On trees with smallest resolvent energies1

Mohammad Ghebleh, Ali Kanso2 , Dragan Stevanovic

Faculty of Science, Kuwait University, Safat 13060, Kuwait
[email protected], [email protected], [email protected]
(Received October 10, 2016)
Abstract
Let Pn1 (a) be the tree obtained by attaching a pendent vertex at position a of
the (n1)-vertex path Pn1 . We prove here a conjecture of Gutman et al. [MATCH
Commun. Math. Comput. Chem. 73 (2015), 267270] that Pn1 (a) has the a-th
smallest resolvent Estrada index among all trees of order n, for 2 a bn/2c, and
show that the analogous result also holds the resolvent energy.

Introduction

Let G = (V, E) be a simple graph. A walk of length k in G is a sequence of its vertices

W : v0 , . . . , vk such that vi vi+1 is an edge of G for each i = 0, . . . , k 1. A walk W is
closed if v0 = vk . Let Mk (G, v) denote the number of closed walks of length k starting and
ending at a vertex v of G. The sequence of numbers Mk (G, v), k 2, provides a certain
glimpse into the density of edges in the vicinity of v. For example, M2 (G, v) is equal to
the degree of v, M3 (G, v) is equal to twice the number of triangles containing v, while
for larger values of k, Mk (G, v) counts a mix of closed walks going up to the distance
bk/2c from v. Since for practical purposes one usually wants to have a single numerical
descriptor instead of an infinite sequence, Estrada and Highman [1, Section 3] proposed
1
2

This work was supported and funded by Kuwait University Research Grant No. SM10/15.
Corresponding author

the use of a weighted series of the closed walk counts

fc (G, v) =

ck Mk (G, v),

k0

where (ck )k0 is a predefined sequence of nonnegative weights that makes the above series
convergent. Values fc (G, v) may then be considered as the closed walk based measure of
vertex centrality, while
fc (G) =

fc (G, v) =

vV

where Mk (G) =

vV

ck Mk (G),

(1)

k0

Mk (G, v) is the total number of closed walks of length k in G,

represents a cumulative closed walk based descriptor of a graph.

The value fc (G) is closely related to the adjacency spectrum of G. Let A(G) be
the adjacency matrix of G, and let 1 (G) n (G) be its eigenvalues. It is a
folklore result that the number of walks of length k between the vertices u and v is equal
to Ak (G)u,v , easily proved by induction on k. Hence the number Mk (G) of closed walks of
length k represents the trace of Ak (G), which is further equal to the k-th spectral moment
Pn k
i=1 i (G), so that (1) may be rewritten as
fc (G) =

n X
X

ck ki (G).

(2)

i=1 k0

Estradas original suggestion [2] for the sequence (ck )k0 was ck = 1/k!, which puts more
emphasis on shorter closed walks and ensures the convergence as
fc (G) =

n X k
X
(G)
i

k!

i=1 k0

This so-called Estrada index EE(G) =

Pn

i=1

n
X

ei (G) .

i=1

ei (G) has been initially applied in measur-

ing the degree of protein folding [2, 3, 4], the centrality of complex networks [5] and the
branching of molecular graphs [6, 7]. It has been steadily gaining popularity in mathematical chemistry community, as Zentralblatt now reports more than a hundred research
articles on the Estrada index.
A more recent suggestion for (ck )k0 was made in [1]. In order to downweight shorter
closed walks, Estrada and Highman suggested the use of ck = 1/(n 1)k , inspired by the
ratio of the numbers of closed walks of length k between the pairs of vertices in G and in
the complete graph Kn . This latter choice defines the so-called resolvent Estrada index
EEr (G) =

X Mk (G)
,
(n 1)k
k0

(3)

which can also be represented in terms of eigenvalues as

EEr (G) =

n
X
i=1

n1
n 1 i (G)

when G is not a complete graph. Some bounds on EEr (G) have been obtained in [8]
and [9]. Chen and Qian [10] showed that the star Sn has the maximum resolvent Estrada
index among trees of order n, followed by a number of trees similar to the star, and on
the other hand, Gutman et al. [11] showed that the path Pn has the minimum resolvent
Estrada index among trees of order n. Let Pn1 (a) be the tree obtained by attaching
a pendent vertex at position a of the (n 1)-vertex path Pn1 . Gutman et al. [11]
further showed that Pn1 (2) has the second smallest resolvent Estrada index for n 4,
while Pn1 (3) has the third smallest resolvent Estrada index for n 6, and proposed the
following conjecture.
Conjecture 1 ([11]). For 2 a bn/2c, the tree Pn1 (a) has the a-th smallest resolvent
Estrada index among trees of order n.
Another suggestion ck = 1/nk+1 appears in [12, 13], where it defines the closely related
resolvent energy
ER(G) =

