Ujhandout

Budapest Semesters in Mathematics
ADVANCED COMBINATORICS HANDOUTS
András Gyárfás
2011 FALL
INTRODUCTION
This handout is based on the Advanced Combinatorics course I have taught through
many years at Budapest Semesters. It contains a basic material covered in each semester
(sections 1-5) and two further parts (fruit salad and advanced menu) from which I have
selected material according to the taste and appetite of the students.
The objective of the course is to introduce hypergraphs together with widely appli-
cable proof methods of combinatorial mathematics.
The material is self contained, apart from basic facts of linear algebra. However,
at least one elementary course - for example counting methods, graph theory, discrete
structures - is strongly recommended.
Each section is supplemented with a set of easy questions (self-test) to check the
understanding of the material and with an exercise set.
At certain places the presentation of the material follows closely the following sources:
Babai, Frankl: Linear Algebra Methods in Combinatorics.
Graham, Rothchild, Spencer: Ramsey Theory.
Spencer: Ten Lectures on the Probability Method.
Van Lint, Wilson: A course in Combinatorics.
Acknowledgement. First of all I express my thanks to the active students of this
course through the years from 1988 to the present time. They understood my English,
read my handwritten notes, found misprints, caught errors, invented new proofs, some
of them even found new results that eventually deserve to be published (or already have
been published). The support of the organizers of the Budapest Semesters (in America
and in Budapest) is also acknowledged together with the help of my home institute,
SZTAKI (Computer and Automation Research Institute of the Hungarian Academy of
Sciences). Thanks to Péter Juhász for implementing the figures.
1
CONTENT
1. BASIC NOTIONS AND EXAMPLES
1.1. Definitions. (Incidence matrices, Duality.)
1.2. Examples of hypergraphs. (Paths and cycles, Linear spaces.)
2. CHROMATIC NUMBER AND GIRTH
2.1. Chromatic number.
2.2. Graphs from the Hall of Fame. (Graphs of Zykov, Mycielski, Tutte, Shift
graph, Kneser graph.)
2.3. How to glue hypergraphs to get graphs? ( Nesetril-Rödl hypergraph.)
3. A LOOK AT RAMSEY THEORY
3.1. Ramsey numbers. (Pigeonholes and parties, Existence of Ramsey numbers,

Upper bound on R(n), Many colors and non-diagonal Ramsey numbers, Exact values of
Ramsey numbers.)
3.2. Van der Waerden numbers.
3.3. Tic-tac-toe and Hales-Jewett theorem.
2
4. COUNTING AND PROBABILITY
4.1. Proofs by counting. (Antichains, intersecting hypergraphs, 3-chromatic uniform

hypergraphs.)
4.2. Probability method. (The Erdős lower bound on R(n), Lower bound for W (k),
Tournaments, Large transitive subtournaments, Existence versus construction, Paradox-
ical tournaments, Hamiltonian paths.)
4.3. Local Lemma. (Erdős-Lovász theorem, Improved lower bound on W (k), Even
cycles in regular digraphs.)
4.4. Jokes (Triangle is 2-chromatic, Spencer’s injections.)
5. LINEAR ALGEBRA METHODS
5.1. The dimension bound. (Oddtown, Fisher inequality, A cubic lower bound for
R(n), Two-distance sets, Cross-intersecting hypergraphs.)
5.2. Homogeneous linear equations. (Partition of Kn , discrepancy.)
5.3. Eigenvalues. (Regular graphs of girth five.)
6. FRUIT SALAD
Distinct representatives, symmetric chain decomposition, properties of hypergraphs with
n vertices and n, n+1, n+2 edges, critical not 2-colorable hypergraphs, sunflower theorem,
sum-free subsets of numbers.
7. ADVANCED MENU
Baranyai theorem, normal hypergraphs and perfect graphs, constructive super polynomial
lower bound for Ramsey numbers, Bursuk conjecture.
3
1 BASIC NOTIONS AND EXAMPLES
1.1 DEFINITIONS
Hypergraph: H = (V, E) where V is a set and E is a collection of subsets of V . The
set V is called the vertex set and E is called the edge set of the hypergraph H . The
word “collection” is used to allow selection of the same subset of V more than once, if
this happens we have multiedges.
Simple hypergraph: no multiedges.
What can be empty? To be precise, there are some details to discuss. Do we allow
that V = ∅ ? Frank Harary has a paper with the pun title: “Is the null-graph a pointless
concept?” Do we allow E = ∅ ? What if V, E are nonempty but e = ∅ for some e ∈ E
? We shall assume that V and E are nonempty but allow empty edges for technical
reasons.
Singletons, isolated vertices. An edge e ∈ E is called singleton if |e| = 1. A vertex
is isolated if no edge contains it. We allow singleton edges and isolated vertices (unless
stated otherwise).
Finite or infinite? In most cases we shall work with finite hypergraphs. However,
sometimes it is natural to consider hypergraphs with |V | = ∞, for example V = Rd , but
usually we shall still assume |E| < ∞.
Degree of v ∈ V : The number of edges containing vertex v (in a hypergraph H = (V, E)
). Notation for the degree: d(v) or dH (v) if not clear from the context.
Maximum and Minimum degree: ∆(H) denotes the largest and δ(H) denotes the
smallest among the degrees of the vertices of H.
Regular hypergraph: d(x) = d(y) for all x, y ∈ V .
t-regular hypergraph: regular hypergraph with common degree t.
Uniform hypergraph: |e| = |f | for all e, f ∈ E.
t-uniform hypergraph: uniform hypergraph with edges of t vertices.
Graphs can be defined as 2-uniform hypergraphs, simple graphs are the simple 2-uniform
hypergraphs. (A singleton edge is similar to the notion of the loop in Graph theory but
a loop contributes 2 and a singleton contributes 1 to the degree of a vertex.) There are
4
some terms used only for graphs. The most important is the adjacency: two vertices
x, y of a graph are called adjacent if {x, y} is an edge of the graph. The complementary
graph (or complement) of a simple graph G = (V, E) is the graph G = (V, E) where E
is the set of pairs of V not in E. Complements of simple uniform hypergraphs can be
defined in a similar way.
Isomorphic hypergraphs: Two hypergraphs are isomorphic if there exists an edge
preserving bijection between their vertex sets.
How to draw a hypergraph? One possibility to represent a hypergraph is to draw
it. In case of graphs this is very usual. It is more complicated to draw a hypergraph.
Figures 1.1. and 1.2. Drawings of hypergraphs
On Figure 1.1 the “potato” represents an edge with four vertices, the other nine
edges have two vertices. Figure 1.2 is a drawing of the so called Fano plane, a famous
3-regular, 3-uniform hypergraph on seven vertices. In that case the edges are represented
with lines with one exception, the circle also represents an edge.
The incidence matrix. A hypergraph H = (V, E) can be represented by its incidence
matrix, defined as follows. Assume that V = {v1 , . . . vn } and E = {e1 , . . . em }. The
incidence matrix of H is the m × n matrix [aij ] where
(
1 if vj ∈ ei
aij =
0 if vj ∈
/ ei
Remember that rows correspond to edges and columns correspond to vertices in the
incidence matrix of the hypergraph. (Of course, this is merely a convention.)
Hypergraphs are rather general structures but there is a theorem valid for all hyper-
graphs.
5
Theorem 1.1 For any hypergraph H = (V, E)
X X
d(x) = |e|
x∈V e∈E
Proof. Interpret the statement for the incidence matrix of H. ¤
Duality. Assume that H is a hypergraph with incidence matrix M . Let M T be the

transpose of M (M T [aij ] := M [aji ]). The hypergraph H∗ with incidence matrix M T
is called the dual of H. The definition obviously shows that (H∗ )∗ = H. There is a
bijection f from the vertices of H to the edges of H∗ and a bijection g from the edges
of E to the vertices of H∗ . There is a dictionary of duality to translate concepts and
theorems. Some lines from this dictionary might look like this:
H H∗
x∈e f (x) 3 g(e)

d(x) |f (x)|
|e| d(g(e))
t-regular t-uniform
t-uniform t-regular
Intersection graph (alias line graph) of a hypergraph H = (V, E) with E = {e1 , . . . , em }
is a graph with vertices x1 , . . . , xm so that {xi xj } is an edge of the graph if and only if
ei ∩ ej 6= ∅. The intersection graph reflects the intersections only between pairs of edges
of a hypergraph, the finer structure of the hypergraph is lost.
Theorem 1.2 Every simple graph is the intersection graph of some hypergraph.
Proof. Exercise 1.1 ¤
1.2 EXAMPLES OF HYPERGRAPHS

Paths and cycles. Assume that k ≥ 2 is an integer. A path in a hypergraph H = (V, E)
is a sequence x1 , e1 , x2 , e2 , . . . , xk−1 , ek−1 , xk where the xi -s are distinct vertices and the
6
ei -s are distinct edges and xi , xi+1 ∈ ei for 1 ≤ i ≤ k − 1. For convenience, a vertex is
considered also as a path (this corresponds to the case k = 1). The length of the path is
k − 1, the number of edges in it.
The definition of a cycle is very similar, it is more symmetric: add ek to the end of
a sequence forming a path, requiring x1 , xk ∈ ek . More precisely, for k ≥ 2, a cycle is
a sequence x1 , e1 , x2 , e2 , . . . , xk−1 , ek−1 , xk , ek where the xi -s are distinct vertices and the
ei -s are distinct edges and xi , xi+1 ∈ ei for 1 ≤ i ≤ k − 1 and x1 , xk ∈ ek . The length
of the cycle is k, the number of vertices (or the number of edges). An important special
case is the cycle of length two (the first object on Figure 1.3).
Paths and cycles can form rather complicated structures in hypergraphs. In the
special case of graphs, paths and cycles are determined by just giving the sequence of
their vertices. For graphs, paths ( and cycles) with k vertices are isomorphic, a rather
standard notation is to use Pk for paths and Ck for cycles (with k vertices). Figure 1.3
gives the drawing of some paths and cycles.
Figure 1.3. Paths and Cycles
The complete uniform hypergraph Knr is the hypergraph with n vertices and with
¡ ¢
all distinct r-element subsets as edges. Therefore Knr has nr edges. We always assume
1 ≤ r ≤ n and (as usual) abbreviate Kn2 as Kn .
A k-partite hypergraph is a hypergraph H = (V, E) such that V is partitioned into k
non-empty sets, called partite classes and every edge e ∈ E intersects each partite class
in at most one vertex. Notice that every hypergraph is k-partite for some k since the
vertex set can be partitioned into one element partite classes. The 2-partite graphs are
called bipartite graphs.
7
Intersecting hypergraphs are hypergraphs in which any pair of edges have a non-
empty intersection. For example, the Fano plane (Figure 1.2) is an intersecting hyper-
graph.
Steiner systems. Let t, k, n be integers satisfying 2 ≤ t ≤ k < n. A Steiner system
S(t, k, n) is a k-uniform hypergraph H = (V, E) with n vertices such that for each t-
element set T ⊂ V there is exacly one edge e ∈ E satisfying T ⊆ e. The complete graph
Kn is the same as a S(2, 2, n) Steiner system. The Steiner systems S(2, 3, n) are called
Steiner triple systems . The Fano plane is an example of a Steiner triple system on 7
vertices. A trivial example of a Steiner system with t = 3 is K43 , the smallest interesting
one is S(3, 4, 8) (Exercise 1.4).
Linear spaces. These are hypergraphs in which each pair of distinct vertices is con-
tained in precisely one edge. To exclude trivial cases, it is always assumed that there
are no empty or singleton edges.
The rest of the section gives examples and results about linear spaces. An obvious
example is a hypergraph with only one edge which contains all vertices, this is called a
trivial linear space . Another example is the so called near pencil with one ’big’ edge
which contains all but one vertices (the other edges have two vertices, see Figure 1.4).
Another straightforward example is the complete graph Kn . A more complicated one,
the affine plane of order 3 is shown in Figure 1.5. This example (like the Fano plane)
is a special case of Steiner triple systems S(2, 3, n). Observe that uniform linear spaces
are precisely the Steiner systems S(2, k, n).
8
Figures 1.4. and 1.5. The near pencil and the affine plane of order 3
Drawings of linear spaces. One can try to represent a linear space by drawing its
vertices as points of the plane so that edges are precisely the lines determined by these
points. If there are at most seven vertices then such a drawing is possible apart from
one exception, the Fano plane. Such a drawing is more clear if only the lines having at
least three points are shown. Figure 1.6 gives the drawings of linear spaces with at most
five vertices according to this convention.
Figure 1.6. Drawings of linear spaces up to five vertices
After observing the examples, it seems that (apart from the trivial one) a linear space
must have at least as many edges as vertices.
Theorem 1.3 If a non-trivial linear space has n vertices and m edges then m ≥ n (de
Bruijn - Erdős).
9
Proof. (Tricky proof, due to Conway.) Let H = (V, E) be a linear space, |V | = n,
|E| = m and suppose 1 < m ≤ n. Select v ∈ V , e ∈ E such that v ∈
/ e. Observe that
this implies d(v) ≥ |e| which (together with m ≤ n ) gives
1 1
≥
n(m − d(v)) m(n − |e|)
Adding these inequalities for all pairs v ∈
/ e one gets
XX 1 XX 1
1= ≥ =1
v∈V e63v
n(m − d(v)) e∈E
m(n − |e|)
v ∈e
/
because the inner sums are never empty (1 < m) and the terms in them do not
depend on e and v (and have m − d(v) resp. n − |e| terms). Therefore all inequalities
must be equalities, in particular m = n. ¤
The next result shows that intersecting linear spaces are rather restrictive.
Theorem 1.4 Assume H = (V, E) is an intersecting linear space not isomorphic to a

trivial one or to a near pencil. Then (for some integer k ≥ 2)
|E| = |V | = k 2 + k + 1, H is (k + 1)-uniform and (k + 1)-regular . (1)
Proof. (Outline.) If the union of at most two edges cover the vertex set then H is
trivial or a near pencil. Otherwise for e, f ∈ E select p ∈ V \ (e ∪ f ). For x ∈ e, let g(x)
be a vertex of f such that x, p, g(x) are in an edge of H. Then g is a bijection therefore
|e| = |f | := k + 1. This shows that H is (k + 1)-uniform. Then (k + 1)-regularity,
|V | = k 2 + k + 1 and |E| = k 2 + k + 1 follows easily. ¤
Intersecting linear spaces with property (1) are called finite planes of order k . Their
existence is known only if k is a prime power. The following important result of Bruck
and Ryser rules out the existence of finite planes for certain values of k: if a finite plane
of order k exists for k ≡ 1 or k ≡ 2 (mod 4) then k = x2 + y 2 has integer solution.
Apart from these results, only the nonexistence of a finite plane of order 10 is known.
The finite planes of order at most eight are unique but this is not true for order nine.
10
Self-test 1.
[T1.1] Show that all 0 − 1 matrices are incidence matrices of hypergraphs. What about
zero rows or columns?
[T1.2] Find the smallest k for which the Fano plane is a k-partite hypergraph. Answer
the same question for the affine plane of order 3.
[T1.3] Is it obvious that the dual of a finite plane is a finite plane?
[T1.4] Is it obvious that the dual of a finite plane is isomorphic to itself?
[T1.5] Formulate Theorem 1.3 in dual form!
[T1.6] What is the intersection graph of the Fano plane?
[T1.7] What is the intersection graph of the affine plane of order 3 ?
[T1.8] Formulate Theorem 1.1 for graphs!
[T1.9] Find how to get the finite plane of order 3 from the affine plane of order 3 (Figure
1.5).
Exercise set 1.
[1.1] Prove by induction that every simple graph is the intersection graph of some hy-
pergraph.
[1.2] Prove the statement in [1.1] by using the notion of duality!
[1.3] Work out the details of the proof of Theorem 1.4
[1.4] Construct S(3, 4, 8) starting from the Fano plane. Explain why does it work!
[1.5] The Fano plane has cyclic representation: shift the set {1, 2, 4} (by adding 1 to
its elements) six times, using arithmetic (mod 7). Find similar representation for the
finite plane of order 3 and order 4. What property of q + 1 positive integers ensures that
they give a cyclic representation of a finite plane of order q?
[1.6] Show that if a Steiner triple system S(2, 3, n) exists then n ≡ 1 or n ≡ 3 (mod 6).
[1.7] Find two divisibility conditions for the existence of S(2, k, n).
[1.8] Find t divisibility conditions for the existence of S(t, k, n).
[1.9] Prove or give counterexample for the following two statements. Regular linear
spaces are uniform. Uniform linear spaces are regular.
[1.10] Give a catalogue of linear spaces with six vertices. (Follow the convention of Figure
1.6.)
11
2 CHROMATIC NUMBER AND GIRTH
2.1 CHROMATIC NUMBER

