Linear Algebra and its Applications 504 (2016) 1–13

Contents lists available at ScienceDirect

Linear Algebra and its Applications

www.elsevier.com/locate/laa

A structure theory for graphs with fixed smallest

eigenvalue
Hyun Kwang Kim a , Jack H. Koolen b,∗ , Jae Young Yang a
a
Department of Mathematics, Pohang Mathematics Institute, POSTECH, Pohang,
South Korea
b
School of Mathematical Sciences, University of Science and Technology of China,
Wen-Tsun Wu Key Laboratory of the Chinese Academy of Sciences, Hefei, Anhui,
China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we will give a structure theory for graphs with
Received 20 April 2015 fixed smallest eigenvalue. In order to do this, the concept of
Accepted 25 March 2016 Hoffman graph (as introduced by Woo and Neumaier) is used.
Available online 6 April 2016
Our main result states that for fixed positive integer λ and any
Submitted by R. Brualdi
graph G with smallest eigenvalue at least −λ, there exist dense
Dedicated to Alan J. Hoffman on the induced subgraphs Q1 , . . . , Qc in G such that each vertex lies
occasion of his 91st birthday in at most λ Qi ’s and almost all edges of G lie in at least one
of the Qi ’s.
MSC: © 2016 Elsevier Inc. All rights reserved.
05C50
05C75
05C62

Keywords:
Smallest eigenvalue
Structure theory
Hoffman graph
Quasi-clique

* Corresponding author.
E-mail addresses: hkkim@postech.ac.kr (H.K. Kim), koolen@ustc.edu.cn (J.H. Koolen),
rafle@postech.ac.kr (J.Y. Yang).

http://dx.doi.org/10.1016/j.laa.2016.03.044
0024-3795/© 2016 Elsevier Inc. All rights reserved.
2 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

1. Introduction

For graph theoretic notations, we refer to [2]. All graphs in this paper are undirected
and simple. In this paper, we will give a structure theory for graphs with fixed small-
est eigenvalue. We will use the concept of Hoffman graph (as introduced by Woo and
Neumaier [7]). First, we will introduce a result of Hoffman.
Let t be a positive integer and let K̃2t be the graph on 2t + 1 vertices consisting of a
complete graph K2t and a vertex ∞ which is adjacent to exactly t vertices of the K2t . It
is easy to see that the smallest eigenvalue of K̃2t goes to −∞ if t goes to ∞. The smallest
√
eigenvalue of a t-claw K1,t equals − t (where t is a positive integer), and hence it will
go to −∞ if t goes to ∞. So if the smallest eigenvalue of a graph G is at least a fixed
real number λ, then there exists a positive integer t = t(λ) such that G contains neither
a K̃2t nor a t-claw K1,t as an induced subgraph. Hoffman [3] showed that the opposite
is also true:

Theorem 1.1. Let G be a graph with smallest eigenvalue λmin (G). Then the following
holds.

(i) For any real number λ ≤ −1, there exists a positive integer t = t(λ) such that
if λmin (G) ≥ λ, then G contains neither a K̃2t nor a t-claw K1,t as an induced
subgraph.
(ii) For any positive integer t, there exists a non-positive real number λ = λ(t) such
that if G contains neither a K̃2t nor a t-claw K1,t as an induced subgraph, then
λmin (G) ≥ λ.

In order to prove Theorem 1.1(ii), Hoffman showed that if G contains neither a K̃2t
nor a t-claw K1,t as an induced subgraph, then there exists a highly structured graph
H with the same vertex set as G and with large distinguished cliques that is close to G
(see, for example, [1, Theorem 7.3.1] for a precise description of the graph H). In this
paper, we will give a proof by using the concept of Hoffman graphs.
The main focus of this paper is to obtain a structure theory for graphs with fixed
smallest eigenvalue. Our main theorem is:

Theorem 1.2. Let λ ≤ −1 be a real number. Then there exists a positive integer dλ such
that if G is a graph with smallest eigenvalue λmin (G) ≥ λ and its minimal valency δ(G) ≥
dλ , then for some integer c, there exist induced subgraphs Q1 , Q2 , . . . , Qc satisfying the
following conditions:

(i) Each vertex x of G lies in at least one and at most −λ Qi ’s;

(ii) For each i, the complement of Qi has maximal valency at most λ2 ;
(iii) For i = j, the intersection V (Qi ) ∩ V (Qj ) contains at most −λ − 1 vertices;
c
(iv) The graph G := (V (G), E(G) \ i=1 E(Qi )) has maximal valency at most dλ .
 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 3

