This action might not be possible to undo. Are you sure you want to continue?
Ngo
SUNY at Buffalo, Fall 2003 Scribe: Hung Q. Ngo
Lecture 3: The Probabilistic Method  Basic ideas
We draw materials mainly from [1, 3].
To show that some (combinatorial) object exists, one can envision working on some probability space
which the object lives in, and show that the probability of such an existence is strictly positive.
We shall often use the following notations:
• δ(G)  the minimum degree of a graph G
• ∆(G)  the maximum degree of a graph G
• deg(v)  the degree of a vertex v ∈ V (G) of a graph G
• χ(v)  the chromatic number of a graph G
• χ
e
(v)  the chromatic index of a graph G
• [n] := ¦1, . . . , n¦
• a ksubset is a subset of size k of some set
• 2
X
 the superset of a set X
•
X
k
 the set of all ksubsets of a set X
• a hypergraph is a pair H = (V, E) where V is a ﬁnite set, and E is a collection of subsets of V .
Naturally, members of V are called vertices, and E are called edges of the hypergraph H
• a uniform hypergraph is a hypergraph all of whose edges have the same size
• an nuniform hypergraph is a hypergraph all of whose edges have size n. (Thus, the normal graph
is a 2uniform hypergraph.)
• S
n
 the symmetric group on [n], i.e. the set of all permutations on [n] (or n symbols)
1 Ramsey numbers
The classical example which Erd˝ os applied the probabilistic method is the socalled Ramsey numbers.
In the simplest form, let R(a, b) be the smallest integer n such that in any 2edgecoloring of K
n
with
RED and BLUE, there exists a RED K
a
or a BLUE K
b
. (You can also think along this line: what’s
the smallest number n so that in any set of n people there must be a mutually acquainted people, or b
mutually strangers. Try it with a = b = 3.)
Proposition 1.1 (Erd˝ os, 1947 [4]). If
n
k
2
1−(
k
2
)
< 1, then R(k, k) > n. Consequently, R(k, k) >
2
k/2
 for all k ≥ 3.
1
Proof. (Note that R(k, k) is the minimum n such that a graph G on n vertices contains a K
k
, or the
complement
¯
G of G has a K
k
.)
Consider K
n
and a random 2coloring on its edges, namely we color an edge BLUE with probability
1/2, and RED with probability 1/2. For any ksubset S of vertices, let E
S
be the event that the induced
subgraph on S is monochromatic. Then, P[E
S
] = 2
1−(
k
2
)
. Thus, the probability that some ksubset
forms a monochromatic subgraph is at most
n
k
2
1−(
k
2
)
. Consequently, when
n
k
2
1−(
k
2
)
< 1 there exists
some 2coloring for which there is no monochromatic K
k
. In other words, K(k, k) > n.
For k ≥ 3, let n = 2
k/2
. Then,
n
k
2
1−(
k
2
)
<
n
k
k!
2
1+k/2
2
k
2
/2
<
2
1+k/2
k!
n
k
2
k
2
/2
< 1.
Note:
• The previous argument can be made a perfectly ﬁne and simple counting argument.
• In most results we shall see, however, the probability is essential. Straightforward counting nor
mally is way to cumbersome or virtually impossible.
• One can give a randomized algorithm to ﬁnd a 2coloring with no monochromatic K
k
based on
the proof above. We have seen that the probability of having some monochromatic K
k
is at most
2
1+k/2
k!
< 1. (In fact, it is < 1 when k is large.) Hence, after m trials of colorings  the probability
of having no monochromatic K
k
is 1 −
2
1+k/2
k!
m
, which goes to 1 as m grows larger. We can
control the error probability easily.
Exercise 1.2. Prove that there is a real p ∈ [0, 1] such that
n
a
p
(
a
2
)
+
n
b
(1 −p)
(
b
2
)
< 1, (1)
then the Ramsey number R(a, b) satisﬁes R(a, b) > n. Use this to show
r(4, b) = Ω
b
3/2
(ln b)
3/2
.