A proper coloring of a hypergraph is a coloring of the vertices so that there are
no monochromatic edges, i.e. each edge of the hypergraph receives at least two distinct
colors. Observe that it is impossible to find a proper vertex coloring if the hypergraph has
a singleton edge. Therefore in this section we shall exclude singleton edges. Of course,
instead of real colors, we shall use positive integers most of the time. However, it is also
traditional to use red and blue for 2-colorings. It is clear that one needs at least two colors
for a proper coloring of a hypergraph. The hypergraphs which have proper 2-colorings
are called 2-colorable hypergraphs. These hypergraphs are very important and we shall
see some of their properties later. The 2-colorable graphs are called bipartite graphs , they
have a well-known characterization (see Excercise 2.1). However, no characterization is
known for 2-colorable hypergraphs.
The chromatic number χ(H) of a hypergraph H is the minimum number of colors
needed for a proper vertex coloring of H.
Greedy coloring algorithm. The simplest way to find a proper coloring of a hyper-
graph H = (V, E) is to order V in an arbitrary way and color the vertices in this order
by assigning the smallest positive integer to the current vertex so that it does not create
a (completely) monochromatic edge.
Theorem 2.1 Every hypergraph H = (V, E) satisfies
χ(H) ≤ ∆(H) + 1
The girth of a hypergraph H = (V, E) , g(H), is the minimum among all cycle lengths
taken over all possible cycles of H. If H has no cycles, we define g(H) = ∞. Observe
that g(H) ≥ 2 and g(H) = 2 if and only if E has two edges intersecting in at least two
vertices.
12
We show later in this section that there are hypergraphs with large girth and large
chromatic number. The existence of such hypergraphs have been proved first by P.Erdős
and A.Hajnal with the probability method, the first explicit construction was given by
L.Lovász. First we look at the case of graphs.
2.2 GRAPHS FROM THE HALL OF FAME

The easiest example of a graph of chromatic number n is the complete graph Kn , it has
girth 3. Can we find an n-chromatic graph G with girth at least 4? This means that G
has no K3 subgraph (triangle). Some famous constructions are shown below.
Zykov graphs. Set Z1 = K1 and construct recursively Zi+1 by taking i disjoint copies
of Zi and taking a disjoint vertex set Ai of |V (Zi )|i elements. Each vertex of Ai is made
adjacent to the vertices of the vector [x1 , x2 , . . . , xi ] where xj is a vertex of the j-th copy
of Zi . Distinct elements of Ai are made adjacent to distinct vectors.
Figure 2.1. Zykov graphs.
Theorem 2.2 The Zykov graph Zi has no triangle and its chromatic number is i.
Proof. First we prove by induction that Zi has no triangle. This is obvious for Z2 ,
assume that true for Zi . Select three vertices from Zi+1 and assume they form a triangle
T . Then, from the construction, no two vertices of T are in Ai+1 and no two vertices of
T are from different copies of Zi and if two vertices of T are in the same copy of Zi then
the third is not from Ai+1 . Therefore the only possibility is that all three vertices of T
are from the same copy of Zi which contradicts the inductive hypothesis.
13
Next we prove that χ(Zi ) = i, it is clear for i = 2. If true for some i, then a proper
coloring of Zi+1 with i + 1 colors can be obtained by taking the same proper i-coloring
on all copies of Zi and coloring A with a new color. But perhaps there is a tricky proper
coloring of Zi+1 with i colors...
Assume there is such a tricky proper coloring of Zi+1 with colors 1, 2, . . . , i. Since
χ(Zi ) = i (by induction), the tricky coloring must use all colors in each copy of Zi . In
particular, there is a vector [x1 , x2 , . . . , xi ] such that for all j, xj is a vertex of the j-th
copy of Zi and xj is colored with color j. There exists a vertex a ∈ Ai adjacent with the
vertices of this vector! (Contradiction). ¤
Mycielski graphs.
This construction is more economic than the Zykov construction. Set M1 = K1 ,
M2 = K2 . To define Mi+1 , let Gi be a copy of Mi . For each vertex v of Gi define a twin
vertex v ∗ which is adjacent to the same set of vertices as v in Gi . (The set of twins are
distinct and disjoint from the vertex set of Gi .) Finally, add a new vertex wi adjacent
to all twin vertices.
Figure 2.2. Mycielski graphs.
Theorem 2.3 The Mycielski graph Mi has no triangles and χ(Mi ) = i.
14
Tutte graphs.
The Tutte graph Ti will be a graph of chromatic number i with girth 6, so Ti is
not only without triangles but without cycles of length four and five. Like the previous
constructions, we start with T1 = K1 and proceed recursively. However, the number of
copies of Ti needed to construct Ti+1 will be enormous...
¡(ni −1)i+1¢
Assume that Ti is already constructed and has ni vertices. Take mi = ni
disjoint copies of Ti , the reason will be explained soon. These copies will be glued
ni
together along the hyperedges of a complete ni -uniform hypergraph Hi = K(n i −1)i+1
whose vertices are disjoint from all copies of Ti . The gluing is done as follows. Due to
the definitions, we can make a one-to-one onto correspondence between the copies of Ti
and the hyperedges of Hi . Assume that the hyperedge ej corresponds to the j-th copy
of Ti (j = 1, 2, . . . , mi ). For every fixed j, define an injection fj from the vertices of ej
to the vertices of the j-th copy of Ti . The new graph edges (in addition to the edges
inside the copies of Ti ) are the pairs {x, fj (x)} where x runs over the vertices of ej and
j runs over {1, 2, . . . , mi }.
Figure 2.3. Tutte graphs.
Theorem 2.4 The Tutte graph Ti is of girth 6 and χ(Ti ) = i.
Proof. Since the vertices of the gluing hypergraph Hi can be colored with a new color
and all copies of Ti can be colored the same way by i colors, χ(Ti ) ≤ i comes easily
by induction. The problem is to show χ(Ti ) ≥ i. This is also done by induction, of
course there is no difficulty to launch it. Assume that we already know this for some i.
Assume indirectly that Ti+1 has a proper coloring with i colors. Then, by the inductive
hypothesis, all copies of Ti in Ti+1 are colored using all the i colors.
15
The vertices of the gluing hypergraph Hi are also colored and by the pigeonhole
principle, there exists a hyperedge, say ej , whose vertices are all colored with the same
color, say color 1! (Check this carefully to enjoy the beauty of Tutte’s idea.) Now we
have a contradiction because the j-th copy of Ti has a vertex y colored with color 1 so
both vertices of the edge {fj−1 (y), y} are colored with color 1.
There is still a nice part ahead: imagine that you walk on a cycle of Ti+1 . Look around
carefully while you complete your tour on the cycle. It can certainly happen that your
tour is completely inside of a copy of Ti . Then, by induction you are convinced that
you visited at least six vertices. Watch your step when you leave a copy, you are in the
gluing hypergraph. Where to go? You are forced to go into a new copy of Ti from where
you can not escape immediately. Therefore you visit at least two vertices in at least two
copies and at least two vertices of the gluing hypergraph, a total of at least six vertices.¤
The Shift graph (of Erdős and Hajnal).
The new feature of this graph is that it is defined without recursion. The vertex set
¡ ¢
of the shift graph SHn is the set of pairs (i, j), 1 ≤ i < j ≤ n so it has n2 vertices.
What about the edges? The vertices (p, q), (r, s) are adjacent if and only if q = r or
p = s.
Figure 2.4. The Shift graph SH5 .
Theorem 2.5 The Shift graph SHn has no triangles and χ(SHn ) = dlog2 (n)e.
Proof. The triangle free property is left as an (easy) exercise. To see that χ(SHn ) ≤
n
dlog2 (n)e, color the vertex (p, q) with color 1 if 1 ≤ p ≤ 2
< q ≤ n. The vertices of SHn
are partitioned into A1 , A2 as follows.
n n
A1 = {(i, j) : 1 ≤ i < j ≤ }, A2 = {(i, j) : < i < j ≤ n}
2 2
16
Observe that x ∈ A1 and y ∈ A2 are not adjacent and A1 , A2 define shift graphs,
isomorphic to SHb n2 c and to SHd n2 e , respectively. Continuing this process, a proper
coloring of SHn is obtained with dlog2 (n)e colors.
To prove the reverse inequality, χ(SHn ) ≥ dlog2 (n)e, assume that the vertices of
SHn are properly colored with t colors. For each i < n, let Si denote the set of vertices
{(i, j) : j = i + 1, i + 2, . . . , n}. We claim that the sets Sk and Sl can not be colored
with the same color set if k 6= l. Indeed, without loss of generality, k < l and the color
of (k, l) is red. Then red is a color used on Sk but red can not be used on Sl because
all vertices of Sl are adjacent to the vertex (k, l). Now the claim is proved and it gives
that n − 1, the number of sets Si is not larger than the number of nonempty subsets of
t colors. It follows that n − 1 ≤ 2t − 1 i.e. n ≤ 2t which implies dlog2 (n)e ≤ t. ¤
Kneser graphs. In 1955 Kneser posed the following problem. For fixed positive
n
integers n, k, consider the hypergraph K2n+k , i.e. all n element subsets of a ground set
of 2n + k elements. Kneser was interested in partitioning the hyperedges into as few
classes as possible so that hyperedges in the same partition class are pairwise intersecting
and conjectured that this minimum is k + 2. One can reformulate this as a problem
to determine the chromatic number (of a very symmetric) graph, the Kneser graph,
n
KN (n, k) whose vertices correspond to the edges of K2n+k , two vertices being adjacent
if and only if the corresponding hyperedges are disjoint. Then the conjecture is that the
chromatic number of this graph is k + 2. It is easy to see that k + 2 is an upper bound
(Exercise 2.10), the problem was whether one can do better. Kneser’s conjecture was
proved by Lovász in 1977 and Bárány simplified it very soon thereafter (but Lovász’s
proof gives more). The main surprise (of both proofs) is the method applied.
Theorem 2.6 (Lovász) It is impossible to partition the n-element subsets of a 2n + k

-element ground set into k + 1 classes so that each class has pairwise intersecting sets.
Proof. Bárány’s short proof uses two ingredients (the first was used in Lovász’s proof
as well):
Borsuk theorem (1933). Assume that surface of the unit sphere, S k = {x ∈ Rk+1 , ||x|| =
1} is covered by the union of k + 1 open sets. Then some of these sets has two antipodal
points.
17
Gale theorem (1956). For positive integers k, n it is possible to place 2n + k points on S k
so that each open hemisphere contains at least n of them. (An open hemisphere H(a)
with center a ∈ S k is the set {x ∈ S k } for which the inner product xa is positive.)
This is the way to put together the two ingredients. Assume that the n-element
subsets of a 2n + k-element are partitioned into k + 1 classes so that each class contains
pairwise intersecting sets. Place the vertices of the ground set on S k as described by
Gale theorem. Define the sets U1 , U2 , . . . , Uk , Uk+1 as follows: let Ui be the set of points
x ∈ S k for which the open hemisphere H(x) contains an n-element set from the i-th
partition class. By Gale theorem, the union of the sets Ui cover S k . The sets Ui are all
open sets, Borsuk theorem says that in some of them, say in Uj , there are two antipodal
points. These points are centers of disjoint open hemispheres, by definition both contain
an n-element set from the j-th partition class - contradiction. ¤
Josh Green, a former BSM student simplified the proof even further by eliminating
Gale’s theorem from the proof! His modification (appeared in 2002 in the American
Mathematical Monthly) is as follows. Instead of placing the vertices of the ground set
on S k by Gale’s theorem, he just placed them on S k+1 as points in general position!
General position here means no k + 2 points are on a great k-sphere, S k ⊂ S k+1 . The
sets U1 , U2 , . . . , Uk , Uk+1 are defined similarly: Ui is the set of points x ∈ S k+1 for which
the open hemisphere H(x) contains an n-element set from the i-th partition class. The
k + 2 sets
k+1
[
k+1
U1 , . . . , Uk+1 , F = S \ Ui
i=1
cover S k+1 . Borsuk’s theorem remains true if open and closed sets are mixed in the
cover (one of the sets, F , is closed). Thus some of these sets contain an antipodal pair
a, −a. This pair can not be in F because the great circle separating the hemispheres
H(a), H(−a) contains at most k+1 points so one of H(a), H(−a) contain at least n
points so a or −a would be in some Ui . Thus - as in the original proof - we have a pair
of antipodal points in some Ui and this contradiction finishes the proof.
18
2.3 HOW TO GLUE HYPERGRAPHS TO GET GRAPHS?
Recall that Tutte graphs have arbitrary large chromatic number and girth six. No simple
construction is known if graphs of girth seven are required with large chromatic number,
although their existence was proved by Erdős in 1961 using his probability method.
The first construction of such graphs was found by Lovász in 1968. Interestingly, his
construction is based on hypergraphs. The following construction is somewhat simpler
(?!). It is also based on hypergraphs.
The Nesetril-Rödl construction.
Assume that three positive integers, p, k, n are given, k, n ≥ 2. Our objective is to
construct a hypergraph H with the following properties: 1. H has no cycles of length at
most p (the girth of H is at least p+1) . 2. H is k-uniform . 3. χ(H) ≥ n. A hypergraph
H with properties 1,2,3 is called a [p, k, n]-hypergraph. Observe that we have already
seen [3, 2, n]-hypergraphs, i.e. graphs without triangles and chromatic number at least
n (Zykov graphs, Mycielski graphs) and [5, 2, n]-hypergraphs, i.e. graphs with girth six
and chromatic number at least n.
The construction of [p, k, n]-hypergraphs is recursive. The first step is to construct a
[1, k, n]-hypergraph. Note that in this case we have no restrictions about cycles because,
by the definition of cycles, every cycle has length at least two. We define (in terms of k
and n ) a number s as follows.
s := (k − 1)(n − 1) + 1.
What is the significance of s? We plan to construct an s-partite [p, k, n]-hypergraph.