1 X Mk (G)
,
n k0 nk

(4)

which can also be represented in terms of eigenvalues as

ER(G) =

n
X
i=1

1
.
n i (G)

Gutman et al. [12, 13] establish a number of bounds on the resolvent energy and characterize trees, unicyclic and bicyclic graphs with smallest and largest resolvent energies.
While researching Estrada index with a former PhD student [15], one of the present
authors made a conjecture analogous to Conjecture 1 about trees with smallest Estrada
indices, although it had not been published. Motivated by similarity in conjectured structure of trees with smallest Estrada and resolvent Estrada indices, and the fact that both
of these indices are defined in terms of the numbers of closed walks, we now turn our
attention to the following partial order of graphs.
Definition 2. For two graphs G and H we write G  H if Mk (G) Mk (H) for each
k 0. Further, G H if G  H and there exists k 0 0 such that Mk0 (G) < Mk0 (H).

Thus
G  H = EE(G) EE(H), EEr (G) EEr (H) and ER(G) ER(H),

(5)

G H = EE(G) < EE(H), EEr (G) < EEr (H) and ER(G) < ER(H).

(6)

while

The partial order  enables the comparison of spectral radii of graphs as well. The
Perron-Frobenius theorem [16] states that 1 (G) = maxi |i (G)|, so that
1 (G) = lim

2k

M2k (G).

(We include only closed walks of even length above, as bipartite graphs do not have closed
walks of odd length.) Thus
G  H = 1 (G) 1 (H).

(7)

However, due to the appearance of the limit above, G H does not necessarily imply
1 (G) < 1 (H).
Certainly, not all trees are comparable by : one of the smallest pairs of incomparable
trees is depicted in Fig. 1. Nevertheless, we will show that trees Pn1 (j) precede almost
all other trees in -order.

Figure 1: A pair of trees incomparable by .

Let us now define a few further trees. First, let Fn be the tree obtained from the
path Pn2 by attaching a pendent path of length two at position 3 in Pn2 (see Fig. 2 for
an example). Next, let Qn denote the set of trees of order n that may be obtained from the

Figure 2: The tree Fn for n = 12.

path Pk for some k n by attaching at most three new pendent edges to the leaves of Pk

and at most one new pendant edge to its internal vertices. Note that this implies that
 

Qn also contains Pn , Pn1 (2), Pn1 (3), . . . , Pn1 n2 . In order to distinguish them from
 

other trees in Qn , we further denote Qn = Qn \ {Pn , Pn1 (2), Pn1 (3), . . . , Pn1 n2 }.
Fig. 3 illustrates a few trees from Q12 in which the attached pendent edges are shown in
red (dashed line).

Figure 3: A few trees from Q12 .

The rest of the paper may be summarized as follows. In the next section we show that
j n k
Pn Pn1 (2) Pn1 (3) Pn1
Fn
2
and further that
Fn  T,
whenever T is a tree of order n such that T
/ Qn . Due to (6), this shows that
 n 

Pn1 (2), Pn1 (3), . . . , Pn1 2
have smaller resolvent Estrada indices (and ordinary
Estrada indices and resolvent energies) than any tree not in Qn . Trees in Qn need not
 

be -comparable to Pn1 (2), Pn1 (3), . . . , Pn1 n2 see the example of incomparable
trees depicted in Fig. 1. In Section 3 we bound the contribution of longer closed walks
to the resolvent Estrada index and resolvent energy in terms of the spectral radius. The
small spectral radius of trees in Qn will then enable us to focus on short closed walks to
show that trees in Qn have larger resolvent Estrada indices and resolvent energies than
 

Pn1 n2 , thus completing the proof of Conjecture 1 and proving an analogous result
for the resolvent energy.

Trees not in Qn

The following lemma appears as Theorem 3.2 in [15].

Lemma 3. Let u be a vertex of a nontrivial connected graph G, and for nonnegative
integers p and q let G(u; p, q) denote the graph obtained from G by attaching two pendent
paths of lengths p and q, respectively, at u. If 1 p q, then
G(u; p 1, q + 1) G(u; p, q).

Repeated application of Lemma 3 further shows that

G(u; 0, p + q) G(u; p, q),

(8)

where G(u; 0, p+q) is the result of replacing two pendent paths of lengths p and q attached
at u with a single path of length p + q.
Corollary 4. For each n 4,
Pn Pn1 (2) Pn1 (3) Pn1

j n k
2

(9)

Proof. Let u be a vertex of the path P2 on two vertices. Then

Pn

P2 (u; 0, n 2),

Pn1 (2)

P2 (u; 1, n 3),

Pn1 (3)

P2 (u; 2, n 4),

...
Pn1

j n k
2

 jnk
lnm

P2 u;
1,
1 .
2
2

The chain of inequalities (9) now follows from Lemma 3.

