PHYSICAL REVIEW E 79, 036114 共2009兲

Entropy of network ensembles

Ginestra Bianconi
The Abdus Salam International Center for Theoretical Physics, Strada Costiera 11, 34014 Trieste, Italy
共Received 20 February 2008; revised manuscript received 13 January 2009; published 27 March 2009兲
In this paper we generalize the concept of random networks to describe network ensembles with nontrivial
features by a statistical mechanics approach. This framework is able to describe undirected and directed
network ensembles as well as weighted network ensembles. These networks might have nontrivial community
structure or, in the case of networks embedded in a given space, they might have a link probability with a
nontrivial dependence on the distance between the nodes. These ensembles are characterized by their entropy,
which evaluates the cardinality of networks in the ensemble. In particular, in this paper we define and evaluate
the structural entropy, i.e., the entropy of the ensembles of undirected uncorrelated simple networks with given
degree sequence. We stress the apparent paradox that scale-free degree distributions are characterized by
having small structural entropy while they are so widely encountered in natural, social, and technological
complex systems. We propose a solution to the paradox by proving that scale-free degree distributions are the
most likely degree distribution with the corresponding value of the structural entropy. Finally, the general
framework we present in this paper is able to describe microcanonical ensembles of networks as well as
canonical or hidden-variable network ensembles with significant implications for the formulation of network-
constructing algorithms.

DOI: 10.1103/PhysRevE.79.036114 PACS number共s兲: 89.75.Hc, 89.75.Fb

I. INTRODUCTION
The quantitative measure of the order present in complex
systems and the possibility to extract information from the
complex of interactions in cellular, technological, and social
networks is a topic of key interest in modern statistical me-
chanics. The field of complex networks 关1,2兴 is having a
rapid development and great success because the concepts
coming from graph theory have wide applicability. Mainly,
the characterization of the structure of different networks al-
lows the scientific community to compare systems of very
different nature. Different statistical mechanics tools have
been devised to describe the different levels of organization
of real networks. A description of the structure of a complex
network is presently performed by measuring different quan-
tities as 共i兲 the density of the links, 共ii兲 the degree sequence
关3兴, 共iii兲 the degree-degree correlations 关4–6兴, 共iv兲 the clus-
tering coefficient 关7,8兴, 共v兲 the k-core structure 关9–11兴, 共vi兲
the community structure 关2,12–14兴, and finally 共vii兲 the na-
ture of the embedding space 关15–17兴. Moreover, significant
characteristics of the network are strength-degree correla-
tions 关18兴 if the network is weighted, and in-degree–out-
degree correlations 关19兴 if the network is directed. These
phenomenological quantities describe the local or nonlocal
topology of the network and do affect dynamical models
defined on them 关1兴.
While many different statistical mechanics models have
been proposed 关20–27兴 to describe how the power-law de-
gree distribution can arise in complex networks, little work
has been done on the problem of measuring the level of
organization and “order” in the frame of theoretical statistical
mechanics. Only recently in the field of complex networks
has attention been addressed to the study of entropy mea-
sures 关28–33兴 able to approach this problem. The entropy of
a given ensemble, giving the normalized logarithm of the
number of networks in the ensemble, has been introduced in
Ref. 关31兴. This quantity can be used to assess the role that
given structural characteristics have in shaping the network.
In fact, given a real network, a subsequent series of random-
ized networks ensembles can be considered, where each sub-
sequent ensemble shares one additional structural character-
istic with the given network. The entropy of these subsequent
network ensembles decreases as we proceed, adding con-
straints, and the difference between the entropies in two sub-
sequent ensembles quantifies how restrictive is the intro-
duced additional constraint. In the first part of this paper we
construct a general statistical mechanics framework for gen-
eralized random network ensembles which satisfy given
structural constraints. We call these ensembles “microcanoni-
cal.” We also describe how to construct canonical network
ensembles with generalized hidden variables following the
lines of the papers 关22–27兴. Subsequently we make an ac-
count of most of the network ensembles that can be formu-
lated: the ensemble of undirected networks with a given
number of links and nodes, and the ensemble of undirected
networks with given degree sequence, with given spatial em-
bedding and community structure. Some of these network
ensembles were already presented in 关31兴 and we report their
derivation here for completeness. This approach is further
extended to weighted networks and directed networks. Fi-
nally we focus our attention on the structural entropy, i.e., the
entropy of an ensemble of uncorrelated undirected simple
networks of a given degree sequence. The structural entropy
of a power-law network with constant average degree is
monotonically decreasing as the power-law exponent ␥ → 2.
This result could appear in contradiction with the wide oc-
currence of power-law degree distribution in complex net-
works. Here we show by a statistical mechanics model that
scale-free degree distributions are the most likely degree dis-
tributions with an associated small value of structural en-
tropy while Poisson degree distributions are the most likely
degree distributions of network ensembles with maximal
structural entropy. This result indicates that the scale-free

1539-3755/2009/79共3兲/036114共10兲 036114-1 ©2009 The American Physical Society

GINESTRA BIANCONI PHYSICAL REVIEW E 79, 036114 共2009兲

degree distribution emerges naturally when considering net-

work ensembles with small structural entropy and therefore a
larger amount of order.
pij = 冏 冏
⳵ ln Z
⳵hij hij=rij=0 ∀ 共i,j兲
. 共5兲