Solution. Randomly color each edge BLUE with probability p (and RED otherwise). The probability of
having a BLUE K
a
or a RED K
b
is at most
n
a
p
(
a
2
)
+
n
b
(1 −p)
(
b
2
)
< 1.
Hence, R(a, b) > n.
To this end, note that
n
a
p
(
a
2
)
+
n
b
(1 −p)
(
b
2
)
< n
a
p
(
a
2
)
/a! +n
b
e
−(
b
2
)p
/b!.
Hence, as long as we maintain that
n
a
p
(
a
2
)
/a! ≤ 1/2 (2)
n
b
e
−(
b
2
)p
/b! ≤ 1/2, (3)
2
then 1 holds true. Put it another way, we want
p ≤
(a!/2)
1/(
a
2
)
n
2/(a−1)
, (4)
p ≥
ln
b!/(2n
b
)
b
2
. (5)
Now, we can just pick an n as large as possible such that
ln
b!/(2n
b
)
b
2
≤
(a!/2)
1/(
a
2
)
n
2/(a−1)
.
It is tedious yet easy to see that n = Θ
(b/ ln b)
3/2
sufﬁces for the case a = 4.
2 Dominating set
Fact 2.1. For p ∈ [0, 1], (1 −p) ≤ e
−p
. The inequality is good for small p.
Fact 2.2 (Linearity of expectation). For any random variables X
1
, . . . , X
k
, E[c
1
X
1
+ + c
k
X
k
] =
c
1
E[X
1
] + +c
k
E[X
k
].
Fact 2.3. To minimize a function f(x) doubly differentiable, where f
(x) ≥ 0 (i.e. f(x) is convex), we
ﬁnd x
0
such that f
(x
0
) = 0. This x
0
is a minima.
Given a graph G = (V, E), a subset S ⊆ V is called a dominating set if every vertex of G is either
in S or adjacent to some vertex in S.
Theorem 2.4. Let G = (V, E) be a graph on n vertices with δ(G) = δ > 1. Then, G has a dominating
set of size at most n
1+ln(δ+1)
δ+1
.
Proof. Pick each vertex of G at random with probability p. Let X be the set of chosen vertices. Let Y
be the subset of V − X where no y ∈ Y has a neighbor in X. Clearly X ∪ Y is a dominating set. We
estimate the average size of X ∪ Y . If the average size (according to p) of X ∪ Y is at most n
1+ln(δ+1)
δ+1
,
then there must exist a choice of X for which [X ∪ Y [ ≤ n
1+ln(δ+1)
δ+1
.
For any v ∈ V ,
P[v ∈ Y ] = P[v and its neighbors were not picked] = (1 −p)
1+deg(v)
≤ (1 −p)
1+δ
.
Hence,
E[[Y [] =
¸
v
P[v ∈ Y ] ≤ n(1 −p)
1+δ
.
Consequently,
E[[X[ +[Y [] ≤ np +n(1 −p)
1+δ
.
The right hand side is minimized at p
0
= 1 − (1 + δ)
1/δ
. Thus, we can prove a slightly stronger result:
there is a dominating set of size at most np
0
+n(1 −p
0
)
1+δ
. This bound, however, is not “clean.”
A cleaner bound can be obtained by noticing that
np +n(1 −p)
1+δ
≤ np +ne
−p(1+δ)
.
The right hand side is minimized at p
1
=
ln(1+δ)
(1+δ)
, yielding the bound stated in the theorem.
3
3 Extremal set theory
A hypergraph H = (V, E) has property B if it is twocolorable, i.e. there exists a twocoloring of the
vertices so that no edge is monochromatic. Obviously the fewer the number of edges, the more likely for
H to have property B. Let m(n) be the least number of edges so that an nuniform hypergraph does not
have property B. We want to ﬁnd a lower bound for m(n).
