Mathematical Biosciences 180 (2002) 293305

www.elsevier.com/locate/mbs

Percolation on heterogeneous networks as a model for epidemics

L.M. Sander
a

a,*

, C.P. Warren a, I.M. Sokolov b, C. Simon c, J. Koopman

Michigan Center for Theoretical Physics, University of Michigan, Ann Arbor, MI 48109-1120, USA b Theoretische Polymerphysik, Unversitt Freiburg, Freiburg, Germany a c Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA d Department of Epidemiology, University of Michigan, Ann Arbor, MI 48109, USA Received 2 August 2001; received in revised form 7 December 2001; accepted 19 December 2001

Abstract We consider a spatial model related to bond percolation for the spread of a disease that includes variation in the susceptibility to infection. We work on a lattice with random bond strengths and show that with strong heterogeneity, i.e. a wide range of variation of susceptibility, patchiness in the spread of the epidemic is very likely, and the criterion for epidemic outbreak depends strongly on the heterogeneity. These results are qualitatively dierent from those of standard models in epidemiology, but correspond to real eects. We suggest that heterogeneity in the epidemic will aect the phylogenetic distance distribution of the disease-causing organisms. We also investigate small world lattices, and show that the eects mentioned above are even stronger. 2002 Elsevier Science Inc. All rights reserved.
Keywords: Epidemic model; Percolation; Heterogeneity; Small world theory

1. Introduction The most commonly used models for the spread of an epidemic assume perfect mixing: i.e. all individuals are able to infect all others. In an inhomogeneous population the susceptibility to infection is replaced by its average. To account for spatial eects many workers use reaction diusion equations to describe traveling waves of infection [1] (see below).
Corresponding author. Address: Department of Physics, University of Michigan, Ann Arbor, MI 48109-1120, USA. Tel.: +1-734 764 4471; fax: +1-734 764 6843. E-mail address: lsander@umich.edu (L.M. Sander). 0025-5564/02/$ - see front matter 2002 Elsevier Science Inc. All rights reserved. PII: S 0 0 2 5 - 5 5 6 4 ( 0 2 ) 0 0 1 1 7 - 7
*

294

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

In both cases important phenomena are not reproduced by the models. For example, it is common experience that in plant diseases [2,3] islands of susceptible individuals can be protected by a band of immune ones, a rewall. The heterogeneity in susceptibilities has induced a spatial structure, and thus a map of the infected areas looks patchy. In principle, of course, a suitable compartmental model could be introduced for this case, but it would involve many special assumptions. As we will see, there is a more natural language for such eects, that of percolation theory. Surprisingly, even in human diseases such as AIDS these eects seem to be important [4]. Such eects do not occur in perfect mixing theories in which susceptibilities are replaced by averages. To deal with populations which are not well mixed, it is natural to consider agents on a regular lattice (or some other graph), which should be thought of as a representation of the transmission network. Many models used in statistical physics are of this nature, and a good deal is known about them; an example is percolation theory [5,6]. A very simple treatment of an epidemic would go as follows: each node of a lattice represents the location of an individual. A site can infect its neighbors if it is itself infected. With probability, p, some sites are immune. This is called the site percolation model. Or, if bonds (representing contacts between neighbors) are present with probability p (some transmissions possible, some forbidden) then we have the bond percolation model. Of course, neither of these extreme cases is a very realistic model for an epidemic. Below, we will see how a conventional epidemic model, with nite infection probability, x, for any adjacent pair of nodes, can always be exactly mapped onto bond percolation with some p that depends on x [6]. The important point, for the moment, is that the mapping exists, and it allows us to use what we know about percolation for this problem. For example, on large scales (which is our interest here) the details of the lattice structure do not matter, and simply reect the fact that we assume that infection is locally transmitted. Also, large-scale spatial structure arises naturally in this approach, even though the distribution of bonds which transmit is not correlated. A lattice is said to percolate when a cluster of linked sites spans large distances. The main result of percolation theory is that there is a critical value of the probability, pc , where percolation sets in. For example, for bond percolation on a square lattice (the case we treat below), pc 0:5. For epidemics among humans it is common to imagine that some agents travel long distances. In this case it is natural to think of percolation on a small world lattice [7]. This is a graph where most sites are linked to their local neighbors, but a few have long bonds that can span the entire system. For a small world graph, pc depends on the number of long bonds, but it still exists. Percolation accounts for the existence of rewalls and uninfected islands because near pc only part of the lattice belongs to the spanning cluster giving rise to a random patchy pattern of infection see Fig. 1. However, a close look makes such an application suspect. Islands and rewalls occur only in a rather narrow transition region dp near the critical probability, pc . In nature it would be unlikely that a ne-tuning of parameters to be near pc would occur. This kind of diculty always plagues attempts to apply theories of critical phenomena in the natural world. In this paper we show that introducing heterogeneity in the probability, x, to transmit the infection across a bond solves this problem in an elegant way because it broadens the transition region for the outbreak, so that a generic epidemic could produce large uninfected islands and rewalls. Furthermore, for a small world lattice, the heterogeneity induced by the long bonds has

