NDMI011: Combinatorics and Graph Theory 1

Lecture #10
Sperner’s theorem. Ramsey numbers
Irena Penev

1 Sperner’s theorem
For a partially ordered set (X, ≤),

• a chain in (X, ≤) is any set C ⊆ X such that for all x1 , x2 ∈ C, we have

that either x1 ≤ x2 or x2 ≤ x1 .1

• a maximal chain in (X, ≤) is a chain C in (X, ≤) such that there is no

chain C 0 in (X, ≤) with the property that C $ C 0 ;

• an antichain in (X, ≤) is any set A ⊆ X such that for all distinct

x1 , x2 ∈ A, we have that x1 6≤ x2 and x2 ≤
6 x1 .

Note that a chain and an antichain in (X, ≤) can have at most one element
in common.2
Here, we are interested in a special case of the above. As usual, for a
set X, we denote by P(X) the power set (i.e. the set of all subsets) of X.
Clearly, for any set X, ⊆P(X) := {(A, B) | A, B ∈ P(X), A ⊆ B} is a partial
order on X. To simplify notation, in what follows, we write (P(X), ⊆)
instead of (P(X), ⊆P(X) ). We apply the above definitions to (P(X), ⊆),
as follows.
For a set X,
1
This definition works both for finite and for infinite X. Note also that ∅ is a chain
in (X, ≤). However, if X is finite and C is a non-empty chain in (X, ≤), then C can be
ordered as C = {x1 , . . . , xt } so that x1 ≤ · · · ≤ xt .
2
Indeed, if distinct elements x1 , x2 belong to a chain of (X, ≤), then x1 ≤ x2 or x2 ≤ x1 .
On the other hand, if they belong to an antichain of (X, ≤), then x1 6≤ x2 and x2 6≤ x1 .
So, distinct elements x1 and x2 cannot simultaneously belong to a chain and an antichain
of (X, ≤).

1
 • a chain in (P(X), ⊆) is any set C of subsets of X such that for all
C1 , C2 ∈ C, we have that either C1 ⊆ C2 or C2 ⊆ C1 .3

• a maximal chain in (P(X), ⊆) is a chain in (P(X), ⊆) such that there

is no chain C 0 in (P(X), ⊆) with the property that C $ C 0 ;

• an antichain in (P(X), ⊆) is any set A of subsets of X such that for

6 A1 .4
all distinct A1 , A2 ∈ A, we have that A1 6⊆ A2 and A2 ⊆

As before, note that a chain and an antichain in (P(X), ⊆) can have at most
one element in common.

Example 1.1. Let X = {1, 2, 3, 4}. The following are chains in (P(X), ⊆):5

• {{2, 4}, {1, 2, 4}};6

• {∅, {1}, {1, 2}, {1, 2, 3}, X}.7

• {∅, {4}, {2, 4}, {1, 2, 4}, X};8

Further, the following are all antichains in (P(X), ⊆):9

• {∅};

• {X};

• {{1, 2}, {2, 3}, {1, 3, 4}};

• {{1, 2}, {1, 3}, {1, 4}, {2, 3}, {2, 4}, {3, 4}}.

Sperner’s theorem. Let n be a non-negative integer, and let X be
an

n
n-element set. Then any antichain in (P(X), ⊆) has at most bn/2c ele-
ments. Furthermore, this bound is tight, that is, there exists an antichain in
n


(P(X), ⊆) that has precisely bn/2c elements.

Proof. First, we note that the set of all bn/2c-element subsets
of X is an
n
antichain in (P(X), ⊆), and this antichain has precisely bn/2c elements.
n


It remains to show that any antichain in (P(X), ⊆) has at most bn/2c
elements.
3
This definition works both for finite and for infinite X. Note also that ∅ is a chain in
(P(X), ⊆). However, if X is finite and C is a non-empty chain in (P(X), ⊆), then C can
be ordered as C = {C1 , . . . , Ct } so that C1 ⊆ · · · ⊆ Ct .
4
Equivalently: A1 \ A2 and A2 \ A1 are both non-empty.
5
There are many other chains in (P(X), ⊆) as well.
6
Note that this chain is not maximal, since we can add (for example) the set {2} to it
and obtain a larger chain.
7
This chain is maximal.
8
This chain is maximal.
9
There are many other antichains in (P(X), ⊆) as well.

