Linear Algebra and its Applications 556 (2018) 210–219

Contents lists available at ScienceDirect

Linear Algebra and its Applications

www.elsevier.com/locate/laa

On the Aα -spectra of graphs

Huiqiu Lin a,∗ , Jie Xue b , Jinlong Shu b
a
Department of Mathematics, East China University of Science and Technology,
Shanghai, PR China
b
Department of Computer Science and Technology, East China Normal University,
Shanghai, PR China

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

Article history: Let G be a graph with adjacency matrix A(G) and let D(G) be
Received 16 June 2017 the diagonal matrix of the degrees of G. For any real α ∈ [0, 1],
Accepted 4 July 2018 Nikiforov [8] defined the matrix Aα (G) as
Available online 10 July 2018
Submitted by R. Brualdi
Aα (G) = αD(G) + (1 − α)A(G).
MSC:
05C50
05C12 In this paper, we give some results on the eigenvalues of Aα (G)
for α > 1/2. In particular, we characterize the graphs with
Keywords: λk (Aα (G)) = αn − 1 for 2 ≤ k ≤ n. Moreover, we show that
Aα -matrix λn (Aα (G)) ≥ 2α − 1 if G contains no isolated vertices.
The k-th largest Aα -eigenvalue © 2018 Elsevier Inc. All rights reserved.
The smallest Aα -eigenvalue

1. Introduction

All graphs considered here are simple and undirected. Let G be a graph with adjacency
matrix A(G), and let D(G) be the diagonal matrix of the degrees of G. For any real
α ∈ [0, 1], Nikiforov [8] defined the matrix Aα (G) as

* Corresponding author.
E-mail address: huiqiulin@126.com (H. Lin).

https://doi.org/10.1016/j.laa.2018.07.003
0024-3795/© 2018 Elsevier Inc. All rights reserved.
 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219 211

Aα (G) = αD(G) + (1 − α)A(G).

It is clear that Aα (G) is the adjacency matrix if α = 0, and Aα (G) is essentially equivalent
to signless Laplacian matrix if α = 1/2.
When α ≥ 1/2, Nikiforov [8] proved that λ2 (Aα (G)) ≤ αn − 1. This implies that
λk (Aα (G)) ≤ λ2 (Aα (G)) ≤ αn − 1 for k ≥ 2. de Lima and Nikiforov [6] showed that
λk (A 12 (G)) = 12 n −1 for k ≥ 2 if and only if G has either k balanced bipartite components
or k + 1 bipartite components. Let ∂1 (G) ≥ · · · ≥ ∂n (G) denote the distance signless
Laplacian spectrum. Lin and Das [7] proved ∂n (G) = n − 2 if and only if Gc contains
either a balanced bipartite graph or at least two bipartite components. In this paper, we
characterize the graphs with λk (Aα (G)) = αn − 1 for k ≥ 2 when α > 1/2, these graphs
are not the same as those for α = 1/2.

Theorem 1.1. Let G be a graph of order n. If α > 1/2, then λk (Aα (G)) = αn −1 for k ≥ 2
if and only if G has k vertices of degree n − 1.

It is well-known that λn (A 12 (G)) = 0 if G is bipartite. So it is interesting to give a

lower bounds of λn (A 12 (G)) when G is non-bipartite. In 2007, Cvetković, Rowlinson and
Simić [2] proposed the following conjecture:

Conjecture 1.2. Let G be a non-bipartite graph with order n. Then

λn (A 12 (G)) ≥ λn (A 12 (G∗ )),

equality holds if and only if G ∼

= G∗ , where G∗ is the unicyclic graph obtained from a
triangle by attaching a path at one of its end vertices.

One year later, the conjecture was confirmed by Cardoso, Cvetković, Rowlinson and
Simić [1]. In [5], de Lima, Oliveira, de Abreu and Nikiforov proved that λn (A 12 (G)) ≤ m
n−
1
2 and in the same paper, they proposed a conjecture on the lower bound of λn (A 2 (G)).
1

Conjecture 1.3. Let G be a graph with order n and size m. Then

m n−2
λn (A 12 (G)) ≥ − .
n−1 2

Guo, Chen and Yu [3] proved a stronger result, λn (A 12 (G)) ≥ n−2

m
− n−1
2 . For α > 1/2,
Nikiforov ([8], Proposition 7) showed that Aα (G) is positive definite if G has no isolated
vertices, that is, λn (Aα (G)) > 0. In the following, we give a lower bound on λn (Aα (G))
when G has no isolated vertices.