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

295

Fig. 1. Snapshot of the nal cluster of recovered after the epidemic has died out on a 256 256 lattice with periodic boundary conditions. The cluster of recovered sites is gray, and the uninfected are black. A strong heterogeneity xmin e10 bond distribution (2) is used.

the same eect, even in the absence of heterogeneity in the susceptibility. Of, course, the two eects will probably be present together, and a small world outbreak is particularly likely to give rise to epidemics that aect part of the population and have clusters of infection. It will be also be very far from the assumptions of conventional, perfect mixing theory. There is a further unexpected result of this study. By denition, an epidemic starts when a sick individual infects more than one other before she recovers, i.e. R0 > 1, where R0 , the reproduction number, is the mean number of infections produced by a site. The usual expression for R0 is xS0s where x is the probability per unit time to infect neighbors, S0 the number which can be infected, and s the time to recovery. We will show that heterogeneity can change this formula by a large amount.

2. Model: percolation with heterogeneity In this section we introduce our model which is a version of the standard susceptible, infected, recovered (SIR) epidemic model. Each site of a square lattice is occupied by an individual which can be infected by neighbors connected by bonds (four in our case), and which is, at any time,

296

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

either S, I or R. None of the qualitative results should depend on the details of the local bonding. For example, introducing next-nearest neighbor coupling will not change any of our conclusions. For each bond we have a probability, x, for an I to infect a neighboring S in a time step. This number represents the product of the infectivity of the I, the susceptibility of the S, and the contact rate. This version of the model has been used by Grassberger [6]. It is the same as Example 6 of Durrett [8]. Here we generalize the model for an inhomogeneous population. We assume that the contact rate is inhomogeneous so that the x for a given bond is chosen from a distribution f x. The xs are independent, identically distributed, random variables. That is, we use a weighted network [9] to model variability. As we will see in Eq. (3), below, we will map our weighted network to a uniform one. We have also considered the case of inhomogeneity of susceptibility rather than contact rate: the simulation results are similar in every respect to what we report here. After exactly s time steps any infected site recovers. We start with one I in the middle of the lattice and the rest S. We say that we have an epidemic when the I and R sites span a large lattice, i.e. reach the edges. A snapshot is shown in Fig. 1. If the recovery time, s, is too small the epidemic will die out, and if s is large enough it will persist. A critical value of s thus plays the role of pc . In Fig. 2 we show the spanning probability, M, for a 200 200 lattice as a function of s for two dierent choices of f x. M is the probability to span starting from a single site. We should note that M is the same as the mass of the innite cluster in percolation theory [5] i.e. the fraction of sites which belong to the spanning cluster. This is true because the probability to span starting at any given site is the probability that the site is a member of the innite cluster.

Fig. 2. The probability, M, that an epidemic started from a point spans the lattice as a function of infection time s for two dierent bond distributions and two dierent lattices: () square lattice, (M) triangular lattice. In both cases the dotted line is strong heterogeneity, and the solid line, weak. Strong heterogeneity broadens the transition region in s. The transition region (e.g. the range 0:1 < M < 0:9) is about 12 times larger for strong heterogeneity than for weak. For fw we use xmax 0:003, fs uses xmin e15 . Averages are over 2000 simulations on a 200 200 lattice.

