call the Bernoulli graphing representing σ. (The name refers to
its close relationship with the Bernoulli shift in Example 18.39.) This lemma also
provides the proof of Theorem 18.37.

Proof. The proof is essentially the same as the proof of Lemma 18.40, since
assigning the random weights to the nodes of a graph G chosen from σ almost
surely destroys all automorphisms. 

Exercise 18.42. Prove that for r > 2, an r-ball has at most Dr nodes, and the
r
number of non-isomorphic r-balls is bounded by DD .
Exercise 18.43. Prove that the sets G•F are closed and open in the metric space
(G• , d• ), they form an open basis, and the space is homeomorphic to a Cantor
set.
Exercise 18.44. Show that a function f : G• → R is continuous if and only if
for every finite rooted graph F ∈ Br there is an ε > 0 such that for all graphs
H ∈ G• with BH,r ∼ = F , we have |f (G) − f (F )| < ε.
344 18. GRAPHINGS

Exercise 18.45. Prove that the following functions, defined for H ∈ G• ,

are Borel: (a) 1(H ∼ = H0 ), where H0 ∈ G and isomorphism is meant
as isomorphism ( of unlabeled
) ( ) graphs; (b) ω(H); (c) χ(H); (d) f (H) =
lim supn→∞ e Br (G) /v Br (G) .
Exercise 18.46. Let H ∈ G• , and consider the probability distribution on G•
concentrated on H. Prove that this distribution is involution invariant if and only
if the automorphism group of H is transitive on the nodes, and for every oriented
edge −→e , the orbit of −
→e (as a directed graph) has equal indegrees and outdegrees.
Exercise 18.47. Let G be a countably infinite graph consisting of a two-way
infinite path with two nodes of degree 1 hanging from every node of the path.
Let G1 and G2 be the two rooted graphs obtained from G by selecting a node of
degree 4 and a node of degree 1 as its root, respectively. Let π be the probability
distribution on G in which π(G1 ) = 1/3 and π(G2 ) = 2/3. Show that π is
involution invariant, but (H, π) is not measure preserving.

18.4. Subgraph densities in graphings

Our next goal is to generalize our central notion, graph homomorphism, to
involution-invariant random graphs and to graphings. Following our general frame-
work, we consider homomorphisms from a small graph into (say) an involution-
invariant random graph, as well as homomorphisms from an infinite graph into
small graphs. In both cases, there will be some nontrivial preparation that is
needed, including proving some results that are important on their own right.
In this section we address the easier task, defining homomorphism densities in
involution-invariant random graphs and graphings. This takes some preparation,
discussing an important consequence of involution invariance.

18.4.1. Mass Transport Principle. Let us consider the set G•• of 2-labeled
connected countable graphs (again, graphs that are isomorphic as 2-labeled graphs
are identified). We can endow this set with a compact topology just like we did for
G• , and then Borel functions are defined.
The following very useful characterization of involution invariance was proved
by Aldous and Lyons [2007] (it was in fact this form how Benjamini and Schramm
first defined involution-invariant measures).
Proposition 18.48 (Mass Transport Principle). Let σ be a probability distri-
bution on G• . Then σ is involution invariant if and only if for every Borel function
f : G•• → R+ the following identity holds:
(∑ ) (∑ )
(18.8) E f (H, v, u) = E f (H, u, v) ,
u u

where (H, v) ∈ G• is randomly chosen from the distribution σ.

Equation (18.8) allows that both sides be infinite. One can generalize the prin-
ciple to functions without the nonnegativity condition, by applying it to separately
to the negative and positive parts (however, one has to make sure that no infinite
expectations occur). The name refers to the following interpretation: if we trans-
port f (H, u, v) amount of mass from node u to node v in the countable graph H
(where this amount depends only on the isomorphism type of (H, u, v), and it de-
pends on it in a Borel measurable way), then on the average, no node gains or loses.
This is trivial for finite graphs, but it does not automatically hold for countable
graphs, since there are distributions on G• that are not involution invariant.
 18.4. SUBGRAPH DENSITIES IN GRAPHINGS 345

One can formulate a related identity for graphings, which shows that the Mass
Transport principle is in a sense a form of Fubini’s Theorem. To illustrate that we
∑ f : S ×S → R∑
can vary the conditions, let us say that a function (where S is a set
of any cardinality) is locally finite, if the sums x∈S f (x, y) and y∈S f (x, y) are
absolutely convergent (this includes that they have a countable number of nonzero
terms).
Proposition 18.49. Let G be a graphing, and let f : V (G) × V (G) → R be a
locally finite Borel function. Assume that f (x, y) = 0 unless y ∈ V (Gx ). Then
∫ ∑ ∫ ∑
f (x, y) dx = f (x, y) dy
y x
V (G) V (G)

If f is the indicator function of edges between two Borel sets A and B, then this
identity gives the basic measure preservation identity 18.2. The Mass Transport
Principle can be used to prove properties of “typical” graphs from an involution
invariant distribution; see Exercises 18.52 and 18.53.
We describe the proof of the graphing version; Proposition 18.48 can be proved
along the same lines.
Proof. It suffices to prove this identity for nonnegative Borel functions, since
we can write a general f as the difference of two such functions, which will also be
locally finite. It suffices to prove it for bounded Borel functions, since we can obtain
an unbounded nonnegative f as the limit of an increasing sequence of bounded Borel
functions. By scaling, we may assume that the range of f is contained in [0, 1]. We
may assume that there is an r ∈ N such that f (x, y) = 0 unless y ∈ BG,r (x), since
we can obtain f as the limit of an increasing sequence of such functions. Finally,
it suffices to consider 0-1 valued Borel functions, since we can write f as
∫ 1
f (x, y) = 1(f (x, y) ≥ t) dt,
0
and here the function 1(f (x, y) ≥ t) is a 0-1 valued Borel function for every t.
A 0-1 valued Borel function corresponds to a Borel subset S ⊆ V (G) × V (G).
Consider the graphing Gr obtained from G by connecting any two nodes at distance
at most r. (This is indeed a graphing by Exercise∫ 18.7.) The
∫ set S is a Borel subset
of E(Gr ), and hence by Exercise 18.27 we have d+ S dλ = d−
S dλ. But this is just
the identity to be proved. 
18.4.2. Homomorphism frequencies. Recall that in a finite graph G,
t∗ (F, G) can be interpreted as the expectation of homu (F ′ , G), where F ′ is ob-
tained from F by labeling one of its nodes, and u is a random node of G. This
can be generalized to homomorphisms into an involution-invariant random graph.
Indeed, let σ be an involution-invariant distribution, and let (H, v) denote a ran-
dom rooted graph from σ. Then homv (F ′ , H) is a bounded nonnegative integer,
and since it depends only on a bounded ( neighborhood
) of the root v, it is a Borel
function of (H, v). So t∗ (F ′ , σ) = E homv (F ′ , H) is well defined. Based on the
finite case, we expect that t∗ (F ′ , σ) is independent of the node labeled in F , and
so we can define t∗ (F, σ) = t∗ (F ′ , σ). This is correct, but not obvious.
Proposition 18.50. Let F ′ and F ′′ be two 1-labeled graphs obtained from the
same unlabeled connected graph F . Let σ be an involution-invariant distribution,
then t∗ (F ′ , σ) = t∗ (F ′′ , σ).
346 18. GRAPHINGS

Proof. Let F ∗ be the 2-labeled graph obtained by labeling both nodes that are
labeled in F ′ or F ′′ . Then
∑ for every rooted graph (H, u) generated
∑ according to σ,
we have homu (F ′ , H) = v homuv (F ∗ , H) and homu (F ′′ , H) = v homvu (F ∗ , H).
Applying the Mass Transport Principle to the function f (H, u, v) = homuv (F ∗ , H),
we get the assertion. 

If we want to define t(F, G) for a graphing G, it suffices to note that the

graphing defines a unique involution invariant distribution σ, and so we can define
t(F, G) = t(F, σ). Explicitly, homu (F ′ ,∫G) is a bounded measurable function of
u ∈ V (G), so the definition t∗ (F, G) = homu (F ′ , G) dλ(u) makes sense. By the
argument above, it follows that this value is independent of the choice of the labeled
node in F .
Exercise 18.51. Show that the distribution concentrated on the Grandmother
Graph (Example 18.36) violates the Mass Transport Principle.
Exercise 18.52. Suppose that an involution-invariant random rooted graph is
almost always infinite. Prove that the expected degree of the root is at least 2
(Aldous and Lyons [2007]).
Exercise 18.53. If G ∈ G• is a random graph from an involution invariant
distribution, then with probability 1, G has 0, 1, 2 or infinitely many ends. (An
end of a graph is defined as an equivalence class of one-way infinite paths, where
two paths are equivalent if they cannot be separated by a finite set of nodes.)

18.5. Local equivalence

Two graphings G1 and G2 are locally equivalent, if they have the same subgraph
densities: t∗ (F, G1 ) = t∗ (F, G2 ) for every connected simple graph F . We can
formulate this notion in terms of the sample distributions. Recall that if G is a
finite graph, then ρG,r (B) (B ∈ Br ) is the probability that the r-neighborhood
of a random node of G is isomorphic to the r-ball B. This definition extends to
graphings verbatim. Two graphings G1 and G2 are locally equivalent if and only
if the neighborhood distributions ρGi ,r are the same for every radius r. This is also
equivalent to saying that they represent the same involution invariant distribution
on the “graph of graphs” G• .
Our goal is to characterize weakly equivalent pairs of graphings. To this end,
we have to introduce some special maps that certify weak equivalence.
Let G1 and G2 be two graphings. We call a measure preserving map
φ : V (G1 ) → V (G2 ) a local isomorphism, if its restriction to almost every con-
nected component of G1 is an isomorphism with one of the connected components
of G2 . To be more precise, if x is a random point of V (G1 ), then φ(x) is a random
point of V (G2 ), and (G1 )x ∼= (G2 )φ(x) with probability 1 (as rooted graphs). So
the involution invariant distributions defined by G1 and G2 are the same. Note,
however, that a local isomorphism may not be invertible.
Example 18.54 (Cycles vs. bicycles). Consider the cyclic graphing Ca for an
irrational real number a, and let C′a be defined by connecting every x ∈ [0, 1] to
x ± (a/2) mod 1/2 if x < 1/2, and to 1/2 + (x ± (a/2) mod 1/2) if x ≥ 1/2.
Informally, C′a consists of two disjoint copies of Ca , shrunk by a factor of 2.
We claim that the map φ : x 7→ 2x mod 1 is a weak isomorphism from C′a
to Ca . It is easy to see that this map is measure preserving. Furthermore, any
connected component G of C′a lies either entirely in [0, 1/2) or entirely in [1/2, 1),
 18.5. LOCAL EQUIVALENCE 347

and therefore φ, restricted to G, is an isomorphism with an appropriate component

of Ga . 
Example 18.55 (Grid II). Let G1 and G2 be the two graphings defined in
Example 18.38 representing the infinite grid, where G1 is defined on [0, 1) and
G2 , on [0, 1)2 . Then the map (x, y) 7→ x + y (mod 1) defines a local isomorphism
from G2 to G1 : it is trivially measure preserving, and it is an isomorphism when
restricted to any connected component of G2 . 
The relation “G1 has a local isomorphism into G2 ” is transitive (since local
isomorphisms can be composed), but not symmetric (since local isomorphism may
not be invertible). To make it symmetric, we define (temporarily, as we shall see)
two graphings to be bi-locally isomorphic, if there is a third graphing that has
local isomorphisms into both. This is now a symmetric relation, but don’t we lose
transitivity? No, we don’t; the next lemma will be a main step in proving this.
Lemma 18.56. Let G1 and G2 be two graphings that both have a local isomorphism
into a third graphing G0 . Then they are bi-locally isomorphic.
Proof. Let φi : Gi → G0 be a local isomorphism. Consider the set
Ω = {(x1 , x2 ) : xi ∈ V (Gi ), φ1 (x1 ) = φ2 (x2 )},
We have Ω ∈ A(G1 ) × A(G2 ), which follows from the easy-to-check formula
∞
∪
(18.9) (V (G1 ) × V (G2 )) \ Ω = φ−1 −1
1 (Ji ) × φ2 (V (G0 ) \ Ji ),
i=1

where {J1 , J2 , . . . } is a countable generating Boolean algebra for A(G0 ). Let A

denote the sigma-algebra obtained by restricting A(G1 ) × A(G2 ) to Ω.
Next, we define an appropriate probability measure on (Ω, A), or more conve-
niently phrased, a coupling measure on V (G1 ) × V (G2 ) concentrated on Ω. Propo-
sition A.7 in the Appendix implies that such a measure exists. We denote by λ its
restriction to (Ω, A).
Finally, we have to define an appropriate graph on the probability space
(Ω, A, λ). We first define it on the product V (G1 ) × V (G2 ) as the categorical prod-
uct G1 × G2 , and then take the induced subgraph on Ω. (The graph G1 × G2 has
degrees bounded by D2 , but its restriction to Ω has degrees at most D, which will
follow from the proof below, but can be checked directly.) The projection map from
V (G1 ) × V (G2 ) onto V (Gi ) can be restricted to Ω, to get a map ψi : Ω → V (Gi ).
We claim that ψi is local isomorphism.
First, we show that ψi is measure preserving. Indeed, for any set A1 ∈ A(G1 ),
we have ψi−1 (A1 ) = (A1 × V (G2 )) ∩ Ω, and hence
( ) ( )
λ ψi−1 (A1 ) = λ (A1 × V (G2 )) ∩ Ω = λ(A1 × V (G2 ))
( )
= λ φ1 (A) ∩ φ2 (V (G2 )) = λ1 (A1 ).
(where λi is the probability measure of the graphing Gi ).
Let (x1 , x2 ) ∈ Ω be a random point from the distribution λ, and let H denote
its connected component. Let Hi = (Gi )xi denote the connected component of Gi
containing xi , and let H0 be the connected component of G0 containing φ1 (x1 ) =
φ2 (x2 ). With probability 1, these three graphs are isomorphic, and the maps φi
give isomorphisms H1 ∼ = H0 ∼ = H2 . For every node v ∈ V (H0 ), let αi (v) ∈ V (Hi )
348 18. GRAPHINGS

( )
be the node with φi αi (v) = v. (Note that φ−1 i (v) is not necessarily a singleton,
but exactly one of( its elements
) belongs to H i .)
Let α(v) = α1 (v), α2 (v) ∈ Ω, then α(u) and α(v) are adjacent in G if and
only if u and v are adjacent in H0 , by the definition of the product graph. So α
is an embedding of H0 into H as an induced subgraph. We want to argue that
α(H0 ) = H. Indeed, if not, then there (is a node ) α(u) that is connected by an
edge of H to a node (w1 , w2 ) ∈ V (H) \ α V (H0 ) . Now α1 (u) is connected to w1
in H1 by the definition of the product graph, and hence z = φ1 (w1 ) is connected
to u in H0 , and so it is a node in H0 . Similarly, φ2 (w2 ) is in H0 . Furthermore,
(w1 , w2 ) ∈ V (H) ⊆ Ω, and hence φ1 (w1 ) = φ2 (w2 ) = z. But then (w1 , w2 ) = α(z),
a contradiction.
It follows that with probability 1, G(x1 ,x2 ) ∼ = (Gi )xi , and ψi provides this
isomorphism, which proves that ψi is a local isomorphism. Thus G1 and G2 are
bi-locally isomorphic. 
Corollary 18.57. Bi-local isomorphism is a transitive relation.
Proof. Figure 18.3 tells the whole story: composing two bi-local isomorphisms,
the middle part can be “flipped up” by Lemma 18.56 to get a single bi-local iso-
morphism. 

Figure 18.3. (a) Bi-local isomorphism is transitive. (b) The re-

lationship between graphings, Bernoulli graphings and Bernoulli
lifts.

We need a construction, introduced by Hatami, Lovász and Szegedy [2012],

which is similar to the Bernoulli graphing defined in Section 18.3.3. For every
graphing G, we define the graphing G+ , which we call the Bernoulli ( ) lift of G. The
points of G+ will be pairs (x, ξ), where x ∈ V (G) and ξ : V Gx → [0, 1]. We
connect (x, ξ) to (y, ζ) if y is a neighbor of x and ξ = ζ (note that if y is a neighbor
of x, then Gx = Gy , so ξ and ζ are weightings of the same graph). Let Ω be the
set of such pairs. We define a sigma-algebra A on Ω generated by the sets φ−1 (A)
(A ∈ A(G)) and ψ −1 (A) (A ∈ B). To define a measure on (Ω, A), it is perhaps
easiest to describe how a random element is generated: we pick a random point x
of G, and then assign independent random weights ξ(u) to the nodes u of (G)x . It
is easy to see that G+ is a graphing.
There is a natural map φ : V (G+ ) → V (G), which simply forgets the weight-
ing. There is also a natural map ψ : V (G+ ) → V (Bσ ), where σ is a involution-
invariant distribution represented by G, (which )forgets all about G except the dis-
tribution σ: this is defined by ψ(x, ξ) = Gx , ξ . The following lemma is straight-
forward to verify.
 18.6. GRAPHINGS AND GROUPS 349

Lemma 18.58. The maps φ : V (G+ ) → V (G) and ψ : V (G+ ) → V (Bσ ) defined
above are local isomorphisms.
Now we are able to prove the main result in this section.
Theorem 18.59. Two graphings are locally equivalent if and only if they are bi-
locally isomorphic.
Proof. The “if” part is trivial by the discussion above. To prove the “only if”
part, let G1 and G2 be two locally equivalent graphings, we want to prove that they
are bi-locally isomorphic. They define the same involution invariant distribution
σ on G, and so they are both locally equivalent to the Bernoulli graphing Bσ .
Lemma 18.58 implies that they are both bi-locally isomorphic to Bσ . Corollary
18.57 implies that they are bi-locally isomorphic. 

Exercise 18.60. Let 0 ≤ a ≤ 1 be an irrational number, and define a graphing

C′′a (related to the graphings Ca and C′a in Example 18.54): we connect every
x ∈ [0, 1] to 1/2 + (x ± (a/2) mod 1/2) if x < 1/2, and to x ± (a/2) mod 1/2
if x ≥ 1/2. Informally, we consider two circles of circumference 1/2, and connect
every point on one to the two points a/2 away from the corresponding point on
the other circle. Prove that C′′a is locally equivalent to Ca , and construct a local
isomorphism C′′a → Ca .

18.6. Graphings and groups

Let (Ω, φ1 , . . . , φm ) be a measure preserving family, where φi : Ω → Ω are
measure preserving maps defined on the whole Ω. The maps φi generate a group
Γ of measure preserving maps.
Conversely, let Γ be a finitely generated group, with generators g1 , . . . , gm . Let
us assume, for simplicity, that together with each gi , its inverse gi−1 is also among
the generators. Let H be the Cayley graph of Γ: V (H) = Γ, and for every x ∈ Γ
and 1 ≤ i ≤ m, we connect x to xgi by an edge. We will get every edge in both
directions, so we may consider G as an undirected graph.
Consider the random rooted graph model which is concentrated on G with root
1 (the identity element of Γ). This is involution invariant (see Exercise 18.46), so it
defines a graphing G (Theorem 18.21). In fact, every oriented edge of H is marked
by a generator of the group, and this marking is inherited by G, and we can use
this marking to construct the graphing in Theorem 18.21.
This correspondence between finitely generated groups, graphings and measure
preserving families explains the interest of group theorists in the limit theory of
bounded degree graphs. We do not elaborate this quite broad and very active area
in this book; see e.g. Kechris and Miller [2004].
 CHAPTER 19

Convergence of bounded degree graphs

Convergence of a graph sequence with bounded degree was perhaps the first
which was formally defined (Benjamini and Schramm [2001]), but it is a more
complex notion than convergence in the dense case. There are more than one non-
equivalent reasonable definitions, which capture different aspects of the notion that
graphs in a sequence are becoming “more and more similar” to each other. We
treat two such notions in this Chapter.

19.1. Local convergence and limit

19.1.1. Distances. Just like in the dense case, we need to introduce some
notions of a distance between two bounded degree graphs before starting the treat-
ment of convergence. We don’t have a good analogue of the cut distance, and
therefore we will have to do with the sampling distance. This is a simpler notion,
but of course less powerful, since knowing that two graphs are close in sampling
distance does not translate into information about their global structure.
Recall that we have defined (in the introduction, informally) the sampling dis-
tance of two graphs F, F ′ ∈ G. To make the definition precise, we start with the
sampling distance of depth r, which is just the variational distance of neighborhood
distributions, and we simply sum these with convenient (but ad hoc) weights:
∑∞
1 r
(19.1) r
δ⊙ (F, F ′ ) = dvar (ρF,r , ρF ′ ,r ), δ⊙ (F, F ′ ) = δ (F, F ′ ).
r ⊙
r=0
2
Note that in the second expression, the term with r = 0 is 0.
Lacking a good analogue of the cut distance, this sampling distance will be our
main tool when comparing graphs. Since we can sample from a graphing just as
well as we can sample from a graph, this distance is defined for two graphings, and
also for a graphing and a graph. Since the sample distributions ρF,r and ρGF ,r are
the same (where GF is the graphing on [0, 1] representing the finite graph F ), we
have
(19.2) r
δ⊙ (F, F ′ ) = δ⊙
r
(GF , GF ′ ), and δ⊙ (F, F ′ ) = δ⊙ (GF , GF ′ ).
r
Using the trivial inequality δ⊙ (G, G′ ) ≤ δ⊙
r+1
(G, G′ ), we get that for every r ≥ 1,
1 r 1
(19.3) r
δ⊙ (G, G′ ) ≤ δ⊙ (G, G′ ) ≤ r + δ⊙
r
(G, G′ ).
2 2
An easy consequence of inequalities 19.3 is that if we want to estimate δ⊙ (G, G′ )
of two finite graphs (which by definition depends on infinitely many radii r) with
an error less than ε > 0, then we can do the following. We choose a positive integer
k > log(3/ε), and then we take sufficiently many samples from both G and G′
so that the empirical distributions φr and φ′r of r-balls in these samples satisfy
351
352 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

dvar (ρG,r , φr ) ≤ ε/6 and dvar (ρG′ ,r , φ′r ) ≤ ε/6 with high probability. We claim
∑k
that A = r=0 2−r dvar (φr , φ′r ) is a good estimate of δ⊙ (G, G′ ). Indeed, with high
probability,
∑k
1 ∑k
1
|δ⊙ (G, G′ ) − A| ≤ r
d var (ρG,r , φ r ) + d (ρ ′ , φ′r )
r var G ,r
r=0
2 r=0
2
∞
∑ 1
+ dvar (ρG,r , ρG′ ,r ).
2r
r=k+1

Here the first term is bounded by (1+1/2+· · ·+1/2k )ε/6 < ε/3, and similar bound
applies for the second term. The last term is bounded by 1/2k+1 + 1/2k+2 + · · · =
1/2k ≤ ε/3. So the total error is less than ε.
Often we have to compare two graphings G1 , G2 that are defined on the same
Borel graph G, and only differ in the invariant distributions π1 , π2 on them. In this
case the sampling distance can be bounded by the variational distance of π1 and
π2 ; it is easy to see that for every r ≥ 1, we have
(19.4) r
δ⊙ (G1 , G2 ) ≤ dvar (π1 , π2 ), and δ⊙ (G1 , G2 ) ≤ dvar (π1 , π2 ).
We will also need the edit distance of graphs/graphings on the same node set.
For two graphs G, G′ ∈ G with V (G) = V (G′ ) = [n], this is defined as
1
d1 (G, G′ ) = |E(G)△E(G′ )|.
n
The difference from the dense case is in the normalization. (We will not need the
“best overlay” version δ1 .) To extend the edit distance to two graphings G, G′ with
V (G) = V (G′ ) = [0, 1], there is a little subtlety. To “count” the edges to be edited,
we use the edge measure defined by (18.5); but these two graphings have different
edge measures, which edge measure to use? After a little thought, the solution is
natural:
d1 (G, G′ ) = ηG (E(G) \ E(G′ )) + ηG′ (E(G′ ) \ E(G)).
We note that ηG (E(G) ∩ E(G′ )) = ηG′ (E(G′ ) ∩ E(G)) (Exercise 18.26).
An easy inequality between the edit distance and sampling distances is stated
in the following proposition.
Proposition 19.1. For any two graphings G and G′ on the same underlying prob-
ability space and r ∈ N, we have
r
δ⊙ (G, G′ ) ≤ 2Dr d1 (G, G′ ),
and
δ⊙ (G, G′ ) ≤ 3d1 (G, G′ )1/ log(2D) .
In particular, these bounds hold for finite graphs.
Proof. Let S = E(G) \ E(G′ ) and S ′ = E(G′ ) \ E(G).
Claim 19.2. Let x be a random point in G, then the probability that the ball
BG,r (x) contains any edge in S is bounded by 2Dr ηG (S).
The number of points x for which BG,r (x) contains a given edge is bounded by
2Dr (this follows by an elementary computation). In the finite case, this implies
the Claim by an easy double counting. For graphings, this double counting can be
justified using the Mass Transport Principle for graphings, Proposition 18.49. For
 19.1. LOCAL CONVERGENCE AND LIMIT 353

( )
two nodes of G, let f (x, y) = degS (y)1 x ∈ BG,r (y) (where degS (y) denotes the
number of edges in S incident with y). Let x be a random point of V (G). Then
{ ( ) } ( ) ( ∑ )
λ x : E BG,r (x) ∩ S ̸= 0 ≤ E |BG,r (x) ∩ S| ≤ E degS (y)
y∈BG,r (x)
(∑ ) (∑ ) ( )
=E f (x, y) = E f (y, x) ≤ 2Dr E degS (x) = 2Dr η(S).
y y

This implies the Claim.

Applying the inequality in Claim 19.2 to S ′ as well, we see that with probability
at least 1 − 2Dr η(S) − 2Dr η ′ (S ′ ), the ball BG,r (x) contains no edge in S and the
ball BG′ ,r (x) contains no edge of S ′ . In this case these two balls are isomorphic,
proving that
( )
δ⊙r
(G, G′ ) = dvar BG,r (x), BG′ ,r (x′ ) ≤ 2Dr η(S) + 2Dr η ′ (S ′ ) = 2Dr d1 (G, G′ ).
This proves the first inequality.
⌈ The second inequality
⌉ follows from the first by
using (19.3) with r = − log(2d1 (G, G′ ))/ log(2D) . 
19.1.2. Locally convergent sequences. A sequence of graphs Gn with
v(Gn ) → ∞ is locally convergent if the r-neighborhood densities ρGn (F ) converge
for every r and every r-ball F . Similarly as for the subgraph sampling in the dense
case, there are equivalent frequency type parameters whose convergence could be
used instead of the neighborhood densities: we could stipulate the convergence of
t∗ (F, Gn ) for every connected graph F , as proved in Proposition 5.6, and also the
convergence of t∗inj (F, Gn ) or t∗ind (F, Gn ) for every connected graph F . All these
versions would lead to the same convergent sequences.
We can also describe convergent sequence of graphs as Cauchy sequences in the
sampling distance. From (19.3) it is easy to check that a graph sequence (Gn ) with
bounded degrees and with v(Gn ) → ∞ is convergent if and only if it is Cauchy
in the sampling distance. Of course, this is essentially just a reformulation of the
definition, and not a structural characterization of convergence as Theorem 11.3
was in the dense case.
For every Gn and every positive integer r, neighborhood sampling provides
a probability distribution ρGn ,r on the set Br of r-balls. By the definition of
convergence, for every fixed r, this distribution tends to a limit distribution σr .
The sequence of these distributions has some special properties. First of all, it is
consistent, in the sense that selecting a random r-ball from σr , and deleting from
it the nodes at distance r from the root, we get an (r − 1)-ball from distribution
σr−1 .
There is another, more subtle consistency property, which is a finite version of
involution invariance for a distribution on rooted countable graphs. Note that an
r-ball contains other (r − 1)-balls, centered at the neighbors of the original root,
and these from these balls we should also be able to recover σr−1 . Since there are
several of these in any given r-ball, we have to be a bit careful with the counting.
As done before, we bias the distribution by the degree of the root: for F ∈ Br ,
define
deg(F )σr (F )
σr∗ (F ) = ∑ ∗
.
H∈Br deg(H)σr (H)
Select a random r-ball F from σr∗ , and a random edge uv from the root u of F . We
can create two random (r − 1)-balls with a root edge: one, we delete from F the
354 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

nodes at distance more than r − 1 from the root u; two, we delete all the nodes at
distance more than r −1 from v, and consider v the root and vu the root edge. If we
get the same distribution on (r − 1)-balls with a root edge with both construction,
and this holds for every r ≥ 1, we say that the sequence (σ1 , σ2 , . . . ) is involution
invariant. To sum up, every convergent graph sequence gives rise to an involution
invariant and consistent probability measure on Br .
We have defined involution invariance for measures on the “graph of graphs”,
and of course the two notions are closely related. From every probability distribu-
tion σ on (G• , A), we get a probability distribution σr on Br by selecting a random
countable graph from σ and taking the r-ball about its root. It is trivial that this
sequence (σ1 , σ2 , . . . ) is consistent.
Conversely, from every consistent sequence (σ1 , σ2 , . . . ) we get a distribution σ
on (G• , A), by defining σ(G•F ) = σr (F ) for every r-ball F . It is also straightforward
to check that (σ1 , σ2 , . . . ) is involution invariant if and only if σ is.
So there is a bijective correspondence between consistent involution invariant
sequences (σ1 , σ2 , . . . ), where σr is a distribution on Br , and involution invariant
probability distributions on (G• , A). Through this correspondence, every locally
convergent graph sequence gives rise to an involution invariant distribution σ on
the sigma-algebra (G• , A). This is the Benjamini–Schramm limit or local limit of
the sequence.
By Theorem 18.37, it follows that there is a graphing G such that ρGn ,r → ρG,r
for every r ≥ 1. We write Gn → G, and say that this graphing “represents” the
limit; but one should be careful not to call it “the” limit; all locally equivalent
graphings represent the same limit object.
Example 19.3 (Cycles III). Consider the sequence of cycles (Cn ). It is easy
to see that the Benjamini–Schramm limit is the involution invariant distribution
concentrated on the two-way infinite path (with any node specified as the root).
The graphing Ca constructed in Example 18.1 represents the limit of this sequence
for any irrational number a. All connected components of this graphing are two-way
infinite paths, so generating a random point x ∈ [0, 1], its connected component
(Ca )x has the Benjamini-Schramm limit distribution.
Every graphing locally equivalent to Ca (i.e., in which almost all connected
components are two-way infinite paths) provides a representation of the limit object.
Example 18.54 shows two different graphings representing this limit. 
Example 19.4 (Grids). Let Gn be the n×n grid in the plane. The r-neighborhood
of a node v is a (2r + 1) × (2r + 1) grid (rooted in the middle), provided v is
farther than r − 1 from the boundary. This holds for (n − 2r)2 of the nodes, which
means almost all nodes if n → ∞. So in the weak limit, every r-neighborhood is
a (2r + 1) × (2r + 1) grid. Hence the Benjamini–Schramm limit of this sequence
is concentrated on the infinite square grid (with a root). We have seen (Example
18.38) how to represent this involution invariant distribution as a graphing. 

Example 19.5 (Penrose tilings). This is a more elaborate example, but interest-
ing in many respects. We can tile the plane with the two rhomboids of the left side
of Figure 19.1. This is no big deal, if we can use them periodically (for example,
as in the middle of Figure 19.1); but we put decorations on the edges, and impose
the restriction that these decorations must match along every common edge (as on
the right side of Figure 19.1); in particular, we are not allowed to combine two of
 19.1. LOCAL CONVERGENCE AND LIMIT 355

Figure 19.1. The Penrose rhomboids, an illegal tiling, and how

they should be attached.

the same kind into a single parallelogram. It turns out that you can tile the whole
plane this way (in fact, in continuum many ways), but there is no periodic tiling.
Figure 19.2 shows the graph obtained from a Penrose tiling of the plane. There
is a related (in fact, equivalent) version, in which we use two deltoids instead of
two rhomboids; such a tiling is also shown in Figure 19.2. A deltoid tiling can be
obtained from a rhomboid tiling by cutting up the rhomboids into a few pieces and
recombining these to form deltoids. (To figure out the details is left as a challenge
to the reader.)
One of the interesting (and nontrivial) features of such tilings is that every
one of them contains each of the two rhomboids with the same frequency. Similar
property holds for every configuration of rhomboids: if a finite configuration F of
tiles can be completed to a tiling at all, then this configuration occurs in every
Penrose tiling with the same frequency. To be precise, if we take a K × K square
about the origin in the plane, and count how many copies of F it contains, then
this number, divided by K 2 , tends to a limit if K → ∞. Moreover, this limit is
independent of the Penrose tiling that we are studying.

Figure 19.2. A piece of a Penrose rhomboid tiling and of a deltoid tiling.

We are not going to dive into the fascinating theory of Penrose tilings, but
point out that their basic properties can be translated into graph limits. Let Gn
be the graph obtained by restricting the graph of a Penrose rhomboid tiling to the
n×n square about the origin. The above properties of the Penrose tiling imply that
this sequence is convergent, and in fact it remains convergent if we interlace it with
a sequence obtained from a different Penrose tiling. In other words, these finite
pieces of any Penrose tiling converge to the same limit. The Benjamini–Schramm
limit will be not the original Penrose tiling, but a probability distribution on all
Penrose tilings. (This illuminates that in Example 19.4 of grids we end up with a
single limiting grid only because grids are periodic.)
356 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

A graphing representation of the limit of Penrose rhomboid tilings can be de-

scribed based on their characterization by de Bruijn [1981] (Figure 19.3); this was
pointed out by M. Bárász. 

Figure 19.3. A graphing describing the limit of Penrose rhom-

boid tilings. The underlying set is the union of parallel slices of a
rhombic icosahedron through its vertices. The edges of the graph-
ing are all translates of the edges of the polytope that connect two
points on these planes.

Example 19.6 (Large girth graphs). Let Gn be a sequence of D-regular graphs

whose girth (length of the shortest cycle) tends to infinity (it is well known that such
graph sequences exist). For every r ≥ 0 and sufficiently large n, the r-neighborhood
BGn ,r (v) of any node v is a rooted tree Tr,D of depth r, in which all the nodes
closer to the root than r have degree D. So the limiting sequence of distributions is
concentrated on these trees Tr,D . The Benjamini–Schramm limit of this sequence
is concentrated on a single countable graph, namely the D-regular tree (it does not
matter where we put the root). We have seen (Example 18.39) how to construct a
graphing representation of this involution invariant distribution. 
Example 19.7 (Various random D-regular graphs). Let G = G(n, D) denote
a random D-regular multigraph. This notion itself is a bit tricky. We can of course
define it as the uniform distribution over all D-regular graphs on n nodes; but
this definition is quite difficult to handle. A more useful definition is called the
configuration model. We start with a set S of nD nodes, partitioned into n sets
S1 , . . . , Sn of size D. We take a random perfect matching on S (we better assume
that nD is even), and then identify every Si into a single node labeled i. This way
we obtain a random D-regular multigraph. If we want a random simple graph, we
reject it if we get a graph with multiple edges, and try again.
It is easy to compute that the expected number of loops, as well as the number
of multiple edges in G(n, D) is bounded by a function of D. More generally, the
expected number of k-cycles is bounded by (D − 1)k /2k + o(1) (when n → ∞).
Hence it follows that for every fixed r and k, almost all nodes will be farther than r
from any cycle of length k or less. In other words, almost all r-neighborhoods will
be D-ary trees of depth r. So this sequence is locally convergent with probability
1, and its local limit is the same infinite D-regular tree as in the previous example.
Finally, random bipartite D-regular graphs will be interesting for us. These
can be generated just like above, except that we assume that n = 2m is even, and
we use a random perfect matching between S1 ∪ · · · ∪ Sm and Sm+1 ∪ · · · ∪ Sn .
 19.1. LOCAL CONVERGENCE AND LIMIT 357

By computations similar to the above, we can see that random D-regular bipartite
graphs tend to the same local limit as random D-regular graphs, the D-regular
rooted tree. 
19.1.3. Which distributions are limits? A big difference from the dense
case is that there is no easy way to construct a sequence of finite graphs that
converges to a given graphing (or involution invariant distribution). In fact, we
don’t know whether all involution invariant distributions arise as limit objects:
Conjecture 19.8 (Aldous–Lyons [2007]). Every involution invariant distribution
on (G• , A) is the limit of a locally convergent bounded-degree graph sequence.
Since every involution invariant distribution can be represented by a graphing
(Theorem 18.37), this is equivalent to asking whether every graphing is the local
limit of a locally convergent sequence of bounded-degree graphs. This conjecture,
which is a central unsolved problem in the limit theory of bounded-degree graphs,
generalizes a long-standing open problem about sofic groups. It is known in some
special cases: when the distribution is concentrated on trees (Bowen [2004], Elek
[2010b]; see Exercise 19.12), and also when the graphing is “hyperfinite” (to be
discussed in Section 21.1).
The following is an interesting reformulation of this conjecture. Let Ar ⊆ RBr
denote the set of all probability distributions ρG,r , where G ranges through all finite
graphs. Let A′r ⊆ RBr denote the set of probability distributions ρG,r , where G
ranges through all graphings. Equivalently, A′r consists of probability distributions
on Br induced by an involution invariant probability distribution on G• . Clearly
Ar ⊆ A′r .
Proposition 19.9. (a) The closure Ar of Ar is a compact convex set. (a) A′r is a
compact convex set.
While most of the time the limit theory of graphs with bounded degree is more
complicated than the dense theory, Proposition 19.9 represents an opposite case: in
the dense case, even the set D2,3 discussed in Section 16.3.2 was non-convex with
a complicated structure.
Proof. (a) Let G1 and G2 be two finite graphs, and consider the graph G =
v(G ) v(G )
G1 2 G2 1 consisting of v(G2 ) copies of G1 and v(G1 ) copies of G2 . Then
1( )
ρG,r (B) = ρG1 ,r (B) + ρG2 ,r (B)
2
for every r-ball B. This implies that Ar is convex. Since it is a bounded closed set
in a finite dimensional space, it is compact.
(b) The fact that A′r is closed follows from general considerations: the set M of
involution-invariant measures, as a subset of the set of all probability measures on
the compact metric space G• , is closed in the weak topology, and so it is compact.
Using that each of the cylinders G•F is open-closed, the projection of M onto RBr is
continuous, and hence the image, which is just A′r , is compact. The convexity of
A′r follows by a construction similar to that in (a). 
The Aldous–Lyons Conjecture is equivalent to saying that Ar = A′r for every r.
So if the conjecture fails to hold, then there is an r ∈ N and a linear inequality on
RBr that is valid for Ar but not for A′r . This would be a linear inequality between
358 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

r-neighborhood densities that holds for every finite graph, but fails to hold for all
graphings, a “positive” consequence of a “negative” fact.
There is a finite version of the Aldous–Lyons conjecture, which was raised by
this author at a conference, and was proved, at least in a non-effective sense, quickly
by Alon [unpublished]:
Proposition 19.10. For every ε > 0 there is a positive integer n such that for
every graph G ∈ G there is a graph G′ ∈ G such that v(G′ ) ≤ n and δ⊙ (G, G′ ) ≤ ε.
Proof. Let r = ⌈log(2/ε)⌉, and let G1 , . . . , Gm be any maximal family of
graphs in G such that δ⊙ r
(Gi , Gj ) > ε/2 for all 1 ≤ i < j ≤ m. Such a family is
finite, since every graph is represented by a point in Ar , which is a bounded set
in a finite dimensional space, and these points are at least ε/2 apart in the total
variation distance. It follows that n = maxi v(Gi ) is finite. By the maximality of
the family, for every graph G there is an i ≤ m such that δ⊙ r
(G, Gi ) ≤ ε/2. We
have v(Gi ) ≤ n, and by (19.3)
1 1 ε
δ⊙ (G, Gi ) ≤ + dr⊙ (G, Gi ) ≤ r + ≤ ε. 
2r 2 2
Unfortunately, no effective bound on n follows from the proof (one can easily
get an explicit bound on m, the number of graphs in the representative family, but
not on the size of these graphs). It would be very interesting to give any explicit
bound (as a function of D and ε), or to give an algorithm to construct H from
G. Ideally, one would like to design an algorithm that would work locally, in the
sampling framework, similarly to the algorithm in Section 15.4.2 in the dense case.
Proposition 19.10 is related to the Aldous–Lyons Conjecture 19.8. Indeed, the
Aldous–Lyons Conjecture implies that for any graphing G there is a finite graph
G whose neighborhood distribution is arbitrarily close; Proposition 19.10 says that
for any finite graph G there is a finite graph H of bounded size whose neighborhood
distribution is arbitrarily close. Suppose that we have a constructive way of finding,
for an arbitrarily large graph G with bounded degree, a graph H of size bounded
by a function of r and ε that approximates the distribution of r-neighborhoods in
G with error ε. With luck, the same construction could also work with a graphing
in place of G, proving the Aldous–Lyons Conjecture.
One route to disproving the Aldous–Lyons Conjecture could be to explicitly find
the sets Ar and A′r for some r, and see that they are disjoint. Since the dimension
of Ar grows very fast with r, it seems useful to consider even simpler questions.
Instead of looking at Ar and A′r , we could fix a finite set {F1 , . . . , Fm } of simple
graphs, assign the vector (t∗ (F1 , G), . . . , t∗ (Fm , G)) to every graph G ∈ G, and
consider the set T (F1 , . . . , Fm ) of all such vectors. We define the set T ′ (F1 , . . . , Fm )
analogously, replacing graphs by graphings. By the same argument as above, the
sets T (F1 , . . . , Fm ) and T (F1 , . . . , Fm ) are convex. The Aldous–Lyons Conjecture
is equivalent to saying that T (F1 , . . . , Fm ) = T (F1 , . . . , Fm ) for every F1 , . . . , Fm .
This leads us to the problem, very interesting on its own right, to determine the
sets T (F1 , . . . , Fm ) and T (F1 , . . . , Fm ), and more generally, to extremal problems
for bounded degree graphs. This should be the title of a chapter, but very little
has been done in this direction. There are, of course, many results in extremal
graph theory that concern graphs with bounded degree; but the limit theory of
bounded degree graphs has not been applied to extremal graph theory in a sense
in which the limit theory of dense graphs has been. One notable exception is the
 19.1. LOCAL CONVERGENCE AND LIMIT 359

result of Harangi [2012], who determined the sets T (K3 , K4 ) and T (K3 , K4 ) for
D-regular graphs. He found the same answer in both cases (so this did not give a
counterexample to the conjecture).

19.1.4. On colored graphs. It will be useful at various points to extend our

constructions and results to colored graphs, where the nodes are colored with b
node-colors and the edges are colored with c edge-colors (where b and c are fixed
positive integers). Colored graphings can be defined analogously, where every node
set and edge set with a given color is Borel. Colored graphs and graphings can be
used to express some properties and additional structures which we want to pass to
the limit. For example, we can express measure preserving families (used in Section
18.2.1 to certify that a measurable graph is measure preserving) by edge-coloring.
We could have formulated all our arguments in the previous chapter and this
one in the more general context of colored graphs and graphings. Alternatively, we
could repeat these arguments now for colored graphs. We will do neither; we point
out in a few sentences how these generalizations would work, and leave it to the
interested reader to think through that the arguments can be extended to colored
graphs.
Sampling from a colored graph results in a distribution of colored balls, and
since there is only a finite number of them, all the arguments above remain valid.
We can extend the notion of convergence to colored graphs of a fixed type (b, c),
i.e., to graphs that are node-colored with b colors and edge-colored with c colors.
The sampling process returns a colored r-ball, which is node-colored with b colors,
edge-colored with c colors, and has a specified root. As before, we denote by ρG,r
the probability distribution on colored r-balls about a random node (where the type
(b, c) is understood). We say that the colored graph sequence is locally convergent,
if the sequence (ρGn ,r (F ) : n = 1, 2, . . . ) converges for every r and every colored
r-ball F .
We can define the “graph of colored graphs”: its nodes will be all connected
colored rooted countable graphs (with the same degree bound as always). Adja-
cency is defined as before; we color the node (H, v) with the color of v, and we color
the edge (H, v)(H, u) with the color of the edge vu.
Every convergent colored graph sequence has a limit object in the form of
an involution invariant probability distribution on the “graph of colored graphs”,
which in turn can be represented by a colored graphing.
One could go a step further, and decorate every node and/or every edge by
an element of a fixed compact Hausdorff space K. (For the dense case, a similar
extension was treated in Section 17.1.) One could extend the notions of involution
invariant distributions and measure preserving graphs to this case, but it would
take more effort, and would have fewer applications. One example of an application
would be the assignment of weights α to the nodes of graphs in G• in the proof
of Theorem 18.37, which could be phrased as using a node-decoration from the
compact space [0, 1]. We will use this more general construction in the next section,
where colored graphs will play and important role.
Exercise 19.11. Let G ∈ G and let S ⊆ E(G), |S| = εv(G). Prove that δ⊙ (G, G\
S) ≤ 4ε1/ log D .
Exercise 19.12. Prove that if σ is an involution-invariant distribution such that
a rooted graph chosen from σ is almost always a tree, then σ is the local limit of
a finite graph sequence (Elek [2010b]).
360 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

Exercise 19.13. Prove that merging two node colors or two edge colors, every
convergent colored graph sequence remains convergent.

19.2. Local-global convergence

Are the notion of convergence and the limit object constructed above informa-
tive enough? The limit graphon of a dense sequence of graphs contains very much
information about the asymptotic properties of the sequence. This is not quite
so for the bounded degree case, unfortunately. Let us illustrate this by a simple
example.