Theorem 3.1 (Erd˝ os, 1963 [5]). Every nuniform hyper graph with < 2
n−1
edges has a property B.
Hence, m(n) ≥ 2
n−1
.
Proof. To prove the existence of a certain type of coloring, we generate random colorings and show the
probability of the existence of the “type” is positive.
Color each vertex of H = (V, E) with two colors at random (probability 1/2 for each color). The
probability that some e ∈ E is monochromatic is 2/2
n
. Hence, the probability that at least one edge in
E is monochromatic is at most [E[/2
n−1
< 1. Consequently, there exists a good coloring.
Exercise 3.2. Suppose n ≥ 4 and let H be an nuniform hypergraph with at most 4
n−1
/3
n
edges. Prove
that there is a vertex coloring of H by 4 colors so that in every edge, all four colors are presented.
The following very wellknown result is called the Erd˝ osKoRado theorem [6]. We present a proof
by Kantona [7].
Theorem 3.3 (Erd˝ osKoRado). Let n ≥ 2k be positive integers. Let T be a family of ksubsets of [n]
for which A, B ∈ T implies A∩ B = ∅. Then,
[T[ ≤
n −1
k −1
.
A direct counting proof of Erd˝ osKoRado theorem. Consider a permutation π ∈ S
n
. Put π(1), . . . , π(n)
on a cycle in the clockwise order. Call the cycle C
π
. A member A ∈ T is said to be consecutive on C
π
if all elements of A occur consecutively on the cycle. It is easy to see that there are at most k members
of T which are consecutive on C
π
for a ﬁxed π. Also, there are only (n − 1)! different cycles, not n!
(why?). Let ( be the set of all such cycles. Then,
¸
C∈C
[¦A ∈ T [ A consecutive on C¦[ ≤ k (n −1)!
Moreover,
¸
C∈C
[¦A ∈ T [ A consecutive on C¦[ =
¸
A∈F
[¦C ∈ ( [ A consecutive on C¦[ = [T[k!(n −k)!.
Hence, [F[ ≤
n−1
k−1
as desired.
A probabilistic proof of Erd˝ osKoRado theorem. Consider π ∈ S
n
. For each i ∈ [n], the probability
that A
π
i
= ¦π(i), . . . , π(i + k − 1)¦ (taken circularly) is a member of T is at most k/n, because, as in
the previous proof, there can be at most k member of T that are consecutive on C
π
.
Moreover, P[A
π
i
∈ T] = [T[/
n
k
. Hence, [T[ ≤
n
k
k/n =
n−1
k−1
.
Exercise 3.4 (Sperner Lemma, 1928 [8]). The maximum size of a family of subsets of [v] none of
whose member is contained in another is
v
v/2
.
The following theorem implies Sperner’s Lemma, although no one noticed it until Tuza (1984, [9]).
4
Theorem 3.5 (Bollob´ as, 1965 [2]). Let A = ¦X
1
, . . . , X
m
¦, and \ = ¦Y
1
, . . . , Y
m
¦ be two set systems
of [n] such that
(i) X
i
∩ Y
i
= ∅, ∀i.
(ii) X
i
∩ Y
j
= ∅, ∀i = j.
Then,
m
¸
i=1
1
X
i
+Y
i

X
i

≤ 1. (6)
Proof. Let x
i
= [X
i
[, and y
j
= [Y
j
[. Consider a random π ∈ S
n
. Let E
i
be the event that elements of
X
i
come before elements of Y
i
in π. Then,
P[E
i
] =
n
x
i
+y
i
x
i
!y
i
!(n −x
i
−y
i
)!
n!
=
1
x
i
+y
i
x
i
.
It’s easy to see that the events E
i
are mutually exclusive. The theorem follows easily.
Corollary 3.6. If x
i
= x, y
i
= x, for all i = 1, . . . , m in the previous theorem, then m ≤
x+y
x
.