The appearance of the power-law degree distribution re- In an ensemble of weighted networks we can define also the
flects the tendency of social, technological, and especially average weight wij of a link between node i and node j given
biological networks toward “ordering.” This tendency is at by

work regardless of the mechanism that is driving their evo-
lution, which can be either off-equilibrium models with a
preferential attachment mechanism 关3兴, or a hidden-variable
mechanism 关22–27兴, or some other statistical mechanics
mechanism 关20,21兴.
II. STATISTICAL MECHANICS
OF NETWORK ENSEMBLES
A network of N labeled nodes i = 1 , 2 , . . . , N is uniquely
defined by its adjacency matrix a of matrix elements aij 艌 0
with aij ⬎ 0 if and only if there is a link between node i and
node j. Simple networks are networks without tadpoles or
double links, i.e., aii = 0 ∀ i and aij = 0 , 1. Weighted networks
describe heterogeneous interactions between the nodes. In
this paper we will consider weighted networks in which the
matrix elements aij can take different integer null or positive
values. Directed networks are described by nonsymmetric
adjacency matrices a ⫽ aT where we have indicated by aT the
transpose of the matrix a.
A structural constraint on a network can always be formu-
lated as a constraint on the adjacency matrix of the graph,
i.e.,
ជ 共a兲 = C
ជ.
F 共1兲
In order to describe microcanonical network ensembles with
given structural constraints in 关31兴 and in the following we
will use a statistical mechanics perspective. Therefore we
define a partition function Z of the ensemble as in the fol-
lowing, i.e.,
Z = 兺 ␦关F
a
ជ 共a兲 − C
ជ 兴exp
冉兺 ij
冊
hij⌰共aij兲 + rijaij ,
where we assume for simplicity that Fជ 共a兲 and aij takes only
integer values, that ␦关·兴 indicates the Kronecker delta, and
that ⌰共x兲 is a step function with ⌰共x兲 = 1 if x ⬎ 0 and ⌰共x兲
= 0 if x = 0. Moreover, in Eq. 共2兲, the auxiliary fields hij , rij
have been introduced as in classical statistical mechanics.
The entropy per node ⌺ of the network ensemble is defined
as
1
⌺= 兩ln共Z兲兩hij=rij=0 ∀ 共i,j兲 .
N III. UNDIRECTED SIMPLE NETWORKS
The marginal probability for a certain value of the element
aij of the adjacency matrix is given by
1 of generality, we can use the definition of Z given by 共2兲 with
␲ij共A兲 = 兺 ␦共aij − A兲␦„Fជ 共a兲 − Cជ ….
Z a
The probability of a link pij is given by
冏 冏
⳵ ln Z
wij = . 共6兲
⳵rij hij=rij=0 ∀ 共i,j兲
In a microcanonical network ensemble all the networks that
satisfy given structural constraints have equal probability.
Therefore the probability of a network G, described by the
adjacency matrix a, is given in the microcanonical ensemble
by
ជ 共a兲 − C
ជ ….
P M 共a兲 = e−N⌺␦„F 共7兲
If we allow for “soft” structural constraints in network en-
sembles we can describe canonical network ensembles. In a
canonical network ensemble each network a has a different
probability given by
PC共a兲 = 兿 ␲ij共aij兲. 共8兲
ij
For ensembles of simple networks Eq. 共8兲 becomes
PC共a兲 = 兿 paijij共1 − pij兲1−aij . 共9兲
ij
If the link probabilities ␲ij共aij兲 are chosen equal to Eqs. 共4兲
and 共5兲, then the structural constraints Fជ 共a兲 = C
ជ are satisfied
on average, i.e.,
ជ 共a兲典
具F ជ 共10兲
PC共a兲 = C ,
where the average 具·典 PC共a兲 is performed over the canonical
ensemble probability 共8兲. For structural constraints that do
not correspond to feasible networks 关34兴, the entropy of the
共2兲
network ensemble is ⌺ = −⬁. Although the statistical me-
chanics problem is always well defined, the calculation of
the partition function by saddle point approximation can be
performed only if the number of constraints F␣ with ␣
= 1 , . . . , M is at most extensive, i.e., M = O共N兲; moreover
phase transitions might emerge as condensation of loops 关1兴
and first-order phase transitions 关22兴. In this paper we are
going to consider only linear constraints of the adjacency
matrix. Further developments of this statistical mechanics
framework will involve perturbative approaches to solve
nonlinear structural constraints.
共3兲
In an undirected simple network the adjacency matrix el-
ements are zero or one 共aij = 0 , 1兲 and tadpoles are forbidden
共aii = 0 ∀ i兲. Therefore for undirected networks, without loss
共4兲
rij = 0, for all i , j. We can consider for these networks differ-
ent types of structural constraints. In the following a few
examples of particular interest are listed.

036114-2
ENTROPY OF NETWORK ENSEMBLES PHYSICAL REVIEW E 79, 036114 共2009兲

共i兲 The ensemble G共N , L兲 of random networks with given B共dᐉ兲 = ⌺i⬍j␹ᐉ共di,j兲aij where dij = d共rជi , rជ j兲 is the distance be-
number of nodes N and links L = ⌺i⬍jaij. In this case the tween nodes i and j in the embedding space and the charac-
statistical mechanics approach gives a “physics” derivation teristic function ␹ᐉ共x兲 = 1 if x 苸 关dᐉ , dᐉ + 共⌬d兲ᐉ兴 and ␹d共x兲 = 0
of the G共N , L兲 random ensemble. In this case we have the otherwise. In this case the structural constraints can be ex-
structural constraint pressed as
ជ = 兺 a − L = 0.
Fជ 共a兲 − C ij 共11兲 F␣共a兲 − C␣ = 兺 a␣ j − k␣ = 0 共17兲
i⬍j j

共ii兲 The configuration model, i.e., the ensemble of net- for ␣ = 1 , . . . , N and
works with given degree sequence 兵k1 , . . . , kN其 with ki
= ⌺ jaij. In this case the structural constraints are given by F␣共a兲 − C␣ = 兺 ␹ᐉ共di,j兲aij − B共dᐉ兲 共18兲
i⬍j
F␣共a兲 − C␣ = 兺 a␣ j − k␣ = 0 共12兲 for ␣ = N + 1 , . . . , N + ⌳.
j

for ␣ = 1 , . . . , N. A. The entropy of the G(N , L) ensemble

共iii兲 The network with given degree sequence 兵k1 , . . . , kN其
The networks in the G共N , L兲 ensemble have given number
and given average nearest neighbor connectivity kNN共k兲
of nodes N and links L. The entropy of this ensemble is given
= 关⌺i,j␦共ki − k兲aijk j兴 / 共kNk兲 of nodes of degree k 共with Nk in-
by the logarithm of the binomial
dicating the number of nodes of degree k in the network兲. In

