Journal of Computational and Applied Mathematics 204 (2007) 166 – 171

www.elsevier.com/locate/cam

Asymptotic estimates for efficiency, vulnerability and cost for

random networks
R. Criado∗,1,2 , B. Hernández-Bermejo1 , J. Marco-Blanco, M. Romance1
Departamento de Matemáticas y Física Aplicadas y CC. de la Naturaleza, ESCET, Universidad Rey Juan Carlos, C/Tulipán s.n., 28933 Móstoles,
Madrid, Spain

Received 15 July 2005; received in revised form 10 March 2006

Abstract
An investigation both analytical and numerical of the values of efficiency, vulnerability and cost functions is provided. Several
comparisons among these functions are made following a pairwise strategy, and taking into account different efficiency, vulnerability
and cost functions. It is shown in all cases that these magnitudes display strong correlations, which are typically piecewise linear in
log-space as well as universal (i.e., independent of the number of nodes n) to a great extent. The analytical part of the work develops
explicit bounds linking all pairs of such parameters. These results are numerically compared with the result of a random testing for
Erdös–Rényi networks.
© 2006 Elsevier B.V. All rights reserved.

MSC: 02.10

Keywords: Network efficiency; Network vulnerability; Random networks; Network topology; Network design

1. Introduction and notation

A huge variety of real systems of interest in science and technology may be described in terms of complex network
properties (for instance, see [1,11] for detailed reviews). In fact, the study of networks has become one of the paradigms
of the science of complexity as well as a fascinating branch of research in applied mathematics, science and engineering.
The investigation of such issues must necessarily embrace a diversity of viewpoints that include different complementary
aspects of the network structure. In this sense, a realistic approach must proceed by defining a suitable set of network
properties (usually reduced to associated parameters which are to be measured over the network) which should comprise
a sufficiently rich and complete picture of the problem under investigation. On the other hand, this should be done while
defining a sufficiently tractable framework: the parameters to be evaluated must be simple and interpretable enough,
as well as computable in practice for real systems. In this sense, very different possible approaches of this kind can
be traced back in the literature [2,3,5,4,6–10]. In this work we shall focus on two magnitudes that have been recently
defined (efficiency E + (·) [5,9,10] and vulnerability VD (·) [4]) as well as a new magnitude that measures the cost of

∗ Corresponding author. Tel.: +34 91 6647445; fax: +34 91 4887338.

E-mail address: [email protected] (R. Criado).
1 Partially supported by Project PPR-2004-16 from Universidad Rey Juan Carlos.
2 Partially supported by Project TIC-2003-9001-C2 from CICyT (Spain).

0377-0427/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2006.04.035
 R. Criado et al. / Journal of Computational and Applied Mathematics 204 (2007) 166 – 171 167

the network, since it is defined as C, (G) = 1 − exp(−n − |E|), where G is a network with n vertices, |E| edges
and ,  are problem dependent parameters expressing the relative weight of the nodes and the edges in the evaluation
of the cost.
The purpose of this work is to provide some insights regarding the correlations existing between efficiency, vul-
nerability and cost. As we shall see, they exhibit unexpected correlations that shall be examined following a pairwise
strategy.

2. Estimates for the relations among efficiency, vulnerability and cost

In this section several upper and lower analytic bounds regarding the correlation existing among efficiency, vulner-
ability and cost functions are provided. All the results presented in this section are going to be developed in the realm
of connected networks.
Our first result gives some estimates for the relation between vulnerability VD (·) and efficiency E + (·).

Theorem 2.1. If G = (V , E) is a connected network with n vertices, then

 
1 n−1 +
p(n) + q(n) − E (G)  log(VD (G)),
n−2 n−2
n−2
log(VD (G))p(n) + q(n)(1 − 2E + (G)) + ,
n
where p(n) = n − 2 + 2/n and q(n) = n(n − 1)/2.

Proof. On the one hand, for every connected network G = (V , E) with n vertices
n − 2 2|E| 1
+ E + (G)
n − 1 n(n − 1) n − 1
(see [4]), which can be used to give an upper bound for |E| in terms of efficiency function and the number of nodes
ensuring that
 
n−1 + 1
|E|q(n) E (G) − (1)
n−2 n−2
and hence
 
1 n−1 +
log(VD (G))p(n) + q(n) − E (G) .
n−2 n−2
On the other hand, if we want to prove the second estimate, we show that
M −m 2 n−2
log(VD (G)) = + n − 2 + − |E|  + p(n) + q(n)(1 − 2E + (G)). 
n n n

Similar estimates can be stated for the relations among V (·) and the efficiency functions (E + or E • ), since
V (G) VD (G) (see [4]) and log(V (G))p(n) − |E|.