We will now extend the chain (9) with an additional term. Note that we exclude the
case n = 6 below, as P5 (3) = F6 .
Lemma 5. For each n 7,
Pn1

j n k
2

Fn .

Proof. We calculate characteristic polynomials of Pn1

(10)
 n 

2

and a subgraph of Fn in

order to derive an important relation between their spectral moments. For this task we
use the result of Schwenk [14], who has shown that
(G, ) = (G uv, ) (G u v, )

(11)

whenever uv is a cut edge of G.

Case 1. Suppose that n is even, n = 2b, b 4, so that Pn1

 n 

2

= P2 (u; b 1, b 1).

We also have Fn
= P3 (v; 2, 2b 5), where v is the vertex of degree three in Fn . Let
P3 (v; 2, b 2) denote a subgraph of Fn , obtained by deleting the farthest b 3 edges from
the branch of Fn of length 2b 5. (Examples of P2 (u; b 1, b 1) and P3 (v; 2, b 2) are
depicted in Fig. 4.)

Figure 4: Trees P2 (u; b 1, b 1) and P3 (v; 2, b 2).

Let u0 be a neighbor of u on one of the branches of length b 1 in P2 (u; b 1, b 1).
Then by (11)
(P2 (u; b 1, b 1), ) = (P2 (u; b1, b1)uu0 , )(P2 (u; b1, b1)uu0 , )
= (Pb+1 , )(Pb1 , ) (P1 , )(Pb1 , )(Pb2 , )
= (Pb1 , ) [(Pb+1 , ) (Pb2 , )] .
On the other hand, denoting by v 0 a neighbor of v on one of the branches of length two
in P3 (v; 2, b), we get
(P3 (v; 2, b 2), ) = (P3 (v; 2, b2)vv 0 , )(P3 (v; 2, b2)vv 0 , )
= (P2 , )(Pb+1 , ) (P1 , )(P2 , )(Pb2 , )
= (P2 , ) [(Pb+1 , ) (Pb2 , )] .
Hence the characteristic polynomials (P2 (u; b 1, b 1), ) and (P3 (v; 2, b 2), )
have a common factor (Pb+1 , ) (Pb2 , ). If Sp() denotes the set of roots of the
polynomial , this implies
Sp((P2 (u; b 1, b 1), )) = Sp((Pb1 , )) Sp((Pb+1 , ) (Pb2 , )),
Sp((P3 (v; 2, b 2), )) = Sp((P2 , )) Sp((Pb+1 , ) (Pb2 , )),
and consequently, for each k 0,
Mk (P2 (u; b 1, b 1)) = Mk (Pb1 ) +

k ,

Sp((Pb+1 ,)(Pb2 ,))

Mk (P3 (v; 2, b 2)) = Mk (P2 ) +

k .

Sp((Pb+1 ,)(Pb2 ,))

Thus
Mk (P2 (u; b 1, b 1)) Mk (P3 (v; 2, b 2)) = Mk (Pb1 ) Mk (P2 ).

(12)

Suppose now that one pendent edge of Pb1 is colored green, while the remaining
b 3 edges are colored red. The expression Mk (Pb1 ) Mk (P2 ) then represents the
number of closed walks of length k in Pb1 that contain at least one red edge.

Suppose further that the edges of Fn = P3 (v; 2, 2b5) are also colored (see Fig. 5): the
last b 3 edges of the branch of length 2b 5 are colored red (dashed line), the preceding
edge of that branch is colored green (dotted line), while the remaining edges are colored
black. The subgraph induced by the black and green edges is then P3 (v; 2, b 2), while
the subgraph determined by the green and red edges is Pb1 . The closed walks of length k

Figure 5: The black and green edges of Fn induce P3 (v; 2, b 2), while the green and red
edges induce Pb1 .
in Fn may now be partitioned into those closed walks that belong to the black-green
subgraph P3 (v; 2, b 2) and those that contain at least one red edge. The closed walks
that contain at least one red edge may be further partitioned into Mk (Pb1 ) Mk (P2 )
walks that contain only green and red edges and those that contain both black and red
edges. Thus from (12) we see that for each k 0,
Mk (Fn ) Mk (P3 (v; 2, b 2) + (Mk (Pb1 ) Mk (P2 )) = Mk (P2 (u; b 1, b 1)).

