J Math Chem

DOI 10.1007/s10910-009-9570-0
ORIGINAL PAPER

The Estrada index of chemical trees

Aleksandar Ilic Dragan Stevanovic

Received: 26 January 2009 / Accepted: 30 June 2009

Springer Science+Business Media, LLC 2009

Abstract Let G be a simple graph with n vertices and let 1 , 2 , . . . , n be the

eigenvalues of its adjacency matrix. The Estrada index

nof G is a recently introduced
e i , proposed as a measure
molecular structure descriptor, defined as E E(G) = i=1
of branching in alkanes. In order to support this proposal, we prove that among the
trees with fixed maximum degree , the broom Bn, , consisting of a star S+1 and a
path of length n 1 attached to an arbitrary pendent vertex of the star, is the unique
tree which minimizes even spectral moments and the Estrada index, and then show the
relation E E(Sn ) = E E(Bn,n1 ) > E E(Bn,n2 ) > > E E(Bn,3 ) > E E(Bn,2 ) =
E E(Pn ). We also determine the trees with minimum Estrada index among the trees
with perfect matching and maximum degree . On the other hand, we strengthen
a conjecture of Gutman et al. [Z. Naturforsch. 62a (2007), 495] that the Volkmann
trees have maximal Estrada index among the trees with fixed maximum degree ,
by conjecturing that the Volkmann trees also have maximal even spectral moments
of any order. As a first step in this direction, we characterize the starlike trees which
maximize even spectral moments and the Estrada index.
Keywords Estrada index Branching Broom graph Chemical trees Volkmann
trees Spectral moments

A. Ilic (B)
Faculty of Sciences and Mathematics, University of Ni, Viegradska 33, 18000 Ni, Serbia
e-mail: [email protected]
D. Stevanovic
University of PrimorskaFAMNIT, Glagoljaka 8, 6000 Koper, Slovenia
D. Stevanovic
Mathematical Institute, Serbian Academy of Science and Arts, Knez Mihajlova 36, 11000 Belgrade,
Serbia
e-mail: [email protected]

123

J Math Chem

1 Introduction
Let G = (V, E) be a simple graph with |V | = n vertices and |E| = m edges. The
spectrum of G consists of the eigenvalues
1
2

n
of its adjacency matrix A. The Estrada index is defined as
E E(G) =

n


ei .

i=1

In the last ten years, the Estrada index found applications in measuring the degree
of protein folding [13], the centrality of complex networks (such as neural, social,
metabolic, proteinprotein interaction networks, and the World Wide Web) [4], and it
was also proposed as a measure of molecular branching, accounting for the effects of
all atoms in the molecule, giving higher weight to the nearest neighbors [5]. Within
groups of isomers, E E was found to increase with the increasing extent of branching
of the carbon-atom skeleton [6]. Also, E E characterizes the structure of alkanes via
electronic partition function [7].
Some mathematical properties of the Estrada index were reported in [814]. It
is natural problem to study chemical trees (trees with maximum degree four) and
generally trees with bounded maximum degree [15,16].
Our goal here is to add some further evidence to support the use of E E as a measure
of branching in alkanes. While the measure of branching cannot be formally defined,
there are several properties that any proposed measure has to satisfy [17], [18]. Basically, a topological index (TI) acceptable as a measure of branching must satisfy the
inequalities
T I (Pn ) < T I (X n ) < T I (Sn )

or

T I (Pn ) > T I (X n ) > T I (Sn ),

for n = 5, 6, . . ., where Pn is the path, and Sn is the star on n vertices. For example,
the first relation is obeyed by the largest graph eigenvalue [19] and E E [20], while the
second relation is obeyed by the Wiener index [21], Hosoya index and graph energy
[22].
We refine the above relation by showing that
E E(Sn ) = E E(Bn,n1 ) > E E(Bn,n2 ) > > E E(Bn,3 ) > E E(Bn,2 )
= E E(Pn ),

(1)

where the broom Bn, is a tree consisting of a star S+1 and a path of length n  1
attached to an arbitrary pendent vertex of the star (see Fig. 1). It is proven in [23]
that among trees with perfect matching and maximum degree equal to , the broom
Bn, uniquely minimizes the largest eigenvalue of adjacency matrix. Also it is shown
that among trees with bounded degree , the broom has minimal Wiener index and