This result refines a result of Hoffman. Our estimate for the number dλ is not very
precise, as we use Ramsey theory to show the existence of dλ . It would be very desirable
to have better estimates for dλ . It would also be desirable to decrease the number dλ in
(iv) to a quadratic or cubic polynomial in λ. In a follow-up paper, we will use Theorem 1.2
to strengthen results of Hoffman [4], and Woo and Neumaier [7].
The paper is organized as follows. In Section 2, we discuss Hoffman graphs as intro-
duced by Woo and Neumaier [7]. In that section, we also give some preliminary results
on Hoffman graphs. In Section 3, we show that if a graph does not contain K̃2m as
an induced subgraph, then there exists an equivalence relation on the maximal cliques
with at least (m + 1)2 vertices. This equivalence relation was first found by Hoffman [3].
By using this equivalence relation, we define the concepts of quasi-clique and associated
Hoffman graph. In Section 4, we show properties of the associated Hoffman graph. In
Section 5, we give a proof of Theorem 1.1(ii) by using the associated Hoffman graph.
This proof is based on idea’s of Hoffman [3]. Finally, in Section 6, we give a proof of
Theorem 1.2.

2. Hoffman graphs

Woo and Neumaier [7] introduced the concept of Hoffman graphs in order to capture
the main idea of Hoffman’s paper [4]. A Hoffman graph h is a pair (H, ), where H =
(V (H), E(H)) is a graph, and  : V (H) → {f, s} is a labeling map, such that the set
Vf (h) := {x ∈ V (H) | (x) = f } is an independent set and any vertex with label
f has a vertex with label s as a neighbor. The vertices with label s are called slim
vertices and the vertices labeled f are called fat vertices. The subgraph of H induced by
Vs (h) := {x ∈ V (H) | (x) = s} is called the slim graph of h.
For a vertex x of h we define N f (x) = Nhf (x) (resp. N s (x) = Nhs (x)) the set of fat
(resp. slim) neighbors of x in h. The set of all neighbors of x is denoted by N (x) = Nh (x),
that is N (x) = N f (x) ∪ N s (x). In a similar fashion, for vertices x and y we define
N f (x, y) = Nhf (x, y) to be the set of common fat neighbors of x and y.
A Hoffman graph h is called t-fat if every slim vertex of h has at least t fat neighbors,
where t is a positive integer. Note that a fat Hoffman graph is a 1-fat Hoffman graph.
A Hoffman graph h1 = (H1 , 1 ) is called an induced Hoffman subgraph of a Hoffman
graph h = (H, ), if H1 is an induced subgraph of H and 1 (x) = (x) for all vertices x
of h1 .
Two Hoffman graphs h = (H, ) and h = (H  ,  ) are called isomorphic if there exists
an isomorphism from H to H  which preserves the labeling.
For a Hoffman graph h = (H, ), let A be the adjacency matrix of H
 
As C
A=
CT O

in a labeling in which the fat vertices come last. The real symmetric matrix S(h) :=
As − CC T is called the special matrix of h. Note that non-isomorphic Hoffman graphs
4 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

may have the same special matrix. The eigenvalues of h are the eigenvalues of the special
matrix S(h).
Let λmin (h) denote the smallest eigenvalue of h. The following result relates the
smallest eigenvalue of a Hoffman graph h and its induced Hoffman subgraph, see [7,
Corollary 3.3].

Lemma 2.1. If g is an induced Hoffman subgraph of a Hoffman graph h, then λmin (g) ≥
λmin (h) holds.

Let h be a Hoffman graph with Vf (h) = {F1 , F2 , . . . , Fc } for some integer c. Let n be a
positive integer. The graph G(h, n) has as vertex set Vs (h) ∪ V1 ∪ . . . ∪ Vc (disjoint union),
where Vi is a set of size n for i = 1, . . . , c, such that each vertex x ∈ Vi is adjacent to
exactly all the vertices in Vs (h) that are neighbors of Fi in h and the induced subgraph
on Vi is the complete graph Kn , for i = 1, . . . , c. Ostrowski and Hoffman (see also [5,
Corollary 2.15]) showed the following result.

