Sie sind auf Seite 1von 4

PHYSICAL REVIEW E 66, 047104 共2002兲

Epidemic spreading in correlated complex networks

Marián Boguñá1 and Romualdo Pastor-Satorras2
Departament de Fı́sica Fonamental, Universitat de Barcelona, Avenida Diagonal 647, 08028 Barcelona, Spain
Departament de Fı́sica i Enginyeria Nuclear, Universitat Politècnica de Catalunya, Campus Nord B4, 08034 Barcelona, Spain
共Received 30 May 2002; published 21 October 2002兲
We study a dynamical model of epidemic spreading on complex networks in which there are explicit
correlations among the node’s connectivities. For the case of Markovian complex networks, showing only
correlations between pairs of nodes, we find an epidemic threshold inversely proportional to the largest
eigenvalue of the connectivity matrix that gives the average number of links, which from a node with connec-
tivity k go to nodes with connectivity k ⬘ . Numerical simulations on a correlated growing network model
provide support for our conclusions.

DOI: 10.1103/PhysRevE.66.047104 PACS number共s兲: 89.75.⫺k, 05.70.Ln, 87.23.Ge

Statistical physics has witnessed in recent years a renewed parameters 关7,12兴. Nevertheless, it represents a first approxi-
interest in graph theory due to the discovery that many natu- mation to real networks, which neglects the possibility of
ral and artificial systems can be described in terms of com- correlations in the connectivity of the nodes. The existence
plex networks, in which the nodes represent typical units, and importance of connectivity correlations has been re-
and the links represent the interactions between pairs of units cently pointed out in the literature. In fact, it has become
关1,2兴. The term complex network has been coined to refer to clear that these correlations are critical in the understanding
networks that typically exhibit two distinct properties: of the hierarchical structure of the Internet 关13–15兴. On the
共i兲 A scale-free 共SF兲 connectivity distribution. If we define other hand, some growing network models have been pro-
the connectivity distribution P(k) as the probability that a posed 关16兴, in which correlations are spontaneously gener-
node is connected to k other nodes, then SF networks are ated and have important effects in the percolation transition.
characterized by a power-law behavior P(k)⬃k ⫺ ␥ , where ␥ In this paper we shall present a study of epidemic spread-
is a characteristic exponent. This property renders classical ing in complex random networks in which there are explicit
models of random graphs 关3兴, described by an exponentially correlations among the node’s connectivities. We will con-
bounded connectivity, inappropriate for the description of sider in particular the subset of undirected Markovian ran-
many real networks. dom networks that are completely defined by their connec-
共ii兲 The small-world property 关4兴 that is defined by an tivity distribution P(k) and the conditional probability
average path length—average distance between any pair of P(k ⬘ 兩 k) that a node of connectivity k is connected to a node
nodes—increasing very slowly 共usually logarithmically兲 with of connectivity k ⬘ . These two functions can have any form
the network size N. 共including SF behavior兲, and are assumed to be normalized
Apart from the empirical characterization of real complex ( 兺 k P(k)⫽ 兺 k ⬘ P(k ⬘ 兩 k)⫽1) and restricted by the connectiv-
networks and the development of new models, accounting ity detailed balance condition
for the observed properties, the interest in this field has been
also driven by the discovery of the profound and subtle ef-
fects that the connectivity has on the behavior of dynamical k P 共 k ⬘ 兩 k 兲 P 共 k 兲 ⫽k ⬘ P 共 k 兩 k ⬘ 兲 P 共 k ⬘ 兲 ⬅ 具 k 典 P 共 k,k ⬘ 兲 , 共1兲
systems defined on top of complex networks. These effects
are particularly interesting in the study of disease transmis-
sion in SF networks, relevant for the understanding of the where the symmetric function (2⫺ ␦ kk ⬘ ) P(k,k ⬘ ) is the joint
spreading of computer viruses 关5兴 and sexually transmitted probability that two nodes of connectivity k and k ⬘ are con-
diseases 关6兴. Indeed, it was first noted in Ref. 关7兴 that in nected. The Markovian nature of this class of networks im-
uncorrelated SF networks with a connectivity exponent ␥ plies that all higher-order correlations can be expressed as a
⭐3, epidemic processes do not possess an epidemic thresh- function of P(k) and P(k ⬘ 兩 k), allowing an exact treatment
old below which diseases cannot produce a macroscopic epi- of epidemic models at the mean-field 共MF兲 level. It is worth
demic outbreak or the development of an endemic state. This noticing, however, that a more detailed description, in terms
feature, observed in several epidemic models 关7–11兴, is of a Langevin equation 共to be reported elsewhere 关17兴兲,
deeply rooted in the presence of very large connectivity fluc- yields exactly the same results, confirming the accuracy of
tuations in infinite SF networks. the MF description. In this framework, the topologically rel-
The study of epidemic spreading in uncorrelated complex evant magnitude is the connectivity matrix C kk ⬘
networks 共that is, in graphs in which the connectivity of any ⫽k P(k ⬘ 兩 k), that measures the average number of links that
node is independent of the connectivity of its neighbors兲 has go from a node with connectivity k to nodes with connectiv-
been proved to be extremely successful, providing, for in- ity k ⬘ . We will show that the epidemic threshold is related to
stance, the first satisfactory explanation of the long-standing the largest eigenvalue of this matrix. Extensive numerical
problem of the generalized low prevalence of computer vi- simulations on a random correlated network model 关16兴 con-
ruses without assuming any global tuning of microscopic firm the predictions of the present analysis. During the