123

J Math Chem
Fig. 1 The broom B11,6

Laplacian-like energy [24]. In [25] and [26] the broom has minimal energy among
trees with fixed diameter or fixed number of pendant vertices.
In particular, we show in Sect. 3 that the broom Bn, has minimal E E among the
trees on n vertices and maximum degree , and we also determine the tree with the
second minimal E E. Next, in Sect. 4 we determine the tree with minimal E E among
n-vertex trees with perfect matching and maximum degree , and prove a relation
analogous to (1).
The question may also be asked which are the most branched alkanes [27]. This
problem was examined in due detail [18] and several molecular structure-descriptors
imply that the most branched alkanes are those represented by the Volkmann trees. In
Sect. 5 we strengthen a conjecture from [6] that the Volkmann trees have maximal E E
among the trees with fixed maximum degree , by conjecturing that the Volkmann
trees also have maximal even spectral moments of any order. As a first step in this direction, we characterize the starlike trees which maximize even spectral moments and E E.
2 Preliminaries
In our proofs, we will use a connection between EE and the spectral moments of a
graph. For k
0, we denote by Mk the kth spectral moment of G,
Mk (G) =

n


ik .

i=1

A walk of length k in G is any sequence of vertices and edges of G,

w = w0 , e1 , w1 , e2 , . . . , wk1 , ek , wk ,
such that ei is the edge joining wi1 and wi for every i = 1, 2, . . . , k. The walk is
closed if w0 = wk . It is well-known (see [28]) that Mk (G) represents the number
of closed walks of length k in G. Obviously, for every graph M0 = n, M1 = 0 and
M2 = 2m. For the third moment we have M3 = 6t, where t is the number of triangles
in graph G. The fourth moment is equal to
M4 = 2

n


di2 2m + 8q,

(2)

i=1

where di is the degree of the ith vertex, and q the number of quadrangles in G.

123

J Math Chem

From the Taylor expansion of e x , we have that the Estrada index and the spectral
moments of G are related by


Mk
.
E E(G) =
k!

(3)

k=0

Thus, if for two graphs G and H we have Mk (G)
Mk (H ) for all k
0, then
E E(G)
E E(H ). Moreover, if the strict inequality Mk (G) > Mk (H ) holds for at
least one value of k, then E E(G) > E E(H ).
The -starlike tree T (n 1 , n 2 , . . . , n  ) is a tree composed of the root v, and the
paths P1 , P2 , . . . , P of lengths n 1 , n 2 , . . . , n  attached at v. The number of vertices
of a tree T (n 1 , n 2 , . . . , n  ) equals n = n 1 + n 2 + + n  + 1. The -starlike
tree is balanced if all paths have almost equal lengths, i.e., |n i n j |  1 for every
1  i  j  . Notice that the broom Bn, = T (n , 1, 1, . . . , 1) is a -starlike
tree.

3 The minimum spectral moments and Estrada index

Let Mk (n, i) be the number of closed walks of length k starting from the vertex vi in
the path Pn = v0 v1 v2 . . . vn . Since every tree is a bipartite graph, we have identity
M2k+1 (n, i) = 0 for 0  i  n. By symmetry, we have Mk (n, i) = Mk (n, n i).
Recall that the sequence of numbers c1 , c2 , . . . , ck is unimodal if there exists no
indices 1  p < q < r  k such that c p > cq < cr .
Lemma 3.1 For every even k > 2 holds

 n 
  n 
Mk (n, 0)  Mk (n, 1)   Mk n,
1  Mk n,
.
2
2
For sufficiently large k, strict inequalities hold.
Proof Let A be the adjacency matrix of the path Pn . It is well-known that the element
(Ak )i, j represents the number of walks of length k from vertex i to vertex j. We will
prove that the matrix Ak has the following property: every diagonal parallel to the
main diagonal is unimodal. First, the matrix Ak is symmetric. Next, by automorphism
f : {v0 , v1 , . . . , vn } {v0 , v1 , . . . , vn } of Pn , given by f (vi ) = vni , we can see
that (Ak )i, j = (Ak )ni,n j . Therefore, we only have to prove that the diagonals that
are parallel to the main diagonal, increase for i + j  n.
We will proceed by mathematical induction on k and prove that for all 1  i, j  n
such that i + j  n, holds:
(Ak )i, j
(Ak )i1, j1 .