1
Theorem 1.4. Let G be a graph of order n. If G has no isolated vertices and 2 < α < 1,
then
212 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219

λn (Aα (G)) ≥ 2α − 1,

the equality holds if and only if G has a component isomorphic to K2 .

The rest of the paper is organized as follows. In Section 2, we study the k-th largest
Aα -eigenvalue, and give the proof of Theorem 1.1. In Section 3, we present the proof of
Theorem 1.4, and we also give some results on the smallest Aα -eigenvalue.

2. The k-th largest eigenvalue of Aα (G)

Let G be a graph with vertex set V (G) and edge set E(G). The degree of a vertex
v ∈ V (G) is denoted by dG (v). We use N (v) to denote the set of vertices of G which
are adjacent to v. We first list some fundamental properties of the matrix Aα (G), which
can be found in [8–10]. Let x be a vector on the vertices of G. We use x(v) to denote
the entry of x corresponding to the vertex v ∈ V (G). Suppose that λ is an eigenvalue
of Aα (G). If x is an eigenvector of Aα (G) with respect to λ, then the eigenequations for
Aα (G) can be written as

λx(v) = αdG (v)x(v) + (1 − α) x(u),
u∈N (v)

where v ∈ V (G). The quadratic form Aα (G)x, x can be represented in several ways:

Aα (G)x, x = (αx(u)2 + 2(1 − α)x(u)x(v) + αx(v)2 ),
uv∈E(G)
 
Aα (G)x, x = (2α − 1) x(u)2 dG (u) + (1 − α) (x(u) + x(v))2 ,
u∈V (G) uv∈E(G)
 
Aα (G)x, x = α x(u)2 dG (u) + 2(1 − α) x(u)x(v).
u∈V (G) uv∈E(G)

For convenience we state below the complete theorem of Weyl and So.

Theorem 2.1. ([11]) Let A and B be n × n Hermitian matrices and C = A + B. Then

λi (C) ≤ λj (A) + λi−j+1 (B)(n ≥ i ≥ j ≥ 1),

λi (C) ≥ λj (A) + λi−j+n (B)(1 ≤ i ≤ j ≤ n).

In either of these inequalities equality holds if and only if there exists a nonzero n-vector
that is an eigenvector to each of the three involved eigenvalues.

Proposition 2.2. ([6]) Let 2 ≤ k < n and A and B be self-adjoint operators of order n.
If for every s = 2, . . . , k,
 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219 213

λs (A) + λn (B) = λs (A + B),

then there exist k − 1 nonzero orthogonal n-vectors x1 , . . . , xk−1 such that

Axs−1 = λs (A)xs−1 , Bxs−1 = λn (B)xs−1 and (A + B)xs−1 = λs (A + B)xs−1

for every s = 2, . . . , k.

Let α ≥ 1/2. We use G − e to denote the graph obtained from G by deleting an

edge e ∈ E(G). Thus, Aα (G) = Aα (G − e) + N , where N is the Aα -matrix of a graph
containing only one edge. In [8], it was proved that the Aα -matrix is positive semi-definite
if α ≥ 1/2. Hence, the matrices Aα (G), Aα (G − e) and N are all positive semi-definite.
By using monotonicity theorem (see [4], Corollary 4.3.3), the following result holds.

Proposition 2.3. Let G be a graph of order n. If e ∈ E(G) and α ≥ 1/2, then

λi (Aα (G)) ≥ λi (Aα (G − e))

for 1 ≤ i ≤ n.

If S ⊆ V (G), then we use G[S] to denote the subgraph of G induced by S.

Proposition 2.4. Let G be a graph with Aα (G) and 0 ≤ α ≤ 1. Let S ⊆ V (G) and |S| = k.
Suppose that dG (w) = d for each vertex w ∈ S, and N (v)\{u} = N (u)\{v} for any two
vertices u, v ∈ S. Then we have the following statements.