Why do we make our life more difficult by imposing further restrictions? The answer is
the same as in many other cases in mathematics: to prove a stronger statement is easier.
Now we are ready to start at p = 1.
Base step: the [1, k, n]-hypergraph. The complete hypergraph H = Ksk is an s-partite
[1, k, n]-hypergraph. The proof of this statement is immediate from the pigeonhole prin-
ciple. The only nontrivial statement is that the chromatic number of H is at least n.
Assume that there is a proper coloring of the vertices of H with n − 1 colors. From the
definition of s there are k vertices of H colored with the same color i.e. the edge of H
19
determined by those k vertices is monochromatic. This contradicts the definition of a
proper coloring.
Recursive step. Assume that [p − 1, K, n]-hypergraphs are already constructed. Unfor-
tunately, there is no misprint here! Although we wish to construct an s-partite [p, k, n]-
hypergraph, we shall use [p − 1, K, n]-hypergraphs with values of K much larger than
k. (This means that graphs (k = 2) can not be constructed without using hyper-
graphs already constructed.) Notice also that s depends on k therefore the [p − 1, K, n]-
hypergraphs we shall use are S-partite where S is not the same as s. It might be a
consolation than n never changes.
In fact, we shall build hypergraphs H1 , . . . , Hs+1 , each of them will be s-partite and
k-uniform. The last one, Hs+1 will be the final aim, the s-partite [p, k, n]-hypergraph.
The definition of H1 . It will be refreshing to see that H1 is not defined recursively.
Moreover, H1 will have pairwise disjoint edges. After these promises, look at the defini-
tion.
The vertex set V of H1 is partitioned into sets X1 , X2 , . . . , Xs where each set Xi
¡s−1¢
has the same number of vertices, namely k−1 . This allows to define pairwise disjoint
¡¢
k-element edges so that for all choices of k partite sets (there are ks choices) there is
exactly one edge which has vertices from the chosen partite classes (one vertex from
each). Figure 2.5 illustrates H1 for k = 3, n = 3 in that case s = (3 − 1)(3 − 1) + 1 = 5,
¡¢
|Xi | = 42 = 6.
20
Figure 2.5. H1
The definition of Hj+1 (1 ≤ j ≤ s). Assume that Hj is already constructed for some
j, 1 ≤ j ≤ s with partite classes X1 , . . . , Xs .
Set K = |Xj |.
Let Fj be a [p − 1, K, n]-hypergraph with vertex set Vj and with edge set Ej . The set
Vj is called the gluing base and the edges in Ej are called the gluing edges because they
serve to keep together many copies of Hj . The definition of Hj+1 is as follows. Take
as many copies of Hj as the number of gluing edges and, for each gluing edge, identify
the j-th partite class of a copy with the gluing edge. For distinct gluing edges distinct
copies of Hj are used. Apart from their j-th classes, all partite classes of the copies are
kept disjoint. The result is an s-partite hypergraph whose t-th partite class is the union
of the t-th partite classes of the copies used, except for t = j when the t-th partite class
is Vj . This finishes the (recursive) definition of Hj which gives Hs+1 in the last step.
Theorem 2.7 The hypergraph Hs+1 is an s-partite [p, k, n]-hypergraph.
21
Figure 2.6. To get Hj+1 from copies of Hj
Proof. The problem is to understand the construction, after that the proof is not
difficult. Only two statements need a proof. The first is that there are no cycles up to
length p. This comes from an argument similar to the one used for Tutte graphs to show
that their smallest cycle is of length at least six. Of course, it is a bit more difficult to
travel through a cycle of a hypergraph than through a cycle of a graph... Prove first for
H1 (trivial) then use induction to prove it for Hs+1 . (Exercise 2.11 asks to elaborate the
details.)
The second statement is that it is impossible to color properly the vertices of Hs+1
with n − 1 colors. Assume there is such a coloring C. Then C colors the vertices of Xs
which is the gluing base of Fs which is at least n-chromatic so there exists a gluing edge in
it monochromatic under C. This gluing edge serves as Xs , the last partite class in a copy
of Hs . Therefore we can select a copy of Hs whose last partite class is monochromatic.
Repeat the argument, C colors the (s − 1)-th partite class, Xs−1 of the selected copy
of Hs which is the gluing base of Fs−1 ... Going back like this, we find a copy of H1 in
which all partite classes are monochromatic under C. And now the choice of s (and the
contradiction) becomes (hopefully) clear (Exercise 2.12). ¤
22
Self-test 2.
[T2.1] What is the chromatic number of Kn ?
[T2.2] What is the chromatic number of Knr ?
[T2.3] Show that the Fano plane and the affine plane of order 3 are 3-chromatic hyper-
graphs.
[T2.4] Find a 3-chromatic graph with girth g.
[T2.5] Give some hypergraphs (graphs) with equality in Theorem 2.1.
[T2.6] Show that (with suitably chosen parameters) the Kneser graph has no triangle
and its chromatic number is arbitrary large.
Exercise set 2.
[2.1] Show that a graph is bipartite if and only if it does not have cycles of odd length.
[2.2] Prove that the Greedy algorithm colors every hypergraph H = (V, E) properly with
at most ∆(H) + 1 colors!
[2.3] Give an example of a bipartite graph G such that for a certain ordering of its
vertices the Greedy algorithm uses 2000 colors for the proper coloring of G.
[2.4] Use the greedy algorithm to prove: if H = (V, E) is a hypergraph (no singleton
edges) such that for all pairs of distinct edges e, f ∈ E we have |e∩f | =
6 1 then χ(H) = 2.
[2.5] Prove that a finite plane of order at least 3 is 2-colorable!
[2.6] Prove that Steiner triple systems are not 2-colorable!
[2.7] Modify the definition of the Zykov graph Zn+1 so that instead of n copies of Zn one
copy of Zi is used for 1 ≤ i ≤ n. Show that it works (no triangles, and it has chromatic
number n + 1).
[2.8] Prove that the Mycielski graph Mn has no triangle and has chromatic number n.
Hint: one possible proof uses the following lemma. Assume that G is an n-chromatic
graph and consider a proper coloring of G with n colors, 1, 2, . . . , n. Then, for each color
i, there exists a vertex v colored with color i such that v is adjacent to at least one vertex
of color j for all j 6= i.
[2.9] Prove that there are no triangles in the Shift graphs.
[2.10] Prove that the chromatic number of the Kneser graph KN (n, k) is at most k + 2.
[2.11] Show that Hs+1 has no cycles of length 2, 3, . . . , p (hint is in the text).
[2.12] Show that the chromatic number of Hs+1 is at least n (outline is in the text).
23
3 A LOOK AT RAMSEY THEORY
3.1 RAMSEY NUMBERS

Pigeonholes and Parties. According to the Pigeonhole Principle, if 2n − 1 persons
are present at a party then either there are n men or n women. The same statement is
not necessarily true if only 2n − 2 persons are present. Of course, similar statement is
true for any 2-partition of the elements of a ground set of 2n − 1 elements. The work of
Ramsey (in 1932) extended the pigeonhole principle by introducing 2-partitions of the
pairs of a ground set. Let R(n) be the smallest integer m with the following property: if
the (unordered) pairs of a m-element ground set are partitioned into two classes in any
¡ ¢
fashion then there exists n elements such that all of the n2 pairs determined by them
are in the same class. It is not obvious that R(n) is well defined but it will be clear soon.
The most famous value is R(3) = 6, in terms of parties and with a symmetric relation
“to know each other” it is usually formulated as follows.
Theorem 3.1 If six persons are present at a party then either three of them pairwise
know each other or three of them pairwise do not know each other. This is not necessarily
true for five persons.
Proof. Select an arbitrary person P . Then there are three other persons P1 , P2 , P3 such
that P either knows all of them or none of them. In the former case, either P, Pi , Pj
pairwise know each other or P1 , P2 , P3 pairwise do not know each other. The latter case
is similar.
The graph C5 (vertices represent persons and edges represent the pairs who know
each other) show the second part of the theorem. ¤
Existence of Ramsey numbers. It is usual to consider partitions as colorings, in case

of two colors red and blue are used most frequently. The objects to be colored can be
t-element subsets of a set. Then, in terms of hypergraphs, Ramsey’s existence theorem
can be formulated as follows.
24
Theorem 3.2 Assume that n, t are integers satisfying n ≥ t. There exists an integer m
t
(depending on n and t) with the following property: if the edges of Km are colored with
red and blue in any fashion then there exists a monochromatic Knt , i.e. n vertices with
¡ ¢
all the nt edges determined by them having the same color.
The smallest m in Theorem 3.2 is called the Ramsey number and denoted by Rt (n).
Then, clearly R1 (n) = 2n − 1 (as mentioned above discussing the Pigeonhole Principle)
and R2 (n) is the same as R(n) introduced above.
Proof. We prove by induction on t. The case t = 1 is the pigeonhole principle (R1 (n) =
2n − 1).
Consider t fixed and assume that Rt (k) exists for any k ≥ t. We are going to prove
that Rt+1 (n) exists.
The heart of the proof is the definition of the “shadow coloring”. If the edges of
t+1
H = Km are 2-colored and a vertex x is removed from H then H − x can be considered
t
as a 2-colored Km−1 , namely: a t-element subset T of H − x is considered to be colored
with the same color as T ∪ x in H. We call briefly this coloring as the shadow coloring
of H − x . Observe that if m is large (depending on k) then the induction hypothesis
guarantees a monochromatic Kkt in the shadow coloring of H − x. In other words, there
exists a set K with k vertices such that the color of each t + 1-element subset containing
x in the set x ∪ K is the same.
Using that k can be chosen arbitrarily in terms of n, one can select a huge m which
allows to find a set S of 2n − 1 vertices, S = {x1 , x2 , . . . x2n−1 }, so that the color of any
t + 1-element subset of S depends only on the smallest index of its elements. By the
pigeonhole principle, there exist n elements of S determining a monochromatic Knt+1 . ¤
Upper bound on R(n).

The next theorem gives the explicit bound for m from the proof above (for the case
t = 2).
Theorem 3.3 R(n) ≤ 4n .
Proof. Set m = 4n = 22n , consider a 2-coloring of the edges of the complete graph Km .
The shadow coloring at any vertex x1 is a vertex coloring. By the pigeonhole principle,
25
one can find 22n−1 vertices all adjacent to x1 in the same color. Select x2 arbitrarily
among those vertices and continue... One can select S = {x1 , x2 , . . . , x2n−1 } so that the
color of any edge {xi , xj } within S depends only on the minimum of i and j. This gives
a monochromatic Kn in S (as in the proof of Theorem 3.2). ¤
Many colors and non-diagonal Ramsey numbers. There are natural generaliza-
tions of the numbers Rt (n). The Ramsey number
Rt (n1 , n2 , . . . , nr )
t
is the smallest integer m with the following property: if the edges of Km are colored
with colors 1, 2, . . . , r in any fashion then, for some i, there exists a Knt i , monochromatic
¡ ¢
in color i (i.e. all the nti edges within some set of ni vertices are colored with color i).
The number Rt (n1 , n2 , . . . , nr ) is sometimes called non-diagonal Ramsey number (for
r colors) emphasizing that the numbers ni can be different. In the same spirit, if all ni
are equal than the term diagonal Ramsey number is used (for example, R(n) = R2 (n, n)
is a diagonal Ramsey number). The existence of non-diagonal Ramsey numbers follows
immediately from the existence of diagonal ones. The order of magnitude of R(3, n)
2n
( log(n) ) was clarified in 1995 (Kim) after a a long process of advances. The role of non-
diagonal Ramsey numbers can be illustrated by the following basic inequality of Erdős
and Szekeres.
Theorem 3.4 R(p, q) ≤ R(p − 1, q) + R(p, q − 1).
Proof. Exercise [3.10] ¤
¡p+q−2¢
Corollary 3.5 R(p, q) ≤ p−1
.
Exact values of Ramsey numbers. It is known that R(3, 4) = 9 and R(4) = 18

(exercises [3.1],[3.2],[3.3]). Further known values: R(3, 5) = 14, R(3, 6) = 18, R(3, 7) =
23, R(3, 8) = 28, R(3, 9) = 36, R(4, 5) = 25.
Paul Erdős thought that it would take enormous effort to determine R(5) but to
determine R(6) is impossible.
26
The only nontrivial exact value known for three colors is R(3, 3, 3) = 17 (see exercise
[3.4]). The only known value for hypergraphs is R3 (4, 4) = 13.
Convex n-gons. The following surprising application of Ramsey numbers is a cele-
brated theorem of Erdős and Szekeres.
Theorem 3.6 There exists a function f (n) with the following property: if f (n) points
of the plane are in general position (no three on a line) then there exist n of these points
forming the vertices of a convex n-gon.
Proof. We show that R4 (5, n) is a good choice for f (n). Consider a set P of R4 (5, n)
points in general position. A four-element subset of P is called convex if the convex hull
of the four points is a quadrangle, otherwise it is called concave. This is a partition of all
four-element subsets of P therefore, from the definition of the Ramsey number R4 (5, n),
either there exist five points of P whose four-element subsets are concave or there exist
n points of P whose four-element subsets are convex. However, the former possibility is
impossible ([T3.2]) and the latter implies that we have n points which are the vertices
of a convex n-gon ([T3.3]). ¤
3.2 VAN DER WAERDEN NUMBERS