1063-651X/2002/66共4兲/047104共4兲/$20.00 66 047104-1 ©2002 The American Physical Society


completion of this work we became aware of two recent a node of connectivity k ⬘ is independent of the connectivity
preprints 关18,19兴 in which the general role of correlations in k of the node from which the link is emanating. Therefore, in
spreading and percolation in complex networks is also high- this case ⌰ k ⫽⌰ nc is independent of k, and can be written as
In order to study the effects of connectivity correlations in 1
具k典 兺
epidemic spreading, we will focus in the standard ⌰ nc⫽ k ⬘ P 共 k ⬘ 兲 ␳ k ⬘共 t 兲 , 共4兲
susceptible-infected-susceptible 共SIS兲 model 关20兴. All the re-
sults, however, can be easily extended to the more general since the probability that a node is pointing to a node of
susceptible-infected-removed-susceptible model 关17兴. In the connectivity k ⬘ is proportional to k ⬘ P(k ⬘ ). Substituting the
SIS model each node in the network represents an individual, expression 共4兲 into Eq. 共3兲, one can solve for the steady-state
and each link represents a connection along which the infec- solution and find the existence of an epidemic threshold ␭ c ,
tion can propagate. Susceptible 共healthy兲 nodes become in- below which there are no solutions with a nonzero value of
fected with probability ␯ if they are connected to one or ⌰ nc. The expression of the epidemic threshold for uncorre-
more infected nodes. On the other hand, infected nodes re- lated random networks is
cover spontaneously with probability ␦ . The ratio of these
two rates defines an effective spreading rate ␭⫽ ␯ / ␦ 共without 具k典
lack of generality, we set ␦ ⫽1). For homogeneous net- c ⫽
␭ nc . 共5兲
works, in which each node has more or less the same number 具 k 2典
of connections, k⯝ 具 k 典 , a general result states the existence For infinite SF networks with ␥ ⭐3, we have 具 k 2 典 ⫽⬁, and
of a finite epidemic threshold, separating an infected 共en- correspondingly ␭ nc c ⫽0. Finally, from the solution of ␳ k ,
demic兲 phase, with a finite average density of infected indi-
one can compute the total prevalence ␳ using the relation ␳
viduals, from a healthy phase, in which the infection dies out
⫽ 兺 k P(k) ␳ k .
exponentially fast. In terms of the average density of infected
For a general network in which the connectivities of the
individuals ␳ (t) 共the prevalence兲 we can describe the SIS
nodes are correlated, the above formalism is not correct,
model in homogeneous networks at a MF level by the fol-
since we are not considering the effect of the connectivity k
lowing rate equation 关8兴
into the expression for ⌰ k . This effect can be taken into
account, however, for Markovian networks, whose correla-
⳵ t ␳ 共 t 兲 ⫽⫺ ␳ 共 t 兲 ⫹␭ 具 k 典 ␳ 共 t 兲关 1⫺ ␳ 共 t 兲兴 . 共2兲
tions are completely defined by the conditional probability