Example 19.14. Let Gn be a sequence of random 3-regular bipartite graphs. Let

Hn consist of two disjoint copies of Gn . The Benjamini–Schramm limit of both
sequences is a distribution concentrated on a single 3-regular rooted tree T3 .
This limit graphing is not uniquely determined. Among others, we have
Bernoulli graphing T3 associated with T3 , but one could take the disjoint union T23
of two such graphings (with the node measure scaled down by 2). It seems that
T3 represents the limit of Gn “better”, while T23 represents the union the limit
of Hn “better”. As another example, if we consider the free group F3 with three
generators acting without fixed points on a probability space, then the correspond-
ing graphing (obtained by connecting every point to its images under any of the
generators) represents the Benjamini–Schramm limit. One feels that the limit of
the sequence (Gn ) is ‘better” represented if the action of the free group is ergodic,
while for the limit of Hn , the space should be split into two invariant subsets of
measure 1/2. 

This example suggests that in the limit object, the underlying σ-algebra carries
combinatorial information. This is in stark contrast with the dense case (cf. Remark
10.1 and the discussion in that section).
In this section we define a notion of convergence for graphs with bounded
degree that is stronger than the local convergence (Hatami, Lovász and Szegedy
[2012]). Among others, if a sequence of graphs is convergent in this stronger sense,
then we can read off from the limit whether the graphs are expanders (up to a
non-expanding part of negligible size).

19.2.1. Nondeterministic sampling distance. First, we define a new ver-

sion of the sampling distance. Let G1 , G2 ∈ G, then their non-deterministic sam-
pling distance of depth r for k colors is defined as the least c > 0 with the following
property: for(every k-coloring α ) 1 of V (G1 ) there exists a k-coloring α2 of V (G2 )
such that δ⊙ r
(G1 , α1 ), (G2 , α2 ) ≤ c, and vice versa. (The sampling distance of
(G1 , α1 ) and (G2 , α2 ) means their sampling distance as colored graphs.) We denote
(r,k)
this distance by δ⊙ (G1 , G2 ). We then take, similarly as before,
∞ ∑
∑ ∞
nd 1 (r,k)
δ⊙ (G1 , G2 ) = δ (G1 , G2 ).
2r+k ⊙
k=0 r=0

It is easy to see that these formulas define a metric on finite graphs.

We can define the non-deterministic sampling distance of two graphings G1 and
G2 similarly, except that we only allow Borel k-colorings, and have to use infimum
 19.2. LOCAL-GLOBAL CONVERGENCE 361

instead of a minimum:
(r,k) { ( )
(19.5) δ⊙ (G1 , G2 ) = inf c : ∀α1 ∃α2 δ⊙
r
(G1 , α1 ), (G2 , α2 ) ≤ c, and
( ) }
∀α2 ∃α1 δ⊙
r
(G1 , α1 ), (G2 , α2 ) ≤ c .
The quantity δ⊙ nd
(G, G′ ) is defined from this just like in the case of graphs.
We say that two graphings G and G′ are locally-globally equivalent, if
δ⊙ (G, G′ ) = 0. A sequence of graphs (Gn ) is locally-globally convergent if it is
nd

a Cauchy sequence in the nondeterministic distance, i.e., δ⊙ nd

(Gn , Gm ) → 0 as
n, m → ∞. We say that its local-global limit is graphing G, if δ⊙ nd
(Gn , G) → 0.
nd k,r
It is clear that we could replace d by d in these definitions, and require the
conditions for all k, r ≥ 1.
We have defined nondeterministic distance and local-global equivalence in terms
of coloring the nodes. We could allow coloring of the edges as well without changing
the notion of equivalence. Let me elaborate this for local-global equivalence.
Proposition 19.15. Suppose that two graphings G and G′ are locally-globally
equivalent. Then for any ε > 0, k ≥ 1, and Borel k-edge-coloring α and Borel
k-point-coloring β of G, there (exists a Borel k-edge-coloring
) α′ and Borel k-point-
′ ′ ′ ′ ′
coloring β of G such that δ⊙ (G, α, β), (G , α , β ) ≤ ε.
Proof. We want to encode the edge-coloring into a node-coloring. The first
trick is to construct (independently of the coloring α) another Borel edge-coloring γ
with 2D −1 colors such that no two adjacent edges have the same γ-color (Theorem
18.4).
( Using
) this, we define a point-coloring ζ: (we color a point) x with the pair
β(x), σ(x) , where σ(x) is the set of all pairs α(x, y), γ(x, y) , where (x, y) ∈
E(G). This point-coloring uses a finite set K = [k] × 2[k]×[2D−1] of colors. From
this point-coloring, we can recover the original coloring easily: for every point x,
β(x) is the first element of ζ(x), and for every edge (x, y), σ(x) ∩ σ(y) has a unique
element (a, b), and the original color of (x, y) was a. (The only role of the coloring
γ was to make sure that this common element of σ(x) and σ(y) is unique.)
By the definition of local-global equivalence, there is a K-coloring ζ ′ of the
points of G′ such that δ⊙ ((G, ζ), (G′ , ζ ′ )) ≤ ε. We define an edge-coloring α′ and a
point-coloring β ′ of G′ as follows. Since ζ(x) ∈ K, we can write ζ(x) = (b(x), s(x)),
where b(x) ∈ [k] and s(X) ⊆ [k] × [2D − 1]. Let β ′ (x) = b(x); let α′ (x, y) be the
smallest a ∈ [k] for which there is a g ∈ [2D − 1] such that (a, g) ∈ s(x) ∩ s(y); if no
such g exists, then let α′ (x, y) = 1 (this will happen only for a small set of edges).
Now comes the key observation: Whenever the r-neighborhoods of point x in
(G, ζ) and of point y in (G′ , ζ ′ ) are isomorphic, then the r-neighborhoods of x in
(G, α, β) and of x′ in (G′ , α′ , β ′ ) are also isomorphic. The rules for obtaining α
and β from ζ and α′ and β ′ from ζ ′ work the same way in both graphs. Hence
δ⊙ ((G, α, β), (G′ , α′ , β ′ )) ≤ δ⊙ ((G, ζ), (G′ , ζ ′ )) ≤ ε. 

19.2.2. Graphings as local-global limits. We have seen that limits of lo-

cally convergent graph sequences can be described as involution-invariant distribu-
tions, and this representation of the limit is unique. We could also represent the
limit by a graphing, but this was not unique, which means that graphings are more
complicated objects than necessary. Why bother with graphings at all, why not use
involution invariant distributions only? One justification for considering graphings
is the following result of Hatami, Lovász and Szegedy [2012].
362 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

Theorem 19.16. For every locally-globally convergent sequence of finite graphs

there is a graphing that is its local-global limit.
For the proof, we need a couple of lemmas about k-colorings. First, let us
discuss continuous k-colorings of a graphing. For this to make sense, we have to
fix a topology on V (G). Of course, we should not use the standard topology of
(say) [0, 1]: this would not admit nontrivial k-colorings. But if we use the local
topology as defined in Section 18.1, we get interesting continuous colorings. Recall
that this topology can be defined by a metric, where two points are 2−r -close if
their r-neighborhoods are isomorphic. This topology is totally disconnected, so
there will be nontrivial continuous k-colorings. A k-coloring will be continuous if
an only if the color of a node can be determined just from the isomorphism type of
its r-neighborhood for some finite r.
We know by Lusin’s Theorem that every Borel function on a compact prob-
ability space can be approximated by a continuous function γ in the sense that
the set {α ̸= γ} has arbitrarily small measure. Here, in general, γ will not be a
k-coloring (if the underlying space of G is the unit interval, for example, then its
range will be an interval, not [k]). However, with an appropriate topology on V (G),
the approximating function can be chosen to be a coloring itself.
Lemma 19.17. Let K be a compact metric space that is totally disconnected, let
π be a probability measure on K, and let α : K → [k] be a Borel k-coloring of
K. Then for every ε > 0 there is a continuous k-coloring δ : K → [k] such that
π{α ̸= δ} ≤ ε.
Proof. By Lusin’s Theorem, there is a continuous function β on K such that
T = {x ∈ K : α = β} has measure at least 1 − ε. Open-closed sets in K separate
any two points, hence by the Stone–Weierstrass Theorem, there is a stepfunction
γ whose steps are open-closed (i.e., γ is continuous), and |β(x) − γ(x)| < 1/3 for
every x. If a step S of γ contains a point y ∈ T , then we fix one such point, and
define δ(x) = α(y) = β(y) for all x ∈ S; else, we define δ(x) = 1. This way we get
a continuous k-coloring δ.
Let x ∈ T and let S be the step of γ containing x. Then α(x) = β(x), and so,
for the point y ∈ S ∩ T used in the definition of δ(x) (which may or may not be x),
we have
2
|α(x) − δ(x)| ≤ |β(x) − γ(x)| + |γ(x) − δ(x)| = |β(x) − γ(x)| + |γ(y) − β(y)| ≤ .
3
Since α(x) and δ(x) are integers, this implies that α(x) = δ(x) for x ∈ T . 

The second lemma we need is similar to Proposition 19.10. It shows that a

uniformly bounded number of k-colorings can approximate all k-colorings (in the
sense of neighborhood statistics) of an arbitrarily large graph.
Lemma 19.18. For every k, r ≥ 1 and ε > 0 there is an integer M = M (k, r, ε) ≥ 1
such that every graph G ∈ G has M k-colorings α1 , . . . , αM such that for every k-
coloring β of G there is an i (1 ≤ i ≤ M ) with
( )
r
δ⊙ (G, β), (G, αi ) ≤ ε.
Of course, M depends on D too, but this is tacitly assumed to be a constant
in all our discussions in this part.
 19.2. LOCAL-GLOBAL CONVERGENCE 363

Proof. Let {F1 , . . . , FM } be any maximal family of k-colored graphs in G such

that δ⊙r
(Fi , Fj ) > ε/2 for all 1 ≤ i < j ≤ M . Since the distributions ρG,r on
k-colored r-balls B belong to a bounded set in a finite dimensional space, such a
family is finite.
Let(G be any )finite graph. For every i ≤ m, select a k-coloring αi of G such
that δ⊙r
(G, αi ), Fi ≤ ε/2, if such a k-coloring exists (call such an i relevant); else,
let αi be an arbitrary k-coloring of G. We claim that the k-colorings αi constructed
this way have the required property. ( )
For every k-coloring β of G, there is an i ≤ M such that δ⊙ r
(G, β), Fi ≤ ε/2,
by the maximality of the family {F1 , . . . , FM }. Then clearly this i is relevant, and
so for the corresponding αi , we have

( ) ( ) ( ) ε ε
r
δ⊙ (G, β), (G, αi ) ≤ δ⊙
r r
(G, β), Fi + δ⊙ Fi , (G, αi ) ≤ + = ε. 
2 2

Proof of Theorem 19.16. We apply Lemma 19.17 with ε = 2−r , and denote
M (k, r, 2−r ) by M (k, r). We fix a set of M (k, r) k-colorings as in Lemma 19.17 for
every graph G ∈ G, and call them its representative
∏∞ k-colorings.
Consider the product space K = k,r=1 [k]M (k,r) ; this is compact and totally
disconnected. We start with constructing a decoration χ = χG : V (G) → K for
every G ∈ G. Given a node v ∈ V (G), we consider the representative k-colorings
α1 , . . . , αM (k,r) of G, and concatenate the sequences (α1 (v), . . . , αM (k,r) (v)) for
k, r = 1, 2, . . . to get χ(v).
Using the decoration χG and the projection map φk,r : K → [k]M (k,r) , we can
manufacture many k-colorings of G as β = ψ ◦ φk,r ◦ χ, where ψ : [k]M (k,r) → [k]
is any map. We call these k-colorings “special”. It follows from the construction
of χ that the representative k-colorings of G are special. Hence for every graph G,
every k, r ≥ 1, and every k-coloring
( α of)V (G), there is a special k-coloring β close
to α, in the sense that δ⊙ (G, α), (G, β) ≤ 2−r .
The graphing HK we construct is similar to the “Graph of Weighted Graphs”
+
H introduced in Section 18.3.3, but instead [0, 1], we use weights from K. We
construct probability measures on HK to get representations of finite graphs and
then, representations of the limit. With the decoration χG , and any choice of a root
v ∈ V (G), the triple (G, v, χG ) is a point of HK . The map τG : v 7→ (G, v, χG )
defines an embedding G → HK onto a connected component of HK (the fact that
this map is injective is clear, since for any two nodes u, v ∈ V (G) one of the k-
colorings in Lemma 19.18 must distinguish
( ) them once r is large enough). Let ζG
be the uniform distribution on τG V (G) . Since G is finite, this distribution is
involution-invariant on HK .
Let (Gn ) be a locally-globally convergent graph sequence. By Prokhorov’s
Theorem (see Appendix A.3.3), we can replace our graph sequence by a subsequence
such that the distributions ζGn converge weakly to a distribution ζ on HK . Since
every ζGn is involution-invariant, so is ζ, and hence G = (HK , ζ) is a graphing.
We claim that Gn → G in the local-global sense. To prove this convergence,
we need the following auxiliary fact.

Claim 19.19. Let β be a continuous (k-coloring of G, )and let βn = β ◦ τGn be the

k-coloring it induces on Gn . Then δ⊙ (Gn , βn ), (G, β) → 0 as n → ∞.
364 19. CONVERGENCE OF BOUNDED DEGREE GRAPHS

To prove this, fix r ≥ 1, and express the frequency of a k-colored r-ball B0 in

(G, β) as an integral:
∫
( )
ρG,β,r (B0 ) = 1 BG,β,r (x) ∼
= B0 dζ(x).
HK

By the definition of βn , we have a similar expression for the frequency of B0 in

(Gn , βn ):
∫
( )
ρGn ,βn ,r (B0 ) = 1 BG,β,r (x) ∼
= B0 dζGn (x).
HK

The main observation we need is that the integrand is continuous. Indeed, suppose
that Hn → H in the topology of HK , where Hn , H ∈ HK are rooted K-decorated
countable graphs. Then for a sufficiently large n, the balls BHn ,r and BH,r are
isomorphic, ( and) moreover, there is an isomorphism σn : BH,r → BHn ,r such
that(χHn σ)n (x) → χH (x) for every x ∈ V (BH,r ). This means that χH (x) and
χ(Hn σn (x)
) agree in more and more coordinates as n grows, which implies that
β σn (x)( → )β(x), since β is continuous. Since β has finite
( range, this implies )
that β σ (x) = β(x) if n is large enough. But then 1 B (H ) ∼
= B =
( n ) G,β,r n 0
∼
1 BG,β,r (H) = B0 if n is large enough, which proves that the integrand is contin-
uous.
Hence it follows by the weak convergence ζGn → ζ that
∫ ∫
1(BG,β,r = B0 ) dζGn −→ 1(BG,β,r ∼
∼ = B0 ) dζ,
HK HK

which proves that ρGn ,βn ,r (B0 ) → ρG,β,r (B0 ) for every k-colored r-ball B0 . This
proves the claim.
Let us return to the proof of the local-global convergence Gn → G. By the
definition of the nondeterministic sampling distance, we have to verify two things
for every r, k ≥ 1: every k-coloring of Gn can be “matched” by a k-coloring of G so
that the distributions of r-neighborhoods are close, and vice versa. Let ε > 0; we
may assume that ε ≥ 2−r , since larger neighborhoods are more difficult to match.
First, let α be a Borel k-coloring of G. Then by Lemma 19.17, there is another
K
Borel k-coloring β such that β is continuous in the ( topology of )H and α = β on
−r
a set of measure at least 1 − ε(2D) . Then δ⊙ (G, α), (G, β) ≤ ε/2 by (19.4).
r

For every n, the k-coloring β gives a k-coloring ( βn of the nodes ) of Gn , under the
embedding τGn . By Claim 19.19, ( we have δ ⊙ )(G n , β n ), (G, β) ≤ ε/2 if n is large
enough. This implies that δ⊙ r
(Gn , βn ), (G, α) ≤ ε.
(r,k)
Second, let n be large enough so that for all m ≥ n, we have δ⊙ (Gn , Gm ) ≤
ε/3, and let αn be a k-coloring
( of Gn . Then ) for every m ≥ n there is a k-coloring
αm of Gm such that δ⊙ r
(Gn , αn ), (Gm , αm ) ≤ ε/3. Furthermore, there is a special
( = ψm ◦ φk,r ◦ χGm of → )[k])
M (k,r)
k-coloring βm ) Gm (with an appropriate ( ψm : [k]
such that δ⊙ (Gm , αm ), (Gm , βm ) ≤ ε/3. It follows that δ⊙ (Gn , αn ), (Gm , βm ) ≤
r r

2ε/3. We can select an infinite subsequence such that ψm = ψ is independent of

m, so that βm (v) depends only on the decoration χGm (v) of the node v ∈ V (Gm ).
) ψ to get an k-coloring of G: we color every x ∈ H
K
We can use the( same map
with β(x) = ψ φr (χ(x)) , where χ(x) is the decoration of root(x). This coloring
 19.2. LOCAL-GLOBAL CONVERGENCE 365

is continuous,
( ) τGm (v) it coincides
and on ( with βm (v).) Claim (19.19) implies that
δ⊙r
(Gm , βm ), (G, β) → 0. Hence δ⊙ r
(Gn , αn ), (G, β) ≤ ε if n is large enough. 
Exercise 19.20. Let F1 and F2 be two finite graphs and let GF1 and GF2 denote
nd nd
the associated graphings. Prove that δ⊙ (F1 , F2 ) = δ⊙ (GF1 , GF2 ).
Exercise 19.21. For an (uncolored) graph G, let Qr,k (G) denote the set of all
neighborhood distributions ρG∗ ,r , where G∗ is a k-colored version of G. Prove
that ( )
δ⊙ (G, G′ ) = dHaus Qr,k (G), Qr,k (G′ ) .
(r,k)
var
Exercise 19.22. (a) Let (Gn ) be a locally-globally convergent graph sequence.
Prove that the numerical sequences α(Gn )/v(Gn ) and Maxcut(Gn )/v(Gn ) are con-
vergent. (b) Show by an example that this does not hold for every locally conver-
gent sequence.
 CHAPTER 20

Right convergence of bounded degree graphs

20.1. Random homomorphisms to the right

Homomorphisms from a large bounded-degree graph G into a small weighted
graph H are the bread and butter of statistical physics, as we have illustrated in
the Introduction (Chapter 2.2). What happens if we go to the limit with G through
bounded degree graphs? Does it make sense to talk about a random homomorphism
from a countable graph into a weighted graph H? Or from a graphing? It is natural
that statistical physicists have worked out a theory that is able to answer these
questions. In this section we reproduce some of these results. We will need these,
among others, in the next section, where we discuss right-convergence.
To start with a trivial example, let G be a countably infinite graph with
bounded degree, and let H be the looped complete graph Kq◦ . Then we can map
the nodes of G independently, which defines a perfectly fine probability distribution
on maps V (G) → V (H). Unfortunately, if we delete any edge from Kq◦ , then the
probability that a random map V (G) → [q] remains a homomorphism is 0. So
we could not define a random homomorphism G → H by taking a random map
V (G) → V (H) and condition on its being a homomorphism. It turns out that ran-
dom homomorphisms from countable graphs into weighted graphs can be defined
in some cases: when the maximum degree of G is small and the edgeweights of H
are close to 1. (We will not attempt to define a random homomorphism from a
graphing.)
It turns out that the construction for a random homomorphism G → H for
infinite graphs G is made possible by another important phenomenon, this time for
finite graphs. In its simplest version, let u and v be two nodes of G that are far
from each other, and consider a random homomorphism G → H. Are the images of
u and v essentially independently distributed? This is not always so; for example,
if G is a connected bipartite graph, then there are two homomorphisms G → K2 ,
and the image of one node determines the images of all of the others. We will start
with showing that under similar conditions as we mentioned above (the maximum
degree of G is small and the edgeweights of H are close to 1), the images of distant
nodes will be essentially independent. This important result, called the Dobrushin
Uniqueness Theorem, will be stated and proved first. There is of course a lot more
in the literature about this theorem and its applications, see e.g. Georgii [1988] or
Simon [1993]. (We have to postpone the explanation of the word “uniqueness” to
the end of the next section.)
Since we have graphons at hand, we can replace H by a graphon W and get
more general results with only a small amount of additional hassle.

20.1.1. Homomorphisms and Markov chains. We start with defining ran-

dom homomorphisms from a finite bounded degree into a weighted graph and into
367
368 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

graphon. Let G = (V, E) be a simple graph and let H be a weighted graph with
nonnegative edgeweights.
We have considered random maps φ : V (H) → V where the probability of φ
was proportional to αφ . It is also quite natural to bias these with the product of
the edgeweights. In other words, let the probability of φ be
homφ (G, H)
(20.1) πG,H (φ) = αφ .
hom(G, H)
In the special case when H is a looped-simple unweighted graph, this is the uniform
distribution on the set Hom(G, H).
Example 20.1 (Ising model). Recall the example from the Introduction (Section
2.2). There is a very large graph G (most often, a grid) whose nodes are the atoms
and whose edges are bonds between these atoms. There is a small graph H, whose
nodes represent the possible states of an atom. (In the case of the Ising model, H
has two nodes only, representing the spins “UP” and “DOWN”.) The nodeweights
αi = e−hi represent the influence of an external field on an atom in state i, and
the edgeweights βij = e−Jij represent the interaction energy between two adjacent
atoms in states i and j (we ignore the dependence on the temperature for this
discussion). A possible configuration is a map σ : V (G) → V (H), and its energy is
∑ ∑
H(σ) = − hσ(u) − Jσ(u),σ(v) .
u uv∈E(G)

In the introduction we were focusing on the partition function Z of the system,

which turned out to be hom(G, H). But the exponential of the energy
e−H(σ) = ασ homσ (G, H)
is also very important, because the system will be in state σ with probability
e−H(σ) ασ homσ (G, H)
= = πG,H (σ).
Z hom(G, H)
So this distribution on homomorphisms introduced above expresses the fundamental
physical state of a material. 
It is not hard to generalize these constructions to the case when we replace H
by a graphon W ̸= 0 on [0, 1] (and of course the formulas become simpler). We will
call maps V → [0, 1], “weightings”. For a measurable set X ⊆ [0, 1]V , we define
∫
tx (G, W ) dx
(20.2) πG,W (X) = X .
t(G, W )
We call a random map drawn from the distribution πG,W a random W -weighting
of G. We could of course replace W by a weakly isomorphic graphon on some
other probability space, and this could be more natural in some cases (think of the
generalization of the Ising model, where the spins can be arbitrarily unit vectors in
R3 ).
Let S ⊆ V , X ⊆ [0, 1]S and suppose that we fix a partial weighting y ∈ [0, 1]V \S .
We can define a kind of conditional distribution
∫
ty,x (G, W ) dx
(20.3) πy (X) = X
ty (G, W )
 20.1. RANDOM HOMOMORPHISMS TO THE RIGHT 369

(The condition x|S = y may have probability 0, but the formula works.) In the
special case when S = V \ {v}, the distribution πy can be identified with a distri-
bution on [0, 1], which we denote by πy,v . It will be important to notice that in this
case the distribution πy is determined by the restriction of y to NG (v).
Is there a more tangible way of defining this distribution? A general tech-
nique of generating random elements of complicated distributions and studying
their properties is to construct a Markov chain with the given stationary distribu-
tion. In this case, there is a rather simple Markov chain M on weightings in [0, 1]V
with this property. (In the special case when W = WKq , this will specialize to
the “heat-bath” chain, or “Glauber dynamics”, on q-colorings of G.) One step of
this Markov chain is described as follows: Given a weighting x, we select a uniform
random node v ∈ V (which we call the pivot node) and reweight it from the distri-
bution πx,v . All other nodeweights remain unchanged. It is not hard to check that
πG,W is a stationary distribution of this Markov chain.
Let us fix a set U ⊆ V and its complement Z = V \ U . We can modify
the Markov chain M by selecting the pivot node v from Z only. This modified
Markov chain preserves the weighting of U ; if we restrict it to the extensions of a
partial weighting a ∈ [0, 1]U , then we get a Markov chain Ma , whose stationary
distribution is πa .
Next, we define a Markov chain M2 on pairs (x, y) ∈ [0, 1]V × [0, 1]V . Given
(x, y), we generate a random pivot node v ∈ Z and modify both x and y according
to M, separately but not independently: using the same pivot node, we generate
a random weight x from the distribution πx,v , and a random weight y from the
distribution πy,v , and couple x and y optimally, so that P(x ̸= y) = dtv (πx,v , πy,v ).
We change the weight of v in x to x, and in y to y. Note that for fixed a, b ∈ [0, 1]U ,
the set of pairs of weightings (x, y) with x|U = a and y|U = b is invariant. Let Ma,b
denote the Markov chain restricted to such pairs.
The stationary distribution of this Markov chain is difficult to construct di-
rectly, but at least it exists:
Lemma 20.2. The Markov chain Ma,b has a stationary distribution with marginals
πa and πb .
This is trivial if Ma,b has a finite number of states (which happens if W is a
stepfunction, i.e., we are studying homomorphisms into a finite weighted graph).
For the general case, the proof follows by more advanced arguments in probability
theory, and is not given here (see Lovász [Notes]).
These Markov chains (especially the simplest chain M) are quite important in
simulations in statistical physics and also in theoretical studies. A lot of work has
been done on their mixing times and other properties. For us, however, the main
consequence of their introduction will be the existence of the stationary distribution
of Ma,b .
20.1.2. Correlation decay. Our next goal is to state and prove the fact,
mentioned above, that πG,W has no long-rage interaction: under appropriate con-
ditions, the weights of two distant nodes in a random W -weighting from πG,W are
essentially independent. We start with an easy observation (the verification is left
to the reader as an exercise).
Proposition 20.3. Let G = (V, E) be a simple graph, and let W be a graphon
of rank 1. Then πG,W is a product measure on [0, 1]V . In other words, if x is a
370 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

random W -weighting of G, then the weights xu (u ∈ V (G)) are independent (as

random variables).
The Dobrushin Uniqueness Theorem, in its combinatorial form, will imply that
if the adjacency matrix of H is close to a rank-1 matrix, then there is almost no
correlation between xu and xv (where x is a random W -weighting of G), provided
the degrees of G are small and the distance of u and v is large. In fact, the theorem
is stronger: it implies that conditioning on the weights of all nodes far from a given
set U ⊆ V has very little influence on the weighting of U .
To state this result, we need the following parameter of graphons. Let r ≤ D,
and consider the star Sr+1 on {0, 1 . . . r} with center 0. Let us define the Dobrushin
value of W as
(20.4) dob(W ) = sup sup dtv (πx , πy ),
r≤D x,y

where x, y ∈ [0, 1][r] range through weightings of the leaves of Sr+1 that differ only
for a single node u ∈ [r]. If dob(W ) is small, then changing the weight of a neighbor
of a node has little influence on the weight of the node. Changing the weight of
neighbors one by one, we get by induction that for any graph G ∈ G, node v ∈ V (G)
and x, y ∈ [0, 1]V , we have
 
(20.5) dtv (πx,v , πy,v ) ≤ dob(W ){u ∈ N (v) : xu ̸= yu }.
Theorem 20.4. Let G = (V, E) be a (finite) graph with all degrees bounded by D,
and let W be a graphon. Then for any partition V = Z ∪ U and any two maps
a, b ∈ [0, 1]U , the distributions πa and πb have a coupling κ such that for every node
v ∈ Z and every pair (x, y) of random W -weightings from the distribution κ, we
have
P(xv ̸= yv ) ≤ (dob(W )D)d(v,U ) ,
where d(v, U ) denotes the distance of v from U in G.
What is important in this theorem is that it gives an exponentially decaying
correlation between the weight of v and the weights of nodes far away, provided
dob(W ) < 1/D.

Proof. We assume that dob(W ) < 1/D (else, there is nothing to prove). Let
κ be the stationary distribution of the Markov chain Ma,b with marginals πa and
πb . So κ is a coupling of these distributions.
Let x, y ∈ [0, 1]V , and let (x′ , y ′ ) be obtained from (x, y) by making one step of
Ma,b , using a random pivot node v ∈ Z. Let n = |Z|. Then for any node w ∈ Z,
n−1 1
(20.6) P(x′w ̸= yw
′
)= P(x′w ̸= yw
′
| v ̸= w) + P(x′w ̸= yw
′
| v = w).
n n
Here P(x′w ̸= yw′
| v ̸= w) = 1(xw ̸= yw ) (since nothing changes at w under this
condition), and
P(x′w ̸= yw
′
| v = w) = P(i ̸= j | v = w) = dtv (πx,w , πy,w ).
Substituting in (20.6), we get
n−1 1
(20.7) P(x′w ̸= yw
′
)= 1(xw ̸= yw ) + dtv (πx,w , πy,w ).
n n
 20.1. RANDOM HOMOMORPHISMS TO THE RIGHT 371

Now let (x, y) be a random pair from κ, and average (20.7) over x and y, to
get
n−1 1 ( )
(20.8) P(x′w ̸= yw
′
)= P(xw ̸= yw ) + E dtv (πx,w , πy,w ) .
n n
By the definition of stationary distribution, (x′ , y′ ) has the same distribution as
(x, y), and hence P(x′w ̸= yw
′
) = P(xw ̸= yw ). Substituting in (20.8), we get
( )
(20.9) P(xw ̸= yw ) = E dtv (πx,w , πy,w ) .
So far, we have not used the Dobrushin parameter dob(W ). By (20.5), we get
( ) ∑
(20.10) E dtv (πx,w , πy,w ) ≤ dob(W ) P(xu ̸= yu ).
u∈N (w)

Define f (u) = P(xu ̸= yu ). We have f (u) ∈ {0, 1} if u ∈ U , and (20.9) and (20.10)
imply that
∑
(20.11) f (w) ≤ dob(W ) f (u)
u∈N (w)

holds for all w ∈ Z. Inequality (20.11) says that the function f is strictly subhar-
monic at the nodes of Z. It is easy to derive from this fact an estimate on f . Let us
start a random walk (v 0 = v, v 1 , . . . ) on G from v ∈ Z, and let T be the (random)
time when this random walk hits U (if the connected component of v does not
intersect U , then f = 0 on this connected component and the conclusion below is
trivial). Consider the random variables Xt = f (v t )(dob(W )D)t . It follows from
(20.11) that these form a submartingale, and hence by the Martingale Stopping
Theorem A.11, we get
( ) ( )
f (v) = X 0 ≤ E(X T ) = E (dob(W )D)T f (v T ) ≤ E (dob(W )D)T .
Since trivially T ≥ d(v, U ), this completes the proof. 
It is important that the coupling κ constructed above is independent of the
node v. This means that if we want to estimate the probability that x|S ̸= y|S for
some subset S ⊆ Z, then we get the same coupling distribution κ, and so we can
use the union bound:
Corollary 20.5. Under the conditions of Theorem 20.4, every S ⊆ Z satisfies
P(x|S ̸= y|S ) ≤ (dob(W )D)d(S,U ) |S|.
Let us formulate some other consequences. First, consider proper q-colorings
of G, i.e., homomorphisms G → Kq . For Sr+1 in the definition of the Dobrushin
parameter, let φ and ψ be two q-colorings of the leaves that differ at node 1 only.
Then πφ,0 is the uniform distribution on the set [q]\φ([r]), and πψ,0 has an analogous
description. These sets have Hamming distance at most 2D, and hence their total
variation distance is at most 1/(q − D). So dob(WKq ) < 1/D is satisfied if q > 2D,
and we get:
Corollary 20.6. Let G = (V, E) be a graph with all degrees bounded by D, and let
q > 2D. Then for any U ⊆ V , any two proper q-colorings α and β of G[U ], and
any v ∈ V \ U , the random extensions φ and ψ of α and β to proper q-colorings of
G satisfy
( D )d(v,Z)
dtv (φ(v), ψ(v)) ≤ .
q−D
372 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

We can generalize this corollary to homomorphisms into any looped-simple

graph H, assuming the maximum degree ∆(H) of its complement H (among looped-
simple graphs) is small:

Corollary 20.7. Let G = (V, E) be a simple graph with all degrees bounded by
D, and let H be a looped-simple graph with 2D∆(H) < v(H). Then for any subset
U ⊆ V , any two homomorphisms α, β : G[U ] → H, and any v ∈ V \U , the uniform
random extensions φ and ψ of α and β to homomorphisms G → H, restricted to
the node v, satisfy

( D∆(H) )d(v,Z)
dtv (φ(v), ψ(v)) ≤ .
v(H) − D∆(H)

20.1.3. The Dobrushin value. Which graphons W have small Dobrushin

value? This property is related to the approximability of W by rank-1 graphons
(see Exercises 20.10 and 20.11, and also Proposition 20.3), but for us, the case when
W is close to the special rank-1 function 1 will be important. For a graphon W ,
define

∫ 1
(20.12) ∆(W ) = sup W (x, y) dy.
x∈[0,1] 0

(the “maximum degree”; note that δ(WH ) = ∆(H)/v(H) if H is a looped-simple

graph). The quantity provides a useful upper bound on the Dobrushin value.

Lemma 20.8. Every graphon W satisfies

∆(W )
dob(W ) ≤ .
1 − D∆(W )

In particular, the Dobrushin condition is satisfied if ∆(W ) < 1/(2D).

Proof. The proof is just computation (although a bit tedious). Let z, w ∈

[0, 1]r be two weightings of the leaves of Sr+1 that differ only at node 1. Let

∏
r ∏
r ∫ 1
g(x) = W (x, zi ) = W (x, wi ) and s(x) = g(y)W (x, y) dy.
i=2 i=2 0

The density functions of the distributions πz,0 and πw,0 are g(x)W (x, z1 )/s(z1 ) and
g(x)W (x, w1 )/s(w1 ), respectively, and hence

∫ 1  W (x, z ) W (x, w ) 
1  1 1 
(20.13) dtv (πz , πw ) = g(x) −  dx.
2 0 s(z1 ) s(w1 )
 20.1. RANDOM HOMOMORPHISMS TO THE RIGHT 373

We may assume without loss of generality that s(z1 ) ≥ s(w1 ), then

∫  W (x, z ) W (x, w ) 
1 1  1 1 
dtv (πz , πw ) = g(x) −  dx
2 0 s(z1 ) s(w1 )
∫  W (x, z ) W (x, z ) W (x, z ) − W (x, w ) 
1 1  1 1 1 1 
= g(x) − +  dx
2 0 s(z1 ) s(w1 ) s(w1 )
∫ (   )
1 1  W (x, z1 ) W (x, z1 )   W (x, z1 ) − W (x, w1 ) 
≤ g(x)  −  
+  dx
2 0 s(z1 ) s(w1 ) s(w1 )
∫ ( )
1 1 W (x, z1 ) W (x, z1 )  W (x, z1 ) − W (x, w1 ) 
= g(x) − +  dx
2 0 s(w1 ) s(z1 ) s(w1 )
∫ 
( W (x, z ) W (x, w ) W (x, z ) − W (x, w ) ) 
1 1 1 1 1 1
= g(x) − + dx
2 0 s(w1 ) s(w1 ) s(w1 )
∫
1 1 W (x, z1 ) − W (x, w1 ) + |W (x, w1 ) − W (x, z1 )|
= g(x) dx
2 0 s(w1 )
∫ ∫
(we have used here that g(x)W (x, z1 )/s(z1 ) = 1 = g(x)W (x, w1 )/s(w1 )). It is
easy to check that W (x, z1 ) − W (x, w1 ) + |W (x, w1 ) − W (x, z1 )| ≤ 2 − 2W (x, w1 ),
and using the trivial fact that g(x) ≤ 1, we get
∫ 1 ∫ 1
1 − W (x, w1 ) 1 ∆(W )
dtv (πx , πy ) ≤ g(x) dx ≤ (1 − W (x, w1 )) dx ≤ .
0 s(w1 ) s(w1 ) 0 s(w1 )
To estimate the denominator, note that
∏
r ∑
r
g(x)W (x, w1 ) = W (x, wi ) ≥ 1 − (1 − W (x, wi )),
i=1 i=1
and so
∫ ∫ ∑
1 1 r
( )
s(w1 ) = g(x)W (x, w1 ) dx ≥ 1− 1 − W (x, zi ) dx ≥ 1 − D∆(W ).
0 0 i=1
This proves the lemma. 
Exercise 20.9. For a random 3-coloring of the path Pn , determine the correlation
between the colors of the endnodes. Does it decay exponentially?
Exercise 20.10. Prove that for every graphon W there is a kernel of rank 1 such
that ∥U − W ∥1 ≤ dob(W ).
Exercise 20.11. Let U > 0 be a kernel of rank 1, and let W be an arbitrary
kernel. Prove that dob(W ) ≤ 4∥(W/U ) − 1∥∞ .
20.1.4. Random homomorphisms from infinite graphs into graphons.
Our goal is to define a random homomorphism G → W , where G is a countable
graph with degrees bounded by D, and W is graphon.
First, some technicalities: the W -weightings of V (G) form the product space
[0, 1]V (G) , which we endow with the product sigma-algebra A. A random homomor-
phism will be defined by a probability measure on A. To specify such a measure,
it suffices to specify its values on cylinder sets obtained by restricting the weight of
a finite number of nodes of G to given Borel sets. In other words, we can specify a
distribution π on W -weightings of G by specifying the distribution of its restriction
π|S to every finite set S ⊆ V (G). Of course, these restrictions must satisfy appro-
priate consistency conditions: if S ⊆ T , then (πT )|S = πS . Once we have a family
374 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

(πS ) of distributions satisfying these consistency relations, then the Extension The-
orem of Kolmogorov gives us a probability distribution π on all W -weightings such
that π|S = πS for all finite sets S ⊆ V (G).
So far, this is quite simple. There are many ways to specify such a family (πS )
of distributions. However, we would like other conditions to be satisfied. Let us
formulate two:
• Markov property. If G1 and G2 are two finite k-labeled graphs, φ is
a random W -weighting of G1 G2 , and we condition on φ|[k] , then φ|V (G1 ) and
φ|V (G2 ) become independent. This is just another way to express the product
formula (5.53). This property can be generalized to infinite graphs. Of course, we
have to exercise some care, since G1 and G2 may be infinite. Let S ⊆ V (G) be a
finite set and suppose that G \ S is the disjoint union of two graphs G1 and G2 . Let
z be a W -weighting of S, and let x denote a random W -weighting of V \ S from
the distribution obtained by conditioning on x|S = α. We require that the random
weightings x|V (G1 ) and x|V (G2 ) be independent. We say that the distribution of x
has the Markov property, if this condition holds for every finite subset S ⊆ V (G)
and every W -weighting z of S.
• Locality. For a finite set S ⊆ V (G), we would like to get a good idea
of the distribution πS by looking at a sufficiently large neighborhood of S. Let
B(S, r) = {v ∈ V (G) : d(v, S) ≤ r} be the r-neighborhood of S, and let xr
denote a random W -weighting of G[B(S, r)]. Then we want that xr |S → x|S in
distribution as r → ∞. We call the distribution of x local if this holds.
These conditions are not too strong, as the following classical theorem shows
(see [1988] and [1993] for slightly different statements of this fact).
Theorem 20.12. Let G be a countable graph with degrees bounded by D, and let
W be a graphon such that dob(W ) < 1/D. Then there is a unique local probability
distribution πG,W on W -weightings of G with the Markov property.
Proof. Let S ⊆ V (G) be a finite set, and let xr be a random W -weighting of
G[B(S, r)].
Claim 20.13. The distribution of xr |S tends to a limit as r → ∞.
We show that these distributions form a Cauchy-sequence in the total variation
distance. Let ε > 0. Since dob(W ) < 1/D, we can choose r large enough so that
(Ddob(W ))r ≤ ε/|S|. Let m, n > r, we claim that the distributions of xm |S and
xn |S are ε-close in total variation distance. Let zn be the restriction of xn to
B(S, n) \ B(S, r), and let x′n be the random weighting of G[B(S, n)], obtained by
conditioning on zn . By the Markov property (we are using it for finite graphs here!),
x′n has the same distribution as xn . We define zm and x′m analogously.
Now we fix any two weightings zn of B(S, n) \ B(S, r) and zm of B(S, m) \
B(S, r), and let yn and ym be obtained by conditioning xn and xm on these partial
weightings. By Theorem 20.4, yn and ym can be coupled so that P(yn (v) ̸=
ym (v)) ≤ ε/|S| for every v ∈ S. This implies that
dtv (yn |S , ym |S ) ≤ ε.
Since this holds for fixed zn and zm , it also holds if they are random restrictions
of xn and xm , so it holds for x′n and x′m . Since these weightings have the same
distribution as xn and xm , the claim follows.
 20.2. CONVERGENCE FROM THE RIGHT 375

Now we are able to define the distribution on W -weightings. For a finite set
S ⊆ V (G), let πS be the limit of the distributions of xr |S as r → ∞. It is easy
to check (using similar arguments as in the proof of Claim 20.13 above), that the
family (πS ) of distributions is consistent, and the distribution πG,W they define has
the Markov property. Uniqueness follows immediately from locality. 
A probability distribution on W -weightings of G is called a Gibbs state if it is
invariant under the Markov chain M of local re-weightings (as used in the proof
of theorem 20.4 in the finite case). It can be proved that under the condition that
dob(W ) < 1/(2D), the Gibbs state is unique.
Remark 20.14. In a sense, the construction of a random homomorphism can be
extended to graphings. The method is similar to the Bernoulli lift of a graphing
(Section 18.5). Given a graphing G and a graphon W on [0, 1] such that dob(W ) <
1/(2D), we define a graphing G[W ] on the Graph of Weighted Graphs H+ . To
describe the probability distribution on G+ , we generate a random element from it
as follows: pick a point x ∈ V (G) and generate a random W -weighting of Gx as
described above. If W ≡ 1, we get the Bernoulli lift.
We cannot randomly map all points of a graphing into [0, 1] in any reasonable
way; this is impossible even if the graphing has no edges. But if we select any
countable subset, this can be mapped, and the graphing G[W ] contains the neces-
sary information. I don’t know of any applications of this construction, but I like
the fact that our two basic limit objects, graphings and graphons, can be combined
this way.

20.2. Convergence from the right

While the theory of convergent sequences of bounded degree graphs lacks some
of the key facts and constructions that apply in the dense case (most notably a
good notion of distance), it is nicer in at least one respect: convergence of a graph
sequence can be characterized by convergence of (appropriately normalized) homo-
morphism numbers into certain fixed graphs so we don’t have to switch to maxi-
mization of multicuts as in the dense case. This result is due to Borgs, Chayes, Kahn
and Lovász [2012]. We note that the necessity of the right-convergence condition
follows only for target graphs that satisfy the Dobrushin condition (but under this
condition, it follows more generally for homomorphism numbers into graphons).
To state this result, let us define, for a simple graph G with bounded degrees
and graphon W , the (sparse, normalized) homomorphism entropy
log t(G, W )
ent∗ (G, W ) = ,
v(G)
In the case when W = WH for some weighted graph H on q nodes, we write
ent∗ (G, H). In this special case, we could replace t(G, H) by hom(G, H) in this
definition: this would mean simply adding log αH to the value, so it is a matter of
taste which version one uses in the definition.
To see the meaning of ent∗ (G, W ), consider the case when W = WH for some
simple graph H. Then log hom(G, H)/v(G) expresses the freedom (entropy) we
have in choosing the image of a node v ∈ V (G) in a homomorphism G → H, and
ent∗ (G, H) (which is always nonpositive) expresses the loss of entropy per node due
to taking homomorphisms instead of all maps.
The main result in this section is the following.
376 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

Theorem 20.15. For any sequence (Gn ) of graphs in G, the following are equiva-
lent:
(i) (Gn ) is locally convergent;
(ii) for every graphon W with dob(W ) ≤ 1/D, the sequence ent∗ (Gn , W ) is
convergent;
(iii) there is an ε > 0 such that for every looped-simple graph H with ∆(H) ≤
εv(H) the sequence ent∗ (Gn , H) is convergent.
The equivalence of conditions (ii) and (iii) is analogous to the equivalence of
conditions (ii) and (iii) in Theorem 12.20, and similarly as there, we could replace
them by any condition “inbetween”, like weighted graphs satisfying the Dobrushin
condition.
In the special case when H = Kq , we have ∆(Kq ) = 1, and hom(G, Kq ) is the
number of q-colorings of G. So it follows that if (Gn ) is convergent and q > 2D,
then the number of q-colorings grows as cv(Gn ) for some c > 1. It is easy to see that
some condition on q is needed: for example, if Gn is the n-cycle and q = 2, then
ent∗ (Gn , K2 ) oscillates between −∞ and ≈ 0 as a function of n.
Lemma 20.8 says that ∆(W ) < 1/(2D) is sufficient for (ii) to apply. This
condition could not be relaxed by more than a constant factor, as the following
example shows.
Example 20.16. Let Gn be a random D-regular graph on 2n nodes, and G′n
be a random bipartite D-regular graph on 2n nodes. The interlaced sequence
(G1 , G′1 , G2 , G′2 , . . . ) is locally convergent with high probability (almost all r-
neighborhoods are D-regular trees if r is fixed and n is large enough). Let H
be obtained from K2◦ by weighting the non-loop edge by 1 and the loops by 2c .
Inequality (5.33) can be generalized to give the bounds
c maxcut′ (G) ≤ ent∗ (G, H) ≤ c maxcut′ (G) + 1.
(Here maxcut′ (G) = Maxcut(G)/v(G) is normalized differently from the normal-
ization in (5.33).) The maximum cut in G′n has Dn/2 edges, but the maximum
cut in Gn has at most Dn/3 edges with high probability (see Bertoni, Campadelli,
Posenato [1997] for a sharp estimate). Hence
cD cD
ent∗ (G′n , H) = , but ent∗ (Gn , H) ≤ 1 +
2 3
If cD/2 − cD/3 = cD/6 > 1, then the sequence (ent∗ (G1 ), ent∗ (G′1 ), ent∗ (G2 ),-
ent∗ (G′2 ), . . . ) cannot be convergent with high probability. So assuming ∆(W ) ≤
7/D would not be enough in (ii). 
While Theorem 20.15 sounds similar to the results in Chapter 12 (in particular
Theorem 12.20), it is both more and less than that theorem. We get a characteriza-
tion of convergence in terms of left and right homomorphisms, but no analogue of
the characterization as a Cauchy sequence in the cut metric. Also, convergence is
not established for all soft-core graphs H, just for those close to a complete graph.
On the other hand, the proof below says more, since it provides explicit formu-
las relating left and right homomorphism numbers. Furthermore, homomorphism
densities into graphons are considered, not just weighted graphs; recall that the
corresponding extension of Theorem 12.20 to graphons if false (Remark 12.22).
 20.2. CONVERGENCE FROM THE RIGHT 377

Supposing that we have a convergent sequence (Gn ) tending to an involution-

invariant distribution σ (or to a graphing G), what is the limit of the homomor-
phism entropies? The answer is not trivial, since there is no good way to “count”
homomorphisms of an infinite graph (or of a graphing) into a weighted graph H.
Even for the number of q-colorings (q > 2D), and for a sequence of D-regular
graphs with girth tending to ∞, where the Benjamini–Schramm limit is a single
(infinite) D-regular rooted tree, the limiting value is nontrivial to determine. A
natural guess would be that starting at the root of the infinite tree, and working
our way out, we have q choices for the color of the root and q − 1 choices for every
other node, which suggests an entropy of ent∗ (Gn , Kq ) → log(1 − 1/q). But this
is not the right answer, which was determined by Bandyopadhyay and Gamarnik
[2008]:
D ( 1)
(20.14) ent∗ (Gn , Kq ) → log 1 − .
2 q
To motivate the following description of the limiting homomorphism entropy
in the general case, consider a finite graph G. For any ordering (v1 , . . . , vn ) of its
nodes, we consider the graphs Gi = G[v1 , . . . , vi ]. Then
1∑
n
t(Gi , W )
(20.15) ent∗ (G, W ) = log .
n i=1 t(Gi−1 , W )
If W = WH for some looped-simple graph H, then the fraction inside the logarithm
is the conditional probability that a random map V (Gi ) → V (H) is a homomor-
phism, given that its restriction to Gi−1 is a homomorphism. In general, it can be
expressed as
t(Gi , W ) ( ∏ )
= Ey Ex W (x, yu ) ,
t(Gi−1 , W )
u∈N (v)

where x is a uniform random number in [0, 1], and y is a random W -weighting of

Gi−1 .
If we try to extend this to infinite graphs, the formula makes sense, but as
Example 20.16 shows, it may give the wrong result. The trick is to average over all
orderings of V (G). We generate an ordering by a random map τ : V (G) → [0, 1].
We denote the set of nodes u ∈ V (G) with τ (u) < τ (v) by Vτ (v), and set Nτ (v) =
N (v) ∩ Vτ (v). We can view (20.15) as another averaging (over a random node of
G). Thus we get
( ∏ )
ent∗ (G, W ) = Ev Eτ log Ey Ex W (x, yu ) ,
u∈Nτ (v)

where x is a uniform random number in [0, 1], where y is a random W -weighting

of G[Vτ (v)].
Now this formula extends to involution-invariant distributions σ. Instead of
a random node v, we consider a random rooted graph (G, v) from σ. Instead
of random bijection V (G) → [n], we consider a random map τ : V (G) → [0, 1].
Assuming G satisfies the Dobrushin condition, so does G[Vτ (v)], and so the random
W -weighting y is well defined. So we can define
( ∏ )
(20.16) ent∗ (σ, W ) = E(G,v) Eτ log Ey Ex W (x, yu ) .
u∈Nτ (v)
378 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

Then we have the following supplement to Theorem 20.15.