123

(4)

J Math Chem

For k = 0 and k = 1, this is obvious. Using the recurrent formulas derived from
matrix multiplication, we have:
(Ak+1 )i1, j1 = (Ak )i1, j2 + (Ak )i1, j
(Ak+1 )i, j = (Ak )i, j1 + (Ak )i, j+1
From the induction hypothesis we have (Ak )i, j1
(Ak )i1, j2 . If i + j + 1  n,
we also have (Ak )i, j+1
(Ak )i1, j . For i + j + 1 = n + 1, i.e., j = n i, we have
that (Ak )i1, j = (Ak )i, j+1 . This proves inequality (4).
For the strict inequality, consider two neighboring rows i 1 and i. Eventually,
the element (Ak )i,0 becomes nonzero and it forces the strict inequality (Ak+1 )i,1 >
(Ak+1 )i1,0 . This causes the chain of strict inequalities (Ak+2 )i,2 > (Ak+2 )i1,1 ,
(Ak+3 )i,3 > (Ak+3 )i1,2 , and finally
(Ak+ni )i,ni > (Ak+ni )i1,ni1 .
The number of closed walks of length k starting from the vertex i equals to the
element (i, i) in matrix Ak . Therefore,
Mk (n, i) = (Ak )i,i
and using inequalities (4) we conclude that Mk (n, i)  Mk (n, i+1) for every 0  i 

 n2  1. If k is large enough, we have strict inequality Mk (n, i) < Mk (n, i + 1). 
Theorem 3.2 Let w be a vertex of the nontrivial connected graph G and for nonnegative integers p and q, let G( p, q) denote the graph obtained from G by attaching
pendent paths P = wv1 v2 . . . v p and Q = wu 1 u 2 . . . u q of lengths p and q, respectively, at w. If p
q
1, then
Mk (G( p, q))
Mk (G( p + 1, q 1)).

(5)

For sufficiently large k, the strict inequalities hold.

Proof Since this transformation does not affect the number of vertices or the edges,
we have that Mk (G( p, q)) = Mk (G( p + 1, q 1)) for k = 0, 1, 2.
Thus, assume that k
3 is fixed. Inequality (5) will be proved if we show that
G( p, q) has more closed walks of length k than G( p + 1, q 1). For that purpose, we
will construct an injection i from the set W k = Wk (G( p + 1, q 1)) of closed walks
of length k in G( p + 1, q 1) to the set Wk = Wk (G( p, q)) of closed walks of length
k in G( p, q). Denote with P  and Q  , the paths wv1 v2 . . . v p v p+1 and wu 1 u 2 . . . u q1 .
Consider an arbitrary closed walk C  from W  . If C  is entirely contained in G
or does not contain the edge (v p , v p+1 ), then i (C  ) = C  is also a closed walk in
G( p, q). If C  is entirely contained in the union of paths P  Q  , we can construct the
corresponding walk i (C  ) in P Q by shifting it for one place, since both unions are
isomorphic to Pp+q+1 . Therefore, we can assume that C  contains the edge (v p , v p+1 )

123

J Math Chem

and some vertices from G. Without loss of generality, assume that the closed walk C 
starts at vertex w and that it is decomposed as follows:
C  = C1 C2 C3 C4 . . . ,

(6)

where Ci is a closed walk that starts at w that is either contained completely in P  Q 
or in G. Let l1 , l2 , l3 , . . . be the lengths of closed walks C1 , C2 , C3 , . . ., respectively.
Since q  p, from Lemma 3.1 we have that
Ml j ( p + q + 1, q 1)  Ml j ( p + q + 1, q).
Thus, one can construct an injection i j mapping the closed walks of length l j starting
at w in P  Q  into the closed walks of length l j starting at w in P Q.