Theorem 2.2. Let h be a Hoffman graph and let n be a positive integer. Then

λmin (G(h, n)) ≥ λmin (h), (1)

and

lim λmin (G(h, n)) = λmin (h). (2)

n→∞

In particular, for any  > 0, there exists a positive integer n such that every graph G
containing G(h, n) as an induced subgraph satisfies

λmin (G) ≤ λmin (h) + .

3. Quasi-cliques

For a graph G, a clique or a complete subgraph is the induced subgraph on a dependent

set. In this section we will define the concept of quasi-cliques for graphs.
Recall that K̃2m is the graph on 2m + 1 vertices consisting of a complete graph K2m
and a distinguished vertex (say ∞) which is adjacent to exactly m vertices of the K2m .
For the rest of this section, let m be a positive integer and let G be a graph that
does not contain K̃2m as an induced subgraph. Let n be a positive integer and define
C(n) := {C | C is a maximal clique of G with at least n vertices}. Define the relation
≡mn on C(n) by C1 ≡n C2 if each vertex x ∈ C1 has at most m − 1 non-neighbors in C2
m

and each vertex y ∈ C2 has at most m − 1 non-neighbors in C1 , where C1 , C2 ∈ C(n).

Lemma 3.1. Let m, n be positive integers such that n ≥ (m +1)2 and let G be a graph that
does not contain K̃2m as an induced subgraph. Then, the above-defined relation ≡m n on
the set C(n) of maximal cliques of G with at least n vertices, is an equivalence relation.
 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 5

Proof. The fact that ≡m n is reflexive and symmetric is clear. We only need to show that
≡n is transitive. Let C1 ≡m
m
n C2 and C2 ≡n C3 for maximal cliques C1 , C2 , C3 ∈ C(n).
m

Let x be a vertex of C1 . We have to show that x is not adjacent to at most m − 1

vertices of C3 . We may assume that x is not a vertex of C3 , as then x is adjacent to all
vertices of C3 except itself. In order to obtain a contradiction, we assume that x is not
adjacent to m distinct vertices z1 , z2 , . . . , zm of C3 . The number of vertices of C2 which
are adjacent to each of the zi ’s is at least #V (C2 ) −(m −1)m ≥ 3m +1 as C2 ≡m n C3 and
#V (C2 ) ≥ n ≥ (m + 1) . The vertex x is adjacent to at least 2m + 2 of those vertices, as
2

C1 ≡m n C2 . This shows that K̃2m is an induced subgraph of G, which is a contradiction

with our assumption that G does not contain K̃2m as an induced subgraph. This shows
that x is not adjacent to at most m − 1 vertices in C3 . As the role of C1 and C3 is
symmetric, this shows the lemma. 2

Let [C]mn denote the equivalence class of C(n) of G under the equivalence relation ≡n
m

containing the maximal clique C of C(n).

Lemma 3.2. Let G be a graph that does not contain K̃2m as an induced subgraph, where
m ≥ 2 is an integer. Let n ≥ (m + 1)2 , and C1 , C2 be maximal cliques of G with at least
n = [C2 ]n . Then the following holds.
n vertices such that [C1 ]m m

(i) At most m − 1 vertices of C1 have at least #V (C2 ) − m + 1 neighbors in C2 and the

other vertices of C1 have at most m − 1 neighbors in C2 ;
(ii) At most m − 1 vertices of C2 have at least #V (C1 ) − m + 1 neighbors in C1 and the
other vertices of C2 have at most m − 1 neighbors in C1 .

Proof. Without loss of generality, there exists a vertex x of C1 such that x has at most
#V (C2 ) − m neighbors in C2 . This means that x has at most m − 1 neighbors in C2 ,
as otherwise, there would be an induced K̃2m in G. Assume there are distinct vertices
y1 , y2 , . . . , ym in C1 , each having at least #V (C2 ) − m + 1 neighbors in C2 . This means
there are at least #V (C2 ) − m(m − 1) > 2m vertices z in C2 that are adjacent to yi for
each i = 1, 2, . . . , m. At least m of them are not neighbors of x. But this means there is
an induced K̃2m in G, a contradiction. This shows (i).
To show (ii), we only need to show that there exists a vertex u in C2 with at most
#V (C1 ) − m neighbors in C1 . In order to show this, let W1 := {x | x ∈ V (C1 ) \ V (C2 )}
and W2 := {y | y ∈ V (C2 ) \ V (C1 )}. Note that #W1 ≥ #V (C1 ) − m + 1, and #W2 ≥
#V (C2 ) − m + 1 both hold. Now we will count the number α of edges xy with x ∈ W1
and y ∈ W2 . Then