(13)

Moreover, strict inequality holds for k 6 as Fn then contains closed walks with both
black and red edges. Hence
Pn1

j n k
2

= P2 (u; b 1, b 1) Fn

holds for even n.

Case 2. Now let n be odd, n = 2b 1, b 4, so that Pn1

 n 

2

= P2 (u; b 1, b 2),

while Fn
= P3 (v; 2, 2b 6). Similarly as in the previous case, we can see that
Mk (P2 (u; b 1, b 1)) Mk (P2 (u; b 1, b 2)) + (Mk (Pb1 ) Mk (Pb2 ))

(14)

by coloring a pendent edge of one of the branches of length b 1 in P2 (u; b 1, b 1) in

red (dashed line), the preceding b 3 edges of the same branch in green (dotted line), and
the remaining edges in black (see Fig. 6). The closed walks of length k of P2 (u; b 1, b 1)
may then be partitioned into the closed walks contained within the black-green subgraph

Figure 6: The black and green edges induce P2 (u; b 1, b 2), while the green and red
edges induce Pb2 .
P2 (u; b 1, b 2) and the closed walks than contain at least one red edge. The latter may
further be partitioned into the Mk (Pb1 ) Mk (Pb2 ) closed walks that contain green and
red edges only and the remaining closed walks that contain both red and black edges.
The inequality in (14) is strict for k 2(b 1) when closed walks with both red and black
edges start to appear in P2 (u; b 1, b 1).
Now from (12) and (14) we get
Mk (P2 (u; b1, b2)) Mk (P2 (u; b1, b1))Mk (Pb1 )+Mk (Pb2 )
= [Mk (P3 (v; 2, b2))+Mk (Pb1 )Mk (P2 )]Mk (Pb1 )+Mk (Pb2 )
= Mk (P3 (v; 2, b 2))+Mk (Pb2 )Mk (P2 ).
By adapting the proof of (13) from the case of Fn
= P3 (v; 2, 2b 5) for even n to the case
of Fn
= P3 (v; 2, 2b 6) for odd n, we see that
Mk (P3 (v; 2, b 2)) + Mk (Pb2 ) Mk (P2 ) Mk (P3 (v; 2, 2b 6)),
which finally yields


Mk Pn1

j n k
2

= Mk (P2 (u; b 1, b 2)) Mk (P3 (v; 2, 2b 6)) = Mk (Fn )

with strict inequality for k 6. Hence

Pn1

j n k
2

Fn

holds for odd n as well.

Remark. As a consequence of walk counting in Lemma 5 we easily get that

j n k
1 Pn1
1 (Fn ).
2
Note that it would be very hard to obtain this inequality in a different way, due to the
diminishing difference between these two spectral radii and the fact that both are bounded

from above by

2 + 5 2.058171027271492 (see [17, 18]). For example, already for

n = 75 we have an agreement in the first eight of their digits after the decimal point:

j n k
1 Pn1
2.05817102402
2
1 (Fn ) 2.05817102727.
Corollary 4 and Lemma 5 yield the chain of inequalities
Pn Pn1 (2) Pn1 (3) Pn1
The fact that Pn , Pn1 (2), Pn1 (3), . . . , Pn1

 n 

2

j n k
2

Fn .

(15)

have smaller resolvent Estrada indices

(and ordinary Estrada indices and resolvent energies) than trees not in Qn now follows
from (5), (15) and the following theorem.
Theorem 6. If T is a tree of order n 6 such that T 6 Qn , then Fn  T .
Proof. If T has no vertices of degree larger than two, then T is the path Pn which belongs
to Qn , yielding a contradiction. Hence T has at least one vertex of degree at least three.
Let v be a vertex of degree d 3, and let T1 , . . . , Td be the connected components
of T v. Let ni be the order of the subtree Ti , i = 1, . . . , d. Within each subtree Ti
we may repeatedly replace any two paths of lengths p and q attached at a vertex of
degree at least three at the largest distance from v with a single path of length p + q,
until the subtree Ti itself becomes the path Pni . Due to (8), each such path replacing
transformation produces a smaller tree in -order. After all the subtrees Ti become paths,
this sequence of transformations produces a tree T 0 such that T 0 v = Pn1 Pnd and
T0  T.
If d > 3, we may continue applying this transformation in T 0 by replacing any pair
of paths Pni and Pnj attached at v with a single path Pni +nj , until the degree of v
becomes three. This sequence of transformations then produces a tree T 00 such that
T 00 v = Pm1 Pm2 Pm3 for some m1 m2 m3 and T 00  T 0 . (If d = 3, then set
T 00 = T 0 .) Here the numbers m1 , m2 and m3 represent the sums of the three subsets that
partition the set {n1 , . . . , nd }, depending on which paths have been replaced together. In
addition, note that T 00 may be equivalently denoted as Pm1 +1 (v; m2 , m3 ), Pm2 +1 (v; m1 , m3 )
or Pm3 +1 (v; m1 , m2 ).