Returning to the Pigeonhole Principle, clearly any 2-coloring of [2k − 1] guarantees a
monochromatic k-element subset. In Ramsey theory one looks for special monochromatic
sets, the first famous example is the case of arithmetic progressions. Van der Waerden
proved in 1927 that for every integer k > 1 there exists an integer m (depending on
k) with the following property: if the numbers in [m] are 2-colored in any fashion then
there is a monochromatic k-term A.P. (arithmetic progression). The smallest m with
this property is called the Van der Waerden number W (k). Clearly, W (2) = 3 because
any two numbers form a 2-term A.P. The next paragraph gives an argument showing
that W (3) is well-defined.
Claim: W (3) ≤ 325. Call a 3-term A.P. almost monochromatic if its first two terms
have the same color. A base block is five consecutive numbers. With these definitions
the following is obvious:
27
Proposition 3.7 A 2-colored base block contains an almost monochromatic 3-term A.P.
Consider a 2-coloring C of [325] and partition [325] into 65 base blocks, B1 , B2 , . . . B65 ,
starting with B1 = [5]. Since there are only 25 ways to color a base block, the Pigeonhole
Principle guarantees the existence of 1 ≤ i < j ≤ 33 such that C colors Bi and Bj
precisely the same way. Now we can select l ≤ 65 so that the three base blocks Bi , Bj , Bl
form a 3-term A.P. of blocks which means that the first elements of the blocks form a 3-
term A.P. Applying Proposition 3.7 for Bi , we can find an almost monochromatic 3-term
A.P. in Bi . If this is monochromatic, the claim follows. Otherwise, w.l.o.g., C colors
it as RRB (red,red,blue). Since Bj is colored precisely like Bi , one can find two 3-term
A.P.-s with identical third term such that one of them is colored RR? and the other is
colored BB?. This finishes the proof of the claim (see Figure 3.1).
Figure 3.1. RR? and BB?
Of course, 325 is just an upper bound on W (3). It is not very difficult to prove that
W (3) = 9, but that proof is not suitable to generalize. The next step shows how to prove
the existence of W (4) from the existence of W (3, r), which is the smallest m with the
property: if [m] is r-colored in any fashion then there exists a monochromatic 3-term
A.P. It is strange that a theorem about 2-colorings can not be proved without proving
it for r-colorings... (This interesting phenomena occurred earlier. The construction of
[p, 2, n]-hypergraphs (graphs!) relied on constructions of [p − 1, K, n]-hypergraphs.)
W (4) < ∞ if W (3, r) < ∞. A base block of length t = 2W (3) − 1 ensures that Propo-
sition 3.7 can be generalized: a 2-colored base block contains an almost monochromatic
4-term A.P. (first three terms of the same color). But now we need three identically
colored base blocks forming a 3-term A.P. of blocks! Since a base block of length t can
be 2-colored in 2t ways we need W (3, 2t ) base blocks to ensure that. To extend it with
28
a fourth base block is just a modest doubling and then we have the possibility to find
two 4-term A.P.-s colored as RRR? and BBB? such that their fourth terms are identical.
In fact, t = 13 works here, (using W (3) = 9 instead of the upper bound 325 from the
proof) we have
W (4) ≤ t(2W (3, 2t ) − 1) < 13W (3, 213 ).
Now the existence of W (4) follows from the existence of W (3, r) where r = 213 . One
can guess that this is not very small, especially not from our proof which will be roughly
2
..
2.
2
where the hight of the tower is r. Concerning the existence of W (3, r) only the first step
of the iteration process is shown.
W (3, 3) < ∞ . Blocks of length 7 are the ’small base blocks’, one can find almost
monochromatic 3-term A.P.-s in them. Blocks of length 2 × 7 × 37 are the ’medium base
blocks’, one can find in them the pattern shown in Figure 3.2. The final step is to find
two identically colored medium blocks and extend them to a 3-term A.P. of medium
blocks (see Figure 3.2). The argument gives
7
W (3, 3) < 2(2 × 7 × 37 )(32×7×3 )
and Exercise [3.14] asks to work out the details.
Figure 3.2. Focusing three copies of medium blocks.
Hopefully the presented argument is convincing enough to accept
Theorem 3.8 (Van der Waerden) For every k, r there exists m (depending on k, r) such
that any r-coloring of [m] contains a monochromatic k-term arithmetic progression.
29
The proof outlined above gives a bound for W (k, 2) which grows as the Ackermann
function. In 1989 Shelah found a better bound, W (k, 2) ≤ wow(k + 1), where “wow” is
the iterated tower function
22
wow(2) = 2, wow(3) = 22 , wow(4) = 22 = 65536
and wow(5) is a tower of 2’s with height 65536 - wow! For comparison, note that
W (4) = 35 and W (3, 3) = 27. It is also known that W (5) = 178. The best lower bound
of W (k, 2) is exponential (see Section 4).
3.3 TIC-TAC-TOE AND HALES-JEWETT THEOREM

The cube and its lines. The cube Ctn is defined as the set of vectors of length n using
t symbols. These vectors are simply called the points of the cube. Clearly, Ctn has tn
points. For example, C2n with symbols 0, 1 is the usual n-dimensional cube.
We shall define lines of the cube Ctn as follows. Assume that the symbol set is
ordered, for example consider the symbols 1, 2, . . . , t with natural order. Each line of Ctn
will contain t points, P1 , P2 , . . . , Pt . Certain coordinates (at least one) are called moving
coordinates. The symbol in Pi at a moving coordinate must be i (the i-th symbol). The
coordinates which are not moving, are called the constant coordinates. At any constant
coordinate all points Pj (1 ≤ j ≤ t) must have the same symbol (but that common
symbol may vary at distinct constant coordinates).
P ¡ ¢
At this point it is useful to check that Ctn has ni=1 ni tn−i lines ([T3.5]). We note
that lines are defined as a subset of ’geometric’ lines. An even wider definition of lines
would be to require only that, at each coordinate, the symbols are all equal or all distinct.
Such a definition is used in the game “SET”, played with cards representing C34 .
The cubes C32 and C23 together with their lines are shown on Figure 3.3.
Theorem 3.9 (Hales-Jewett) For arbitrary positive integers t, r there exists a positive
integer n (depending on t, r) with the following property: for every r-coloring of the
points of the cube Ctn there exists a monochromatic line (i.e. a line whose points are
colored with the same color).
30
Figure 3.3. The cubes C32 and C23 and their lines.
Corollary 3.10 Van der Waerden theorem
Tic-tac-toe. In the r-player Tic-tac-toe game the players color the points of Ctn (players
use distinct colors) and their objective is to create a monochromatic line (using some
definition of a line). Since the lines defined above are the most restrictive, Theorem 3.9
implies that there is no draw if r persons play with fixed board size t in sufficiently high
dimenson n. In case of two persons, the first player always wins (again, if n is large
enough).
Self-test 3.
[T3.1] Show that R(n) > (n − 1)2 .
[T3.2] Assume that five points of the plane are in general position (no three on a line).
Show that four of them are the vertices of a convex quadrangle.
[T3.3] Assume that n points of the plane are in general position and all subsets of four
points determine a convex quadrangle. Prove that the n points are the vertices of a
convex n-gon.
[T3.4] Consider the k-uniform hypergraph A(k, n) whose vertex set is [n] and whose
edges are the k-term A.P.-s. Formulate Van der Waerden’s theorem as a statement
about the chromatic number of this hypergraph.
P ¡ ¢
[T3.5] Show that Ctn has ni=1 ni tn−i lines.
Exercise set 3.
31
[3.1] Prove that R(3, 4) ≤ 9.
[3.2] Prove that R(3, 4) ≥ 9.
[3.3] Prove that R(4, 4) ≤ 18. (The reverse inequality comes from the following famous
coloring of the edges of K17 : the vertex set is [17] and the edge (i, j) is red if i − j is a
square (mod 17) otherwise it is blue.)
[3.4] Prove that R(3, 3, 3) ≤ 17.
[3.5] Show that R(3, 3, . . . , 3) ≤ 3r! (assuming r arguments). Extra part: improve the
bound to ber!c + 1 (this is the current record set eighty years ago).
[3.6] Show that R(3, 3, . . . , 3) > 2r (assuming r arguments). Extra: Ideas for improve-
ment?
[3.7] Use the result of exercise [3.5] to prove the following: if n ≥ 3r! then no matter
how [n] is partitioned into r classes, there is a solution of the equation x + y = z with
all three numbers from the same class (x = y is permitted in the solution).
[3.8] Show that if the edges of a countably infinite complete graph are colored red or
blue in any fashion then there is a infinite monochromatic complete subgraph.
[3.9] (Kiran Kedlaya, former BSM student) Show that R(n) ≤ R3 (6, n).
[3.10] Prove Theorem 3.4!
[3.11] Prove the corollary of Theorem 3.4!
[3.12] Prove that R3 (n, n) is also a good choice for f (n) in Theorem 3.6 . Hint: color
triples according to the parity of the number of points inside the triangle spanned by
them.
[3.13] Assume (a real life situation) that ’knowing each other’ is not necessarily sym-
metric. Prove that at a party with nine persons there are either three persons among
which nobody knows any other or there are three persons A, B, C such that A knows B,
B knows C and A knows C (a transitive triple). Show that the same statement is not
necessarily true for a party of eight persons.
[3.14] Show that
7
W (3, 3) < 2(2 × 7 × 37 )(32×7×3 )
[3.15] Show that Hales-Jewett theorem implies Van der Waerden theorem.
32
[3.16] Show that
Rt+1 (n1 , n2 ) ≤ Rt (Rt+1 (n1 − 1, n2 ), Rt+1 (n1 , n2 − 1)) + 1
33
4 COUNTING AND PROBABILITY
4.1 PROOFS BY COUNTING

In this section some results from extremal hypergraph theory are given with proofs which
are based on counting. The extremal problems are of the following type: what is the
maximum or minimum number of edges in hypergraphs with a certain property?
Antichains. An antichain is a hypergraph H = (V, E) in which no edge contains any
other edge (a chain is a hypergraph in which for each pair of edges, one contains the
other.)
Theorem 4.1 (Sperner) Assume that V = [n] and H = (V, E) is an antichain. Then
µ ¶
n
|E| ≤ .
b n2 c
Proof. (Lubell)We say that a permutation x1 x2 . . . xn of V = [n] extends the edge e ∈ E

if e = {x1 , x2 , . . . , xt } for some t. Observe that
|e|!(n − |e|)! permutations extend e∈E
and no permutation extends distinct edges. Therefore

X
|e|!(n − |e|)! ≤ n!
e∈E
which is equivalent with

X µ n ¶−1
≤1
e∈E
|e|
¡ n
¢−1 ¡n¢
Each term is at least bn c
because k
is maximum for k = b n2 c. This gives
2
µ ¶−1
n
|E| n ≤1
b2c
and the proof is finished. ¤
Intersecting hypergraphs. What is the maximum number of edges of an intersecting

hypergraph H = (V, E) on vertex set V = [n]? We have to assume that H is simple,
34
otherwise the question is foolish. With this assumption, the answer is not difficult, guess
it before you look at exercise [4.2]. What happens if we ask the same question for simple
t-uniform hypergraphs? If n ≤ 2t − 1 then the answer is immediate ([T4.2]). For larger
values of n, the answer is provided by
Theorem 4.2 (Erdős-Ko-Rado) Assume that H = (V, E) is a simple intersecting t-

¡ ¢
uniform hypergraph on vertex set V = [n] and n ≥ 2t. Then |E| ≤ n−1
t−1
.
Proof. (Katona)We say that a cyclic permutation of [n] extends an edge e ∈ E if the
vertices of e appear in consecutive positions of the cyclic permutation. There are t!(n−t)!
cyclic permutations extending e. On the other hand, a cyclic permutation can be the
extension of at most t edges. This is not so obvious, exercise [4.3] asks for the proof.
The number of cyclic permutations of [n] is (n − 1)! therefore
|E|t!(n − t)! ≤ t(n − 1)!
which, like a miracle, gives the statement to be proved. ¤
3-chromatic uniform hypergraphs. What is the minimum number of edges in a t-

uniform 3-chromatic hypergraph? If t = 2, i.e. we have a graph, the answer is 3 because
a graph with just two edges is clearly 2-chromatic and the triangle is 3-chromatic. If
t = 3 then the answer is 7, the Fano plane gives an example of a 3-uniform 3-chromatic
hypergraph, but it is not so easy to prove that all 3-uniform hypergraphs with six edges
are 2-chromatic. For the next step, when t = 4, the answer is not known. It is rather
surprising that one needs at least 2t−1 edges as the following famous theorem with equally
famous proof claims.
Theorem 4.3 (Erdős - Selfridge) A t-uniform hypergraph with less than 2t−1 edges is
2-chromatic.
Proof. Assume that H = (V, E) is a t-uniform hypergraph with vertex set V = [n].
For each edge e ∈ E there are 2 × 2n−t = 2n−t+1 distinct 2-colorings of V which makes e
monochromatic. This implies that at most |E|2n−t+1 2-colorings of V are monochromatic
on some edges of H. However, from the assumption |E| < 2t−1 , that number is smaller
35
than 2n , the number of 2-colorings of V . This means that there exists a 2-coloring of V
under which no edge of H is monochromatic! ¤
4.2 PROBABILITY METHOD

Counting versus probability. First we repeat the proof of Theorem 4.3 in terms of the
probability method. Consider the probability space of all possible (2n ) 2-colorings of [n]
where each coloring has the same ( 21n ) probability. The probability of a monochromatic
1
edge is 2t−1
therefore the probability of the event that some edge is monochromatic is
|E|
not larger than 2t−1
which is smaller than 1 from the assumption of Theorem 4.3. Thus
the complementary event has positive probability which finishes the proof.
In all subsequent proofs such a simple probability space is used. It has only one
parameter N , the number of elements of the probability space, each element has proba-
1
bility N
. The probability of subsets (events) of the probability space is the cardinality
of the subset divided by N . We shall use the obvious fact that the probability of union
of events is at most the sum of the probabilities of the events. (Equality if and only if
the events are pairwise disjoint.) In many applications the probability space is colorings
of vertices or edges of a graph or hypergraph. We shall prove highly nontrivial theorems
from a trivial property of the model, namely that only the empty set has probability
zero.
The Erdős lower bound on R(n). The best lower bound of R(n) we have seen so far
is quadratic in n (Self-test [T3.1]). A cubic one is to be constructed in the next section
which can be generalized to a construction providing a lower bound which grows faster
than any polynomial of n (with fixed degree). However, no construction is known which
reaches cn (with some constant c > 1). Still, the probability method gives such a bound
easily.
¡m¢
< 2( 2 )−1 .
n
Theorem 4.4 (Erdős) m < R(n) if n
Proof. The probability space is the 2-colorings of the edges of Km . This space has N =
2( 2 ) elements. A subset of n vertices in K spans a monochromatic complete subgraph
m
m
1
with probability p = . Therefore the probability of having a monochromatic Kn is
2( 2 )
n −1
36
¡m¢
at most p n
< 1 from the assumption of the theorem. This means that the probability
of having no monochromatic Kn is positive. Therefore there exists a 2-coloring of the
edges of Km such that there is no monochromatic Kn . This clearly implies that m is a
lower bound for R(n). ¤
The lower bound m for R(n) was stated in a strange way in Theorem 4.4 to show
the idea of the proof clearly. To get an actual bound, one has to calculate the largest m
(in terms of n) satisfying the condition
µ ¶
m
< 2( 2 )−1 .
n
n
Such calculations often require upper bound on binomial coefficients. The better the
estimate, the better the bound. If we start with the trivial upper bound
µ ¶
m
< mn
n
then our lower bound is
n 1 1
m = 2 2 × 2− n × 2− 2
¡m¢
([T4.3]). If Sterling’s formula is used for the asymptotic of n
then
n n
m = 22 × √
e 2
is obtained so there is no earthshaking difference between the trivial and the best
estimate of m.
n
Corollary 4.5 (Erdős) For n ≥ 3, 2 2 < R(n).
Lower bound for W (k).