α ≤ (m − 1)(#W2 ) + (#W1 )(m − 1)

≤ (m − 1)(#W2 ) + (#V (C1 ))(m − 1)

holds.
6 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

Now, as #V (C1 ) ≥ n ≥ (m + 1)2 ≥ 5m − 5 and #W2 ≥ n − m − 1 ≥ 4m − 4, we find

#V (C1 )(m − 1) ≤ #W2 (#V (C1 ) − 3m + 3), and hence

α ≤ (m − 1)(#W2 ) + (#V (C1 ))(m − 1) ≤ (#W2 )(#V (C1 ) − 2m + 2)

holds. So there exists a vertex u of C2 that has at most #V (C1 ) − m + 1 neighbors in

C1 . This concludes the proof of the lemma. 2

The next lemma is needed later to show that a quasi-clique is well-defined.

Lemma 3.3. Let m, n be positive integers such that n ≥ (m + 1)2 and G be a graph that
does not contain K̃2m as an induced subgraph. Let x be a vertex of G and let C be a
maximal clique with at least n vertices. Assume that x has at most m − 1 non-neighbors
in C. Then for all maximal cliques C  with at least n vertices in [C]m
n , the vertex x has

at most m − 1 non-neighbors in C .

Proof. We may assume that x is not a vertex of C  , as then it is clear that the lemma
holds. Assume that x is not adjacent to the distinct vertices z1 , z2 , . . . , zm of C  , in order

to obtain a contradiction. As C ≡m n C , the set of vertices {y ∈ V (C) | y ∼ zi for all
i = 1, . . . , m} contains at least n − (m − 1)m ≥ 3m + 1 vertices. Among these 3m + 1
vertices at least 2m + 2 are adjacent to x, and hence G has an induced subgraph K̃2m .
This gives a contradiction. 2

Let m, n be positive integers such that n ≥ (m + 1)2 and let G be a graph that does
not contain K̃2m as an induced subgraph. Let C be a maximal clique of G with at least
n vertices. We define the quasi-clique Q([C]m n ), with respect to the pair (m, n), as the
induced subgraph of G on the set {x ∈ V (G) | x has at most m − 1 non-neighbors in C}.

Note that by Lemma 3.3, for any vertex x of Q([C]m n ) and any maximal clique C of

n , the vertex x has at most m − 1 non-neighbors in C .
[C]m

4. Associated Hoffman graphs

Let m, n be positive integers such that n ≥ (m + 1)2 and let G be a graph

that does not contain K̃2m as an induced subgraph. Let G have equivalence classes
[C1 ]m m m
n , [C2 ]n , . . . , [Ct ]n on the maximal cliques with at least n vertices under the equiv-
alence relation ≡n . The associated Hoffman graph g = g(G, m, n) is a Hoffman graph
m

satisfying the following conditions:

(i) Vs (g) = V (G), Vf (g) = {F1 , F2 , . . . Ft };

(ii) The slim graph of g equals G;
(iii) The fat vertex Fi is adjacent to each vertex in the quasi-clique Q([Ci ]m
n ) for i =
1, 2, . . . , t.
 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 7

The following proposition shows an important property of the associated Hoffman

graph.

Proposition 4.1. Let G be a graph and let m ≥ 2, φ ≥ 1, σ ≥ 1, p ≥ 1 be integers. There

exists a positive integer n = n(m, φ, σ, p) ≥ (m + 1)2 such that for any integer q ≥ n, and
any Hoffman graph h with at most φ fat vertices and at most σ slim vertices, the graph
G(h, p) is an induced subgraph of G, provided that the graph G satisfies the following
conditions:

(i) The graph G does not contain K̃2m as an induced subgraph,

(ii) Its associated Hoffman graph g = g(G, m, q) contains h as an induced Hoffman
subgraph.