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

297

2.1. Strong and weak heterogeneity We have studied two classes of fs which we call weak and strong heterogeneity. An example of weak heterogeneity is fw x 1=xmax ; 0 6 x 6 xmax : 1 Other functions with a rather narrow range of x give similar eects, and the results do not dier very much from the case with homogeneity, all identical strength bonds [6]. Strong heterogeneity refers to broad distribution such as fs x C=x xmin 6 x 6 1; C 1=j lnxmin j: 2 Any strongly skewed distribution should give similar results to those quoted here. Distribution functions of this type were studied [10] in random resistor networks, and are known, in that context, to give dierent behavior from the ordered case. Random resistor networks have the property of always tuning themselves to be at threshold for suciently strong heterogeneity because the electrical current will always nd a way across the system by the weakest resistance path. In the case of epidemics there is no such principle, but the fact that the epidemic will typically follow a tortuous path across the system is central to our results. The black portion of Fig. 1 corresponds to sites that are protected by the weak bonds at the edge of the infected cluster, the rewall. Eects of this type occur only in the transition region of Fig. 2. The transition region is very broad for strong heterogeneity. Firewalls and large islands of uninfected sites arise purely from the statistics of the system. 2.2. Relationship to random percolation Ordinary bond percolation has some bonds present and some absent. This is quite dierent from our approach where all of our sites correspond to individuals that could be infected. In our model, given enough time, infection can take place across any bond, no matter how weak. If x takes on a single value, i.e. f x dx x0 , then the probability to infect a neighbor before recovery is p 1 1 x0 s [6]. Thus p is the fraction of bonds completed in a given epidemic. If we have heterogeneity we generalize this by writing Z p 1 1 xs f x dx: 3 For a square lattice with pc 0:5, so we expect that when p in Eq. (3) is greater than 0.5 we will have an epidemic. For other transmission networks the value, 0.5, would change, but the result would still be valid. We verify this in Fig. 3 where plot M against p for various choices of f x and compare them to ordinary bond percolation on the same lattice. We have also done a simulation for percolation on a triangular lattice. The value of pc changed, but the data collapse still holds (not shown). 2.3. Threshold for outbreak We can discuss the threshold for outbreak of an epidemic by using the usual SIR equations [1] for the perfect mixing case:

298

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

Fig. 3. Data collapse of the spanning probability M as a function of p, compared with bond percolation. This veries Eq. (3) and shows the models relationship to simple bond percolation.

dS=dt xSI;

dI=dt xSI I=s:

Clearly, we have an outbreak if R0 xS0s > 1. Eq. (4) describes a situation where recovery takes place at rate 1=s rather than a xed time to recover. It is not dicult to show that the outbreak conditions are exactly the same in the two cases. We can think of Eq. (4) as describing dynamics on a fully conjugated graph every node sees every other equally. Up to now we have been considering a case with only local transmission, where an infected individual sees and directly infects only its neighbors. The small world graphs, below, interpolate between the two cases. Because of the uniform transmission, once R0 > 1 the epidemic will inevitably spread from any point, which is quite dierent from a case with local transmission and heterogeneity, though, as in the case of percolation, only a fraction of the Ss will eventually be infected for any nite R0 . We can convert Eq. (4) into a local (uniform) spatial model by adding diusion terms: dS=dt Dr2 S xSI; dI=dt Dr2 I xSI I=s: 5 Now the number of susceptibles, for example, is a function of space, Sr; t. The spread of the infection is by diusion of the I and S with diusion constant D, as represented by the rst term in the equations. Murray shows that the criterion for outbreak is of the same form [1], where St 0 now refers to the initial number of susceptibles which can be infected by the single infected site, a number of order unity. For the Grassberger model (or weak heterogeneity) we have a similar looking result: we put p 1 1 xsc pc 1=2. For small x, R0 xsc O1. Strong heterogeneity is quite dierent. For the common case of xmin ( 1, we can estimate the integral in Eq. (3) using the fs x of Eq. (2). Using 1 xs % expxs, we have an integral related to the exponential integral whose asymptotic expansion gives

