Random Graphs

Electrical Engineering 126 (UC Berkeley)

Spring 2018

1 Introduction
In this note, we will briefly introduce the subject of random graphs, also
known as Erdös-Rényi random graphs. Given a positive integer n and
a probability value p ∈ [0, 1], the G(n, p) random graph is an undirected
n


graph on n vertices such that each of the 2 edges is present in the graph
independently with probability p. When p = 0, G(n, 0) is an empty graph on
n vertices, and when p = 1, G(n, 1) is the fully connected graph on n vertices
(denoted Kn ). Often, we think of p = p(n) as depending on n, and we are
usually interested in the behavior of the random graph model as n → ∞.
A bit more formally, G(n, p) defines a distribution over the set of undirected
graphs on n vertices. If G ∼ G(n, p), meaning that G is a random graph with
the G(n, p) distribution, then for every fixed graph G0 on n vertices with m
n
edges, P(G = G0 ) := pm (1 − p)( 2 )−m . In particular, if p = 1/2, then the
probability space is uniform, or in other words, every undirected graph on n
vertices is equally likely.
Here are some warm-up questions.
Question 1. What is the expected number of edges in G(n, p)?
Answer 1. There are n2 possible edges and the probability that any given

edge appears
in the random graph is p, so by linearity of expectation, the

answer is n2 p.
Question 2. Pick an arbitrary vertex and let D be its degree. What is the
distribution of D? What is the expected degree?
Answer 2. Each of the n − 1 edges connected to the vertex is present
independently with probability p, so
D ∼ Binomial(n − 1, p). For every
d ∈ {0, 1, . . . , n − 1}, P(D = d) = n−1
d
pd (1 − p)n−1−d , and E[D] = (n − 1)p.
Question 3. Suppose now that p(n) = λ/n for a constant λ > 0. What is
the approximate distribution of D when n is large?

1
Answer 3. By the Poisson approximation to the binomial distribution, D is
approximately Poisson(λ). For every d ∈ N, P(D = d) ≈ exp(−λ)λd /d!.

Question 4. What is the probability that any given vertex is isolated?

Answer 4. All of the n − 1 edges connected to the vertex must be absent,

so the desired probability is (1 − p)n−1 .

2 Sharp Threshold for Connectivity

We will sketch the following result (see [1]):

Theorem 1 (Erdös-Rényi, 1961). Let

ln n
p(n) := λ
n
for a constant λ > 0.

• If λ < 1, then P{G(n, p(n)) is connected} → 0.

• If λ > 1, then P{G(n, p(n)) is connected} → 1.

In the subject of random graphs, threshold phenomena like the one above
are very common. In the above result, nudging the value of λ slightly around
the critical value of 1 causes drastically different behavior in the limit, so it
is called a sharp threshold. In such cases, examining the behavior near the
critical value leads to further insights. Here, if we take p(n) = (ln n + c)/n
for a constant c ∈ R, then it is known that


P G n, p(n) is connected → exp{− exp(−c)},

see [2, Theorem 7.3]. Notice that the probability increases smoothly from 0
to 1 as we vary c from −∞ to ∞.
Why is the threshold p(n) = (ln n)/n? When p(n) = 1/n, then the
expected degree of a vertex is roughly 1 so many of the vertices will be joined
together (a great deal is known about the evolution of the so-called giant
component, see e.g. [2]), but it is too likely that one of the vertices will have no
edges connected to it, making it isolated (and thus the graph is disconnected).

2
Proof of Theorem 1. First, let λ < 1. If Xn denotes the number of isolated
nodes in G(n, p(n)), then it suffices to show that P(Xn > 0) → 1, i.e., there is
an isolated node with high probability (this will then imply that the random
graph is disconnected).

• E[Xn ]: Define Ii to be the indicator random variable of the event that

the ith vertex
Pn is isolated.PUsing linearity of expectation and symmetry,
n
E[Xn ] = i=1 E[Ii ] = i=1 P(node i is isolated) = nq(n), where we
define q(n) := P(a node is isolated) = [1 − p(n)]n−1 .
Observe that
n−1
ln E[Xn ] = ln n + (n − 1) ln{1 − p(n)} ∼ ln n − λ ln n → ∞,
n
since λ < 1. Here, if f and g are two functions on N, then the notation
f (n) ∼ g(n) means f (n)/g(n) → 1 (asymptotically, f and g have the
same behavior). The above line also uses the first-order Taylor expansion
ln(1 − x) = −x + o(x) as x → 0.
Thus E[Xn ] → ∞ which is reassuring, since we want to prove that
P(Xn > 0) → 1, but in order to prove the probability result we will
need to also look at the variance of Xn .

