ORIE 6334 Spectral Graph Theory September 1, 2016

Lecture 4
Lecturer: David P. Williamson Scribe: Kun Dong

Once again, we’ll start by proving a general theorem about eigenvalues, and then
show its application to some graph problems.

1 Eigenvalue Interlacing Theorem

The following theorem is known as the eigenvalue interlacing theorem.

Theorem 1 (Eigenvalue Interlacing Theorem) Suppose A ∈ Rn×n is symmet-

ric. Let B ∈ Rm×m with m < n be a principal submatrix (obtained by deleting both i-th
row and i-th column for some values of i). Suppose A has eigenvalues λ1 ≤ · · · ≤ λn
and B has eigenvalues β1 ≤ · · · ≤ βm . Then
λk ≤ βk ≤ λk+n−m for k = 1, · · · , m
And if m = n − 1,
λ1 ≤ β1 ≤ λ2 ≤ β2 ≤ · · · ≤ βn−1 ≤ λn
 
B XT
Proof: WLOG, assume A = . Let {x1 , · · · , xn } be eigenvectors of A,
X Z
{y1 , · · · , ym } be eigenvectors of B. We define the following vector spaces:
  
w n
V = span(xk , · · · , xn ), W = span(y1 · · · , yk ), W
f= ∈ R ,w ∈ W
0
Tf
Sincedim(V  ) = n − k + 1 and dim(W ) = dim(W ) = k, there exists w e∈V W and
f
w
w
e= for some w ∈ W . Then
0
  
T
 T  B XT w
w
e Awe= w 0 = wT Bw
X Z 0

xT Ax xT Bx
Recall λk = min and β k = max . Then we see that
x∈V xT x x∈W xT x

eT Aw
w wT Bw
λk ≤ ≤ βk
e
=
eT w
w e wT w
0
This lecture was drawn from some notes of Embree http://www.caam.rice.edu/~caam440/
chapter2.pdf and Spielman’s 2012 lecture notes, Lecture 3: http://www.cs.yale.edu/homes/
spielman/561/2012/lect03-12.pdf.

4-1
 The proof of the other inequality is similar. We now define the vector spaces
  
w n
V = span(x1 , · · · , xk+n−m ), W = span(yk · · · , ym ), W
f= ∈ R ,w ∈ W
0
T
 ) =k + n − m, dim(W ) = dim(W ) = m − k + 1, there exists w
Since dim(V e∈V W
f
w
and we= for some w ∈ W . We have w eT Aw e = wT Bw as before. It follows
0
that
xT Ax eT Aw
w wT Bw xT Bx
λk+n−m = max T ≥ T ≥
e
= min = βk ,
x∈V x x w
e we wT w x∈W xT x
completing the proof. 2

2 Clique and Chromatic Number

We now use the eigenvalue interlacing theorem to prove some statements about two
particular graph quantities, the clique number and the chromatic number.
Definition 1 The clique number of G, ω(G), is the size of the largest S ⊆ V such
that for all i, j ∈ S, (i, j) ∈ E.
Example:

ω(G) = 4

Definition 2 The chromatic number χ(G) is the fewest number of colors needed such
that we can assign one color to each vertoex and for all (i, j) ∈ E, i, j are assigned
different colors.
Observation 1 χ(G) ≥ ω(G).
The observation follows since every vertex in the maximum clique needs to be
assigned a different color: if two vertices in the clique are assigned the same color,
then since there is an edge between them, the two endpoints of that edge are not
assigned different colors.
Consider the complete graph on n nodes G ≡ Kn ; that is, there is an edge between
every pair of vertices. Then ω(G) = n = χ(G). The adjacency matrix of G is
A = J − I where J is the matrix of all ones. Let
 
1
 .. 
e =  . .
1

4-2
Then
Ae = (J − I)e = ne − e = (n − 1)e
Therefore e is an eigenvector for eigenvalue n − 1.
For any vector v such that he, vi = 0, Av = (J − I)v = 0 − v = −v. This means
any v such that he, vi = 0 is an eigenvector of eigenvalue −1. Thus the spectrum of
A is n − 1 with multiplicity 1 and −1 with multiplicity n − 1.
Now consider an arbitrary graph G, and let λ1 ≥ · · · ≥ λn be the eigenvalues to
its adjacency matrix.