k
Theorem 4.6 W (k) ≥ 2 2 .
Proof. Outline: ¡ n¢
2
< 1.
2k−1
Tournaments. A tournament Tn is an oriented Kn , which means that each edge of the
complete graph Kn gets a one-way orientation. The edge from vertex v to vertex w is
37
denoted by (v, w). We can interpret a tournament Tn as vertices represent players and
(v, w) means that v beats w in the game.
There are two non-isomorphic T3 -s, one is the cyclic triangle and the other is the
transitive triangle. ([T4.4]: define the isomorphism of tournaments) A tournament is
called transitive if the presence of the edges (i, j) and (j, k) imply that (i, k) is also
an edge. A tournament Tn is transitive if and only if its vertices can be labeled with
1, 2, . . . , n so that all edges point from smaller label to larger label (exercise [4.4]). This
shows that transitive tournaments on the same number of vertices are isomorphic.
Large transitive subtournaments.
Let f (k) be the smallest n such that any Tn contains a transitive Tk . It is obvious
that f (2) = 2, it is easy that f (3) = 4 ([T4.5]). Then f (4) = 8 and f (5) = 14.
k−1
Theorem 4.7 2 2 ≤ f (k) ≤ 2k−1 .
Proof. The lower bound comes from the probability method. The probability space is
the set of all tournaments on [n], there are N = 2( 2 ) elements in the probability space.
n
If ¡n¢
k
k!
<1
2( )
k
2
then there exists a Tn which does not contain a transitive subtournament on k vertices.
Work out the details! (Exercise [4.5])
The upper bound comes easily by induction (Exercise [4.6]). ¤
Existence versus construction. Theorem 4.7 shows that there exists a tournament
T127 which does not contain a transitive T15 because the probability of having a transitive
T15 in every T127 is smaller than one. In a “real world” problem, it would be important to
know how much smaller? The following situation is from the book of Erdős and Spencer:
Probabilistic methods in Combinatorics.
“A computer scientist is told to construct , on a computer, the matrix of a tournament
on 127 players that does not contain a transitive subtornament on 15 players. Imagine
that if his matrix does contain a transitive subtournament on 15 players, his company
(or country or mutual fund) will suffer grave financial consequences. He learns from this
book that such a tournament exists. But this is not sufficient - his boss wants a specific
38
matrix stored in the computer. Our scientist calculates that it may cost millions of
dollars (or tens of millions of forints ) to find such a T by exhaustive methods. However,
like many of us, he is on a limited budget. What is he going to do? If his moral character
is sufficiently low he would construct T at random, pack his bags to prepare for a hasty
departure, and hope for the best. Given our experiences with computer scientists, we
surmise that this would be the “real world” solution. If, on the other hand, he is of
outstanding moral character his only recourse is to ask for a refund on this book. We
would, of course, refuse.”
This was written more than twenty years ago so certain statements (definitely the
conversion rate of dollars to forints) may not be valid any more. The quoted paragraph
continues:
“The “real world” situation is different if it is required that the tournament does
not contain a transitive subtournament on 16 players. Then a “random” tournament
will have the desired property with probability ≥ .998. This probability of success is
sufficient in many practical situation - our computer scientist can randomize and relax.”
Paradoxical tournaments. Who are the best k players in a tournament? This can
not be answered easily because, for every k, there are tournaments in which every set of
k players is beaten by somebody. Tournaments of this property are called k-paradoxical
.
Theorem 4.8 For every positive integer k, there exist k-paradoxical tournaments.
Proof. At this point, the reader can easily prove it (exercise [4.7]) by interpreting the
inequality µ ¶
n 1
(1 − k )n−k < 1
k 2
¤
39
Hamiltonian Paths in Tournaments.
A famous theorem of Rédei (1937) says that the players of any tournament can be
placed in a line so that each player won the match against her left hand neighbor. More
formally, every tournament has a (directed) Hamiltonian path. The simple proof is
left as an exercise ([4.8]). (A more difficult result is that the number of Hamiltonian
paths is odd in every tournament.) How many Hamiltonian paths can we found in a
tournament? If the tournament is transitive then there is only one. On the other hand,
there are tournaments with many Hamiltonian paths.
n!
Theorem 4.9 For every n there is a tournament Tn with at least 2n−1
Hamiltonian
paths.
Proof. (Outline.) A given permutation of V = [n] gives a Hamiltonian path with

1 n!
probability 2n−1
. This implies that the expected number of H. paths is 2n−1
. Therefore
there exists a Tn with at least that many H. paths. ¤
4.3 LOCAL LEMMA

All the proofs we have done by the probability method can be described using the
following scheme. There are events A1 , A2 , . . . , An in a probability space and we know
that for A = A1 ∪ A2 ∪ . . . ∪ An , prob(A) < 1. Then prob(A) > 0 which means that the
event A exists (nonempty). We have also seen that requiring prob(A) < ² will ensure
the existence of A with probability 1 − ².
Another possibility to ensure prob(A) > 0 is to assume mutual independence of the
¡ ¢
events Ai and prob(Ai ) < 1 (for all i). In this case prob Ai > 0 and
Ãn ! n
¡ ¢ \ Y
prob A = prob Ai = Ai > 0
i=1 i=1
follows from mutual independence. However, mutual independence is almost never

present in applications. The Local Lemma, which is introduced in a paper of Erdős
and Lovász, allows a limited number of dependencies among the events Ai . Initially
40
the lemma was applied to a specific problem (to prove Theorem 4.11 below) but later
became an important tool applicable in many situations.
Dependency bound. The set of events {A1 , A2 , . . . , An } has dependency bound D if,
for each i, Ai is dependent on at most D other events Aj . More precisely, this means
that for each i, there is a set Si ⊆ [n] such that |Si | ≤ D, i ∈
/ Si and Ai is independent of
any boolean combination of the events {Aj : j ∈
/ Si }. (Two events A, B are independent
if prob(A ∩ B) = prob(A)prob(B).)
Theorem 4.10 (Local Lemma) Assume that A1 , A2 , . . . , An are events with dependency
1
bound D and prob(Ai ) ≤ 4D
for 1 ≤ i ≤ n. Then
Ãn !
\
prob Ai > 0
i=1
Proof. We shall use occasionally AB for A ∩ B, pr for prob. It is convenient to use the
notion of conditional probability. For non-empty B define probability of A with condition
pr(AB)
B as pr(A|B) := pr(B)
. Then A, B are independent if and only if pr(A|B) = pr(A).
Claim: for arbitrary S ⊆ [n] and i ∈
/S
Ã !
\ 1
P = pr Ai | Aj ≤
j∈S
2D
1
The claim is proved by induction on |S|, for S = ∅ it is true because pr(Ai ) ≤ 4D
from the
condition of the theorem. From the assumptions, S can be partitioned into S1 and S2 so
that |S1 | ≤ D and Ai is independent of any boolean combination of {Aj : j ∈ S2 , j 6= i}.
pr(AB|C)
Using the identity pr(A|BC) = pr(B|C)
write P as
³ T T ´
pr Ai j∈S1 Aj | j∈S2 Aj
P = ³T T ´ (2)
pr j∈S1 Aj | j∈S2 Aj
The numerator in (2) is bounded from above by

Ã !
\ 1
pr Ai | Aj = pr(Ai ) ≤
j∈S
4D
2
41
and the denominator in (2) is bounded from below by
Ã ! Ã !
[ \ X \ X 1 1 1
1 − pr Aj | Aj ≥ 1 − pr Aj | Aj ≥ 1 − ≥1− =
j∈S j∈S j∈S j∈S j∈S
2D 2 2
1 2 1 2 1
and the claim follows.

The proof is finished because
Ã n ! n n µ ¶
\ Y Y 1
pr Aj = (Ai |A1 ∩ A2 ∩ . . . ∩ Ai−1 ) ≥ 1− > 0.
j=1 i=1 i=1
2D
Erdős - Lovász theorem.
Theorem 4.11 Assume that H = (V, E) is a t-uniform hypergraph in which each edge
intersects at most 2t−3 other edges (t ≥ 3). Then H is 2-chromatic.
Proof. Assume that the vertex set of H is [n]. Consider the probability space of all
2-colorings of V . For each edge ei ∈ E, let Ai be the event that ei is monochromatic.
1
Clearly prob(Ai ) = 2t−1
and D = 2t−3 is a good dependency bound for the events
because Ai is independent from any Boolean combinations of those Aj -s for which ej
does not intersect ei . Thus the condition for the Local Lemma is satisfied with equality
1 1
( 2t−1 = 4D
). Then the Local Lemma says
Ã !
\
prob Ai > 0
ei ∈E
i.e. there exists an event (2-coloring of V ) under which no edge is monochromatic. ¤
Improved lower bound on W (k).
2k
Theorem 4.12 W (k) > 8k
.
Proof. The probability space is all 2-colorings of [n]. For each k-term A.P. S, AS is the
1
event that S is monochromatic. Clearly, prob(AS ) = 2k−1
. Since a fixed S can intersect
42
at most nk other S’s (Ex.[4.9]), D = nk is a good dependency bound. The condition of
2k
the Local Lemma is satisfied with equality if n = 8k
. ¤
Even cycles in regular digraphs.
A directed graph is r-regular if the indegree and the outdegree of each vertex is r. It
was a long-standing conjecture that for sufficiently large r any r-regular digraph contains
an even (directed) cycle. A surprising proof was found by Alon.
Theorem 4.13 For r ≥ 8 any r-regular digraph has an even cycle.
Proof. The probability space is the 2-colorings of the vertex set of the given r-regular
digraph G. A coloring is ’good’ if for each vertex x of G there exists an edge (x, y) such
that the color of x and y are different. A coloring is ’bad’ at x if there is no vertex y such
that the color of x and y are different. Let Ax be the event that a coloring is bad at x.
1
Then prob(Ax ) = 2r
. On the other hand, D = r2 is good dependency bound for the set
of events {Ax : x ∈ V } (check! Ex.[4.10] ). To apply the Local Lemma, the condition
1 1
2r
≤ 4r2
is needed which is valid since r ≥ 8. Now the Local Lemma says that there
exists a good coloring which immediately gives an even cycle ([T4.6]). ¤
4.4 JOKES.
The triangle is 2-chromatic!
Theorem 4.14 Joke 1. The graph K3 is 2-chromatic.
Proof. Let V = [3] be the vertex set of the triangle T . The probability space is the 2-
colorings of V (N = 8 elements). Let Aij be the event that the edge ij is monochromatic.
1
Clearly, prob(Aij ) = 2
for all the three choices of index pairs. Any pair of events is
independent, for example
1
prob (A12 A13 ) = prob (A12 ) prob (A13 ) =
4
1
therefore D = 2
is a good dependency bound. The condition for the application of the
Local Lemma holds with equality because
1 1
prob (Aij ) = =
2 4D
43
Therefore there exists a coloring of V so that no edge is monochromatic. ¤
Spencer’s injections.
A function f : S → T is called injective if its domain is S and it is one-to-one.
¡|S|¢
Theorem 4.15 Joke 2 (Spencer) Assume that S is a finite set. If |T | > 2
then there
exists an injective function f : S → T .
Proof. The probability method is used with probability space of all f : S → T functions.
The event Axy is that f (x) = f (y). Clearly
1
prob (Axy ) =
|T |
for all x, y ∈ S. Since

Ã ! ¡|S|¢
[ X 1
prob Axy ≤ = 2 <1
x,y∈S x,y∈S
|T | |T |
it follows that Ã !
\
prob Axy >0
x,y∈S
so an injective f exists. ¤
One can try to improve the theorem above by the Local Lemma. Indeed, an order of
magnitude better result comes from it.
Theorem 4.16 Joke 3 (Spencer) Assume that S is a finite set. If |T | ≥ 8|S| then there
exists an injective function f : S → T .
Proof. Follow the previous proof, but apply the Local Lemma. Clearly, D = 2(|S| − 2)
is a good dependency bound for the events Axy . The condition |T | ≥ 8|S| makes the
Local Lemma applicable. ¤
Self-test [T4.8] asks for further improvements along this line...
44
Self-test 4.
[T4.1] Is it true that an antichain is a simple hypergraph?
[T4.2] What is the maximum number of edges of a t-uniform intersecting simple hyper-
graph on vertex set [n], if n ≤ 2t − 1 ?
[T4.3] Check the calculation of the lower bound in Theorem 4.4 if the upper bound mn
¡ ¢
is used for m
n
.
[T4.4] Define isomorphism for tournaments.
[T4.5] Show that any T4 contains a transitive T3 .
[T4.6] Show that a digraph with a good coloring (defined in the proof of Theorem 4.13
has an even cycle!
[T4.7] Find the mistake in the proof of Joke 1.
[T4.8] Find the strongest theorem generalizing the results of Jokes 2 and 3.
Exercise set 4.
[4.1] Let H be an antichain with fk edges of k elements. Prove the so called LYM
inequality:
n
X µ ¶−1
n
fk ≤1
k=0
k
[4.2] Prove that an intersecting simple hypergraph on vertex set [n] has at most 2n−1
edges!
[4.3] Prove that any cyclic permutation extends at most t edges! (See the proof of
Theorem 4.2.)
[4.4] Prove that if Tn is transitive then its vertices can be labeled with 1, 2, . . . , n so that
all edges point from smaller label to larger label.
[4.5] Work out the details for the lower bound in Theorem 4.7
[4.6] Work out the inductive proof for the upper bound in Theorem 4.7.
[4.7] Prove Theorem 4.8 by interpreting the stated inequality!
[4.8] Prove that every tournament has a directed Hamiltonian path.
[*4.9] Prove that a k-term A.P. in [n] can intersect at most kn other k-term A.P.-s (in
[n]).
[4.10] Show that D = r2 is a good dependency bound in the proof of Theorem 4.12.
45
5 LINEAR ALGEBRA METHOD
5.1 THE DIMENSION BOUND