, i
0 are contained in G, while the walks
Now, if we assume that the walks C2i+1



C2i , i
1 are contained in P Q , then
i (C  ) = C1 il2 (C2 ) C3 il4 (C4 ) . . .
is a corresponding closed walk in G( p, q). The mapping i is an injection by its
construction, and that fact proves (5).
For k large enough and the closed walks which are decomposable by (6) into a
fixed number of parts, at least one of the numbers l2 , l4 , . . . is large enough such
that the strict inequality Ml2 j ( p + q + 1, q 1) < Ml2 j ( p + q + 1, q) holds. Then


Mk (G( p + 1, q 1)) < Mk (G( p, q)) holds as well.
Corollary 3.3 Let w be a vertex of the nontrivial connected graph G and for nonnegative integers p and q, let G( p, q) denote the graph obtained from G by attaching
pendent paths P = wv1 v2 . . . v p and Q = wu 1 u 2 . . . u q of lengths p and q, respectively, at w. If p
q
1, then
E E(G( p, q)) > E E(G( p + 1, q 1)).
Proof The result follows directly from (3) and previous theorem.




Theorem 3.4 Let T

= Bn, be an arbitrary tree on n vertices with the maximum

vertex degree . Then
E E(Bn, ) < E E(T ).

(7)

Proof Fix a vertex v of degree  as a root and let k > 2 be an even number. Let
T1 , T2 , . . . , T be the trees attached at v. We can repeatedly apply the transformation
from Corollary 3.3 at any vertex of degree at least three with largest eccentricity from
the root in every tree Ti , as long as Ti does not become a path. From Corollary 3.3
it follows that each application of this transformation strictly decreases its Estrada
index.

123

J Math Chem

When all trees T1 , T2 , . . . , T turn into paths, we can again apply Corollary 3.3
at the vertex v as long as there exists at least two paths of length at least two,
further decreasing the Estrada index. In the end of this process, we arrive at the broom


Bn, .

= T (n  1, 2, 1, . . . , 1) be a starlike tree obtained from Bn, by
Let Bn,
removing the last edge from the longest path and attaching it to another pendent vertex.

has the second minimal E E among trees
From the above proof, we also get that Bn,
with maximum vertex degree .

be an arbitrary tree on n vertices with the maxiCorollary 3.5 Let T

= Bn, , Bn,
mum vertex degree . Then

) < E E(T ).
E E(Bn,

It was conjectured in [29] that the path Pn has minimum Estrada index among
connected graphs on n vertices. While the conjecture has been proved recently in [20],
Corollary 3.3 enables us to give a new proof of this conjecture. Namely, deleting an
edge from a graph decreases the number of closed walks, and by (3) also decreases the
Estrada index. Thus, any graph will have larger Estrada index than its spanning tree,
showing that the graph with minimum Estrada index has to be a tree itself. Next, from
Theorem 3.4 we know that the minimum Estrada index among trees on n vertices
is achieved for one of the brooms Bn, . If  > 2, we can apply the transformation from Corollary 3.3 at the vertex of degree  in Bn, and obtain Bn,1 . Thus,
E E(Bn, ) > E E(Bn,1 ) for  > 2, which shows that
E E(Sn )=E E(Bn,n1 ) > E E(Bn,n2 ) > > E E(Bn,3 ) > E E(Bn,2 )=E E(Pn ).
In [30] the authors estimate the Estrada index of a path Pn and proved that
E E(Pn ) (n + 1)I0 cosh 2,
where I0 =
of

1
k=0 (k!)2

{0}

= 2.27958530. Since, the spectra of graph Bn,3

= Z n consists



(2k + 1)
| k = 0, 1, 2, . . . , n 2 ,
2 cos
2(n 1)

we have the following approximation E E(Z n ) n I0 from [31]. The path Pn is the
unique graph with maximal degree 2, and therefore Z n has the second minimal Estrada
index among trees on n vertices.
4 The minimum Estrada index of trees with perfect matchings
It is well known that if a tree T has a perfect matching, then the perfect matching M
is unique: namely, a pendent vertex v has to be matched with its unique neighbor w,
and then M {vw} forms the perfect matching of T v w.