Proof. Without loss of generality, we may assume that p ≥ 3m. Suppose that h has
exactly φ fat vertices and σ  slim vertices. We will show the proposition by induction
on φ .
Now we show that the proposition is true for φ = 1. Let n = n (m, 1, σ  , p) :=
max{(m + 1)2 , σ  m + p} and q ≥ n . Let h be an induced Hoffman subgraph of g(G, m, q)
with exactly one fat vertex F corresponding to the quasi-clique Q = Q([C]m q ), where C
is a maximal clique of G with at least q vertices. Let x1 , x2 , . . . , xσ be the slim vertices
of h. The slim graph of h is an induced subgraph of G.
If a vertex xi is adjacent to F , xi has at most m −1 non-neighbors in C. If a vertex xi is
not adjacent to F , xi has at most m−1 neighbors in C. By deleting these σ  (m−1) vertices
from C \{x1 , x2 , . . . , xσ }, we can obtain a clique with p vertices since n −σ  −σ  (m −1) ≥
p holds. This shows the proposition when φ = 1.
We now consider the case φ ≥ 2. Let n = max{n (m, φ − 1, σ  , p + m − 1), σ  m +
p(m − 1)(φ − 1) + p} and q ≥ n .
Let h be an induced Hoffman subgraph consist of fat vertices F1, . . . , Fφ and slim
vertices x1 , . . . , xσ . Now, as h is an induced Hoffman subgraph of g = g(G, m, q), the
slim graph of h is an induced subgraph of G. Let Qi = Q([Ci ]m q ) be the quasi-clique of
G corresponding to the fat vertex Fi for i = 1, 2, . . . , φ , where Ci is a maximal clique of
G with at least q vertices.
By the induction hypothesis, we have subcliques Di of Ci for i = 1, 2, . . . , φ − 1 such
that each Di has exactly p + m − 1 vertices, V (Di ) ∩ V (Dj ) = ∅ and there are no edges
between Di and Dj for 1 ≤ i < j ≤ φ − 1, and the vertex x is adjacent to all vertices
of Di if x ∼ Fi and x has no neighbors in Di if x  Fi for i = 1, 2, . . . , φ − 1 and
 = 1, 2, . . . , σ  .
Let Dφ be the maximal subclique of Cφ such that for all  = 1, 2, . . . , σ  , the following
holds: if x ∼ Fφ , then x is adjacent to all vertices of Dφ , and, if x  Fφ , then x
is not adjacent to any vertex of Dφ . In a similar manner to the case φ = 1, the clique
Dφ has at least n − σ  m vertices. By Lemma 3.2, each Di has at most m − 1 vertices
which has at least #V (Cφ ) − m + 1 neighbors in Cφ for i = 1, 2, . . . , φ − 1. Let Di be
8 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

Fig. 1. k4 , k4 and c4 .

the cliques obtained by removing m − 1 vertices from Di for i = 1, 2, . . . , φ − 1. Then

each Di has exactly p vertices and all vertices in Di have at most m − 1 neighbors in
Cφ . For i = 1, 2, . . . , φ − 1, each vertex of Di has at most m − 1 neighbors in Dφ .
After removing all these neighbors from Dφ , we can find a clique Dφ  with exactly p
vertices since n − σ  m − p(m − 1)(φ − 1) ≥ p holds. Hence, the subgraph induced by
φ
{x1 , x2 , . . . , xσ } ∪ i=1 V (Di ) equals G(h, p).
Take n = max1≤φ ≤φ, 1≤σ ≤σ {n (m, φ , σ  , p)}. This shows the proposition. 2

Now we will apply Proposition 4.1 to show that certain Hoffman graphs cannot be
induced Hoffman subgraphs of the associated Hoffman graph. For this, we need to in-
troduce three different Hoffman graphs.

Definition 4.2. Let r be a positive integer. Then we have:

(i) The Hoffman graph kr is the Hoffman graph with exactly r fat vertices F1 , F2 , . . . , Fr
and a slim vertex x such that x ∼ Fi for all i.
(ii) The Hoffman graph kr is the Hoffman graph with one fat vertex F and r slim
vertices x1 , x2 , . . . , xr such that xi ∼ F for all i and none of the xi ’s are adjacent.
(iii) The Hoffman graph cr is the Hoffman graph with exactly two fat vertices F1 , F2 and
r slim vertices x1 , x2 , . . . , xr , and the vertex xi is adjacent to all the other vertices
for i = 1, 2, . . . , r.

In Fig. 1, you can see pictures of kr , kr and cr for r = 4.

Lemma 4.3. Let m and t be positive integers. Let G be a graph that contains neither K̃2m ,
nor K1,t as an induced subgraph.