Oddtown. Oddtown is a little town somewhere (probably near to Chicago) whose
inhabitants like to form clubs. They insist to have clubs with an odd number of members
to settle all matters easily by a majority vote. Their other rule is that each pair of clubs
must share an even number of members, because they can prove
Theorem 5.1 Oddtown (under the stated rules) has no more clubs than the number of
inhabitants.
Proof. Assume that the inhabitants form the vertices and the clubs form the edges of
the hypergraph H = (V, E) . Set V = [n] and E = {e1 , e2 , . . . , em }. Consider the linear
space of n-component vectors over the two-element field F2 = {0, 1}. The edges of H are
viewed as elements of this linear space by looking at ei as a row vector in the incidence
matrix of H. The two rules of Oddtown can be expressed nicely with the standard inner
product:
(
1 if i = j
ei ej =
0 if i 6= j
Claim: e1 , e2 , . . . , em are linearly independent.
To see this, assume that
m
X
λi ei = 0 (λi ∈ F2 )
i=1
and take the inner product of both sides by ej . Using the property of the inner product
displayed above, we get λj = 0 and this is valid for any j proving the claim.
Therefore m ≤ the dimension of the linear space = n. ¤
The main motive of this cute proof is that the dimension is an upper bound for
the number of linearly independent elements in any linear space. One can easily find a
similar theorem and proof for “Modprimetown” (exercise [5.1]). A further trick is needed
46
for the proof of the “Modprimepowertown” theorem (exercise [5.2]). Surprisingly, it is
unknown whether a similar “Modsixtown” theorem is true or not. But presently there is
the strong belief in Modsixtown that they will never have more clubs than the number
of inhabitants.
Fisher inequality.
Theorem 5.2 Assume that H = (V, E) is a simple hypergraph without empty edges and
each pair of distinct edges intersect in precisely λ elements. Then |E| ≤ |V |.
Proof. There are some easy special cases treated separately. For λ = 0 the edges
are pairwise disjoint, so the theorem follows immediately. If there is an edge e ∈ E
with |e| = λ then the theorem follows easily ([T5.1]). Assume these cases are excluded,
|V | = n, |E| = m.
The edges of H are considered again as row vectors from the incidence matrix. The
intersection property of the edges can be expressed with the inner product as follows
(with suitable positive integers γi ):
(
λ + γi if i = j
ei ej =
λ if i 6= j
Claim: The vectors ei are linearly independent in the linear space of n-component
P
vectors of real numbers (i.e. in Rn ). Assume, on the contrary, that mi=1 αi ei = 0 and
compute the inner product of both sides by ej . We get that for all j (j ∈ [m])
λβ + αj γj = 0
Pm
where β = i=1 αi . Expressing αj (γj 6= 0), we get that for all j ∈ [m]
λ
αj = − β (3)
γj
proving the claim if β = 0. If β 6= 0 then adding (5) for all j ∈ [m] we get
Ã m !
X 1
β = −λ β
j=1
γ j
47
which is contradiction (λ > 0, β 6= 0). This proves the claim and the theorem follows
from the dimension argument. ¤
Corollary 5.3 Theorem 1.3 (de Bruijn- Erdős theorem).

Proof. Self-test [T5.2]
A cubic lower bound for R(n).
¡n−1¢
Proposition 5.4 (Zs.Nagy) R(n) > 3
Proof. Here is the tricky coloring of the edges of the complete graph K = K(n−1)
3
with red and blue. Associate the vertices of K with the edges (triples) of the complete
3
hypergraph T = Kn−1 . Color an edge of K red if the corresponding two triples of T
intersect in precisely one vertex. Otherwise color the edge blue.
A red complete subgraph of K corresponds to triples of T pairwise intersecting in
precisely one element. From Theorem 5.2, there are at most n − 1 such triples. A blue
complete subgraph of K corresponds to triples of T pairwise intersecting in zero or two
elements. Since three is an odd number, we have an oddtown situation, theorem 5.1
shows that there are at most n − 1 triples again. Therefore in our coloring of the edges
of K there is no monochromatic Kn . ¤
We note here that the above construction can be generalized to give superpolynomial
lower bound for R(n). For a fixed prime p define a red-blue coloring by associating ver-
tices to (p2 − 1)-element subsets of an m-element set and coloring the edge between two
vertices red if the corresponding subsets has intersection size congruent to −1 (mod p).
Otherwise the edge is blue. Observe that proposition 5.4 gives this coloring for p = 2.
However, it is not as easy to show that under this coloring there are no large monochro-
¡m¢
matic complete subgraphs [not larger than 2 p−1 ].
Two-distance sets. A one-distance set in Rn is a set of points (vectors) such that
any two of them have the same distance. It is easy to see that a one-distance set has at
most n + 1 points and equality is possible (exercise [5.3]). How many points can form a
two-distance set in Rn , i.e. a set such that the distances between the points take only
¡ ¢
two values? It is easy to find a two-distance set with n2 points ([T5.3]). The next
theorem gives an upper bound not far from this lower bound.
48
(n+1)(n+4)
Theorem 5.5 A two-distance set of Rn has no more than 2
points.
Proof. Assume that {a1 , . . . , am } is a two-distance set of Rn with distances δ1 and δ2 .

Using
Ã n ! 21
X
||x|| = x2k
k=1
the distance of x, y is ||x − y||. Define the polynomials (in n variables) fi (x) for i ∈ [m]
as
fi (x) = (||x − ai ||2 − δ12 )(||x − ai ||2 − δ22 )
Claim: the polynomials fi are linearly independent over R.

Assume that for real λi
m
X
λi fi (x) = 0
i=1
and substitute aj for x! All but the j-th term vanishes and the j-th term is λj (δ1 δ2 )2
which can be zero only if λj = 0. Since j was arbitrary, the claim is proved.
On the other hand, looking at the definition of fi , all fi can be written as a linear
combination of polynomials of five types:
Ã n !2 Ã n !
X X
x2k ; x2k xj ; xi xj ; xi ; 1
k=1 k=1
thus (adding the number of these polynomials type-to-type) all fi belong to a linear
space of dimension at most
n(n + 1) (n + 1)(n + 4)
1+n+ +n+1=
2 2
so their number, m, can not exceed the claimed value. ¤
Cross-intersecting hypergraphs. Assume that a p-uniform hypergraph (V, E) and a

q-uniform hypergraph (V, F) have the same vertex set V = [n] and both hypergraphs
have m edges, E = {e1 , e2 , . . . , em }, F = {f1 , f2 , . . . , fm }. We call the pair (E, F) cross-
intersecting if ei ∩ fj = ∅ if and only if i = j.
49
¡p+q¢
Theorem 5.6 If (E, F) is cross intersecting then m ≤ p
.
Proof. We say that an ordering of [n] is compatible with the pair (ei , fi ) if all elements
of ei precede all elements of fi in the ordering. There are
µ ¶
n n!
p!q!(n − p − q)! = ¡p+q¢
p+q p
orderings compatible with a pair (ei , fi ) and the cross-intersecting property implies that
each ordering of [n] is compatible with at most one pair (ei , fi ). Therefore
n!
m ¡p+q¢ ≤ n!
p
proving the theorem. ¤
There are two questions to answer. Why do we look at cross-intersecting hyper-

graphs? The answer is that they come along in many extremal problems. An interesting
application is given in exercise [5.4]. The second question is why do we have this nice
proof here, why not in section 4.1 (PROOFS BY COUNTING)? Because there is another
proof, by Lovász, which uses the dimension bound and gives a stronger theorem.
Represent the vertices of the cross-intersecting pair (E, F) by vectors of Rp+1 placed
in general position, i.e. so that any p + 1 of them are linearly independent. The vector
associated with v ∈ V is denoted by
s(v) = (s0 (v), s1 (v), . . . , sp (v)) ∈ Rp+1
For i = 1, 2, . . . , , m define the polynomial

Y
gi (x) = s(v)x
v∈fi
(product of inner products) which is a homogeneous polynomial of degree q in p + 1

¡ ¢
variables. The dimension of this linear space is d = p+q
p
(exercise [5.5]).
The vectors corresponding to the vertices of ej generate a subspace Aj of dimension
p, let aj be a nonzero vector orthogonal to Aj . Since the vectors s(v) are in general
position, s(v) ∈ Aj if and only if v ∈ ej . This means that gi (aj ) = 0 if and only if ej
50
and fi have non-empty intersection. Since we have cross-intersecting hypergraphs, this
means
(
0 if i 6= j
gi (aj ) =
6= 0 if i = j
and this condition implies that the polynomials gi are linearly independent (exercise
[5.6]). Therefore their number, m, is at most the dimension d, proving theorem 5.4. In
fact, the linear independence of the gi -s follow from the weaker assumption
(
0 if i < j
gi (aj ) =
6= 0 if i = j
(exercise [5.7]) and it gives a generalization of Theorem 5.4. for so called skew cross-
intersecting hypergraphs.
5.2 HOMOGENEOUS SYSTEMS OF LINEAR EQUATIONS

Partitioning into complete bipartite graphs
Assume that we want to partition the edge set of the complete graph Kn into as
few complete bipartite graphs as possible. One possibility is to decompose it into n − 1
stars. One can also try to halve the vertex set, take a complete bipartite graph between
these parts and continue. This works also but again gives n − 1 parts. After some futile
attempts to beat these constructions, one feels that perhaps n − 1 is the best. This is
true but it is very surprising that presently nobody can prove it without using linear
algebra.
Theorem 5.7 (Graham-Pollack) If the complete graph Kn is partitioned into m (edge

disjoint) complete bipartite graphs then m ≥ n − 1.
Proof. (Tverberg) Assume that [n] is the vertex set of Kn and we have a partition into
m complete bipartite graphs in the form [Ai , Bi ], where i ∈ [m] and the edges of the
51
complete bipartite graphs are between Ai and Bi . Consider n variables xj (j ∈ [n]) and
define 2m linear forms as follows:
X X
Li = xj and Mi = xj
j∈Ai j∈Bi
The condition that the complete bipartite graphs form edge disjoint partition is trans-
lated as
X
xi xj = L1 M1 + L2 M2 + . . . + Lm Mm (4)
1≤i<j≤n
Assume indirectly that m ≤ n − 2. Then the system of m + 1 homogeneous linear

equations with n variables
Li = 0 for i = 1, 2, . . . , m
and
x1 + x2 + . . . + xn = 0
has a nontrivial solution a1 , a2 , . . . , an because m + 1 < n. We get a surprising

contradiction:
X
0 < a21 + a22 + . . . + a2n = (a1 + a2 + . . . + an )2 − 2 ai aj = 0
1≤i<j≤n
Discrepancy of hypergraphs.
The discrepancy of a hypergraph is a measure of the most balanced bipartition of
the vertex set. Let Ψ denote a function from V to {−1, +1} and let H = (V, E) be a
hypergraph. Then the discrepancy of H is defined as follows.
X
disc(H) = minΨ maxe∈E | Ψ(x)|
x∈e
Theorem 5.8 (Beck-Fiala) Assume H is a hypergraph with maximum degree t. Then

disc(H) ≤ 2t − 1.
Proof. Assume V = [n], E = {e1 , e2 , . . . , em }. We allow Ψ to take real values from the
interval [−1, +1]. A vertex x is called bad for Ψ if Ψ(x) ∈ (−1, +1). An edge ej is called
52
bad for Ψ if it has more than t bad vertices. We shall define a procedure to get a Ψ
with no bad edges and at each step the following condition will be preserved for all bad
edges:
X
Ψ(x) = 0 (5)
x∈ej
The procedure starts with Ψ ≡ 0 which clearly satisfies (5). For the general step,
assume that there exist bad edges e1 , e2 , . . . , er for Ψ and let 1, 2, . . . , s be the bad vertices
of H for Ψ.
Claim: r < s. (Exercise [5.9])
Consider the unknowns y1 , y2 , . . . , ys . Using the claim, the system of r homogeneous
linear equations
X
yi = 0
i∈ej
has a non-trivial solution y1 , y2 , . . . , ys . Select λ so that for all i, 1 ≤ i ≤ s the values

Ψ(i) + λyi are all in [−1, +1] and at least one of them becomes +1 or −1. Then a new
Ψ can be defined changing the values Ψ(i) to Ψ(i) + λyi for all i, 1 ≤ i ≤ s. It is left as
exercise [5.10] to show that the required λ exists and the new Ψ satisfies condition (5).
The procedure eventually terminates with a Ψ with no bad edges because the set
{i ∈ V : Ψ(i) ∈ {+1, −1}} increases at each step. At this point the final Ψ is defined by
setting all values Ψ to 1 at the remaining bad vertices. It is easy to see (exercise [5.11])
P
that this final Ψ satisfies | x∈ej Ψ(x)| ≤ 2t − 1 for all j, 1 ≤ j ≤ m. ¤
5.3 EIGENVALUES
Regular graphs of girth five.
Consider an r-regular (simple) graph of girth five. At least how many vertices do we
need in such a graph? An arbitrary vertex is adjacent to r further vertices and all of
these are adjacent to r − 1 further vertices. Could it be possible that we double-counted
some vertices so far? The answer is no, thus we have at least r2 + 1 vertices in such a
graph. (Check this, [T5.4]). But is it possible to find r-regular graphs with girth five
53
which have only r2 + 1 vertices? For r = 2 the five-cycle, C5 , is a trivial example. For
r = 3 the famous Petersen graph is such an example.
Figure 5.1. The Petersen graph.
Unfortunately such beautiful graphs are rare as the next theorem shows.
Theorem 5.9 (Hoffman-Singleton) If G is an r-regular graph with girth five and has
r2 + 1 vertices then r ∈ {2, 3, 7, 57}.
Proof. Set n = r2 + 1, V (G) = [n] and consider the adjacency matrix A = [aij ] of G
defined as the n × n matrix
(
1 if ij is an edge of G
aij =
0 if ij is not an edge of G
Claim. A2 + A − (r − 1)In = Jn where In is the n × n identity matrix and Jn is the
n × n all-ones matrix.
To prove the claim, notice that the diagonal entries of A2 are r so the equation is
valid for the diagonal entries (the diagonal of A is zero). Assume i 6= j. If (ij) is not an
edge of G then there is precisely one vertex of G is adjacent to both i and j (no C4 in G)
therefore the ij entry in A2 is 1. If (ij) is an edge of G then no vertex of G is adjacent
to both i and j (no C3 in G) therefore the ij entry in A2 is 0. In both cases the claimed
matrix equation is correct.
54
By the principal axis theorem the symmetric matrix A has an orthogonal basis from
eigenvectors. Let f be the all-ones column vector with n elements. Then Af = rf hence
f is an eigenvector with eigenvalue r. Consider an eigenvector e orthogonal to f , by
definition, Ae = λe, fe = 0 (with inner product notation). Multiply the equation of the
claim by e from the right which gives
λ2 e + λe − (r − 1)e = 0
implying that λ2 + λ − (r − 1) = 0. Solving this equation we get that the possible

eigenvalues are
1 √
λ12 = (−1 ± 4r − 3).
2
Let mi be the multiplicity of λi (i ∈ [2]). The sum of the multiplicities of eigenvalues is
n, therefore
1 + m1 + m2 = n = r2 + 1 (6)
and the sum of the eigenvalues is the trace of A which gives
r + m1 λ1 + m2 λ2 = 0. (7)
Substituting the actual values of λi to (7) we get