If 2 m1 then from the repeated application of Lemma 3 we have

Pm1 +1 (v; 2, m2 + m3 2)  Pm1 +1 (v; m2 , m3 ) = T 00 ,
P3 (v; m1 , m2 + m3 2) = Pm1 +1 (v; 2, m2 + m3 2),
Fn
= P3 (v; 2, m1 + m2 + m3 2)  P3 (v; m1 , m2 + m3 2).
Hence Fn  T 00  T 0  T .
Let us now consider when we can direct path replacing transformations so as to arrive
at 2 m1 . Suppose, without loss of generality, that 1 n1 n2 nd . First, if
d 6 then we can set m1 = n1 + n2 2, m2 = n3 + n4 2 and m3 = n5 + + nd 2.
Next, if d = 5 and n5 2 then we can similarly set m1 = n1 + n2 2, m2 = n3 + n4 2
and m3 = n5 2. The remaining case n1 = = n5 = 1 corresponds to T 00 being a star,
which has the largest number of closed walks of any even length among all trees of the
given order [19], so that necessarily Fn  T 00 . Further, if d = 4 and 2 n3 then we can
set m1 = n1 + n2 2, m2 = n3 2 and m3 = n4 2. Finally, if d = 3 then m1 = n1 ,
m2 = n2 , m3 = n3 , so that 2 m1 is equivalent to 2 n1 .
From the previous paragraph we see that we have Fn  T 00 unless it happens that
for each vertex of degree at least three in T holds either d = 3 and n1 = 1 or d = 4
and n1 = n2 = n3 = 1. Since ni = 1 corresponds to a pendent edge, we see that the
tree obtained after removing all pendent edges from such T necessarily results in some
path Pk : vertices of degree three in T have their degree reduced to two, while vertices of
degree four have their degree reduced to one (and consequently there may be at most two
vertices of degree four in T ). This, however, implies that T Qn , which is contradictory
to the assumption of this theorem.

Trees in Qn

As we can see from the example depicted in Fig. 1, trees in Qn need not be -comparable
 

to Pn1 n2 , so that Conjecture 1 has to be tackled in a different way for them. Thanks
to the fact that trees in Qn have small spectral radius, the following simple bound turns
out to be sufficient for its resolution.
Lemma 7. For graphs G and H of order n, let max{1 (G), 1 (H)} . Then



k0
 X M (G) M (H) 
n(n 1)



k
k
.



 n1 n1
(n 1)k
kk0

(16)

Proof. Since |i (G)| 1 (G) for each i = 1, . . . , n, we have that



n
n
X
 X

k
|Mk (G)| = 
i (G) 
|i (G)|k nk


i=1

i=1

and, similarly, |Mk (H)| nk . Since both Mk (G) and Mk (H) are nonnegative, we get
nk Mk (H) Mk (G) Mk (H) Mk (G) nk .
Hence


 X M (G) M (H) 
X |Mk (G) Mk (H)|


k
k





(n 1)k
(n 1)k
kk0
kk0
X  k
n
n1
kk0

k0
n(n 1)

=
.
n1 n1
A tree is a bipartite graph which does not contain closed walks of odd length, so that
its resolvent Estrada index is equal to
EEr (T ) = n +

X Mk (T )
M4 (T )
M6 (T )
2(n 1)
+
+
+
,
(n 1)2 (n 1)4 (n 1)6 k8 (n 1)k

thanks to M0 (T ) = n and M2 (T ) = 2(n 1). For two trees T and S of order n we

therefore have
EEr (T ) EEr (S) =

M4 (T ) M4 (S) M6 (T ) M6 (S) X Mk (T ) Mk (S)

+
+
.
k
(n 1)4
(n 1)6
(n

1)
k8

If is a common upper bound for the spectral radii of T and S, we get from Lemma 7
that
"
EEr (T ) EEr (S)

n(n 1)
M4 (T ) M4 (S) M6 (T ) M6 (S)
+

(n 1)4
(n 1)6
n1

M4 (T ) M4 (S) M6 (T ) M6 (S)
n(n 1)
+
+
(n 1)4
(n 1)6
n1

n1

8

n1

8 #

This yields the following useful lemma.