Supplement 20.17. Let (Gn ) be a locally convergent sequence of graphs with de-
grees at most D, and let σ be the involution-invariant distribution representing its
limit. Let W be a graphon with dob(W ) < 1/D. Then
ent∗ (Gn , W ) → ent∗ (σ, W ).
Let us illustrate that the rather hairy formula (20.16) does allow us to determine
the limiting values of homomorphism entropies, at least in a simple case.
Example 20.18. Suppose that Gn is D-regular and the girth of Gn tends to
infinity. Let H = Kq , so that hom(G, H) = chr(G, q). Then Gn tends to the
involution-invariant distribution concentrated on the infinite D-regular tree (at least
we don’t have to take expectation over this). Specializing (20.16), we get
( ∏ )
ent∗ (σ, W ) = Eτ log Ey Ex 1(x ̸= yu ) .
u∈Nτ (v)

Here y is a random coloring with colors from [q], and x is a random color. Whatever
Vτ (v) is, y assigns uniform and independent colors to the nodes in Nτ (v), since our
graph is a tree. Hence for every x,
( ∏ ) ( q − 1 )|Nτ (v)|
Ey 1(x ̸= yu ) = ,
q
u∈Nτ (v)

and hence
(( )
q − 1 )|Nτ (v)| ( 1) D ( 1)
ent∗ (σ, W ) = Eτ log = Eτ (Nτ (v)) log 1 − = log 1 − .
q q 2 q
So we get the theorem of Bandyopadhyay and Gamarnik (20.14). 
Proof of Theorem 20.15. (i)⇒(ii) Let G = (V, E) be a simple graph with
degrees bounded by D. We may assume that αH = 1. We use the formula (20.15)
derived above, and concentrate on the innermost expression
( ∏ )
s(v, τ, y) = Ex W (x, yu ) .
u∈Nτ (v)

The Dobrushin Uniqueness Theorem 20.4 implies that we don’t change the expres-
sion by much if we restrict everything to the r-neighborhood Nr (v). To be precise,
let c = Ddob(H) < 1, and define Gr = G[Nr (v)], Vτr (v) = Nr (v) ∩ Vτ (v), and let
sr denote the function s defined for the graph Gr . Let z be a random W -weighting
of G[Vτr (v)], then Theorem 20.4 implies that the distributions of y and z, when
restricted to v and its neighbors, are closer that (D + 1)cr−1 is total variation
distance. This implies that
 r 
Ez s (v, τ, z) − Ey s(v, τ, y) ≤ (D + 1)cr−1 ,
and hence
 ∗ 
(20.17) ent (G, H) − Ev Fr (v) ≤ (D + 1)cr−1 ,
where
( )
(20.18) Fr (v) = Eτ log Ez sr (v, τ, z) .
(We can take expectation over the same τ , since it induces a uniform random
permutation of V (Gr ) as well as of V (G).)
 20.2. CONVERGENCE FROM THE RIGHT 379

Let us note that in (20.18) Fr (v) depends only on the r-ball B = Nr (v), and
we can denote it by F (B). This allows us to express Ev Fr (v) in terms of the
distribution σG,r of r-neighborhoods in G. Thus (20.17) implies
 ∑ 
 ∗ 
(20.19) ent (G, H) − σG,r (B)F (B) ≤ (D + 1)cr−1 .
B∈Br

Now, let (Gn ) be a locally convergent sequence tending to an involution-

invariant distribution σ. Then (20.19) implies that
∑
lim sup ent∗ (Gn , H) ≤ σr (B)F (B) + (D + 1)cr−1 ,
n
B∈Br

and hence ∑
lim sup ent∗ (Gn , H) ≤ lim inf σr (B)F (B).
n r
B∈Br
A similar argument proves that lim inf n ≥ lim supr , which implies that both limits
exist.
(ii)⇒(iii) is trivial.
(iii)⇒(i) We switch to the natural logarithm, since we are going to use analytic
formulas (this only means that all formulas are multiplied by ln 2). We express the
logarithm of t(G, H) as
∑
(20.20) ln t(G, H) = ℓ(G[S], H),
S≤V (G)

where by Möbius inversion,

∑
(20.21) ℓ(G, H) = (−1)|V (G)|−|S| ln t(G[S], H).
S⊆V (G)

Using that ln t(., H) is an additive graph parameter for any fixed H, it is easy to
see that ℓ(F, H) = 0 unless F is a connected graph together with isolated nodes
(cf. Exercise 4.2). The term corresponding to the edgeless graph is 0, and so we
can modify (20.20) so that the summation runs over connected induced subgraphs
of G. Collecting terms with isomorphic graphs, we get
∑ ind(F, G) ℓ(F, H)
(20.22) ent∗ (G, H) = · ,
v(G) aut(F )
F

where the summation ranges over all isomorphism types of connected graphs F ;
but of course, only a finite number of terms are non-zero for any fixed G.
So we can express the homomorphism entropies ent∗ (Gn , H) as linear combina-
tions of the induced subgraph densities ind(F, Gn )/v(Gn ). This suggests a heuristic
for the proof: We show that the system of equations (20.22) can be inverted, to ex-
press the induced subgraph densities as linear combinations of the homomorphism
entropies. It follows then that if the homomorphism entropy into any given graph
converges to some value, then so does the frequency of each induced subgraph.
This heuristic is of course very naive: (20.22) is an infinite system of equations,
and so to do anything with it we need tail bounds; furthermore, the coefficient
ℓ(F, H) is defined by the hairy formula (20.21), which has all the unpleasant features
one can think of: it has an exponential number of terms, these terms alternate in
sign, and the terms themselves are logarithms of simpler functions.
380 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

The identities developed in Section 5.3.1 come to rescue. We can get rid of
the logarithms using Corollary 5.22. Substituting the formula for ln t(G, H) in the
definition of ℓ(G, H), we get a lot of cancellation, which leads to the formula
∑∞ ∑ ∑
(−1)m e(Ji )
(20.23) ℓ(F, H) = (−1) i
m!
m=1 J1 ,...,Jm ∈Conn(F )
∪i V (Ji )=V (F )

( )∏
k
× cri L(J1 , . . . , Jm ) t(Jr , H).
r=1

(It is not clear at this point that this is any better than (20.21), but be patient.)
Next we turn to inverting the expression (20.22). Let m ≥ 1 and let
{F1 , . . . , FN } be the set of all connected simple graphs with 2 ≤ v(Fi ) ≤ m. Let
q > m/ε, add q − v(Fi ) ≥ 1 new isolated nodes to Fi , and take the complement to
get a looped-simple graph Hi on [q] with loops added at all nodes. We weight each
node of Hi by 1/q. Every node in Hi has degree at least q − m, so ∆(H i ) ≤ εq.
Consider any graph G with all degrees at most D. We write (20.22) in the form
∑N
ind(Fi , G) ℓ(Fi , Hj )
(20.24) ent∗ (G, Hj ) = · + R(G, Hj ),
i=1
v(G) aut(Fi )
where
∑ ind(F, G)ℓ(F, Hj )
(20.25) R(G, Hj ) =
aut(F )v(G)
v(F )>m

is a remainder term.
We can view (20.24) as a system of N equations in the N unknowns xi =
( )N
inj(Fi , G)/v(G). Let A = ℓ(Fi , Hj )/aut(Fi ) i,j=1 be the matrix of this system,
and let s, R ∈ RN be defined by sj = ent∗ (G, Hj ) and Rj = R(G, Hj ), then we
have AT x = s−R. Assuming that A is invertible (which we will prove momentarily),
let B = (AT )−1 . Then the system can be solved: x = B(s − R), or

ind(Fi , G) ∑
N
(20.26) = Bij ent∗ (G, Hj ) + ri (G),
v(G) j=1

where
∑
N
ri = ri (G) = Bij R(G, Hj )
j=1
is a remainder term.
We have to show that the matrix A is invertible (at least if q is large enough)
and estimate the remainder terms. We use (20.23):
∞
∑ ∑ ∑
(−1)k e(Ji )
(20.27) ℓ(F, Hi ) = (−1) i
k!
k=1 J1 ,...,Jk ∈Conn(F )
∪V (Ji )=V (F )

( )∏
k
× cri L(J1 , . . . , Jk ) t(Jr , Hi ).
r=1
 20.2. CONVERGENCE FROM THE RIGHT 381

By the construction of Hi , we have t(Jr , Hi ) = q −v(Jr ) hom(Jr , Fi ), and so

∏
k ∑ ∏
k
t(Jr , Hi ) = q − r v(Jr )
t(Jr , Fi ).
r=1 r=1

Note that for a nonzero term the exponent of q is less than −v(F ) except for k = 1
and V (J1 ) = V (F ), and that the last product does not depend on q. Hence for any
simple graph F ,
∑
(20.28) ℓ(F, Hi ) = q −v(F ) (−1)e(J)−1 t(J, Fi ) + O(q −v(F )−1 ).
J∈Csp(F )

(Here and in what follows, the constants implied in the big-O notation may depend
( )N
on m, but not on q and G). By Proposition 5.43, the matrix M = t(Fi , Fj ) i,j=1
is nonsingular. Let L be the N × N matrix with entries Lij = 1(Fi ∈ Csp(Fj )),
and let P and Q denote the diagonal matrices with entries Pii = (−1)e(Fi )−1 and
Qii = q v(Fi ) aut(Fi ), respectively. Clearly L, P and Q are nonsingular. By (20.28),
we have
QAT = LT P M + O(q −1 ),
which implies that A is nonsingular if q is large enough. Furthermore,
Bij = q v(Fi ) aut(Fi )((M T P L)−1 )ij + O(q v(Fj )−1 ),
and so
(20.29) |Bij | = O(q v(Fi ) ) = O(q m ).
Using this, the remainder terms can be estimated as follows:
∑∞ ∑ ind(F, G)
|R(G, Hj )| ≤ |ℓ(F, Hj )|
r=m+1
aut(F )v(G)
v(F )=r
∞
∑ ∑ ind(F, G)
= O(q −r )
r=m+1 v(F )=r
aut(F )v(G)
∞
∑
(20.30) = 2Dr O(q −r ) = O(q −m−1 ).
r=m+1

and
∑
N
(20.31) ri (G) = Bji R(G, Hj ) = O(q m )O(q −m−1 ) = O(q −1 ).
j=1

So we have proved that in (20.26), for fixed m, the error term ri tends to 0 as
q → ∞.
The rest of the proof is standard analysis: Assume that ent∗ (Gn , H) → Sj
(n → ∞) for every looped-simple graph H with ∆(H) ≤ ε. Consider any simple
graph Fi on m nodes. Equation (20.26) implies that
 ind(F , G ) ∑ N  ∑ N
 
|Bji ||ent∗ (Gn , Hj ) − Sj | + |ri (Gn )|.
i n
(20.32)  − Bji Sj  ≤
v(Gn ) j=1 j=1

Let δ > 0 be given, and choose q large enough so that |ri (Gn )| ≤ δ/2 for every n
(recall that the big-O in (20.31) does not depend on G). Since ent∗ (Gn , Hj ) → Sj ,
382 20. RIGHT CONVERGENCE OF BOUNDED DEGREE GRAPHS

the first term on the right side of (20.32) is at most δ/2 if n is large enough. It
follows that ind(F, Gn )/v(Gn ) is a Cauchy sequence, which means that the sequence
(Gn ) is locally convergent. 
The proof of the Supplement is based on similar arguments and not given here
in detail. The proof method used above for (iii)⇒(i) can also be used to prove a
somewhat weaker version of (ii), replacing the Dobrushin condition dob(W ) < 1/D
by 8D∆(W ) < 1. In fact, the expression (20.22) yields itself more directly to a
proof of (i)⇒(ii) than to a proof of (b): naively, if the frequency of any induced
subgraph converges to some value, then so do the homomorphism entropies. The
main issue is to obtain good tail bounds, which can be done similarly as in the
proof above, as long as we are satisfied with proving the convergence for very small
∆(W ); but if we want a bound that is sharp up to a constant, then we need more
technical computations. We refer to the paper of Borgs, Chayes, Kahn and Lovász
[2012] for these details.
Remark 20.19. It is a natural question to ask which sequences of bounded de-
gree graphs are right-convergent in the sense that their homomorphism entropies
converge for all soft-core target graphs. Gamarnik [2012] studies this problem for
sparse random graphs, but the general question is unsettled. It is also natural to
ask whether local-global convergence can be characterized by any right-convergence
condition.

Exercise 20.20. Let G and G′ be two graphs on the same set of nodes [n], such
that |E(G)△E(G′ )| ≤ εn. Prove that δ⊙ nd
(G1 , G2 ) ≤ 2ε.
Exercise 20.21. Let H be a weighted graph with positive edgeweights and (Gn ),
a bounded degree graph sequence for which the sequence (ent∗ (Gn , H)) is conver-
gent. Let G′n be obtained from Gn by deleting o(v(Gn )) nodes and edges. Prove
that (ent∗ (G′n , H)) is convergent.
Exercise 20.22. Let H be a weighted graph with at least one positive edgeweight.
Prove that the sequence ent∗ (Pn Pm , H) is convergent as n, m → ∞, and the
same holds for the sequence ent∗ (Cn Pm , H), provided n is restricted to even
numbers.
Exercise 20.23. Let H be a weighted graph whose edges with positive
weight form a connected and nonbipartite graph. Prove that the sequence
ent∗ (Cn Pm , H) is convergent as n, m → ∞.
Exercise 20.24. Let σ be an involution-invariant measure. Show how to express
s(σ) in terms of the associated Bernoulli graphing.
 CHAPTER 21

On the structure of graphings

21.1. Hyperfiniteness
A notion related to Følner sequences in the theory of amenable groups is “hyper-
finiteness” for general graph families with bounded degree, which can be extended
to graphings in a natural way. This notion was introduced (in different settings) by
Kechris and Miller [2004], Elek [2007b] and Schramm [2008]. Hyperfiniteness of a
graph family has a number of important consequences, like testability of many graph
properties. Quoting an informal remark by Elek, hyperfinite bounded-degree graph
families and graphings behave as nicely as dense graph sequences and graphons do.

21.1.1. Hyperfinite graph families. A graph G ∈ G is called (ε, k)-

hyperfinite (ε ∈ (0, 1), k ∈ N), if we can delete εv(G) edges and get a graph in
which every connected component has at most k nodes. Let H ⊆ G be any fam-
ily of finite graphs (recall that all degrees are bounded by D). We say that H is
hyperfinite, if for every ε > 0 there is a k = k(ε) > 0 such that every G ∈ H is
(ε, k)-hyperfinite.
We could talk about deleting nodes instead of edges. Indeed, deleting one
endnode of every edge in S results in even smaller components. Conversely, if
deleting a set T of nodes results in a graph with small components, then deleting the
set S of edges incident with any node in T leaves small components, and |S| ≤ D|T |.
We will see that hyperfinite families are very well-behaved, often as well as
dense graphs for analogous questions. How special are they? The examples below
show that several important graph families are hyperfinite. (In fact, one has to work
to construct a family that is not hyperfinite.) It is also likely that many large real-
life networks can be thought of as hyperfinite, showing the potential applicability
of the theory of hyperfinite families.
Example 21.1 (Trees). The family of trees in G is hyperfinite. Indeed, select an
endpoint r as the root in a tree T and fix an integer k ≥ 1. It is easy to see that if
v(T ) > k, then there is always an edge such that the branch rooted at it has at least
k/(D −1) but at most k nodes. If we delete recursively such edges, then the number
of edges deleted is at most (D − 1)(v(T ) − 1)/k, and every connected component of
the remaining forest has at most k nodes. So our tree is ((D − 1)/k, k)-hyperfinite.

Example 21.2 (Grids). From an n×m grid G, delete the edges inside every M -th
vertical and horizontal ribbon of squares (starting from the top and from the left,
say). The number of edges deleted is at most m((n − 1)/M ) + n((m − 1)/M ) <
2v(G)/M , and every connected component of the remaining graph has at most M 2
nodes. So this grid is (M 2 , 2/M ) hyperfinite. 

383
384 21. ON THE STRUCTURE OF GRAPHINGS

Example 21.3 (Planar graphs). More generally, the family of planar graphs
with degree bounded by D is hyperfinite. Indeed, let G be such a graph on n
nodes. The Lipton–Tarjan
√ Planar Separation Theorem [1979] says that G has a
set S of at most 3 n nodes such that every connected component of G − S has at
most 2n/3 nodes. We repeat this with every connected component of the remaining
graph until all components have at most K nodes.
The estimation of the number of deleted nodes is somewhat tricky and it is left
to the reader as Exercise 21.21. 
Example 21.4 (Random regular graphs). Let us generate a random D-regular
graph Gn on n nodes, for every even n, by choosing one of the D-regular graphs
uniformly at random. The family of graphs obtained is not hyperfinite with prob-
ability 1 if D ≥ 3. The following heuristic argument to prove this can be made
precise rather easily. If Gn is (ε, k)-hyperfinite, then V (G) can be split into two sets
of size between n/2 −k and n/2+ k in such a way that the number of edges between
the two classes is at most εn. On the other hand, let {S1 , S2 } be a partition of
[n] into two classes of size about n/2, and let Z denote the number of edges in
Gn connecting the two classes. The expected number of such edges is about Dn/4.
Furthermore, Z is highly concentrated around its mean, and so the probability that
Z ≤ εn is o(2−n ) if ε is small enough. There are fewer than 2n such partitions of
[n], so with high probability all of these have more than εn edges connecting the
two classes. 
Example 21.5 (Expanders). We call a family E ⊆ G of graphs an expander
family if there is a c > 0 such that for every graph G ∈ E and every S ⊆ V (G)
with |S| ≤ v(G)/2, we have eG (S, V (G) \ S) ≥ c|S|. An infinite family of expander
graphs is not hyperfinite. Indeed, T ⊆ E(G) and G−T has components G1 , . . . , Gr ,
and all of these have fewer than v(G)/2 nodes, then
1∑ ( ) ∑
r r
|T | = eG V (Gi ), V (G) \ V (Gi ) ≥ cv(Gi ) = cv(G).
2 i=1 i=1
It can be shown that the family of random D-regular graphs in Example 21.4 is an
expander family with probability 1.
The following explicit construction for an expander graphing was given (in a
different context) by Margulis [1973]. Consider the space R2 /Z2 , a.k.a. the torus.
Let us connect every point (x, y) to the points (x ± y, y) and (x, y ± x) (additions
modulo 1; we can leave out the axes if we don’t want loops). This graph is the
support of the measure preserving family consisting of the two maps (x, y) 7→
(x + y, y) and (x, y) 7→ (x, x + y), and hence it is a graphing. Furthermore, this
graphing is an expander, and hence not hyperfinite. This is not easy to prove; for
a proof based on Fourier analysis, see Gabber and Galil [1981]. 
Example 21.6. A special case of a hyperfinite family is a family of graphs with
subexponential growth, familiar from group theory. To be precise, for a function
f : N → N we say that a family H of graphs has f -bounded growth, if for any
graph G ∈ H, any v ∈ V (G) and any m ∈ N, the number of nodes in the m-
neighborhood of v is at most f (m). We say that H has
( subexponential
) growth, if it
has f -bounded growth for some function f such that ln f (m) /m → 0 (m → ∞).
It was asked by Elek and proved by Fox and Pach [unpublished] that this property
implies hyperfiniteness (Exercise 21.22). 
 21.1. HYPERFINITENESS 385

The following important class of hyperfinite families, generalizing Example 21.3

was found by Benjamini, Schramm and Shapira [2010].
Proposition 21.7. Every minor-closed family of graphs in G that does not contain
all graphs is hyperfinite.
Proof (sketch). Alon, Seymour and Thomas [1990] proved that the Pla-
nar Separator Theorem extends to every minor-closed property not containing all
graphs. The same argument as described in Example 21.3 can be carried through
to give hyperfiniteness. 
21.1.2. Hyperfinite graphings. Hyperfiniteness can be generalized to
graphings; in fact, we don’t have to talk about classes of graphings, the notion
makes sense for a simple graphing, and leads to some nontrivial questions.
A graphing G is (ε, k)-hyperfinite (ε ∈ (0, 1), k ∈ N), if there is a Borel set
S ⊆ E(G) with η(S) ≤ ε such that every connected component of G − S has at
most k nodes. A graphing G is hyperfinite, if for every ε > 0 there is a positive
integer k such that G is (ε, k)-hyperfinite.
We could relax this definition and ask for a set S ⊆ E(G) with η(S) ≤ ε such
that every connected component of G − S is finite. But this would not change the
notion of hyperfiniteness. Indeed, suppose that G satisfies the relaxed condition;
we show that it satisfies this stronger condition as well. We choose a set S ′ for ε/2
in place of ε. Let Vm ⊂ V (G) \ S ′ be the set of points contained in some connected
component
∑ of G−S ′ with m nodes. Then Vm is measurable (Exercise 18.10), and we
have m η(E(G[V∑ m ])) = η(E(G)) ≤ D. It follows that there is a∪positive integer
K such that m≥K η(E(G[Vm ])) ≤ ε/2, and so the set S = S ′ ∪ m≥K E(G[Vm ])
is a set of measure at most ε such that every connected component of G − S has
fewer than K nodes.
Similarly as for graphs, we could define the same notion by the existence of a
Borel set of points S ⊆ V (G) such that λ(S) ≤ ε and every connected component
of G − S is finite.
For a graphing, (ε, k)-hyperfiniteness can be expressed in a local-global way,
through a red-blue coloring of the edges such that the η-measure of red edges is
less than ε, and every connected k-node subgraph contains a red edge. Using this,
it is easy to verify the following fact.
Proposition 21.8. Let G and G′ be locally-globally equivalent graphings. If G is
(ε, k)-hyperfinite, then G′ is (ε′ , k)-hyperfinite for every ε′ > ε.
This proposition does not remain true for locally equivalent graphings (see Ex-
ample 21.12 below). The following important and surprisingly non-trivial theorem,
which is a version of a result of Schramm (see Theorem 21.13 below) shows that
hyperfiniteness (without parameters) is preserved by local equivalence.
Theorem 21.9. Let G and G′ be locally equivalent graphings. If G is hyperfinite,
then so is G′ .
The proof is best understood if we introduce a fractional version of hyperfinite-
ness. The motivation comes from combinatorial optimization. Let G be a graphing,
and let R denote the set of subsets Y ⊆ V = V (G) that induce a connected sub-
graph of G and have at most k elements. This can be considered as a subset of
U = V ∪ V 2 ∪ · · · ∪ V k , and it is a Borel set if we endow U with the natural sigma-
algebra it inherits from V . Let S ⊆ E(G) be a Borel set such that every connected
386 21. ON THE STRUCTURE OF GRAPHINGS

component of G \ S has at most k nodes. For x ∈ V , let Cx denote the node set of
the connected component of G \ S containing x. The sets Cx partition V . Let x be
a random point of G, then Cx is a random member of R, which has the following
two properties:
(1) If we select Cx first, and then select a uniform random point y ∈ Cx (note
that Cx is finite!), then y is distributed according to λ.
(2) If ∂(X) denotes the number of edges of G connecting X to V (G)\X (where
X is a finite subset of V (G)), then
( ∂(C ) )
x
E = η(S) ≤ ε.
|Cx |
Both of these properties can be easily verified using the Mass Transport Prin-
ciple (similarly to the proof of Proposition 18.50).
This motivates the next definition: we call a probability distribution τ on R a
fractional partition (into parts in R), if selecting Y ∈ R according to τ , and then
a point y ∈ Y uniformly, we get a point distributed according to λ; and we define
the boundary value of τ as
( ∂(Y) )
∂(τ ) = E .
|Y|
We say that G is fractionally (ε, k)-hyperfinite, if there is a fractional partition τ
such that ∂(τ ) ≤ ε. It follows from the discussion above that every (ε, k)-hyperfinite
graphing is fractionally (ε, k)-hyperfinite. The converse is not true (cf. Example
21.12), but we have the following weak converse:
Lemma 21.10. If a graphing is fractionally (ε, k)-hyperfinite, then it is
(ε log(8D/ε), k)-hyperfinite.
Proof. We use the Greedy Algorithm to construct a partition from a frac-
tional partition τ that establishes that G is fractionally (ε, k)-hyperfinite. Similar
algorithms are well known in combinatorial optimization, but here we have to be
careful, since we are going to construct an uncountable family of sets, and have
to make sure that the partition we obtain has the property that the set of edges
connecting different classes is Borel (and of course has small measure). We do our
construction in r = ⌊log(2D/ε)⌋ phases.
We start with R0 = U0 = ∅. In the j-th phase, let Uj,0 = Uj−1 be the
union of previously selected sets. Let Rj,1 be the set of sets Y ∈ R such that
∂(Y ) < ε2j−1 |Y \ Uj,0 |. Let Qj,1 be a maximal Borel set of sets Rj,1 such that
the sets Y \ Uj,0 are disjoint. Such a set exists by the following construction. Let
H0 be the intersection graph of Rj,0 . It is easy to see that H0 is a Borel graph
with bounded degree, and so it contains a maximal stable set that is Borel (this is
implicit in the proof of Theorem 18.3, see Exercise 18.11). Let Uj,1 be the union of
Uj,0 and the sets in Qj,1 .
The phase is not over; we select a maximal Borel family Qj,2 of sets Y ∈ R
such that the sets Y \ Uj,1 are disjoint and ∂(Y ) < ε2j−1 |Y \ Uj,1 |. We let Uj,2 be
the union of Uj,1 and the sets in Qj,2 . We repeat this k + 1 times, to finish the j-th
phase (after a while, we may not be adding anything). Let Qj be the family of sets
Y ∈ R selected in the j-th phase, and let Uj = Uj,k be their union.
We repeat this for j = 1, . . . , r. Let Q = Q1 ∪ · · · ∪ Qr be the set of all sets
Y ∈ R selected. For every Y ∈ Qj , let Y 0 = Y \ Uj−1 (this is the set of nodes first
 21.1. HYPERFINITENESS 387

covered by Y ). Let T0 be the set of all edges incident with any node of V \ Ur ,
and let T1 denote the set of all edges connecting any set Y ∈ Q to its complement.
Clearly every connected component of G − (T0 ∪ T1 ) has at most k nodes.
Next we show that
(21.1) ∂(Y ) ≥ ε2j−1 |Y \ Uj |
for every Y ∈ R (selected or not) and 1 ≤ j ≤ r. Suppose (by way of contradiction)
that ∂(Y ) < ε2j−1 |Y \ Uj |. Then Y ̸⊆ Uj , and hence it was not selected in the j-th
phase or before. But Y was eligible for selection throughout the j-th phase, and if
it was not selected, then (by the maximality of the family selected) Y must contain
a point of Uj,i \ Uj,i−1 for i = 1, . . . , k + 1, which is impossible since |Y | ≤ k. This
proves (21.1).
We want to bound the measure of T0 ∪T1 . We start with T0 . Select a random set
Y ∈ R from the distribution τ , and a random point y ∈ Y . Then y is distributed
according to λ, and hence by (21.1) and the definition of the fractional partition τ
we have
( |Y \ U | ) ( ∂(Y) )
j
(21.2) λ(V \ Uj ) = P(y ∈ / Uj ) = E ≤E ≤ 21−j
|Y| ε2j−1 |Y|
for every 1 ≤ j ≤ r. In particular, we have λ(V \ Ur ) ≤ 21−r ≤ 2ε/D by the choice
of r, and hence
∫
η(T0 ) ≤ deg(x) dx ≤ Dλ(V \ Ur ) ≤ 2ε.
V \Ur

Turning to T1 , we select a random point y of G again, and consider the set

Y ∈ Q that is the first set added containing y. Since sets added at the same time
are disjoint, this is well-defined, unless y ∈/ Ur , in which case we take Y = {y}. We
consider Y a random set, from some distribution
alpha). We can generate a random y by generating Y first according to α, and
then selecting a uniform random element of Y0 . Counting every edge in T1 with
its endpoint that was selected first (breaking ties arbitrarily), we have
( ∂(Y) ) ∑r
( )
η(T1 ) ≤ E ≤ ε 2j−1 λ(Uj ) − λ(Uj−1 )
|Y0 | j=1
∑
r
( )
=ε 2j−1 λ(V \ Uj−1 ) − λ(V \ Uj ) .
j=1

Doing partial summation and using (21.2) again,

∑
r
η(T1 ) ≤ ε 2j−1 λ(V \ Uj ) ≤ εr ≤ ε log(2D/ε).
j=1

Hence
η(T0 ∪ T1 ) ≤ 2ε + ε log(2D/ε) = ε log(8D/ε). 
Theorem 21.9 follows from our characterization of local equivalence (Theorem
18.59), Lemma 21.10 and the following rather simple couple of facts.
Proposition 21.11. Let φ : G1 → G2 be a local isomorphism between graphings
G1 and G2 .
388 21. ON THE STRUCTURE OF GRAPHINGS

(a) If G2 is (ε, k)-hyperfinite, then so is G1 .

(b) If G1 is fractionally (ε, k)-hyperfinite, then so is G2 .
Proof. (a) Let S ⊆ E(G2 ) be a Borel set such that every connected component
of G2 \ S is has at most k nodes. Then S ′ = φ−1 (S) is a Borel set in E(G1 ) with
η1 (S ′ ) = η2 (S). We claim that almost every connected component of G1 \ S ′ has
at most k nodes. Indeed, let X be the union of components of G1 \ S ′ with more
than
( k nodes,
) then for almost all x ∈ X, φ is an isomorphism on (G1 )x , and hence
φ (G1 )x is a connected subgraph of G2 \ S, which is a contradiction unless X has
measure 0.
(b) Let τ1 be a probability distribution on the set R1 of the connected induced
subgraphs of G1 with at most k nodes that is a fractional partition with ∂(τ ) ≤ ε.
Select Y ∈ R1 randomly according to τ1 ; then Z = φ(Y) is a random connected
induced subgraph of G2 with at most k nodes. Let τ2 denote the distribution of Z.
We claim that τ2 is a fractional partition. Indeed, we can generate a uniform
random point of Z by generating a uniform random point y ∈ Y, and taking z =
φ(y). (We use here that φ is a local isomorphism, and so it yields and isomorphism
between Y and Z.) Since y is distributed according to λG1 , it follows that z is
distributed according to λG2 .
By a similar argument, we have
( ∂(Z) ) ( ∂(Y) )
E =E ≤ ε.
|Z| |Y|
This proves (b). 
We note that the stronger assertion, namely that (ε, k)-hyperfiniteness of a
graphing would be invariant under local equivalence is false. As a simple example,
consider the two graphings Ca and C′′a (Exercise 18.60). These are locally equiv-
alent, but C′′a contains a perfect matching that is a Borel set (the set of edges of
the form (x, 1/2 + (x − a/2 mod 1/2)), where x ≤ 1/2), and hence it is (1/2, 2)-
hyperfinite. No such perfect matching exists in Ca (indeed, such a perfect matching
would give a measure-preserving involution by Lemma 18.19, and its complement
would be another one, contradicting the argument in Example 18.22). Hence Ca is
not (1/2, 2)-hyperfinite.
In this example, (ε, k)-hyperfiniteness is almost preserved, in the sense that Ca
is not (1/2, 2)-hyperfinite, but (1/2 + ε, 2)-hyperfinite for every ε > 0. We will see
that this is a general phenomenon among hyperfinite graphings (Corollary 21.18),
but not among all graphings, as the following example shows.
Example 21.12. For a 3-regular graph G, define a new graph G△ as follows. We
replace every node by a triangle (call these principal triangles), and then replace
every old edge by a copy of K4− , with the two nodes of degree 2 identified with the
endpoints of the edge (Figure 21.1). The graph G△ has 6n nodes and 21n/2 edges.
If G is bipartite, then G△ is (3/4, 3)-hyperfinite. Indeed, in this case V (G△ )
can be covered by disjoint triangles: we select all principal triangles corresponding
to nodes in one color class, and all non-principal triangles disjoint from them. The
number of remaining edges is |E(G△ )| − |V (G△ )| = 9n/2 = (3/4)|V (G△ )|.
Now let G be nonbipartite, and let S be a minimum set of edges such that every
connected component of G△ \ S has at most 3 nodes. Let A be the set of nodes in
G for which the corresponding principal triangle is a component of G△ \ S. It is
 21.1. HYPERFINITENESS 389

Figure 21.1. (3/4, 3)-hyperfiniteness is not even approximately

preserved by local equivalence

easy to check that for every edge of G induced by A or induced by V (G) \ A, there
is at least one node of degree at most one among the corresponding four nodes of
G△ \ S. The remaining nodes of G△ \ S have degree at most 2. Hence
1( )
|E(G△ ) \ S| ≤ |E(G)| − eG (A, V \ A) + 2(|V (G△ )| − |E(G)| + eG (A, V \ A))
2
9n
≤ + Maxcut(G),
2
and so
9n
|S| ≥ 6n − Maxcut(G) > ,
2
since G is nonbipartite. So G△ is (3/4, 3)-hyperfinite. For the appropriate choice
of G, we can prove more: let Gn be a random D-regular graph and G′n , a random
D-regular bipartite graph on n nodes, then (G′n )△ is (3/4, 3)-hyperfinite. On the
other hand, Maxcut(Gn ) < 1.41n with high probability (McKay [1982], see also
Hladky [2006]), and we get that (Gn )△ is not even (4/5, 3)-hyperfinite.
Let G and G′ be local-global limit graphings of the sequences (Gn )△ and
(Gn ) , respectively (or of appropriate subsequences), then G and G′ are locally
′ △

equivalent, G is (3/4, 3)-hyperfinite, but G′ is not even (4/5, 3)-hyperfinite. 

Our argument proving Theorem 21.9, which was motivated by an argument
of Schramm [2008], says that if a graphing is (ε, k)-hyperfinite, then every graph-
ing locally equivalent to it is (ε′ , k)-hyperfinite with a somewhat larger ε′ than ε
(namely, ε′ = O(ε log(1/ε))). We could have based another proof on the graph
partitioning algorithm of Hassidim, Kelner, Nguyen, and Onak; [2009] this would
show that if a graphing is (ε, k)-hyperfinite, then every graphing locally equivalent
to it is (ε, k ′ )-hyperfinite with some larger k ′ .
The following theorem was proved (in a different formulation, using involution
invariant random rooted graph models) by Schramm [2008].
Theorem 21.13. Let (Gn ) be a sequence of graphs in G, converging to a graphing
G. Then G is hyperfinite if and only if the family {Gn : n = 1, 2, . . . } is hyperfinite.
Proof. The “if” part is not hard. Suppose that (Gn ) is hyperfinite. Let ε > 0,
we want to show that G is hyperfinite. Let ε > 0, and let k ≥ 1 be chosen so that
for every n that is large enough there is a set Sn ⊆ V (Gn ) with |Sn | ≤ εv(Gn )
such that every connected component of Gn − Sn has at most k nodes. Consider
the pairs (Gn , Sn ) as graphs with their nodes 2-colored, and choose a subsequence
that is convergent as a sequence of colored graphs. The limit can be represented
390 21. ON THE STRUCTURE OF GRAPHINGS

by a colored graphing (G′ , S). It follows from the definition of convergence that
|Sn |/v(Gn ) → λ(S) (where λ is the node measure in (G′ , S)), and also that almost
all connected components of G′ − S have at most k nodes. Hence the uncolored
graphing G′ is hyperfinite. Since G′ and G are locally equivalent, it follows by
Theorem 21.9 that G is hyperfinite.
To prove the “only if” part, we invoke Theorem 19.16. By selecting an ap-
propriate subsequence, we may assume that the sequence (Gn ) is locally-globally
convergent, and so it has a limit graphing G′ such that δ⊙ nd
(Gn , G′ ) → 0 for ev-
ery k ≥ 0. Clearly G and G are locally equivalent, and hence G′ is hyperfinite
′

by Theorem 21.9. This means that for every ε > 0 there is an m ≥ 1 such that
V (G′ ) has a Borel 2-coloring with read and blue such that every connected m-node
subgraph contains a red point, and λ′ {red points} ≤ ε. These properties can be
read off from the 1-balls and the m-balls, respectively. It follows by the assump-
tion that Gn → G′ in the local-global sense that for a large enough n, Gn has
a 2-coloring such that the set Rn of red nodes satisfies |Rn | ≤ 2εv(Gn ), and the
number of m-neighborhoods that contain a connected blue subgraph with m + 1
nodes is at most εv(Gn ). Adding the roots of these m-neighborhoods to Rn , we
get a set Rn′ ⊆ V (Gn ) with |Rn′ | ≤ 3εn such that every connected component of
Gn − Rn has at most m nodes. 
We state another result, in a sense dual to Theorem 21.13:
Theorem 21.14. A graphing is hyperfinite if and only if it is the limit of a hyper-
finite graph sequence.
Proof. In view of Theorem 21.13, it suffices to prove that every hyperfinite
graphing is the limit of a locally convergent graph sequence. (So the Aldous–Lyons
conjecture holds for hyperfinite graphings.) Let G be a hyperfinite graphing, and
let ε > 0. Let S be a subset of edges with η(S) = ε such that every connected
component of G − S is finite. Proposition 19.1 implies that
(21.3) δ⊙ (G, G − S) ≤ 4ε1/ log(2D) .
For every graph F ∈ G, let aF be the measure ∑ of points in G − S whose
connected component is isomorphic
∑ to F . Since F a F = 1, we can
∑ choose a finite
set H of graphs such that F ∈H
/ a F ≤ ε/D. Let n > (D/ε) F ∈H v(F ), and
nF = ⌊aF n/v(F )⌋ (so that the rationals nF v(F )/n approximate the real numbers
aF with common denominator). For every F ∈ / H, let us delete the edges of all
connected components of G − S isomorphic to F . For every F ∈ H, let us delete
the edges of a set of connected components of G − S isomorphic to F so that the
remaining connected components cover a set of measure nF v(F )/n; it is not hard
to see that this can be done so that a Borel graph remains. The measure of the set
T of deleted edges can be bounded as follows:
∑ D ∑ D( nF v(F ) ) ε D ∑ v(F )
η(T ) ≤ aF + aF − ≤ + ≤ ε.
2 2 n 2 2 n
F ∈H
/ F ∈H F ∈H

Hence it follows just like above that

(21.4) δ⊙ (G − S, G − S − T ) ≤ 4ε1/ log(2D) .
Let G be a graph on n nodes consisting of nF copies of each F , together with
sufficiently many isolated nodes. Then G − S − T and G have the same connected
 21.1. HYPERFINITENESS 391

components, with the same frequencies; hence δ⊙ (G − S − T, G) = 0, and so by

(21.3) and (21.4),
δ⊙ (G, G) ≤ δ⊙ (G, G − S) + δ⊙ (G − S, G − S − T ) ≤ 8ε1/ log(2D) .
So G can be approximated arbitrarily well in the δ⊙ distance by finite graphs. 
Kaimanovich [1997] proved the following characterization of hyperfinite graph-
ings, which we quote without proof. (For a proof, see also Elek [2012a].)
Theorem 21.15. A graphing G is not hyperfinite if and only if it has a subgraphing
F such that ηG (E(F)) > 0 and there is an ε > 0 such that ∂F (Y ) ≥ εv(Y ) for every
finite connected subgraph Y of F.
We conclude our discussion of hyperfiniteness with a result of Hatami, Lovász
and Szegedy [2012] and Elek [2012a], showing that in the hyperfinite world, local
and local-global are equivalent.
Theorem 21.16. Any two locally equivalent hyperfinite atom-free graphings are
locally-globally equivalent.
As a preparation for the proof, we prove a somewhat stronger statement in a
special case.
Lemma 21.17. Let G and G′ be two locally equivalent graphings such that all
components of them are finite and have at most k nodes (k ≥ 1). Then for every
Borel m-coloring β of G′ there is a Borel m-coloring γ of G such that (G, γ) and
(G, β) are locally equivalent as colored graphings.
Proof. By Theorem 18.59, we may assume that there is a local isomorphism
φ : G′ → G or a local isomorphism φ : G → G′ . The second alternative is trivial,
so we assume the first.
Applying Theorem 18.3 to the graphing obtained by filling up every connected
component of G to a complete graph (which results in a Borel graph, cf. Exercise
18.7), we get that there is a Borel k-coloring α : V (G) → [k] such that any two
nodes in the same component have different colors.
For every isomorphism type F of k-colored graphs with at most k nodes, let
UF be the union of all connected components C of G for which C ∼ = F . It is easy
to see that the UF are Borel sets.
Next, we pull back the coloring to G′ : let UF′ = φ−1 (UF ) and α′ = φ ◦ α. It
follows from the definition of local isomorphism that every connected component
of G′ [UF′ ] is isomorphic to F as a colored graph (with colors according to α′ ).
Consider the given m-coloring β of G′ . On every connected component C of
UF , the colors according to α′ are all different, and hence β can be represented as
′

β = α′ ◦ fC with an appropriate map fC : [k] → [m]. For every isomorphism class

′
F of k-colored graphs with at most k nodes and every map f : [k] → [m], let UF,f
∼
be the union of all connected components C for which C = F and fC = f . This
partitions every set UF′ into mk sets UF,f , which are Borel (as it is easy to see).
′
The images φ(UF,f ) are not necessarily Borel sets, but we can partition every
′
k
set UF into m Borel sets UF,f so that λG (UF,f ) = λG′ (UF,f ). Let us color x ∈ UF,f
with color f (α(x)), to get an m-coloring γ. Then the whole component of a point
′
x ∈ UF,f is colored by γ the same way as the component of any y ∈ UF,f . This
proves that γ satisfies the conclusion of the Lemma. 
392 21. ON THE STRUCTURE OF GRAPHINGS

Proof of Theorem 21.16. As above, we may assume that there is a local

isomorphism φ : G′ → G. We want to prove that for every m-coloring β of
G′ there is an m-coloring of G for which the sampling distance of these colored
graphings is arbitrarily small; and vice versa. In fact, the “vice versa” part is trivial
(we can just pull back the k-coloring of G by φ). The first assertion, however, takes
more work.
Let ε > 0. By hyperfiniteness, there is a set S ⊆ E(G) and a k ≥ 1 such that
ηG (S) ≤ ε′ , where ε′ = 14 εlog(2D+2) , and every connected component of G \ S has
at most k nodes. Let S ′ = φ−1 (S), then ηG′ (S ′ ) = ηG (S) ≤ ε′ and φ is a local
isomorphism from G′ \ S ′ to G \ S.
Since β is a Borel m-coloring of G′ , Lemma 21.17 implies that there is a Borel
m-coloring γ of G \ S such that (G \ S, γ) and (G′ \ S ′ , β) are locally equivalent as
colored graphings. By Proposition 19.1, we have
ε
δ⊙ ((G, γ), (G \ S, γ)) ≤ 2d1 ((G, γ), (G \ S, γ))1/ log(2D+2) ≤ ,
2
and similar inequality holds for G′ . Hence
δ⊙ ((G, γ), (G′ , β)) ≤ δ⊙ ((G, , γ), (G \ S, γ)) + δ⊙ ((G \ S, γ), (G′ \ S ′ , β))
ε ε
+ δ⊙ ((G′ \ S ′ , β), (G′ , β)) ≤ + 0 + = ε. 
2 2
This theorem has some interesting consequences. We have seen (Example 21.12)
that a graphing may be (ε, k) hyperfinite and a locally equivalent graphing may
not be (16ε/15, k) hyperfinite. However, if the graphing is hyperfinite, this cannot
occur:
Corollary 21.18. Let G and G′ be two hyperfinite locally equivalent graphings,
and assume that G is (ε, k)-hyperfinite. Then G′ is (ε′ , k)-hyperfinite for every
ε′ > ε.
The second corollary shows that the two notions of convergence discussed in
sections 19.1 and 19.2 are equivalent for hyperfinite graph sequences.
Corollary 21.19. Every locally convergent hyperfinite graph sequence is locally-
globally convergent.
Proof. Let (Gn ) be a locally convergent hyperfinite sequence, then it converges
locally to a hyperfinite graphing G by Theorem 21.13. If (Gn ) does not converge
locally-globally to G, then it has a locally-globally convergent subsequence whose
limit graphing G′ is not locally-globally equivalent to G. Since G and G′ are
locally equivalent, this contradicts Theorem 21.16. 