Remark 2.2. Theorem 2.1 can be used to give the asymptotic behavior of log VD (·) with respect to E + (·) and we
obtain that for every connected network G with n vertices

−q(n)E + (G) log VD (G) − q(n)(2E + (G) − 1),

where f (n)g(n) means that f (n)C · g(n), for some universal constant C > 0 independent of n.

A similar technique can be used to give estimates for the relation among the cost C, (·) and the efficiency, that is,
by using inequality (1) and since log(1 − C, (G)) = −n − |E|, we can prove the following result:
168 R. Criado et al. / Journal of Computational and Applied Mathematics 204 (2007) 166 – 171

Theorem 2.3. If G = (V , E) is a connected network with n vertices, , 0 and q(n) = n(n − 1)/2 (as before), then
 
1 n−1 +
−n + q(n) − E (G)  log(1 − C, (G)),
n−2 n−2

log(1 − C, (G)) − n + q(n)(1 − 2E + (G)).

Note that the asymptotic behavior of log(1−C, (·)) with respect to E + (·) is the same that the behavior of log VD (·),
since Theorem 2.3 implies that

−q(n)E + (G) log(1 − C, (G)) − q(n)(2E + (G) − 1).

Finally, the next result gives estimates for the relation among the cost C, (·) and the vulnerability.

Theorem 2.4. If G = (V , E) is a connected network with n vertices, , 0 and q(n) = n(n − 1)/2 (as before), then:

(i) If  1
2
( + )n − ( + 1) +  log(VD (G)) − log(1 − C, (G)),
n
log(VD (G)) − log(1 − C, (G))( − 1)q(n) + ( + 1)n − 1.

(ii) If 0 1
2
( − 1)q(n) + ( + 1)n − 2 +  log(VD (G)) − log(1 − C, (G)),
n
log(VD (G)) − log(1 − C, (G))( + )n − .

Proof. After some algebra, a direct computation shows that

M −m 2
log VD (G) − log(1 − C, (G)) = + ( − 1)|E| + ( + 1)n − 2 +
n n
and if G is connected then it is well known that 0 M − m q(n) and

( − 1)(n − 1)( − 1)|E|( − 1)q(n) if 1,

( − 1)q(n)( − 1)|E|( − 1)(n − 1) if 0 1.

Combining these inequalities the result holds. 

Remark 2.5. If we regard at the asymptotic behavior of log VD (·) with respect to log(1−C, (·)), Theorem 2.4, shows
that if we denote for every connected network G with n vertices we denote (G) = log(VD (G)) − log(1 − C, (G)),
then, if  1

( + )n − ( + 1)(G)( − 1)q(n) + ( + 1)n − 1,

while if 0  1

( − 1)q(n) + ( + 1)n − 2(G)( + )n − .

If we take  = 1, then p1 (n)(G)p2 (n), with p1 (n), p2 (n) linear functions, that is, if  = 1, then there is a linear
relation between VD (·) and C, (·) in the logarithmic space. This result can be directly derived from the fact that when
 = 1, log VD (·) and log(1 − C, (·)) are essentially the same, up to the range of the degree distribution and it is well
known that 0 (M − m)/n1.
 R. Criado et al. / Journal of Computational and Applied Mathematics 204 (2007) 166 – 171 169

log (E)
-1 -0.8 -0.6 -0.4 -0.2
0
-20
-40
-60
-80

log (V)
-100
-120
-140
-160
-180
-200

Fig. 1. Joint plot for log E + vs. log VD for random networks with n = 5, 10, 15, 20 and 25 vertices.

log (1-C)
-80 -60 -40 -20
0

-10

-20

log (V)
-30

-40

-50

-60

Fig. 2. log(1 − C) vs. log VD for random networks with n = 5, 10, 15, 20 and 25 vertices.