• var Xn : We claim that

var Xn
P(Xn = 0) ≤ .
E[Xn ]2

Here are two ways to see this. First, from the definition of variance,

var Xn = E[(Xn − E[Xn ])2 ]

= E[Xn ]2 P(Xn = 0) + (1 − E[Xn ])2 P(Xn = 1) + · · ·
≥ E[Xn ]2 P(Xn = 0).

The second way is to use Chebyshev’s Inequality:

var Xn
P(Xn = 0) ≤ P(|Xn − E[Xn ]| ≥ E[Xn ]) ≤ .
E[Xn ]2

The use of the variance is often called the Second Moment Method.
We must show that the ratio (var Xn )/ E[Xn ]2 → 0. Since I1 , . . . , In are

3
 not independent, we must use var Xn = n var I1 + n(n − 1) cov(I1 , I2 ).
Since I1 is a Bernoulli random variable, var I1 = q(n)[1 − q(n)], and by
definition cov(I1 , I2 ) = E[I1 I2 ] − E[I1 ] E[I2 ] = E[I1 I2 ] − q(n)2 .
In order to find E[I1 I2 ], we interpret it as a probability:
E[I1 I2 ] = P(nodes 1, 2 are isolated).
In order for this event to happen, 2n − 3 edges must be absent:
q(n)2
P(nodes 1, 2 are isolated) = [1 − p(n)]2n−3 = .
1 − p(n)
So, cov(I1 , I2 ) = q(n)2 /[1 − p(n)] − q(n)2 = p(n)q(n)2 /[1 − p(n)], and
var Xn nq(n)[1 − q(n)] + n(n − 1)p(n)q(n)2 /[1 − p(n)]
=
E[Xn ]2 n2 q(n)2
1 − q(n) n − 1 p(n)
= + .
nq(n) n 1 − p(n)
Since nq(n) = E[Xn ] → ∞, the first term tends to 0, and since p(n) → 0,
the second term tends to 0 as well.
Next, let λ > 1. The key idea for the second claim is the following: the graph
is disconnected if and only if there exists a set of k nodes, k ∈ {1, . . . , bn/2c},
such that there is no edge connecting the k nodes to the other n − k nodes in
the graph. We can apply the union bound twice.


P G n, p(n) is disconnected
bn/2c
[ 
=P {some set of k nodes is disconnected}
k=1
bn/2c
X
≤ P(some set of k nodes is disconnected)
k=1
bn/2c 
X n
≤ P(a specific set of k nodes is disconnected)
k=1
k
bn/2c  
X n
= [1 − p(n)]k(n−k) .
k=1
k
The rest of the proof is showing that the above summation tends to 0 via
tedious calculations, which will be given in the Appendix.

4
Appendix: Tedious Calculations
Here, we will argue that
bn/2c   bn/2c  
X n k(n−k)
X n
[1 − p(n)] ≤ exp{−k(n − k)p(n)}
k=1
k k=1
k
bn/2c  
X n
= n−λk(n−k)/n
k=1
k

tends to 0 as n → ∞. One way to do this is to break up the summation into

two parts. Since λ > 1, choose n∗ so that λ(n − n∗ )/n > 1, which means we
can take n∗ = bn(1 − λ−1 )c. The first part of the summation is
n∗   n∗ n∗
X n X X ∗ )/n−1]
n−λk(n−k)/n ≤ n−k[λ(n−k)/n−1] ≤ n−k[λ(n−n
k=1
k k=1 k=1
−[λ(n−n∗ )/n−1]
n
≤ → 0.
1 − n−[λ(n−n∗ )/n−1]
For the second part of the summation, we will use the bound
∞
nk  n k k k  n k X k j  en k
 
n
≤ = ≤ = .
k k! k k! k j=0 j! k

Using this bound:

bn/2c bn/2c  bn/2c 
X n X en1−λ(n−k)/n k X en1−λ(n−k)/n k
−λk(n−k)/n
n ≤ ≤
k=n∗ +1
k k=n∗ +1
k k=n∗ +1
n∗ + 1
bn/2c bn/2c 
X  en−λ(n−k)/n k X en−λ/2 k
≤ ≤
k=n∗ +1
1 − λ−1 k=n∗ +1
1 − λ−1

For n sufficiently large, en−λ/2 /(1 − λ−1 ) < δ for some δ < 1.
∞ ∗
X δn
k
≤ δ = →0
k=n∗
1−δ

since n∗ → ∞.

5
References
[1] Daron Acemoglu and Asu Ozdaglar. Erdös-Rényi graphs and branching
processes. url: http://economics.mit.edu/files/4621.
[2] Béla Bollobás. Random graphs. Second. Vol. 73. Cambridge Studies
in Advanced Mathematics. Cambridge University Press, Cambridge,
2001, pp. xviii+498. isbn: 0-521-80920-7; 0-521-79722-5. doi: 10.1017/
CBO9780511814068.