In this equation we have neglected higher-order terms, since P(k ⬘ 兩 k). In this case, it is easy to realize that the correct
we are interested in the onset of the endemic state, close to factor ⌰ k can be written as
the point ␳ (t)⬃0. Also, we have neglected correlations
among nodes. That is, the probability of infection of a new ⌰ k共 t 兲 ⫽ 兺 P 共 k ⬘兩 k 兲 ␳ k ⬘共 t 兲 , 共6兲
node—the second term in Eq. 共2兲—is proportional to the k ⬘
infection rate ␭, to the probability that a node is healthy, 1
that is, the probability that a link in a node of connectivity k
⫺ ␳ (t), and to the probability that a link in a healthy node
is pointing to an infected node is proportional to the prob-
points to an infected node. This last quantity, assuming the
ability that any link points to a node with connectivity k ⬘ ,
homogeneous mixing hypothesis 关21兴, is approximated for
times the probability that this node is infected, ␳ k ⬘ (t), aver-
homogeneous networks as 具 k 典 ␳ (t). From Eq. 共2兲 the exis-
aged over all the nodes connected to the original node. Equa-
tence of an epidemic threshold ␭ c ⫽ 具 k 典 ⫺1 关20兴 can be
tions 共3兲 and 共6兲 define together the MF equation describing
proved, such that ␳ ⫽0 if ␭⬍␭ c , while ␳ ⬃(␭⫺␭ c ) if ␭
the SIS model on Markovian complex networks.
⭓␭ c .
The exact solution of this set of equations can be difficult
For general complex networks, in which large connectiv-
to find, depending on the particular form of P(k ⬘ 兩 k). How-
ity fluctuations and correlations might be allowed, we must
ever, it is possible to extract the value of the epidemic thresh-
relax the homogeneous hypothesis made in writing Eq. 共2兲
old by analyzing the stability of the steady-state solutions. Of
and work, instead, with the relative density ␳ k (t) of infected
course, the healthy state ␳ k ⫽0 is one solution. For small ␳ k ,
nodes with given connectivity k; i.e. the probability that a
we can linearize Eq. 共3兲, obtaining
node with k links is infected. Following Refs. 关7,8兴, the rate
equation for ␳ k (t) can be written as d ␳ k共 t 兲
⯝ 兺
L kk ⬘ ␳ k ⬘ 共 t 兲 . 共7兲
d ␳ k共 t 兲 k⬘
⫽⫺ ␳ k 共 t 兲 ⫹␭k 关 1⫺ ␳ k 共 t 兲兴 ⌰ k 共 t 兲 . 共3兲
dt In the previous equation we have defined the Jacobian matrix
L⫽ 兵 L kk ⬘ 其 by
In this case, the creation term is proportional to the spreading
rate ␭, the density of healthy sites 1⫺ ␳ k (t), the connectivity L kk ⬘ ⫽⫺ ␦ kk ⬘ ⫹␭k P 共 k ⬘ 兩 k 兲 , 共8兲
k, and the variable ⌰ k (t) that stands for the probability that a
link emanating from a node of connectivity k points to an where ␦ kk ⬘ is the Kronecker delta function. The solution ␳ k
infected site. In the case of an uncorrelated random network, ⫽0 will be unstable if there exists at least one positive ei-
considered in Refs. 关7,8兴, the probability that a link points to genvalue of the Jacobian matrix L. Let us consider the con-


