G. Caldarelli,
1
A. Capocci,
2
P. De Los Rios,
3,4
and M. A. Mun˜oz
5
1
INFM UdR ROMA1 Dipartimento Fisica, Universita` di Roma ‘‘La Sapienza,’’ Piazzale Aldo Moro 2 00185, Roma, Italy
2
De´partement de Physique, Universite´ de FribourgPe´rolles, CH1700 Fribourg, Switzerland
3
Institut de Physique The´orique, Universite´ de Lausanne, CH1004 Lausanne, Switzerland
4
INFM UdR Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy
5
Instituto de Fı ´sica Teo´rica y Computacional Carlos I, Universidad de Granada, Facultad de Ciencias, 18071Granada, Spain
(Received 15 July 2002; published 3 December 2002)
A new mechanism leading to scalefree networks is proposed in this Letter. It is shown that, in many
cases of interest, the connectivity powerlaw behavior is neither related to dynamical properties nor to
preferential attachment. Assigning a quenched ﬁtness value x

to every vertex, and drawing links
among vertices with a probability depending on the ﬁtnesses of the two involved sites, gives rise to what
we call a goodgetricher mechanism, in which sites with larger ﬁtness are more likely to become hubs
(i.e., to be highly connected).
DOI: 10.1103/PhysRevLett.89.258702 PACS numbers: 89.75.Hc, 89.75.Da
Complex networks [1–3] are attracting much interest
as one of the most promising tools to describe a large
variety of social [4], biological [5,6], and technological
systems, as the Internet [7–10] or the World Wide
Web (WWW) [11]. Networks are abstract mathematical
objects composed by vertices (sites) connected by arcs
(links). In the aforementioned examples, vertices can
represent people, proteins, species, routers, or html docu
ments, while arcs correspond to acquaintances, physical
interactions, predation relationships, cable connections,
or hyperlinks, respectively. In recent developments,
scalefree (SF) networks have emerged in many different
contexts, as the WWW, the Internet, Email and
scientiﬁccitation networks, protein and gene interaction
networks, etc., and have become paradigmatic [2,3]. In all
these examples, the degree k of a vertex, i.e., the number
of arcs linking it to other vertices, is powerlaw distribu
ted, P(k) ·k
¡
. SF networks also present the socalled
smallworld phenomenon [12], that is, by a few selected
jumps (that can be either short or long range) it is possible
to reach very different regions of the system and appa
rently distant environments.
To understand how SF networks arise, the concepts of
growing networks and of preferential attachment have
been introduced [2]. In particular, in the best known SF
network model, i.e., the Baraba´siAlbert (BA) one, net
works grow at a constant rate (modeling the fact that new
Web pages are continuously created, new proteins emerge
by mutation, and so forth), and new vertices are attached
to older ones with a probability ¡(k) which is a (linearly
[13]) growing function of the number of preexisting links,
k, at every site. In this way, a highly connected vertex is
more likely to receive further links from newly arriving
vertices: This is the socalled ‘‘rich get richer’’ rule.
In some other, recently proposed, models of protein
interaction networks [14,15] and of the WWW, new ver
tices (proteins and Web pages, respectively) are added by
copying (replicating) existing vertices, borrowing some
of their links, and adding some new others. It has been
shown that this mechanism leads also to an effective
preferential attachment mechanism.
Yet, although in some contexts preferential attachment
can be a very reasonable assumption, in many others it is
certainly not. In particular, in some situations, the infor
mation about the degree of each and every single vertex is
not available to newly added sites, neither in a direct nor
in an effective way. Instead, it is reasonable that two
vertices are connected when the link creates a mutual
beneﬁt (here we restrict ourselves to bidirectional links)
depending on some of their intrinsic properties (authori
tativeness, friendship, social success, scientiﬁc relevance,
interaction strength, etc.). Therefore, it is reasonable to
expect that for some of these systems the P(k) scalefree
behavior (when existing) has an origin unrelated to pref
erential attachment.
In order to explore this simple idea, we propose the
following networkbuilding algorithm: (i) Start by creat
ing a total (large) number N of vertices. At every vertex 
a ﬁtness x

, which is a real number measuring its impor
tance or rank, is assigned. Fitnesses are random numbers
taken from a given probability distribution ¡(x). (ii) For
every couple of vertices, , ], a link is drawn with a
probability ](x

, x
]
) (] a symmetric function of its argu
ments) depending on the ‘‘importance’’ of both vertices,
i.e., on x