冤冢 冣冥
this case the structural constraints are given by N共N − 1兲
F␣共a兲 − C␣ = 兺 a␣ j − k␣ = 0 共13兲 N⌺0 = ln 2 共19兲
j L
for ␣ = 1 , . . . , N and 共we always assume distinguishable nodes in the networks
关29兴兲. The probability pij of a given link 共i , j兲 is given by
F␣共a兲 − C␣ = 兺 ␦共ki − k兲aijk j − kNkkNN共k兲 共14兲 p共0兲
ij = L / 关N共N − 1兲 / 2兴 for every pair of nodes i , j. The en-
ij
semble G共N , p兲 is the canonical ensemble corresponding to
for ␣ = N + 1 , . . . , N + K, with K indicating the maximal con- the microcanonical G共N , L兲 ensemble.
nectivity of the network.
共iv兲 The network ensemble with given degree sequence B. The entropy of the configuration ensemble
and given community structure. In this case we assume that
In the configuration ensemble we consider all the net-
each node is assigned a community 兵q1 , . . . , qN其 and we fix
works with given degree sequence. Using 共2兲 and 共12兲 the
the number of links between nodes of different communities
partition function of the ensemble can be explicitly written as
A共q , q⬘兲 = ⌺i⬍j␦共qij − q兲␦共qij − q⬘兲aij with qij = min共qi , q j兲 and
qij = max共qi , q j兲. In this case the structural constraints are
given by
Z1 = 兺 兿 ␦冉ki − 兺j aij冊exp冉 兺
兵aij其 i i⬍j
hijaij冊 . 共20兲

F␣共a兲 − C␣ = 兺 a␣ j − k␣ = 0 共15兲 Expressing the ␦ functions in the integral form with Lagrang-
j ian multipliers ␻i for every i = 1 , . . . , N, we get
for ␣ = 1 , . . . , N and

F␣共a兲 − C␣ = 兺 ␦共qij − q兲␦共qij − q⬘兲aij − A共q,q⬘兲 共16兲

Z1 = 冕 冉
D␻ exp − 兺 ␻iki
i
冊兿 i⬍j
共1 + e␻i+␻ j+hij兲, 共21兲

i⬍j
where D␻ = ⌸id␻i / 共2␲兲. We solve this integral by saddle
for ␣ = N + 1 , . . . , N + Q共Q + 1兲 / 2 with Q equal to the number point equations accounting also for second-order terms of the
of different features of the nodes. Here and in the following expansion. The entropy of this ensemble of networks can be
in order to have an extensive number of constraints we as- approximated in the large-network limit N Ⰷ 1 with
sume Q = O共N1/2兲.
1
1 ⯝ − 兺 ␻i ki + 兺 ln共1 + e 兺 ln共2␲␣i兲
쐓 쐓
共v兲 The ensemble of networks with given degree sequence N⌺und * ␻i +␻ j
兲−
and dependence of the link probability on the distance of i i⬍j 2 i
the nodes in an embedding geometrical space. In this en-
共22兲
semble we consider a fixed spatial distribution of nodes in
space 兵rជ1 , . . . , rជN其 and we consider all the networks compat- with 1 = ln共Z1兲
N⌺und and the Lagrangian multipliers ␻i satis-
ible with the given degree sequence and the number of links fying the saddle point equations
linking nodes in a given distance interval. In this case 쐓 쐓
we take ⌳ distance intervals Iᐉ = 关dᐉ , dᐉ + 共⌬d兲ᐉ兴 with ᐉ e ␻i +␻ j
ki = 兺 , 共23兲
= 1 , . . . , ⌳, and we fix the number of links linking nodes in a j⫽i
쐓
1 + e ␻i +␻ j
쐓

given distance interval. The structural constraints that we

consider are therefore provide by the values of the vector and the coefficients ␣i approximated with

036114-3
GINESTRA BIANCONI PHYSICAL REVIEW E 79, 036114 共2009兲

쐓 쐓
e ␻i +␻ j the following expression in the large-N limit, i.e.,
␣i ⯝ 兺 쐓 쐓 . 共24兲 2典/具k典兲2/4
j 共1 + e␻i +␻ j 兲2 共2L − 1兲!!e−共具k
Nuncorr ⬀ . 共29兲
兿 i k i!
c
The probability of a link i , j in this ensemble is given by
쐓 쐓
e ␻i +␻ j The factor 共2L − 1兲!! accounts for the total number of wirings
p共1兲 = . 共25兲 of the links. In fact if we want to construct a network, given
ij ␻i쐓+␻쐓j
1+e a certain distribution of half edges through the N nodes of the
In this ensemble pij ⫽ f共␻i兲f共␻ j兲; consequently the model re- network, as a first step we take a half edge and we match it
tains some “natural” correlations 关27兴 given by the degree with one of the 2L − 1 other half edges of the network. Sec-
sequence. In fact these are nothing else than the correlations ond, we match a new half edge with one of the 2L − 3 re-
of the configuration model 关35兴. maining half edges. Repeating this procedure, we get one out
The canonical model corresponding to the configuration of 共2L − 1兲!! possible wirings of the links. This number in-
model is then a hidden-variable model where each node i is cludes also the wiring of the links which gives rise to net-
assigned a hidden variable ␻i and the probability for each works with double links. To estimate the number of such
link follows 共25兲. Similar expressions were already derived undesired wirings we assume that the network is random,
in different papers 关22,25,27兴 but with a different interpreta- i.e., that the probability that a node with ki half edges con-
tion. Here the hidden variables ␻i are simply fixing the av- nects to a node with k j half edges is a Poisson variable with
erage degree of each node. Moreover we note that the deri- average kik j / 共具k典N兲. In this hypothesis the probability ⌸ that
vation of 关26兴 guarantees that in the canonical model the the network does not contain double links is equal to 关36兴
connectivity of each node is distributed according to a Pois-
son distribution with average ⌺ j pij.
The form of the probability pij is such that, when inferring
⌸=兿 1+
i⬍j
冉 具k典N
冊
kik j −k k /具k典N
e ij
2 2
⬃ e−共具k 典/具k典兲 /4 . 共30兲