Exercise 3.7. Prove that Boll ´ obas’ Theorem implies Sperner’s Lemma.
Exercise 3.8. Let T = ¦(A
i
, B
i
), 1 ≤ i ≤ m¦ be a family of pairs of subsets of the set of integers such
that [A
i
[ = a, [B
i
[ = b, for all i, A
i
∩ B
i
= ∅, and (A
i
∩ B
j
) ∪ (A
j
∩ B
i
) = ∅ for all i = j. Show that
m ≤
(a +b)
a+b
a
a
b
b
. (7)
4 Coding Theory
Note that the following theorems hold for qary codes, also. We only stated the binary versions for clarity
of presentation.
Theorem 4.1 (Kraft inequality). Let ( be a ﬁnite collection of ﬁnite binary strings such that no string
is a preﬁx of another. Let n
i
be the number of strings of length i in (. Then,
¸
i
n
i
2
i
≤ 1.
Proof. Let mbe the length of longest string in the collection. Flip a fair coin mtimes to generate a binary
string s of length m. No two strings in ( can both be preﬁxes of s at the same time. For a particular
codeword w ∈ ( (i.e. a string in () of length i, the probability that s has w as a preﬁx is 1/2
i
. These
possibilities are mutually exclusive for all codewords w, hence the total probability is ≤ 1, namely
¸
i
n
i
2
i
≤ 1,
as desired.
Kraft showed that the converse is also true, i.e. given the numbers n
i
satisfying the inequality, then
there exists a preﬁxfree code with n
i
codewords of length i.
McMillan later discovered that the inequality is also the necessary condition for a code to be uniquely
decipherable, which is the content of the following theorem. (Sufﬁciency, again, follows from Kraft’s
result.)
5
Theorem 4.2 (KraftMcMillan inequality). Let ( be a ﬁnite collection of ﬁnite binary strings such that
no two distinct concatenations of two ﬁnite sequences of codewords result in the same binary sequence.
Let n
i
be the number of strings of n
i
in (. Then,
¸
i
n
i
2
i
≤ 1.
A typical proof. Let m be the largest string length. A trivial combinatorial reasoning shows that
m
¸
i=1
n
i
2
i
k
≤ mk, (8)
for all integers k. Since (mk)
1/k
→ 1 as k → ∞, the desired inequality follows.
One can show (8) with a somewhat messy probabilistic argument.
References
[1] N. ALON AND J. H. SPENCER, The probabilistic method, WileyInterscience Series in Discrete Mathematics and Opti
mization, WileyInterscience [John Wiley & Sons], New York, second ed., 2000. With an appendix on the life and work of
Paul Erd˝ os.
[2] B. BOLLOB ´ AS, On generalized graphs, Acta Math. Acad. Sci. Hungar, 16 (1965), pp. 447–452.
[3] B. BOLLOB ´ AS, Random graphs, vol. 73 of Cambridge Studies in Advanced Mathematics, Cambridge University Press,
Cambridge, second ed., 2001.
[4] P. ERD¨ OS, Some remarks on the theory of graphs, Bull. Amer. Math. Soc., 53 (1947), pp. 292–294.
[5] P. ERD˝ OS, On a combinatorial problem, Nordisk Mat. Tidskr., 11 (1963), pp. 5–10, 40.
[6] P. ERD˝ OS, C. KO, AND R. RADO, Intersection theorems for systems of ﬁnite sets, Quart. J. Math. Oxford Ser. (2), 12
(1961), pp. 313–320.
[7] G. O. H. KATONA, A simple proof of the Erd˝ osChao KoRado theorem, J. Combinatorial Theory Ser. B, 13 (1972),
pp. 183–184.
[8] E. SPERNER, Ein satz uber untermengen einer endlichen Menge, Math. Z., 27 (1928), pp. 544–548.
[9] Z. TUZA, Hellytype hypergraphs and Sperner families, European J. Combin., 5 (1984), pp. 185–187.
6
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.