Exercise 21.20. Let G ∈ G be an (ε, k)-hyperfinite graph, and let 0 ≤ δ ≤ ε.

Prove that there exists an (ε − δ, k) hyperfinite graph G′ for which δ⊙ (G, G′ ) ≤
4δ 1/ log D . State and prove the analogous assertion for graphings.
Exercise 21.21. Prove that for every planar √ graph √ G on n nodes and every
integer K ≤ n, one can delete at most 60n/ K − 30 n nodes from G so that
every connected component of the remaining graph has at most K nodes. (The
strange formula is given as help, to facilitate induction.)
Exercise 21.22. Prove that every family of graphs in G with subexponential
growth is hyperfinite.
Exercise 21.23. Formulate and prove a version of Corollary 21.18 for finite
graphs.
 21.2. HOMOGENEOUS DECOMPOSITION 393

21.2. Homogeneous decomposition

Hyperfinite graphs can be decomposed into bounded size graphs by deleting
a small fraction of the edges. How far can we simplify a general bounded degree
graph by deleting a small fraction of the edges? It was proved recently by Angel and
Szegedy [unpublished], and independently by Elek and Lippner [2011], that every
graph with degrees bounded by D can be decomposed into a bounded number of
“homogeneous” parts by deleting small number of edges.
To be precise, let us call a subset U ⊆ V (G) an (ε, δ)-island, if eG (U, V (G)\U ) ≤
δv(G), |U | ≥ εv(G), and δ⊙ (G[U ], G) ≥ ε. We say that a graph G ∈ G (ε, δ)-
homogeneous, if it contains no (ε, δ)-island. Clearly, an (ε, δ)-island is also an
(ε′ , δ ′ )-island if ε′ ≤ ε and δ ′ ≥ δ. Hence an (ε, δ)-homogeneous graph is also
(ε′ , δ ′ )-homogeneous, if ε′ ≤ ε and δ ′ ≥ δ.
Example 21.24. An m × m grid G = Pm Pm is (ε, δ)-homogeneous if δ ≤ ε2 /18
and m is large enough. Indeed, suppose that U ⊆ V (G) is an (ε, δ)-island. Consider
any r ≥ 0. We claim that most nodes of G[U ] are “orderly” in the sense that they
have the same r-neighborhood as a node in an infinite grid. Indeed, if v ∈ U is
not orderly, then either it is closer to the boundary than r, or the r-ball around it
contains one of the edges leaving U . It is easy to check that any edge leaving U can
be counted at most 2r2 times, hence the number of “disorderly” nodes is at most
4rm + 2r2 δm2 < 3r2 δm2 (if m is large enough). The proportion of “disorderly”
nodes in the whole grid G is even smaller, and hence
r 3r2 δm2 3r2 δ
δ⊙ (G, G[U ]) < 2
= .
εm ε
Summing,
∞
∑ 3r2 δ 18δ
δ⊙ (G, G[U ]) < 2−r = ≤ ε.
r=0
ε ε
This shows that U is not an (ε, δ)-island. 
Theorem 21.25. For every ε > 0 there is a δ > 0 such that from every graph G ∈ G
we can delete εv(G) edges in such a way that every component of the remaining
graph is (ε, δ)-homogeneous.
The dependence of δ on ε is explicit, and only moderately bad: choose r =
r
1 + ⌈log(1/ε)⌉ (so that 2r ≈ ε/2), let b = |Br | ≤ DD be the number of r-balls
(Exercise 18.42), and define δ = ε5 /(4Dr b) = 2−1/ε
O(log D)
.
Proof. The proof follows the argument of Angel and Szegedy, which is rem-
iniscent of the proof of the Regularity Lemma. By (19.3), an (ε, δ)-island U will
satisfy dvar (ρG[U ],r , ρG,r ) ≥ ε/2.
For a graph G ∈ G with connected components G1 , . . . , Gk , we define
∑
k
v(Gi ) ∑
f (G) = ρGi ,r (B)2 .
i=1
v(G)
B∈Br

Trivially, 0 ≤ f (G) ≤ 1.
Let, say, G1 , . .∑
. , Gm be those components of G that are not (ε, δ)-homogeneous,
and suppose that i=1 v(Gi ) = p > (ε/2D)n. Let Vi′ ⊆ V (Gi ) be an (ε, δ)-island,
m

and let Vi = V (Gi ) \ Vi′ . Let Ci be the set of edges connecting Vi and Vi′′ , then
′′

|Ci | ≤ δ|Vi′ |. Finally, let G′ be obtained from G by removing the edges in the sets
394 21. ON THE STRUCTURE OF GRAPHINGS

Ci . We want to show that if many of the parts are not ε-homogeneous, then f (G′ )
is substantially larger than f (G). To keep the notation in check, set G′i = G[Vi′ ],
G′′i = G[Vi′′ ], n = v(G), ni = v(Gi ), n′i = v(G′i ) etc. Since the radius r is fixed, we
don’t have to show it in notation, and write ρi = ρGi ,r , ρ′i = ρG′i ,r etc.
Fix any i ∈ [k] and any B ∈ Br , and consider the difference of their contribu-
tions to f (G′ ) and f (G):
n′i ′ n′′ ni
(21.5) ρi (B)2 + i ρ′′i (B)2 − ρi (B)2
n n n
n′i ( ′ )2 n′′ ( )2
= ρi (B) − ρi (B) + i ρ′′i (B) − ρi (B)
n n
2 ( ′ ′ ′′ ′′
)
+ ρi (B) ni ρi (B) + ni ρi (B) − ni ρi (B) .
n
Here the first term will provide the gain, the second is nonnegative, while the third
is an error term. To estimate the “gain” term, first we sum over all balls B:
∑( )2 1 (∑ ′ )2 4 ε2
ρ′i (B) − ρi (B) ≥ |ρi (B) − ρi (B)| = dvar (ρ′i , ρi )2 ≥ .
b b b
B B

Summing over i and using that n′i ≥ εni by the definition of an island, we get
∑ n′i ( ′ )2 ∑ ni ε2 ε4
ρi (B) − ρi (B) ≥ ε ≥ .
n i
n b Db
i,B

To estimate the error term, we argue that the quantity |n′i ρ′i (B) + n′′i ρ′′i (B) −
ni ρi (B)| is the increase or decrease in the number of neighborhoods isomorphic to
B when the edges in Ci are deleted; since deletion of an edge can change at most
2Dr balls with radius r, we have
∑ 
n′i ρ′i (B) + n′′i ρ′′i (B) − ni ρi (B) ≤ 2Dr |Ci |,
B

and so the total contribution of the error term is at most

2Dr ∑ ε4
|Ci | ≤ 2Dr δ < .
n i 2Db

of f (G) increases by at least ε4 /(2Db). The number

It follows that the value ∑
of edges deleted is at most i |Ci | ≤ δn. We can repeat this until the number
of nodes in non-(ε, δ)-homogeneous components drops below εn/D. This happens
after at most 2Db/ε4 repetitions (since f (G) ≤ 1), and when we get stuck, the
number of deleted edges is at most (2Dbδ/ε4 )n < (ε/2)n, and the number of nodes
in those components that are not (ε, δ)-homogeneous is less than (ε/D)n. Deleting
all edges in these components means the deletion of no more than (ε/2)n further
edges. This turns these remaining components into isolated nodes, which count as
(ε, δ)-homogeneous components. This proves the theorem. 

21.2.1. The quest for a regularity lemma. Is there a good analogue of

the Regularity Lemma for bounded degree graphs? The Regularity Lemma, as
discussed in Chapter 9, does not say anything about non-dense graphs.
What do we expect from such a lemma? If we think about the many uses of
the dense Regularity Lemma, there is no single answer to this question.
 21.2. HOMOGENEOUS DECOMPOSITION 395

• It gives a partition of the nodes such that most bipartite graphs between
different classes are homogeneous (random-like). Several extensions of the Regular-
ity Lemma to sparse graphs in this sense are known (see e.g. Kohayakawa [1997],
Gerke and Steger [2005], Scott [2011]), but they are more-or-less meaningless, or
very weak, for graphs that have bounded degree.
• It gives a decomposition of the graph into simpler, homogeneous subgraphs.
Theorem 21.25 describes such a decomposition. However, this result is clearly not
the ultimate word: the (ε, δ)-homogeneous pieces it produces can still have a very
complicated structure.
• It implies that an arbitrarily large (simple, dense) graph can be “scaled
down” to a graph whose size depends on the error bound only, and which is almost
indistinguishable from the original by sampling. Proposition 19.10 shows that such
a “downscaling” is also valid for bounded degree graphs; Unfortunately, it is non-
effective, and provides no algorithm for the construction of the smaller graph.
• It provides an approximate code for the graph, which has bounded size (de-
pending on the error we allow), from which basic parameters of the graph can be
reconstructed, and from which graphs can be generated on an arbitrary number of
nodes that are almost indistinguishable from the original graph by sampling. In
this sense, a Regularity Lemma may exist, and should be very useful once we learn
how to work with it. While not quite satisfactory, I feel that the results mentioned
above justify cautious optimism.
 CHAPTER 22

Algorithms for bounded degree graphs

The algorithmic theory of large graphs with bounded degree is quite extensive.
Similarly as in the case of dense graphs, we can formulate the problems of parameter
estimation, property distinction, property testing, and computing a structure.
However, it seems that the theory in the bounded degree case is lacking the
same sort of general treatment as dense graphs had, in the form of useful general
conditions for parameter estimations (like Theorem 15.1), treatment of property
distinction in the limit space (Section 15.3), and the use of regularity partitions
and representative sets in the design of algorithms (Section 15.4). The most im-
portant tools that are missing are analogues of the Regularity Lemma and of the
cut distance.
Our discussions in this chapter, accordingly, will be more an illustration of
several interesting and nontrivial results than a development of a unifying theory.
But even so, graph limit theory provides a useful point of view for these results.

22.1. Estimable parameters

We call a graph parameter defined on bounded degree graphs estimable, if it is
bounded, and for every ε > 0 there is a positive integer k and an “estimator” func-
tion g : (Bk )k → R such that for every graph G ∈ G and uniform, independently
chosen random nodes v1 , . . . , vk ∈ V (G), we have
( )
(22.1) P f (G) − g(BG,k (v1 ), . . . , BG,k (vk ))| > ε ≤ ε.
In other words, g estimates f from a sample chosen according to the rules of sam-
pling from a bounded degree graph. (For convenience, we use the same ε to bound
the error in the function value and the probability of a large error; also the same k
for the number of samples and the radius of balls we explore around the sampling
points. In specific algorithms, one may want to distinguish these values, but this
would not alter the notion of estimability.)
In the dense case, we did not need a separate estimator function g; we could
use g = f . This is not the case here.
Example 22.1. Let f (G) be the fraction of nodes of G of degree 1. If G is a
3-regular graph with large girth, then every sample BG,k (v) is a tree with more
than half of its nodes of degree 1; but G itself has no nodes of degree 1. 
It is also easy to see that it would not be enough to use just one sample ball.
Example 22.2. Let G be a 2-regular graph on n nodes, G′ , a 3-regular graph on n
nodes, and GG′ , their disjoint union. Let the parameter to estimate be the average
degree. In a single sample you see only nodes of degree 2 or nodes of degree 3, no
matter how far you explore the graph. No matter how large neighborhoods you
397
398 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

take, and what function of them you compute, this single bit of information (degree
2 or degree 3) will not distinguish three possibilities (G, G′ and GG′ ). 
On the other hand, some other facts extend from the dense case with more or
less difficulty. The following theorem of Elek [2010a] connects parameter estimation
with convergence (recall that the analogous result for dense graphs was trivial).
Theorem 22.3. A bounded graph parameter f is estimable if and only( if for )ev-
ery locally convergent graph sequence (Gn ), the sequence of numbers f (Gn ) is
convergent.
Proof. The “only if” part is easy: from similar graphs we get similar sam-
ples and so we compute similar estimates. Let us make this precise. Suppose
that f is estimable, and let (Gn ) be a locally convergent graph sequence. Let
0 < ε < 1/8, we want to show that |f (Gn ) − f (Gm )| < ε if n, m are large enough.
By the definition of estimability, we have a positive integer k and an estimator
function g : (Bk )k → R such that (22.1) holds. If n, m are large enough, then
δ⊙ (Gn , Gm ) ≤ 1/(4k2k ), and hence dvar (ρk,Gn , ρk,Gm ) ≤ 1/(4k). This means that
we can couple a random node v ∈ V (Gn ) with a random node u ∈ V (Gm ) so that
BGn ,k (v) ∼= BGm ,k (u) with probability at least 1 − 1/(4k). If we sample k indepen-
dent nodes v1 , . . . , vk from Gn and k independent nodes u1 , . . . , uk from Gm , then
with probability more than 3/4, we have BGm ,k (u1 ) ∼ = BGn ,k (v1 ), . . . , BGm ,k (uk ) ∼ =
BGn ,k (vk ). With positive probability, we have (simultaneously BGm ,k (u1 )) ∼ =
BGn ,k (v ), . . . , B (u ) ∼
= B (v ),  f (G ) − g B (v ), . . . , B (v )  ≤
1 Gm(,k k Gn ,k k n) Gn ,k 1 Gn ,k k
ε and f (Gm ) − g BGm ,k (u1 ), . . . , BGm ,k (uk )  ≤ ε. But in this case we have
|f (Gn ) − f (Gm )| ≤ 2ε, which we wanted to prove.( )
The converse is a bit trickier. Suppose that f (Gn ) is convergent for every
locally convergent graph sequence (Gn ). Given ε > 0, we want to find a suitable
positive integer k and construct an estimator g : (Bk )k → R. The condition on
f implies that for every ε > 0 there is an ε′ > 0 such that if δ⊙ (G, G′ ) ≤ ε′ then
|f (G) − f (G′ )| ≤ ε. Let r be chosen so that 21−r < ε′ , and let k > 2r/(εε′ ).
The estimator we construct will only depend on the r-balls around the roots
of the k-balls. So we will construct a function g : (Br )k → R. For every sequence
b = (B1 , . . . , Bk ) ∈ (Br )k , let ρb denote the distribution of a randomly chosen
element of the sequence. We define the estimator as follows:

 ′
f (G) where G is any graph with dvar (ρG,r , ρb ) ≤ ε /4,
g(b) = if such a graph exists,


0 otherwise.
To show that this is a good estimator, let G ∈ G be any graph, ( and let v1 , . . . , vk )∈
V (G) be uniformly chosen random nodes, and let b = BG,r (v1 ), . . . , BG,r (vk ) .
By the choice of k, elementary probability theory gives that with probability at
least 1 − ε, we have dvar (ρb , ρG,r ) ≤ ε′ /4. If this happens, then in the definition
of g(b) the first alternative applies, and so g(b) = f (G′ ) for some graph G′ that
satisfies dvar (ρG′ ,r , ρb ) ≤ ε′ /4. This implies that dvar (ρG′ ,r , ρG,r ) ≤ ε′ /2. Then we
have by (19.3)
1 ε′
δ⊙ (G, G′ ) ≤ r + ≤ ε′ .
2 2
By the definition of ε′ , this implies that |f (G) − f (G′ )| ≤ ε. 
 22.1. ESTIMABLE PARAMETERS 399

Corollary 22.4. For every estimable graph parameter f there exists a graphing pa-
rameter fb that is continuous in the δ⊙ distance such that f (Gn ) → fb(G) whenever
Gn → G.
Notice that continuity in the δ⊙ distance implies invariance under local equivalence.

Proof. It is easy to see (using Theorem 22.3) that the parameter fb is uniquely
determined for graphings that represent limits of convergent graph sequences, and
it is continuous in the δ⊙ distance. However, we don’t know if all graphings are like
that (cf. Conjecture 19.8). To complete the proof, we can use Tietze’s Extension
Theorem to extend the definition of fb to all graphings. 
This possible non-uniqueness of the extension may be connected with the fact
that it is typically not easy to see the “meaning” of the extension of quite natural
graph parameters (cf. Supplement 20.17).
Our discussion in Section 20.2 shows that parameters of the type G 7→
ent∗ (G, H) = log t(G, H)/v(G) are estimable provided the weighted graph H is
sufficiently dense. In the next two sections we describe a couple of further interest-
ing examples of estimable graph parameters.
Not all natural parameters of graphs are estimable (see Examples 20.16 and
22.5). However, it was shown by Elek [2010a] that if we restrict ourselves to testing
properties on hyperfinite graphs, then many of these become testable. The method
is similar to property testing for hyperfinite graphs, which will be discussed later.
Example 22.5 (Independence ratio). Recall that α(G) denotes the maximum
size of a stable set in graph G. The independence ratio α(G)/v(G) is not estimable.
Let Gn be a random D-regular graph on 2n nodes, and G′n be a random bipartite
D-regular graph on 2n nodes. It is clear that α(G′n ) = n. In contract, α(Gn ) ≤
(1 − 2cD )n with high probability, where cD > 0 depends only (Bollobás [1980]).
The interlaced sequence (G1 , G′1 , G2 , G′2 , . . . ) is locally convergent (as discussed in
Example 19.7), but the independence ratios oscillate between 1/2 and something
less than 12 − cD , so they don’t converge.
However, it we restrict ourselves to the sequence Gn , then the independence
ratios form a convergent sequence; this is a recent highly nontrivial result of Bayati,
Gamarnik and Tetali [2011]. 
22.1.1. Number of spanning trees. Lyons [2005] proved that the number of
spanning trees tree(G), suitably normalized, is an estimable parameter of bounded
degree graphs. He in fact proved a more general result, allowing the degrees to be
unbounded, as long as the average degree remains bounded and the degrees don’t
vary too much; we treat the bounded case only, and refer for the exact statement
of the more general result to the paper.
Let G be a connected graph with n nodes and m edges, whose degrees are
bounded by D as always in this part of the book. It is easy to see that tree(G) ≤
Dv(G) ; a bit sharper,
∏
tree(G) ≤ deg(v),
v∈V (G)
whence
1 1 ∑
log tree(G) ≤ log deg(v).
n n
v∈V (G)
400 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

The right hand side is clearly bounded and estimable, which reassures us that we
have the right normalization.
( )
Theorem 22.6. The graph parameter log tree(G) /v(G) is estimable for connected
bounded degree graphs.
Proof. Let G be a connected graph with all degrees bounded by D. It will
be convenient to choose D generously so that all degrees are in fact at most D/2.
We add D − deg(v) loops to each node v, to make the graph regular (here, a loop
adds only 1 to the degree) and also to make sure that its adjacency matrix A is
positive semidefinite. This does not change the number of spanning trees, and from
the samples of the original graph the samples of this augmented graph are easily
generated just by adding loops.
We start with developing formulas for tree(G) and its logarithm. Here we face
an embarrassment of riches: there are many formulas for tree(G) in the literature,
and possibly others would also work. We use (5.41), which we write as
1 ( t∗ (Cr , G) 1)
∑∞
1 n−1 log n
(22.2) log tree(G) = log D − − (log e) − .
n n n r=1
r Dr n
For every fixed r, the quantity t∗ (Cr , G) − 1/n is estimable. Since the other terms
in (22.2) are trivially estimable, we are almost done. But the problem is that we
have an infinite sum, and we need a convergent majorant. (This is where it becomes
important that we have subtracted 1/(nr) in every term!)
Lemma 22.7. For any r ≥ 0 and v ∈ V (G),
1 homv (Cr• , G) 1 2D1/3
≤ ≤ + .
n Dr n (r + 1)1/3
(It may help with the digestion of this formula that Dr = homv (Pr• , G), and so
the ratio in the middle expresses the probability that a random walk started at v
returns to v after r steps. Since the endpoint of a random walk becomes more and
more independent of the starting point, this probability tends to 1/n by elementary
properties of random walks. The main point is that the upper bound gives a uniform
bound on the rate of this convergence.)
Averaging over all nodes v, the lemma implies that
t∗ (Cr , G) 1 2D1/3
(22.3) 0≤ r
− ≤ .
D n (r + 1)1/3
This gives a convergent majorant, independent of G, for the infinite sum in (22.2),
which proves that (1/n) log tree(G) is estimable. 
Proof of Lemma 22.7. Let P = (1/D)A (this is the transition matrix of the
random walk on G), and yr = P r 1v (this is the distribution of a random walk after
r steps). Clearly t∗v (Cr• , G)/Dr = 1T v P 1v = 1v yr . Since P is positive semidefinite,
r

we see from here that the values yr (v) are monotone decreasing, and since P r → n1 J
(where J is the all-1 matrix), yr (v) → n1 as r → ∞. This implies the lower bound
in the lemma.
To get the upper bound, we note that
∑∞ ∑∞
1
(22.4) ytT (I − P )yt + ytT (P − P 2 )yt = 1 −
t=0 t=0
n
 22.1. ESTIMABLE PARAMETERS 401

Indeed, the matrices I − P and P − P 2 are positive semidefinite, hence all terms
here are nonnegative. Furthermore, P yt = yt+1 , so if we stop the sums at m steps,
then the middle terms telescope out, and we are left with y0T y0 − ym+1 T
ym+1 , where
y0 y0 = 1 and ym+1 ym+1 → 1/n.
T T

From (22.4) it follows that there is a t ≤ r such that ytT (I − P )yt ≤ 1/(r + 1).
Let x = 12 (yt (v) + n1 ), and let u be the closest node to v with yt (u) ≤ x. Consider a
shortest path v0 v1 . . . vk , where v0 = v and vk = u. Since yt (v0 ), . . . , yz (vk−1 ) ≥ x,
we must have k ≤ 1/x. On the other hand,
( )2 ( )2
yt (v) − x ≤ yt (v0 ) − yt (vk )
( ) ( )2
= (yt (v0 ) − yt (v1 ) + · · · + yt (vk−1 ) − yt (vk ))
( )
≤ k (yt (v0 ) − yt (v1 ))2 + · · · + (yt (vk−1 ) − yt (vk ))2
Dk D
≤ DkytT (I − P )yt ≤ ≤ .
r+1 x(r + 1)
Hence
D
(22.5) (yt (v) − x)2 x ≤ .
r+1
Substituting the definition of x, we get
( 1 )3 8D
yt (v) − ≤ 8(yt (v) − x)2 x ≤ .
n r+1
Since we know that yr (v) ≤ yt (v), this proves the lemma.
We note that Lyons gets a better estimate, with 1/2 in the exponent of r + 1
rather than 1/3, but the simpler bound above was good enough for our purposes.

( )
Once we know that the graph parameter log tree(G) /v(G) is estimable,
we
( also know ) that if Gn is a locally convergent graph sequence, then
log tree(Gn ) /v(Gn ) tends to a limit. From the proof, it is not difficult to fig-
ure out what the limiting graphing parameter (or involution-invariant-distribution-
parameter) is. We formulate the answer for a graphing, but it is easy to translate
this to the Benjamini–Schramm model. Given a graphing G and a number D such
that D/2 is an upper bound on the degrees, pick a random node x, and start a
random walk from x, where you have to add D − deg(y) loops to node y as you
go along. Let Xr be the indicator that the random walk returns to x after r steps
(not necessarily the first time). With this notation, we have
∑∞
log tree(Gn ) 1
(22.6) −→ log D − E(Xr ).
v(Gn ) r=1
r

The expression on the right describes the limit as a function of the limiting graphing.
(Note that its value may be −∞.)

Exercise
∑ 22.8. Suppose that we want to estimate the variance of the degrees, i.e.,
v∈V (G) (deg(v) − d0 )2 /v(G), where d0 is the average degree. (a) Show that this
parameter is (estimable. (b) Prove that ) we
( cannot estimate it using an estimator
)
of the form g BG,k (v1 ), . . . , BG,k (vk ) = h(BG,k (v1 ))+· · ·+h(BG,k (vk )) /k with
any function h : Bk → R.
402 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

22.2. Testable properties

22.2.1. Distinguishing properties. Similarly as in the dense case (Section
15.2), our next step is discuss the problem of distinguishing two disjoint graph
properties. The setup is very similar to the dense case. Let P1 , P2 ⊆ G be two
graph properties with P1 ∩ P2 = ∅. We call properties P1 and P2 distinguishable
by sampling, if there exist positive integers r and k, and a property Q of k-tuples
of r-balls, such that for every graph G ∈ G and random nodes v1 , . . . , vk ,
 2

( ) ≥ 3 , if G ∈ P1 ,
P (BG,r (v1 ), . . . , BG,r (vk )) ∈ Q

≤ 1 , if G ∈ P .
2
3
The following analogue of Theorem 15.8 is due to Benjamini, Schramm and
Shapira [2010].
Theorem 22.9. Two graph properties P1 , P2 ⊆ G are distinguishable by sampling
if and only if δ⊙ (P1 , P2 ) > 0.
Proof. The necessity of the condition is easy to prove, along the same lines
as it was done for dense graphs in the proof of Theorem 15.8; we don’t go into the
details. The sufficiency is done differently, and unfortunately it is not constructive.
Suppose that δ⊙ (G1 , G2 ) = ε > 0. Similarly as in the proof of Proposition
19.10, we can select a finite set Q1 ⊆ P1 of graphs such that for every graph
G ∈ P1 there is a graph H ∈ Q1 such that δ⊙ (G, H) ≤ ε/4. Now if we want to
decide whether G ∈ P1 or G ∈ P2 , then we compute δ⊙ (G, H) with error less than
ε/4 for all H ∈ Q1 . (This can be done with error probability less than 1/3 by taking
a large enough sample of large enough balls.) If there is an H for which we find
that δ⊙ (G, H) ≤ ε/2, we conclude that G ∈ P1 . If no such H exists, we conclude
that G ∈ P2 . It is straightforward to check that if G ∈ Pi , then the answer will be
correct with probability larger than 2/3. 
22.2.2. Testable properties and their closures. Now we come to testing
a single property P. Just as in the dense case, we either want to conclude that a
given graph G does not have the property, or that one can change a small number
of adjacencies so that the property is restored. To be more precise, let P ε = {G ∈
G : d1 (G, P) > ε}. We say that a property P of bounded degree graphs is testable
if for every ε > 0 there are integers r = r(ε) ≥ 1 and k = k(ε) such that sampling
k neighborhoods of radius r from a graph G ∈ G, we can compute “YES” or “NO”
so that:
(a) if G ∈ P, then the answer is “YES” with probability at least 2/3;
(b) if G ∈
/ P ε , then the answer is “NO” with probability at least 2/3.

Example 22.10 (Forests). Let us look at a simple example that illustrates some of
the difficulties in designing algorithms for property testing for graphs with bounded
degree, even for monotone properties. Suppose that we want to test whether a graph
G is a forest. Our first thought might be to test whether a random ball contains a
cycle. Certainly, if it does, then the graph is not a forest. But drawing a conclusion
in the other direction is not justified: if the graph G has large girth, then every ball
will be tree, and G would be very far from being a forest. This shows that (unlike
in the dense case) P is not a good test property for itself. If in addition to this
 22.2. TESTABLE PROPERTIES 403

we estimate the average degree and eliminate small components, we can design a
test for being a forest (Goldreich and Ron [2008]). To fill in the details makes an
interesting exercise. 
We can use limit objects to give the following condition for testability (the
proof is immediate).
Proposition 22.11. A graph property P is not testable if and only if there exists
an ε > 0 and two convergent sequences of graphs (Gn ) and (Hn ) with Gn ∈ P and
d1 (Hn , P) > ε that have a common local limit.
22.2.3. Hyperfinite properties. Hyperfiniteness is particularly important
in property testing. Just as in the dense case, property testing is about the interplay
between the sampling distance and the edit distance, and these two distances are
intimately related for hyperfinite graphs.
Benjamini, Schramm and Shapira show that hyperfiniteness is in a sense
testable. This does not make sense as said, since hyperfiniteness is a property
of a family of graphs, not of a single graph. But if we quantify hyperfiniteness,
then we can turn it into a meaningful statement.
Proposition 22.12. For every ε there is an ε′ such that for any positive integer
k, the properties P1 = {(ε′ , k)-hyperfinite} and P1 = {not (ε, k)-hyperfinite} are
distinguishable.
Proof. Suppose that this is false, then by Theorem 22.9 there exist an ε > 0,
a sequence εn → 0, graphs Gn , G′n ∈ G and positive integers kn such that Gn is
(εn , kn )-hyperfinite, G′n is not (ε, kn )-hyperfinite, and δ⊙ (Gn , G′n ) → 0. Then the
sequence (Gn ) is hyperfinite. Let us select a convergent subsequence, then the limit
graphing of this subsequence is hyperfinite by Theorem 21.13. But the sequence
(G′n ) has the same limit, so it must be hyperfinite, by the same theorem. This
implies that there is a positive integer k such that all members of the sequence
are (ε, k)-hyperfinite. Since (G′n ) is not (ε, kn )-hyperfinite, it follows that kn < k
for all n. This implies that almost all connected components of G have at most k
elements, but then it follows from Gn → G that all but a o(1) fraction of connected
components of G′n have at most k elements, which implies that (G′n ) is an (ε, k)-
hyperfinite sequence. 
Benjamini, Schramm and Shapira [2010] proved an important analogue of The-
orem 15.24: every minor-closed property of bounded degree graphs is testable. As
noted by Elek, the theorem can be extended to any monotone hyperfinite graph
property.
Theorem 22.13. Every monotone hyperfinite property of graphs with bounded de-
gree is testable.
The property of being a forest is certainly minor-closed, so the example discussed
at the end of the introduction above is a special case. As another special case,
planarity of bounded degree graphs is testable.
Proof. Let P be a monotone hyperfinite graph property, and suppose that it is
not testable. Then there exist an ε > 0 and two sequences of graphs (Gn ) and (Fn )
such that Gn ∈ P, d1 (Fn , P) > ε and δ⊙ (Gn , Fn ) → 0. We may assume that both
sequences are locally convergent, and so they have a common weak limit graphing
404 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

G. Since P is hyperfinite, so is the sequence (Gn ). Theorem 21.13 implies that

G is hyperfinite, and applying this theorem again, we get that (Fn ) is hyperfinite.
By Theorem 21.19, the interlaced sequence G1 , F1 , G2 , F2 , . . . is locally-globally
convergent, and hence we may assume that Gn , Fn → G in the local-global sense.
Since G is hyperfinite, it has a Borel 2-coloring α : V (G) → [2] such that
λ(α−1 (1)) ≤ ε′ and there is a k ∈ N such that every connected subgraph with
k + 1 nodes contains a point v with α(v) = 1. By local-global convergence,
if n is large enough, then Gn has a 2-coloring αn : V (Gn ) → [2] such that
δ⊙k
((G, α), (Gn , αn )) ≤ ε′ /k. In particular, |αn−1 (1)| ≤ 2ε′ v(Gn ) and the union
of connected subgraphs with k + 1 nodes that contain no node with color 1 has at
most ε′ v(Gn ) nodes. The graph Fn has a 2-coloring βn with similar properties. It
follows that δ⊙ k
((Gn , αn ), (Fn , βn )) ≤ 2ε′ .
′
Let Gn be obtained from Gn by deleting all edges incident with any node of
Sn as well as all edges in connected components of Gn [Tn ] that have more than k
nodes. This way we delete at most 3ε′ Dv(Gn ) edges. Furthermore, every connected
component of G′n has at most k nodes. We define Fn′ analogously.
It is important that whether or not an edge is deleted is determined locally,
which implies that whenever v ∈ V (Gn ) and u ∈ V (Fn ) satisfied BGn ,αn ,k (v) ∼ =
BFn ,βn ,k (u), then also BG′n ,αn ,k (v) ∼ = BFn′ ,βn ,k (u), which means simply that the
connected component of G′n containing v is isomorphic to the connected component
of Fn′ containing u. Hence δ⊙ k
(G′n , Fn′ ) ≤ δ⊙
k
((Gn , αn ), (Fn , βn )) ≤ 2ε′ . Let Yk
denote the set of connected graphs with at most k nodes (up to isomorphism), let
aY denote the number of connected components of G′n isomorphic to Y ∈ Yk , and
let bY be defined analogously. In these terms, we have
∑  aY v(Y ) bY v(Y ) 
 −  ≤ 4ε′ .
v(Gn ) v(Fn )
Y ∈Yk
( )
We may assume that v(Gn ) ≥ v(Fn ). Let cY = min bY , ⌊aY v(Fn )/v(Gn )⌋ .
Let us keep cY copies of every Y ∈ Yk in Fn′ and delete the edges of the rest, to get
a graph Fn′′ . The number of edges to delete is bounded by 2ε′ Dv(Fn ) + Dk|Yk | <
(ε/2)v(Fn ) if n is large enough, and so d1 (Fn , Fn′′ ) ≤ 3ε′ D + ε/2 < ε. Furthermore,
Fn′′ is isomorphic to a subgraph of Gn , and hence Fn′′ ∈ P by monotonicity. This
implies that d1 (Fn , P) < ε, a contradiction. 

Corollary 22.14. Every minor-closed property of graphs with bounded degree is

testable.

Monotonicity of the property P was used in the proof above only in a somewhat
annoying technical way, and one would like to extend the argument to all hyperfinite
properties P. One must be careful though: the property that “G is a planar graph
with an even number of nodes” is hyperfinite, but not testable (a large grid with
an even number of nodes cannot be distinguished from a large grid with an odd
number of nodes by neighborhood sampling). But the method works with a little
twist: suppose that two graphs F and G have the same number of nodes, and we
know that G ∈ P, and the sampling distance of G and F is small; then it follows
that the edit distance of F from P is small. This implies that P is testable in a
non-uniform sense. For an exact formulation and details, see Newman and Sohler
[2011].
 22.3. COMPUTABLE STRUCTURES 405

Exercise 22.15. Let G and G′ be (ε, k)-hyperfinite graphs with the same number
of nodes n. Prove that they can be overlayed so that
1
|E(G)△E(G′ )| ≤ 2(1 + Dk )ε + Dδ⊙ k
(G, G′ ).
n

22.3. Computable structures

Suppose that we want to compute some structure on a very large graph with
bounded degree: say, a maximum matching, a maximum flow, a spanning tree, a
maximum cut, a 3-coloring. Even if we can compute (approximately) the appropri-
ate number, what does it mean to compute this object? Similarly as in the dense
case, the answer is not obvious.
One possible answer is similar to what we did in the dense case. We offer a
service: if somebody comes with a question about a particular node v (“How is this
node matched in your maximum matching?”), we can answer this question just by
inspecting a bounded neighborhood of v (determine the mate u of v in the matching,
or conclude that v is unmatched). These answers must be consistent, so for example
in the case of matchings, if somebody comes with the request concerning the node
u, we must match it with v. Furthermore, the proportion of unmatched nodes
should be at most ε higher than for the true maximum matching.
In the bounded-degree world, however, there is another, equivalent model, that
is perhaps easier to understand and analyze. Let us place an “agent” on every
node of the graph G. These agents are allowed to communicate with each other,
but only with their neighbors along the edges and for a bounded time. At the end,
they have to decide whether they would be matched with any neighbor at all, and
if so, to which of their neighbors. This model is called distributed computing.
The two models are essentially equivalent.
• Suppose that the agents can compute something; then in the “service” model,
if somebody comes with a question about a node v, we just look up what our agent
responsible for v has computed. All the information the agent has collected from
the neighbors can be gathered by exploring a bounded neighborhood of the node.
• Suppose that we can provide the service correctly. Then we can instruct each
agent to do the computation we would have done if the node they are responsible
for were queried. Of course, the agent has to gather all the information about the
neighborhood we would have explored, but this can be done by communicating
with his/her neighbors. Of course, we also have to instruct them to provide and
communicate the information that is needed for this. All of this takes a bounded
number of bits. We assume here that the agents can generate a name for themselves
that identifies them, at least locally. We don’t go into other details of this model,
like whether the communication between agents is synchronized (sending a bit on
every tick of the clock).
We will in fact use a kind of hybrid description of the algorithms, where the
agents are allowed to explore their neighborhood to a bounded depth (including
the colors and weights of the other agents in this neighborhood, which we have to
use in some cases). Using results of Nguyen and Onak [2008] and Csóka [2012a] we
illustrate the power and some of the subtleties of these models.
Symmetry. There is a fundamental difficulty with distributed algorithms: sym-
metry. Suppose that we want to construct a matching in a very long cycle. All our
agents see the same neighborhood of any given radius, so they will all compute the
406 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

same answer: that they want to remain unmatched (which gives an empty match-
ing, very far from being optimal). Symmetry does not allow them to give any other
answer, at least deterministically.
We can break the symmetry and find a matching close to the optimum, if we
allow the agents to flip coins. We can consider the coinflips generated by any agent
as a real number between 0 and 1, and call this the local random seed of the agent.
This takes an infinite number of coinflips, but only a finite and bounded number of
them has to be generated during the run of the algorithm.
Preprocessing. In our examples for computing a structure in the dense case,
preprocessing (computing a representative set) played a large role. In the bounded
degree case, there is less room for preprocessing. We can think of two kinds of
preprocessing:
• We can do some preliminary computation (perhaps randomized) indepen-
dently of the graph, and inform the agents about the result. If we are lazy, we can
just let the agents do this computation for themselves. The only information they
need for this is the random seed we use during the computation. So it suffices to
generate a random number in [0, 1] and tell it to all the agents. We call this number
the global random seed. (Note: they could generate a random number themselves,
but this would not be the same for all agents!)
• We can do preliminary computation using information about the graph. This
could be based on the distribution of r-balls in G for some fixed r (which is the
realistic possibility for us to obtain information about the graph), but perhaps we
have some other information about the graph (like somebody tells us that it is
connected). Again, we can let the agents work, just have to pass on to them the
information about the graph they need. In the strongest form, we let the agents
know what the graph is (up to isomorphism).
The task. Assume that our agents have to compute a decoration f : V (G) → C,
where C is a finite set. Not all decorations will be feasible, but we assume that
the feasibility criterion is local, i.e., there is an r ∈ N and a set of feasible C-
decorated r-neighborhoods such that a decoration is feasible if and only if every
r-neighborhood is feasible.
The goal is to find an “optimal” decoration. The decoration is evaluated locally
in the following sense: we associate a value ω(B) ∈ [0, 1] to every C-decorated
r-ball F ∈ F , and we want to minimize the average value of r-balls. Setting
ω(v) = ω(BG,r (v), f |BG,r (v) ), the cost of the decoration is defined by
1 ∑
w(f ) = ω(v).
v(G)
v∈V (G)

The agents want to compute a decoration f for which w(f ) is as small as possible.
Example 22.16 (Proper coloring). Suppose that we want to compute a proper
k-coloring of G. Then we choose C = [K] for some very large K, the feasibility
criterion is that the coloring should be proper (clearly this can be verified from the
1-neighborhoods), and we evaluate the coloring by imposing a penalty of 1 on every
node with color larger than k. 
Example 22.17 (Maximum matching). Suppose that we want to compute a
maximum matching in G. Then we can take C = [D + 1]. Decoration with i,
 22.3. COMPUTABLE STRUCTURES 407

i ≤ D, means that the node is matched with its i-th largest neighbor (in the order
of their local seeds); decoration with D + 1 means that the node is unmatched.
The feasibility criterion is clearly local. We impose a penalty of 1 for decoration by
D + 1. 
Example 22.18 (Max-flow-min-cut). Suppose that we are given a 3-coloring
of the nodes of a graph G by red, white and green. We consider all edges to
have capacity 1, and would like to find a maximum flow from the red nodes to
the green nodes. This means a decoration of every node v by a rational vector
f (v) = (f1 , . . . , fD ), where fi is the flow it is sending to its i-th highest weighted
neighbor (this can be negative or positive). The sum of entries of f (v) is the gain
γ(v) of the node. Feasibility means that the flow on any edge uv, indicated in the
decoration of u, is the negative of the flow on this edge indicated in the decoration
of v; furthermore, the gain is 0 at every white node, nonnegative at every red node,
and nonpositive at every green node. The objective function is the sum of γ(v)
over the red nodes.
Computing the minimum cut fits in the framework quite easily too: We decorate
every node by either “LEFT” or “RIGHT”. Feasibility means that all red nodes
are decorated by “LEFT” and all green nodes are decorated by “RIGHT”. The
objective value is half of the average number of neighbors of a node on the other
side. 

The computational model. We compare four settings, getting increasingly more

powerful:
(A) The agents don’t get any preprocessing information, and have to work
deterministically;
(B) The agents don’t get any preprocessing information, but have access to
their own random number generator;
(C) The agents have access to their own random number generator, and in
addition they get the same global seed chosen uniformly from [0, 1];
(D) The agents have access to their own random number generator, to the
public random number g0 as in (B), and in addition they know the graph up to
isomorphism (but they don’t know at which node of the graph they sit).
Note that in models (A)-(C), the agents can see their own r-neighborhood, and
can hear about other r-neighborhoods at a bounded distance from them, but they
will not be able to learn global statistics. The agents themselves will not know, for
example, the average degree of the graph. In model (D), one may be concerned
how an arbitrarily large graph can be communicated to our agents; instead, we
could say that this model allows any kind of information (any number of graph
parameters) to be passed to the agents: the most natural would be neighborhood
statistics, but the model allows non-testable graph parameters and properties like
the chromatic number or connectivity to be used.
We will see that there are nontrivial algorithms in the weakest model (A); that
(B) is strictly stronger than (A), and (C) is strictly stronger than (B); but every
problem solvable in (D) is also solvable in (C) with an arbitrarily small increase in
the cost.
22.3.1. Matchings. We start with describing an algorithm in model (B), de-
signed by Nguyen and Onak, [2008] to find an (almost) maximum matching.
408 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

Algorithm 22.19.
Input: A graph G with maximum degree D and no isolated nodes in the agent
model, and an error bound ε.
Output: A random matching M such that with probability at least 1 − ε, |M | ≥
(1 − ε)ν(G).
As in most matching algorithms, we start with the empty matching, and aug-
ment it using augmenting paths: these are paths that start and end at unmatched
nodes, and every second edge of them belongs to the current matching M . Aug-
menting along such a path interchanging the matching edges and non-matching
edges) increases the size of the matching by 1. Of course, in our setting we will
have to augment simultaneously along many disjoint augmenting paths, to make
measurable progress. We will augment along augmenting paths of length at most
k = ⌈3/ε⌉, which we call short augmenting paths.
This is again done in rounds. It will be convenient to assume that in each
round, a new local seed is generated for every agent v. (They could get this from
a single random real number in [0, 1], by using all the bits in even position in the
first round, half of the remaining bits in the second, etc.) This way the rounds
will be independent of each other in the probabilistic sense. Augmentation along
many disjoint augmenting paths will be carried out simultaneously by our agents.
It is clear that agents looking at their neighborhoods with radius k will discover all
short augmenting paths. The problem is that there will be conflicts: these short
augmenting paths are not disjoint. To this end, we define when a path is better
than another, and we will augment only along those paths that are better than any
path intersecting them.
To be precise, we define that path P is better than path Q if walking along
both paths, starting from their endnodes with higher local seed, the first node that
is different has higher local seed in P than in Q. We will augment along paths
that are better than any path intersecting them; we call such a path locally best. If
we allow agents to explore their neighborhoods with radius 2k, then every locally
best short augmenting path will be discovered by at least one agent, who will carry
out the augmentation (i.e., send a message to the agents along the path how their
mates are to be changed). Several agents may do so for a given path, but there will
be no conflict between their messages.
The above is repeated q = 4D2k ⌈log(1/ε)⌉ times, then we stop and output the
current matching.
The idea in the analysis is that in a particular phase, we either find many
good short paths, and hence make substantial progress, or the number of all short
augmenting paths is small, in which case we have an almost maximum matching.
Let as call a node eligible (at a certain phase) if at least one short augmenting
path starts at it (such a node is of course unmatched). Let Mi be matching after the
i-th round, and let Xi denote the number of eligible nodes. Let M ′ be a maximum
matching, and consider the set Mi ∪ M ′ . This set of edges consists of the common
edges of Mi and M ′ , and cycles and paths whose edges alternate between M and
M ′ . Every cycle contains the same number of edges from Mi and M ′ . Paths that
contain more edges from M ′ than from Mi have to end with edges in M ′ at both
ends, and so they are augmenting paths. Thus the number of augmenting paths
is at least |M ′ | − |Mi | = ν(G) − |Mi |. The number of augmenting paths among
these that have length more than k is less than 2|M ′ |/k, so there are at least
 22.3. COMPUTABLE STRUCTURES 409

( )
(1 − 2/k)ν(G) − |Mi | short augmenting paths, and Xi ≥ 2 − (4/k) ν(G) − 2|Mi |
eligible nodes.
Let u be an eligible node after phase i. All the short augmenting paths inter-
secting any of the short augmenting paths starting at u stay within BG,2k (u). Since
|BG,2k (u)| ≤ D2k , there is a chance of at least p = 1/D2k that u is the node with
highest local seed among them. Then the best path starting at u will be augmented
upon, and hence u has a chance of at least p to become matched in that round.
This means that
(  ) 1 (k − 2 )
E |Mi+1 |  Mi ≥ |Mi | + pXi ≥ |Mi | + p ν(G) − |Mi |
2 k
(here expectation is taken over random choices in the (i + 1)-st round), which we
can write as
(k − 2  ) (k − 2 )
E ν(G) − |Mi+1 |  Mi ≤ (1 − p) ν(G) − |Mi | .
k k
Taking expectation over Mi , we get
(k − 2 ) (k − 2 )
E ν(G) − |Mi+1 | ≤ (1 − p)E ν(G) − |Mi | .
k k
Hence
(k − 2 ) (k − 2 ) k−2
E ν(G) − |Mq | ≤ (1 − p)q E ν(G) − |M0 | = (1 − p)q ν(G).
k k k
By Markov’s Inequality, this implies that
(k − 2 k−2 )
P ν(G) − |Mq | > ν(G) ≤ k(1 − p)q ≤ ke−pq ≤ ε.
k k2
So with probability at least 1 − ε, we have
k−2 k−2 (k − 1)(k − 2)
|Mq | ≥ ν(G) − 2
ν(G) = ν(G) ≥ (1 − ε)ν(G).
k k k2
This proves that the algorithm works as claimed.
We have seen that without local seeds, there is no way to approximately com-
pute a maximum matching. So the matching problem can be solved in model (B)
but not in (A).