(i) There exists a positive integer p1 ≥ (m + 1)2 such that for any integer q1 ≥ p1 , the
associated Hoffman graph g(G, m, q1 ) does not contain kt as an induced Hoffman
subgraph.
(ii) There exists a positive integer p2 ≥ (m + 1)2 such that for any integer q2 ≥ p2 , the
associated Hoffman graph g(G, m, q2 ) does not contain kt as an induced Hoffman
subgraph.
(iii) There exists a positive integer p3 ≥ (m + 1)2 such that for any integer q3 ≥ p3 , the
associated Hoffman graph g(G, m, q3 ) does not contain cm as an induced Hoffman
subgraph.
 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 9

Proof. (i): The graph G does not contain the t-claw, which is isomorphic to G(kt , 1). By
Proposition 4.1, the existence of p1 follows.
(ii): The proof is very similar to the proof of (i).
(iii): The proof is the same as for (i) except that now we use the fact that G does not
contain K̃2m as an induced subgraph. 2

5. Proof of Theorem 1.1(ii)

Before proving Theorem 1.1(ii), recall the well-known Ramsey’s Theorem.

Theorem 5.1. (See [6].) Let k, l be two positive integers. There exists a minimum positive
integer R(k, l) such that for any graph G on n ≥ R(k, l) vertices, the graph G contains
a dependent set of vertices of size k or an independent set of vertices of size l.

For two positive integers k, l, the number R(k, l) of the above result is called the
Ramsey number.

Theorem 5.2. Let m and t be positive integers. Then there exist positive integers s ≥
(m + 1)2 and κ(m, t) such that, if G is a graph that contains neither K̃2m , nor a t-claw
as an induced subgraph, and Q1 , Q2 , . . . , Qc are the quasi-cliques of G with respect to the
pair (m, s), then the following conditions hold:

(i) Each vertex in G has at most t − 1 fat neighbors in g(G, m, s);

(ii) The complement of Qi has maximal valency at most κ(m, t) for i = 1, 2, . . . , c;

(iii) The intersection V (Qi ) V (Qj ) contains at most κ(m, t) vertices for 1 ≤ i < j ≤ c;
q
(iv) The graph G := (V (G), E(G) \ i=1 E(Qi )) has maximal valency at most κ(m, t).

Proof. By Lemma 4.3(i), there exists a positive integer p1 ≥ (m+1)2 such that for any in-
teger q1 ≥ p1 , the associated Hoffman graph g(G, m, q1 ) does not contain kt as an induced
Hoffman subgraph. In a similar manner, using Lemma 4.3(ii) and (iii), we see that there
exist positive integers p2 , p3 both at least (m +1)2 , such that for any integers q2 ≥ p2 and
q3 ≥ p3 , the associated Hoffman graphs g(G, m, q2 ) and g(G, m, q3 ) do not contain kt and
cm as induced Hoffman subgraphs, respectively. Let s := max{p1 , p2 , p3 , 1 + t(m − 1)}.
Let Q1 , Q2 , . . . , Qc be the quasi-cliques of G with respect to the pair (m, s), and let
g := g(G, m, s) be the associated Hoffman graph. Let κ(m, t) := R(s, t) − 1, as defined
in Theorem 5.1. Then (i) follows from the fact that g does not contain kt as an induced
Hoffman subgraph.
For (ii), suppose that the complement of Qi has a vertex x with valency at least
R(s, t). Then x would not be adjacent to all vertices in a clique C of order s or x would
not be adjacent to all vertices in an independent set T of size t. In the first case, we find
a maximal clique C  containing C but it contradicts the definition of quasi-clique since
x has s non-neighbors in C  . In the second case, take a maximal clique C  with order at
10 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

least s from Qi . Each vertex of T has at most m − 1 non-neighbors in C  . We can find a

vertex y ∈ C  which is adjacent to all vertices of T since s − t(m − 1) ≥ 1 holds. Hence,
the graph G contains a t-claw as an induced subgraph. This gives a contradiction.
To show (iii), first notice that the induced subgraph on V (Qi ) does not have an
independent set of size t since kt is not an induced Hoffman subgraph, for i = 1, 2, . . . , c.

Now let 1 ≤ i < j ≤ c, and consider the subgraph H induced by V (Qi ) V (Qj ).
As cm is not an induced Hoffman subgraph of g, it follows that H does not contain
a complete subgraph with m vertices and an independent set of size t. It follows that

#(V (Qi ) V (Qj )) < R(m, t) ≤ κ(m, t). This shows (iii).
Finally, (iv) follows from the fact that the graph G does not contain a complete
subgraph with s vertices and a t-claw. 2

Now we are in the position to show Theorem 1.1(ii).

