ORIE 6334 Spectral Graph Theory August 23, 2016

Lecture 1
Lecturer: David P. Williamson Scribe: Matthew Zalesak

1 Eigenvectors, Eigenvalues, and Graph Theory

Consider A ∈ IRn×x , A = AT (symmetric). If Ax = λx, then for x ∈ IRn , λ ∈ IR,

• x is an eigenvector;

• λ is the corresponding eigenvalue.

Eigenvectors and eigenvalues have applications in differential equations, machanics, fre-

quency analysis, and many others.
An undirected graph G is represented as a tuple (V, E) consisting of a set of vertices V
and a set of edges E. We are interested in paths, flows, cuts, colorings, cliques, spanning
trees, etc. of the graph G.
This semester we will ask what graphs and eigenvalues have to do with each other.

2 An Introductory Example
The diameter D of a graph is minimum length you would have to be able to travel to
guarantee that you could go from any node in the graph to any other node. Formally,
D = maxi,j∈V [length of shortest i-j path]. A graph is said to be d-regular if all nodes are
of degree d, where degree is defined as the number of edges incident on each vertex. The
below graph has diameter 2 but is not d-regular since some nodes are of degree 2 and some
are of degree 3.

For our introductory example1 , we will consider d-regular graphs of diameter 2 with as
many nodes as possible. By starting at any node i, the graph could look like
1
This material is taken from the article A.J. Hoffman, R.R. Singleton, “On Moore Graphs with Diameters
2 and 3,” IBM Journal, pp. 497–504, November 1960.

1-1
 1

d(d-1)

In this graph, note that there are no connections between adjacent nodes in the first
layer since we want to maximize the number of nodes in the graph. Also, the connections
between the leaf nodes are omitted from the diagram. Based on the diagram, such a graph
would have n = 1 + d + d(d − 1) = d2 + 1 nodes.
Let A = (aij ) be the adjacency matrix of G, defined as
(
1 if (i, j) ∈ E
aij =
0 otherwise

If B = A2 , then
X
bij = aik akj = number of walks of 2 steps in graph G from i to j.
k

Since this is a d-regular graph, we have that bii = d since starting at i we can reach d
vertices in one step and then immediately return.
Starting at i we reach every other vertex in G in exactly 0 steps, exactly 1 step, or
exactly 2 steps (exclusive or). Thus,

I + A + A2 − dI = J,

where I is the identity matrix and J is the matrix of all ones.

We’ll need the following facts from linear algebra.

Fact 1 For A ∈ IRn×n symmetric, the following are true:

• All of the eigenvalues of A are real.

• There exist eigenvalues λ1 , . . . , λn (called the spectrum) and eigenvectors x1 , . . . , xn

such that hxi , xj i = xTi xj = 0 for i 6= j.

• The trace tr(A) = ni=1 aii = ni=1 λi .

P P

1-2
 We’ve discussed the notion of a d-regular graph with diameter two, but does such a
graph exist? To answer this problem we can use linear algebra to tell us what such a graph
would might look like.
First notice that Ae = de, where e = (1, . . . , 1)T is the vector of all ones. This is true
since A is d-regular. Thus, e is an eigenvector of A and d is the eigenvalue. Also notice that

A2 e = A(Ae) = A(de) = d(Ae) = d2 e.

Thus,

(I + A + A2 − dI)e = Je
e + de + d2 e − de = ne

So n = d2 + 1, though we already knew that.

Now let v be any other eigenvector of A orthogonal to e. Then v T e = 0, and thus
Jv = 0. We have that Av = λv for some eigenvalue λ. Also, A2 v = A(Av) = A(λv) = λ2 v.
Thus

(I + A + A2 − dI)v = Jv
v + λv + λ2 v − λv = 0
=⇒ 1 + λ + λ2 − d = 0
√
So for all eigenvalues not corresponding to e, we have λ = −1± 2 4d−3 .
Given what we now know about the eigenvalues, what can we tell? We can invoke the
trace! Notice that tr(A) = 0: aii = 0 for all i since there are no self-loops in the graph.
Now we consider two possible cases.
√
1. If 4d − 3 is irrational,
√ √
then in order for the trace to sum to zero, the eigenvalues −1+ 2 4d−3 and −1− 4d−3
2
must each have multiplicity n−1
2 . Plugging this in gives
 √ √ 
n − 1 −1 + 4d − 3 −1 − 4d − 3
tr(A) = 0 = d + +
d 2 2
n−1
=d−
2
d2
=d−
2
=⇒ d = 0 or d = 2.

So the only possible graphs would be:

(a) d = 0 A single node, which does not have diameter two.

1-3
 (b) d = 2 In this case n = 5, which gives the 5 cycle; the 5-cycle is a 2-regular graph
of diameter 2.

√
2. If 4d − 3 is rational,
−1+s
then let s2 = 4d − 3. Let m be the multiplicity of the eigenvalue 2 . Then
   
−1 + s −1 − s
tr(A) = d + m + (n − 1 − m)
2 2
=0

Using the fact that d = 41 (s2 + 3), we get n − 1 = d2 = 1 4 2

16 (s + 6s + 9). So continuing
    
1 −1 + s 1 4 −1 − s
= (s2 + 3) + m + 2
(s + 6s + 9) − m
4 2 16 2

After simplifying the algebra, we find that

−s5 − s4 − 6s3 + 2s2 + (32m − 9)s + 15 = 0.

By the rational root theorem, we know that any solution to this polynomial must be
a factor of 15. Thus, we can enumerate all possible roots.
Possibilities

(a) s = 1, d = 1, n = 2
A 1-regular graph on 2 nodes is a single edge, but its diameter is not 2.

(b) s = 3, d = 3, n = 10
We can represent this graph in two ways:

1-4
 This first representation shows what the graph would look like using the same
type of diagram we used earlier.

This second representation is often called the Petersen representation (Image

source: Wikipedia), and the graph is called the Petersen graph. Petersen found
it in the course of trying to find the smallest cubic (that is, 3-regular) bridgeless
graph that could not be 3-edge-colored. A bridgeless graph is one such that the
graph is still connected after removing any edge. A 3-edge-colorable graph is one
in which we can color every edge with one of three colors such that at each vertex,
all incident edges have different colors. The Petersen graph is also the smallest
cubic bridgeless graph that does not have a Hamiltonian cycle. Knuth has called
the Petersen graph ”a remarkable configuration that serves as a counterexample
to many optimistic predictions about what might be true for graphs in general.”
(c) s = 5, d = 7, n = 50
This graph is known to exist and is called the Hoffman-Singleton graph (Hoffman,
Singleton 1960).

1-5
 (Image source: Wikipedia)

(d) s = 15, d = 57, n = 3250.

Does this graph exist? We don’t know. This is a good research question!

This example is to give you a small taste of how eigenvectors can be useful in graph
theory. By looking at the spectrum of d-regular graphs of diameter 2 with as many nodes
as possible, we were able to come up with very strong restrictions on the possible values of
d.

1-6