2
 Claim 1. There are precisely n! maximal chains in (P(X), ⊆).
Proof of Claim 1. Clearly, any maximal chain in (P(X), ⊆) is of the form
{∅, {x1 }, {x1 , x2 }, . . . , {x1 , x2 , . . . , xn }}, where x1 , . . . , xn is some ordering of
the elements of X. There are precisely n! such orderings, and so the number
of maximal chains in (P(X), ⊆) is n!. 
Claim 2. For every set A ⊆ X, the number of maximal chains
of (P(X), ⊆) containing A is precisely |A|!(n − |A|)!.
Proof of Claim 2. Set k = |A|. As in the proof of Claim 1, we have that any
chain in (P(X), ⊆) is of the form {∅, {x1 }, {x1 , x2 }, . . . , {x1 , x2 , . . . , xn }},
where x1 , . . . , xn is some ordering of the elements of X; this chain contains
A if and only if A = {x1 , . . . , xk } (and therefore, X \ A = {xk+1 , . . . , xn }).
The number of ways of ordering A is k!, and the number of ways of ordering
X \ A is (n − k)!. So, the total number of chains of (P(X), ⊆) containing A
is precisely k!(n − k)!. 
Now, fix an antichain A in (P(X), ⊆). We form the matrix M whose
rows are indexed by the elements of A, and whose columns are indexed by
the maximal chains of (P(X), ⊆), and in which the (A, C)-th entry is 1 if
A ∈ C and is 0 otherwise.10 Our goal is to count the number of 1’s in the
matrix M in two ways.
First, by Claim 2, for any A ∈ A, the number of maximal chains of
(P(X), ⊆) containing A is precisely |A|!(n − |A|)!; so, the number of 1’s in
the row of M indexed by A is precisely |A|!(n − |A|)!. Thus, the number of
1’s in the matrix M is precisely
X
|A|!(n − |A|)!.
A∈A
On the other hand, by Claim 1, the number of columns of M is precisely n!.
Furthermore, no chain of (P(X), ⊆) contains more than one element of the
antichain A, and so no column of M contains more than one 1. So, the total
number of 1’s in the matrix M is at most n!. We now have that
P
|A|!(n − |A|)! ≤ n!,
A∈A
and consequently,
P |A|!(n−|A|)!
n! ≤ 1.
A∈A
On the other hand, for all A ⊆ X (and in particular, for all A ∈ A), we have
that
(∗)
|A|!(n−|A|)! 1
1 1
n! = = n
n! ≥ n ,
(|A|)
|A|!(n−|A|)! (bn/2c )
where (*) follows from the fact that nk ≤ bn/2c
n
for all k ∈ {0, . . . , n}.11

10
Here, A ∈ A, C is a maximal chain in (P(X), ⊆), and the (A, C)-th entry of M is the
entry in the row indexed by A and column indexed by C.
11
See subsection 2.2 of Lecture Notes 1.

3
We now have that
P |A|!(n−|A|)! P 1 1
1 ≥ n! ≥ n ≥ |A| n ,
A∈A A∈A (bn/2c) (bn/2c)

n


which yields |A| ≤ bn/2c . This completes the argument.

2 The Pigeonhole principle

The Pigeonhole Principle. Let n1 , . . . , nt (t ≥ 1) be non-negative integers,
and let X be a set of size at least 1 + n1 + · · · + nt . If (X1 , . . . , Xt ) is any
partition of X,12 then there exists some i ∈ {1, . . . , t} such that |Xi | > ni .13

Proof. Suppose otherwise, and fix a partition (X1 , . . . , Xt ) such that |Xi | ≤
ni for all i ∈ {1, . . . , t}. But then

1 + n1 + · · · + nt ≤ |X| = |X1 | + · · · + |Xt | ≤ n1 + · · · + nt ,

a contradiction.

As an immediate corollary, we obtain the following.

Corollary 2.1. Let n and t be positive integers. Let X be an n-element

set, and let (X1 , . . . , Xt ) be any partition of X.14 Then there exists some
i ∈ {1, . . . , t} such that |Xi | ≥ d nt e.