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

299

1 xs fs x dx % C lnsxmin =j lnxmin j;

where C % 0:577 is Eulers constant. Using this estimate as we did in the weak disorder case p x above, we nd sc % 0:56= xmin . To estimate R0 we replace x by its average, . The naive estimate mentioned above for the mean reproduction number at the threshold for outbreak gives p 7 R0 sc / 1= xmin j lnxmin j: x Eq. (7) shows that for a heterogeneous system the estimate for the mean reproduction number at threshold is very large, and not of order unity. This means that for such systems the eective R0 is very much smaller than the estimate sc which is an appropriate formula only for perfect-mixing x systems. This statement may seem reminiscent of the well-known fact [11] that for diseases such as AIDS the naive formula for R0 does not work if there are very active agents, because active agents have an outsized eect in spreading the disease. However, the results are, in fact, in opposite directions. Here, the spatial correlations between strong bonds make them less eective than they might seem to be. A clump of very infective bonds does not do as much damage as we might suspect since the infected sites are sharing the same victims. In the transition region there are always bottlenecks to the spread of the infections, and thus random spatial correlations are always important. We can understand the broadening of the transition region due to heterogeneity in the same way. Using the expressions in Eqs. (3) and (6) we nd that the width of the transition region, ds, is given by ds=sc j lnxmin jdp. We can dene the transition region, dp, as the range over which M increases from, say, 0.1 to 0.9; it is of order 0.05. We see that ds=sc is large as xmin ! 0. This is illustrated in Fig. 2. 2.4. Tracking the spread of the epidemic Our theory is based on an assumption, that of strong heterogeneity, which is dicult to verify directly. Whether we are dealing with the heterogeneous case or not, the sort of detailed geographic information that we would need to verify that we are dealing with an epidemic near a percolation threshold is rarely available. However, we show here that there is a measure of these eects which could be collected, and which is quite interesting in its own right, namely the genetic dierence (the dierence in the sequence of base pairs in the DNA) between the infectious agents during the course of the transmission of the infection through the population. The point of this measurement would be to deduce information about the spread of the infection from the genetic makeup of the pathogens. We argue that the genetic dierence, which could be measured, is related to the phylogenetic distance which is related to the transmission pattern, and which we now dene. Phylogenetic distance between diseases of individuals is a measure the genetic dierences between those two diseases. Assuming one initial infected site, the diseases of any two infected or recovered individuals in the population will share a most recent common ancestor, the most recent ancestor to which both of the infections may be traced back through infectors. Suppose that there have been n1 transmissions from this ancestor to one of the infected/recovered individuals and n2 transmissions back to the other infected/recovered. The phylogenetic distance is dened to be n1 n2 .

300

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

The basis of this is the fact that the relationship between the genetic diversity of infectious agents and transmission history is strong enough to permit inferences about past transmission histories in a population from genetic diversity at one point in time [12]. For example, we might assume that the genetic diversity of the agents is proportional to the number of transmissions of the disease, which we can easily model. However, whether the diversity is better correlated with the number of transmissions to the most recent common ancestor or to the time to the most recent common ancestor is controversial [13]. Bottleneck events that x randomly generated diversity account for most of the population diversity seen in infectious agents. Most likely for infectious agents undergoing few selective events within a host the number of transmissions mostly determines the diversity, as these are the most frequent source of bottleneck events. For HIV, however, so many events occur within the host that time, not transmission events, seems to be the source of diversity [13]. For our model, time and the number of transmissions to the most recent common ancestor are highly correlated since infection has a xed duration. Time between transmission events would add noise to the relationships we discuss. However, even in the infection where such noise is maximal, namely HIV, the relationship between diversity and transmission history remains strong enough allow one to infer past transmission dynamics from current patterns of diversity [12]. In our calculations we can keep track of the number of transmissions of the disease. In Fig. 4 we show the histogram of the number of such transmissions from a common ancestor for two examples, an epidemic near threshold, and one well above threshold. For near threshold behavior, the complex path that the disease takes is reected in the large spread of the distribution, while above threshold there is a denite peak, a characteristic number of transmissions. We can also collect data for distances between individuals at a given time. These show similar behavior.

Fig. 4. The frequency of occurrence of path distances from a recovered site through its infectors back to the origin averaged over 1000 simulations on a 200 200 lattice. Note that the overall number of paths is larger well above threshold; near threshold there are many bottlenecks in the spread of the epidemic.