, x
]
. Some remarks are in order before proceed
ing further.
(i) The concept of ‘‘verteximportance’’ or ﬁtness
has been already introduced successfully in the ﬁeld of
complex networks, but as an additional ingredient on top
of the BA network [16]. Contrarily, here we put the
emphasis on ﬁtness itself, by eliminating the preferential
attachment rule. (ii) A trivial realization of the above
rules is the standard Erdo˝sRe´nyi model [17], where
](x

, x
]
) is constant and equal to ¡ for all vertex couples.
VOLUME 89, NUMBER 25 P H YS I CA L R E VI E W L E T T E RS 16 DECEMBER 2002
2587021 00319007,02,89(25),258702(4)$20.00 2002 The American Physical Society 2587021
This particular choice does not produce SF networks, but
in what follows we will show that other realizations of the
general rules do so. (iii) The model, as deﬁned, is static,
but it can straightforwardly be considered a dynamic one
by adding new vertices at every time step and linking
them to the existing ones according to the above attaching
rule. (iv) It is also easy to generalize the model to include
asymmetric or directed links. (v) A somehow similar
static model was studied by Goh et al. [18].
A general expression for P(k) can be easily derived.
Indeed, the mean degree of a vertex of ﬁtness x is simply
k(x) ÷ N
Z
«
0
](x, y)¡(y)Jy ÷ NE(x) (1)
[with x

÷ (0, «)]. Assuming E(x) to be a monotonous
function of x, and for large enough N, we have the simple
relation,
P(k) ÷ ¡
E
1
k
N
J
Jk
E
1
k
N
. (2)
For ﬁnite values of N corrections to this equation emerge
[19]. As a particular example, consider ](x

, x
]
) ÷
(x

x
]
),x
2
M
, where x
M
is the largest value of x in the net
work. Then
k(x) ÷
Nx
x
2
M
Z
«
0
y¡(y)Jy ÷ N
(x)x
x
2
M
, (3)
and we have the simple relation,
P(k) ÷
x
2
M
N(x)
¡
x
2
M
N(x)
k
. (4)
A particularly simple realization of the model emerges if
we consider powerlaw distributed ﬁtnesses. This choice
can be naturally justiﬁed by arguing that power laws
appear rather generically in many contexts when one
ranks, for example, people according to their incomes
or cities according to their population, etc. This is the so
called Zipf law which establishes that the rank R(x)
behaves as R(x) · x
n
in a quite universal fashion [20].
The reason for the ubiquitous presence of the Zipf law
yields on the multiplicative nature of the intrinsic ﬂuctua
tions which generically leads to ﬂat distributions in loga
rithmic space and, consequently, to power laws [20].
Clearly, if ¡(x) ·x
¡
(Zipf ’s behavior, with ¡ ÷ 1 +
1,n [20]) then, using Eq. (4), also the degree distribution
P(k) is a power lawand the network shows SF behavior. In
Fig. 1, we show the degree distributions from simulations
with ¡ ÷ 2.5, 3, 4 (corresponding to Zipf exponents n ÷
2,3, 1,2, 1,3); the asymptotic behavior is, in all cases,
well described by Eq. (4). This result is hardly surprising:
From SF ﬁtnesses we generate SF networks, but still it
provides a new generic path to SF networks and takes into
account the widespread occurrence of the Zipf ’s behavior
in nature. In order to extend this result and check whether
SF networks can be generated even when ¡(x) is not SF
itself, we consider an exponential distribution of ﬁtnesses,
¡(x) ÷ e
x
(representing a random, Poisson distribution)
and ](x

, x
]
) ÷ 0jx

+x
]
z(N), where 0(x) is the usual
Heaviside step function. This represents processes where
two vertices are linked only if the sum of their ﬁtnesses is
larger than a given threshold z(N). Using these rules, we
obtain analytically (and conﬁrm in computer simula
tions) that P(k) ·k
2
[21]. This leads to the nontrivial
result that even nonscalefree ﬁtness distributions can
generate scalefree networks (see Fig. 2). Also different
implementations of the threshold rule, such as ](x