123

J Math Chem

Fig. 2 The tree A16,6

Let An, be a -starlike tree T (n 2, 2, 2, . . . , 2, 1) consisting of a central vertex v, a pendant edge, a pendant path of length n 2 and  2 pendant paths of
length 2, all attached at v (see Fig. 2).
Theorem 4.1 The tree An, has minimum E E among trees with perfect matching
and maximum degree .
Proof Let T be an arbitrary tree with perfect matching and let v be a vertex of degree
, with neighbors v1 , v2 , . . . , v . Let T1 , T2 , . . . , T be the maximal subtrees rooted
at v1 , v2 , . . . , v , respectively, such that neither of these trees contains v. Then at most
one of the numbers |T1 |, |T2 |, . . . , |T | can be odd (if Ti and T j have odd number of
vertices, then their roots vi and v j will be unmatched). Actually, since the number of
vertices in T is even, there exists exactly one among T1 , T2 , . . . , T with odd number
of vertices.
Using Corollary 3.3, we may transform each Ti into a pendant path attached at v
while simultaneously decreasing E E and keeping the existence of a perfect matching. Assume that T has odd number of vertices, while the remaining trees have even
number of vertices. We apply similar transformation to the one in Theorem 3.2, but
instead of moving one edge, we move two edges in order to keep the existence of a
perfect matching. Thus, if p
q
2 then from Corollary 3.3
E E(G( p, q)) > E E(G( p + 2, q 2)).
Using this transformation we may reduce T to one vertex, the trees T2 , . . . , T1 to
two vertices, leaving T1 with n 2 vertices, and thus obtaining An, . Since we have
been strictly decreasing E E at all times, we conclude that An, indeed has minimum
E E among the trees with perfect matching.


If  > 2, we can apply the transformation from Corollary 3.3 (by moving two
vertices) at the vertex of degree  in An, to obtain An,1 . Thus, it follows that
E E(An,n/2 ) > E E(An,n/21 ) > > E E(An,3 ) > E E(An,2 ) = E E(Pn ).
5 The maximum spectral moments and Estrada index
The complete -ary tree is defined as follows. Start with the root having  children.
Every vertex different from the root, which is not in one of the last two levels, has

123

J Math Chem
Fig. 3 The complete 4-ary tree
of order 21

exactly  1 children. In the last level, while not all nodes have to exist, the nodes
that do exist fill the level consecutively (see Fig. 3). Thus, at most one vertex on the
level second to last has its degree different from  and 1.
In [6] the authors proposed these trees to be called Volkmann trees, as they represent
alkanes with minimal Wiener index [32]. Volkmann trees also have maximal greatest
eigenvalue among trees with maximum degree , as shown in [33].
The computer search among trees with up to 24 vertices revealed that the complete
-ary trees attain the maximum values of E E and even spectral moments of orders
up to 16 among the trees with the maximum vertex degree .
It is well-known that an integer sequence d1
d2

dn
1 represents the
degrees sequence of a tree if and only if d1 + d2 + + dn = 2n 2. Consider an
arbitrary tree T with vertex degrees
 = d1
d2

dn1
dn = 1.
From (2), the fourth spectral moment of trees depends only on the sum of squares
of degrees,
known as the first Zagreb index [34]. In order to maximize the sum

also
n
Z (T ) = i=1
di2 , assume that there exist two vertices i and j such that 1 < di 
d j < . Simple transformation di = di 1, d j = d j + 1 strictly increases M4 , as
Z (T  ) Z (T ) = (di 1)2 + (d j + 1)2 di2 d 2j = 2(d j di + 1)
2.
Therefore, the fourth spectral moment is maximized if at most one vertex of a tree has
degree different from 1 and , which is satisfied for the complete -ary tree.
Based on this argument and the above-mentioned empirical observations, we pose
the following
Conjecture 5.1 For any k
2, the complete -ary tree has maximum spectral
moment M2k among trees on n vertices with maximum degree .
This conjecture implies the conjecture of Gutman et al. [6], thanks to Eq. (3).
While we do not have the proof of the above conjecture at the moment, we make a
humble step forward by characterizing the starlike trees which maximize even spectral
moments and the Estrada index.