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

301

If this sort of information could be collected, a signature of near threshold epidemics would be in hand. This might even have implications for intervention: a near threshold epidemic must follow a tortuous path, and could be susceptible to disruption. 3. Small world eects In epidemics of human diseases some (usually a small number) of individuals travel long distances. We use a small-world lattice to account for these contacts by adding a small fraction, /, of long bonds connecting randomly chosen sites [7]. The number of long bonds in our case is 2Np/ because there are 2N bonds in a square lattice, where N is the number of sites. The result is shown in Fig. 5. Even a very small / breaks up the large uninfected regions. The percolation threshold is lowered, and below the old pc there are many small regions of epidemic connected by long bonds. However, the eects of heterogeneity remain. Long distance contacts are not the same as perfect mixing: rewalls still occur over a substantial range of parameters. In this section we will investigate these matters by doing percolation simulations rather than dynamic epidemic simulations as above. These are very much faster, and allow us to get far better statistics. In fact, we use the very fast algorithm of Newman and Zi [14] which allows us to do very large simulations. To translate to epidemic language, use the results of Eq. (3), above. We should also note that these results are new: previous work (e.g., [7,15]) on small world lattices

Fig. 5. Snapshot of the nal cluster of recovered on a 256 256 small-world lattice with periodic boundary conditions and / 0:01.

302

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

considered long bonds superimposed on an underlying one-dimensional lattice. This model can be solved exactly but is not easily generalized to higher dimensions. The treatment below can be generalized. 3.1. Shift of the percolation threshold To understand these results we extend the work of Newman and Watts [15] to our case. The infected islands of Fig. 5 can be considered to be nodes of a random graph whose edges are the long bonds. In percolation language, these islands are the nite connected clusters which occur below the square lattice threshold. The random graph percolates if there are twice as many such clusters as long bonds [16]; this occurs for some p < pc , which we will now nd. We need to count the nodes of the random graph. To do this we use the average cluster size P P dened as  s2 ns = sns where ns is the number of clusters of size s. This size-weighted avn erage is the correct average because it counts the probability that a given site (the end of a long bond) belongs to a cluster of size s [5]. Scaling theory gives  Kjp pc jc where K is a constant, n and c is a critical exponent equal to 43=18 2:39 for our case [5]. We did a percolation simulation for a large square lattice 1024 1024, and found that  ts this formula within a few percent if n we take the non-universal parameter K to be 0.132. n The number of clusters is Ni N =, where N is the total number of sites. The number of long bonds that connect two sites that are parts of clusters is B 2Np/. Setting Ni =2 B gives a criterion for percolation on a small world lattice with no adjustable parameters: 4/K0:5 c 1; p p 8 where  is the percolation threshold for the small world lattice. p We also need to redene the percolation threshold since it clearly has no relationship with spanning in the presence of long bonds. We return to the denition of percolation on a random graph, namely the appearance of a giant component [16], that is, a connected component of the graph containing a nite fraction of all the sites in the limit N ! 1. We plot in Fig. 6, for various N, the fraction of the lattice occupied by the largest cluster for a small world lattice, and compare to the case for / 0, the parent square lattice. We can see that there does seem to be a limiting fraction for large N which we identify as the giant component. Interestingly, the nature of the scaling towards the limit (the nite-size scaling) seems to be quite dierent for the two cases. Furthermore, Mp is linear near threshold for the small world case; i.e. the critical exponent, b, is unity, as it should be for random graph percolation. We found  by p linear extrapolation. In Fig. 7 we show the depression of the percolation threshold as a function of /. We also show a t to the formula in Eq. (8) using the known value of c and the measured K. The agreement with our measured values of  is essentially perfect. p 3.2. Broadening of the transition region As we see in Fig. 5 there are still islands surrounded by a rewall in the small world case. Further, we see from Fig. 6 the transition region, even as a function of p is quite a lot broader than for ordinary percolation. Expressing the plot as a function of s using Eq. (3) would further

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

303

Fig. 6. The giant cluster for a small world and an ordinary lattice for various lattice sizes. N L2 where N is the number of sites.

Fig. 7. Dependence of the percolation threshold on / } compared with the scaling theory prediction on a 1024 1024 lattice.

emphasize the eect, so that a small world epidemic with heterogeneity would have a very broad transition region.