Proof of Theorem 1.1(ii). Let the integers s and κ(t, t) be defined as in Theorem 5.2,
and g := g(G, m, s) be the associated Hoffman graph. Note that λmin (G) ≥ λmin (g), by
Lemma 2.1.
Let S = S(g) be the special matrix of g and x be a vertex of G. We will count the
number of vertices y of G such that Sxy = 0. First we count the number of y’s such that
Sxy = 1. Then such y is adjacent to x and, x and y have no common fat neighbor. There
are at most κ(t, t) of such y’s.
Next, we count the number of y’s such that Sxy ≤ −1. Note that x is contained in at
least one and at most t − 1 fat vertices. We need to consider three cases: y = x, y ∼ x
and y  x. If y ∼ x, then x, y have at least two common fat neighbors and hence there
 
are at most t−12 × κ(t, t) such y’s. If y  x, then y and x have at least one common
fat neighbor in g. There are at most (t − 1) × κ(t, t) such y’s by Theorem 5.2(ii). So
 
there are at most 1 + ( t−12 + 1)κ(t, t) + (t − 1)κ(t, t) y’s such that the entry Sxy = 0.
As 1 − t ≤ Sxy ≤ 1 for all y, we know that the absolute value of the row sum of S
corresponding to x satisfies:

      
t−1
| Sxy | ≤ |Sxy | ≤ 1+ + 1 κ(t, t) + (t − 1)κ(t, t) (t − 1).
y y
2

Hence, by Perron–Frobenius Theorem,


λmin (g) ≥ − max |Sxy |
x
y
    
t−1
≥− 1+ + 1 κ(t, t) + (t − 1)κ(t, t) (t − 1).
2

This shows the theorem. 2

 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 11

6. Graphs with fixed smallest eigenvalue

Now we will look at graphs with fixed smallest eigenvalue. Then we need the following
theorem which is a slightly modified version of Theorem 5.2 and whose proof is essentially
the same as of Theorem 5.2. The details are left for the reader.

Theorem 6.1. Let p, m and t be positive integers. Then there exist positive integers
s ≥ max{(m + 1)2 , p} and κ(p, m, t) such that if G is a graph satisfying the following
conditions:

(i) Each vertex of G has valency at least κ(p, m, t);

(ii) G does not contain K̃2m as an induced subgraph;
(iii) G does not contain t-claw as an induced subgraph;

and let Q1 , Q2 , . . . , Qc be the quasi-cliques of G with respect to the pair (m, s), then the
graph G satisfies the following conditions:

(a) Each vertex of G lies in a complete subgraph with at least s vertices, and hence the
associated Hoffman graph g(G, m, s) is a fat Hoffman graph;
(b) Each vertex of G has at most t − 1 fat neighbors in g(G, m, s);
(c) The complement of Qi has valency at most κ(p, m, t) for i = 1, 2, . . . , c;

(d) The intersection V (Qi ) V (Qj ) contains at most κ(p, m, t) vertices for 1 ≤ i <
j ≤ c;
c
(e) The graph G := (V (G), E(G) \ i=1 E(Qi )) has maximal valency at most κ(p, m, t).

For fixed smallest eigenvalue, we also obtain the following result.

Theorem 6.2. Let λ ≤ −1 be a real number. Then there exist positive integers mλ , nλ
and dλ such that if G is a graph satisfying the following conditions:

(i) Each vertex has valency at least dλ ;

(ii) λmin (G) ≥ λ.

and let Q1 , Q2 , . . . , Qc be the quasi-cliques of G with respect to the pair (mλ , nλ ), then
the graph G satisfies the following conditions:

(a) Each vertex lies in a complete subgraph with at least nλ vertices;

(b) Each vertex in G has at most −λ fat neighbors in g(G, mλ , nλ );
(c) The complement of Qi has valency at most λ2 for i = 1, 2, . . . , c;

(d) The intersection V (Qi ) V (Qj ) contains at most −λ−1 vertices for 1 ≤ i < j ≤ c;
c
(e) The graph G := (V (G), E(G) \ i=1 E(Qi )) has maximal valency at most dλ .
12 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13