(1) If G[S] is a clique, then (d + 1)α − 1 is an eigenvalue of Aα (G) with multiplicity at

least k − 1.
(2) If G[S] is an independent set, then dα is an eigenvalue of Aα (G) with multiplicity
at least k − 1.

Proof. Let S = {v1 , v2 , . . . , vk }. Clearly, dG (v1 ) = · · · = dG (vk ) = d. Let x1 , x2 , . . . , xk−1

be vectors such that
⎧
⎪
⎨ xi (v1 ) = 1,
xi (vi+1 ) = −1,
⎪
⎩ x (v) = 0 if v ∈ V (G)\{v , v },
i 1 i+1

for i = 1, . . . , k − 1. Suppose that G[S] is a clique. One can easily obtain that

Aα (G)xi = ((dG (v1 ) + 1)α − 1, 0, . . . , 0, 1 − (dG (vi ) + 1)α, 0, . . . , 0) = ((d + 1)α − 1)xi ,

for i = 1, . . . , k − 1. Hence, (d + 1)α − 1 is an eigenvalue of Aα (G) and x1 , . . . , xk−1 are

eigenvectors of Aα (G) corresponding to (d + 1)α − 1. Moreover, since x1 , . . . , xk−1 are
linearly independent, the multiplicity of (d + 1)α − 1 is at least k − 1.
214 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219

When G[S] is an independent set, we have

Aα (G)xi = (dG (v1 )α, 0, . . . , 0, −dG (vi )α, 0, . . . , 0) = dαxi ,

for i = 1, 2, . . . , k − 1. Since x1 , . . . , xk−1 are linearly independent, it follows that dα is

an eigenvalue of Aα (G) with multiplicity at least k − 1. Thus we complete the proof. 2

The graph Ks ∨(n −s)K1 is called complete split graph, denoted by CSs,n−s . According
to Proposition 2.4 and [[8], Proposition 37], we determine all Aα -eigenvalues of CSs,n−s .

Corollary 2.5. The Aα -spectrum of CSs,n−s contains nα − 1 with multiplicity s − 1, sα

with multiplicity n − s − 1 and the remaining two Aα -eigenvalues are

nα + s − 1 ± (nα + s − 1)2 − 4sα(s − 1) + 4s(n − s)(1 − 2α)
.
2

At the end of this section, we give the proof of Theorem 1.1.

Proof of Theorem 1.1. If G has at least k vertices of degree n − 1, then it follows from
Proposition 2.4 that nα − 1 is an eigenvalue of Aα (G) with multiplicity at least k − 1.
Since CSk,n−k is a subgraph of G, we have λ1 (Aα (G)) ≥ λ1 (Aα (CSk,n−k )) > nα − 1,
and so λk (Aα (G)) = αn − 1. On the other hand, we assume that λk (Aα (G)) = αn − 1.
Hence, λ2 (Aα (G)) = · · · = λk (Aα (G)) = αn − 1. If k = n, then G ∼ = Kn (otherwise,
λn (Aα (G)) ≤ λn (Aα (CSn−2,2 ) < nα−1)). Suppose k ≤ n −1. Let G be the complement
c

of G. Note that

Aα (G) + Aα (Gc ) = Aα (Kn ).

By Theorem 2.1, we have

λi (Aα (G)) + λn (Aα (Gc )) ≤ λi (Aα (Kn )) for 2 ≤ i ≤ n.

Since λk (Aα (G)) = αn − 1 and λk (Aα (Kn )) = αn − 1, it follows that λn (Aα (Gc )) = 0.
Hence,

λi (Aα (G)) + λn (Aα (Gc )) = λi (Aα (Kn ))

for every i = 2, . . . , k. According to Proposition 2.2, we can see that there exist k − 1
nonzero orthogonal n-vectors x1 , . . . , xk−1 such that