, x
]
) ÷
0jx
n

+x
n
]
z(N)
n
 (where n is an integer number) give
rise to the same inverse square behavior (although, in
some cases, with logarithmic corrections).
In a future publication we will explore, in a more
systematic way, the necessary and sufﬁcient conditions
for the ﬁtness distribution and attaching rule under which
wellbehaved SF networks are generated.
Let us stress that the model, as deﬁned, has a diverging
average connectivity in the large N limit, as can be easily
inferred from Eq. (1); i.e., it is severely accelerated [22].
Nevertheless, we can introduce in a rather natural way an
upper cutoff accounting for the fact that every site has
limited information on the rest of the world and, there
fore, connection is attempted with a ﬁnite number, m, of
different sites. Alternatively, vertices can be linked
with the above rule and, after that, links are kept with
probability ¡ (so that, for example, ¡N ÷ m). By includ
ing this modiﬁcation, the N factor in Eq. (1) is substituted
by m, and the connectivity is ﬁnite in the thermodynamic
limit. In order to generate different accelerated net
works (with the averaged connectivity not reaching a
stationary value but growing with N in different pos
sible ways [22]), other selection rules can be easily
implemented.
FIG. 1. Graph representations of four networks produced,
respectively, with (a) ¡ ÷ 2.5, (b) ¡ ÷ 3.0, (c) ¡ ÷ 4.0, and
(d) ¡(x) ÷ e
x
with a threshold rule. Graphs have been pro
duced with the Pajek software.
VOLUME 89, NUMBER 25 P H YS I CA L R E VI E W L E T T E RS 16 DECEMBER 2002
2587022 2587022
To have a more extensive picture of the nature of the
networks under consideration, we have studied the follow
ing topological properties [3], interest in which has been
triggered by recent studies on the Internet structure
[10,23]: (i) the average distance (J), measuring the aver
age minimum number of arcs needed to connect two
given sites; (ii) the average neighbor connectivity k
nn
(k),
measuring the average degree of vertices neighbor of a
kdegree vertex; (iii) the clustering coefﬁcient c(k) that
measures the degree of interconnectivity of nearest
neighbors of kdegree vertices. More speciﬁcally, the
clusteringcoefﬁcient c

of a vertex , whose degree is
k

, is the ratio between the number of edges e

in the
subgraph identiﬁed by its neighbors and the maximum
possible number of edges in the subgraph. That is c

÷
2e

,k

(k

1) [2]. c(k) is obtained by averaging c

for all
vertices with ﬁxed degree k; (iv) the probability distribu
tion of the betweenness, l

, deﬁned as the total number of
minimum paths between any couple of vertices in the
network passing through vertex  [24]. This quantity gives
a measure of the amount of trafﬁc passing through a
vertex. We studied, as in the aforementioned papers,
both the probability distribution P(l) and its ﬁrst mo
ment (l),N.
Computer simulations of our model show that networks
with powerlaw distributed ﬁtnesses, and different values
of ¡, show nearly constant k
nn
(k) and c(k), just as occurs
for the original BA model [2]. The distribution of be
tweenness decays as a power law with an exponent ¡
l

2.2 for ¡ ÷ 2.5 and ¡ ÷ 3, and ¡
l
 2.6 for ¡ ÷ 4. This
is in good agreement with that conjectured in Ref. [18]:
All networks with 3 ¯ ¡ > 2 can be classiﬁed in only
two groups according to the value of ¡
l
( ¡
l
÷ 2 and
¡
l
÷ 2.2, respectively), while for larger values of ¡,
larger nonuniversal values of ¡
l
are reported.
The exponential case behaves in a different way: For a
network of size N ÷ 10
4
, z ÷ 10, and m ÷ N, we ﬁnd
(J) ÷ 2, (c) ’ 0.1, and (l),N ’ 0.1, but a powerlaw be
havior is found for the clustering magnitudes, i.e., (k
nn
) ·
k
0.85
and c(k) · k
1.6
. The betweenness distribution,
instead, shows an unexpected behavior, giving a power
law tail with an exponent ¡
l
 1.45 (see Fig. 4). It is
worth remarking that our model having ¡ ÷ 2 is not
included in the previously discussed classiﬁcation of be
tweenness exponents [18] (Fig. 3).
Having explored the most basic properties of the model
and some particular realizations, let us comment now on
possible applications.
Email networks [25] are a good candidate to be repre
sented by our model. In this case growth may occur, but
agents (Email senders) do not have any access or knowl
edge of the degree of the receivers. Rather than prefer
ential attachment there should be some intrinsic feature of
the receiver playing a role in the phenomenon.
To further emphasize the utility of this new mecha
nism, let us mention the following possibility: One can
imagine situations where a Poisson network is seen as SF
just because the exploration method implicitly imple
ments a probabilistic rule depending on the ﬁtnesses
(this applies, for example, when links are detected by
‘‘picking’’ them one by one, but not if the network is
10
−3
10
−2
10
−1
10
0
10
1
10
2
10
3
10
−3
10
−2
10
−1
10
0
ρ(x)=x
−2.5
ρ(x)=x
−3.0
ρ(x)=x
−4.0
ρ(x)=e
−x
γ
b
=2.2
γ
b
=2.6
γ
b
=1.45
b/N
P
(
b
’
>
b
)
FIG. 4 (color online). Integrated betweenness distribution,
P(l
/
> l) · l
1¡
l
, for different ﬁtness distributions.
10
0
10
1
10
2
10
3
10
4
10
−3
10
−2
10
−1
10
0
10
1
10
2
10
3
10
4
k
nn
y~k
−0.85
c(k)
y~k
−1.6
10
0
10
1
10
2
10
3
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
data
y ~ k
−2
k
P
(
k
)
k
c
(
k
)
,
k
n
n
FIG. 3 (color online). Average neighbor connectivity k
nn
against k, for networks generated using ¡(x) ·e
x
and a
threshold rule. Results are averaged over 100 realizations of
size 10
4
. Inset: The degree distribution P(k); the straight line
corresponds to k
2
.
10
0
10
1
10
2
10
3
10
−5
10
−4
10
−3
10
−2
10
−1
10
0
ρ(x)~x
−2.5
y~k
−2.5
ρ(x)~x
−3.0
y~k
−3.0
ρ(x)~x
−4.0
y~k
−4.0
k
P
(
k
)
FIG. 2 (color online). Degree distribution for networks gen
erated using ¡(x) ·x
¡
with ¡ ÷ 2.5, 3, 4, and ](x