√
2r − (m1 + m2 ) + (m1 − m2 )s = 0 where s= 4r − 3.
Using (6) this equation changes to
2r − r2 + (m1 − m2 )s = 0. (8)
If s is irrational then m1 − m2 = 0 implying r = 2. Otherwise s is a positive integer and

s2 +3
r= 4
. Plug this into (8) to get
s4 − 2s2 − 16(m1 − m2 )s − 15 = 0.
This equation implies that s is a divisor of 15, its possible values are 1, 3, 5, 15. Therefore
s2 +3
r= 4
∈ {1, 3, 7, 57} but r = 1 is impossible (girth is ∞). ¤
55
Self-test 5.
[T5.1] Prove Fisher inequality if some edge has λ vertices.
[T5.2] How do you get de Bruijn-Erdős theorem from Fisher inequality?
¡ ¢
[T5.3] Construct a two-distance set in R2 with n2 points.
[T5.4] Show that an r-regular graph of girth five has at least r2 + 1 vertices.
Exercise set 5.
[5.1] Assume that p is a prime and H = (V, E) is a hypergraph in which the edge sizes
are not divisible by p but each pair of edges has intersection size divisible by p. Prove
that |E| ≤ |V |.
[5.2] The same as exercise [5.1] but replace the prime by a power of prime. Hint: show
that the edges as row vectors are linearly independent in the linear space of n-component
vectors over the field of rational numbers .
[5.3] Show that a one-distance set of Rn has at most n + 1 points.
[5.4] A transversal of a hypergraph H = (V, E) is a set T ⊆ V such that T ∩ e 6= ∅ for
all e ∈ E. The transversal number of H is defined as
τ (H) = min{|T | : T is a transversal of H}
A Hypergraph H is called p-critical if τ (H) = p but τ (H − e) < p for each e ∈ E.

Prove (using Theorem 5.6) that a (p + 1)-critical t-uniform hypergraph has at most
¡p+t¢
t
edges. Is equality possible (for every p, t)?
[5.5] Prove that the linear space of homogeneous polynomials of degree q in p+1 variables
¡ ¢
has dimension p+q p
.
[5.6] Let S be an arbitrary set and F is a field. Consider the linear space of all functions
from S into F. Let ai ∈ S and gi are functions satisfying
(
0 if i 6= j
gi (aj ) =
6= 0 if i = j
(Diagonal criterion.) Prove that the functions are linearly independent.
[5.7] The same as [5.6] but from the weaker condition, called triangular criterion:
(
0 if i < j
gi (aj ) =
6= 0 if i = j
56
[5.8] Assume that we want to cover the edge set of the complete graph Kn by the edges of
complete bipartite graphs (allowing to cover an edge many times). How many bipartite
graphs do we need?
[5.9] Show that r < s at an arbitrary step in the procedure of the proof of Theorem 5.8.
[5.10] Prove that λ can be defined as required in the proof of Theorem 5.8 and condition
(*) is preserved for the new Ψ.
P
[5.11] Show that the final Ψ in the proof of Theorem 5.8 satisfies | x∈ej Ψ(x)| ≤ 2t − 1
for all j, 1 ≤ j ≤ m.
57
6 FRUIT SALAD
6.1 Distinct representatives of sets.

A system of distinct representatives (SDR) of a hypergraph with m edges e1 , e2 , . . . , em
is a set of distinct vertices v1 , v2 , . . . , vm such that vi ∈ ei for all i.
Theorem 6.1 A hypergraph has an SDR if and only if | ∪i∈S ei | ≥ |S| for all S ⊆ [m].
This important result can be formulated for bipartite graphs as follows.
Theorem 6.2 A bipartite graph [A, B] has a matching from A to B if and only if
|Γ(S)| ≥ |S| for every S ⊆ A. (Here Γ(S) denotes the set of vertices adjacent to some
vertex of S.)
Outline of a short proof by induction on |A|: if strict inequality holds for all nonempty
proper subsets S, delete an edge and apply induction; if there is equality for some S
then apply induction for the bipartite graphs [S, Γ(S)] and [A \ S, B \ Γ(S)]. ¤
The proof above does not provide a fast algorithm to find a matching or a subset
S ⊆ A violating the condition. Such a proof can be obtained by using the alternating
path method .
6.2 Symmetric chain decomposition.

¡ n
¢
The subsets of [n] can pe partitioned into bn/2c
chains. (Apply Theorem 6.2.) This gives
another proof for Sperner’s theorem (Theorem 4.1 in the handout). A nicer partition is
also possible: a chain in [n] is called symmetric if it contains sets of cardinality a, . . . , n−a
for some a, (0 ≤ a ≤ bn/2c).
Theorem 6.3 The subsets of [n] can be partitioned into symmetric chains.
Proof. Assume we have a SCD on [n − 1] and to each chain C associate C ∗ by adding

the element n to each member of C. Then remove the top element of C ∗ and place it as
the top element of C. ¤
58
6.3 A property of n sets on n elements.
A vertex deletion of a hypergraph means that we remove a vertex x from the vertex set
and each edge e is replaced by the edge e \ x. In other words, column x is removed from
the incidence matrix of the hypergraph.
Theorem 6.4 Asssume that a simple hypergraph has n vertices and n edges. Then there
is a vertex deletion which gives a simple hypergraph.
Proof. Assume that the edges are e1 , . . . , en and define a graph G with vertex set [n]
and with edges ij if and only if the symmetric difference of the edges ei and ej is a single
vertex x. If Theorem 6.4 is not true then we have n edges in G. However, G can not
contain any cycle - prove it! - thus we have a contradiction. ¤
6.4 A property of n + 1 sets on n elements.

Theorem 6.5 Assume that e1 , . . . , en+1 are nonempty edges of a hypergraph on vertex
set [n]. There exists nonempty disjoint subsets I1 , I2 of [n + 1] such that
∪i∈I1 ei = ∪i∈I2 ei
Proof. Consider the edges as 0 − 1 vectors in Rn . They are linearly dependent. This
gives easily the proof. ¤
6.5 A property of n + 2 sets on n elements.

Theorem 6.6 Assume that e1 , . . . , en+2 are nonempty edges of a hypergraph on vertex
set [n]. There exists nonempty disjoint subsets I1 , I2 of [n + 2] such that
∪i∈I1 ei = ∪i∈I2 ei
moreover
∩i∈I1 ei = ∩i∈I2 ei
59
Proof. Consider the vectors in R2n whose first n coordinates are determined by ei and
the last n by the complement of ei . These vectors are in an n + 1-dimensional subspace
of R2n therefore they are linearly dependent. Finish the proof. ¤
No proofs are known without linear algebra.
6.6 Critical hypergraphs.

Call a hypergraph critical if it is not 2-colorable but the edge set within any proper
subset of vertices is 2-colorable (like the Fano plane).
Theorem 6.7 A critical hypergraph has at least as many edges as vertices.
Proof. Assign real variables xj to vertices and consider the system of homogeneous
P
linear equations j∈ei xj = 0. ¤
The critical hypergraphs have a surprising connection with digraphs without even
cycles. Assume that D is a digraph, define the hypergraph H(D) as follows. The
vertices of H are the vertices of D and the edges of H are the sets x ∪ Γ+ (x) where x is
a vertex of D and Γ+ (x) is the ’outset’ of x.
Theorem 6.8 For a strong digraph D, H(D) is critical if and only if D has no even
directed cycles.
6.7 Sunflower theorem.

A sunflower with s petals is a hypergraph with s edges such that the intersection of any
two edges are the same set, called the kernel of the sunflower. Notice that the kernel
can be empty...
Theorem 6.9 An r-uniform hypergraph with more than r!(s − 1)r+1 edges contains a
sunflower with s petals.
Proof. Induction on r. Take a maximum family of pairwise disjoint edges, e1 , . . . , ek .

If k ≥ s we are done. Otherwise
| ∪ki=1 ei | ≤ r(s − 1)
60
and there exists a vertex of degree more than
r!(s − 1)r+1
= (r − 1)!(s − 1)r
r(s − 1)
and we are done by induction. ¤
6.8 Sum-free subsets of numbers.

Theorem 6.10 Every set B of n positive integers contain A ⊆ B such that |A| > n/3
and A is sum-free (the sum of any two elements of A is not in A).
Proof. Let p = 3k + 2 be a prime larger than any number in B = {b1 , . . . , bn }. Set

C = {k + 1, . . . , 2k + 1}. Then C is a sum-free subset of the cyclic group Zp . For each
bi , the set
Sbi = {xbi : x ∈ [p − 1]} (mod p) = [p − 1].
Since Sbi covers [3k + 1], their union for bi ∈ B covers C n(k + 1) times. Therefore for
(k+1)n
some x ∈ [p − 1] at least 3k+1
> n/3 bi satisfy xbi ∈ C (mod p) and these bi -s form a
sum-free subset A in B. ¤
Remarks. Theorem 6.1 is Hall’s theorem (1935). The SCD theorem (Theorem 6.3) is
from de Bruijn, Tengbergen and Kruyswijk (1951). Theorem 6.4 is from Bondy (1972),
Theorem 6.5 is from Tverberg (1971), Theorem 6.6 is from Lindstrom (1993), Theorems
6.7 and 6.8 are from Seymour (1974). The sunflower theorem (Theorem 6.9) is from
Erdős and Rado (1960) - the bound given is not sharp. In fact, it seems to be a very
difficult problem to decide whether for s=3 the bound can be improved to cr . The
complete r-partite hypergraph with two vertices in each partite class and with edge
multiplicity r − 1 shows that 2r (r − 1) edges do not give a sunflower with three petals.
Theorem 10 is from Erdős (1965). It is not known whether n/3 is the best bound, Alon
and Kleitman constructed a set B of n positive integers with no larger sum-free subset
than 12n/29.
61
7 ADVANCED MENU
7.1 Factorization of hypergraphs.

Assume that r is a divisor of n, a factorization of Knr is a partition of the edges into
¡n−1¢
r−1
parts so that each part consists of n/r pairwise disjoint edges. The case r = 1 is
trivial, the case r = 2 is easy: represent the vertices as a regular polygon (1, 2, . . . , n − 1)
and its center (n). The edges 1 + i, n + i for 0 ≤ i ≤ n/2 − 1 define a factor and one
can rotate it to get a factorization. The general solution is provided by the following
theorem (Baranyai, 1973).
Theorem 7.1 The complete r-uniform hypergraph Knr has a factorization.
Proof. The first step is the following rounding lemma of Baranyai:
Lemma 7.2 Let A be a nonnegative real matrix with integral row and column sums.
Then the elements of A can be replaced by their floor or ceiling so that all row and
column sums retain their values.
Proof. A non-integral entry (NIE) of A gives another NIE in the same row which gives
another NIE in the same column which gives another NIE in the same row, etc. This
procedure gives an alternating cycle of NIE-s along which we can replace the elements
to get at least one integral entry and keep all row and column sums. Then repeat the
procedure. ¤
¡n−1¢
Set M = r−1
(the number of factors) and m = n/r (the number of edges in each
factor). Define an m-partition of a set X as a partition of X into m parts, such that ∅
is allowed with multiplicities. Example: ∅, ∅, 1 is a 3-partition of [1].
Lemma 7.3 For every integer k ∈ [0, n] there exist m-partitions of [k], A1 , A2 , . . . , AM
¡ n−k ¢
with the following property (*): each subset S ⊆ [k] occurs in exactly r−|S| partitions.
Example: n = 6, r = 2, m = 3, M = 5, k = 2 then A1 = A2 = A3 = A4 = {∅, 1, 2},

¡ ¢
A5 = {∅, ∅, {1, 2}}. Then S = ∅ occurs 6−2 = 6 times, S = {1} and S = {2} occurs
¡6−2¢ ¡6−2¢ 2−0
2−1
= 4 times, S = {1, 2} occurs 2−2 = 1 times.
62
Observe that Lemma 7.3 gives the theorem: for k = n it says that S ⊆ [n] occurs in
¡ n−n ¢
exactly r−|S| partitions - this number is zero except for |S| = r when it is one!!
Lemma 7.3 is proved by induction on k. The base is k = 0 when all partitions
consist of m empty sets. Assume that for some k < n the m-partitions A1 , A2 , . . . , AM
satisfy (*). Define matrix A as an M × 2k matrix where the rows are associated with
the partitions the columns are associated with the 2k subsets of [k]. The entry of A
associated to row Ai and column S is 0 if S ∈
/ Ai , otherwise it is
r − |S|
n−k
r−|S|
except when S = ∅: in that case the entry n−k
is taken with the multiplicity of S in Ai .
In the example above, A is a 5 × 4 matrix, the columns are associated to ∅, 1, 2, {1, 2}.
The first four rows are [1/2, 1/4, 1/4, 0] and the fifth row is [1, 0, 0, 0] because ∅ has
multiplicity two in A5 .
The row sums of A:
X r − |S| 1 X
= (mr − |S|) = 1
S∈A
n − k n − k S∈A
i i
and the column sums of A:

X r − |S| µ ¶ µ ¶
r − |S| n − k n−k−1
= = .
n−k n − k r − |S| r − |S| − 1
{i:S∈Ai }
Using Lemma 7.2, round A to integral B keeping row and column sums. Since the row
sums are all ones, there is precisely one 1 in each row of B, this corresponds to a block
Si ∈ Ai .
Change A1 , A2 , . . . , AM by replacing Si with Si ∪ {k + 1} (and keep all other blocks
unchanged). This defines A∗1 , A∗2 , . . . , A∗M and we show that they satisfy (*) for k + 1.
Assume S ∗ ⊆ [k + 1].
Case 1. k + 1 ∈ S ∗ . Now S ∗ appeared 0 times in the Ai -s and appears in the A∗i -s as
many times as the column sum of B which is equal to the column sum of A thus
µ ¶ µ ¶
n−k−1 n−k−1
=
r − (|S ∗ | − 1) − 1 r − |S ∗ |
63
showing property (*) for k + 1.
¡ n−k
¢
/ S ∗ . Now S ∗ occurred
Case 2. k + 1 ∈ r−|S ∗ |
times in the Ai -s (inductive hypothesis)
but disappeared as many times as extended by k + 1 i.e. by column sum of B which is
equal to column sum of A thus
µ ¶ µ ¶ µ ¶
n−k n−k−1 n−k−1
− =
r − |S ∗ | r − |S ∗ | − 1 r − |S ∗ |
showing property (*) for k + 1.