22.3.2. Maximum flow: an algorithm in the weakest model. An al-

gorithm based on similar ideas was developed by Csóka [2012a] to find an almost
maximum flow (Example 22.18). This algorithm too looks for short augmenting
paths. We do not describe the details here, but point out one interesting feature.
Using the random local seeds, the agents compute a flow that is almost optimal,
in a way similar to the maximum matching algorithm described above. A different
choice of seeds would give a different flow. But the expected flow values (expectation
taken over all random seeds) also give a valid, almost maximum flow, which is
independent of any random seeds. This expectation can be computed locally, from
the neighborhoods with radius 2r. Thus to compute the maximum flow, we don’t
need the random seeds after all: deterministic agents can compute it. (This does
not contradict our arguments about the curse of symmetry, because matchings are
not invariant under all automorphism of the given graph, but there is a “canonical”
maximum flow that is invariant under automorphisms preserving the sets of sources
and sinks: the average of all maximum flows.)
410 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

Csóka also describes an algorithm in model (C) to find an almost minimum

cut, and proves that for this, a public random number is needed; in other words,
the problem cannot be solved in model (B).
22.3.3. Knowing the graph does not help. Our next goal is to show that
model (D), which seems to be a lot stronger than (C), is in fact equivalent to it (up
to an arbitrarily small increase of the cost; Csóka [2012a]).
Theorem 22.20. Suppose that there is an algorithm in model (D) by which the
agents compute, for every graph G, a feasible decoration f with cost c(G). Then
for every ε > 0 there is an algorithm in model (C) by which the agents compute,
for every graph G, a feasible decoration f with cost at most c(G) + ε.
Proof. Recall that Ar denotes the set of all probability distributions ρG,r ,
where G ranges through finite graphs. By Proposition 19.9, its closure Ar is convex.
Let us fix a large graph G. We may assume that c(G) is the optimal cost of a
decoration the agents can compute in model (D). If the agents mistakenly believe
that they are working on the graph F , then they will compute (in model (D)) a
decoration fF of G (which is random, as a function of the public random seed
and the private random seeds). The expectation of ω(fF ) will depend on the true
distribution ρG,r . Setting ωF (v) = ω(BG,r (v), fF |BG,r (v) ), we get by the linearity
of expectation,
( ) 1 ∑ ( )
E w(fF ) = E ωF (v) .
v(G)
v∈V (G)
The last expectation depends only on F and on BG,2r (v) (since the distribution of
fF (u) depends only on the r-neighborhood of u, and to compute the distribution
of ωF (v) it suffices to know the joint distribution of( the decorations
) fF (u) in the
r-neighborhood of v). For B ∈ B2r , let a(F, B) = E ωF (root(B)) , then
( ) ∑
E w(fF ) = ρG,2r (B)a(F, B) = LF (ρG,2r ),
B∈B2r

where LF : A2r → R is a homogeneous linear function. Clearly LF (ρG,2r ) ≥ c(G)

for any F , and our assumption about the quality of the output of (D) implies that
LG (ρG,2r ) = c(G).
Next, we define a function u : A2r → R by
u(ρ) = lim inf c(Gn ),
where the limes inferior is to be taken over all sequences (Gn ) for which ρGn ,2r → ρ.
A similar argument as in the proof of Proposition 19.9 shows that the function u
is convex. For any graph G, considering the special sequence (Gn : n = 1, 2, . . . ),
we get
(22.7) u(ρG,2r ) ≤ c(G).
We claim that for every ρ ∈ A2r there is a graph F such that
(22.8) LF (ρ) < u(ρ) + ε.
Indeed, let (Gn ) be a sequence of graphs such that ρn = ρGn ,2r → ρ and c(Gn ) →
u(ρ). Then
LGn (ρ) = LGn (ρn ) + o(1) = c(Gn ) + o(1) = u(ρ) + o(1).
Thus F = Gn can be chosen in (22.8) for a sufficiently large n.
 22.3. COMPUTABLE STRUCTURES 411

} geometry. Consider the set K ⊆ R × R,

Br
Next, we engage
{ in a little convex
defined by K = (ρ, y) : y ≥ u(ρ) . It is{clear that this set is convex.
} For every
graph F , we consider the halfspace HF = (ρ, y) : y ≤ LF (ρ) − ε . We claim that
∩
K∩ HF = ∅.
F

Indeed, suppose that (ρ, y) is a point contained in the left side, then (ρ, y) ∈ K,
and (22.8) implies that there is an F ∈ G such that y ≥ u(ρ) > LF (ρ) − ε. On the
other hand, (ρ, y) ∈ HF implies that y ≤ LF (ρ) − ε, a contradiction.
Hence by Helly’s Theorem, there is a finite set of ∩graphs F1 , . . . , Fm ∈ A2r ,
where m ≤ |Br | + 1, such that K ∩ H = ∅, where H = i≤m HFi . Since K and H
are convex, there is a halfspace defined by a linear inequality y−L(ρ) ≤ b containing
H but disjoint from K. This means two things:
(a) The inequality y ≥ u(ρ) (ρ ∈ A2r ) implies that y − L(ρ) > b. This last
condition means that u(ρ) > L(ρ) + b for every ρ ∈ A2r .
(b) The linear inequalities y − LFi (ρ) ≤ ε imply the inequality y − ∑L(ρ) ≤ b.
By
∑ the Farkas Lemma, there are nonnegative numbers α i such that i αi = 1,
i αi LFi = L, and b ≥ −ε. The numbers α i form a probability distribution α on
[m].
Now we can give the following instruction to our agents: Use the even bits of
the public random number g0 to pick an i ∈ [m] from the distribution α. (All
agents will pick the same i.) Then pretend that you are working on the graph Fi ,
and compute the decoration fFi according to algorithm (C) (using the remaining
bits of g0 as the public random number). Then (using (22.7) in the last step) the
agents achieve a cost that is almost as good as the cost they could achieve knowing
the graph:
( ) ∑ ( ) ∑
E w(fFi ) = αj E w(fFj ) = αj LFj (ρG,2r ) = L(ρG,2r )
j j
≤ u(ρG,2r ) − b ≤ u(ρG,2r ) + ε ≤ c(G) + ε. 
22.3.4. Computable structures and Borel sets. We conclude this chap-
ter with sketching a connection between algorithmic problems and measure theory.
Elek and Lippner [2010] give another algorithm for computing an almost maximum
matching in a large bounded degree graph. Their approach is based on the connec-
tions with Borel graphs, which were discussed in Section 18.1. Instead of describing
a second matching algorithm, we only illustrate the idea on a simpler example, by
showing how the proof that every Borel graph with degrees at most D has a Borel
coloring with D + 1 colors (Theorem 18.3) can be turned into an algorithm.
In the proof of Theorem 18.3, we start with constructing a countable Borel
coloring. This part of the argument can be translated easily. We have to select
an explicit countable basis for the Borel sets in [0, 1); for example, we can choose
intervals of the form [a/b, (a + 1)/b), where 0 ≤ a < b are integers. We have to
assign a positive integer index to each of these intervals, say (2a + 1)2b . Then every
agent picks the interval with smallest index that contains his local seed but not the
local seed of any of his neighbors. Now the agents have indices (they can forget the
seeds from now on). Trivially, adjacent agents have different indices.
Next, every agent whose index is smaller than the indices of his neighbors
changes his index to 1, and labels himself FINISHED. (In the proof of Theorem
412 22. ALGORITHMS FOR BOUNDED DEGREE GRAPHS

18.3, only those with index 2 not adjacent to any node with index 1 did so in the first
round; but it is easy to see that eventually all nodes with a locally minimal index
will change to 1). Next, all those agents whose index is smaller than the indices of
all their unfinished neighbors change their indices to the smallest possible, etc.
At this point comes an important difference: for Borel coloring, we could repeat
this infinitely many times, but here we have a time bound. Those nodes that
managed to change their indices are now properly colored with D+1 colors; however,
there will be some who are stuck with their large original indices. Down-to-earth
work starts here to show that their number is a small fraction of v(G). We don’t
go into the details (see Exercises 22.22, 22.23).
Exercise 22.21. Describe how Algorithm 22.19 can be simplified in two simpler
versions of the problem: (a) we only want to find a maximal (non-extendable)
matching (of course, with an error); (b) somebody marks a matching for us, and
we have to test whether it is maximum (again, with some error).
Exercise 22.22. Prove that any constant time distributed algorithm that con-
structs a legitimate coloring of a cycle will use, with high probability, more than
100 colors.
Exercise 22.23. Prove that for every ε > 0 there is a k ≥ 1 such that the
algorithm (with k rounds) as described above will produce a coloring in which
fewer than εv(G) nodes have color larger than D.
 Part 5

Extensions: a brief survey

 CHAPTER 23

Other combinatorial structures

The ideas of characterizing homomorphism functions, connection ranks, regu-

larity lemmas and limit objects have been extended to several combinatorial struc-
tures besides graphs. Some of these extensions are rather involved and deep, like
the limit theory of hypergraphs; others can be described as “analogous” (at least
after finding the right definitions). Without attempting to be complete, we survey
several of these extensions.

23.1. Sparse (but not very sparse) graphs

The obvious big gap in our treatment of limits of growing graph sequences is
any sequence of graphs with density tending to 0, but maximum degree tending
to infinity. Some interesting examples are the point-line incidence graphs of finite
projective planes (about n3/2 edges, if n is the number of nodes), and d-cubes
(n log n edges).
Some work has been done. We have mentioned extensions of the Regularity
Lemma to sparser graphs by Kohayakawa [1997], Gerke and Steger [2005], and Scott
[2011]). While the case of bounded degree graphs is open, these results are highly
nontrivial for sparse (but not very sparse) graphs, and have important applications.
They are very likely to play an important role in the limit theory of such graphs. In
a substantial paper, Bollobás and Riordan [2009] investigate many of the techniques
discussed in this book and elsewhere, mostly from the point of view of extending
them from the case of dense graphs to sparser classes.
Lyons [2005] extended the convergence theory of bounded degree graphs to
graph sequences with bounded average degree, under a condition called tightness
(this guarantees that the sequence of sample distributions has a limit distribution).
The following example shows that some condition like this is necessary: the average
degree of a subdivision G′ of any graph G (dense or not) is bounded by 4. Clearly,
to properly describe the limit of the graph sequence (G′n ), the description must
contain essentially the same information as the limit of the sequence (Gn ). So
limits of graphs with bounded average degree are as complex as limits of any graph
sequence (dense or sparse).
Graphons and graphings generalize dense graphs and bounded degree graphs,
respectively, and they can be considered as the two extremes as far as edge density
goes. One common feature is that we can do a random walk on each of them. More
precisely, there is a Markov chain on a graphon, as well as on a graphing, and we
are going to show that this Markov chain contains all the necessary information
about these objects.
415
416 23. OTHER COMBINATORIAL STRUCTURES

Let W : Ω2 → [0, 1] be a graphon with density ω = t(K2 , W ). We can define

a Markov chain on Ω by
∫
1
Pu (A) = W (u, v) dv
dW (u)
A
(this is defined for almost all u). This Markov chain has a stationary distribution,
defined by ∫
1
π(A) = W (x, y) dx dy.
ω
A×[0,1]
It is also easy to check that this Markov chain is reversible. The step distribution
of this Markov chain is proportional to the integral measure of W . Note that the
Markov chain does not change if we scale W , so we have to remember the “density”
ω if we want to preserve all information about W . But the Markov chain together
with the density does determine the graphon.
Next, consider a graphing G on Ω. We can define a Markov chain by
degA (u)
Pu (A) = .
deg(u)
Then the random walk defined by this chain is just the random walk on this graph
in the usual sense. The measure preservation condition (18.2) says that this Markov
chain is reversible. A stationary measure of this random walk is λ∗ (as defined in
Section 18.2), its step distribution is η/ω. So the step distribution of the Markov
chain is the same as the probability measure on the edges of a graphing. The
graphing is determined by this Markov chain. It is a fascinating open problem
whether Markov chains can be used to define convergence and limit objects for
graph sequences that are neither dense nor of bounded degree.

23.2. Edge-coloring models

23.2.1. Edge-connection matrices. We consider multigraphs with loops.
It will be useful to allow a single edge with no endpoints; we call this graph the
circle, and denote it by ⃝.
We can define edge-connection matrices that are analogous to the connection
matrices defined before: Instead of gluing graphs together along nodes, we glue
them together along edges. To be precise, we define a k-broken graph as a k-labeled
graph in which the labeled nodes have degree one. (It is best to think of the labeled
nodes as not nodes of the graph at all, rather, as points where the k edges sticking
out of the rest of the graph are broken off.) We allow that both ends of an edge be
broken off.
For two k-broken graphs G1 and G2 , we define G1 ∗ G2 by gluing together the
corresponding broken ends of G1 and G2 . These ends are not nodes of the resulting
graph any more, so G1 ∗ G2 is different from the graph G1 G2 we would obtain by
gluing together G1 and G2 as k-labeled graphs. We can glue together two copies of
an edge with both ends broken off; the result is the circle ⃝. One very important
difference is that while G1 G2 is k-labeled, G1 ∗ G2 has no broken edges any more,
and so it is not k-broken but 0-broken. This fact leads to considerable difficulties
in the treatment of edge models.
For every graph parameter f and integer k ≥ 0, we define the edge-connection
matrix M ′ (f, k) as follows. The rows and columns are indexed by isomorphism
 23.2. EDGE-COLORING MODELS 417

types of k-broken graphs. The entry in the intersection of the row corresponding
to G1 and the column corresponding to G2 is f (G1 ∗ G2 ). Note that for k = 0, we
have M (f, 0) = M ′ (f, 0), but for other values of k, connection and edge-connection
matrices are different.
Let G be a finite graph. An edge-coloring model is determined by a mapping
h : Nq → R, where q is positive integer. We call h the node evaluation function.
Here we think of [q] as the set of possible edge colors; for any coloring of the edges
and d ∈ Nq , we think of h(d) as the “value” of a node incident with dc edges
with the color c (c ∈ [q]). In statistical physics this is called a vertex model: the
edges can be in one of several states, which are represented by the color; an edge-
coloring represents a state of the system, and (assuming that h > 0) ln h(d) is the
contribution of a node (incident with dc edges with color c) to the energy of the
state.
There are many interesting and important questions to be investigated in con-
nection with edge-coloring models; we will only consider what in statistical physics
terms would be called its “partition function”. To be more precise, for an edge-
coloring φ : E(G) → [q] and node v, let degc (φ, v) denote the number of edges e
incident with node v with color φ(e) = c. So the vector deg(φ, v) ∈ Nq is the “local
view” of node v. The edge-coloring function of the model is defined by
∑ ∏ ( )
col(G, h) = h deg(φ, v) .
φ: E(G)→[q] v∈V (G)

Recall that we allow the graph ⃝ consisting of a single edge with no endpoints; by
definition, col(⃝, h) = q. We also allow that q = 0, in which case col(G, h) = 1 if G
has no edges, and col(G, h) = 0 otherwise. We could of course allow complex valued
node evaluation functions, in which case the value of the edge-coloring function can
be complex.
Example 23.1 (Number of perfect matchings). The number of perfect match-
ings can be defined by coloring the edges by two colors, say black and white, and
requiring that the number of black edges incident with a given node be exactly
one. This means that this number is col(., h), where h : N2 → R is defined by
h(d1 , d2 ) = 1(d1 = 1). The number of all matchings could be expressed similarly.

Example 23.2 (Number of 3-edge-colorings). This number is col(., h), where
h : N3 → R is defined by h(d1 , d2 , d3 ) = 1(d1 , d2 , d3 ≤ 1). 
Example 23.3 (Spectral decomposition of a graphon). Recall the definition
(7.18) and expression (7.25) for t(F, W ) in terms of the spectrum of TW . We
can consider χ as a coloring of E(F ) with colors 1, 2, . . . . Then Mχ (v) depends
only
( on the) numbers of edges with different colors, and so we can write Mχ (v) =
h deg(χ, v) , and we get
∑ ∏ ∏ ( )
t(F, W ) = col(G, λ, h) = λχ(e) h deg(χ, v) .
χ: E(G)→[q] e∈E(F ) v∈V (G)

However, this is not a proper edge-coloring model, since the value of the circle,
which is the number of colors, is infinite in general. 
The following facts about the edge-connection matrices of edge-coloring func-
tions are easy to prove along the same lines as Proposition 5.64:
418 23. OTHER COMBINATORIAL STRUCTURES

Proposition 23.4. For every edge-coloring model h : Nq → R, the graph param-

eter col(., h) is multiplicative, its edge-connection matrices M ′ (f, k) are positive
semidefinite, and rk(M ′ (f, k)) ≤ q k . 
B. Szegedy [2007] showed that the first two of these properties suffice to give a
characterization of edge-coloring functions.
Theorem 23.5. A graph parameter f can be represented as an edge-coloring func-
tion if and only if it is multiplicative and M ′ (f, k) is positive semidefinite for all
k ≥ 0. 
The proof of this theorem is quite involved and not reproduced here; it is based
on ideas similar to those used in Section 6.6 to prove Theorem 5.57, but using quite
a bit more involved tools: The use of the Nullstellensatz and simple semidefiniteness
arguments must be replaced by real versions of the Nullstellensatz (Positivstellen-
satz), and the simple symmetry arguments must be replaced by deeper results
from the representation theory of algebras. (As a historical comment, the proof of
Theorem 23.5 came first, and Schrijver’s proof of Theorem 5.57 was motivated by
this.)
Draisma, Gijswijt, Lovász, Regts and Schrijver [2012] give characterizations
of complex valued edge-coloring functions. Let us state without proof a result of
Schrijver [2012], which shows that a condition on the growth of the edge-connection
rank (along with minor other constraints) can characterize complex edge-coloring
models.
Theorem 23.6. A complex valued graph parameter f is an edge-coloring func-
tion
( of a complex
) model if and only if it is multiplicative, f (⃝) is real, and
rk M ′ (f, k) ≤ f (⃝)k for every k. 
23.2.2. Tensor algebras. There is a surprisingly close connection between
edge-coloring models and rather general multilinear algebra. Given an edge-coloring
model (with color set [q]), we can think of the nodes as little gadgets with wires (or
legs) sticking out corresponding to the edges incident with it. If we assign colors to
the wires, the gadget outputs a number (this could be real or complex). The graph
parameter defined by this edge-coloring model is the expectation of the product of
these numbers, one for each node, where the edge-coloring is random.
Formulating the question like this, we see two restrictions that look artificial:
— There is only one gadget for each degree. Why not have several?
— We have assumed that the output of a gadget depends only on the number
of legs with each color; in other words, it is invariant under the permutation of the
legs. Why not drop this condition?
If we relax these conditions, then every gadget would be described by a real
array (Hi1 ,...,id : ir ∈ [q]), where d is the number of legs. In other words, the gadget
is described by a tensor with d slots over Rq . Furthermore, we have to indicate for
each node v which gadget is sitting there, and how its legs correspond to the edges
incident with v.
In more mathematical terms, we have a graph where a tensor is associated with
every node, and an index associated with every edge, so that the slots (indices) of
the tensor correspond to the edges incident with the node. Let us call such a graph
a tensor network. The corresponding graph parameter is evaluated by taking the
 23.2. EDGE-COLORING MODELS 419

product of these tensors, and then summing over all choices of the indices. Note
that every index occurs twice, so we could call this “tracing out” every index.
These tensor networks play an important role in several areas of physics, but
we can’t go into this topic in this book.
This setup allows for a more general construction. If we have a tensor network
with k broken edges, then the value associated with the graph will depend on the
color of these edges, in other words, it will be described by an array (Ai1 ,...,ik : ir ∈
[q]). So the graph with k broken edges can be considered as a gadget itself.
We can break down the procedure of assembling a tensor network from the
gadgets (with or without broken edges) into two very simple steps:
(a) We can take the disjoint union of two gadgets; if the gadgets have k and
l legs, respectively, the union has k + l legs. In terms of multilinear algebra, this
means to form the tensor product of two tensors.
(b) We can fuse two legs of a gadget. If (Ai1 ,...,ik : ir ∈ [q]) is the tensor
describing the gadget, and (say) we fuse legs k − 1 and k, then we get the tensor
∑
Bi1 ,...,ik−2 = Ai1 ,...,ik−2 ,j,j .
j∈[q]

In multilinear algebra slang, we trace out the last two indices.

It is easy to see that with these operations, we can construct every tensor
network with or without broken edges, and we get the corresponding tensor.
Supposing that we have a starting kit of gadgets, we can look at the set of all
tensors that can be realized by assembling tensor graphs with broken edges from
these gadgets. In the spirit of linear algebra, we take all linear combinations of the
obtained tensors with the same number of slots. Every tensor obtained this way
will be called an assembled tensor.
It is clear from (a) and (b) above that the set of assembled tensors has the
following structure: For every k, there is a linear space Tk of tensors over Rq with
k slots. For every A ∈ Tk and B ∈ Tl , the tensor product A ⊗ B ∈ Tk+l . For every
A ∈ Tk , and any two indices in A, tracing out these two indices results in a tensor
in Tk−2 . We call such a set of tensors a traced tensor algebra.
Conversely, every traced tensor algebra arises as the set of assembled tensors:
for every number k of slots, we select a basis of the space Tk , and use the resulting
set of tensors as the starting kit.
It is quite fruitful to use this connection; one can obtain results that are new
both for graphs and for tensor algebras. We describe one important result with
combinatorial connections.
Given a starting kit K, how can we decide about a tensor whether it can be
assembled from this kit? In other words, is it contained in the traced tensor algebra
generated by K? A beautiful answer to this question was found by Schrijver [2008a],
which we describe in graph-theoretic terms (the proof uses the representation theory
of algebras, and we do not give it here; cf. also Schrijver [2008b, 2009]).
Recall that we work over a fixed vector space Rq . Every q × q real matrix A is a
gadget in itself, with two legs. If it is symmetric, then the legs are interchangeable,
but in general we have to talk about a “left leg” (corresponding to the row index)
and a “right leg” (corresponding to the column index). Connecting the gadgets for
matrices A and B in series gives a gadget representing the matrix AB.
420 23. OTHER COMBINATORIAL STRUCTURES

Orthogonal matrices will play a special role. One observation is that if A is

orthogonal, then AAT = I (the identity matrix), and so if we have a gadget graph
with no broken edges, and replace any edge by a path of length 3 with A and AT
sitting on it:
←→

If we replace every edge by this path of length 3, then the value of the graph does
not change. However, we can group together every original gadget B with the
orthogonal matrices next to, to get a gadget B A , which—in multilinear algebra
terms—is obtained from B by applying the linear transformation A to every slot.
If we replace every gadget B in the kit by B A , then the value of the tensor network
does not change.

Figure 23.1. Replacing every edge by a path with the same or-
thogonal transformation at both inner nodes (just facing the op-
posite direction), and regrouping does not change the value.

Now consider a tensor network with broken edges. If we replace every tensor B
in the kit by B A , then the matrices A and AT along the unbroken edges still cancel
each other, but on the broken edges, one copy still remains. In other words, if we
apply the same orthogonal transformation to every slot of every tensor in the kit,
then the tensor defined by a tensor network with broken edges undergoes the same
transformation.

Figure 23.2. Applying the same orthogonal transformation to all

slots of all tensors in the kit results in applying the same orthogonal
transformation to the slots of the assembled tensor.

In particular, if all tensors in the kit have the property that a particular or-
thogonal transformation applied to all their slots leaves them invariant, then the
 23.3. HYPERGRAPHS 421

same holds for every assembled tensor. The theorem of Schrijver [2008a] asserts
that this is the only obstruction to assembling a given tensor.
Theorem 23.7. Let T be a traced tensor algebra generated by a set S of tensors,
including the identity tensor 1(i = j) (i, j ∈ [q]). Then a tensor T is in T if
and only if it is invariant under every orthogonal transformation that leaves every
tensor in S invariant. 
The special case when the generating tensors are symmetric describes edge-
coloring models. This can be viewed as an analogue of Theorem 6.38, with the role
of the edges and nodes interchanged. Regts [2012] showed how Theorem 23.7 yields
an exact formula for the edge-connection rank of edge-coloring models.
Example 23.8 (Number of perfect matchings revisited). The tensor model
for this graph parameter is a bit more complicated than in Example 23.1. We have
2 edge colors (which will be convenient to call 0 and 1), so we work over R2 ; but
we need to specify a tensor for every degree d, expressing that exactly one edge is
black:
Ti1 ,...,id = 1(i1 + · · · + id = 1).
It is easy to see that no orthogonal transformation, applied to all slots, leaves
this tensor invariant, so it follows from Theorem 23.7 that every tensor can be
assembled from this kit. (We note that the tensor is invariant under permuting
the slots; however, this symmetry is not preserved under composition of tensor
networks.) 
Example 23.9 (Number of 3-edge-colorings revisited). To construct a tensor
model for the number of 3-edge-colorings, we work over R3 . We again need to specify
a tensor for every degree expressing that the edges have different colors:
Ti1 ,...,id = 1(i1 , . . . , id are different)
(for d > 3, we get the 0 tensor). Permuting the colors (i.e., the coordinates in
the underlying vector space R3 ) leaves this tensor invariant, and these are the only
orthogonal transformations of R3 with this property. Theorem 23.7 implies that a
tensor is invariant under the permutations of the coordinates of R3 if and only if it
can be assembled from this kit. 

23.3. Hypergraphs
When talking about generalizing results on graphs, the first class of structures
that comes to mind is hypergraphs (at least to a combinatorialist). So it is per-
haps surprising that to extend the main concepts and methods developed in this
book (quasirandomness, limit objects, Regularity Lemma, and Counting Lemma)
to hypergraphs is highly nontrivial. Even the “right” formulation of the Regularity
Lemma took a long time to find, and in the end both the Regularity Lemma and
the limit object turned out quite different from what one would expect as a naive
generalization. Nevertheless, the issue is essentially solved now, thanks to the work
of Chung, Elek, Graham, Gowers, Rödl, Schacht, Skokan, Szegedy, Tao and others.
A full account of this work would go way beyond the possibilities of this book, but
we will give a glimpse of the results.
By an r-uniform hypergraph, or briefly r-graph, ( we
) mean a pair H = (V, E),
where V = V (H) is a finite set and E = E(H) ⊆ Vr is a collection of r-element
422 23. OTHER COMBINATORIAL STRUCTURES

subsets. The elements of V are called nodes, the elements of E are called edges.
So 2-graphs are equivalent to simple graphs. We can define the homomorphism
number hom(G, H) of an r-graph G into an r-graph H in the natural way, as the
number of maps φ : V (G) → V (H) for which φ(A) ∈ E(H) for every A ∈ E(G).
The homomorphism density of G in H is defined as one expects, by the formula
hom(G, H)
t(G, H) =
|V (G)||V (H)|
Quasirandomness can be defined by generalizing the condition on the density
of quadrilaterals. We need to define a couple of special hypergraph classes. Let Knr
( )
denote the complete r-uniform hypergraph on [n] (i.e., E(Knr ) = [n] r
r ). Let Lk be
the “complete r-partite hypergraph” defined on the node set V1 ∪· · ·∪Vr , where the
Vi are disjoint k-sets, and the edges are all r-sets containing exactly one element
from each Vi . Clearly t(Krr , H) = t(Lr1 , H) is the edge density of H. It is not
r
hard to prove that t(Lrk , H) ≥ t(Krr , H)k for every H (this generalizes inequality
2.9 from the Introduction). We define the quasirandomness of H as the difference
r
qr(H) = t(Lr2 , H) − t(Krr , H)2 .
A sequence (Hn ) of hypergraphs is called quasirandom with density p if
r
t(Krr , Hn ) → p and qr(Hn ) → 0, or equivalently, t(Lr2 , Hn ) → p2 . It was proved
by Chung and Graham [1989] that this implies that t(G, Hn ) → pe(G) for every
r-graph G, so the equivalence of conditions (QR2) and (QR3) for quasirandomness
in the Introduction (Section 1.4.2) generalizes nicely.
As a first warning that not everything extends in a straightforward way, let
us try to generalize (QR5). A first guess would be to consider disjoint sets
X1 , . . . , Xr ⊆ V , and then stipulate that the number of edges with one endpoint in
each of them is p|X1 | . . . |Xr | + o(nr ). (For simplicity of presentation, we assume
that v(Hn ) = n.) This property is indeed valid for every quasirandom sequence,
but it is strictly weaker than quasirandomness. It is not well-defined what the
“right” generalization is; we state one below, which is a version of a generalization
found by Gowers. Several other equivalent conditions are given by Kohayakawa,
Rödl and Skokan [2002].
Proposition 23.10. A sequence (Hn ) of hypergraphs is quasirandom with density
p if and only if for every (r−1)-graph Gn on V (Hn ), the number of( edges
) of Hn that
induce a complete subhypergraph in Gn is t(Krr−1 , Gn )t(Krr , Hn ) nr + o(nr ). 
In the case of simple graphs (r = 2), let Hn be a simple graph with edge density
p. The 1-graph Gn means simply a subset of V (Hn ), and K21 is just a 2-element
set. So the condition says that the number of edges of the graph Hn induced by
the set Gn is asymptotically
( ) ( ( ) ( )
n |Gn | )2 2e(Hn ) n |Gn |
t(K21 , Gn )t(K22 , Hn ) = ∼ p ,
2 n n2 2 2
and so we get condition (Q4). For general r, the condition can be rephrased as
follows: for a random r-set X ⊆ V , the events that X is complete in Gn and X is
an edge in Hn are asymptotically independent.
The last remark takes us to another complication.
Example 23.11. Let G(n, 1/2) be a random graph and let Tn denote the 3-graph
formed by the triangles in G(n, 1/2). Then Tn is a 3-graph with density 1/8, which
is random in some sense, but it is very different from the random 3-graph Hn on
 23.3. HYPERGRAPHS 423

[n] obtained by selecting every edge independently with probability 1/8. In fact,
the sequence (Hn ) is quasirandom with probability 1 (this is not hard to see), while
Tn has a very small intersection with every quasirandom 3-graph by Proposition
23.10. Also, Tn has some special features, like no 4-set of nodes contains exactly 3
edges of Hn .
On the other hand, Tn is totally homogeneous. It has no special global struc-
ture; more concretely: on any two disjoint k-sets we see independent copies of the
same random hypergraph. If we want to generalize the Regularity Lemma, it has to
reflect the difference between Tn and Hn , and similarly for the generalization of the
notion of graphons. Which of these sequences should tend to a constant function?

We show how to overcome this difficulty, starting with the construction of the
limit object. We say that a sequence of r-graphs (Hn ) is convergent, if v(Hn ) → ∞
and t(F, Hn ) has a limit as n → ∞ for every r-graph F . Let t(F ) denote this
limit. How to represent this limit function, in other words, what is the hypergraph
analogue of a graphon? The natural guess would be a symmetric r-variable function
W : [0, 1]r , which would represent the limit by
∫ ∏
t(F, W ) = W (xi1 , . . . , xir ) dx.
[0,1]r {i ,...,i }∈E(F )
1 r

The example of the hypergraphs Hn and Tn above show that this cannot be right.
The only reasonable candidate for their limit object would be the function W ≡ 1/8,
which represents correctly the limiting densities for the sequence Hn , but not for
the sequence Tn . We could make life even more complicated, and consider the
intersection Hn ∩ Tn , which is a random 3-graph with expected density 1/64, and
the limiting densities are even more complicated.
For r > 3, one could construct a whole zoo of homogeneous random hyper-
graphs, generalizing the construction of Hn and Tn . After several steps of general-
ization, one arrives at the following: we generate a random coloring of Knj for every
0 ≤ j ≤ r (with any number of colors). To decide whether an r-subset X ⊆ [n]
should be an edge, we look at the colors of its subsets, and see if this coloring be-
longs to some prescribed family of colorings of 2X . (We assume that the prescribed
family is invariant under permutations of X.)
While this example warns us of complications, it also suggests a way out: we
describe the limit not in the r-dimensional but in the 2r -dimensional space. In fact,
the limit object turns out to be a subset, rather than a function, which is a gain
(it is of course very little relative to the increase in the number of coordinates).
[r]
Consider the set [0, 1]2 (so we have a coordinate xI for every I ⊆ [r]; the
coordinate for ∅ will play no role, we can think of it as 0). Let us note that the
[r]
symmetric group Sr acts on the power set 2[r] , and hence also on [0, 1]2 . Let
[r]
U ⊆ [0, 1]2 be a measurable set that is invariant under the action of Sr . We call
such a set a hypergraphon.
For every hypergraphon U , we define the density of an r-graph F as follows.
We assign independent random variables XS , uniform in [0, 1], to every subset
S ⊆ V (F ) with |S| ≤ r. For every edge A = {a1 , . . . , ar } ∈ E(F ), and every
I ⊆ [r], we denote by AI the subset {ai : i ∈ I}, and we consider the point
[r]
X(A) ∈ [0, 1]2 defined by (X(A))I = XAI (this depends on the ordering of A, but
424 23. OTHER COMBINATORIAL STRUCTURES

this will not matter thanks to our symmetry assumption about U ). Now we define
( )
t(F, U ) = P X(A) ∈ U for all A ∈ E(F ) .
To illuminate the meaning of this formula a little, consider the case r = 2.
Then we have U ⊆ [0, 1]3 , where the three coordinates correspond to the sets {1},
{2} and {1, 2} (as we remarked above, the empty set plays no role). For a graphon
W , we define the set UW = {(x1 , x2 , x12 ) ∈ [0, 1]3 : x12 ≤ W (x1 , x2 )}. Then it is
easy to see that t(F, UW ) = t(F, W ) for any simple graph F .
Elek and Szegedy [2012] prove the following.
Theorem 23.12. For every convergent sequence (Hn ) of r-graphs there is a hy-
pergraphon U such that t(F, Hn ) → t(F, U ) for every r-graph F . 
The limit graphon is essentially unique up to some “structure preserving trans-
formations”, which are more difficult to define than in the case of graphs and we
don’t go into the details. Elek and Szegedy [2012] give several applications of
Theorem 23.12. For a given hypergraphon U , they define U -random hypergraphs
and prove that they converge to U . They derive from it the Hypergraph Removal
Lemma due to Frankl and Rödl [2002], Gowers [2006], Ishigami [2006], Nagle, Rödl
and Schacht [2006] and Tao [2006a]. As a refreshing exception, the statement of
this lemma is a straightforward generalization of the Removal Lemma for graphs
(Lemma 11.64); the proof of Elek and Szegedy is similar to our second proof in
Section 11.8. They also derive the Hypergraph Regularity Lemma using Theorem
23.12, using a stepfunction approximation of hypergraphons.
This brings us to the Hypergraph Regularity Lemma, a very important but
also quite complicated statement. There are several essentially equivalent, but not
trivially equivalent forms, due to Frankl and Rödl [1992], Gowers [2006, 2007],
Rödl and Skokan [2004], Rödl and Schacht [2007a, 2007b]. Proving the appropriate
Counting Lemma for these versions is a further difficult issue, and I will not go into
it. But I must not leave this topic without stating at least one form, based on the
formulation of Elek and Szegedy [2012], which in fact generalizes the strong form
of the Regularity Lemma (Lemma 9.5).
We have to define what we mean by “regularizing” a hypergraph. For ε, δ > 0
and k ∈ N, we define an (α, β, k)-regularization of a r-graph H on [n] as follows. For
every i ∈ [r], we partition the complete hypergraph Kni into r-graphs Gi,1 , . . . , Gi,k .
Let us think of the edges in Gi,j as colored with color j. This defines a partition
P of the edges of Knr , where two r-sets are in the same class if the colorings of
their subsets are isomorphic. The family {Gi,j : i ∈ [r], j ∈ [k]}, together with an
r-graph G on [n] will be called an (α, β, k)-regularization of H, if
(a) every r-graph Gi,j has quasirandomness at most α, and
(b) G is the union of some of the classes of P, and
( )
(c) |E(H)△E(G)| ≤ β nr .
Now we can state one version of the Hypergraph Regularity Lemma.
Lemma 23.13 (Strong Hypergraph Regularity Lemma). For every r ≥ 2
and every sequence ϵ = (ε0 , ε1 , ...) of positive numbers there is a positive integer
kϵ such that for every r-graph H there is an integer k ≤ kϵ such that H has an
(εk , ε0 , k)-regularization.
The main point is that to regularize H, we have to partition not only its node
set, but also the set of i-tuples for all i ≤ r. Just like in the graph case, we could
 23.4. CATEGORIES 425

demand that the i-graphs Gi,j have almost the same number of edges for every
fixed i. Of course, the prize we have to pay for stating a relatively compact version
is that it takes more work to apply it; but we don’t go in that direction.
The extension of the theory exposed in this book to hypergraphs is not com-
plete, and there is space for a lot of additional work. Just to mention a few loose
ends, it seems that no good extension of the distance δ has been found to hy-
pergraphs (just as in the case of limit objects or the regularity lemma, the first
natural guesses are not really useful). Another open question is to extend these re-
sults to nonuniform hypergraphs, with unbounded edge-size. The semidefiniteness
conditions for homomorphism functions can be extended to hypergraphs (see e.g.
Lovász and Schrijver [2008]), but perhaps this is just the first, “naive” extension.
One area of applications of these conditions is extremal graph theory. The work of
Razborov [2010] shows that generalizations of graph algebras and of the semidefi-
niteness conditions can be useful in extremal hypergraph theory. However, we have
seen that graph algebras can be defined in the setting of gluing along nodes and
also along edges, and this indicates that for hypergraphs a more general concept of
graph algebras may be useful.

23.4. Categories
The categorial way of looking at mathematical structures is quite prevalent
in many branches of mathematics. In graph theory, the use of categories (as a
language and also as guide for asking question in a certain way) has been practiced
mainly by the Prague school, and has lead to many valuable results; see e.g. the
book by Hell and Nešetřil [2004].
One can go a step further and consider categories (with appropriate finiteness
assumptions) as objects of combinatorial study on their own right. After all, cat-
egories are rather natural generalizations of posets, and there is a huge literature
on the combinatorics of posets. However, surprisingly little has happened in the
direction of a combinatorial theory of categories; some early work of Isbell [1991],
Lovász [1972] and Pultr [1973], and the more recent work of Kimoto [2003a, 2003b]
can be cited.
Working with graph homomorphisms, we have found not only that the cate-
gorial language suggests very good questions and a very fruitful way of looking at
our problems, but also that several of the basic results about graph homomorphism
and regularity can be extended to categories in a very natural way. The goal of this
section is to describe these generalizations, and thereby encourage a combinatorial
study of categories. (Appendix A.8 summarizes some background.)

23.4.1. Cancellation laws. Counting homomorphisms has been a main tool

for proving cancellation laws for finite relational structures in Section 5.4, and
it is not surprising that these results can be extended to locally finite categories
(Lovász [1972], Pultr [1973]). The following two theorems generalize Theorem 5.34,
Proposition 5.35(b) and Lemma 5.38 to categories.
Theorem 23.14. Let a and b be two objects in a locally finite category such that the
direct powers a×k and b×k exist and are isomorphic. Then a and b are isomorphic.
Theorem 23.15. Let a, b, c be three objects in a locally finite category K such that
the direct products a × c and b × c exist and are isomorphic.
426 23. OTHER COMBINATORIAL STRUCTURES

(a) If both a and b have at least one morphism into c, then a and b are isomor-
phic.
(b) There exists an isomorphism from a × c to b × c that commutes with the
projections of a × c and b × c to c. 
So if there is any isomorphism σ in Figure 23.3, then there is one for which the
diagram commutes.

Figure 23.3.