, x
]
) ÷
(x

x
]
),x
2
M
; power laws with the corresponding analytical values
are explicitly drawn in straight lines. Data obtained for net
works with 10
4
vertices, averaged over 100 realizations.
VOLUME 89, NUMBER 25 P H YS I CA L R E VI E W L E T T E RS 16 DECEMBER 2002
2587023 2587023
explored by crawling on it). Let us think, for example, of
the case with threshold type of attaching rule. If only
links with corresponding ﬁtnesses over threshold are
‘‘seen’’ by the exploration method then, for example, an
Erdo˝sRe´nyi network with exponentially distributed ﬁt
nesses can be seen as SF (with, obviously, a connectivity
upper cutoff related to the maximum connectivity of the
underlying network; in cases in which this connectivity is
high, one can generate hubs in the ‘‘apparent’’ SF net
work). In particular, this scenario could be of relevance to
protein networks. Let us argue why.
The way comprehensive protein networks have been
obtained until now is through a baitprey method, named
‘‘twohybrid’’ method: Two proteins are hybridized with
two fragments of a transcription factor (a protein that
binding to a gene promotes its transcription into the
corresponding RNA). The spliced promoter does not
bind to the gene, transcription is inhibited, and the cor
responding RNA is absent. Yet, if the two proteins inter
act, they bring together the two promoter fragments
allowing it to bind to the gene and transcription to start.
The presence of the corresponding RNA signals the in
teraction between the two proteins. We can imagine that
the interaction strength between the two proteins has to be
above a given threshold, else the typical promoter binding
time will be too short for the RNA polymerase to bind to
the gene and initiate transcription. In turn, it is reasonable
to assume that the interaction strength is a function of
some properties of the two proteins (such as, for example,
their hydrophobicity, or their accessible surface area).
This possibility has still to be checked through an analy
sis of the detailed physics behind the twohybrid method.
In summary, we have presented an alternative model to
justify the ubiquity of SF networks in nature. It is a
natural generalization of the standard Erdo˝sRe´nyi. The
main result is that emergence of SF properties is not
necessarily linked to the ingredients of growth and
preferential attachment. Instead, static structures charac
terized by quenched disorder (for different disorder dis
tributions) and threshold phenomena may generate effects
very similar to those measured in the real data. In par
ticular, we recover the powerlaw behavior of degree,
betweenness, and clusteringcoefﬁcient distributions. We
believe that this model is particularly suitable for situ
ations where the degree value of nodes is not publicly
available.
We thank R. PastorSatorras for very useful comments
and suggestions, and P. Hurtado for a critical reading of
the manuscript. The authors thank FET Open Project IST
200133555 COSIN, P. D. L. R. thanks the OFESBern
(Grant No. 02.0234), and M. A. M. acknowledges ﬁnan
cial support fromthe MCYT (Grant No. BFM20012841).
[1] S. H. Strogatz, Nature (London) 410, 268 (2001).
[2] R. Albert and A.L. Baraba´si, Rev. Mod. Phys. 74, 47
(2002), and references therein.
[3] S. N. Dorogovtsev and J. F. F. Mendes, Adv. Phys. 51,
1079 (2002).
[4] M. E. J. Newman, Proc. Natl. Acad. Sci. U.S.A. 98, 404
(2001).
[5] H. Jeong, S. P. Mason, A.L. Baraba´si, and Z. N. Oltvai,
Nature (London) 411, 41 (2001).
[6] R.V. Sole´ and J. M. Montoya, Proc. R. Soc. London B
268, 2039 (2001); J. Camacho et al., Phys. Rev. Lett. 88,
228102 (2002).
[7] M. Faloutsos, P. Faloutsos, and C. Faloutsos, Pro
ceedings of the ACM, SIGCOMM [Comput. Commun.
Rev. 29, 251 (1999).
[8] G. Caldarelli, R. Marchetti, and L. Pietronero, Europhys.
Lett. 52, 386 (2000).
[9] R. PastorSatorras and A. Vespignani, Phys. Rev. Lett. 86,
3200 (2001).
[10] R. PastorSatorras, A. Vazquez, and A. Vespignani, Phys.
Rev. Lett. 87, 258701 (2001).
[11] R. Albert, H. Jeong, and A.L. Baraba´si, Nature
(London) 401, 130 (1999).
[12] D. Watts, SmallWorlds (Princeton University Press,
Princeton, New Jersey, 1999).
[13] SF networks are generated only if ¡(k) ·k; sublinear
¡(k) produce skewed but not SF networks; P. L.
Krapivsky and S. Redner, Phys. Rev. E 63, 066123 (2001).
[14] A. Vazquez, A. Flammini, A. Maritan, and
A. Vespignani, condmat/0108043.
[15] R.V. Sole´, R. PastorSatorras, E. D. Smith, and T. Kepler,
Adv. Compl. Systems 5, 42 (2002).
[16] G. Bianconi and A.L. Baraba´si, Europhys. Lett. 54, 436
(2001).
[17] P. Erdo˝s and A. Re´nyi, Bull. Inst. Int. Stat. 38, 343
(1961).
[18] K.I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. 87,
278701 (2001); K.I. Goh et al., condmat/0205232.
[19] P. L. Krapivsky and S. Redner, condmat/0207107.
[20] M. Marsili and Y.C. Zhang, Phys. Rev. Lett. 80, 2741
(1998);
[21] Being more precise, we obtain a power law plus a delta
function of amplitude exp(z) at K ÷ N, corresponding
to the fact that sites with x