the values of the hidden variables ␻i for a canonical network

Finally in the expression 共29兲 for Nc there is an additional
in this ensemble by maximum likelihood methods, we obtain
term which takes into account the number of wirings of the
␻i⬘ = ␻i in the large-network limit 关27兴.
links giving rise to equivalent networks without double links.
The case in which there is a structural cutoff in the net-
work ki ⬍ 冑具k典N is of particular interest. In this case we can
This term is given by the number of possible permutations of
the half edges at each node, i.e., ⌸iki!. We note here that a
approximate Eq. 共23兲 by e␻i ⯝ ki / 冑具k典N, ␣i ⯝ ki. In this limit
쐓
similar result was derived by mathematicians for the case in
the network is uncorrelated. and the probabilities of a link which the maximal connectivity K ⬍ N1/3 关37兴 and an in-
are given by p共1兲,uncorr
ij = kik j / 共具k典N兲, since ␻i* ⬍ 0. We call the equality was proved for the case K ⬎ N1/3 关38兴. Now we ex-
entropy of these uncorrelated ensembles the structural en- tend these results by statistical mechanics methods to uncor-
tropy ⌺S, and we can evaluate it from the explicit expression related networks with maximal connectivity K ⬍ 冑具k典N.

N⌺S ⯝ − 兺 ln关ki/冑具k典N兴ki −
1 1 k ik j
i
兺 ln共2␲ki兲 + 2 兺ij 具k典N
2 i
C. The entropy of network ensembles
with fixed degree correlations

1 k2i k2j We consider now network ensembles, with given degree

−兺 + ¯ = − 兺 共ln ki − 1兲ki correlations and given average degree of neighboring nodes,
4 共具k典N兲2
ij i that satisfy the constraints defined in Eqs. 共13兲 and 共14兲. We

−
1
兺
2 i
1
ln共2␲ki兲 + 具k典N关ln共具k典N兲 − 1兴 −
2
1 具k2典
4 具k典
冉 冊 2 can proceed to the evaluation of the probability of a link p共2兲
and the calculation of the entropy of the ensemble as in the
configuration model. In this case we have to introduce the
ij

+ ¯ . 共26兲 Lagrangian multipliers ␻i fixing the degree of node i and the

Lagrangian multipliers Ak fixing the average degree of nodes
Expression 共26兲 gives for the number of networks NS in the of degree k.
ensemble The partition function of this ensemble can be evaluated

Nuncorr
S ⯝
共具k典N兲!!
兿 i k i!
exp −冋 冉 冊1 具k2典
4 具k典
2
册
+ O共ln N兲 . 共27兲
at the saddle point, giving for the entropy of the ensemble, in
the thermodynamic limit,

2 ⯝ − 兺 ␻i ki − 兺 Ak knn共k兲kNk
N⌺und * *
From combinatorial arguments we can derive an expres- i k
sion NCuncorr for the number of uncorrelated networks with a
1
+ 兺 ln共1 + e␻i +␻ j +kiAk j+k jAki兲 − 兺 ln共2␲␣i兲
* * * *
given degree sequence which agrees with the above estimate
共27兲 in the limit N Ⰷ 1, i.e., i⬍j 2 i
1
ln Nuncorr
c = ln Nuncorr
S + O共ln N兲. 共28兲 − 兺 ln共2␲␣k兲
2 k
共31兲
In fact by combinatorial arguments we can show that the
number of networks with given degree sequence is given by where ␻쐓i and A쐓k satisfy the saddle point equations

036114-4
ENTROPY OF NETWORK ENSEMBLES PHYSICAL REVIEW E 79, 036114 共2009兲

쐓 쐓共q,q 兲
* *
e␻i +␻ j +k jAki+kiAk j
*
e␻i +␻ j+w ⬘
ki = 兺 , A共q,q⬘兲 = 兺 ␦共qគ ij − q兲␦共q̄ij − q⬘兲 쐓 쐓 쐓共q,q 兲
,
j⫽i
*
1 + e␻i +␻ j +k jAki+kiAk j
* * *
i⬍j 1 + e␻i +␻ j +w ⬘

共37兲
␻*+␻*+k jA* +kiA*
1 e i j
兺 ␦共ki − k兲 兺 and with ␣i , ␣q,q⬘ that can be approximated by
ki kj
kNN共k兲 = kj ,
kNk i ␻*+␻*+k jA* +kiA*
j⫽i 1 + e i j ki kj 쐓 쐓 쐓共q ,q̄ 兲
e␻i +␻ j +w គ ij ij
共32兲 ␣i ⯝ 兺 쐓 쐓 쐓共q ,q̄ 兲 2
,
j 共1 + e␻i +␻ j +w គ ij ij 兲
and where ␣i , ␣k are approximately equal to the expressions
쐓 쐓 쐓共q,q 兲
* * * *
e␻i +␻ j +w ⬘
e␻i +␻ j +k jAki+kiAk j
␣i ⯝ 兺 , ␣q,q⬘ ⯝ 兺 ␦共qគ ij − q兲␦共q̄ij − q⬘兲 쐓 쐓 쐓共q,q 兲 2
.
* * * *
共1 + e␻i +␻ j +w ⬘ 兲
j 共1 + e␻i +␻ j +k jAki+kiAk j兲2 i⬍j