Aα (G)xi−1 = λi (Aα (G))xi−1 , Aα (Gc )xi−1 = λn (Aα (Gc ))xi−1 and

Aα (Kn )xi−1 = λi (Aα (Kn ))xi−1
 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219 215

for every i = 2, . . . , k. Let V ∗ = {v ∈ V (G) : dG (v) = n − 1}. Since λn (Aα (Gc )) = 0, Gc

contains isolated vertex, and so V ∗
= ∅. Suppose V ∗ = {v1 , v2 , . . . , vt }. In the following,
we only need to show t ≥ k. For 1 ≤ i ≤ k − 1, since Aα (Gc )xi = λn (Aα (Gc ))xi and
λn (Aα (Gc )) = 0, we have
 
0 = Aα (Gc )xi , xi  = (2α − 1) dGc (u)xi (u)2 + (1 − α) (xi (u) + xi (v))2 .
u∈V (Gc ) uv∈E(Gc )

Since α > 1/2, one can see that xi (u) = 0 for all u ∈ V (G)\V ∗ . Moreover, since xi is also

an eigenvector of λi+1 (Aα (Kn )) = αn − 1, we have 1 xi = 0 and hence v∈V ∗ xi (v) = 0,
that is, xi (v1 ) = −(xi (v2 ) + · · · + xi (vt )). Suppose that y1 , . . . , yt−1 are n-vectors such
that yi (v1 ) = −1, yi (vi+1 ) = 1 and yi (v) = 0 if v ∈ V (G)\{v1 , vi+1 }, for i = 1, . . . , t − 1.
Thus, it is easy to see that

xi = xi (v2 )y1 + xi (v3 )y2 + · · · + xi (vt )yt−1

for i = 1, . . . , k − 1. Therefore, x1 , . . . , xk−1 can be represented by using y1 , . . . , yt−1 .

Since these two vector families are both linearly independent, it follows that t ≥ k, thus
we complete the proof. 2

3. The smallest eigenvalue of Aα (G)

In this section, we study the smallest eigenvalue of Aα (G). First, we will give the
proof of Theorem 1.4. Before proceeding, the following lemma is needed.

Lemma 3.1. Let T be a tree of order n ≥ 2. If 1

2 < α < 1, then λn (Aα (T )) ≥ 2α − 1 with
equality if and only if T ∼
= K2 .

Proof. Clearly, λ2 (Aα (K2 )) = 2α − 1. It suffices to show that λn (Aα (T )) > 2α − 1 if

n ≥ 3. We prove this by induction on the order n. Recall that the smallest eigenvalue of
star K1,n−1 is

1 
λn (Aα (K1,n−1 )) = (αn − α2 n2 + 4(n − 1)(1 − 2α)).
2

It is easy to check that 12 (αn − α2 n2 + 4(n − 1)(1 − 2α)) > 2α − 1, if 12 < α < 1 and
n ≥ 3. Hence, the result follows when T is a star. So, in the following we may assume
that T  K1,n−1 . Then there exists a non-pendent edge e such that T − e = T1 ∪ T2 and
|V (T1 )| ≥ |V (T2 )| ≥ 2.
Case 1. |V (T1 )| ≥ 3 and |V (T2 )| ≥ 3.
By the induction hypothesis, we have λ|V (T1 )| (Aα (T1 )) > 2α−1 and λ|V (T2 )| (Aα (T2 )) >
2α − 1. Note that λn (Aα (T − e)) = min{λ|V (T1 )| (Aα (T1 )), λ|V (T2 )| (Aα (T2 ))}. According
to Proposition 2.3, it follows that
216 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219

λn (Aα (T )) ≥ λn (Aα (T − e)) = min{λ|V (T1 )| (Aα (T1 )), λ|V (T2 )| (Aα (T2 ))} > 2α − 1,

as required.
Case 2. |V (T1 )| = 2 and |V (T2 )| = 2.
Thus T ∼= P4 . By a simple calculation, we have

1 1 2 1 1 2
λ4 (Aα (P4 )) = min{α + − 4α − 8α + 5, 2α − − 8α − 12α + 5}.
2 2 2 2
√
It is easy to check that both α + 12 − 12 4α2 − 8α + 5 > 2α − 1 and 2α − 1
−
√ 2
2 8α − 12α + 5 > 2α − 1, hence λ4 (Aα (P4 )) > 2α − 1.
1 2

Case 3. |V (T1 )| ≥ 3 and |V (T2 )| = 2.