above threshold are con
nected to every other site. This is irrelevant for large
enough z and N.
[22] S. N. Dorogovtsev and J. F. F. Mendes, in Handbook of
Graphs and Networks: From the Genome to the Internet,
edited by S. Bornholdt and H. G. Schuster (Wiley, Berlin,
2002), p. 320.
[23] A. Vazquez, R. PastorSatorras, and A. Vespignani, Phys.
Rev. E 65, 066130 (2002); condmat/0206084.
[24] See M. E. J. Newman, Phys. Rev. E 64, 016131 (2001); 64,
016132 (2001), and references therein.
[25] H. Ebel, L.I. Mielsch, and S. Bornholdt, Phys. Rev. E 66,
035103 (2002).
VOLUME 89, NUMBER 25 P H YS I CA L R E VI E W L E T T E RS 16 DECEMBER 2002
2587024 2587024
Graph representations of four networks produced. we have the simple relation. y ydy NF x (1) 0 [with xi 2 0. the mean degree of a vertex of ﬁtness x is simply Z1 k x N f x. 1. As a particular example. 1]. (v) A somehow similar static model was studied by Goh et al. [18]. consider f xi . d ÿ1 k ÿ1 k P k F F : (2) N dk N For ﬁnite values of N corrections to this equation emerge [19]. N UMBER 25 PHYSICA L R EVIEW LET T ERS 16 DECEMBER 2002 This particular choice does not produce SF networks. 2 x2 x M P k M k : Nhxi Nhxi FIG.VOLUME 89. Assuming F x to be a monotonous function of x. xj xi xj =x2 . (iv) It is also easy to generalize the model to include asymmetric or directed links. but in what follows we will show that other realizations of the general rules do so. with (a) . but it can straightforwardly be considered a dynamic one by adding new vertices at every time step and linking them to the existing ones according to the above attaching rule. Indeed. is static. and for large enough N. Then Nx Z 1 hxix k x 2 y ydy N 2 . (iii) The model. A general expression for P k can be easily derived. (3) xM 0 xM and we have the simple relation. where xM is the largest value of x in the netM work. respectively. as deﬁned.
(b) . 2:5.
(c) . 3:0.
4:0. and (d) x eÿx with a threshold rule. consequently. Graphs have been produced with the Pajek software. people according to their incomes or cities according to their population. for example. This choice can be naturally justiﬁed by arguing that power laws appear rather generically in many contexts when one ranks. (4) A particularly simple realization of the model emerges if we consider powerlaw distributed ﬁtnesses. Clearly. This is the socalled Zipf law which establishes that the rank R x behaves as R x / xÿ in a quite universal fashion [20]. if x xÿ. etc. to power laws [20]. The reason for the ubiquitous presence of the Zipf law yields on the multiplicative nature of the intrinsic ﬂuctuations which generically leads to ﬂat distributions in logarithmic space and.
(Zipf ’s behavior. with .
(4). also the degree distribution P k is a power law and the network shows SF behavior. 1. 1 1= [20]) then. we show the degree distributions from simulations with . using Eq. In Fig.
as deﬁned. in a more systematic way. vertices can be linked with the above rule and. 4 (corresponding to Zipf exponents 2=3. m. in all cases. other selection rules can be easily implemented. xj xi xj ÿ z N. in some cases. By including this modiﬁcation. Using these rules. In order to extend this result and check whether SF networks can be generated even when x is not SF itself. we obtain analytically (and conﬁrm in computer simulations) that P k kÿ2 [21]. (4). Alternatively. and the connectivity is ﬁnite in the thermodynamic limit. In order to generate different accelerated networks (with the averaged connectivity not reaching a stationary value but growing with N in different possible ways [22]).e. 1=2. i. x eÿx (representing a random. This represents processes where two vertices are linked only if the sum of their ﬁtnesses is larger than a given threshold z N. This result is hardly surprising: From SF ﬁtnesses we generate SF networks. as can be easily inferred from Eq. Let us stress that the model. well described by Eq. the asymptotic behavior is. (1) is substituted by m. has a diverging average connectivity in the large N limit. 