共38兲
* * * *
e␻i +␻ j +k jAki+kiAk j p共c兲
␣k ⯝ 兺 ␦共ki − k兲 兺 k2j . 共33兲 In this ensemble the probability for a link ij between a
␻*+␻*+k jA* +kiA* 2 node i and a node j is equal to
i j⫽i 共1 + e i j ki kj 兲
쐓 쐓 쐓共q ,q̄ 兲
The probability p共2兲 of the link 共i , j兲 in this ensemble is given e␻i +␻ j +w គ ij ij
by
ij
p共c兲
ij = 쐓 쐓 쐓共q ,q̄ 兲
. 共39兲
1 + e␻i +␻ j +w គ ij ij
* * *
e␻i +␻ j +k jAki+kiAk j Assigning each node a hidden variable ␪i and to each pair of
p共2兲
ij = . 共34兲 communities the symmetric matrix V共q , q⬘兲, we can con-
␻*+␻*+k jA* +kiA*
1+e i j ki kj
struct the hidden variable or canonical ensemble by extract-
This formula generalizes the hidden-variable formula of the ing each link with probability
configuration model to networks with strong degree-degree
␪i␪ jV共qi,q j兲
correlations. In particular, in order to build a canonical net- pij = . 共40兲
work with strong degree-degree correlation we can consider 1 + ␪i␪ jV共qi,q j兲
nodes with hidden variables ␪i and G␪ and a probability pij to
have a link between a node i and a node j given by
E. The entropy of network ensembles
␪i␪ j共G␪i兲␪ j共G␪ j兲␪i with given distance between the nodes
pij = . 共35兲
1 + ␪i␪ j共G␪i兲␪ j共G␪ j兲␪i Finally, we consider the ensemble of undirected networks
living in a generic embedding space and with structural con-
straints described by Eqs. 共17兲 and 共18兲. Following the same
D. The entropy of network ensembles with given degree steps as in the previous cases, we find that the entropy for
sequence and given community structure such an ensemble in the large network limit is given by
⌳
The partition function of network ensembles with given
degree sequence 共15兲 and given community structure 共16兲 N⌺d ⯝ − 兺 ki␻쐓i − 兺 B共dᐉ兲g쐓共dᐉ兲
can be evaluated by saddle point approximation in the large- i ᐉ=1

network limit as long as Q = O共N1/2兲. Following the same

steps as in the previous case we find that the entropy for this 冋 冉
+ 兺 ln 1 + exp ␻쐓i + ␻쐓j + 兺 ␹ᐉ共dij兲g쐓共dᐉ兲
i⬍j ᐉ
冊册
ensemble is given by ⌳
1 1
N⌺c ⯝ − 兺 ki␻쐓i − − 兺 ln共2␲␣i兲 − 兺 ln共2␲␣ᐉ兲 共41兲
兺 A共q,q⬘兲w共q,q⬘兲쐓 2 i 2 ᐉ=1
i q艋q⬘
with the Lagrangian multipliers 兵␻쐓i 其 , 兵g쐓d 其 satisfying the
1
+ 兺 ln共1 + e ␻쐓i +␻쐓j +w쐓共qគ ij,q̄ij兲
兲 − 兺 ln共2␲␣i兲 saddle point equations

冉 冊
i⬍j 2 i
exp ␻쐓i + ␻쐓j + 兺 ␹ᐉ共dij兲g쐓共dᐉ兲
1
− 兺 ln共2␲␣q,q⬘兲 共36兲 ki = 兺
ᐉ

冉 冊
,
1 + exp ␻쐓i + ␻쐓j + 兺 ␹ᐉ共dij兲g쐓共dᐉ兲
2 q⬍q
⬘ j⫽i
ᐉ
쐓
with the Lagrangian multipliers 兵␻쐓i 其 , 兵wq,q ⬘
其 satisfying the
saddle point equations 쐓 쐓共d 兲
e␻i +␻ j+g ᐉ

쐓 쐓 쐓共q ,q̄ 兲
B共dᐉ兲 = 兺 ␹ᐉ共dij兲 쐓 쐓 쐓共d 兲
, 共42兲
e␻i +␻ j +w គ ij ij 1 + e␻i +␻ j +g ᐉ
ki = 兺
i⬍j
쐓 쐓 쐓共q ,q̄ 兲
,
j⫽i 1 + e␻i +␻ j +w គ ij ij and the variables ␣i , ␣q,q⬘ approximated by the expressions

036114-5
GINESTRA BIANCONI PHYSICAL REVIEW E 79, 036114 共2009兲

冉
exp ␻쐓i + ␻쐓j + 兺 ␹ᐉ共dij兲g쐓共dᐉ兲
ᐉ
冊 semble the structural constraints are given by

␣i ⯝ 兺
j
冋 1 + exp 冉 ␻쐓i + ␻쐓j + 兺 ␹ᐉ共dij兲g쐓共dᐉ兲
ᐉ
冊册 2, F共a兲␣ − C␣ = 兺 ⌰共a␣ j兲 − k␣ = 0
j
共49兲

for ␣ = 1 , . . . , N and
␻쐓i +␻쐓j +g쐓共dᐉ兲
e
␣ᐉ ⯝ 兺 ␹d共dij兲 쐓 쐓 쐓共d 兲 2
. 共43兲 F共a兲␣ − C␣ = 兺 a␣ j − s␣ = 0 共50兲
i,j 共1 + e␻i +␻ j +g ᐉ 兲 j

The probability for a link between node i and j is equal to for ␣ = N + 1 , . . . , 2N.
␻i쐓+␻쐓j +g쐓共dᐉ兲
e
ij = 兺 ␹ᐉ共dij兲
p共d兲 쐓 쐓 쐓d
. 共44兲 A. The entropy of weighted network ensembles
ᐉ 1 + e␻i +␻ j +g ᐉ with given total strength S
Therefore the hidden-variable model associated with this en- The entropy of this ensemble is given by
semble corresponds to a model where we fix the hidden vari-