Lemma 8. Let T and S be two trees of order n such that max{1 (T ), 1 (S)}. If
(n 1)2 [M4 (T ) M4 (S)] + [M6 (T ) M6 (S)] >
then EEr (T ) > EEr (S).

n8
,
(n 1)(n 1 )

Similarly to Lemmas 7 and 8, we can prove the following two lemmas on the resolvent
energy.
Lemma 9. For graphs G and H of order n, let max{1 (G), 1 (H)} . Then


 k0
 X M (G) M (H) 

n


k
k
.


k+1

 n n
n

(17)

kk0

Lemma 10. Let T and S be two trees of order n such that max{1 (T ), 1 (S)}. If
n2 [M4 (T ) M4 (S)] + [M6 (T ) M6 (S)] >

8
,
n

then ER(T ) > ER(S).

Next we have to provide an upper bound on the spectral radius of trees in Qn . We
first observe that, as a tree obtained by attaching at most three new pendent edges to
the leaves of a path and at most one new pendant edge to its internal vertices, every tree
in Qn is an induced subgraph of the tree Q depicted in Fig. 7. To bound the spectral

Figure 7: A tree containing all trees in Qn as its subtrees.

radius of Q, we will resort to the following result from [20], restated in more familiar
terms.
Theorem 11 ([20]). If G is a connected graph with adjacency matrix A, then the system
of inequalities
Ax x

(18)

has a solution for real and non-negative x if and only if 1 (G) .

Let A be the adjacency matrix of Q and recall that for each vertex u of Q
(Ax)u =

xv ,

vu

where the sum goes over all neighbors of u in Q. It is straightforward to check that the

system (18) is satisfied for = 1 + 2 and the vector x, whose components are depicted

in Fig. 7. As a matter of fact, equality holds in (18) for each vertex of Q, except for the
two vertices of degree four for which we have strict inequality. Hence
1 (Q) ,
and since the spectral radius is edge-monotone, also 1 (T ) for each tree T Qn .
Now that we have a common upper bound on the spectral radius of trees in Qn , we
can move on to estimate their numbers of closed walks of lengths 4 and 6. From [21] we
have that
X

M4 (G) = 2

d2u 2m + 8q,

(19)

uV (G)

where du is the degree of the vertex u, m is the number of edges and q is the number
of quadrangles in G. Certainly, m = n 1 and q = 0 in a tree T , so that (19) for trees
reduces to

M4 (T ) = 2

d2u n + 1 .

uV (G)

Closed walks of length six may, in general, go over the cycles of length 4 and 6 in a
graph. However, as we are interested in trees only, each closed walk u = u0 , u1 , . . . , u6 = u
may have only one of the forms depicted in Fig. 8. They are classified according to whether
the vertices u2 and u4 are equal to u and to each other (vertices that are not specified
beneath each drawing may be mutually equal even if they are depicted as different in the
drawing).

Figure 8: Five types of closed walks of length six in trees.

Clearly, there are

uV (T )

d3u walks of the first type u2 = u = u4 . Walks of the second

type u2 = u 6= u4 are most easily counted with the help of adjacent vertices u and u3 :

there are du choices for u1 and since u4 6= u, there are du3 choices for the vertex u4 , so
that the total number of walks of this type is
X

du (du3 1) = 2

uV (T ) {u3 : uu3 E(T )}

du du3

uu3 E(T )

d2u .

uV (T )

as each edge of T gets counted twice in the first double sumonce as (u, u3 ) and once
as (u3 , u), while each degree du gets subtracted du times. In the same way, counting over
the pairs of adjacent vertices u and u1 (instead of u3 ), we see that there are also
X

du du1

uu1 E(T )

d2u

uV (T )

walks of the third type u2 6= u = u4 . Closed walks of the fourth type may be counted by
the adjacent pair (u, u1 ): for each choice of the neighbor u1 of u, there are du1 1 choices
for each of u2 6= u and u4 6= u, so that the total number of walks of this type is
X

(du1 1)2 =

uV (T ) {u1 : uu1 E(T )}

du1 (du1 1)2 =

u1 V (T )

d3u1 2

u1 V (T )

d2u1 + 2m.

u1 V (T )

Closed walks of the fifth type may be counted by the adjacent pair (u1 , u2 ): for each
choice of the neighbor u1 of u, there are du1 1 choices for u2 and du2 1 choices for u3 .
Hence the total number of walks of this type is
X

uV (T ) {u1 : uu1 E(T )} {u2 : u1 u2 E(T )}

= 2

du1 du2

u1 u2 E(T )

d2u1

u1 V (T )

(du2 1) = 2

(du1 1)(du2 1)

u1 u2 E(T )

d2u3

+ 2m.

u3 V (T )

Finally, adding these quantities together we see that

X

M6 (T ) = 2

d3u 6

uV (T )

d2u + 6

uV (T )

du dv + 4m.