Proof. By the Pigeonhole Principle, we need only show that n ≥ 1+t(d nt e−1).
If t | n,15 then d nt e = nt , and we have that

1 + t(d nt e − 1) ≤ 1 + t( nt − 1) = n − t + 1 ≤ n,

which is what we needed. Suppose now that t 6 | n, so that d nt e − 1 = b nt b.

Then let m = b nt c and ` = n − mt; since t 6 | n, we have that ` ≥ 1. But now

1 + t(d nt e − 1) = 1 + t(b nt c) = 1 + tm ≤ ` + tm ≤ n,

and we are done.

We remark that Corollary 2.1 is also often referred to as the Pigeonhole

Principle.
12
Here, we allow the sets X1 , . . . , Xt to possibly be empty.
13
If one thinks of elements of X as “pigeons” and sets X1 , . . . , Xt as “pigeonholes,” then
the Pigeonhole Principle states that some pigeonhole Xi receives more than ni pigeons.
14
Here, we allow the sets X1 , . . . , Xt to possibly be empty.
15
“t | n” means that n is divisible by t.

4
3 Ramsey numbers
A clique in a graph G is any set of pairwise adjacent vertices of G. The clique
number of G, denoted by ω(G), is the maximum size of a clique of G.
A stable set (or independent set) in a graph G is any set of pairwise
non-adjacent vertices of G. The stability number (or independence number)
of G, denoted by α(G), is the maximum size of a stable set in G.
Proposition 3.1. Let G be a graph on at least six vertices. Then either
ω(G) ≥ 3 or α(G) ≥ 3.
Proof. Let u be any vertex of G. Then |V (G) \ {u}| ≥ 5, and so (by the
Pigeonhole Principle) either u has at least three neighbors or it has at least
three non-neighbors.
Suppose first that u has at least three neighbors. If at least two of those
neighbors, say u1 and u2 , are adjacent, then {u, u1 , u2 } is a clique of G of
size three, and we deduce that ω(G) ≥ 3. On the other hand, if no two
neighbors of u are adjacent, then they together form a stable set of size at
least three, and we deduce that α(G) ≥ 3.
Suppose now that u has at least three non-neighbors. If at least two of
those non-neighbors, say u1 and u2 , are non-adjacent, then {u, u1 , u2 } is a
stable set of G of size three, and we deduce that α(G) ≥ 3. On the other
hand, if the non-neighbors of u are pairwise adjacent, then they together
form a clique of size at least three, and we deduce that ω(G) ≥ 3.

For a graph G and a vertex u, NG (u) is the set of all neighbors of u in

G, and NG [u] = {u} ∪ NG (u).
Theorem 3.2. Let k and ` be positive integers, and let G be a graph on at
least k+`−2 16 Then either ω(G) ≥ k or α(G) ≥ `.


k−1 vertices.
Proof. We may assume inductively that for all positive integers k 0 , `0 such
0 0 −2
that k 0 + `0 < k + `, all graphs G0 on at least k k+`


0 −1 vertices satisfy either
0 0 0
ω(G ) ≥ k or α(G ) ≥ ` . 0

If k = 1 or ` = 1, then the result 17 So, we may assume

is immediate.
that k, ` ≥ 2. Now, set n = k−1 , n1 = k−1 , and n2 = k+`−3
k+`−2 k+`−3



k−2 ; then
n = n1 + n2 , and consequently, n − 1 = 1 + (n1 − 1) + (n2 − 1). Fix any
vertex u ∈ V (G), and set N1 = V (G) \ NG [u] and N2 = NG (u).
u

N1 N2
16
Note that k−1 = k+`−2
k+`−2



`−1
.
17
Indeed, it is clear that ω(G) ≥ 1 and α(G) ≥ 1. So, if k = 1, then ω(G) ≥ k; and if
` = 1, then α(G) ≥ `.