冤冢 冣冥
ables ␪i and rជi and and the vector W共dᐉ兲 is given and we N共N − 1兲
+S
draw a link between node i and node j according to 2
N⌺W
1 = ln .
␪i␪ jW共dᐉ兲 N共N − 1兲
pij = 兺 ␹ᐉ共dij兲 . 共45兲 2
ᐉ 1 + ␪i␪ jW共dᐉ兲
The average value of the weight of the link from i to j is
IV. WEIGHTED NETWORKS given by
Many networks have, in addition to a nontrivial topologi- S
cal structure, weighted links. We will assume in this paper wij = 具aij典W
1 = 共51兲
N共N − 1兲
that the weight of a link can take only integer values and
consequently a link between a node i and node j is charac- 2
terized by an integer number aij 艌 1. This is not a very strin-
and the probability of a link between node i and j is equal to
gent constraint since any finite network with weights of the
links taking rational numbers can be easily reduced to a net- S
work of integer weights. In a weighted network the degree ki pW,1
ij = . 共52兲
and the strength si of the node i are defined as N共N − 1兲
S+
2
ki = 兺 ⌰共aij兲,
j⫽i Therefore the simple networks with adjacency matrix 共共Aij兲兲
that can be constructed starting from the weighted networks
si = 兺 aij , 共46兲 with adjacency matrix 共共aij兲兲, by putting Aij = ⌰共aij兲 ∀ i , j,
j⫽i are uncorrelated. The canonical ensemble is given by Eq. 共8兲
with
where ⌰共x兲 = 0 if x = 0 and ⌰共x兲 = 1 is x ⬎ 0. It is possible to
define series of weighted networks by considering networks e␻aij
with fixed total strength, with given strength sequence, with ␲ij共aij兲 = 共53兲
1 − e␻
given strength and degree sequence and proceeding by add-
ing additional features as in the unweighted case. Here and in and ␻ = −ln关1 + N共N − 1兲 / 共2S兲兴.
the following we study the most relevant cases.
共i兲 The network ensemble with given total strength S. The B. The entropy of weighted network ensembles
structural constraint in this case is equal to with given strength sequence
F共a兲 − C = 兺 aij − S = 0. 共47兲 To calculate the entropy of undirected networks with a
i⬍j given strength sequence of degrees 兵si其 we proceed by the
共ii兲 The network ensemble with given strength sequence saddle point approximation as in previous cases. We find that
s1 , . . . , sN. The structural constraints are for this ensemble the entropy of this ensemble of networks is given by
given by
1
1 ⯝ − 兺 ␻i si − 兺 ln共1 − e 兺 ln共2␲␭i兲
쐓 쐓 쐓
␻i +␻ j
N⌺W 兲−
F共a兲␣ − C␣ = 兺 a␣ j − s␣ = 0, 共48兲 i i⬍j 2 i
j
共54兲
for ␣ = 1 , . . . , N.
共iii兲 The network ensemble with given degree sequence with the Lagrangian multipliers ␻쐓i satisfying the saddle
兵k1 , . . . , kn其 and strength sequence 兵s1 , . . . , sN其. For this en- point equations

036114-6
ENTROPY OF NETWORK ENSEMBLES PHYSICAL REVIEW E 79, 036114 共2009兲

쐓 쐓 쐓 쐓 쐓 쐓
e ␻i +␻ j e−共␻i +␻ j 兲+共␺i +␺ j 兲
si = 兺 쐓 쐓 共55兲 具aij典W
2 = 쐓 쐓 쐓 쐓 쐓 쐓 共64兲
j⫽i 1 − e ␻i +␻ j 共esii +␺ j + e−共␻i +␻ j 兲 − 1兲共e−␻i −␻ j − 1兲
and with ␭i being the eigenvectors of the Jacobian of the and the probability of a link between node i and j is equal to
function 쐓 쐓
e ␺i +␺ j
pW,2 = 共65兲
F = 兺 ln共1 − e−␻i−␻ j兲. 共56兲 ij 쐓 쐓
e␺i +␺ j + e−共␻i +␻ j 兲 − 1
쐓 쐓

i⬍j
The canonical ensemble 共8兲 in this case can be constructed
The average value of the weight of the link from i to j is
by assigning to every possible link 共i , j兲 the weight aij with
given by
the probability
쐓 쐓
e ␻i +␻ j 쐓 쐓
e共␺i +␺ j 兲⌰共aij兲e共␻i +␻ j 兲aij
쐓 쐓
具aij典W
1 = . 共57兲 ␲ij共aij兲 = . 共66兲
␻i쐓+␻쐓j
1−e ␺i쐓+␺쐓j 1
1+e 쐓 쐓
and the probability of a link between node i and j is equal to e −␻i −␻ j − 1
쐓 쐓
␻i +␻ j
pW,1
ij = e . 共58兲 V. DIRECTED NETWORKS
The canonical ensemble 共8兲 in this case can be constructed An undirected network is determined by a symmetric ad-
by assigning to every possible link 共i , j兲 the weight aij with jacency matrix, while the matrix of a directed network is in
the probability general nonsymmetric. Consequently the degrees of freedom
쐓 쐓 of a directed network are more than the degrees of freedom
e共␻i +␻ j 兲aij of an undirected network. In the following we only consider
␲ij共aij兲 = 쐓 쐓 . 共59兲
1 − e ␻i +␻ j the following network ensembles.
共i兲 Network ensemble with fixed total number of directed
links. The structural constraint in this case is equal to
F共a兲 − C = 兺 aij − S = 0.
C. The entropy of weighted network ensembles
共67兲
with given strength and degree sequence ij
The entropy of weighted networks with a given strength
共ii兲 Network ensemble with given directed degree se-
and degree sequence 兵si , ki其 in the large-size network limit is
quence 兵kin1 , k1 , . . . , kN , kN 其. The structural constraints in
out in out
given by
this case are
2 ⯝ − 兺 ␻i si − 兺 ␺i ki − 兺
쐓 쐓
N⌺W F共a兲␣ − C␣ = 兺 a␣ j − k␣out = 0 共68兲
i i i