304

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

In fact, the addition of long bonds introduces another type of heterogeneity. A few long bonds do not make the system more like the random mixing case, where parameters are replaced by their averages. Random mixing and average behavior is only attained in the limit where every site is linked to every other, i.e. very large /. The main eect of long bonds for small / is to allow an outbreak which is limited to a small cluster of susceptible sites to be transferred to another, distant cluster in conditions where the epidemic could not spread uniformly with local coupling. 4. Summary Our model for percolation with strong heterogeneity is intended to model extreme variability in a population. Our calculations illustrate two eects: there is a mechanism which produces patches of uninfected but susceptible individuals without ne-tuning of parameters. And for this case R0 depends on the heterogeneity. Whether variability of this sort occurs in real epidemics is not known. However, it is recognized that variability in susceptibility increases when epidemics recur. Our model may apply best to that case. We can consider the extension of this work to other sorts of networks. For example, it has recently been pointed out [17] the contact network on which sexually transmitted diseases spread may be of a dierent nature from what we have considered. Rather it might be a scale-free network [18], in which the connectivities of the bonds have a power law distribution. We could do our simulations in this context with minor changes. Acknowledgements L.M.S. would like to thank R.M. Zi and M.E.J. Newman for helpful discussions. J.K. is supported by EPA grant CR 827427-01-0, and by NIH project grant from NIAID no. F002732.

References
[1] J.D. Murray, Mathematical Biology, Springer, New York, 1993. [2] M.J. Jeger, Spatial Components of Plant Disease Epidemics, Prentice Hall, Engelwood Clis, 1989. [3] D.J. Bailey, W. Otten, C.A. Gilligan, Saprotrophic invasion by the soil-borne fungal plant pathogen Rhizctonia solani and percolation thresholds, New Phytol. 146 (2000) 535. [4] S.R. Friedman et al., Network related mechanisms may help explain long-term HIV-1 seroprevalence levels that remain high but do not approach population-group saturation, Am. J. Epidemiol. 152 (2000) 913. [5] D. Stauer, A. Aharony, Introduction to Percolation Theory, Taylor and Francis, Bristol, PA, 1994. [6] P. Grassberger, On the critical behavior of the general epidemic process and dynamic percolation, Math. Biosci. 63 (1983) 157. [7] D.J. Watts, Small Worlds: the Dynamics of Networks between Order and Randomness, Princeton University, Princeton, NJ, 1999. [8] R. Durrett, Stochastic spatial models, SIAM Rev. 41 (1999) 677. [9] R. Albert and A.-L. Barabsi, Statistical Mechanics of Complex Networks, cond-mat/0106096. a [10] V. Ambegaokar, B.I. Halperin, J.S. Langer, Hopping conductivity in disordered systems, Phys. Rev. B 4 (1971) 2612.

L.M. Sander et al. / Mathematical Biosciences 180 (2002) 293305

305

[11] R.M. May, R.M. Anderson, The transmission dynamics of human immunodeciency virus (HIV), Philos. Trans. Roy. Soc. B 321 (1988) 239. [12] O. Pybus, A. Rambaut, P.H. Harvey, An integrated framework for the inference of viral population history from reconstructed genealogies, Genetics 155 (2000) 1429. [13] S. Wolinsky, G.H. Learn, Levels of diversity with and among host individuals, in: K.A. Crandall (Ed.), The Evolution of HIV, Oxford University, Oxford, 1999, p. 275. [14] M.E.J. Newman, R.M. Zi, Ecient Monte Carlo algorithm and high-precision results for percolation, Phys. Rev. Lett. 85 (2000) 4104. [15] M.E.J. Newman, D.J. Watts, Scaling and the percolation threshold in the small-world network model, Phys. Rev. E 60 (1999) 7332. [16] N. Alon, J.H. Spencer, The Probabilistic Method, Wiley, New York, 1999. [17] F. Liljeros, C.R. Edling, L.A. Nunes Amaral, H. Stanley, Y. Aberg, The web of human sexual contacts, Nature 411 (2001) 907. [18] A.-L. Barabasi, R. Albert, Emergence of scaling in random networks, Science 286 (1999) 509.