5
Since (N1 , N2 ) is a partition of V (G) \ {u}, and since |V (G) \ {u}| ≥ n − 1 =
1 + (n1 − 1) + (n2 − 1), the Pigeonhole Principle guarantees that either
|N1 | ≥ n1 or |N2 | ≥ n2 .
Suppose first that |N1 | ≥ n1 , i.e. |N1 | ≥ k+(`−1)−2


k−1 . Then by the
induction hypothesis, either ω(G[N1 ]) ≥ k or α(G[N1 ]) ≥ ` − 1. In the
former case, we have that ω(G) ≥ ω(G[N1 ]) ≥ k, and we are done. So
suppose that α(G[N1 ]) ≥ ` − 1. Then let S be a stable set of G[N1 ] of size
` − 1. Then {u} ∪ S is a stable set of size ` in G, we deduce that α(G) ≥ `,
and again we are done.
Suppose now that |N2 | ≥ n2 , i.e. |N2 | ≥ (k−1)+`−2


k−2 . Then by the
induction hypothesis, either ω(G[N2 ]) ≥ k − 1 or α(G[N2 ]) ≥ `. In the latter
case, we have that α(G) ≥ α(G[N2 ]) ≥ `, and we are done. So suppose that
ω(G[N2 ]) ≥ k − 1. Then let C be a clique of G[N2 ] of size k − 1. But then
{u} ∪ C is a clique of size k in G, we deduce that ω(G) ≥ k, and again we
are done.

For positive integers k and `, we denote by R(k, `) the smallest number

n such that every graph G on at least n vertices satisfies either ω(G) ≥ k or
α(G) ≥ `. The existence of R(k, `) follows immediately from Theorem 3.2.
Numbers R(k, `) (with k, ` ≥ 1) are called Ramsey numbers.
It is easy to see that for all k, ` ≥ 1, we have that18
R(1, `) = 1 R(k, 1) = 1

R(2, `) = ` R(k, 2) = k
Furthermore, we have R(3, 3) = 6. Indeed, by Proposition 3.1, R(3, 3) ≤ 6.
On the other hand, ω(C5 ) = 2 and α(C5 ) = 2, and so R(3, 3) > 5. Thus,
R(3, 3) = 6. The exact values of a few other Ramsey numbers are known,19
but no general formula for R(k, `) is known. Note however, that Theorem
3.2
gives an upper bound for Ramsey numbers, namely, R(k, `) ≤ k+`−2 k−1 for
all k, ` ≥ 1.
We complete this section by giving a lower bound for the Ramsey number
R(k, k).
Theorem 3.3. For all integers k ≥ 3, we have that R(k, k) > 2k/2 .
Proof. Since ω(C5 ) = 2 and α(C5 ) = 2, we see that R(3, 3) > 5 > 23/2 and
R(4, 4) > 5 > 24/2 . Thus, the claim holds for k = 3 and k = 4. From now
on, we assume that k ≥ 5.
Let G be a graph on n := b2k/2 c vertices, with adjacency as follows:
between any two distinct vertices, we (independently) put an edge with
probability 12 (and a non-edge with probability 21 ).
18
Check this!
19
For example, it is known that R(4, 4) = 18. On the other hand, the exact value of
R(5, 5) is still unknown.

6
 For any set of k vertices of G, the probability that this set is a clique is
k
( 12 )(2) ;
there are nk subsets of V (G) of size k, and the probability that at


k
least one of them is a clique is at most nk ( 12 )(2) . So, the probability that


k
ω(G) ≥ k is at most nk ( 12 )(2) . Similarly, the probability that α(G) ≥ k


k
is at most n ( 1 )(2) . Thus, the probability that G satisfies at least one of


k 2
ω(G) ≥ k and α(G) ≥ k is at most
k k
n
( 12 )(2) ≤ 2( en k 1 (2)


2 k k ) (2) by Theorem 2.1
from Lecture Notes 1

k/2
2( e2 k )k
≤ 2k(k−1)/2
because n = b2k/2 c

k/2
= 2( k2e2
(k−1)/2 )
k

√
< 2( e k 2 )k

< 1 because k ≥ 5

Thus, the probability that G satisfies neither ω(G) ≥ k nor α(G) ≥ k is

strictly positive. So, there must be at least one graph on n = b2k/2 c vertices
whose clique number and stability number are both strictly less than k. This
proves that R(k, k) > b2k/2 c; since R(k, k) is an integer, we deduce that
R(k, k) > 2k/2 .