123

J Math Chem

Theorem 5.2 The balanced -starlike tree has maximum even spectral moments and
maximum E E among -starlike trees of order n.
Proof Let T = T (n 1 , . . . , n  ) be an arbitrary -starlike tree. If there exists i and
j, 1 i, j , such that |n i n j | > 1, we can strictly increase its even spectral
moments by applying Theorem 3.2 repeatedly until we obtain -starlike tree with
n +n
n +n
paths of lengths  i 2 j  and  i 2 j  instead of n i and n j .
The maximality of E E in such trees is shown analogously using Corollary 3.3. 

Acknowledgments This work is supported by Research Grant 144007 and 144015G of Serbian Ministry
of Science and Technological Development and the research program P1-0285 of the Slovenian Agency
for Research. We gratefully acknowledge the insightful comments and suggestions from the two referees
that helped in improving this article.

References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.

E. Estrada, Chem. Phys. Lett. 319, 713 (2000)

E. Estrada, Bioinformatics 18, 697 (2002)
E. Estrada, Proteins 54, 727 (2004)
E. Estrada, Phys. Rev. E 75, 016103 (2007)
E. Estrada, J.A. Rodrguez-Valzquez, M. Randic, Int. J. Quantum Chem. 106, 823 (2006)
I. Gutman, B. Furtula, V. Markovic, B. Gliic, Z. Naturforsch. 62a, 495 (2007)
E. Estrada, Chem. Phys. Lett. 463, 422 (2008)
E. Estrada, J.A. Rodrguez-Valzquez, Phys. Rev. E 71, 056103 (2005)
E. Estrada, J.A. Rodrguez-Valzquez, Phys. Rev. E 72, 046105 (2005)
I. Gutman, E. Estrada, J.A. Rodrguez-Valzquez, Croat. Chem. Acta 80, 151 (2007)
I. Gutman, S. Radenkovic, A. Graovac, D. Plavic, Chem. Phys. Lett. 447, 233 (2007)
B. Zhang, B. Zhou, Z. Naturforsch. 61a, 536 (2006)
I. Gutman, S. Radenkovic, Kragujevac J. Sci. 29, 67 (2007)
B. Zhou, MATCH Commun. Math. Comput. Chem. 60, 485 (2008)
C. Heuberger, S.G. Wagner, J. Graph Theory 58, 49 (2008)
C. Heuberger, S.G. Wagner, J. Math. Chem. 46, 214 (2009)
Z. Markovic, V. Ivanov-Petrovic, I. Gutman, J. Mol. Struct. (Theochem) 629, 303 (2003)
M. Fischermann, I. Gutman, A. Hoffmann, D. Rautenbach, D. Vidovic, L. Volkmann, Z. Naturforsch.
57a, 49 (2002)
L. Lovsz, J. Pelikn, Period Math. Hung. 3, 175 (1973)
H. Deng, MATCH Commun. Math. Comput. Chem. 62 (2009)
R.C. Entringer, D.E. Jackson, D.A. Snyder, Czech. Math. J. 26, 283 (1976)
I. Gutman, Theor. Chim. Acta 45, 79 (1977)
W. Lin, X. Guo, J. Math. Chem. 42, 1057 (2007)
D. Stevanovic, MATCH Commun. Math. Comput. Chem. 61, 407 (2009)
W. Yan, L. Ye, Appl. Math. Lett. 18, 1046 (2005)
A. Yu, X. Lv, Linear Algebra Appl. 418, 625 (2006)
M. Lepovic, I. Gutman, J. Chem. Inf. Comput. Sci. 38, 823 (1998)
D. Cvetkovic, M. Doob, H. Sachs, Spectra of GraphsTheory and Applications (Academic Press,
New York, 1980)
J.A. de la Pea, I. Gutman, J. Rada, Linear Algebra Appl. 427, 70 (2007)
I. Gutman, A. Graovac, Chem. Phys. Lett. 436, 294 (2007)
Y. Ginosar, I. Gutman, T. Mansour, M. Schork, Chem. Phys. Lett. 454, 145 (2008)
M. Fischermann, A. Hoffmann, D. Rautenbach, L. Szkely, L. Volkmann, Discrete Appl.
Math. 122, 127 (2002)
S. Simic, D. Toic, MATCH Commun. Math. Comput. Chem. 54, 351 (2005)
I. Gutman, K.C. Das, MATCH Commun. Math. Comput. Chem. 50, 83 (2004)

123