1=3). it is severely accelerated [22]. after that. therefore. connection is attempted with a ﬁnite number. xj xn xn ÿ z Nn (where n is an integer number) give i j rise to the same inverse square behavior (although. the necessary and sufﬁcient conditions for the ﬁtness distribution and attaching rule under which wellbehaved SF networks are generated. This leads to the nontrivial result that even nonscalefree ﬁtness distributions can generate scalefree networks (see Fig. 2). 2:5. Also different implementations of the threshold rule. of different sites. In a future publication we will explore. 3. (1). such as f xi . Poisson distribution) 2587022 and f xi . pN m). links are kept with probability p (so that. with logarithmic corrections). for example. but still it provides a new generic path to SF networks and takes into account the widespread occurrence of the Zipf ’s behavior in nature. 2587022 . Nevertheless. the N factor in Eq. we consider an exponential distribution of ﬁtnesses. we can introduce in a rather natural way an upper cutoff accounting for the fact that every site has limited information on the rest of the world and.. where x is the usual Heaviside step function.
85 y~k c(k) −1.5 y~k −3.0 y~k −2.VOLUME 89. Degree distribution for networks generated using x xÿ. N UMBER 25 10 0 PHYSICA L R EVIEW LET T ERS ρ(x)~x −2.5 16 DECEMBER 2002 10 −1 10 −2 10 P(k) 10 −4 10 10 −6 10 −5 0 1 2 3 −3 0 10 10 10 4 3 data −2 y~k 10 −1 2 10 −2 P(k) c(k).6 y~k 0 1 2 3 4 10 −5 10 0 10 1 k 10 2 10 3 10 −3 10 10 10 k 10 10 FIG.0 ρ(x)~x −3.0 ρ(x)~x −4. 2 (color online).0 y~k −4.knn 10 10 1 10 10 k 10 10 10 −3 0 10 −4 10 10 −1 −2 knn −0.
with .
Inset: The degree distribution P k. 3 (color online). 3. deﬁned as the total number of minimum paths between any couple of vertices in the network passing through vertex i [24]. This quantity gives a measure of the amount of trafﬁc passing through a vertex. To have a more extensive picture of the nature of the networks under consideration. for networks generated using x eÿx and a threshold rule. bi . the clusteringcoefﬁcient ci of a vertex i. power laws with the corresponding analytical values M are explicitly drawn in straight lines. 4. (iv) the probability distribution of the betweenness.23]: (i) the average distance hdi. and f xi . averaged over 100 realizations. and different values of . Data obtained for networks with 104 vertices. FIG. c k is obtained by averaging ci for all vertices with ﬁxed degree k. Computer simulations of our model show that networks with powerlaw distributed ﬁtnesses. More speciﬁcally. (ii) the average neighbor connectivity knn k. Results are averaged over 100 realizations of size 104 . whose degree is ki . xj xi xj =x2 . as in the aforementioned papers. measuring the average minimum number of arcs needed to connect two given sites. Average neighbor connectivity knn against k. interest in which has been triggered by recent studies on the Internet structure [10. 2:5. measuring the average degree of vertices neighbor of a kdegree vertex. (iii) the clustering coefﬁcient c k that measures the degree of interconnectivity of nearest neighbors of kdegree vertices. we have studied the following topological properties [3]. both the probability distribution P b and its ﬁrst moment hbi=N. the straight line corresponds to kÿ2 . We studied. is the ratio between the number of edges ei in the subgraph identiﬁed by its neighbors and the maximum possible number of edges in the subgraph. That is ci 2ei =ki ki ÿ 1 [2].