23.4.2. Connection matrices and algebras of morphisms. For the next

theorem, we assume that K is a locally finite category that has a zero object, a left
generator, pushouts, and epi-mono decompositions. Let f be a real valued function
defined on the objects, invariant under isomorphism. We say that f is multiplicative
over coproducts, if f (a ⊕ b) = f (a)f (b) for any two objects a and b.
For every object a, we define a (possibly infinite) symmetric matrix M (f, a),
whose rows and columns( are indexed
) by morphisms in Kain , and whose entry in row
α and column β is f t(α ∨ β) (since α ∨ β is determined up to isomorphism, this
is well defined). Note that specializing to the category of graph homomorphisms,
M (f, a) corresponds to the multiconnection matrix; to get the simple connection
matrix, we have to restrict the row and column indices to monomorphisms.
One can extend the characterization of homomorphism functions in Corollary
5.58 to categories (Lovász and Schrijver [2010]); this theorem will also contain the
dual characterization Theorem 5.59.
Theorem 23.16. Let K be a locally finite category that has a zero object z, a left
generator, pushouts, and epi-mono decompositions. Let f be a function defined
on the objects, invariant under isomorphism. Then there is an object b such that
f = |K(., b)| if and only if the following conditions are fulfilled:
(F1) f (z) = 1,
(F2) f is multiplicative over coproducts, and
(F3) M (f, a) is positive semidefinite for every object a.
We note that if there is an epimorphism from a to b, then M (f, b) is a submatrix
of M (f, a). Thus it would be enough to require the semidefiniteness condition for
a left-cofinal subset of elements a.
Corollary 23.17. Conditions (F1)–(F3) of the theorem imply that (a) the values
of f are non-negative integers, (b) the rank of M (f, a) is finite for every a.
Statement (a) of this corollary contrasts it with Theorem 5.54, where (thanks
to the weights) the function values can be arbitrary real numbers. An analogue
of (b) must be imposed as an additional condition e.g. in the characterization in
Theorem 5.54, while in this version it follows from the other assumptions.
 23.4. CATEGORIES 427

The conditions are very similar to those in Theorem 5.54, except that there
the graphs cannot have loops and the matrices are indexed by monomorphisms
only. As a consequence, the characterization concerns homomorphism numbers
into weighted graphs, which has not been extended to categories so far.
The proof of Theorem 23.16 is built on similar ideas as the proof of Theorem
5.54 in Chapter 6, using algebras associated with the category. Since it is instructive
how such algebras can be defined, we describe their construction below; for the
details of the proof, we refer to the paper of Lovász and Schrijver [2010].
For two objects a and b in a locally finite category K, a formal linear com-
bination (with real coefficients) of morphisms in K(a, b) will be called a quantum
morphism. Quantum morphisms between a and b form a finite dimensional linear
space Q(a, b). Let
∑ ∑
x= xφ φ ∈ Q(a, b) and y= yψ ψ ∈ Q(b, c),
φ∈K(a,b) ψ∈K(b,c)

then we define ∑
xy = xφ yψ φψ ∈ Q(a, c).
φ∈K(a,b)
ψ∈K(b,c)
With this definition, quantum morphisms form a category Q on the same set of
objects as K. (Of course, Q is not locally finite any more, but it is locally finite
dimensional.)
We can be more ambitious and take formal linear combinations of morphisms
in Kaout (for a fixed object a), to get a linear space Qout
a . This space will be infinite
dimensional in general, but it has interesting finite dimensional factors. For each
object a, the pushout operation ∧ defines a semigroup on Kaout . Let Qout a denote
its semigroup
⊕ algebra of all formal finite linear combinations of morphisms in Kaout .
So Qa = b Q(a, b).
out

( every)function f : Ob(K) → R defines an inner

Just as in the case of graphs,
product on Qouta , by ⟨α, β⟩ = f h(α∧β) . Condition (F3) in Theorem 23.16 implies
that this inner product is positive semidefinite. Factoring out its kernel, we get a
Frobenius algebra, which is finite dimensional (this takes a separate argument,
since unlike in the proof of Theorem 5.54, this is not assumed directly). The proof
of Theorem 23.16, just like the proof of Theorem 5.54, is built on studying the
idempotent bases in these algebras.
Example 23.18 (Graph algebras). If the category is the category of graph
homomorphisms, and a is the k-labeled graph with k nodes and no edges, then
Qout
a is the gluing algebra of k-multilabeled graphs. 
Example 23.19 (Flag algebras). Razborov’s “flag algebras” [2007] can be de-
fined in our setting as follows. We consider the category of embeddings (injective ho-
momorphisms) between graphs. Fixing a graph F (which Razborov calls a “type”),
the morphisms from F correspond to graphs with a specified subgraph isomorphic
with F (which Razborov calls a “flag”). The pushout of two such morphisms re-
sults in an object obtained by gluing together the two graphs along the image of F ,
which is exactly how Razborov defines the product in flag algebras. So flag algebras
are the algebras Qout
F in the category of monomorphisms between graphs. This is
a subalgebra of the algebra QoutF defined in terms of all homomorphisms between
graphs. 
428 23. OTHER COMBINATORIAL STRUCTURES

23.4.3. Regularity Lemma for categories. There are more results on

graph homomorphisms that extend quite naturally to the categorial setting. Let
us state a generalization of the Regularity Lemma—both in its weak and original
form (Lovász [Notes]).
To motivate the definitions below, consider a weighted graph G. This can be
viewed as a weighting of the edges of a complete graph, i.e., as a quantum morphism
K2 → K̃n , which is symmetric, i.e., it is invariant under swapping the two nodes
of K2 . Regularity lemmas try to find a partition (a morphism Kn◦ → K̃k ), and
weighting of edges of d = K̃k (a quantum morphism in Q(a, d)), such that “pulling
back” these weights to Kn◦ , we get a good approximation in the cut norm. How to
translate to the categorial language that the cut norm of a weighted graph is small?
It means that for every morphism Kn◦ → K2◦ , if we push forward the edgeweights,
then the resulting edgeweights of K2◦ are all small (this says that versions (a) and (c)
of the cut norm in Exercise 8.4 are small, but these are all equivalent up to absolute
constant factors). These considerations motivate the following general definitions.
Let α ∈ K(a, b) and β ∈ K(c, b). We define a quantum morphism αβ ∗ ∈ Q(a, c)
by ∑
αβ ∗ = φ.
φ∈K(a,c): φβ=α

This operation extends linearly to define xy ∗ for x ∈ Q(a, b) and y ∈ Q(c, b). It is
not hard to check that x(zy)∗ = (xy ∗ )z ∗∑
, and ⟨x, yz ∗ ⟩ = ⟨xz, y⟩.
For every quantum morphism x = φ xφ φ ∈ Q(a, b) and every object c, we
define the c-norm of x by
∥xβ∥∞
∥x∥c = max .
β∈K(b,c) |K(a, b)|

This norm generalizes the cut norm: if a = K2 and c = K2◦ , then a symmetric
quantum morphism x ∈ Q(a, b) is a weighting of the edges of b, and it is not hard
to see that ∥x∥ /2 ≤ ∥x∥c ≤ ∥x∥ .
Let cm denote the m-th direct power of the object c. The first inequality in
the following lemma generalizes the Frieze–Kannan Weak Regularity Lemma 9.3,
while the second implies the Original Regularity Lemma of Szemerédi 9.2.
Lemma 23.20. Let K be a locally finite category having finite direct products. Let
a, b and c be three objects in K, and let m ≥ 1. Then for every x ∈ Q(a, b) there
exists a morphism φ ∈ K(b, cm ) and a quantum morphism y ∈ Q(a, cm ) such that
1
∥x − yφ∗ ∥c ≤ √ ∥x∥2
m
and
1
∥x − yφ∗ ∥c2m ≤ √ ∥x∥2 .
log∗ m
The Weak Regularity Lemma is obtained, as described above, by taking a = K2
and c = K2◦ and applying the first bound. Note that a morphism in K(b, cm )
corresponds to a partition of V (G) into 2m classes. The Original Regularity Lemma
can be derived from the second bound similarly. Strong versions can be generalized
as well, but for the details we refer to Lovász [Notes].
There are many unsolved questions here: can the Counting Lemma be general-
ized to categories? Do the notions of convergence and limit objects be formulated
 23.5. AND MORE... 429

in an interesting way? Could these results shed new light on hypergraph limits and
regularity lemmas? Or perhaps even on sparse regularity lemmas?
Exercise 23.21. Let K be a locally finite category, and let c be an object. Prove
that every monomorphism in K(c, c) is an isomorphism.
Exercise 23.22. Let K be a locally finite category, and let c, d be two objects.
Suppose that there are monomorphisms in K(c, d) and in K(d, c). Prove that c
and d are isomorphic.
Exercise 23.23. Let K be a locally finite category, and let c and d be two objects.
For any two morphisms α ∈ K(a, a′ ) and β ∈ K(b, b′ ), let Nα,β denote the number
of 4-tuples of morphisms (φ, ψ, µ, ν) (φ ∈ K(c, a), ψ ∈ K(c, b), µ ∈ K(a′ , d), ν ∈
K(b′ , d)) such that φαµ = ψβν. Prove that the matrix N = (Nα,β ), where α and
β range over all morphisms of the category, is positive semidefinite.
Exercise 23.24. Let a and b be two objects in a locally finite category. Suppose
that the direct powers a × a and b × b exist and are isomorphic. Prove that a and
b are isomorphic.
Exercise 23.25. Let a, b, c, d be four objects in a locally finite category K such
that the direct products a × c, b × c, a × d and b × d exist, a × c and b × c are
isomorphic, and d has at least one morphism into c. Prove that a × d and b × d
are isomorphic.

23.5. And more...

There are many types of discrete structures for which one can try to define
convergence and limit objects for growing sequences. This is typically not straight-
forward, as one can see from the case of simple graphs with (say) Θ(n3/2 ) edges.
However, this approach has been successful in some cases.
It is a natural question to extend the theory of graph limits to directed graphs.
Let us assume that these graphs are simple, so that there are no loops and there
is at most one edge between two nodes in a given direction. Diaconis and Janson
show that at least some of the theory can be developed based on the theory of
exchangeable arrays (see Section 11.3.3). The limit object is a bit more complicated,
it can be described by four measurable functions W0,0 , W0,1 , W1,0 , W1,1 : [0, 1]2 →
[0, 1] such that W0,0 and W1,1 are symmetric, W0,1 (x, y) = W1,0 (y, x) and W0,0 +
W0,1 + W1,0 + W1,1 = 1. The function W0,1 (x, y) measures the density of edges
from an infinitesimal neighborhood of x to an infinitesimal neighborhood of y etc.
Some further remarks and observations can be found scattered in papers, but no
comprehensive treatment seems to be known. Perhaps most of the extension is
rather straightforward (but be warned: the theory of existence of homomorphisms
between digraphs is much more involved—one can say richer—than for undirected
graphs; see Hell and Nešetřil [2004]).
Posets can be considered as special digraphs, but they are sufficiently important
in many contexts to warrant a separate treatment. Janson [2011a, 2012] starts a
limit theory of posets. The treatment is based on methods similar to the limit
theory of dense graphs in this book, but there are some analytic complications and
interesting special features, for which we refer to the paper.
Going away from graphs, let us consider the set Sn of permutations of the set [n].
Cooper [2004, 2006] defined and characterized quasirandomness for permutations,
and proved a regularity lemma for them. Hoppen, Kohayakawa, Moreira, Ráth
and Menezes Sampaio [2011, 2011] defined convergent sequences of permutations,
and described their limit objects. Given a permutation π ∈ Sn and a subset A =
430 23. OTHER COMBINATORIAL STRUCTURES

{a1 , . . . , ak } ⊆ [n], we can define a permutation π[A] ∈ Sk by letting π[A]i < π[A]j
iff πai < πaj . For a permutation τ ∈ Sk , let Λ(τ, π) denote the number( ) of sets A with
π[A] = τ , and define the density of τ in π by t(τ, π) = Λ(τ, π)/ nk . A sequence
of permutations π1 , π2 , . . . (on larger and larger sets) is convergent, if for every
permutation τ , the number t(τ, πn ) tends to a limit as n → ∞. Every convergent
permutation sequence has a limit object in the form of a coupling measure on [0, 1]2 ,
which is uniquely determined. Král and Pikhurko [2012] have used this machinery
of limit objects to prove a conjecture of Graham on permutations.
I have already mentioned the limit theory of metric spaces due to Gromov
[1999]. While developed with quite different applications in mind, this turns out to
be closely related to our theory of graph limits. Gromov considers metric spaces
endowed with a probability measure, and defines distance, convergence and limit
notions for them. A simple graph G can be considered as a special case, where
the distance of two adjacent nodes is 1/2, the distance of two nonadjacent nodes
is 1, and the probability distribution on the nodes is uniform. Under this corre-
spondence, our notion of graph convergence is a special case of Gromov’s “sample
convergence” of metric spaces. Vershik [2002, 2004] considers random metric spaces
on countable sets, and defines and proves their universality. He also characterizes
isomorphism of metric spaces with measures in terms of sampling, analogously to
Theorem 13.10. In a recent paper, Elek [2012b] explores this connection and shows
how Gromov’s notions imply results about graph convergence, and also how results
about graph limits inspire answers to some questions about metric spaces. Perhaps
Gromov’s theory can be applied to graph sequences that are not dense, using the
standard distance between nodes in the graph.
One of the earliest limit theories is John von Neumann’s theory of continuous
geometries. The idea here is that if we look at higher and higher dimensional vector
spaces over (say) the real field, then the obvious notion of their limit is the Hilbert
space. But, say, we are interested in the behavior of subspaces whose dimension
is proportional to the dimension of the whole space. Going to the Hilbert space,
this condition becomes meaningless. Neumann constructed a limit object, called
a continuous geometry, in which the “dimensions” of subspaces are real numbers
between 0 and 1. This construction can be extended to certain geometric lattices
(Björner and Lovász [1987]), but its connection with the theory in this book has
not been explored.
Perhaps most interesting from the point of view of quasirandomness and limits
are sequences of integers, due to their role in number theory. (After all, Szemerédi’s
Regularity Lemma was inspired by his solution of the Erdős–Turán problem on
arithmetic progressions in dense sequences of integers.) Often sequences are con-
sidered modulo n; this gives a finite group structure to work with, while one does
not lose much in generality. Ever since the solution of the Erdős–Turán problem
for 3-term arithmetic progressions by Roth [1952], through the general solution by
Szemerédi [1975], through the work of Gowers [2001] on “Gowers norms”, to the
celebrated result of Green and Tao [2008] on arithmetic progressions of primes, a
central issue has been to define and measure how random-like a set of integers is.
I will not go into this large literature; Tao [2006c] and Kra [2005] give accessible
accounts of it. What I want to point out is the exciting asymptotic theory of struc-
tures consisting of an abelian group together with a subset of its elements, and
more generally, abelian groups with a function defined on them. There has been a
 23.5. AND MORE... 431

lot of parallel developments in this area, most notably the work of Green, Tao and
Ziegler [2011] and of Szegedy [2012a]. Not surprisingly, the latter is closer to the
point of view taken in this book, and develops a theory of limit objects of functions
on abelian groups, which is full of surprises but also with powerful results. (For
example, to describe the limits of abelian groups, non-abelian groups are needed!)
The theory has connections with number theory, ergodic theory, and higher-order
Fourier analysis. This explains why I cannot go into the details, and can only refer
to the papers.
 APPENDIX A

Appendix

A.1. Möbius functions

Let L be a finite lattice (for us, it will be either the lattice of all subsets, or the
lattice of all partitions, of a finite set V ). The Möbius function of the lattice is a
function µ : L × L → R, defined by the equations
∑
µ(x, y) = 0 if x  y, µ(x, z) = 1(x = y).
x≤z≤y

This is perhaps easier to understand in a matrix algebra setting. Let M(L) denote
the set of L × L matrices A in which Axy = 0 for any two lattice elements x  y. It
is easy to see that M(L) is closed under addition, matrix multiplication and matrix
inverse (if an inverse exists), and so it is a matrix algebra. One special matrix of
importance is the zeta matrix Z ∈ M(L) defined by Zxy = 1(x ≤ y). Clearly Z is
invertible, and M = Z −1 is a matrix with integer entries, called the Möbius matrix.
The entries of M give the Möbius function: Mxy = µ(x, y).
For ∑every function f : L → C, we define its (upper) summation function
g(x) = y≥x f (y). From g, we can recover function f by the formula f (x) =
∑
y≥x µ(x, y)g(y). This is again better seen in a matrix form: we consider f and
g as vectors in CL , then g = Zf , which is equivalent to f = Z −1 g = M g. Of
course, we can turn the lattice upside down, and derive similar formulas for the
lower summation.
The following simple but very useful matrix identity is due to Lindström [1969]
and Wilf [1968]. Let f : L → R be any function, Af be the L × L matrix with
(Af )xy = f (x ∨ y). Then
(A.1) Af = Zdiag(M f )Z T .
An important consequence of this identity states that Af is positive semidefinite if
and only if the Möbius inverse of f is nonnegative.
Example A.1. If L is the lattice of subsets of a finite set S, then µ(X, Y ) =
(−1)|Y \X| for all X ⊆ Y ⊆ S. Möbius inversion is equivalent to the inclusion-
exclusion formula in this case. 
Example A.2. Consider the lattice of partitions Πn of the finite set [n], where
the bottom element is the discrete partition P0 (with n classes), the top element
is the indiscrete partition P1 (with one class), and P ≤ Q means that P refines
Q. The Möbius function of this lattice is given by the Frucht–Rota–Schützenberger
Formula
∏
(A.2) µP = µ(0, P ) = (−1)n−|P | (|S| − 1)!
S∈P

433
434 A. APPENDIX

where |P | denotes the number of classes in the partition P. (This easily implies a
formula for µ(Q, P ), but we won’t need it.)
For the partition lattice, we need some simple identities: for every P ∈ Πn ,
∑
(A.3) (x)|R| = x|P |
R≥P

By Möbius inversion,
∑
(A.4) µP x|P | = (x)n ,
P
and from the Lindström–Wilf Formula,
∑
(A.5) µP µQ x|P ∨Q| = (x)n .
P,Q


See Van Lint and Wilson [1992] for more on the Möbius function of a lattice.

A.2. The Tutte polynomial

Several important graph invariants can be expressed in terms of the Tutte
polynomial of the graph G = (V, E) (which may have loops and multiple edges).
The quickest way to define this is by the following formula. Let c(A) (A ⊆ E)
denote the number of connected components of the graph (V, A) (including the
isolated nodes); in particular, c(E) is the number of components of G. We define
∑
(A.6) tut(G; x, y) = (x − 1)c(A)−c(E) (y − 1)c(A)+|A|−v(G) .
A⊆E

This definition does not in any way indicate the many uses this polynomial has.
The recurrence relation
(A.7) tut(G; x, y) = tut(G − e; x, y) − tut(G/e; x, y),
where e ∈ E(G) is any edge that is not a cut-edge or a loop, says much more (here
G/e denotes the graph obtained from G by contracting e, i.e., deleting one copy
of e and identifying its endpoints). If the G has i loops and j cut-edges, and no
other edges, then tut(G; x, y) = xi y j . The Tutte polynomial is multiplicative over
connected components. There are many graph invariants that satisfy recurrence
(A.7) (or some very similar recurrence), and these can be expressed as substitutions
into the Tutte polynomial (or some slight modification of it).
One often uses the following version of the Tutte polynomial, sometimes called
the cluster expansion polynomial:
∑
(A.8) cep(G; u, v) = uc(A) v |A| .
A⊆E(G)

This differs from the usual Tutte polynomial T (x, y) on two counts: first, instead
of the variables x and y, we use u = (x − 1)(y − 1) and v = y − 1; second, we scale
by uc(E) v |V | .
The cluster expansion polynomial satisfies the following identities: (a)
cep(G; u, v) = vcep(G/e; u, v) + cep(G − e; u, v) for all edges e that are not loops;
(b) cep(G; u, v) = qcep(G − i; u, v) if i is an isolated node; cep(G; u, v) = ue(G) if
G is a graph consisting of a single node. These relations determine the value of
 A.2. THE TUTTE POLYNOMIAL 435

the polynomial for any substitution. (See e.g. Welsh [1993] for more on the Tutte
polynomial.)
Chromatic polynomial. Let G = (V, E) be a multigraph with n nodes. For every
nonnegative integer q, we denote by chr(G, q) the number of q-colorations of G (in
the usual sense, where adjacent nodes must be colored differently). Clearly chr(G, q)
does not depend on the multiplicities of edges (as long as these multiplicities are
positive), and chr(G, q) = 0 if G has a loop.
Let chr0 (G, k) denote the number of k-colorations of G in which all colors occur.
Then clearly

∑
v(G) ( )
q
(A.9) chr(G, q) = chr0 (G, k) .
k
k=0

This implies that chr(G, q) is a polynomial in q with leading term q n and constant
term 0, which is called the chromatic polynomial of G. One can evaluate this poly-
nomial for non-integral values of q, when it has no direct combinatorial meaning.
We define chr(K0 , q) = 1.
It is easy to see that if q is a positive integer, then for every e ∈ E(G),
(A.10) chr(G, q) = chr(G − e, q) − chr(G/e, q).
Since this equation for polynomials holds for infinitely many values of q, it holds
identically. If i is an isolated node of G, then we have chr(G, q) = qchr(G − i; q).
From these recurrence relations a number of properties of the chromatic polyno-
mial are easily proved, for example, that its coefficients alternate in sign. Most
importantly, they imply that the chromatic polynomial is a special substitution of
the cluster expansion polynomial: chr(G, q) = cep(G; q, −1). From formula (A.8)
we get
∑
(A.11) chr(G; q) = (−1)|A| q c(A) .
A⊆E(G)

The coefficient of the linear term in the chromatic polynomial is called the chromatic
invariant of the graph. It will be convenient to consider this quantity with an
adjusted sign
∑ ′
cri(G) = (−1)e(G )−v(G)+1 ,
G′
′
where G ranges through all connected spanning subgraphs of G. It follows from
(A.10) that if G is a simple graph, then for every e ∈ E(G),
(A.12) cri(G) = cri(G − e) + cri(G/e).
This implies by induction that cri(G) > 0 if G is connected and cri(G) = 0 if G is
disconnected.
Spanning trees. Let tree(G) denote the number of spanning trees in the graph
G. This parameter has played an important role in the development of algebraic
graph theory; formulas for its computation go back to the work of Kirchhoff in the
mid-19th century. The number of spanning trees satisfies the recurrence relation
(A.13) tree(G) = tree(G − e) + tree(G/e)
436 A. APPENDIX

for every edge that is not a loop. It is best to define tree(K1 ) = 1 and tree(K0 ) =
0. One gets by direct substitution in (A.6) that for every connected graph G,
tree(G) = tut(G; 1, 1).
There are many other expressions in the literature for tree(G). Perhaps the
best known is Kirchhoff’s Formula (also called the Matrix Tree Theorem) saying
that tree(G) is equal to any cofactor of the Laplacian LG = AG − DG (here AG is
the adjacency matrix of G and DG is the diagonal matrix composed of the degrees).
There are many useful inequalities for tree(G), of which we mention two: the
trivial bound
∏
(A.14) tree(G) ≤ dG (u),
u
and the relation with the chromatic invariant, which follows easily by induction
from the recurrences (A.13) and (A.12):
(A.15) 0 ≤ chr(G) ≤ tree(G).

Nowhere zero flows. Let flo(G, q) denote the number of nowhere-zero q-flows.
To be precise, we fix an orientation of the edges for any graph G, and count maps
−
→
f : E (G) → Zq such that the sum of flow values on edges entering a given node
is equal to the sum of flow values on edges leaving the node. This number is given
by |tut(0, q − 1)|.

A.3. Some background in probability and measure theory

A.3.1. Probability spaces. We have to fix some terminology. A probability
space is a triple (Ω, A, π), where A is a sigma-algebra on the set Ω, and π is a
probability measure on π. We say that the space is separating, if for any two
elements of Ω there is a set in A containing exactly one of them. The space is
countably generated, if there is a countable subset J of A generating A (in other
words, A is the smallest sigma-algebra containing J ). It is often convenient to
assume (which we can do for free), than J is a set algebra, i.e., it is closed under
intersection and complementation. An atom of the space is a singleton with positive
measure.
Two probability spaces (Ωi , Ai , πi ) (i = 1, 2) are isomorphic if there is an
invertible map φ : Ω1 → Ω2 that gives a bijection between A1 and A2 and preserves
the measure. Two probability spaces (Ωi , Ai , πi ) (i = 1, 2) are isomorphic up to
nullsets if one can delete sets Xi ⊆ Ωi of measure 0 so that the remaining probability
spaces are isomorphic.
From the point of view of basic constructions in probability (independence,
expectation and variance of random variables etc.) the underlying probability space
does not matter much, at least as long as it is atom-free. But for more advanced
technical work, one likes to work with a robust class of them with nice properties. A
Borel sigma-algebra (also called standard Borel space) is a sigma-algebra isomorphic
to the sigma-algebra of Borel subsets of a Borel set in R. It can be shown that this
definition would not change if instead of subsets of R we allowed subsets of Rn ,
or indeed, of any separable complete metric space. A Borel probability space is
a probability space defined on a Borel sigma-algebra. Equivalently (this is non-
trivial), it is isomorphic up to nullsets to the disjoint union of a closed interval
(with the Borel sets and the Lebesgue measure) and a countable set of atoms. Every
Borel space is countably generated and separating. Every finite probability space
 A.3. SOME BACKGROUND IN PROBABILITY AND MEASURE THEORY 437

is Borel. A standard probability space (with small variations, also called a Lusin,
Lebesgue or Rokhlin space) is the completion of a Borel probability space (i.e., we
add all subsets of sets of measure 0 to the sigma-algebra). Standard probability
spaces have many useful properties, some of which will be mentioned below; in a
sense, they behave as you would expect them to behave.
In this sense Borel (or standard) spaces are quite special. On the other hand,
they are general enough so that we can restrict our attention to them; this is due
to the following fact:
Proposition A.3. Every probability space on a countably generated separating
sigma-algebra can be embedded into a Borel space in the sense that it is isomor-
phic up to nullsets to the restriction of a Borel space to a subset with outer measure
1. 
A.3.2. Measure preserving maps. Let (Ωi , Ai , πi ) (i = 1, 2) be probability
spaces. A map φ : (Ω1 , A1 , π(1 ) → (Ω)2 , A2 , π2 ) is measure preserving, if φ−1 (A) ∈
A1 for every A ∈ A2 , and π1 φ−1 (A) = π2 (A). (So the name is a bit misleading,
because it is φ−1 rather than φ that preserves measure.) A measure preserving
map is not necessarily bijective; for example, the map [0, 1] → [0, 1] defined by
x 7→ 2x mod 1 is measure preserving. We say that a measure preserving map φ is
invertible, if it is bijective and φ−1 is also measure preserving.
If φ : (Ω1 , A1 , π1 ) → (Ω2 , A2 , π2 ) is measure preserving, then for every inte-
grable function f : (Ω2 , A2 , π2 ) → R we have
∫ ∫
( )
(A.16) f φ(x) dπ1 (x) = f (x)dπ2 (x).
Ω1 Ω2

Let S [0,1] denote the semigroup of measure preserving maps [0, 1] → [0, 1], and
let S[0,1] be the group of invertible measure preserving maps [0, 1] → [0, 1].
One of the most important properties of standard probability spaces is that
under mild conditions, their measure preserving images are also standard.
Proposition A.4. Let (Ω1 , A1 , π1 ) be a standard probability space and let
(Ω2 , A2 , π2 ) be another probability space where A2 has a countable subset sepa-
rating any two points of Ω2 . Let φ : Ω1 → Ω2 be a measure preserving map.
Then (Ω2 , A2 , π2 ) is standard, and Ω′2 = Ω2 \ φ(Ω1 ) has measure 0. Furthermore,
if φ is bijective, then φ−1 is an isomorphism (Ω′2 , A2 |Ω′2 , π2 |Ω′2 ) → (Ω1 , A1 , π1 ). In
particular, φ−1 is also measure preserving.
Remark A.5. It is usually a matter of taste or convenience whether we decide
to work on a complete space or on a countably generated space. One tends to be
sloppy about this, and just say, for example, that the underlying probability space
is [0, 1], without specifying whether we mean the sigma-algebra of Borel sets or of
Lebesgue measurable sets.
Often, one implicitly assumes that the Borel sigma algebra is defined as the
set of Borel sets in a Polish space, and uses topological notions like open sets or
continuous functions to define measure theoretic notions. This is sometimes un-
avoidable (see e.g. the definition of weak convergence below), but the same Borel
sigma-algebra can be defined by very different topological spaces, and this is im-
portant in some cases even in this book. I will use this topological representation
only where it is necessary.
438 A. APPENDIX

A.3.3. The space of measures. Let T be a topological space, and let P(T )
denote the set of probability measures on the Borel subsets of T . We say that a
sequence of measures µ1 , µ2 , · · · ∈ P(T ) converges weakly to a probability measure
µ ∈ P(T ), if
∫ ∫
f dµn → f dµ (n → ∞)
T T
for every continuous bounded function f : S → R. Most often we need this notion
in the case when T is a compact metric space, so we don’t have to assume the
boundedness of f . This notion of convergence defines a topology on P(T ), which
we call the topology of weak convergence.
By Prokhorov’s Theorem (see e.g. Billingsley [1999]; this is not the most general
form), for a compact metric space K, the space P(K) is compact in the topology
of weak convergence, and also metrizable. (One can describe explicit metrizations,
like the Levy-Prokhorov metric, but we don’t need them.)
There is an important warning about weak convergence: it is not a purely
measure theoretic notion, but topological. In other words, we can have a sequence
of measures on a Borel sigma-algebra (Ω, B) that is weakly convergent if we put one
topology on Ω with the given Borel sets, but not convergent if we put another such
topology on Ω. Sometimes we play with this, and change the topology (without
changing its Borel sets) to suit our needs.

A.3.4. Coupling. A coupling measure between two probability spaces

(Ωi , Ai , πi ) (i = 1, 2) is a probability measure µ on the sigma-algebra (Ω1 , A1 ) ×
(Ω2 , A2 ) whose marginals are π1 and π2 , i.e., µ(A1 ×Ω2 ) = π1 (A1 ) and µ(Ω1 ×A2 ) =
π2 (A2 ) for all Ai ∈ Ai . In terms of random variables, a coupling measure is the
distribution of a pair (X1 , X2 ), where Xi has distribution πi . The simplest cou-
pling measure is the product measure π1 × π2 , corresponding to choosing X1 and
X2 independently. If (Ω1 , A1 , π1 ) = (Ω2 , A2 , π2 ), then the measure on the diago-
nal {(x, x) : x ∈ Ω1 } defined by µ{(x, x) : x ∈ A} = π1 (A) is another coupling
measure.
Suppose that (Ωi , Ai , πi ) is the sigma-algebra of Borel sets in a compact metric
space Ki . It is easy to see that if we fix the marginal distributions, the set of
coupling measures forms a closed (and hence compact) subspace of P(K1 × Ki ).
This space is in fact much nicer than the space of all measures, as the following
proposition shows (for a proof, see [Notes]).
Proposition A.6. Let K1 , K2 be compact metric spaces and let (Ki , Bi , λi ) be
probability spaces on their Borel sets. Let µ1 , µ2 , . . . and µ be coupling measures
between (K1 , B1 , λ1 ) and (K2 , B2 , λ2 ). Then the following are equivalent:
(i) µn → µ weakly;
(ii) µn (B1 × B2 ) → µ(B1 × B2 ) for all sets Bi ∈ Bi ;
∫ ∫
(iii) K1 ×K2 f dµn → K1 ×K2 f dµ for every function f : K1 × K2 → R that is
the limit of a uniformly convergent sequence of stepfunctions;
( (iv) There
) are measurable
( functions
) fn , f : [0, 1] → K1 ×K2 such that µn (X) =
λ fn−1 (X) , µ(X) = λ f −1 (X) , and fn → f almost everywhere.
The following construction of coupling measures follows from Proposition 3.8
of Kellerer [1984].
 A.3. SOME BACKGROUND IN PROBABILITY AND MEASURE THEORY 439

Proposition A.7. Let (Ωi , Ai , πi ) (i = 0, 1, 2) be standard probability spaces. Let

φi : Ωi → Ω0 (i = 1, 2) be measure preserving{ maps. Then there exists a}coupling
µ of (Ω1 , A1 , π1 ) and (Ω2 , A2 , π2 ) such that µ (x1 , x2 ) : φ1 (x1 ) = φ2 (x2 ) = 1.
A.3.5. Markov chains. Markov chains are very basic material in probability
theory, but usually they are defined in a more restrictive setting than what we need,
so let us give a brief introduction.
A Markov chain is described by a σ-algebra (Ω, A) (the state space), together
with a system of probability distributions (Pu : u ∈ Ω) on (Ω, A) such that Pu (A)
is a measurable function of u for each A ∈ A (the transition distributions). We call
a probability distribution π on (Ω, A) stationary, if
∫
Pu (A) dπ(u) = π(A)
A
for all A ∈ A. If the state space is finite, then the Markov chain always has a
stationary distribution. In the general case, this is not always true. One sufficient
condition for the existence is that (Ω, A) is the sigma-algebra of Borel sets in a
compact Hausdorff space K, and the map u 7→ Pu is continuous as a map from K
into P(K) with the weak topology.
The more usual description of a Markov chain as a sequence of random variables
is obtained if we also specify a starting distribution σ on (Ω, A). We start with an
X0 ∈ Ω from the distribution σ, and generate Xn+1 as a random element of Ω from
the distribution PXn . We will call the sequence (X0 , X1 , X2 , . . . ) a random walk on
Ω. If the Markov chain has a stationary distribution π, and X0 is randomly chosen
according to π, then every Xn is also from the stationary distribution, and we call
the sequence (X0 , X1 , X2 , . . . ) a stationary walk.
Every Markov chain defines a probability measure ψ on Ω × Ω by
∫
ψ(A × B) = Pu (B) dπ(u).
A
We can think of ψ(A × B) as the frequency with which a stationary walk steps from
A to B. We call the measure ψ the step distribution of the Markov chain. We say
that the Markov chain is reversible if ψ(A × B) = ψ(B × A) for any two measurable
sets A, B.
A.3.6. Martingales. A (finite or infinite) sequence (X1 , X2 , . . . ) of real val-
ued random variables is called a martingale, if for all k ≥ 0 we have E(|Xk |) < ∞,
and E(Xk+1 | X1 , . . . , Xk ) = Xk . More generally (and not quite logically), the se-
quence is called a supermartingale, if E(Xk+1 | X1 , . . . , Xk ) ≤ Xk . A submartingale
is defined analogously.
It is often convenient to define X0 = E(X1 ) (so this is a random variable that
is concentrated on a single value). Clearly all expectations in a martingale are the
same: X0 = E(X1 ) = E(X2 ) = . . . For a supermartingale, the expectations form a
non-increasing sequence.
Example A.8. Let Y1 , Y2 , . . . be independent random variables such that E(Yk ) =
0. Then the random variables Xk = Y1 + · · · + Yk form a martingale.
The condition that E(Yk ) = 0 can of course be arranged, as soon as the expec-
tations exist, by subtracting its expectation from each Yk , which does not influence
the independence of these variables. So the results of martingale theory can be
440 A. APPENDIX

applied to the partial sums of any sequence of independent random variables with
finite expectations. 
Many applications in combinatorics use martingales through the following con-
struction.
Example A.9 (Doob’s Martingale). Let (Ω, A, π) be a probability space and
let f : Ω → R be an integrable function. Let Y1 , . . . , Yn be independent random
elements of Ω from the distribution π, and let Xk = E(f (Y1 , . . . , Yn | Y1 , . . . , Yk ).
Then (X1 , . . . , Xn ) is a martingale. 
Example A.10. Let f : [0, 1] → R be an integrable function, and let P1 , P2 , . . .
be a sequence of partitions of [0, 1] into a finite number of measurable parts such
that Pn+1 is a refinement of Pn . Let Y ∈ [0, 1] be a uniform random point, and
consider the sequence Xk = fPk (Y ). Then (X1 , X2 , . . . ) is a martingale. Instead of
[0, 1], we could of course consider any probability space, for example, [0, 1]2 , which
shows the connection of martingales with the stepping operator. 
There are (at least) three theorems on martingales that are relevant for com-
binatorial applications; these play an important role in our book as well.
Let (X0 , X1 , . . . ) be a sequence of random variables. A random variable T with
nonnegative integral values is called a stopping time (for the sequence (X0 , X1 , . . . )),
if for every k ≥ 0, the event T = k, conditioned on X1 , . . . , Xk , is independent of
the variables Xk+1 , Xk+2 . . . (In computer science, this is often called a stopping
rule: we decide whether we want to stop after k steps depending on the values of
the variables we have seen before, possibly using some new independent coin flips).
The Martingale Stopping Theorem (a.k.a. Optional Stopping Theorem) has
many versions, of which we state one:
Theorem A.11. Let (X1 , X2 , . . . ) be a supermartingale for which |Xm+1 − Xm |
is bounded (uniformly for all m), and let T be a stopping time for which E(T ) is
finite. Then E(XT ) ≤ X0 .
For a martingale, we have equality in the conclusion, and for a submartingale,
we have the reverse inequality in the conclusion.
The following fact is called the Martingale Convergence Theorem (again, we
don’t state it in its most general form).
Theorem A.12. Let (X1 , X2 , . . . ) be a martingale such that supn E(|Xn |) < ∞.
Then (X1 , X2 , . . . ) is convergent with probability 1.
Applying this theorem to the martingale in Example A.10, we get that if f is
integrable, then the functions fPk tend to a limit almost everywhere. This limit
may not be the function f itself, but it is equal to f almost everywhere if any two
points of [0, 1] are separated by one of the partitions Pn (cf. also Proposition 9.8).
If we want to prove that a random variable is highly concentrated around its
average, most of the time we use Azuma’s Inequality (or one of its corollaries).
Theorem A.13. Let (X1 , X2 , . . . ) be a martingale such that |Xm+1 − Xm | ≤ 1 for
every m ≥ 0. Then
( )
P Xm > X0 + λ < e−λ /(2m) .
2

Applying Azuma’s Inequality to the martingale (−X1 , −X2 , . . . ), we can bound

the probability that Xm < X0 − λ. Applying it to the martingale in Example A.8,
 A.4. MOMENTS AND THE MOMENT PROBLEM 441

we get the following inequality, which (up to minor variations) is called Bernstein’s,
Chernov’s or Hoeffding’s:

Corollary A.14. Let X1 , X2 , . . . be i.i.d. random variables, and assume that

|Xi | ≤ 1. Then
( )
1 ∑
m
Xm − E(X1 ) > ε < e−ε m/2 .
2
P
m i=1

For us, it will be most convenient to use the following corollary of Azuma’s
Inequality, obtained by applying it to the martingale in Example A.9:

Corollary A.15. Let (Ω, A, π) be a probability space, and let f : Ωn → R be a mea-

surable function such that |f (x1 , . . . , xn ) − f (y1 , . . . , yn )| ≤ 1 whenever (x1 , . . . , xn )
and (y1 , . . . , yn ) differ in one coordinate only. Let x be a random point of Ωn
(chosen according to the product measure). Then
( ( ) )
P f (x) − E f (x) > εn < e−ε n/2 .
2

There are also reverse martingales. A sequence (X1 , X2 , . . . ) of real valued

random variables is called a reverse martingale, if for all k ≥ 0 we have E(|Xk |) <
∞, and E(Xk | Xk+1 , Xk+2 , . . . ) = Xk+1 . A finite reverse martingale is just a
martingale backwards, but infinite reverse martingales are different. While reverse
martingales don’t seem to be as important as martingales, there is a very important
example.

Example A.16. Let (Y1 , Y2 , . . . ) be i.i.d. real valued random variables with
E(|Yi |) < ∞, and let Xk = (Y1 + · · · + Yk )/k. Then (X1 , X2 , . . . ) is a reverse
martingale. So partial sums Y1 + · · · + Yk form a martingale, but dividing by the
number of terms, we get a reverse martingale. (The latter is a bit trickier to verify.)


The Martingale Convergence Theorem has an analogue for reverse martingales

(which holds under more general conditions and is easier to prove):

Theorem A.17. Every reverse martingale is convergent with probability 1.

Applying this theorem to Example A.16, we can derive the Strong Law of Large
Numbers. We refer to the book of Williams [1991] for more on martingales.

A.4. Moments and the moment problem

Throughout this book, we deal with kernels and graphons, which are functions
in two variables. We define subgraph densities in them, we consider weak isomor-
phism and its correspondence with measure preserving changes in the variables,
approximate them by stepfunctions, just to name a few analytic techniques with
graph-theoretic significance. In this Appendix we summarize some analogous no-
tions and results for functions in a single variable (see Feller [1971], Diaconis and
Freedman [2004] for more). Some of these are used in the study of kernels, some
others should serve as motivation for the problems and results in the main body of
the book.
442 A. APPENDIX

Let us consider the space L∞ [0, 1] of bounded measurable functions f : [0, 1] →

[0, 1]. For such a function, we consider its moments
∫1
Mk (f ) = f (x)k dx (k = 0, 1, 2, . . . ).
0

The moment sequence of a function determines it up to a measure preserving trans-

formation:
Proposition A.18. Two bounded measurable functions f, g ∈ L∞ [0, 1] have the
same moments if and only if there are measure preserving maps φ, ψ ∈ S [0,1] such
that f ◦ φ = g ◦ ψ almost everywhere. Equivalently, there is a function h ∈ L∞ [0, 1]
and maps φ, ψ ∈ S [0,1] such that f = h ◦ φ and g = h ◦ ψ.
What makes this correspondence substantially easier to handle in the one-
variable case than in the two-variable case is that in each equivalence class of weak
isomorphism there is a special element:
Proposition A.19 (Monotone Reordering Theorem). For every measurable
function f : [0, 1] → R+ there is a monotone decreasing function h : [0, 1] → R+
and a map φ ∈ S [0,1] such that f = h ◦ φ. The function h is uniquely determined
up to a set of measure 0.
Moment sequences can be characterized; this is called the Hausdorff Moment
Problem. Given a sequence (a0 , a1 , . . . ) of nonnegative, we define two infinite matri-
∑k ( )
ces H(a) and M (a) by H(a)n,k = j=0 (−1)j kj an+j and M (a)n,k = an+k . Using
this notation, moment sequences can be characterized in different ways:
Proposition A.20. For a sequence (an ) of nonnegative numbers, the following are
equivalent:
(i) (an ) is the moment sequence of a function in L∞ [0, 1];
(ii) a0 = 1 and H(a) ≥ 0 (entry by entry);
(iii) a0 = 1 and M (a) is positive semidefinite.
We call a function f ∈ L∞ [0, 1] a stepfunction if its range is finite. The set
f −1 (x) for any x in the range is called a step of f . Note that its monotone reordering
(in the sense of Proposition A.19) is then a stepfunction in the more usual sense,
whose steps are intervals.
Moment sequences of stepfunctions can be expressed as finite sums of the form
∑
r
Mk (f ) = λ(Si )f (xi )k ,
i=1

where the Si are the steps and xi ∈ Si . Conversely, every exponential sum
∑
k
s(k) = ai bki
i=1
∑
with ai > 0 and i ai = 1 can be thought of as the moment sequence of a stepfunc-
tion. An infinite sum of this type can also be represented as the moment sequence
of a function (with countably many “steps”). Proposition A.18 implies that the
values s(k) of such an exponential sum uniquely determine the numbers ai and bi .
 A.4. MOMENTS AND THE MOMENT PROBLEM 443

This fact is “self-refining” in the sense that the following seemingly stronger
statement easily follows from it.
Proposition A.21. Let ai , bi , ci , di be nonzero real numbers (i = 1, 2, . . . ), such
that bi ̸= bj and d∑
i ̸= dj for i ̸= ∑
j. Assume that there is a k0 ≥ 0 such that for all
∞ ∞
k ≥ k0 , the sums i=1 ai bki and i=1 ci dki are convergent and equal. Then the two
sums are formally equal, i.e., there is a permutation π of N such that ai = cπ (i)
and bi = dπ (i).
Returning to stepfunctions with a finite number of steps, we note that they can
be characterized in terms of their moment matrices:
Proposition A.22. A function is a stepfunction if and only if its moment matrix
has finite rank. In this case, the rank of the moment matrix is the number of steps.
Stepfunctions are determined by a finite number of moments, and this fact
characterizes them. To be more precise,
Proposition A.23. (a) Let f ∈ L∞ [0, 1] be a stepfunction with m steps, and let
g ∈ L∞ [0, 1] be another function such that Mk (f ) = Mk (g) for k = 0, . . . , m. Then
f and g have the same moments.
(b) For every function g ∈ L∞ [0, 1] and every m ≥ 0 there is a stepfunction
f ∈ L∞ [0, 1] with m steps so that Mk (f ) = Mk (g) for k = 0, . . . , m − 1.
These results can be extended to functions f : [0, 1] → [0, 1]d quite easily; we
only formulate those that we are using in the book. Such a function is called a
stepfunction if its range is finite. Moments don’t form a sequence, but an array
with d indices (a d-array for short). For a = (a1 , . . . , ad ) ∈ Nd , the corresponding
moment of f = (f1 , . . . , fd ) is defined by
∫1
Ma (f ) = f1 (x)a1 . . . fd (x)ad dx.
0

For an array A : N → R, we define its moment matrix M = M (A) as the

d

infinite symmetric matrix whose rows and columns are indexed by vectors in Nn ,
and Mu,v = Au+v . Semidefiniteness of the moment matrix does not characterize
moment sequences if d ≥ 2, but they do at least when the function values are
bounded by 1 (Berg, Christensen and Ressel [1976], Berg and Maserick [1984]):
Proposition A.24. A d-array A is the moment array of some measurable function
f : [0, 1] → [−1, 1]d if and only if A0...0 = 1, M (A) is positive semidefinite and
|Av | ≤ 1 for all v ∈ Nd . Furthermore, f is a stepfunction if and only if M (A) has
finite rank, and the rank of M (A) is equal to the number of steps of f .
Again, stepfunctions are determined by their moments:
Proposition A.25. (a) Let f : [0, 1] → [0, 1]d be a stepfunction with m steps,
and let g : [0, 1] → [0, 1]d be another function such that Ma (f ) = Ma (g) for
a ∈ {0, . . . , m}d . Then there are measure preserving maps φ, ψ ∈ S [0,1] such that
f ◦ φ = g ◦ ψ almost everywhere. In particular, g is a stepfunction, and Ma (f ) =
Ma (g) for a ∈ Nd .
(b) For every function f : [0, 1] → [0, 1]d and every finite set S ⊆ Nd , there is
a stepfunction g : [0, 1] → [0, 1]d with at most |S| + 1 steps so that Ma (f ) = Ma (g)
for all a ∈ S.
444 A. APPENDIX

A next question would be to define moments for functions f : [0, 1]d → [0, 1].
It is not enough to use here sequences or arrays. For d = 2, the right amount of
information is contained in a graph parameter, and the subgraph densities t(F, f )
show many properties analogous to the classical results described above. Theorem
13.10, Theorem 11.52 together with Proposition 14.61, Theorem 5.54 and Theorem
16.46 are analogues of Theorems A.18, A.20, A.22, and A.23(a). Other results
(e.g, the Monotone Reordering Theorem A.19 or Theorem A.23(b) do not seem to
generalize to d = 2 in any natural way. The case d ≥ 3 clearly corresponds to
hypergraphs, where, as discussed in Chapter 23.3, new difficulties arise, and many
of the interesting questions are open.

A.5. Ultraproduct and ultralimit

Ultrafilters. Let ω ⊆ 2N . We say that ω is an ultrafilter, if it is a filter (closed
under supersets), it is closed under finite intersections, and for every X ⊆ N, either
X ∈ ω or N \ X ∈ ω, but not both. It follows that N ∈ ω and ∅ ∈ / ω. (See Bell and
Slomson [2006] for more on ultrafilters and other constructions below.)
A trivial example of an ultrafilter is the set of subsets X ⊆ N containing a
given element n ∈ N; such an ultrafilter is called principal. There are non-principle
ultrafilters; their existence can be proved using Zorn’s Lemma (i.e., the Axiom of
Choice). From now on, we fix a non-principal ultrafilter ω (it does not matter which
one).
It is sometimes convenient to call the sets in ω Big, and the sets in N \ ω, Small.
(We capitalize to make a distinction from the informal use of these words.) The
following properties are not hard to prove:
Proposition A.26. (a) The union of a finite number of Small sets is Small.
(b) The intersection of a finite number of Big sets is Big.
(c) Every finite set is Small.