(20)

uvE(T )

Now we can move forward to compare the closed walks of lengths 4 and 6 between
 

 

Pn1 n2 and the trees in Qn . Tree Pn1 n2 has one vertex of degree 3, three leaves
and n 4 vertices of degree 2, so that


M4 Pn1

j n k
2

= 6n 6 and M6 Pn1

j n k
2

= 20n 14.

If T is a tree in Qn , then it has one of the three types depicted in Fig. 3, depending on
the number of vertices of degree 4 that it contains:

Type i) For a tree T Qn with no vertices of degree 4, if it has k 2 vertices of degree 3,

then it has k + 2 leaves and n 2k 2 vertices of degree 2, so that
M4 (T ) = 6n + 4k 10 6n 2.

(21)

It further contains k edges between vertices of degrees 3 and 1, two edges that
connect path leaves to vertices of degree either 2 or 3, and n k 3 edges that
connect vertices of degrees either 2 or 3 on the path. The edge on the right from each
P
vertex of degree 3 on the path will contribute at least 3 2 to the sum uvE(T ) du dv .
Regardless of how the vertices of degree 3 are distributed, there will be at least k 1
such edges, while the remaining n 2k 2 will contribute at least 2 2. Hence
P
uvE(T ) du dv 4n + k 10, so that
M6 (T ) 20n + 18k 56 20n 20.

(22)

Hence
(n 1)2 [M4 (T ) M4 (S)] + [M6 (T ) M6 (S)] 4(n 1)2 6.
The inequality
4(n 1)2 6 >
for = 1 +

n8
(n 1)(n 1 )

2 is satisfied for n 9, so that from Lemma 8 we have


j n k
EEr (T ) > EEr Pn1
2

for each tree T of this type with at least 9 vertices.

Type ii) For a tree T Qn with one vertex of degree four, if it has k 0 vertices of degree
three, then it has k + 4 leaves and n 2k 5 vertices of degree two, so that
M4 (T ) = 6n + 4k + 2 6n + 2;
Similarly as in the previous case, we get

uvE(T )

(23)

du dv 4n + k 4, so that

M6 (T ) 20n + 18k 32 20n 32.

Hence
(n 1)2 [M4 (T ) M4 (S)] + [M6 (T ) M6 (S)] 8(n 1)2 18.

(24)

The inequality
n8
8(n 1) 18 >
(n 1)(n 1 )
2

is satisfied for n 8, so that from Lemma 8


j n k
EEr (T ) > EEr Pn1
2
for each tree T of this type with at least 8 vertices.
Type iii) For a tree T Qn with two vertices of degree four, if it has k 0 vertices of degree
three, then it has k + 6 leaves and n 2k 8 vertices of degree two, so that
M4 (T ) = 6n + 4k + 14 6n + 14.
In this case we get

uvE(T )

(25)

du dv 4n + k + 2, so that

M6 (T ) 20n + 18k + 112 20n + 112.

(26)

Hence
(n 1)2 [M4 (T ) M4 (S)] + [M6 (T ) M6 (S)] 20(n 1)2 + 126.
The inequality
20(n 1)2 + 126 >

n8
(n 1)(n 1 )

is satisfied for n 6, so that from Lemma 8

j n k

EEr (T ) > EEr Pn1
2
for all trees T of this type.
Hence it remains to computationally check trees of the first type with at most 8 vertices
and trees of the second type with at most 7 vertices. There are 11 such trees in total,
 

depicted in Fig. 9, all of which have resolvent Estrada index larger than that of Pn1 n2 .
This resolves Conjecture 1 for trees in Qn as well, so that we finally have
Theorem 12. For 2 a bn/2c, the tree Pn1 (a) has the a-th smallest resolvent Estrada
index among trees of order n.
Lemma 10, together with the estimates (21)-(26) for the numbers of closed walks of
lengths 4 and 6, implies that


ER(T ) > ER Pn1

j n k
2

Figure 9: Trees in Qn to which Lemma 8 cannot be applied. Their resolvent

 n Estrada


indices and resolvent energies are shown beneath the drawings. Trees Pn1 2 , 6
n 8, are shown for comparison.
holds for trees in Qn of the first type with at least 8 vertices, of the second type with at
least 7 vertices, and for all trees of the third type (see Fig. 3). The remaining four trees in
 

Q6 and Q7 also have resolvent energy larger than that of the corresponding Pn1 n2 ,
as shown in Fig. 9. Hence we also have
Theorem 13. For 2 a bn/2c, the tree Pn1 (a) has the a-th smallest resolvent energy
among trees of order n.