This proves Lemma 7.3 and Theorem 7.1. ¤
7.2 Normal hypergraphs and perfect graphs.

A partial hypergraph of H = (V, E) is a hypergraph H0 = (V, E 0 ) where E 0 ⊆ E. Thus
partial hypergraphs are defined by removing edges. The chromatic index of a hypergraph
is the minimum number of colors needed to color the edges so that each color class has
pairwise disjoint edges. (Such edge colorings are called proper.)
The chromatic index is denoted by q(H). Clearly, q(H) ≥ ∆(H) for every hypergraph
(∆ is the maximum degree). Example: for the Fano plane q = 7 and ∆ = 3.
A hypergraph is called normal if q = ∆ for every partial hypergraph. The following
lemma of Lovász captures an important property of normal hypergraphs.
Lemma 7.4 Assume that e is an edge of a normal hypergraph H. Form a new hyper-
graph H+ by adding a new copy of e. Then H+ is also a normal hypergraph.
Proof. We prove by induction on the number of edges of H. The only problem is to

prove q = ∆ for H+ . Let e+ be the new edge. If e contains a vertex of maximum degree
in H then
∆(H+ ) = ∆(H) + 1, q(H+ ) = q(H) + 1
and the proof is done. Otherwise consider a proper coloring of H with ∆(H) = t colors,
assume that e is red. Remove all red edges and add e+ . This is a partial hypergraph of
H with maximum degree t − 1 so there is a proper coloring with t − 1 colors. Adding
64
back the red edges, we have a proper coloring of H+ with ∆(H+ ) = t colors and the
lemma is proved.
The next lemma shows that normal hypergraphs behave nicely for another pair of
parameters. Let ν(H) denote the maximum number of pairwise disjoint edges in a
hypergraph H (matching number) and let τ (H) denote the transversal number, the
minimum number of vertices needed to intersect all edges of H (see Exercise 5.4 in
handout). The obvious inequality τ ≥ ν becomes equality for normal hypergraphs.
Lemma 7.5 For normal hypergraphs τ = ν.
Proof. Assume that H is normal, set t = ν(H). We prove by induction on t that

τ (H) = t. Let Ex denote the set of edges of H incident to vertex x.
Base: t = 1. Since H is intersecting, q equals to the number of edges, which by
normality equals to the maximum degree. Thus τ (H) = 1.
Inductive step. We prove that there exists a vertex x for which ν(H \ Ex ) < t, then
induction proves the theorem.
Assume indirectly that for each vertex x, ν(H \ Ex ) = t. Assume that H has n
vertices x1 , x2 , . . . , xn . For each xi let Fi be a maximum matching of H \ Ex . Consider
the hypergraph
H 0 = F1 ∪ . . . ∪ F n
and note that

(a) H0 is normal by Lemma 7.4.
(b) H0 has tn edges.
(c) ∆(H0 ) < n.
In a proper coloring of H0 each color class has at most t edges thus, by (b), q(H0 ) ≥ n.
However, by (c) q(H0 ) 6= ∆(H0 ) which contradicts (a) and the proof of Lemma 7.5 is
finished ¤
Intersection graphs of hypergraphs. Every simple graph is the intersection graph

of a suitable hypergraph (exercises 1.1 and 1.2 in the handout). For a given (simple)
graph G, the dual of G is a suitable hypergraph. This is a special case of the following
65
construction. Assume that on the vertex set of G we have a hypergraph H whose hyper-
edges span complete subgraphs in G and each edge of G is covered by some hyperedge
of H. It is easy to see that the dual of H is a suitable hypergraph, i.e. the intersection
graph of H ∗ is G. The cited special case is when the edge set of G defines H. Another
(not very economic) choice of H is to select all complete subgraphs of G as hyperedges.
The best choice is to select the minimum number of complete subgraphs needed to cover
the edge set of G. However, as we shall see, there is another useful choice for H called
the clique-hypergraph of G: the edges of H are the complete subgraphs of G maximal
for inclusion.
Example. Consider the graph G obtained from a triangle T = K3 by adding three
new vertices, each adjacent to a different pair of vertices of T . (G has six vertices and
nine edges.) The dual of G has 9 vertices, the dual of all complete subgraphs of G
has 19 vertices, the dual of a minimum cover of G has 3 vertices and the dual of the
clique-hypergraph of G has 4 vertices.
The advantage of the clique-hypergraph is that its dual has the Helly-property: pair-
wise intersecting edges have a nonempty intersection.
The perfect graph theorem. Let ω(G) denote the number of vertices in the largest
complete subgraph of G. A graph is called perfect if χ = ω for all induced subgraphs.
The perfect graph theorem (PGT) is the following result of Lovász:
Theorem 7.6 A graph is perfect if and only if its complement is perfect.
Proof. Assume that G is perfect, i.e. ω = χ for all induced subgraphs. This implies
that ∆ = q for all partial hypergraphs of H ∗ , the dual of the clique hypergraph H of G -
in other words, H ∗ is normal (Helly property is needed!). Lemma 15 gives that ν = τ for
all partials of H ∗ which gives that ω = χ for all induced subgraphs of the complement
of G. ¤
Remark. Lovász’s proof of PGT is from 1972. It was a conjecture of Berge (1960)
who also had a stronger conjecture, the perfect graph conjecture (PGC) claiming that a
graph is perfect if and only if it does not contain odd cycles of length at least five and
66
their complements (as induced subgraphs!). The PGC is recently proved by Chudnovsky,
Robertson, Seymour and Thomas. However, the proof is very long structural analysis
and definitely does not make the PGT obsolete. An introduction to the subject is
Golumbic: Algorithmic Graph Theory and Perfect graphs (1980).
7.3 Constructive super-polynomial lower bound for Ramsey

numbers.
The idea of the cubic lower bound of R(n) (Proposition 5.1) can be generalized as follows.
The next theorem is a variant of a result of Ray-Chaudhuri and Wilson (1969, published
proof in 1975). The theorem is from Deza,Frankl and Singhi (1983). The simple proof
is from Alon, Babai and Suzuki (1991).
Theorem 7.7 Assume that p is a prime, L is a set of s integers, and we have a hyper-
graph with n vertices and with m edges e1 , e2 , . . . , em satisfying the following conditions:
for each i, |ei | ∈
/L (mod p); for each i, j, i 6= j, |ei ∩ ej | ∈ L (mod p). Then
µ ¶ µ ¶ µ ¶
n n n
m≤ + + ... + .
0 1 s
Proof. The claimed upper bound is the dimension of polynomials in n variables with
degree at most s which are linear in all variables (products of at most s distinct variables
form a basis). The proof idea is to define m linearly independent polynomials of that
type. Apart from a tiny technical problem, the following definition works:
Y
fi (x) := (xei − l)
l∈L
where ei is the 0 − 1 incidence vector of the edge ei , xei is the inner product (mod p).
Linear independence follows by substituting ej to fi and applying the diagonal criterion
(exercise 5.6). The tiny problem is that to get linear polynomials, the domain of fi must
be restricted to {0, 1}n . ¤
Theorem 17 gives a super-polynomial lower bound for the Ramsey number as follows.
Let p be a prime and n > 2p2 . Associate the vertices of the complete graph K with the
67
p2 − 1-element subsets of an n-element set. Color an edge of K red if the corresponding
subsets has intersection size not congruent to −1 (mod p) otherwise color it blue. Apply
Theorem 17 with L = {0, 1, 2, . . . , p − 2} for the red color. Then apply Theorem 17 with
a prime q > p2 − 1 and with L = {p − 1, 2p − 1, . . . , p2 − p − 1} for the blue color.
7.4 Hypergraphs in geometry: the fall of Borsuk conjecture.

For any prime p define the graph Gp as follows. The vertices of Gp are the 2p − 1-
element subsets of a 4p − 1-element set. Two vertices of G are adjacent if and only if
the corresponding subsets intersect in precisely p − 1 elements.
Lemma 7.8 For the graph Gp the following holds:
α(Gp ) < 1.75484p−1 , χ(Gp ) > 1.13974p−1
Proof. Theorem 7.7 is applied with L = {0, 1, . . . , p − 2}. If S is a set of independent

vertices in Gp then from Theorem 7.7 we get
p−1 µ ¶ µ ¶
X 4p − 1 4p − 1
|S| ≤ ≤2 ≤ 1.75484p−1
i=0
i p−1
(the last step uses Sterling formula). The inequality for the chromatic number follows
|V (Gp )| 24p−1
from χ(Gp ) ≥ α(Gp )
≥ 1.7548p−1
. ¤
Borsuk conjectured (in 1933) that every set of Rd with diameter one can be parti-
tioned into d + 1 sets of diameter less than one. The conjecture was proved for d ≤ 3
and for all d if the set is special (centrally symmetric, have smooth boundary). Let f (d)
be the minimum for which every set of diameter one can be partitioned into f (d) sets
of diameter smaller than one. Borsuk’s conjecture was f (d) ≤ d + 1. This fails badly as
proved by Kahn and Kalai in 1992:
√
2d
Theorem 7.9 f (d) > 1.1397
¡4p−1¢
Proof. Let p be a prime, d = 2
. Let H be a set of 4p−1 elements. The hypergraph
KK(d) is defined as follows.
68
The vertex set of KK(d) is formed by the d pairs of elements of H. Formally
V (KK) = {axy : x, y ∈ H}
The edges eZ of KK are associated with 2p − 1-element subsets Z of H as follows:
eZ = {axy : x ∈ Z, y ∈
/ Z}
thus eZ is defined by the pairs of H split by Z.

¡ ¢
Notice that KK(d) has d vertices, 4p−12p−1
edges and (2p − 1)2p-uniform. Represent
KK(d) in Rd with the rows of its incidence matrix. Observe that the distance of the
points representing eZ1 and eZ2 is
p
2((2p − 1)2p − |eZ1 ∩ eZ2 |)
therefore the maximum distance, i.e. the diameter of KK(d) in Rd is realized if and
only if
|eZ1 ∩ eZ2 | = min{|eZ ∩ eZ 0 | : Z, Z 0 ⊆ H} = λ
Claim : |eZ1 ∩ eZ2 | = λ if and only if |Z1 ∩ Z2 | = p − 1.

The claim follows from observing that |eZ1 ∩ eZ2 | is the number of pairs of H split
by both Z1 and Z2 . Thus, with r = |Z1 ∩ Z2 |,
λ = min{r(r + 1) + (2p − 1 − r)2 : 0 ≤ r ≤ 2p − 1}
and it is easy to check that r = p − 1 minimizes the expression.

Using the claim, the partition of KK(d) in Rd into sets of smaller diameter is equiv-
alent with partitioning the 2p − 1-element sets Z in H into classes so that no class has
two sets intersecting in p − 1 elements. This is at least χ(Gp ) so we get
√
f (d) ≥ χ(Gp ) > 1.13974p−1 > 1.1379 2d
.
69

Ujhandout

Uploaded by

Copyright:

Available Formats

Ujhandout

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ujhandout

Uploaded by

Copyright:

Available Formats

Budapest Semesters in Mathematics

ADVANCED COMBINATORICS HANDOUTS

1. BASIC NOTIONS AND EXAMPLES

1.1. Definitions. (Incidence matrices, Duality.)

1.2. Examples of hypergraphs. (Paths and cycles, Linear spaces.)

2. CHROMATIC NUMBER AND GIRTH

2.1. Chromatic number.

2.3. How to glue hypergraphs to get graphs? ( Nesetril-Rödl hypergraph.)

3. A LOOK AT RAMSEY THEORY

3.1. Ramsey numbers. (Pigeonholes and parties, Existence of Ramsey numbers,

3.2. Van der Waerden numbers.

3.3. Tic-tac-toe and Hales-Jewett theorem.

4.1. Proofs by counting. (Antichains, intersecting hypergraphs, 3-chromatic uniform

4.4. Jokes (Triangle is 2-chromatic, Spencer’s injections.)

5. LINEAR ALGEBRA METHODS

5.2. Homogeneous linear equations. (Partition of Kn , discrepancy.)

5.3. Eigenvalues. (Regular graphs of girth five.)

Figures 1.1. and 1.2. Drawings of hypergraphs

Proof. Interpret the statement for the incidence matrix of H. ¤

Duality. Assume that H is a hypergraph with incidence matrix M . Let M T be the

x∈e f (x) 3 g(e)

Proof. Exercise 1.1 ¤

1.2 EXAMPLES OF HYPERGRAPHS

Figure 1.3. Paths and Cycles

Figure 1.6. Drawings of linear spaces up to five vertices

Theorem 1.4 Assume H = (V, E) is an intersecting linear space not isomorphic to a

|E| = |V | = k 2 + k + 1, H is (k + 1)-uniform and (k + 1)-regular . (1)

2.1 CHROMATIC NUMBER

Theorem 2.1 Every hypergraph H = (V, E) satisfies

Proof. Exercise 2.2 ¤

2.2 GRAPHS FROM THE HALL OF FAME

Figure 2.1. Zykov graphs.

Figure 2.2. Mycielski graphs.

Theorem 2.3 The Mycielski graph Mi has no triangles and χ(Mi ) = i.

Proof. Exercise 2.8 ¤

Figure 2.3. Tutte graphs.

Theorem 2.4 The Tutte graph Ti is of girth 6 and χ(Ti ) = i.

Figure 2.4. The Shift graph SH5 .

Theorem 2.6 (Lovász) It is impossible to partition the n-element subsets of a 2n + k

What is the significance of s? We plan to construct an s-partite [p, k, n]-hypergraph.

Theorem 2.7 The hypergraph Hs+1 is an s-partite [p, k, n]-hypergraph.

3.1 RAMSEY NUMBERS

Existence of Ramsey numbers. It is usual to consider partitions as colorings, in case

Upper bound on R(n).

Theorem 3.3 R(n) ≤ 4n .

Theorem 3.4 R(p, q) ≤ R(p − 1, q) + R(p, q − 1).

Proof. Exercise [3.10] ¤

Proof. Exercise [3.11] ¤

Exact values of Ramsey numbers. It is known that R(3, 4) = 9 and R(4) = 18

3.2 VAN DER WAERDEN NUMBERS

Figure 3.1. RR? and BB?

and Exercise [3.14] asks to work out the details.

Figure 3.2. Focusing three copies of medium blocks.

Hopefully the presented argument is convincing enough to accept

3.3 TIC-TAC-TOE AND HALES-JEWETT THEOREM

Corollary 3.10 Van der Waerden theorem

Proof. Exercise [3.15] ¤

Rt+1 (n1 , n2 ) ≤ Rt (Rt+1 (n1 − 1, n2 ), Rt+1 (n1 , n2 − 1)) + 1

4.1 PROOFS BY COUNTING