冉 冊
j
1
+ 兺 ln 1 + e␺i +␺ j
쐓 쐓
for ␣ = 1 , . . . , N and
−␻i쐓−␻쐓j
e −1
F共a兲␣ − C␣ = 兺 a j␣ − k␣in = 0
i⬍j
共69兲
1
+ 兺 2N 兺 ln共2␲␭ᐉ兲
j
共60兲
2 ᐉ=1 i for ␣ = N + 1 , . . . , 2N.
with the Lagrangian multipliers satisfying the saddle point A. The entropy of directed network ensembles
equations with fixed number of directed links
쐓 쐓 If we consider the number of directed networks Ndir
0 with
e ␺i +␺ j
ki = 兺 쐓 쐓 쐓 쐓 , 共61兲 given number of nodes and fixed number of directed links
j⫽i e␺i +␺ j + e−共␻i +␻ j 兲 − 1 we find

si = 兺 쐓 쐓
쐓 쐓
e−共␻i +␻ j 兲+共␺i +␺ j 兲
쐓 쐓
쐓 쐓

쐓 쐓 , 共62兲
N⌺dir
0 = ln 冋冉 N共N − 1兲
Ldir
冊册 . 共70兲

j⫽i 共esii +␺ j + e−共␻i +␻ j 兲 − 1兲共e−␻i −␻ j − 1兲 In this case the probability of a directed link is given by
and with ␭ᐉ being the eigenvectors of the Jacobian of the L
function pij = . 共71兲
N共N − 1兲

F = 兺 ln 1 + e␺i+␺ j
i⬍j
冉 e −␻i−␻ j
1
−1
冊 共63兲 B. The entropy of directed network ensembles
with given degree sequence
calculated at the values 兵␻쐓i , ␺쐓i 其. The average weight of the To calculate the entropy of directed networks with a given
link 共ij兲 is given by degree sequence of in- and out-degrees 兵kout i , ki 其 we just
in

036114-7
GINESTRA BIANCONI PHYSICAL REVIEW E 79, 036114 共2009兲

have to impose the constraints on the incoming and outgoing quence ⌺und1 given by Eq. 共22兲 decreases with the value of
connectivity, the power-law exponent ␥ when we compare network en-

␦冉 kout
i − 兺 aij冊 兿 ␦
sembles with the same average degree 关31兴. Therefore scale-
Zdir
1 = 兺 兿
兵a 其 i j i
free networks have much smaller entropy than homogeneous
ij networks. This fact seems to be in contrast with the fact that
冉i − 兺 a ji exp
⫻ kin
j
冊 冉兺 冊 ij
hi,jaij . 共72兲 scale-free networks are the underlying structure of a large
class of complex systems. The apparent paradox can be eas-
ily resolved if we consider that many networks are the result
Following the same approach as for the undirected case, we of a nonequilibrium dynamics. Therefore they do not have to
find that the entropy of this ensemble of networks is given by satisfy the maximum entropy principle. Nevertheless, in or-
der to give more insight and comment on the universal oc-
1 ⯝ − 兺 ␻i ki − 兺 ki ␻
ˆ i + 兺 ln共1 + e␻i +␻ˆ j 兲
쐓 out in 쐓 쐓 쐓
N⌺dir currence of power-law networks in this section we derive the
i i i⫽j
most likely degree distribution of given structural entropy
1 when the total number of nodes and links is kept fixed. By
− 兺 ln关共2␲兲2␣ini ␣out
2 i i 兴 共73兲 structural entropy we define the entropy ⌺S given by Eq. 共26兲
of uncorrelated networks with fixed degree distribution. In
with the Lagrangian multipliers satisfying the saddle point particular we construct a statistical model very closely re-
equations lated to the urn or “ball in the box” models 关21,39兴.
쐓 쐓 We consider degree distributions 兵Nk其 = ⌺i␦共k − ki兲 which
e␻i +␻ˆ j
kout
i = 兺 쐓 쐓 ,
arise from the random distribution of the 2L half edges
j⫽i 1 + e␻i +␻ˆ j through the N nodes of the network. The number of ways
N兵Nk其 in which we can distribute the 共2L兲 half edges in order
쐓 쐓 to have an 兵Nk其 degree distribution is
e␻ j +␻ˆ i
i =兺
kin 쐓 쐓 , 共74兲 共2L兲!
j⫽i 1 + e␻ j +␻ˆ i N兵Nk其 = . 共78兲
with
兿k 共kNk兲!
쐓 쐓
e␻i +␻ˆ j We want to find the most likely degree distribution that cor-
i ⯝兺
␣out
␻i쐓+␻ˆ 쐓j 2
, responds to a given value of the structural entropy.
j⫽i 共1 + e 兲 Using the tools of statistical mechanics, we define a nor-
malized partition function Z as
쐓 쐓
e␻ j +␻ˆ i
␣in
i ⯝ 兺 쐓 쐓 . 共75兲
Z=
1
兺 ⬘N兵Nk其e␤N⌺S共兵Nk其兲 共79兲
j⫽i 共1 + e␻ j +␻ˆ i 兲2 C 兵Nk其
The probability for a directed link from i to j is given by with C = 共2L兲!exp关␤共2L兲!!兴. The role of the parameter ␤ in
␻쐓i +␻ˆ 쐓j Eq. 共79兲 is to fix the average value of the structural entropy
e
p1,dir
ij = 쐓 쐓 . 共76兲 ⌺S. When ␤ → ⬁ the structural entropy ⌺S is maximized and
1 + e␻i +␻ˆ j when ␤ → ␤min = 1 the structural entropy ⌺S is minimized.
In Eq. 共79兲 the sum ⌺⬘ over the 兵Nk其 distributions is ex-
If ␻i + ␻
ˆ j ⬍ 0 ∀ i, j = 1 , . . . , N, the directed network becomes
out in 冑 tended only to 兵Nk其 for which the total number of nodes N
uncorrelated and we have p1,dir ij = ki k j / 具kin典N. Given this and the total number of links L in the network is fixed, i.e.,
solution the condition for having uncorrelated directed net-
works is that the maximal in-degree Kin and the maximal
out-degree Kout should satisfy KinKout / 冑具kin典N ⬍ 1. The en-
兺k Nk = N,
tropy of the directed uncorrelated network is then given by
1 具k2in典 具kout
2
典 兺k kNk = 2L. 共80兲
uncorr
N⌺1,dir ⯝ ln共具kin典N兲! − 兺 ln共kin out
i !ki !兲 − ,
i 2 具kin典 具kout典
To enforce these conditions we introduce in Eq. 共79兲 the ␦
共77兲 functions in the integral form, providing the expression
which has a clear combinatorial interpretation as it happens
also for the undirected case. Z=
1
共2L兲!
冕 冕 冕 冕
d␭
2␲
dS
d␮
2␲
d␯
冋
兺 exp − ␤ 兺k Nk ln k!
2␲ 兵Nk其
VI. NATURAL DEGREE DISTRIBUTION
CORRESPONDING TO A GIVEN − 冉 冊
␤ S
4 具k典
2