and m N. z 10. show nearly constant knn k and c k. P b0 > b / b1ÿ b . Email networks [25] are a good candidate to be represented by our model. The betweenness distribution. but a powerlaw behavior is found for the clustering magnitudes.45 −2. [18]: All networks with 3 > 2 can be classiﬁed in only two groups according to the value of b ( b 2 and b 2:2. we ﬁnd hdi 2. instead. when links are detected by ‘‘picking’’ them one by one. i.e. In this case growth may occur. It is worth remarking that our model having 2 is not included in the previously discussed classiﬁcation of betweenness exponents [18] (Fig. This is in good agreement with that conjectured in Ref. 4).. and hbi=N ’ 0:1. while for larger values of .5 10 −3 10 −3 10 −2 10 −1 10 0 10 1 10 2 10 3 b/N FIG. 4 (color online). let us mention the following possibility: One can imagine situations where a Poisson network is seen as SF just because the exploration method implicitly implements a probabilistic rule depending on the ﬁtnesses (this applies. and b 2:6 for 4. larger nonuniversal values of b are reported. Rather than preferential attachment there should be some intrinsic feature of the receiver playing a role in the phenomenon. just as occurs for the original BA model [2]. 3). Integrated betweenness distribution. shows an unexpected behavior. Having explored the most basic properties of the model and some particular realizations. but not if the network is 0 10 P(b’>b) 10 −1 10 −2 ρ(x)=x −3. for example. hci ’ 0:1.0 ρ(x)=x −x ρ(x)=e γb=2.. for different ﬁtness distributions.0 ρ(x)=x −4.2 γb=2. The distribution of betweenness decays as a power law with an exponent b 2:2 for 2:5 and 3. To further emphasize the utility of this new mechanism. 2587023 . let us comment now on possible applications. The exponential case behaves in a different way: For a network of size N 104 . giving a powerlaw tail with an exponent b 1:45 (see Fig.6 γb=1. respectively). hknn i / 2587023 kÿ0:85 and c k / kÿ1:6 . but agents (Email senders) do not have any access or knowledge of the degree of the receivers.
066130 (2002). [7] M. A. Phys. This is irrelevant for large enough z and N. and P. Rev. Lett. 64. Dorogovtsev and J. L. ´ [16] G. 1999). Phys. Nature (London) 410. 016131 (2001). Systems 5. and A. F. 02. and C. Sci. Rev. and A. Bornholdt and H. of the case with threshold type of attaching rule. Phys.I. In particular. betweenness. SmallWorlds (Princeton University Press. 320. Rev. J. Smith. in cases in which this connectivity is high. Zhang. A. Lett. Newman. Europhys. If only links with corresponding ﬁtnesses over threshold are ‘‘seen’’ by the exploration method then. 035103 (2002). Erdos and A.L. J. Yet. R. Renyi. R. It is a ˝ ´ natural generalization of the standard ErdosRenyi. Rev.L. Phys. A. 3200 (2001). In summary. p. M. 436 (2001). Kim. and references therein. In particular. ˝ ´ [17] P. 386 (2000). Stat. F. Vazquez. Camacho et al. P. for example. [3] S.V.. and D. R. and S. Proc. BFM20012841). Mod.I. it is reasonable to assume that the interaction strength is a function of some properties of the two proteins (such as. Krapivsky and S. [13] SF networks are generated only if p k k. Europhys. H. N UMBER 25 PHYSICA L R EVIEW LET T ERS 16 DECEMBER 2002 explored by crawling on it). Montoya. ´ [2] R. F. in Handbook of Graphs and Networks: From the Genome to the Internet. [1] S. Sole and J. a connectivity upper cutoff related to the maximum connectivity of the underlying network. PastorSatorras and A. Newman. Adv. Sole . 016132 (2001). E 63. Vespignani. condmat/0206084. Soc. Compl.A. Phys. Inst. 251 (1999). Rev. 1079 (2002). [20] M. 87. 