3. A random testing of the estimates

At this point our purpose is to briefly provide a numerical illustration regarding the sharpness of the bounds obtained
in the previous section. The specific cost function that has been used for the computations is given by C(G) = 1 −
exp{−0.1 n − 0.2 |E|}. In all cases, for every pair of magnitudes different values of n are selected, and for each n a
sample of 1000 random graphs is generated. Additionally, the correlations have been analyzed in logarithmic space,
in which they become more manageable. The testing shows the existence of very significant correlations between all
pairs of magnitudes involved. Since the results for the arithmetic and geometric efficiencies are qualitatively similar,
as it is also for both vulnerabilities, only the results for one of them are given here for the sake of conciseness.
For instance, in Fig. 1 the joint plot for E + and VD is displayed for n equal to 5, 10, 15, 20 and 25. Notice the
sharp linear correlation in the limit of strongly connected graphs (where the asymptotic slope depends on n) in which
vulnerability is small and efficiency is close to 1. On the contrary, the dispersion is higher in the case of weakly
connected networks, but in such case the asymptotic slope is independent of n. In spite that cost and vulnerability
are not functions of each other, now the log-linear correlation is sharp and has a slope independent of n, as Fig. 2
shows. Curiously, this connection seems to be independent of the statistical measure employed in order to define the
170 R. Criado et al. / Journal of Computational and Applied Mathematics 204 (2007) 166 – 171

log (E)
-1 -0.8 -0.6 -0.4 -0.2
0

-20

-40

log (V)
-60

-80

Fig. 3. The plots in logarithmic scale for VD vs. E + for random networks with n = 15 vertices together with its estimates.

log (1-C)
-180 -160 -140 -120 -100 -80 -60 -40 -20
0
-20
-40
-60
log (V)

-80
-100
-120
-140
-160
-180

Fig. 4. log(1 − C) vs. log VD for the case n = 20,  = 0.95 and  = 0.1 with its upper and lower estimates.

vulnerability (similar results are obtained for V , actually). In this case it is also remarkable that this strong correlation
is edge-independent.
Complementary to the previous one, Fig. 3 displays the plots in logarithmic scale for VD vs. E + with n = 15 together
with its estimates. Note that the upper bound is sharp for efficient graphs and the lower bound shows a finite gap for
weakly connected graphs. In Fig. 4 we can see the plot for log(1 − C) vs. log VD for the case n = 20,  = 0.95 and
 = 0.1 with its upper and lower bounds. The greater is  the narrower these bounds are: for  = 1 they coincide and
fit exactly the data points.
A detailed analysis of these results is not given here for the sake of conciseness as far as several key issues remain
open, specially those related to the universality of the asymptotic behaviors, the comparison with additional kinds of
networks and the interplay between the kind of network considered and the sharpness of the bounds. A throughout
investigation of these questions will be the aim of future work.

References

[1] R. Albert, A.L. Barabási, Statistical mechanics of complex networks, Rev. Modern Phys. 74 (2002) 47–97.
[2] R. Cohen, K. Erez, D. ben-Avraham, S. Havril, Resilience of the internet to random breakdowns, Phys. Rev. Lett. 85 (21) (2000) 4626–4628.
[3] R. Cohen, K. Erez, D. ben-Avraham, S. Havril, Breakdown of the internet under intentional attacks, Phys. Rev. Lett. 86 (16) (2001) 3682–3685.
 R. Criado et al. / Journal of Computational and Applied Mathematics 204 (2007) 166 – 171 171

[4] R. Criado, J. Flores, B. Hernández-Bermejo, J. Pello, M. Romance, Effective measurement of network vulnerability under random and intentional
attacks, J. Math. Modelling Algorithms 4 (3) (2005) 307–316.
[5] R. Criado, A. García del Amo, B. Hernández-Bermejo, M. Romance, New results on computable efficiency and its stability for complex
networks, J. Comput. Appl. Math. 192 (1) (2006) 59–74.
[6] P. Crucitti, V. Latora, M. Marchiori, A. Rapisarda, Efficiency of scale-free networks: error and attack tolerance, Physica A 320 (2003) 622–642.
[7] A.H. Dekker, B.D. Colbert, Network Robustness and Graph Topology, Proceedings of ACSC04, the 27th Australasian Computer Science
Conference, 18–22 January 2004, Dunedin, New Zealand.
[8] P. Holme, Beom Jun Kim, Chang No Yoon and Seung Kee Han, Attack vulnerability of complex networks, Phys. Rev. E. 65 (2002), art. no.
056109.
[9] V. Latora, M. Marchiori, Efficient behaviour of small-world networks, Phys. Rev. Lett. 87 (2001), art. no. 198701.
[10] V. Latora, M. Marchiori, How the science of complex networks can help developing strategies against terrorism, Chaos Solitons Fractals 20
(2004) 69–75.
[11] M.E.J. Newman, The structure and function of complex networks, SIAM Rev. 45 (2003) 167–256.