That is T − e = T1 ∪ K2 . We may assume that e = uv and K2 = vw. Hence

Aα (T ) = Aα (T1 ∪ K2 ) + M

where M = Aα (K2 ∪ (n − 2)K1 ). The eigenvalues of M are 1, 2α − 1, 0, . . . , 0. According

to Theorem 2.1, we have

λn (Aα (T )) ≥ λn (Aα (T1 ∪ K2 )) + λn (M ).

By the induction hypothesis, we have λ|V (T1 )| (Aα (T1 )) > 2α−1, then λn (Aα (T1 ∪K2 )) =
λ2 (Aα (K2 )) = 2α − 1. And since λn (M ) = 0, therefore

λn (Aα (T )) ≥ λn (Aα (T1 ∪ K2 )) + λn (M ) = 2α − 1.

If λn (Aα (T )) = 2α−1, then it follows from Theorem 2.1 that λn (Aα (T )), λn (Aα (T1 ∪K2 ))
and λn (M ) share a common eigenvector (say x), that is,

Aα (T )x = λn (Aα (T ))x, Aα (T1 ∪ K2 )x = λn (Aα (T1 ∪ K2 ))x and M x = λn (M )x.

Since M x = λn (M )x and λn (M ) = 0, we have αx(u) + (1 − α)x(v) = 0 and αx(v) + (1 −

α)x(u) = 0, this implies that x(u) = x(v) = 0. Since Aα (T1 ∪ K2 )x = λn (Aα (T1 ∪ K2 ))x
and λn (Aα (T1 ∪ K2 )) = 2α − 1, we have (2α − 1)x(v) = αx(v) + (1 − α)x(w), then
x(w) = 0. Let y be an (n − 2)-vector obtained from x by deleting the entries x(v), x(w).
Since Aα (T1 ∪ K2 )x = (2α − 1)x and x(v) = x(w) = 0, it is easy to check that Aα (T1 )y =
(2α − 1)y. It leads to that 2α − 1 is an eigenvalue of Aα (T1 ), this is contrary to the
induction hypothesis. Hence λn (Aα (T )) > 2α − 1, which completes the proof of this
lemma. 2

Proof of Theorem 1.4. Suppose that G contains k components, say, G1 , . . . , Gk . Then

λn (Aα (G)) = min{λ|V (Gi )| (Aα (Gi ))|i = 1, . . . , k}.

 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219 217

Let Ti be a spanning tree of Gi for 1 ≤ i ≤ k. It follows from Proposition 2.3 that

λ|V (Gi )| (Aα (Gi )) ≥ λ|V (Ti )| (Aα (Ti )).

Now the theorem follows by Lemma 3.1. 2

The smallest Aα -eigenvalue was also studied in [8]. In particular, the author presented
the following problem: when α ∈ (0, 1/2), determine how small the least Aα -eigenvalue
can be for a graph of given order (see [8] Problem 30). When 1/2 < α < 1, it is interesting
to find out which connected graph with order n minimizes the smallest eigenvalue of
Aα (G). Based on our numerical experiments, we propose the following conjecture.

1
Conjecture 3.2. Let G be a connected graph of order n. If 2 < α < 1, then

λn (Aα (G)) ≥ λn (Aα (K1,n−1 )),

the equality holds if and only if G ∼

= K1,n−1 .

Finally, we will give an upper bound for the smallest Aα -eigenvalue of a bipartite
graph.

1
Theorem 3.3. Let G be a bipartite graph of order n. If 2 < α < 1, then

λn (Aα (G)) ≤ λn (Aα (K n2 , n2  ))

with equality if and only if G ∼

= K n2 , n2  .

Proof. Let G be a bipartite graph with partitions V1 and V2 . Suppose that |V1 | = a and
|V2 | = b. Without loss of generality, we may assume that a ≥ b.

Fact 1. If a − b ≥ 2, then λn (Aα (Ka,b )) < λn (Aα (Ka−1,b+1 )).