Proof. By Theorem 1.1(i), there exist positive integers t = tλ and m = mλ such that G
contains neither a t-claw nor a K̃2m as an induced subgraph.
Note that λmin (k−λ+1 ) = −−λ + 1 = λ − 1 < λ for the graph k−λ+1 . By
Theorem 2.2, there exists a positive integer b such that λmin (G(k−λ+1 , b)) < λ. Hence,
by Proposition 4.1, there exists a positive integer p1 such that for any integer q1 ≥ p1 , the
associated Hoffman graph g(G, m, q1 ) does not contain k−λ+1 as an induced Hoffman
subgraph.
Let H be a graph with  := λ2 + 2 vertices such that at least one vertex has valency
 − 1. Then, by the Perron–Frobenius Theorem (see [2, Theorem 8.8.1]), the largest
√
eigenvalue of H is at least  − 1 > −λ. Now, consider the fat Hoffman graph b = b(H)
with one fat vertex such that the slim graph of b(H) is equal to the complement of
√
H. Then S(b) = −A(H) − I. It follows that λmin (b) ≤ −  − 1 − 1 < λ. Hence, by
Theorem 2.2 and Proposition 4.1, there exists a positive integer pH ≥ (m + 1)2 such
that for any integer q2 ≥ pH , the associated Hoffman graph g(G, m, q2 ) does not contain
b(H) as an induced Hoffman subgraph. Let p2 be the maximum of all pH , where the
maximum is taken over all graphs H having exactly  vertices and containing a vertex
with valency  − 1.
Let y be any fat Hoffman graph with exactly two fat vertices F1 , F2 and −λ slim
vertices x1 , x2 , . . . , x−λ , such that each vertex xi is adjacent to F1 and F2 , for i =
1, 2, . . . , −λ. Then the diagonal elements of S(y) are equal to −2 and the rest of the
entries of S(y) are in {−2, −1}. This shows that λmin (y) ≤ λ − 1 < λ. Hence, by
Theorem 2.2 and Proposition 4.1, there exists a positive integer p(y) ≥ (m + 1)2 such
that for any integer q3 ≥ p(y), the associated Hoffman graph g(G, m, q3 ) does not contain
y as an induced Hoffman subgraph. Let p3 be the maximum of all such p(y)’s.
Take n = nλ := max{p1 , p2 , p3 } and take dλ := κ(n, m, t) as defined in Theorem 6.1.
Then, by the definition of n, we see that (b), (c) and (d) hold. By Theorem 6.1, it follows
that (a) and (e) hold.
This shows the theorem. 2

Remark 6.3. Note that the bounds in (b), (c), (d) are sharp. It would be very helpful to
improve the bound in (e).

Note that Theorem 1.2 is an immediate consequence of Theorem 6.2.

Acknowledgements

The research of H.K. Kim was supported by the Basic Science Research Program
through the National Research Foundation of Korea (NRF) funded by the Ministry
of Education (2014047764). J.H. Koolen is partially supported by the National Natural
Science Foundation of China (No. 11471009). He also acknowledges the financial support
of the Chinese Academy of Sciences under its ’100 talent’ program. We would also like
 H.K. Kim et al. / Linear Algebra and its Applications 504 (2016) 1–13 13

to thank Aida Abiad, Qianqian Yang and the anonymous referee for their comments on
an earlier version of the paper.

References

[1] D. Cvetković, P. Rowlinson, S. Simić, Spectral Generalizations of Line Graphs, London Math. Soc.
Lecture Note Ser., vol. 314, Cambridge Univ. Press, 2004.
[2] C.D. Godsil, G. Royle, Algebraic Graph Theory, Graduate Texts in Mathematics, Springer, New
York, 2000.
[3] A.J. Hoffman, On spectrally bounded graphs, in: N. Srivastava, et al. (Eds.), A Survey of Combina-
torial Theory, North Holland, 1973, pp. 277–283. √
[4] A.J. Hoffman, On graphs whose least eigenvalue exceeds −1 − 2, Linear Algebra Appl. 16 (1977)
153–165.
[5] H.J. Jang, J. Koolen, A. Munemasa, T. Taniguchi, On fat Hoffman graphs with smallest eigenvalue
at least −3, Ars Math. Contemp. 7 (2014) 105–121.
[6] F.P. Ramsey, On a problem of formal logic, Proc. Lond. Math. Soc. 48 (1930) 264–286.
√
[7] R. Woo, A. Neumaier, On graphs whose smallest eigenvalue is at least −1 − 2, Linear Algebra
Appl. 226–228 (1995) 577–591.