nectivity matrix C, defined by C kk ⬘ ⫽k P(k ⬘ 兩 k). Using the

symmetry condition, Eq. 共1兲, it is easy to check that if v k is
an eigenvector of C, with eigenvalue ⌳, then P(k) v k is an
eigenvector of the transposed matrix CT with the same ei-
genvalue. From here it follows immediately that all the ei-
genvalues of C are real. Let ⌳ m be the largest eigenvalue of
C. Then, the origin will be unstable whenever ⫺1⫹␭⌳ m
⬎0, which defines an epidemic threshold

␭ c⫽ , 共9兲

above which the solution ␳ k ⫽0 is unstable, and another non-

zero solution takes over as the actual steady state—the en- FIG. 1. Total prevalence ␳ for the SIS model in Callaway’s
demic state. network and in a MR uncorrelated network with the same connec-
It is instructive to see how this general formalism recov- tivity distribution.
ers previous results 关7,8兴, implicitly obtained for random un-
correlated networks. For any random network, in which there exponential decay of P(k,k ⬘ ) for large values of k and k ⬘ , as
are no correlations among the connectivities of the nodes, we can be seen from Eq. 共10兲. The estimate of the largest eigen-
have that the connectivity matrix is given by C kk ⬘
nc value for a connectivity matrix of size 500⫻500 is ⌳ m
⫽k P(k ⬘ 兩 k)⬅kk ⬘ P(k ⬘ )/ 具 k 典 . It is easy to check that the ma- ⬇6.476 56, which yields an epidemic threshold ␭ c ⬇0.15.
trix 兵 C k ⬘ k 其 has unique eigenvalue ⌳ m nc
⫽ 具 k 2 典 / 具 k 典 , corre- This value is stable for connectivity matrices of size larger
than 300⫻300. The epidemic threshold thus obtained is to
sponding to the eigenvector v nc k ⫽k, from where we recover be compared with the prediction for an uncorrelated network,
the above established result, Eq. 共5兲.
c ⫽ 具 k 典 / 具 k 典 ⫽0.20.
␭ nc 2
In order to check the theoretical prediction Eq. 共9兲, we
Figure 1 shows the results of numerical simulations of the
have performed numerical simulations of the network model
SIS model on Callaway’s network, as well as on random
proposed by Callaway et al. 关16兴. This model is defined in a
networks with the same connectivity distribution, generated
very simple way: each time step we add a new node, and,
using the Molloy and Reed 共MR兲 algorithm 关22,23兴. The MR
with probability ␦ , two nodes are randomly selected and
algorithm generates a random network with a prescribed
joined with a link. This model yields a random network with
connectivity distribution and no correlations among nodes,
a exponential connectivity distribution, P(k)⫽(2 ␦ ) k /(1
and thus it is expected to yield an epidemic threshold given
⫹2 ␦ ) k⫹1 , which shows correlations among the connectivi-
ties of the nodes. These correlations can be analytically com- c . Simulations were performed for a fixed value ␦ ⫽1
by ␭ nc
in networks of size up to N⫽107 , averaging over at least 100
puted by means of the joint probability P(k,k ⬘ ), which ful-
fills the recursion relation 关16兴 different starting configurations, performed on at least 10
different realizations of the network. Figure 1 depicts the
2␦ steady-state prevalence as a function of the spreading rate ␭.
P 共 k,k ⬘ 兲 ⫽ 关 P 共 k⫺1,k ⬘ 兲 ⫹ P 共 k,k ⬘ ⫺1 兲兴 For the MR network, the function ␳ (␭) shows a clear linear
1⫹4 ␦
behavior. The epidemic threshold estimated from a least-
P 共 k⫺1 兲 P 共 k ⬘ ⫺1 兲 squares fitting is ␭ c ⫽0.21⫾0.01, in excellent agreement
⫹ . with the prediction for an uncorrelated network. On the other
1⫹4 ␦
hand, Callaway’s network exhibits a very different behavior,
This is a linear equation, that can be solved using the gener- which might be compatible with the presence of a transition
ating functional technique, yielding the joint probability of an infinite order. In fact, Fig. 1 is reminiscent of the be-
havior found in Ref. 关16兴 for the giant component of the

P 共 k,k ⬘ 兲 ⫽ P 共 k 兲 P 共 k ⬘ 兲
1⫹4 ␦ 1⫹2 ␦
共 2 ␦ 兲 2 1⫹4 ␦
冉 冊 k⫹k ⬘ network as a function of the parameter ␦ . In that work, the
size of the giant component was fitted to an stretched expo-
nential with exponent 1/2. Guided by this intuition, in Fig. 2
k⫺1 k ⬘ ⫺1 we perform a fit of the stationary prevalence in Callaway’s
共 k⫹k ⬘ ⫺k ⫺k ⫺2 兲 ! model to the form ␳ (␭)⬃exp关⫺␣(␭⫺␭c)⫺1/2兴 关16兴. The fit
⫻ 兺 兺 1 2
k ⫽0 k ⫽0 共 k⫺k ⫺1 兲 ! 共 k ⬘ ⫺k ⫺1 兲 !
1 2
yields a prefactor ␣ ⫽1.52⫾0.02, and an epidemic threshold
1 2
␭ c ⫽0.11⫾0.02, smaller by a factor 2 than the value corre-
⫻ 冉 冊
1⫹4 ␦
1⫹2 ␦
k 1 ⫹k 2
. 共10兲
sponding to an uncorrelated random network, and in quite
good agreement with the prediction from the largest eigen-
value of the connectivity matrix.
We can numerically estimate the largest eigenvalue of the In summary, we have shown that, in the presence of cor-
connectivity matrix C by generating a finite matrix P(k,k ⬘ ) relations, the epidemic threshold in complex networks is de-
from Eq. 共10兲 and using the relation 共1兲 to find P(k ⬘ 兩 k). The termined by the connectivity matrix C, and not by the con-
correctedness of this numerical estimate is ensured by the nectivity distribution P(k), as happens in uncorrelated