Ultraproduct of sets. Let (Vi : i ∈ N) be a sequence of sets. We say that two

sequences (ai : i ∈ N) and (bi : i ∈ N) (ai , bi ∈ Vi ) are ω-equivalent, if they differ
only on a Small set of indices, i.e., if {i : ai = bi } ∈ ω. (It is easy to see that
this is an equivalence relation.) The ultraproduct of the sets ∏ Vi (with respect to
the ultrafilter ω) is obtained from their cartesian product ∏i∈N Vi by identifying
ω-equivalent∏sequences. We denote this ultraproduct by ∏ ω Vi . Formally, the
elements of ω Vi are ω-equivalence classes of sequences in i∈N Vi ; we denote the
ω-equivalence class containing a sequence a by [a].
It is not hard to see that the cardinality of the ultraproduct of a sequence of
finite non-singleton sets is continuum. ∏
Let Ui ⊆ Vi . Consider the set U of sequences (a1 , a2 , . . . ) ∈ i∈N Vi such that
ai ∈ Ui for a Large set of indices i. It is clear that if a sequence belongs to U then
so does every ω-equivalent sequence; the set of ω-equivalence
∏ classes contained in
U will be denoted (with a little abuse of notation) by ω Ui .
Ultraproduct of structures. Let Ai = (Vi , Ri1 , . . . , Rik ) be relational struc-
tures of the same type, where Rij is a relation∏on Vi with a finite number rj of
variables for j = 1, . . . , k. Their ultraproduct ω Ai is ∏ defined as the relational
structure (V, R1 , . . . , Rk ) of the same type, where V = ω Vi and for any rj se-
quences xi = (xi1 , xi2 , . . . ) (i = 1, . . . , rj ) we have ([x1 ], . . . , [xrj ]) ∈ Rj if and only
 A.6. VAPNIK–CHERVONENKIS DIMENSION 445

if (xi1 , . . . , xirj ) ∈ Rij for a large set of indices i. It is easy to see that this definition
is correct in the sense that ([x1 ], . . . , [xrj ]) ∈ Rj depends only on the equivalence
classes [x1 ], . . . , [xrj ] and not on which representative xi is chosen from [xi ].
A very important property of ultraproduct of structures is stated in the follow-
ing theorem:
Proposition A.27 (Loś’s Theorem). If every structure Ai (i = 1, 2, . . . ) satisfies
a first order sentence Φ, then their ultraproduct also satisfies Φ.
As a special case, we can look at a sequence of finite simple graphs Gi =
(Vi , Ei ), i.e., finite sets Vi with a symmetric irreflexive binary relation Ei . The
ultraproduct of them is also a simple graph: the symmetry and irreflexivity of the
relation on the ultraproduct is easy to check (or it follows from Loś’s Theorem,
since these properties of the relation can be expressed by a first-order sentence:
∀x∀y(xy ∈ E ↔ yx ∈ E), and ∀x(xx ∈ / E)). If all the graphs have degrees bounded
by D, then so does their ultraproduct, since this property can be expressed by a
first-order sentence.
Ultralimit of a numerical sequence. As a nice application of an ultrafilter
ω we can associate a “limit” to every bounded sequence of numbers. (This is a
special construction for a Banach limit of bounded sequences.) Let (a1 , a2 , . . . )
(ai ∈ [u, v]) be a bounded sequence of real numbers. We say that a real number
a is the ultralimit of the sequence (in notation limω ai = a) if for every ε > 0, the
set {i : |ai − a| > ε} is Small. (Note: ordinary convergence to a would require
that this set is finite.) It is not hard to prove that every bounded sequence of real
numbers has a unique ultralimit. Furthermore, if limω ai = a and ai ∈ [u, v] for
every i, then a ∈ [u, v].
Ultraproduct of measures.
∏ Let (Vi , Ai ) be a sigma-algebra for i∏= 1, 2, . . . .
The sets of the form ω Ai (Ai ∈ Ai ), considered as subsets of V = ω Vi , form
a Boolean algebra B (they are closed under finite union, intersection, and∏com-
∏ algebra B generates a sigma-algebra on V = ω Vi ,
plementation). The Boolean
which we denote by A = ω Ai .
Next, suppose that there is a probability measure πi on (Vi , Ai ); then we define
a setfunction on B by
(∏ )
π Ai = lim πi (Ai ).
ω
ω
It is not hard to see that π is finitely additive, and a bit harder to see that it is
a measure on B, i.e., if B1 , B2 · · · ∈ B and ∩∞ n=1 Bn = ∅, then limn π(Bn ) = 0.
Trivially π(V ) = 1. It follows by Carathéodory’s Measure Extension Theorem (see
e.g. Halmos [1950]) that π extends to a probability measure on A (which we also
denote by π). Thus (V, A, π) is a probability space,∏which we call the ultraproduct
of the probability spaces (Vi , Ai , π). We write π = ω πi . (This is a special case of
a Loeb space; see Loeb [1979].)

A.6. Vapnik–Chervonenkis dimension

In probability theory, we often have to prove that out of a large number of
“bad” events, with positive probability none happens. The trivial method (which is
sufficient surprisingly often) is to use the union bound: we can draw this conclusion
provided the sum of probabilities of the bad events is less than one. There are of
course many less trivial tricks that can be used to improve this method, and one of
446 A. APPENDIX

them is the use of the Vapnik–Chervonenkis dimension. We describe this method

only in a simple special case when we need it.
For any set V and family of subsets H ⊆ 2V , a set S ⊆ V is called shattered, if
for every X ⊆ S there is a Y ∈ H such that X = Y ∩ S. The Vapnik–Chervonenkis
dimension or VC-dimension dimVC (H) of a family of sets is the supremum of cardi-
nalities of shattered sets (Vapnik and Chervonenkis [1971]). For us, this dimension
will be always finite.
We recall two basic facts about VC-dimension.
Proposition A.28 (Sauer–Shelah Lemma). If a family H of subsets of an m-
element set has VC-dimension k, then
( )
m
|H| ≤ 1 + m + · · · + .
k
For a family H of sets, we denote by τ (H) the minimum cardinality of a set
meeting every member of H. The following basic fact about VC-dimension, is based
on the results of Vapnik and Chervonenkis [1971]. It was proved by Haussler and
Welzl [1987] in a slightly weaker form than stated here, and by Komlós, Pach and
Woeginger [1992] in a slightly stronger form.
Proposition A.29. Let J be a probability space and H, a family of measurable
subsets of J such that every A ∈ H has measure at least ε. Suppose that H has
finite VC-dimension k. Then τ (H) ≤ (8k/ε) log(1/ε).
The following fact follows easily from these basic results (Lovász and Szegedy
[2010b]):
Proposition A.30. Let H be a family of measurable sets in a probability space with
VC-dimension k such that π(A△B) ≥ ε for all A, B ∈ H. Then |H| ≤ (80k)k ε−20k .

A.7. Nonnegative polynomials

We say that a real polynomial p ∈ R[x1 , . . . , xk ] is nonnegative, in notation
p ≥ 0, if p(x1 , . . . , xk ) ≥ 0 for every real x1 , . . . , xk .
A characterization of nonnegative polynomials is an important problem. A first
guess would be that a polynomial is nonnegative if and only if it is a sum of squares
of real polynomials. This is indeed true for one and two variables, but fails to hold
for three; the polynomial
z 6 + x4 y 2 + x2 y 4 − 3x2 y 2 z 2
is a counterexample. Hilbert conjectured in 1900 in his famous lecture about 23
mathematical problems (this was number 17), and Artin proved in 1927, that the
answer becomes affirmative if we allow rational functions instead of polynomials.
Proposition A.31 (Artin’s Theorem). Every nonnegative polynomial can be
written as the sum of squares of rational functions.
In the same talk, Hilbert suggested (problem number 10) to give an algorithm
to solve diophantine equations. It will be useful to contrast this with the problem
of solving algebraic equations in reals:
Problem A.32. Given a polynomial p ∈ Z[x1 , . . . , xk ], decide whether or not
p(x1 , . . . , xk ) ≥ 0 for every real x1 , . . . , xk .
 A.8. CATEGORIES 447

Problem A.33. Given a polynomial p ∈ Z[x1 , . . . , xk ], decide whether or not

p(x1 , . . . , xk ) ≥ 0 for every integral x1 , . . . , xk .
Problem A.32 is decidable; this follows e.g. from the general result of Tarski
that the first order theory of real numbers is decidable. Problem A.33 is unde-
cidable; this follows from the solution of Hilbert’s Tenth Problem by Matiyasevich
[1970].

A.8. Categories
As in other sections of this Appendix, I only summarize some basic notation,
definitions and examples that are necessary to understand certain parts of the book.
For more definitions and facts in category theory, see e.g. Adámek, Herrlich and
Strecker [2006].
A category K consists of a set of objects Ob(K), and, for any two objects
a, b ∈ Ob(K), a set K(a, b) of morphisms. Two morphisms α ∈ K(a, b) and
β ∈ K(b, c) have a product αβ ∈ K(a, c), and this multiplication is associative
(whenever defined). For α ∈ K(a, b), we set t(α) = a (tail of α) and h(α) = b (head
of α). Let Kain [Kaout ] denote the set of morphisms with h(α) = a [t(α) = a]. (If you
want to think about morphisms as maps, then please note that I am writing the
product so that the maps should be applied in the order going from left to right.)
For every object a, we have the identity morphism ida ∈ K(a, a), which has
the property that ida α = α for every α ∈ Kaout and αida = α for every α ∈ Kain .
An isomorphism between objects a and b is a morphism ξ ∈ K(a, b) which has an
inverse ζ ∈ K(b, a) such that ξζ = ida and ζξ = idb . Two objects are isomorphic if
there is an isomorphism between them.
For every object a, we introduce an equivalence relation on Kain by α ≃ β if and
only if β = αγ for some isomorphism γ. We say that α and β are left-isomorphic. It
is also clear that if α1 , α2 ∈ K(a, b), φ ∈ K(b, c), and α1 and α2 are left-isomorphic,
then so are α1 φ and α2 φ.
We can delete any object from a category and still have a category, so to really
work in a category (in particular, to prove the existence of a particular object), we
must assume the existence of certain objects and morphisms. The existence of these
is usually easily verified when we want to apply our results to specific categories,
in particular, to the category of graphs.
Terminal and zero objects. An object is terminal, if every object has a
unique morphism into it. Any two terminal objects are isomorphic, and we will
assume that the terminal object, if it exists, is unique.
Dually, an object is a zero object, if it has a unique morphism into any object.
Product and coproduct. A set of morphisms πi ∈ K(c, ai ) (i ∈ I) is called
a product, if for every set of morphisms φi ∈ K(d, ai ) (i ∈ I) there is a unique
morphism ξ ∈ K(d, c) such that φi = ξπi for all i ∈ I. We also say that the
object c is the product of objects ai . It is easy to see that the product is uniquely
determined up to isomorphism. For two objects a and b, we denote by a × b their
product. We write a×k for the k-fold product a × · · · × a. We say that the category
has products, if every finite set of objects has a product.
Coproducts are defined by turning the arrows around: A set of morphisms
σi ∈ K(ai , c) (i ∈ I) is called a coproduct, if for every set of morphisms φi ∈ K(ai , d)
(i ∈ I) there is a unique morphism ξ ∈ K(c, d) such that φi = σi ξ. For two objects
448 A. APPENDIX

a and b, we denote by a ⊕ b their coproduct, and by a⊕k , the k-fold coproduct of a

with itself.
Pullback and pushout. For two morphisms α ∈ K(a, c) and β ∈ K(b, c),
a pair of morphisms α′ ∈ K(d, a) and β ′ ∈ K(d, b) is called a pullback of (α, β)
if α′ α = β ′ β, and whenever ξ ∈ K(e, a) and ζ ∈ K(e, b) are two morphisms such
that ξα = ζβ, then there is a unique morphism η ∈ K(e, d) such that ηα′ = ξ
and ηβ ′ = ζ. The morphism α′ α = β ′ β will be denoted by α ∨ β. The pullback
is uniquely determined up to isomorphism: if σ ∈ K(d, d1 ) is an isomorphism,
then (α′ σ, β ′ σ) is also a pullback, and we get all pullbacks this way. We say that
the category has pullbacks if every pair of morphisms into the same object has a
pullback.
The pushout of two morphisms α ∈ K(c, a) and β ∈ K(c, b) is defined analo-
gously to the pullback, just reversing the arrows: we want a pair of morphisms α′
and β ′ such that αα′ = ββ ′ and for any two morphisms ξ ∈ K(a, e) and ζ ∈ K(b, e)
such that αξ = βζ there exists a unique morphism η ∈ K(d, e) such that α′ η = ξ
and β ′ η = ζ. The morphism αα′ = ββ ′ will be denoted by α ∧ β.
It is easy to check that for α1 , α2 , β ∈ Kaout , if α1 ≃ α2 , then β ∧ α1 ≃ β ∧ α2 .
So the operation ∧ is well defined on left-equivalence classes of morphisms, and the
object h(α ∧ β) is determined up to isomorphism. Similar remarks can be made
about pullbacks.

Figure A.1. Pushout and pullback

Pushouts generalize coproducts, and pullbacks generalize products: If the cat-

egory has a zero element z and pushouts, then for any two objects a and b, the
pushout of the (unique) morphisms α ∈ K(z, a) and β ∈ K(z, b) points to the co-
product of a and b, and dually, the product can be obtained from the pullback of
the unique morphisms into a terminal object.
Generators. An object g is a right-generator, if for any two objects a, b and
any two different morphisms α, β ∈ K(a, b), there is a morphism η ∈ K(b, g) such
that αη ̸= βη. A left-generator is defined dually.
Epi-mono decomposition. A morphism φ ∈ K(a, b) is a monomorphism, if
for any two morphisms α, β ∈ K(c, a) such that αφ = βφ, it follows that α = β. An
epimorphism, is defined dually, reversing the arrows. We say that the category has
epi-mono decompositions, if every morphism has a decomposition into the product
of an epimorphism and a monomorphism.
Monomorphisms of K are closed under product, and hence they form a category
Kmon . The category Kepi is defined similarly.
Locally finite categories. A category K is called locally finite, if K(a, b) is
finite for all a, b. Categories of homomorphisms between finite objects (graphs,
 A.8. CATEGORIES 449

groups etc.) are locally finite. These categories have many pleasant properties,
for example, if two objects have monomorphism into each other, then they are
isomorphic.
Let us continue with some examples.
Example A.34. The category of finite simple graphs with loops (where morphisms
are homomorphisms, i.e., adjacency-preserving maps) has all the nice properties
defined above.
It is trivial that this category is locally finite and has epi-mono decompositions.
The terminal object is the single node with a loop, and the zero object is the empty
graph. The looped complete graph on 2 nodes can serve as a right-generator object,
and the single node is a left-generator.
To construct the pullback of two homomorphisms α : a → c and β : b → c,
take the direct (categorical) product d of the two graphs a and b, together with
its projections πa and πb onto a and b, respectively, and take the subgraph d′
of d induced by those nodes v for which (πa α)(v) = (πb β)(v), together with the
restrictions of πa and πb onto d′ . (This essentially the same construction as used
in the proof of Lemma 5.38 and in the statement of Theorem 5.59.)
To construct the pushout of two homomorphisms α : c → a and β : c → b,
take the disjoint union d of the two graphs a and b, and identify the nodes α(x) and
β(x) for every node x of c. Note that in the case when c is the edgeless graph on [k],
then this is just the product of two k-multilabeled graphs, as defined in Chapter 4.
Coproduct in this category means disjoint union. Product means the direct
product of two graphs as defined in Section 3.2. 
Example A.35. Reversing the arrows in the category of finite simple graphs with
loops (Example A.34) gives another category with the above properties, since the
collection of these properties is invariant under reversing arrows. 
These examples can be extended to simplicial maps between finite simplicial
complexes, homomorphisms between directed graphs, hypergraphs, etc.
Example A.36 (Partially ordered sets). Let (P, ≤) be a partially ordered set.
For every pair x, y ∈ P such that x ≤ y, we define a unique morphism φx,y .
There is only one way to define the composition: φx,y φy,z = φx,z , which makes
sense because of the transitivity of the relation. This category is locally finite, and
every morphism is both a monomorphism and an epimorphism, so it has (trivial)
epi-mono decompositions.
If the poset is a lattice with lowest element 0 and highest element 1, then 0
is a zero object and 1 is a terminal object. Furthermore, for any two morphisms
φc,a and φc,b , the morphisms (φa,a∨b , φb,a∨b ) form their pushout. Every element is
a left and right generator (in a vacuous way). 
 Bibliography

M. Abért and T. Hubai: Benjamini-Schramm convergence and the distribution of chromatic roots
for sparse graphs, http://arxiv.org/abs/1201.3861
J. Adámek, H. Herrlich and G.E. Strecker: Abstract and Concrete Categories: The Joy of Cats,
Reprints in Theory and Applications of Categories 17 (2006), 1–507.
S. Adams: Trees and amenable equivalence relations, Ergodic Theory Dynam. Systems 10 (1990),
1–14.
R. Ahlswede and G.O.H. Katona: Graphs with maximal number of adjacent pairs of edges, Acta
Math. Hung. 32 (1978), 97–120.
R. Albert and A.-L. Barabási: Statistical mechanics of complex networks, Rev. Modern Phys. 74
(2002), 47–97.
D.J. Aldous: Representations for partially exchangeable arrays of random variables, J. Multivar.
Anal. 11 (1981), 581–598.
D.J. Aldous: Tree-valued Markov chains and Poisson-Galton-Watson distributions, in: Microsur-
veys in Discrete Probability (D. Aldous and J. Propp, editors), DIMACS Ser. Discrete Math.
Theoret. Comput. Sci. 41, Amer. Math. Soc., Providence, RI. (1998), 1–20.
D. Aldous and R. Lyons: Processes on Unimodular Random Networks, Electron. J. Probab. 12,
Paper 54 (2007), 1454–1508.
D.J. Aldous and M. Steele: The Objective Method: Probabilistic Combinatorial Optimization
and Local Weak Convergence, in: Discrete and Combinatorial Probability (H. Kesten, ed.),
Springer (2003) 1–72.
N. Alon (unpublished)
N. Alon, R.A. Duke, H. Lefmann, V. Rödl and R. Yuster: The algorithmic aspects of the regularity
lemma, J. Algorithms 16 (1994), 80–109.
N. Alon, E. Fischer, M. Krivelevich and M. Szegedy: Efficient testing of large graphs, Combina-
torica 20 (2000) 451–476.
N. Alon, W. Fernandez de la Vega, R. Kannan and M. Karpinski: Random sampling and approx-
imation of MAX-CSPs, J. Comput. System Sci. 67 (2003) 212–243.
N. Alon, E. Fischer, I. Newman and A. Shapira: A Combinatorial Characterization of the Testable
Graph Properties: It’s All About Regularity, Proc. 38th ACM Symp. on Theory of Comput.
(2006), 251–260.
N. Alon and A. Naor: Approximating the Cut-Norm via Grothendieck’s Inequality SIAM J. Com-
put. 35 (2006), 787–803.
N. Alon, P.D. Seymour and R. Thomas: A separator theorem for non-planar graphs, J. Amer.
Math. Soc. 3 (1990), 801–808.
N. Alon and A. Shapira: A Characterization of the (natural) Graph Properties Testable with
One-Sided Error, SIAM J. Comput. 37 (2008), 1703–1727.
N. Alon, A. Shapira and U. Stav: Can a Graph Have Distinct Regular Partitions? SIAM J. Discr.
Math. 23 (2009), 278–287.
N. Alon and J. Spencer: The Probabilistic Method, Wiley–Interscience, 2000.
N. Alon and U. Stav: What is the furthest graph from a hereditary property? Random Struc.
Alg. 33 (2008), 87–104.

451
452 BIBLIOGRAPHY

O. Angel and B. Szegedy (unpublished)

D. Aristoff and C. Radin: Emergent structures in large networks,
http://arxiv.org/abs/1110.1912
S. Arora, D. Karger and M. Karpinski: Polynomial time approximation schemes for dense instances
of NP-hard problems, Proc. 27th ACM Symp. on Theory of Comput. (1995), 284–293.
T. Austin: On exchangeable random variables and the statistics of large graphs and hypergraphs,
Probability Surveys 5 (2008), 80–145.
T. Austin and T. Tao: On the testability and repair of hereditary hypergraph properties, Random
Struc. Alg. 36 (2010), 373–463.
E. Babson and D. Kozlov: Topological obstructions to graph colorings, Electr. Res. Announc.
AMS 9 (2003), 61–68.
E. Babson and D.N. Kozlov: Complexes of graph homomorphisms, Isr. J. Math. 152 (2006),
285–312.
E. Babson and D.N. Kozlov: Proof of the Lovász conjecture, Annals of Math. 165 (2007), 965–
1007.
A. Bandyopadhyay and D. Gamarnik: Counting without sampling. Asymptotics of the log-
partition function for certain statistical physics models, Random Struc. Alg. 33 (2008), 452–
479.
A.-L. Barabási: Linked: The New Science of Networks, Perseus, Cambridge, MA (2002).
A.-L. Barabási and R. Albert: Emergence of scaling in random networks, Science 286 (1999)
509–512.
M. Bayati, D. Gamarnik and P. Tetali: Combinatorial Approach to the Interpolation Method
and Scaling Limits in Sparse Random Graphs, Proc. 42th ACM Symp. on Theory of Comput.
(2010), 105–114.
J.L. Bell and A.B. Slomson: Models and Ultraproducts: An Introduction, Dover Publ. (2006).
I. Benjamini, L. Lovász: Global Information from Local Observation, Proc. 43rd Ann. Symp. on
Found. of Comp. Sci. (2002), 701–710.
I. Benjamini, G. Kozma, L. Lovász, D. Romik and G. Tardos: Waiting for a bat to fly by (in
polynomial time), Combin. Prob. Comput. 15 (2006), 673–683.
I. Benjamini and O. Schramm: Recurrence of Distributional Limits of Finite Planar Graphs,
Electronic J. Probab. 6 (2001), paper no. 23, 1–13.
I. Benjamini, O. Schramm and A. Shapira: Every Minor-Closed Property of Sparse Graphs is
Testable, Advances in Math. 223 (2010), 2200–2218.
C. Berg, J.P.R. Christensen and P. Ressel: Positive definite functions on Abelian semigroups,
Math. Ann. 223 (1976), 253–272.
C. Berg and P.H. Maserick: Exponentially bounded positive definite functions, Ill. J. Math. 28
(1984), 162–179.
A. Bertoni, P. Campadelli and R. Posenato: An upper bound for the maximum cut mean value,
Graph-theoretic concepts in computer science, Lecture Notes in Comp. Sci., 1335, Springer,
Berlin (1997), 78–84.
P. Billingsley: Convergence of Probability Measures, Wiley, New York (1999).
A. Björner and L. Lovász: Pseudomodular lattices and continuous matroids, Acta Sci. Math.
Szeged 51 (1987), 295-308.
G.R. Blakley and P.A. Roy: A Hölder type inequality for symmetric matrices with nonnegative
entries, Proc. Amer. Math. Soc. 16 (1965) 1244–1245.
M. Boguñá and R. Pastor-Satorras: Class of correlated random networks with hidden variables,
Physical Review E 68 (2003), 036112.
B. Bollobás: Relations between sets of complete subgraphs, in: Combinatorics, Proc. 5th British
Comb. Conf. (ed. C.St.J.A. Nash-Williams, J. Sheehan), Utilitas Math. (1976), 79–84.
 BIBLIOGRAPHY 453

B. Bollobás: A probabilistic proof of an asymptotic formula for the number of labelled regular
graphs, Europ. J. Combin. 1 (1980), 311–316.
B. Bollobás: Random Graphs, Second Edition, Cambridge University Press, 2001.
B. Bollobás, C. Borgs, J. Chayes and O. Riordan: Percolation on dense graph sequences, Ann.
Prob. 38 (2010), 150–183.
B. Bollobás, S. Janson and O. Riordan: The phase transition in inhomogeneous random graphs,
Random Struc. Alg. 31 (2007) 3–122.
B. Bollobás, S. Janson and O. Riordan: The cut metric, random graphs, and branching processes,
J. Stat. Phys. 140 (2010) 289–335.
B. Bollobás, S. Janson and O. Riordan: Monotone graph limits and quasimonotone graphs, In-
ternet Mathematics 8 (2012), 187–231.
B. Bollobás and V. Nikiforov: An abstract Szemerédi regularity lemma, in: Building bridges,
Bolyai Soc. Math. Stud. 19, Springer, Berlin (2008), 219–240.
B. Bollobás and O. Riordan: A Tutte Polynomial for Coloured Graphs, Combin. Prob. Comput.
8 (1999), 45–93.
B. Bollobás and O. Riordan: Metrics for sparse graphs, in: Surveys in combinatorics 2009,
Cambridge Univ. Press, Cambridge (2009) 211–287.
B. Bollobás and O. Riordan: Random graphs and branching processes, in: Handbook of large-scale
random networks, Bolyai Soc. Math. Stud. 18, Springer, Berlin (2009), 15–115.
C. Borgs: Absence of Zeros for the Chromatic Polynomial on Bounded Degree Graphs, Combin.
Prob. Comput. 15 (2006), 63–74.
C. Borgs, J. Chayes, J. Kahn and L. Lovász: Left and right convergence of graphs with bounded
degree, http://arxiv.org/abs/1002.0115
C. Borgs, J. Chayes and L. Lovász: Moments of Two-Variable Functions and the Uniqueness of
Graph Limits, Geom. Func. Anal. 19 (2010), 1597–1619.
C. Borgs, J. Chayes, L. Lovász, V.T. Sós and K. Vesztergombi: Counting graph homomorphisms,
in: Topics in Discrete Mathematics (ed. M. Klazar, J. Kratochvil, M. Loebl, J. Matoušek,
R. Thomas, P. Valtr), Springer (2006), 315–371.
C. Borgs, J.T. Chayes, L. Lovász, V.T. Sós and K. Vesztergombi: Convergent Graph Sequences
I: Subgraph frequencies, metric properties, and testing, Advances in Math. 219 (2008), 1801–
1851.
C. Borgs, J.T. Chayes, L. Lovász, V.T. Sós and K. Vesztergombi: Convergent Graph Sequences
II: Multiway Cuts and Statistical Physics, Annals of Math. 176 (2012), 151–219.
C. Borgs, J.T. Chayes, L. Lovász, V.T. Sós and K. Vesztergombi: Limits of randomly grown graph
sequences, Europ. J. Combin. 32 (2011), 985–999.
C. Borgs, J.T. Chayes, L. Lovász, V.T. Sós, B. Szegedy and K. Vesztergombi: Graph Limits and
Parameter Testing, STOC38 (2006), 261–270.
L. Bowen: Couplings of uniform spanning forests, Proc. Amer. Math. Soc. 132 (2004), 2151–2158.
G. Brightwell and P. Winkler: Graph homomorphisms and long range action, in Graphs, Mor-
phisms and Statistical Physics, DIMACS Ser. Disc. Math. Theor. CS, American Mathematical
Society (2004), 29–48.
W.G. Brown, P. Erdős and M. Simonovits: Extremal problems for directed graphs, J. Combin.
Theory B 15 (1973), 77–93.
W.G. Brown, P. Erdős and M. Simonovits: On multigraph extremal problems, Problèmes combi-
natoires et théorie des graphes, Colloq. Internat. CNRS 260 (1978), 63–66.
O.A. Camarena, E. Csóka, T. Hubai, G. Lippner and L. Lovász: Positive graphs,
http://arxiv.org/abs/1205.6510
S. Chatterjee and P. Diaconis: Estimating and understanding exponential random graph models,
http://arxiv.org/abs/1102.2650
S. Chatterjee and S.R.S Varadhan: The large deviation principle for the Erdős–Rényi random
graph, Europ. J. Combin. 32 (2011), 1000–1017.
454 BIBLIOGRAPHY

F.R.K. Chung: From Quasirandom graphs to Graph Limits and Graphlets

http://arxiv.org/abs/1203.2269
F.R.K. Chung, R.L. Graham and R.M. Wilson: Quasi-random graphs, Combinatorica 9 (1989),
345–362.
F.R.K Chung and R.L. Graham: Quasi-Random Hypergraphs, Proc. Nat. Acad Sci. 86 (1989),
pp. 8175–8177.
F.R.K Chung and R.L. Graham: Quasi-random subsets of Zn , J. Combin. Theory A 61 (1992),
64–86.
D.L. Cohn: Measure Theory, Birkheuser, Boston, 1980.
D. Conlon and J. Fox: Bounds for graph regularity and removal lemmas,
http://arxiv.org/abs/1107.4829
D. Conlon, J. Fox and B Sudakov: An approximate version of Sidorenko’s conjecture, Geom.
Func. Anal. 20 (2010), 1354–1366.
J. Cooper: Quasirandom Permutations, J. Combin. Theory A 106 (2004), 123–143.
J. Cooper: A permutation regularity lemma, Electr. J. Combin. 13 R# 22 (2006).
E. Csóka: Local algorithms with global randomization on sparse graphs (manuscript)
E. Csóka: An undecidability result on limits of sparse graphs,
http://arxiv.org/abs/1108.4995
E. Csóka: Maximum flow is approximable by deterministic constant-time algorithm in sparse
networks, http://arxiv.org/abs/1005.0513
N.G. de Bruijn: Algebraic theory of Penrose’s non-periodic tilings of the plane I-II, Indag. Math.
84 (1981) 39–52, 53–66.
P. Diaconis and D. Freedman: On the statistics of vision: the Julesz Conjecture, J. Math. Psy-
chology 2 (1981) 112–138.
P. Diaconis and D. Freedman: The Markov Moment Problem and de Finetti’s Theorem: Part I,
Math. Zeitschrift 247 (2004), 183–199.
P. Diaconis, S. Holmes and S. Janson: Threshold graph limits and random threshold graphs,
Internet Math. 5 (2008), 267–320.
P. Diaconis, S. Holmes and S. Janson: Interval graph limits, http://arxiv.org/abs/1102.2841
P. Diaconis and S. Janson: Graph limits and exchangeable random graphs, Rendiconti di Matem-
atica 28 (2008), 33–61.
R.L. Dobrushin: Estimates of semi-invariants for the Ising model at low temperatures, in: Topics
in Statistical and Theoretical Physics (AMS Translations, Ser. 2) bf 177 (1996), pp. 59–81.
W. Dörfler and W. Imrich: Über das starke Produkt von endlichen Graphen, Österreich. Akad.
Wiss. Math.-Natur. Kl. S.-B. II 178 (1970), 247–262.
J. Draisma, D.C. Gijswijt, L. Lovász, G. Regts and A. Schrijver: Characterizing partition functions
of the vertex model, J. of Algebra 350 (2012), 197–206.
G. Elek: On limits of finite graphs, Combinatorica 27 (2007), 503–507.
G. Elek: The combinatorial cost, Enseign. Math. bf53 (2007), 225–235.
G. Elek: L2 -spectral invariants and convergent sequences of finite graphs, J. Functional Analysis
254 (2008), 2667–2689.
G. Elek: Parameter testing with bounded degree graphs of subexponential growth, Random Struc.
Alg. 37 (2010), 248–270.
G. Elek: On the limit of large girth sequences, Combinatorica 30 (2010), 553–563.
G. Elek: Finite graphs and amenability, http://arxiv.org/abs/1204.0449
G. Elek: Samplings and observables. Convergence and limits of metric measure spaces ,
http://arxiv.org/abs/1205.6936
G. Elek and G. Lippner: Borel oracles. An analytical approach to constant-time algorithms, Proc.
Amer. Math. Soc. 138 (2010), 2939–2947.
 BIBLIOGRAPHY 455

G. Elek and G. Lippner: An analogue of the Szemerédi Regularity Lemma for bounded degree
graphs, http://arxiv.org/abs/0809.2879
G. Elek and B. Szegedy: A measure-theory approach to the theory of dense hypergraphs, Advances
in Math. 231 (2012), 1731–1772.
P. Erdős: On sequences of integers no one of which divides the product of two others and on some
related problems, Mitt. Forsch.-Inst. Math. Mech. Univ. Tomsk 2 (1938), 74–82;
P. Erdős: On some problems in graph theory, combinatorial analysis and combinatorial number
theory, in: Graph Theory and Combinatorics, Academic Press, London (1984), 1–17.
P. Erdős, L. Lovász and J. Spencer: Strong independence of graphcopy functions, in: Graph
Theory and Related Topics, Academic Press (1979), 165-172.
P. Erdős and A. Rényi: On random graphs I, Publ. Math. Debrecen 6 (1959), 290–297.
P. Erdős and A. Rényi: On the evolution of random graphs, MTA Mat. Kut. Int. Közl. 5 (1960),
17–61.
P. Erdős and M. Simonovits: A limit theorem in graph theory, Studia Sci. Math. Hungar. 1
(1966), 51–57.
P. Erdős and A.H. Stone: On the structure of linear graphs, Bull. Amer. Math. Soc. 52 (1946),
1087–1091.
W. Feller, An Introduction to Probability Theory and its Applications, Second edition, Wiley, New
York (1971).
E. Fischer: The art of uninformed decisions: A primer to property testing, The Computational
Complexity Column of the Bulletin of the European Association for Theoretical Computer
Science 75 (2001), 97-126.
E. Fischer and I. Newman: Testing versus Estimation of Graph Properties, Proc. 37th ACM Symp.
on Theory of Comput. (2005), 138–146.
D.C. Fisher: Lower bounds on the number of triangles in a graph, J. Graph Theory 13 (1989),
505–512.
J. Fox: A new proof of the graph removal lemma, Annals of Math. 174 (2011), 561–579.
J. Fox and J. Pach (unpublished)
F. Franek and V. Rödl: Ramsey Problem on Multiplicities of Complete Subgraphs in Nearly
Quasirandom Graphs, Graphs and Combin. 8 (1992), 299–308.
P. Frankl and V. Rödl: The uniformity lemma for hypergraphs, Graphs and Combin. 8 (1992),
309–312.
P. Frankl and V. Rödl: Extremal problems on set systems, Random Struc. Alg. 20 (2002),
131–164.
M. Freedman, L. Lovász and A. Schrijver: Reflection positivity, rank connectivity, and homomor-
phisms of graphs, J. Amer. Math. Soc. 20 (2007), 37–51.
A. Frieze and R. Kannan: Quick approximation to matrices and applications, Combinatorica 19
(1999), 175–220.
O. Gabber and Z. Galil: Explicit Constructions of Linear-Sized Superconcentrators, J. Comput.
Syst. Sci. 22 (1981), 407–420.
D. Gaboriau: Invariants l2 de relations d’equivalence et de groupes, Publ. Math. Inst. Hautes.
Ètudes Sci. 95 (2002), 93–150.
D. Gamarnik: Right-convergence of sparse random graphs, http://arxiv.org/abs/1202.3123
D. Garijo, A. Goodall and J. Nešetřil: Graph homomorphisms, the Tutte polynomial and “q-state
Potts uniqueness”, Elect. Notes in Discr. Math. 34 (2009), 231–236.
D. Garijo, A. Goodall and J. Nešetřil: Contractors for flows, Electronic Notes in Discr. Math. 38
(2011), 389–394.
H.-O. Georgii, Gibbs Measures and Phase Transitions, de Gruyter, Berlin, 1988.
S. Gerke and A. Steger: The sparse regularity lemma and its applications, Surveys in Combina-
torics (2005), 227–258.
456 BIBLIOGRAPHY

E.N. Gilbert: Random graphs, Ann. Math. Stat. 30 (1959), 1141–1144.

B. Godlin, T. Kotek and J.A. Makowsky: Evaluations of Graph Polynomials, 34th Int. Workshop
on Graph-Theoretic Concepts in Comp. Sci., Lecture Notes in Comp. Sci. 5344, Springer,
Berlin–Heidelberg (2008), 183–194.
C. Godsil and G. Royle: Algebraic Graph Theory, Grad. Texts in Math. 207, Springer, New York
(2001).
O. Goldreich (ed): Property Testing: Current Research and Suerveys, LNCS 6390, Springer, 2010.
O. Goldreich, S. Goldwasser and D. Ron: Property testing and its connection to learning and
approximation, J. ACM 45 (1998), 653–750.
O. Goldreich and D. Ron: A sublinear bipartiteness tester for bounded degree graphs, Combina-
torica 19 (1999), 335–373.
O. Goldreich and D. Ron: Property testing in bounded degree graphs, Algorithmica 32 (2008),
302–343.
O. Goldreich and L. Trevisan: Three theorems regarding testing graph properties, Random Struc.
Alg. , 23 (2003), 23–57.
A.W. Goodman: On sets of aquaintences and strangers at any party, Amer. Math. Monthly 66
(1959) 778–783.
W.T. Gowers: Lower bounds of tower type for Szemerédi’s Uniformity Lemma, Geom. Func. Anal.
7 (1997), 322–337.
W.T. Gowers: A new proof of Szemerédi’s theorem, Geom. Func. Anal. 11 (2001), 465–588.
W.T. Gowers: Quasirandomness, counting and regularity for 3-uniform hypergraphs, Combin.
Prob. Comput. 15 (2006), 143–184.
W.T. Gowers: Hypergraph regularity and the multidimensional Szemeredi theorem, Annals of
Math. 166 (2007), 897–946.
B. Green and T. Tao: The primes contain arbitrarily long arithmetic progressions. Ann. of Math.
167 (2008), 481–547.
B. Green, T. Tao and T. Ziegler: An inverse theorem for the Gowers U s+1 [N ]-norm. Electron.
Res. Announc. Math. Sci. 18 (2011), 69–90.
G.R. Grimmett and D.R. Stirzaker: Probability and Random Processes, Oxford University Press
(1982).
M. Gromov: Metric structures for Riemannian and non-Riemannian spaces, Birkhäuser (1999).
A. Grzesik: On the maximum number of five-cycles in a triangle-free graph, J. Combin. Theory
B 102 (2012), 1061–1066.
P. Halmos: Measure theory, D. van Nostrand and Co., 1950.
V. Harangi: On the density of triangles and squares in regular finite and unimodular random
graphs, http://www.renyi.hu/~harangi/papers/d3d4.pdf
P. de la Harpe and V.F.R. Jones: Graph Invariants Related to Statistical Mechanical Models:
Examples and Problems, J. Combin. Theory B 57 (1993), 207–227.
A. Hassidim, J.A. Kelner, H.N. Nguyen and K. Onak: Local Graph Partitions for Approximation
and Testing, Proc. 50th Ann. IEEE Symp. on Found. Comp. Science (2009), 22–31.
H. Hatami: Graph norms and Sidorenko’s conjecture, Israel J. Math. 175 (2010), 125–150.
H. Hatami, J. Hladky, D. Kral, S. Norine and A. Razborov: On the number of pentagons in
triangle-free graphs, http://arxiv.org/abs/1102.1634
H. Hatami, J. Hladky, D. Kral, S. Norine and A. Razborov: Non-three-colorable common graphs
exist, http://arxiv.org/abs/1105.0307
H. Hatami, L. Lovász and B. Szegedy: Limits of local-global convergent graph sequences
http://arxiv.org/abs/1205.4356
H. Hatami and S. Norine: Undecidability of linear inequalities in graph homomorphism densities,
J. Amer. Math. Soc. 24 (2011), 547–565.
 BIBLIOGRAPHY 457

F. Hausdorff: Summationsmethoden und Momentfolgen I–II, Math. Z. 9 (1921), 74–109 and

280–299.
D. Haussler and E. Welzl: ε-nets and simplex range queries, Discr. Comput. Geom. 2 (1987),
127–151.
J. Haviland and A. Thomason: Pseudo-random hypergraphs. Graph theory and combinatorics
(Cambridge, 1988). Discrete Math. 75 (1989), 255–278.
J. Haviland and A. Thomason: On testing the ”pseudo-randomness” of a hypergraph. Discrete
Math. 103 (1992), 321–327.
P. Hell and J. Nešetřil: On the complexity of H-colouring, J. Combin. Theory B 48 (1990), 92–110.
P. Hell and J. Nešetřil: Graphs and Homomorphisms, Oxford University Press, 2004.
J. Hladky, D. Kral and S. Norine: Counting flags in triangle-free digraphs, Electronic Notes in
Discr. Math. 34 (2009), 621–625.
J. Hladky: Bipartite subgraphs in a random cubic graph, Bachelor Thesis, Charles University,
2006.
S. Hoory and N. Linial: A counterexample to a conjecture of Lovász on the χ-coloring complex,
J. Combin. Theory B 95 (2005), 346–349.
D. Hoover: Relations on Probability Spaces and Arrays of Random Variables, Institute for Ad-
vanced Study, Princeton, NJ, 1979.
C. Hoppen, Y. Kohayakawa, C.G. Moreira, B. Rath and R. Menezes Sampaio: Limits of permu-
tation sequences, http://arxiv.org/abs/1103.5844
C. Hoppen, Y. Kohayakawa, C.G. Moreira and R. Menezes Sampaio: Limits of permutation
sequences through permutation regularity, http://arxiv.org/abs/1106.1663
Y.E. Ioannidis and R. Ramakrishnan: Containment of conjunctive queries: Beyond relations as
sets, ACM Transactions on Database Systems, 20 (1995), 288–324.
J. Isbell: Some inequalities in hom sets, J. Pure Appl. Algebra 76 (1991), 87–110.
Y. Ishigami: A Simple Regularization of Hypergraphs, http://arxiv.org/abs/math/0612838
C. Jagger, P. Štovı́ček and A. Thomason: Multiplicities of subgraphs, Combinatorica 16 (1996),
123–141.
S. Janson: Connectedness in graph limits, http://arxiv.org/abs/math/0802.3795v2
S. Janson: Graphons, cut norm and distance, couplings and rearrangements,
http://arxiv.org/pdf/1009.2376.pdf
S. Janson: Poset limits and exchangeable random posets, Combinatorica 31 (2011), 529–563.
S. Janson: Limits of interval orders and semiorders, J. Combinatorics 3 (2012), 163–183.
S. Janson: Graph limits and hereditary properties, http://arxiv.org/abs/1102.3571
S. Janson, T. Luczak and A. Ruczynski: Random Graphs, Wiley, 2000.
V. Kaimanovich: Amenability, hyperfiniteness, and isoperimetric inequalities, C. R. Acad. Sci.
Paris Sr. I Math. 325 (1997), 999–1004.
O. Kallenberg: Probabilistic Symmetries and Invariance Principles, Springer, New York, 2005.
Z. Kallus, P. Hága, P. Mátray, G. Vattay and S. Laki: Complex geography of the internet network,
Acta Phys. Pol. B 52, 1057–1069.
G.O.H. Katona: Continuous versions of some extremal hypergraph problems, in: Combinatorics,
(Keszthely, Hungary, 1976), Coll. Math. Soc. J. Bolyai 18 (J. Bolyai Math. Soc., Budapest,
1978), 653–678.
G.O.H. Katona: Continuous versions of some extremal hypergraph problems II, Acta Math. Hung.
35 (1980), 67–77.
G.O.H. Katona: Probabilistic inequalities from extremal graph results (a survey), Ann. Discrete
Math. 28 (1985), 159–170.
A. Kechris and B.D. Miller: Topics in orbit equivalence theory, Lecture Notes in Mathematics
1852. Springer-Verlag, Berlin, 2004.
458 BIBLIOGRAPHY

A. Kechris, S. Solecki and S. Todorcevic: Borel chromatic numbers, Advances in Math. 141
(1999) 1–44.
H.G. Kellerer: Duality Theorems for Marginal Problems, Z. Wahrscheinlichkeitstheorie verw.
Gebiete 67 (1984), 399–432.
K. Kimoto: Laplacians and spectral zeta functions of totally ordered categories, J. Ramanujan
Math. Soc. 18 (2003), 53–76.
K. Kimoto: Vandermonde-type determinants and Laplacians of categories, preprint No. 2003-24,
Graduate School of Mathematics, Kyushu University (2003).
J. Kock: Frobenius Algebras and 2D Topological Quantum Field Theories, London Math. Soc.
Student Texts, Cambridge University Press (2003).
Y. Kohayakawa: Szemerédi’s regularity lemma for sparse graphs, in: Sel. Papers Conf. Found. of
Comp. Math., Springer (1997), 216–230.
Y. Kohayakawa, B. Nagle, V. Rödl and M. Schacht: Weak hypergraph regularity and linear
hypergraphs, J. Combin. Theory B 100 (2010), 151–160.
Y. Kohayakawa and V. Rödl: Szemerédi’s regularity lemma and quasi-randomness, in: Recent
Advances in Algorithms and Combinatorics, CMS Books Math./Ouvrages Math. SMC 11,
Springer, New York (2003), 289–351.
Y. Kohayakawa, V. Rödl and J. Skokan: Hypergraphs, quasi-randomness, and conditions for
regularity, J. Combin. Theory A 97 (2002), 307–352.
I. Kolossváry and B. Ráth: Multigraph limits and exchangeability, Acta Math. Hung. 130 (2011),
1–34.
J. Komlós, J. Pach and G. Woeginger: Almost Tight Bounds for epsilon-Nets, Discr. Comput.
Geom. 7 (1992), 163–173.
J. Komlós and M. Simonovits: Szemerédi’s Regularity Lemma and its applications in graph the-
ory, in: Combinatorics, Paul P. Erdős is Eighty (D. Miklós et. al, eds.), Bolyai Society
Mathematical Studies 2 (1996), pp. 295–352.
S. Kopparty and B. Rossman: The Homomorphism Domination Exponent, Europ. J. Combin.
32 (2011), 1097–1114.
S. Kopparty: Local Structure: Subgraph Counts II, Rutgers University Lecture Notes,
http://www.math.rutgers.edu/~sk1233/courses/graphtheory-F11/hom-inequalities.pdf
D. Kozlov: Combinatorial Algebraic Topology, Springer, Berlin, 2008.
B. Kra: The Green-Tao Theorem on arithmetic progressions in the primes: an ergodic point of
view, Bull. of the AMS 43 (2005), 3–23.
D. Král and O. Pikhurko: Quasirandom permutations are characterized by 4-point densities,
http://arxiv.org/abs/1205.3074
D. Kunszenti-Kovács (unpublished)
M. Laczkovich: Closed sets without measurable matchings, Proc. of the Amer. Math. Soc. 103
(1988), 894–896.
M. Laczkovich: Equidecomposability and discrepancy: a solution to Tarski’s circle squaring prob-
lem, J. Reine und Angew. Math. 404 (1990), 77–117.
M. Laczkovich: Continuous max-flow min-cut theorems, Report 19. Summer Symp. in Real Anal-
ysis, Real Analysis Exchange 21 (1995–96), 39.
J.B. Lasserre: A sum of squares approximation of nonnegative polynomials, SIAM Review 49
(2007), 651–669.
J.L.X. Li and B. Szegedy: On the logarithimic calculus and Sidorenko’s conjecture,
http://arxiv.org/abs/math/1107.1153v1
B. Lindström: Determinants on semilattices, Proc. Amer. Math. Soc. 20 (1969), 207–208.
R. Lipton and R.E. Tarjan: A separator theorem for planar graphs, SIAM Journal on Applied
Mathematics 36 (1979), 177–189.
 BIBLIOGRAPHY 459