Nikiforov [8] showed that λn (Aα (Ka,b )) = 1
2 (αn − α2 n2 + 4ab(1 − 2α)). Then we
have

λn (Aα (Ka−1,b+1 )) − λn (Aα (Ka,b ))

1  2 2 

= α n + 4ab(1 − 2α) − α2 n2 + 4(a − 1)(b + 1)(1 − 2α)

2
> 0,

as required.

Fact 2. λn (Aα (K n2 , n2  − e)) < λn (Aα (K n2 , n2  )).
218 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219

For convenience, let H ∼

= K n2 , n2  . Suppose that x is a unit eigenvector of Aα (H)
corresponding to λn (Aα (H)). Hence,

λn (Aα (H)) = Aα (H)x, x

 
= (2α − 1) (x(u)2 + x(v)2 ) + (1 − α) (x(u) + x(v))2 .
uv∈E(H) uv∈E(H)

Since λn (Aα (H)) > 0, there exists an edge e = uv such that (2α − 1)(x(u)2 + x(v)2 ) +
(1 − α)(x(u) + x(v))2 > 0. Therefore,

λn (Aα (H − uv)) − λn (Aα (H))

≤ Aα (H − uv)x, x − Aα (H)x, x
= −(2α − 1)(x(u)2 + x(v)2 ) − (1 − α)(x(u) + x(v))2
< 0.

This completes the proof of Fact 2.

According to Fact 1, Fact 2 and Proposition 2.3, we can see that: if a − b ≥ 2, then
λn (G) ≤ λn (Aα (Ka,b )) < λn (Aα (K n2 , n2  )); if a − b ≤ 1 and G  K n2 , n2  , then
λn (G) ≤ λn (Aα (K n2 , n2  − e)) < λn (Aα (K n2 , n2  )). Thus we complete the proof. 2

Acknowledgements

The authors would like to express their gratitude to the anonymous reviewers for
their kind suggestions on the original manuscript. The first author was supported by
the National Natural Science Foundation of China (No. 11401211) and Fundamental
Research Funds for the Central Universities (No. 222201714049). The third author was
supported by the National Natural Science Foundation of China (No. 11471121).

References

[1] D.M. Cardoso, D. Cvetković, P. Rowlinson, S.K. Simić, A sharp lower bound for the least eigenvalue
of the signless Laplacian of a non-bipartite graph, Linear Algebra Appl. 429 (2008) 2770–2780.
[2] D. Cvetković, P. Rowlinson, S.K. Simić, Eigenvalue bounds for the signless Laplacian, Publ. Inst.
Math. (Beograd) 81 (95) (2007) 11–27.
[3] S. Guo, Y. Chen, G. Yu, A lower bound on the least signless Laplacian eigenvalue of a graph, Linear
Algebra Appl. 448 (2014) 217–221.
[4] R.A. Horn, C.R. Johnson, Matrix Analysis, Cambridge University Press, Cambridge, 1985.
[5] L.S. de Lima, C.S. Oliveira, N.M.M. de Abreu, V. Nikiforov, The smallest eigenvalue of the signless
Laplacian, Linear Algebra Appl. 435 (2011) 2570–2584.
[6] L.S. de Lima, V. Nikiforov, On the second largest eigenvalue of the signless Laplacian, Linear
Algebra Appl. 438 (2013) 1215–1222.
[7] H. Lin, K.Ch. Das, Characterization of extremal graphs from distance signless Laplacian eigenvalues,
Linear Algebra Appl. 500 (2016) 77–87.
[8] V. Nikiforov, Merging the A- and Q-spectral theories, Appl. Anal. Discrete Math. 11 (2017) 81–107.
[9] V. Nikiforov, G. Pastén, O. Rojo, R.L. Soto, On the Aα -spectra of trees, Linear Algebra Appl. 520
(2017) 286–305.
 H. Lin et al. / Linear Algebra and its Applications 556 (2018) 210–219 219

[10] V. Nikiforov, O. Rojo, A note on the positive semidefiniteness of Aα (G), Linear Algebra Appl. 519
(2017) 156–163.
[11] W. So, Commutativity and spectra of Hermitian matrices, Linear Algebra Appl. 212/213 (1994)
121–129.