Concluding remarks

We have resolved Conjecture 1 in Sections 2 and 3 by applying different approaches to

trees not in Qn and trees in Qn . The fact that trees not in Qn are -comparable to Fn
 

enabled us to conclude that Pn , Pn1 (2), . . . , Pn1 n2
are the smallest trees for any
invariant defined as a weighted series of closed walk numbers with nonnegative coefficients,
including the Estrada index, the resolvent Estrada index, the resolvent energy and the
spectral radius (allowing equality instead of strict inequality in the latter case).

On the other hand, trees in Qn are, in principle, not -comparable to Pn1

 n 

2

. The

appropriate choices of coefficients ck = 1/(n1)k in the definition of the resolvent Estrada

index and ck = 1/nk+1 in the definition of the resolvent energy enabled us to use a simple
spectral bound on the tail of defining series in Lemmas 7 and 9, and to focus solely on
closed walks of lengths 4 and 6 in Lemmas 8 and 10. Such approach, unfortunately, cannot
be used with the Estrada index or the spectral radius, so that new methods have to be
found in order to deduce results for the Estrada index and the spectral radius analogous
to those presented here for the resolvent Estrada index and the resolvent energy.

References
[1] E. Estrada, D.J. Highman, Network properties revealed through matrix functions,
SIAM Rev. 52 (2010), 696714.
[2] E. Estrada, Characterization of 3D molecular structure, Chem. Phys. Lett. 319 (2000),
713718.
[3] E. Estrada, Characterization of the folding degree of proteins, Bioinformatics 18
(2002), 697704.
[4] E. Estrada, Characterization of the amino-acids contributions to the folding degree of
proteins, Proteins: Structure, Function and Bioinformatics 54 (2004), 727737.
[5] E. Estrada, Topological structural classes of complex networks, Phys. Rev. E 75
(2007), 016103.
[6] E. Estrada, J.A. Rodrguez-Valazquez, M. Randic, Atomic branching in molecules Int.
J. Quantum Chem. 106 (2006), 823832.
[7] I. Gutman, B. Furtula, V. Markovic, B. Glisic, Alkanes with greatest Estrada index,
Z. Naturforsch. 62a (2007), 495498.
[8] M. Benzi, P. Boito, Quadrature rule-based bounds for functions of adjacency matrices,
Linear Algebra Appl. 433 (2010), 637652.
[9] X. Chen, J. Qian, Bounding the resolvent Estrada index of a graph, J. Math. Study
45 (2012), 159166.

[10] X. Chen, J. Qian, On resolvent Estrada index, MATCH Commun. Math. Comput.
Chem. 73 (2015), 163174.
[11] I. Gutman, B. Furtula, X. Chen, J. Qian, Graphs with Smallest Resolvent Estrada
Indices, MATCH Commun. Math. Comput. Chem. 73 (2015), 267270.
[12] I. Gutman, B. Furtula, E. Zogic, E. Glogic, Resolvent energy of graphs, MATCH
Commun. Math. Comput. Chem. 75 (2016), 279290.
[13] I. Gutman, B. Furtula, E. Zogic, E. Glogic, Resolvent energy, in: I. Gutman, X. Li
(eds.), Energies of graphstheory and applications, University of Kragujevac, Kragujevac, 2016, pp. 277290.
[14] A.J. Schwenk, Computing the characteristic polynomial of a graph, in: Graphs and
Combinatorics, Proc. Capital Conf., Washington, D.C. 1973, Lecture Notes in Mathematics 406, Springer-Verlag, Berlin, 1974, pp. 153172.
[15] A. Ilic, D. Stevanovic, The Estrada index of chemical trees, J. Math. Chem. 47
(2010), 305314.
[16] D. Stevanovic, Spectral radius of graphs, Academic Press, Amsterdam, 2015.
[17] D.M. Cvetkovic, M. Doob, I. Gutman, On graphs whose spectral radius does not
p

exceed 2 + 5, Ars Combin. 14 (1982), 225239.

[18] A.E. Brouwer, A. Neumaier, The graphs with spectral radius between 2
p

and 2 + 5, Linear Algebra Appl. 114/115 (1989), 273276.

[19] H. Deng, A proof of a conjecture on the Estrada index, MATCH Commun. Math.
Comput. Chem. 62 (2009), 599606.
[20] W.E. Pruitt, Eigenvalues of non-negative matrices, Ann. Math. Stat. 35 (1964), 1797
1800.
[21] I. Gutman, Some topological properties of benzenoid systems, Croat. Chem. Acta 46
(1974), 209215.