P.A. Loeb: An introduction to non-standard analysis and hyperfinite probability theory, Prob-
abilistic Analysis and Related Topics 2 (A.T. Bharucha-Reid, editor), Academic Press, New
York (1979), 105–142.
D. London: Inequalities in quadratic forms, Duke Mathematical Journal, 33 (1966), 511–522.
L. Lovász: Operations with structures, Acta Math. Hung. 18 (1967), 321–328.
L. Lovász: On the cancellation law among finite relational structures, Periodica Math. Hung. 1
(1971), 145–156.
L. Lovász: Direct product in locally finite categories, Acta Sci. Math. Szeged 23 (1972), 319–322.
L. Lovász: Kneser’s conjecture, chromatic number, and homotopy, J. Combin. Theory A 25
(1978), 319-324.
L. Lovász: Combinatorial Problems and Exercises, Akadémiai Kiadó - North Holland, Budapest,
1979; reprinted by AMS Chelsea Publishing (2007).
L. Lovász: Connection matrices, in: Combinatorics, Complexity and Chance, A Tribute to Do-
minic Welsh Oxford Univ. Press (2007), 179–190.
L. Lovász: The rank of connection matrices and the dimension of graph algebras, Europ. J. Com-
bin. 27 (2006), 962–970.
L. Lovász: Very large graphs, in: Current Developments in Mathematics 2008 (eds. D. Jerison,
B. Mazur, T. Mrowka, W. Schmid, R. Stanley, and S. T. Yau), International Press, Somerville,
MA (2009), 67–128.
L. Lovász: Subgraph densities in signed graphons and the local Sidorenko conjecture, Electr.
J. Combin. 18 (2011), P127 (21pp).
L. Lovász: Notes on the book: Large networks, graph homomorphisms and graph limits,
http://www.cs.elte.hu/~lovasz/book/homnotes.pdf
L. Lovász and A. Schrijver: Graph parameters and semigroup functions, Europ. J. Combin. 29
(2008), 987–1002.
L. Lovász and A. Schrijver: Semidefinite functions on categories, Electr. J. Combin. 16 (2009),
no. 2, Special volume in honor of Anders Björner, Research Paper 14, 16 pp.
L. Lovász and A. Schrijver: Dual graph homomorphism functions, J. Combin. Theory A , 117
(2010), 216–222.
L. Lovász and M. Simonovits: On the number of complete subgraphs of a graph, in: Combina-
torics, Proc. 5th British Comb. Conf. (ed. C.St.J.A.Nash-Williams, J.Sheehan), Utilitas Math.
(1976), 439–441.
L. Lovász and M. Simonovits: On the number of complete subgraphs of a graph II, in: Studies in
Pure Math., To the memory of P. Turán (ed. P. Erdös), Akadémiai Kiadó (1983), 459-495.
L. Lovász and V.T. Sós: Generalized quasirandom graphs, J. Combin. Theory B 98 (2008),
146–163.
L. Lovász and L. Szakács (unpublished)
L. Lovász and B. Szegedy: Limits of dense graph sequences, J. Combin. Theory B 96 (2006),
933–957.
L. Lovász and B. Szegedy: Szemerédi’s Lemma for the analyst, Geom. Func. Anal. 17 (2007),
252–270.
L. Lovász and B. Szegedy: Contractors and connectors in graph algebras, J. Graph Theory 60
(2009), 11–31.
L. Lovász and B. Szegedy: Testing properties of graphs and functions, Israel J. Math. 178 (2010),
113–156.
L. Lovász and B. Szegedy: Regularity partitions and the topology of graphons, in: An Irregular
Mind, Szemerédi is 70, J. Bolyai Math. Soc. and Springer-Verlag (2010), 415–446.
L. Lovász and B. Szegedy: Finitely forcible graphons, J. Combin. Theory B 101 (2011), 269–301.
L. Lovász and B. Szegedy: Random Graphons and a Weak Positivstellensatz for Graphs, J. Graph
Theory 70 (2012) 214–225.
460 BIBLIOGRAPHY

L. Lovász and B. Szegedy: Limits of compact decorated graphs http://arxiv.org/abs/1010.5155

L. Lovász and B. Szegedy: The graph theoretic moment problem
http://arxiv.org/abs/1010.5159
L. Lovász and K. Vesztergombi: Nondeterministic property testing,
http://arxiv.org/abs/1202.5337
N. Lusin: Leçons par les Ensembles Analytiques et leurs Applications, Gauthier–Villars, Paris
(1930).
R. Lyons: Asymptotic enumeration of spanning trees, Combin. Prob. Comput. 14 (2005) 491–522.
R. Lyons and F. Nazarov: Perfect Matchings as IID Factors on Non-Amenable Groups, Europ.
J. Combin. 32 (2011), 1115–1125.
J.A. Makowsky: Connection Matrices for MSOL-Definable Structural Invariants, in: Logic and
Its Applications, ICLA 2009 (eds. R. Ramanujam, S. Sarukkai) (2009), 51–64.
W. Mantel: Problem 28, solution by H. Gouwentak, W. Mantel, J. Teixeira de Mattes, F. Schuh
and W.A. Wythoff, Wiskundige Opgaven 10 (1907), 60–61.
G.A. Margulis: Explicit construction of concentrators, Probl. Pered. Inform. 9 (1973), 71–80;
English translation: Probl. Inform. Transmission, Plenum, New York (1975).
Yu.V. Matiyasevich: The Diophantineness of enumerable sets. Dokl. Akad. Nauk SSSR, 191:279–
282, 1970.
J. Matoušek: Using the Borsuk-Ulam Theorem: Lectures on Topological Methods in Combina-
torics and Geometry, Springer (2003).
B.D. McKay: Maximum Bipartite Subgraphs of Regular Graphs with Large Girth, Proc. 13th SE
Conf. on Combin. Graph Theory and Computing, Boca Raton, Florida (1982), 436.
R. McKenzie: Cardinal multiplication of structures with a reflexive relation, Fund. Math. 70
(1971), 59–101.
J.W. Moon and L. Moser: On a problem of Turán, Mat. Kut. Int. Közl. 7 (1962), 283–286.
H.P. Mulholland and C.A.B. Smith: An inequality arising in genetical theory, Amer. Math.
Monthly 66 (1959), 673–683.
V. Müller: The edge reconstruction hypothesis is true for graphs with more than n · log2 n edges,
J. Combin. Theory B 22 (1977), 281–283.
B. Nagle, V. Rödl and M. Schacht: The counting lemma for regular k-uniform hypergraphs,
Random Struc. Alg. 28 (2006), 113–179.
J. Nešetřil and P. Ossona de Mendez: How many F’s are there in G? Europ. J. Combin. 32
(2011), 1126–1141.
Ilan Newman and Christian Sohler : Every Property of Hyperfinite Graphs is Testable, Proc. 43th
ACM Symp. on Theory of Comput. (2011) 675–684.
H. N. Nguyen and K. Onak, Constant-time approximation algorithms via local improvements,
Proc. 49th Ann. IEEE Symp. on Found. Comp. Science (2008), 327–336.
V. Nikiforov: The number of cliques in graphs of given order and size, Trans. Amer. Math. Soc.
363 (2011), 1599–1618.
G. Palla, L. Lovász and T. Vicsek, Multifractal network generator, Proc. NAS 107 (2010), 7640–
7645.
F.V. Petrov, A. Vershik: Uncountable graphs and invariant measures on the set of universal
countable graphs, Random Struc. Alg. 37 (2010), 389–406.
O. Pikhurko: An analytic approach to stability, Discrete Math, 310 (2010) 2951–2964.
B. Pittel: On a random graph evolving by degrees, Advances in Math. 223 (2010), 619–671.
A. Pultr: Isomorphism types of objects in categories determined by numbers of morphisms, Acta
Sci. Math. Szeged 35 (1973), 155–160.
C. Radin and M. Yin: Phase transitions in exponential random graphs,
http://arxiv.org/abs/1108.0649v2
 BIBLIOGRAPHY 461

B. Ráth and L. Szakács: Multigraph limit of the dense configuration model and the preferential
attachment graph, Acta Math. Hung. 136 (2012), 196–221.
A.A. Razborov: Flag Algebras, J. Symbolic Logic, 72 (2007), 1239–1282.
A.A. Razborov: On the minimal density of triangles in graphs, Combin. Prob. Comput. 17 (2008),
603–618.
A.A. Razborov: On 3-hypergraphs with forbidden 4-vertex configurations, SIAM J. Discr. Math.
24 (2010), 946–963.
G. Regts: The rank of edge connection matrices and the dimension of algebras of invariant tensors,
Europ. J. Combin. 33 (2012), 1167–1173.
C. Reiher: Minimizing the number of cliques in graphs of given order and edge density (manu-
script).
N. Robertson and P. Seymour: Graph Minors. XX. Wagner’s conjecture, J. Combin. Theory B
92, 325–357.
V. A. Rohlin: On the fundamental ideas of measure theory, Translations Amer. Math. Soc., Series
1, 10 (1962), 1–54. Russian original: Mat. Sb. 25 (1949), 107–150.
K. Roth: Sur quelques ensembles d’entiers, C. R. Acad. Sci. Paris 234 (1952), 388–390.
V. Rödl and M. Schacht: Regular partitions of hypergraphs: Regularity Lemmas, Combin. Prob.
Comput. 16 (2007), 833–885.
V. Rödl and M. Schacht: Regular partitions of hypergraphs: Counting Lemmas, Combin. Prob.
Comput. 16 (2007), 887–901.
V. Rödl and J. Skokan: Regularity lemma for k-uniform hypergraphs, Random Struc. Alg. 25
(2004), 1–42.
R. Rubinfeld and M. Sudan: Robust characterization of polynomials with applications to program
testing, SIAM J. Comput. 25 (1996), 252–271.
I.Z. Ruzsa and E. Szemerédi: Triple systems with no six points carrying three triangles, in: Com-
binatorics, Proc. Fifth Hungarian Colloq., Keszthely, Bolyai Society–North Holland (1976),
939–945.
R.H. Schelp and A. Thomason: Remark on the number of complete and empty subgraphs, Combin.
Prob. Comput. 7 (1998), 217–219.
O. Schramm: Hyperfinite graph limits, Elect. Res. Announce. Math. Sci. 15 (2008), 17–23.
A. Schrijver: Tensor subalgebras and first fundamental theorems in invariant theory, J. of Algebra
319 (2008) 1305–1319.
A. Schrijver: Graph invariants in the edge model, in: Building Bridges Between Mathematics and
Computer Science (eds. M. Grötschel, G.O.H. Katona), Springer (2008), 487–498.
A. Schrijver: Graph invariants in the spin model, J. Combin. Theory B 99 (2009) 502–511.
A. Schrijver: Characterizing partition functions of the vertex model by rank growth (manuscript)
A. Scott and A. Sokal: On Dependency Graphs and the Lattice Gas, Combin. Prob. Comput. 15
(2006), 253–279.
A. Scott: Szemerédi’s regularity lemma for matrices and sparse graphs, Combin. Prob. Comput.
20 (2011), 455–466.
A.F. Sidorenko: Classes of hypergraphs and probabilistic inequalities, Dokl. Akad. Nauk SSSR
254 (1980), 540–543.
A.F. Sidorenko: Extremal estimates of probability measures and their combinatorial nature (Rus-
sian), lzv. Acad. Nauk SSSR 46 (1982), 535–568.
A.F. Sidorenko: Inequalities for functionals generated by bipartite graphs (Russian) Diskret. Mat.
3 (1991), 50–65; translation in Discrete Math. Appl. 2 (1992), 489–504.
A.F. Sidorenko: A correlation inequality for bipartite graphs, Graphs and Combin. 9 (1993),
201–204.
A.F. Sidorenko: Randomness friendly graphs, Random Struc. Alg. 8 (1996), 229–241.
B. Simon: The Statistical Mechanics of Lattice Gasses, Princeton University Press (1993).
462 BIBLIOGRAPHY

M. Simonovits: A method for solving extremal problems in graph theory, stability problems, in:
Theory of Graphs, Proc. Colloq. Tihany 1966, Academic Press (1968), 279–319.
M. Simonovits: Extremal graph problems, degenerate extremal problems, and supersaturated
graphs, in: Progress in Graph Theory, NY Academy Press (1984), 419–437.
M. Simonovits and V.T. Sós: Szemerédi’s partition and quasirandomness, Random Struc. Alg. 2
(1991), 1–10.
M. Simonovits and V.T. Sós: Hereditarily extended properties, quasi-random graphs and not
necessarily induced subgraphs. Combinatorica 17 (1997), 577–596.
M. Simonovits and V.T. Sós: Hereditary extended properties, quasi-random graphs and induced
subgraphs, Combin. Prob. Comput. 12 (2003), 319–344.
Ya.G. Sinai: Introduction to ergodic theory, Princeton Univ. Press (1976).
B. Szegedy: Edge coloring models and reflection positivity, J. Amer. Math. Soc. 20 (2007),
969–988.
B. Szegedy: Edge coloring models as singular vertex coloring models, in: Fete of Combinatorics
(eds. G.O.H. Katona, A. Schrijver, T. Szőnyi), Springer (2010), 327–336.
B. Szegedy: Gowers norms, regularization and limits of functions on abelian groups,
http://arxiv.org/abs/1010.6211
B. Szegedy (unpublished).
E. Szemerédi: On sets of integers containing no k elements in arithmetic progression”, Acta
Arithmetica 27 (1975) 199–245.
E. Szemerédi: Regular partitions of graphs, Colloque Inter. CNRS (J.-C. Bermond, J.-C. Fournier,
M. Las Vergnas and D. Sotteau, eds.) (1978) 399–401.
T. Tao: A variant of the hypergraph removal lemma, J. Combin. Theory A 113 (2006), 1257–1280.
T.C. Tao: Szemerédi’s regularity lemma revisited, Contrib. Discrete Math. 1 (2006), 8–28.
T.C. Tao: The dichotomy between structure and randomness, arithmetic progressions, and the
primes, in: Proc. Intern. Congress of Math. I (Eur. Math. Soc., Zürich, 2006) 581–608.
A. Thomason: Pseudorandom graphs, in: Random graphs ’85 North-Holland Math. Stud. 144,
North-Holland, Amsterdam, 1987, 307–331.
A. Thomason: A disproof of a conjecture of Erdős in Ramsey theory, J. London Math. Soc. 39
(1898), 246–255.
P. Turán: Egy gráfelméleti szélsőértékfeladatról. Mat. Fiz. Lapok bf 48 (1941), 436–453.
J.H. van Lint and R.M. Wilson: A Course in Combinatorics, Cambridge University Press (1992).
V. Vapnik and A. Chervonenkis: On the uniform convergence of relative frequencies of events to
their probabilities, Theor. Prob. Appl. 16 (1971), 264–280.
A.M. Vershik: Classification of measurable functions of several arguments, and invariantly dis-
tributed random matrices, Funkts. Anal. Prilozh. 36 (2002), 12–28; English translation: Funct.
Anal. Appl. 36 (2002), 93–105.
A.M. Vershik: Random metric spaces and universality, Uspekhi Mat. Nauk 59 (2004), 65–104;
English translation: Russian Math. Surveys 59 (2004), 259–295.
D. Welsh: Complexity: Knots, Colourings and Counting, London Mathematical Society Lecture
Notes 186. Cambridge Univ. Press, Cambridge, 1993.
D. Welsh and C. Merino: The Potts model and the Tutte polynomial, Journal of Mathematical
Physics, 41 (2000), 1127–1152.
H. Whitney: The coloring of graphs, Ann. of Math. 33 (1932), 688–718.
H.S. Wilf: Hadamard determinants, Möbius functions, and the chromatic number of a graph,
Bull. Amer. Math. Soc. 74 (1968), 960–964.
D. Williams: Probability with Martingales, Cambridge Univ. Press, 1991.
E. Witten: Topological quantum field theory, Comm. Math. Phys. 117 (1988), 353–386.
E. Zeidler: Nonlinear functional analysis and its applications. Part I, Springer Verlag, New York,
1985.
 BIBLIOGRAPHY 463

A.A. Zykov: On Some Properties of Linear Complexes, Mat. Sbornik 24 (1949), 163–188; Amer.
Math. Soc. Transl. 79 (1952), 1–33.
 Author Index

Abért, Miklós, xiv Cooper, Joshua, 429

Adámek, Jiři, 447 Csóka, Endre, xiv, 405, 409, 410
Adams, Scot, 332
Ahlswede, Rudolph, 28 de la Harpe, Pierre, 34, 63
Albert, Réka, 10 Diaconis, Persi, 8, 14, 157, 178, 180, 184,
Aldous, David, 23, 180, 184, 338, 341, 344, 193, 218, 221, 253, 257, 260, 441
346, 357, 358 Dobrushin, Roland, 67, 367, 370, 372, 378
Alon, Noga, xiv, 8, 128, 143, 151, 154, 160, Dörfler, Willibald, 72
230, 272, 274, 276, 279, 294, 358, 385 Draisma, Jan, 418
Angel, Omer, 393 Duke, Richard, 279
Aristoff, David, 261
Arora, Sanjeev, 19, 265 Elek, Gábor, xiv, 23, 180, 184, 199, 338,
Artin, Emil, 446 341, 357, 359, 383, 384, 391, 393, 398,
Austin, Timothy, 279 399, 411, 421, 424, 430
Erdős, Paul, xi, 8, 10, 13, 16, 28, 29, 73, 74,
Babson, Eric, 80 157, 260, 293, 297, 307, 430
Bandyopadhyay, Antar, 377, 378
Barabási, Albert-László, 10 Feller, William, 441
Bayati, Mohsen, 399 Fischer, Eldar, 20, 143, 271, 272, 276
Bell, John, 444 Fisher, David, 27
Benjamini, Itai, xi, 6, 23, 338, 351, 354, Fox, Jacob, 142, 143, 198, 296, 384
356, 360, 385, 402, 403 Franek, Frantisek, 297, 298
Berg, Christian, 443 Frankl, Peter, 424
Bertoni, Alberto, 376 Freedman, David, 8, 14, 157, 441
Billingsley, Patrick, 185, 438 Freedman, Michael, xi, xiii, 47, 64, 75
Björner, Anders, 430 Frieze, Alan, 127, 142, 145, 146, 230, 279,
Blakley, George, 28 428
Boguñá, Marian, 157 Frucht, Robert, 433
Bollobás, Béla, 8, 26, 27, 123, 135, 141,
157, 158, 221, 247, 286, 287, 399, 415 Gabber, Ofer, 384
Borgs, Christian, xiii, 67, 73, 151, 153, 158, Gaboriau, Damien, 336
160, 164, 169, 174, 185, 187, 195, 201, Galil, Zvi, 384
210, 218, 221, 264, 265, 317, 375, 382 Gamarnik, David, 377, 378, 382, 399
Bowen, Lewis, 357 Garijo, Delia, 97
Brightwell, Graham, 79 Gerke, Stefanie, 395, 415
Brown, William, 307 Gijswijt, Dion, 418
Gilbert, Edgar, 8
Campadelli, Paola, 376 Godlin, Benny, 46, 51
Chatterjee, Sourav, 260 Goldreich, Oded, xii, 5, 19, 20, 263, 265,
Chayes, Jennifer, xi, xiii, 16, 67, 73, 151, 403
153, 158, 160, 164, 169, 174, 185, 187, Goldwasser, Shafi, xii, 5, 19, 265
195, 201, 210, 218, 221, 264, 265, 317, Goodall, Andrew, 97
375, 382 Goodman, Al, 26
Chervonenkis, Alexey, 446 Gowers, Timothy, 15, 142, 421, 422, 424,
Christensen, Jens, 443 430
Chung, Fan, 9, 197, 308, 421, 422 Graham, Ronald, 9, 197, 308, 421, 422
Conlon, David, 142, 143, 296 Green, Benjamin, 430

465
466 AUTHOR INDEX

Grimmet, Geoffrey, 191 Kunszenti-Kovács, Dávid, xiv, 239

Gromov, Mikhail, 430
Grothendieck, Alexander, 128 Laczkovich, Miklós, 138, 337, 338
Grzesik, Andrzej, 307 Laki, Sándor, 22
Lasserre, Jean, 303
Hága, Péter, 22 Lefmann, Hanno, 279
Halmos, Paul, 445 Li, Xiang, 296
Harangi, Viktor, 359 Lindström, Bernt, 433
Hassidim, Avinathan, 389 Linial, Nati, 80
Hatami, Hamed, 30, 239–241, 287, 295–297, Lippner, Gábor, xiv, 338, 393, 411
299, 303, 307, 348, 360, 361, 391 Lipton, Richard, 384
Hell, Pavol, 55, 425, 429 Loeb, Peter, 445
Herrlich, Horst, 447 London, David, 28
Hilbert, David, 446 Loś, Jerzy, 445
Hladky, Jan, 297, 389 Lovász, László, 6, 16, 27, 30, 47, 64, 67–70,
Holmes, Susan, 193, 218, 253 72–78, 80, 98, 101, 125, 139, 141, 143,
Hoory, Shlomo, 80 149, 151, 153, 157, 158, 160, 164, 166,
Hoover, Douglas, 180, 184 167, 169, 174, 178, 180, 185, 187, 195,
Hoppen, Carlos, 429 201, 210, 218, 221, 226, 232, 245, 253,
257, 264, 265, 268, 280, 294, 296, 303,
Imrich, Wilfried, 72 308, 309, 311, 317, 318, 324, 348, 360,
Ioannidis, Yannis, 299 361, 375, 382, 391, 418, 425–428, 430,
Isbell, John, 425 446
Ishigami, Yoshiyasu, 424 Luczak, Tomasz, 8
Lusin, Nikolai, 330
Jagger, Chris, 297, 298
Lyons, Russell, xiv, 23, 67, 338, 341, 344,
Janson, Svante, xiv, 8, 122, 123, 131, 157,
346, 357, 358, 399, 415
158, 178, 180, 184, 193, 218, 221, 247,
248, 253, 257, 429 Mátray, Péter, 22
Jones, Vaughan, 34, 63 Makowski, Janos, 46, 51
Julesz, Bela, 14 Mantel, W., 25, 26
Margulis, Grigory, 384
Kahn, Jeff, 67, 73, 375, 382
Maserick, Peter, 443
Kaimanovich, Vadim, 391
Matiyasevich, Yuri, 299, 447
Kallenberg, Olav, 184, 221
Matoušek, Jiři, 80
Kallus, Zsófia, 22
McKay, Brandan, 389
Kannan, Ravindran, 127, 142, 145, 146,
McKenzie, Ralph, 72
160, 230, 279, 428
Menezes Sampaio, Rudini, 429
Karger, David, 19, 265
Merino, Criel, 64
Karpinski, Marek, 20, 160, 265
Miller, Benjamin, 330, 336, 349, 383
Katalin, Vesztergombi, 280
Moon, John, 293
Katona, Gyula O.H., 26–30, 287, 288
Moreira, Carlos Gustavo, 429
Kechris, Alexander, 330, 336, 349, 383
Moser, Leo, 293
Kellerer, Hans, 438
Mulholland, Hugh P., 28
Kelner, Jonathan, 389
Müller, Vladimir, 69
Kimoto, Kazufumi, 425
Kock, Joachim, 84 Nagle, Brendan, 424
Kohayakawa, Yoshiharu, 395, 415, 422, 429 Naor, Assaf, 128
Kolmogorov, Andrey, 374 Nešetřil, Jarik, xiv, 55, 59, 97, 425, 429
Kolossváry, István, 318 Newman, Ilan, 271, 276, 404
Komlós, János, 446 Nguyen, Huy, 24, 389, 405, 407
Kopparty, Swastik, 59, 299 Nikiforov, Vlado, 27, 141, 289
Kotek, Tomer, 46, 51 Norine, Serguey, 30, 287, 297, 299, 303, 307
Kozlov, Dmitri, 80
Kozma, Gadi, 6 Onak, Krzysztof, 24, 389, 405, 407
Král, Daniel, 297 Ossona de Mendez, Patrice, 59
Kra, Bryna, 430
Krivelevich, Michael, 143, 272 Pach, János, 384, 446
Kruskal, Joseph, 26–28, 30, 287, 288 Paley, Raymond, 9
 AUTHOR INDEX 467

Palla, Gergely, 158 Steger, Angelika, 395, 415

Pastor-Satorras, Romualdo, 157 Stirzaker, David, 191
Penrose, Roger, 355 Stone, Arthur, 29, 293
Peres, Yuval, xiv Štovı́ček, Pavel, 297, 298
Petrov, Fedor, 178 Strecker, George, 447
Pikhurko, Oleg, xiv, 135, 139, 153, 155 Sudakov, Benjamin, 296
Posenato, Roberto, 376 Sudan, Madhu, 19
Pultr, Aleš, 425 Szakács, László, 318, 326
Szegedy, Balázs, xi, xiii, 16, 30, 77, 78, 98,
Radin, Charles, 261 108, 125, 139, 141, 143, 149, 157, 166,
Ramakrishnan, Raghu, 299 167, 178, 180, 184, 199, 226, 232, 234,
Ráth, Balázs, 318, 429 245, 253, 257, 268, 294–296, 303, 308,
Razborov, Alexander, xii, 25, 27, 284, 288, 311, 318, 324, 348, 360, 361, 391, 393,
289, 297, 307, 425, 427 418, 421, 424, 431, 446
Regts, Guus, xiv, 418, 421 Szegedy, Márió, 143, 272
Reiher, Christian, 27, 289 Szemerédi, Endre, xii, 13, 15, 21, 141, 142,
Rényi, Alfréd, 8, 10, 157, 260 167, 180, 198, 199, 276, 279, 428, 430
Ressel, Paul, 443
Riordan, Oliver, 123, 135, 157, 158, 221, Tao, Terence, 13, 141, 279, 421, 424, 430
247, 415 Tardos, Gábor, 6
Robertson, Neil, 49 Tarjan, Robert, 384
Rödl, Vojtech, 279, 297, 298, 421, 422, 424 Tetali, Prasad, 399
Romik, Dan, 6 Thomas, Robin, 385
Ron, Dana, xii, 5, 19, 265, 403 Thomason, Andrew, 9, 286, 292, 297, 298
Rossman, Benjamin, 59 Todorcevic, Stevo, 330
Rota, Gian-Carlo, 433 Trevisan, Lucca, 263
Roth, Klaus, 430 Turán, Pál, 13, 25–27, 430
Roy, Prabir, 28 Tutte, William, 48, 64, 434, 453
Rubinfeld, Ronitt, 19
Ruczinski, Andrzej, 8 van Lint, Jakobus, 434
Ruzsa, Imre, 198 Vapnik, Vladimir, 446
Varadhan, Srinivasa, 260
Sauer, Norbert, 446 Vattay, Gábor, 22
Schacht, Matthias, 421, 424 Vershik, Anatoly, 178, 430
Schelp, Richard, 286, 292 Vesztergombi, Katalin, xi, xiii, 16, 151, 153,
Schramm, Oded, xi, xiv, 23, 338, 351, 354, 160, 164, 169, 174, 185, 187, 195, 201,
356, 360, 383, 385, 389, 402, 403 202, 210, 221, 264, 265, 317
Schrijver, Alexander, xi–xiii, 53, 64, 75, 76, Vicsek, Tamás, 158
108, 174, 185, 418, 419, 421, 425–427 von Neumann, John, 430
Schützenberger, Marcel-Paul, 433
Scott, Alexander, 67, 395, 415 Walker, Kevin, xiv
Seymour, Paul, 49, 385 Welsh, Dominic, xiv, 47, 64, 435
Shapira, Asaf, 154, 274, 276, 385, 402, 403 Whitney, Hassler, 74
Shelah, Saharon, 446 Wilf, Herbert, 433
Sidorenko, Alexander, 17, 29, 295, 297 Williams, David, 441
Simonovits, Miklós, xiv, 9, 27, 29, 293, 295, Wilson, Richard, 9, 197, 308, 434
307 Winkler, Peter, 79
Sinai, Yakov, 121 Witten, Edward, 86
Skokan, Jozef, 421, 422, 424 Woeginger, Gerhard, 446
Slomson, Alan, 444
Yin, Mei, 261
Smith, Cedric, 28 Yuster, Raphael, 279
Sohler, Christian, 404
Sokal, Alan, 67 Zeidler, Eberhard, 312
Solecki, Slawomir, 330 Ziegler, Tamar, 431
Sós, Vera T., xi, xiii, xiv, 9, 16, 151, 153,
160, 164, 169, 174, 185, 187, 195, 201,
210, 221, 264, 265, 309, 317
Spencer, Joel, xi, 8, 16, 73, 74
Stav, Uri, 154, 294
 Subject Index

K-graphon, 322 edge-connection matrix, 416

edit distance, 11, 128, 352
adjacency matrix, 39 epi-mono decomposition, 448
automorphism, 234 epimorphism, 448
average ε-net, 228 equipartition, 37
Azuma’s Inequality, 163, 265, 440, 441 equitable partition, 37
equivalence graphon, 255
ball distance, 339
expansion
Bernoulli lift, 348
proper, 94
Bernoulli shift, 342
blowup of graph, 40 flag algebra, 289, 427
bond, 38 formula
Borel coloring, 330 first order, 50
Borel graph, 329 monadic second order, 50
node-monadic second order, 50
Cauchy sequence, 174
free energy, 265
chromatic invariant, 435
Frobenius algebra, 84
chromatic polynomial, 47, 435
Frobenius identity, 84
circle graph, 416
cluster of homomorphisms, 80
grandmother graph, 341
component map, 341
graph
concatenation of graphs, 85
H-colored
conjugate in concatenation algebra, 85
isomorphic, 71
connection matrix, 43
ε-homogeneous, 141
dual, 76
ε-regular, 142
flat, 43
k-broken, 416
connection rank, 44, 107
bi-labeled, 85
connector, 94
decorated, 120
constituent of quantum graph, 83
directed, 429
continuous geometry, 430
edge-weighted, 38
contractor, 94
eulerian, 63, 89
convergent graph sequence, 173
expander, 384
Counting Lemma, 167
flat, 39
Inverse, 169
fully labeled, 39
cut
gaudy, 51
maximum, 64
H-colored, 71
cut distance, 12, 128, 132
hyperfinite, 383
cut metric, 128
labeled, 39
cut norm, 127, 131
looped-simple, 38
distance from a norm, 135 measure preserving, 332
distinguishable by sampling, 266 multilabeled, 39
dough folding map, 342 norming, 239
partially labeled, 39, 85
edge-coloring, 417 planar, 383
edge-coloring function, 417 random
edge-coloring model, 417 evolving, 8

469
470 SUBJECT INDEX

involution invariant, 340 graphon variety, 250

multitype, 8 r-graph, 421
randomly weighted, 77 Grothendieck norm, 128
seminorming, 239 ground state energy, 65
series-parallel, 88 microcanonical, 204
signed, 39
simple, 38 Hölder property, 240
simply labeled, 39 Hadwiger number, 54
W-random, 157 half-graph, 13
weakly norming, 239 Hamilton cycle, 47
weighted, 38 Hausdorff Moment Problem, 442
Graph of Graphs, 339 homomorphism, 55
graph parameter, 41 homomorphism density, 6, 58
additive, 41 homomorphism entropy, 202, 204, 375
contractible, 95 homomorphism frequency, 59
estimable, 263 hypergraph
finite rank, 44 r-uniform, 421
gaudy, 52 complete, 422
isolate indifferent, 257 complete r-partite, 422
isolate-indifferent, 41 hypergraph sequence
maxing, 41 quasirandom, 422
minor-monotone, 42
idempotent
multiplicative, 41
degree, 91
normalized, 41
resolve, 89
reflection positive, 44
support, 102
flatly, 44
idempotent basis, 89
simple, 41
internet, 3
graph property, 41
intersection graph, 38
distinguishable by sampling, 402
interval graph, 218
hereditary, 42, 247
isomorphism up to a nullset, 121
minor-closed, 42
monotone, 41
kernel, 115, 217
random-free, 247
connected, 122
robust, 273
direct sum, 122
testable, 272, 402
pullback, 217
graph sequence
pure, 222
bounded growth, 384 regular, 250
hyperfinite, 383 tensor product, 123
locally convergent, 16, 353 twin-free, 219
quasirandom, 9, 187, 197 unlabeled, 132
multitype, 10 weakly isomorphic, 121
subexponential growth, 384 kernel variety, 250
graphing, 23, 332 simple, 250
bilocally isomorphic, 347
cyclic, 330 left-convergent, 173
cylic, 334 left-generator, 448
hyperfinite, 385 limit
local isomorphism, 346 dense graph sequence, 180
locally equivalent, 346 weak, 175
random rooted component, 341 line-graph, 38, 331
stationary distribution, 333 local distance, 331
graphon, 16, 115, 217
degree function, 116 Möbius function, 433
finitely forcible, 308 Möbius inverse, 42
infinitesimally finitely forcible, 314 Möbius matrix, 433
triangle-free, 247 Markov chain, 416, 439
graphon property, 247 reversible, 439
testable, 268 stationary distribution, 439
 SUBJECT INDEX 471

step distribution, 439 separating, 436

matching standard, 437
perfect, 84 product of graphs, 40
measure gluing, 42
ergodic, 259 property testing, 5, 19
unimodular, 340
measure preserving, 336 quantum graph, 83
Minkowski dimension, 231 k-labeled, 84
moment matrix, 443 loopless, 84
monomorphism, 448 simple, 84
Monotone Reordering Theorem, 253, 442 quantum morphism, 427
morphism quasirandomness, 422
coproduct, 447 quotient
left-isomorphic, 447 graph
product, 447 fractional, 212
pullback, 448 quotient graph, 40, 141, 144
pushout, 448 simple, 40
multicut, 64 quotient set, 211
restricted, 204 fractional, 212
multigraph, 38 restricted, 211
multiplicative over coproducts, 426
Rado graph, 178
near-blowup of graph, 40 random graph
neighborhood distance, 222 ultimate, 256
neighborhood sampling, 22 random graph model, 175
networks, 3 consistent, 175
node countable, 177
labeled, 28 local, 175
node cover number, 46 local countable, 177
node evaluation function, 417 random graphon model, 256
norm random variable
invariant, 135 exchangeable, 184
random walk, 439
object rank of kernel, 254
terminal, 447 Reconstruction Conjecture, 69
zero, 447 (α, β, k)-regularization, 424
oblivious testing, 272 Regularity Lemma, xii, 14, 142, 145, 199
orientation Strong, 143, 148
eulerian, 49, 97, 119 Weak, 142
overlay Removal Lemma, 197, 198, 273
fractional, 129 right-generator, 448
root, 339
Pólya urn, 191
Paley graph, 9, 237 S-flow, 63
parameter estimation, 18 sample concentration, 158, 159
partially ordered set, 429 sampling distance, 12, 351
partition nondeterministic, 360
fractional, 210, 212 Sampling Lemma
legitimate, 225 First, 160
partition function, 265 Second, 164, 165
Penrose tiling, 355 Schatten norm, 135
perfect matching, 46, 417 semidefinite programming, 304
preferential attachment graph sequence
growing, 191 well distributed, 185
prefix attachment graph, 188 sequence of distributions
probability space, 436 consistent, 353
atom, 436 involution invariant, 354
Borel, 436 similarity distance, 226
countably generated, 436 square-sum, 282
472 SUBJECT INDEX

stability number, 46
stable set polynomial, 65
stationary walk, 439
stepfunction, 115, 442
stepping operator, 144
subgraph sampling, 5
support graph, 336

template graph, 8, 141

tensor network, 418
tensoring method, 240
test parameter, 263
test property, 266, 268
threshold graph
random, 193
threshold graphon, 187, 253
topology
local, 331
weak, 226
tree
spanning, 48
tree-decomposition, 107
tree-width, 107
Turán graph, 25
Tutte polynomial, 48
twin nodes, 39, 101
twin points, 219
twin reduction, 40

ultrafilter, 444
principal, 444
ultralimit, 445
ultrametric, 331
ultraproduct, 444
uniform attachment graph, 188
unlabeling, 86

Vapnik–Chervonenkis dimension,
VC-dimension, 446
variation distance, 12
Voronoi cell, 228

weak convergence, 139

zeta matrix, 433

 Notation Index

1, 1A indicator function, 37 δ (U, W ) cut distance, 132

d (U, W ) cut norm distance, 131
A Borel sets in the graph of graphs, 339 deg, degA , degG A degree, 331
A ≽ 0 positive semidefinite, 37 degc (φ, v) edge-colored degree, 417
a × b product in category, 447 deg(H) degree of root, 339
a ⊕ b coproduct in category, 448 dG (X, Y ) edge density between sets, 141
A · B dot product of matrices, 37 dHaus (A, B) Hausdorff distance, 208
α ∨ β pullback, 448 d• (H1 , H2 ) ball distance, 339
α ∧ β morphisms in pushout, 448 dimVC (H) Vapnik–Chervonenkis
αφ nodeweight of homomorphism, 56 dimension, 446
AG adjacency matrix, 39 (r,k)
δ⊙ (G1 , G2 ), δ⊙ nd (G , G )
1 2
α(G) stability number, 41 nondeterministic sampling distances,
A(G, x) characteristic polynomial, 67 360
αH nodeweight-sum, 58 δN (U, W ), δ1 (U, W ), δ2 (U, W ) distances
αi (H) nodeweight, 38 derived from norm, 135
A+ sigma-algebra of weighted rooted dob(W ) Dobrushin value, 370
graphs, 343 r (F, F ′ ), δ (F, F ′ ) sampling distance, 351
δ⊙ ⊙
Aut(W ) automorphism group, 234
δsamp (G, G′ ), δsamp (W, W ′ ) sampling
a1 W1 ⊕ a2 W2 ⊕ . . . direct sum of kernels,
distance, 12, 158
122
dsim (s, t) similarity distance, 277
BG,r (v) ball of radius r about v, 38 d◦ (u, v) local distance, 331
B(H, r) neighborhood of root, 339 dvar (α, β) variation distance, 12
βi,j (H) edgeweight, 38 dW (x) normalized degree, 116
B m m-bond, 38
E+ , E− signed edge-sets, 39
B m•• labeled bond, 39
e(G) = |E(G)| number of edges, 38
Br r-balls, 339
η, ηG edge measure, 333
C complex numbers, 37 eG (X, Y ) set of edges between X and Y , 38
Ca cyclic graphing, 330 ent(G, H) homomorphism entropy, 202
cep(G; u, v) cluster expansion polynomial, ent∗ (G, H) homomorphism entropy
434 typical, 204
chr(G, q) chromatic polynomial, 435 ent∗ (G, W ) sparse homomorphism entropy,
chr0 (G, k) number of colorations, 435 375
Cn cycle, 38 Eul eulerian property, 63
−
→
col(G, h) edge-coloring function, 417 eul(F ) number of eulerian orientations, 49
Conn(G) connected subgraphs, 38
cri(G) chromatic invariant, 435 f ↑ , f ↓ , f ⇓ Möbius inverses, 42
Csp(G) connected spanning subgraphs, 38 F1 F2 gluing product, 42
C(U, H), C(U, W ) overlay functional, 205 F † , F ‡ , F ♮ labeled quantum graphs from
cut(G, B) maximum multicut, 64 simple graph, 283
F ◦ G concatenation of graphs, 85
d1 (G, G′ ) edit distance, 128, 352 Fb signed graph from simple graph, 57
d (G, G′ ) labeled cut distance, 128, 129 F (K), Fn (K) compact decorated graphs,
δb (G, G′ ), δ (G, G′ ) unlabeled cut 318
distance, 129 Fksimp simple graphs on [k], 38
d (G, G′ , X) overlay cut distance, 129 Fk,m bi-labeled graphs, 85, 86

473
474 NOTATION INDEX

Fk• , FS• k-labeled and S-labeled graphs, 39 I(W ) induced subgraphs of graphon, 247
Fkmult multigraphs on [k], 38
[k] = {1, 2, . . . , k}, 12
Fkstab k-labeled graphs with stable labeled
K1•• 2-multilabeled graph on one node, 39
set, 95
K(a, b) morphisms, 447
flo(G, q) nowhere-zero q-flows, 436 • , K •• partially labeled complete
Ka,b
F ∗ conjugate in concatenation algebra, 85 a,b
bipartite graphs, 39
G1 G2 Cartesian sum, 40 Kain , Kaout morphisms into and out of a, 447
G1 × G2 categorical product, 40 χ(G) chromatic number, 41
G1  G2 strong product, 40 Kn complete graph, 38
◦ looped complete graph, 38
G1 ∗ G2 gluing k-broken graphs, 416 Kn
• , K •• partially labeled complete graphs,
G→ edge-rooted graphs, 340 Kn n
G• , G • rooted graphs, 339 39
G•F extensions of F , 339 Kn r complete hypergraph, 422

G(H) randomized weighted graph, 157

L(C) intersection graph, 38
G(k, G) random induced subgraph, 157
L(G) line-graph, 38, 331
G×k categorical power, 40
λi (G), λ′i (G), λi (W ), λ′i (W ) eigenvalues,
G(m) blow-up of G, 40
195
G(n, p), G(n, m), G(n; H) random graphs, 8
limω ultralimit, 445
G+ Bernoulli lift, 348
ln, log, log∗ natural, binary and iterated
G+ weighted rooted graphs, 343
logarithm, 37
Gpa
n growing preferential attachment graph,
Lrn complete r-partite hypergraph, 422
191
L(W ) space of functions txy , 311
Gpfx
n prefix attachment graph, 188
G/P quotient graph, 40, 83, 141, 211 Maxcut(G), maxcut(G) maximum cut, 64
GP averaged graph, 141 M (f, k), M simp (f, k), M mult (f, k),
G/ρ fractional quotient graph, 212 M flat (f, k), M (f, N) connection
G[S] induced subgraph, 38 matrices, 43
Gsimp simple version of G, 38 M ′ (f, k) edge-connection matrix, 416
G(n, W ), G(S, W ) W -random graphs, 157 M G Möbius inverse of ZG, 83
Gua
n uniform attachment graph, 188 A , M A , M A , M A matrices of
Mhom inj surj aut
[[G]] unlabeling, 39 homomorphism numbers, 73
[[G]]S removing labels not in S, 39 Mk (f ) moments of a function, 442
Gx connected component containing x, 338 µ(x, y), µP Möbius function, 433, 434
h(α) head of morphism, 447 N, N∗ nonnegative integers and positive
H(a, B)) weighted graph with nodeweight integers, 37
vector a and edgeweight matrix B, 38 N (f, k) dual connection matrix, 76
H graph of graphs, 339 ν(G) matching number, 41
H+ graph of weighted graphs, 343 NG (v) = N (v) neighbors of v, 38
hom(F, G) homomorphism number, 6, 56, (n)k = n(n − 1) . . . (n − k + 1), 37
77, 319 Nk (f ) annihilator of quantum graphs, 84
Hom(F, G) set of homomorphisms, 56 ∇(v) edges incident with v, 38
Hom(F, G) graph of homomorphisms, 79
Hom(F, G) complex of homomorphisms, 80 Ob(K) objects of category, 447
homφ (F, G) weight of homomorphism, 56, ω(G) size of maximum clique, 41
58 On edgeless graph, 38
hom∗ (G, H) typical homomorphism
number, 204 Pa path graphing, 330
hom(F, X), inj(F, X) homomorphism πG,H , πG,W , πy distribution on
polynomials, 108 homomorphisms, 368
H(n, W ), H(S, W ) W -random weighted Π(α) partitions of [0, 1], 205
graphs, 157 Π(n), Π(n, α) partitions of [n], 204
Π∗ (n, α) fractional partitions, 212
I(G) set of stable sets, 65 pm(G) number of perfect matchings, 41
ind(F, G) number of embeddings, 56 Pn path, 38
P • ••
inj(F, G) injective homomorphism number, ∏n , Pn partially labeled paths, 39
56, 57 ω ultraproduct, 444
I(W ) entropy functional, 260 P closure of graph property, 247
 NOTATION INDEX 475

P(T ) probability measures on Borel sets, W(K) K-graphons, 322

438 W n , W ◦n , W ⊗n powers of kernels, 123
W/P quotient graph, 144
Qq (G), Qa (G) quotient sets, 211 WP (x, y) stepping operator, 144
Q∗q (G), Q∗a (G) fractional quotient sets, 212
Qk k-labeled quantum graphs, 84 x ≥ 0 (for U ) nonnegativity of quantum
Qk /f , Qk,k /f factor algebras, 84 graphs, 281
Qstab
k k-labeled quantum graphs with stable x ≡ y (mod f ) congruence of quantum
labeled set, 95 graphs, 84
qr(H) quasirandomness, 422
Z, Zq integers and integers modulo q, 37
R, R+ reals and nonnegative reals, 37 ZG sum of quotients, 83
Rcε graphons far from R, 270
r(f, k) connection rank function, 44
ρG,r neighborhood sample distribution, 22
root(H) root of graph, 339
rW (a, b) similarity distance, 226
rW (x, y) neighborhood distance, 222

S [0,1] , S[0,1] measure preserving maps, 437

σG,k subgraph sample distribution, 5
Sn star, 38
σ + distribution of weighted graphs, 343
Spec(W ) spectrum, 124
σ ∗ (A) probability measure scaled by
degrees, 339
stab(G) number of stable sets, 46
stab(G, x) stable set polynomial, 65
surj(F, G) surjective homomorphism
number, 56

t(α) tail of morphism, 447

t(F, G), tinj (F, G), tind (F, G)
homomorphism densities, 58, 319
t(F, W ), tind (F, W ) homomorphism
densities in graphons, 116, 117
t(F, w) decorated homomorphism density,
120, 323
τ (G) node cover number, 46
τ (H) node-cover of hypergraph, 446
Tk tensors with k slots, 419
T (n, r) Turán graph, 25
tree(G) number of spanning trees, 435
t∗ (F, G), t∗inj (F, G), t∗ind (F, G)
homomorphism frequencies, 59
tut(G; x, y) Tutte polynomial, 434
TW kernel operator, 124
tx (F, W ) labeled homomorphism density,
117

U W, U ◦ W, U ⊗ W products of kernels, 123

U [X] submatrix of kernel, 160

v(G) = |V (G)| number of nodes, 38

W, W0 , W1 spaces of kernels and graphons,

115
f W
W, f0 , Wf1 unlabeled kernels, 132
W φ (x, y) variable transformation, 121
WH graphon from graph, 115