Claim 2 λ1 ≥ ω(G) − 1.
Proof: For the largest clique S in G, let B be the principal submatrix with
columns and rows corresponding to S. Let m = |S| = ω(G), then B = Jm − Im . If
β1 is the largest eigenvalue of B, then β1 = ω(G) − 1. By the Interlacing Theorem,
λ1 ≥ β1 = ω(G) − 1. 2
We can in fact prove something slightly stronger. The following theorem strength-
ens that bound of the claim since ω(G) ≤ χ(G).

Theorem 3 (Wilf 1967) χ(G) ≤ bλ1 c + 1

Before we can prove this
P we need a lemma. Let d(i) be the degree of node i,
d(i)
∆ = maxi d(i) and dave = i∈Vn .

Lemma 4 dave ≤ λ1 ≤ ∆ for G connected.

Proof: P P
xT Ax eT Ae i,j aij d(i)
λ1 = max T ≥ T = = i∈V = dave
x x x e e n n
Let x1 be the eigenvector. By Perron-Frobenius Theorem, we can assume x1 > 0. Let
i∗ = arg maxi x(i), then
(Ax1 )(i∗ ) = λ1 x1 (i∗ )
P
(Ax1 )(i∗ ) j:(j,i∗ )∈E x1 (j)
X x1 (j) X
=⇒ λ1 = = = ≤ 1 = d(i∗ ) ≤ ∆
x1 (i∗ ) x1 (i∗ ) ∗
x 1 (i∗)
∗
j:(j,i )∈E j:(j,i )∈E

2
Notice that we needed connectivity in the proof above to invoke the Perron-
Frobenius theorem for the inequality λ1 ≤ ∆; we did not need it for the lower bound
dave ≤ λ1 .
As an aside, we can use the above lemma to prove the following, though we will
not do so here.

Claim 5 G is connected and λ1 = ∆, then G is ∆-regular.

We also observe the following.

4-3
Observation 2 χ(G) ≤ ∆ + 1.
This is true because if we color the graph greedily, we will never get stuck: if we
color a vertex, it has at most ∆ neighbors that have already been colored, and so we
can color it with the (∆ + 1)st color.
Proof of Theorem 3:
The proof is by induction on n.
Base case n = 2,

λ1 = 1, χ1 (G) = 2

λ1 = 0, χ1 (G) = 1

Inductive step: Suppose the theorem holds on all graphs with at most n − 1 vertices.
By the Lemma, G has a vertex of degree less than bλ1 c. Remove this vertex v and call
the resulting graph G0 . Let B be its adjacency matrix and β1 be its largest eigenvalue.
By the Interlacing Theorem, β1 ≤ λ1 . By induction, we can color G0 with bβ1 c + 1
colors, which is less than bλ1 c + 1 colors. We can then finish coloring G by coloring
v with one of the bλ1 c + 1 colors since degree of v is less than bλ1 c. 2
We now give one more result, a lower bound on the clique number. In order to
prove it, we assume the following theorem of Motzkin and Straus.

Theorem 6 (Motzkin-Straus 1965)

 
1 1 X
1− = Maximize xi xj
2 ω(G)
(i,j)∈E
subject to:
X
xi ≤ 1
i∈V
xi ≥ 0 ∀i ∈ V.

Then we can prove the following

Pn
Theorem 7 (Wilf 1986) Let S = i=1 x1 (i) where x1 is the eigenvector for the
2
largest eigenvalue and kx1 k = 1. Then ω(G) ≥ S 2S−λ1 .
2

Proof: Let x be solution to Motzkin-Straus program. Then

1
1− = xT Ax;
ω(G)
notice that the factor of 12 is dropped because xT Ax counts every edge twice. By the
Perron-Frobenius Theorem, we can assume x1 ≥ 0 and xS1 is feasible for the program.
This implies
1 x1 T x1 λ1
1− = xT Ax ≥ A = 2
ω(G) S S S

4-4
so that
λ1 1
1− 2
≥ ,
S ω(G)
which implies
S 2 − λ1 1
2
≥ ,
S ω(G)
or
S2
ω(G) ≥ 2 .
S − λ1
2
Using the Cauchy-Schwartz inequality, we can get that
n
X √ √
S= x1 (i) ≤ kx1 k 12 + · · · + 12 = n.
i=1

Plugging this bound into the inequality above gives us

n
ω(G) ≥ .
n − λ1

4-5