connectivity matrix. At this respect, it might be surprising

that some complex networks, such as the Barabási-Albert
共BA兲 graph 关24兴, are exactly described at the uncorrelated
level given by Eq. 共3兲 with ⌰ k independent of k 关7,8兴. This
fact must be taken as an evidence of the lack of correlations
in the BA model; lack that, on the other hand, has been
already checked numerically in Ref. 关14兴. The formalism
presented in this paper represents a refinement over previous
works because it includes the effects of correlations between
pairs of nodes and, in this sense, it is exact for Markovian
networks. Real networks, such as the Internet, however, pos-
sess a more complex correlation structure. Our formalism
will provide an improved approximation to epidemic dynam-
ics in these cases, but it still remains the task of ascertaining
FIG. 2. Fit of the prevalence for the SIS model in Callaway’s the effects of higher-order correlations. Further work is nec-
network to the form ␳ (␭)⬃exp关⫺␣(␭⫺␭c)⫺1/2兴 . The fit yields the essary in this direction.
values ␣ ⫽1.52⫾0.02 and ␭ c ⫽0.11⫾0.02.
This work has been partially supported by the European
networks. This fact implies that the previously predicted null Commission—Fet Open project Grant No. COSIN IST-2001-
epidemic threshold for SF networks with ␥ ⭐3 might be 33555. R.P.-S. acknowledges financial support from the Min-
shifted in correlated graphs, attaining a positive value de- isterio de Ciencia y Tecnologı́a 共Spain兲. We thank A. Vespig-
pending on the nature of the correlations as given by the nani for helpful comments and discussions.

关1兴 R. Albert and A.-L. Barabási, Rev. Mod. Phys. 74, 47 共2002兲. 关13兴 R. Pastor-Satorras, A. Vázquez, and A. Vespignani, Phys. Rev.
关2兴 S.N. Dorogovtsev and J.F.F. Mendes, Adv. Phys. 51, 1079 Lett. 87, 258701 共2001兲.
共2002兲. 关14兴 A. Vázquez, R. Pastor-Satorras, and A. Vespignani, Phys. Rev.
关3兴 P. Erdös and P. Rényi, Publ. Math. 共Debrecen兲 5, 17 共1960兲. E 65, 066130 共2002兲.
关4兴 D.J. Watts and S.H. Strogatz, Nature 共London兲 393, 440 关15兴 K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. 88, 108701
共1998兲. 共2002兲.
关5兴 R. Pastor-Satorras and A. Vespignani, in Handbook of Graphs 关16兴 D.S. Callaway, J.E. Hopcroft, J.M. Kleinberg, M.E.J. New-
and Networks: From the Genome to the Internet, edited by S. man, and S.H., Strogatz, Phys. Rev. E 64, 041902 共2001兲.
Bornholdt and H. G. Schuster 共Wiley-VCH, Berlin, 2002兲; 关17兴 M. Boguñá, R. Pastor-Satorras, and A. Vespignani 共unpub-
e-print cond-mat/0205260. lished兲.
关6兴 F. Liljeros, C.R. Edling, L.A.N. Amaral, H.E. Stanley, and Y. 关18兴 M.E.J. Newman, e-print cond-mat/0205405.
Aberg, Nature 共London兲 411, 907 共2001兲. 关19兴 V.M. Eguı́luz and K. Klemm, Phys. Rev. Lett. 89, 108701
关7兴 R. Pastor-Satorras and A. Vespignani, Phys. Rev. Lett. 86, 共2002兲.
3200 共2001兲. 关20兴 O. Diekmann and J. Heesterbeek, Mathematical Epidemiology
关8兴 R. Pastor-Satorras and A. Vespignani, Phys. Rev. E 63, 066117 of Infectious Diseases: Model Building, Analysis and Interpre-
共2001兲. tation 共Wiley, New York, 2000兲.
关9兴 R.M. May and A.L. Lloyd, Phys. Rev. E 64, 066112 共2001兲. 关21兴 R.M. Anderson and R.M. May, Infectious Diseases in Humans
关10兴 Y. Moreno, R. Pastor-Satorras, and A. Vespignani, Eur. Phys. J. 共Oxford University Press, Oxford, 1992兲.
B 26, 521 共2002兲. 关22兴 M. Molloy and B. Reed, Random Struct. Algorithms 6, 161
关11兴 M.E.J. Newman, e-print cond-mat/0201433. 共1995兲.
关12兴 S.R. White, Virus Bulletin Conference, Munich 1998 关23兴 M. Molloy and B. Reed, Combinatorics, Probab. Comput. 7,
共unpublished兲 共 295 共1998兲.
Problems/ Problems.html兲. 关24兴 A.-L. Barabási and R. Albert, Science 286, 509 共1999兲.