228102 (2002). Proc.S. obviously. and A. an ˝ ´ ErdosRenyi network with exponentially distributed ﬁtnesses can be seen as SF (with. we have presented an alternative model to justify the ubiquity of SF networks in nature. and the corresponding RNA is absent. Acad. we obtain a power law plus a delta function of amplitude exp ÿz at K N. named ‘‘twohybrid’’ method: Two proteins are hybridized with two fragments of a transcription factor (a protein that binding to a gene promotes its transcription into the corresponding RNA).C. D. 2587024 2587024 . 80. PastorSatorras. condmat/0207107. Int. Lett. 278701 (2001). 98. 066123 (2001). K. Natl. The presence of the corresponding RNA signals the interaction between the two proteins. Phys. Instead. 42 (2002). The spliced promoter does not bind to the gene. and A. ´ [15] R. Faloutsos. Rev. corresponding to the fact that sites with xi above threshold are connected to every other site. 29. and T. L. Nature (London) 411. The main result is that emergence of SF properties is not necessarily linked to the ingredients of growth and preferential attachment. R. Hurtado for a critical reading of the manuscript.L. PastorSatorras. S. Kahng. 41 (2001). acknowledges ﬁnancial support from the MCYT (Grant No. This possibility has still to be checked through an analysis of the detailed physics behind the twohybrid method. E. 52. Phys. Maritan. Vespignani. condmat/0108043.L. for example. and L. R. N. Redner. Vespignani. sublinear p k produce skewed but not SF networks. 2039 (2001). Baraba si. 268 (2001). P. Lett. and clusteringcoefﬁcient distributions. Let us think. [25] H.V. Faloutsos. condmat/0205232. Flammini. [9] R. Vazquez. B. [8] G. We believe that this model is particularly suitable for situations where the degree value of nodes is not publicly available. 258701 (2001). Bianconi and A. E. Rev. We can imagine that the interaction strength between the two proteins has to be above a given threshold. Mielsch. thanks the OFESBern (Grant No. L. [23] A. [4] M. The authors thank FET Open Project IST200133555 COSIN. 54. J. Proceedings of the ACM.0234). 2741 (1998). PastorSatorras. we recover the powerlaw behavior of degree. Goh et al. this scenario could be of relevance to protein networks. The way comprehensive protein networks have been obtained until now is through a baitprey method. U. 86. 130 (1999). ´ [6] R. Nature (London) 401. Faloutsos. static structures characterized by quenched disorder (for different disorder distributions) and threshold phenomena may generate effects very similar to those measured in the real data.VOLUME 89. Vespignani. Strogatz. one can generate hubs in the ‘‘apparent’’ SF network). 87. Ebel. N. ´ [11] R. Berlin. Phys. [10] R. Jeong. Mendes. G. Watts. if the two proteins interact. 88. Rev. Phys. 51. Marchetti. Pietronero. Bull. E. 404 (2001). PastorSatorras for very useful comments and suggestions. or their accessible surface area). Marsili and Y. [21] Being more precise. New Jersey. P. Lett. N. [19] P. Baraba si. their hydrophobicity. 38. Redner. Goh. Phys. edited by S. Oltvai. they bring together the two promoter fragments allowing it to bind to the gene and transcription to start. M. Kepler. Dorogovtsev and J.I. SIGCOMM [Comput. L. Let us argue why. 74. [14] A. E 66. transcription is inhibited. Lett. Commun. P. Albert. Krapivsky and S. ´ [5] H. H. London B 268. Vazquez. Mason. Albert and A. Rev. and Z. E 65. Baraba si. Rev. Schuster (Wiley. Adv. E 64. A. and references therein. and M.. Bornholdt. Phys. Jeong. Princeton. D. We thank R. [12] D. Baraba si. else the typical promoter binding time will be too short for the RNA polymerase to bind to the gene and initiate transcription. A. 47 (2002). In turn. [24] See M. 343 (1961). 2002). F. Rev. Mendes. Caldarelli. Lett. [18] K. for example. [22] S.