冉
− 兺 ln关共kNk兲!兴 − i␭ 2L − 兺 Nkk
k k
冊
冊册
STRUCTURAL ENTROPY

For power-law networks with power-law exponent ␥

苸 共2 , 3兲 the entropy of the networks with fixed degree se-
冉 k
冊 冉
− i␮ N − 兺 Nk − i␯ NS − 兺 k2Nk
k
. 共81兲

036114-8
ENTROPY OF NETWORK ENSEMBLES PHYSICAL REVIEW E 79, 036114 共2009兲

This last expression can be further simplified as in the fol-

lowing, i.e., Zk共L,ᐉ,N兲 = 冕 冕 冕 冕
dS
d␭
2␲
d␮
2␲
d␯
2␲
exp关Nf k共␭, ␮, ␯,S,ᐉ兲兴

Z= 冕 冕 冕 冕
dS
d␭
2␲
d␮
2␲
d␯
2␲ and
共87兲

冋 ␤ S
冉 冊 2
f k共␭, ␮, ␯,ᐉ兲 = − i共具k典 − ᐉ/N兲␭ − i␮关1 − ᐉ/共kN兲兴

再 冉
⫻exp − i␭2L − i␮N − i␯NS −

冠兺
4 具k典

册
1
+ ln exp sNs i␭ + i␮/s + i␯s
共sNs兲!
+ 兺 ln Gk共␭, ␮, ␯兲

冊冎冡
s⫽k

=
k

冕 冕 冕 冕
dS
d␭ d␮ d␯
exp关Nf共␭, ␮, ␯,S兲兴, 共82兲
−
␤
s
ln共s!兲 − i␯共S − kᐉ/N兲 − 冉 冊
␤ S2
2 具k典
2
1
N
.

2␲ 2␲ 2␲ 共88兲
where If we develop Eq. 共86兲 for ᐉ Ⰶ L and we use the Stirling

冋 冉 冊册
approximation for factorials, we get the result that each vari-
1 ␮ ␤ able Lk is a Poisson variable with mean 具Lk典 satisfying
Gk共␭, ␮, ␯兲 = 兺 kNk i␭ + i + i␯k − ln共k!兲 .
Nk 共kNk兲! k k
具Lk典
共83兲 = 具Nk典 ⯝ k−␤−1ei␭+␤+i␮/k+i␯k , 共89兲
k
Assuming that the sum over all Nk can be approximated by where we assume that the minimal connectivity of the net-
the sum over all Lk = kNk = 1 , 2 , . . . , ⬁ we get ln Gk共␭ , ␮ , ␯兲 work is k ⬎ 0. The average 具Nk典 is a power-law distribution
= exp关i␭ + i␮ / k − 共␤ / k兲ln共k!兲 + i␯k兴 and with lower and upper effective cutoffs −i␮ and 1 / 共i␯兲 fixing

冉 冊
the average degree 具k典, with the Lagrangian parameter ␭ fix-
␤ S 2
ing the normalization constant, and finally with ␤ fixing the
f共␭, ␮, ␯,S兲 = − i具k典␭ − i␮ − i␯S −
4 具k典 structural entropy. The distribution of P共Nk兲 is given by
1
+ 兺 ei␭+i␮/k−共␤/k兲ln共k!兲+i␯k
N k
共84兲 P共Nk兲 =
k
共kNk兲!
2
e−␤Nk ln共k!兲+i␭kNk+i␮Nk+i␯共k兲 . 共90兲

where 具k典 = 2L / N indicates the average degree of the net- In the limit ␤ → ⬁ Eq. 共90兲 is extremely peaked around
work. By evaluating 共82兲 at the saddle point, deriving the the average degree k ⯝ k쐓 = O共具k典兲 of the network and the
argument of the exponential respect to ␭ and ␯, we obtain degree distribution Nk decays at large value of Nk as a Pois-
son distribution, i.e.,
1 1 i␭+i␮/k−共␤/k兲ln共k!兲+i␯k
1= 兺
N k k
e , P共Nk兲 ⯝
1 쐓 쐓 쐓 쐓 쐓
ekNk关−␤ ln共k !兲/k +i␭k +i␮/k +i␯k 兴 . 共91兲
共kNk兲!
Therefore for ␤ → ⬁ the network is Poisson-like. In the
1
具k典 = 兺 ei␭+i␮/k−共␤/k兲ln共k!兲+i␯k ,
N k
opposite limit of small structural entropy and ␤ small the
P共Nk兲 distribution 共90兲 develops a fat tail decaying like a
power law 共89兲 with an exponent ␥ = ␤ + 1. Therefore the
natural distribution with a small value of the structural en-
1
S= 兺 k2ei␭+i␮/k−共␤/k兲ln共k!兲+i␯k ,
N k
tropies is decaying as a power law and smaller values of the
power-law exponent correspond to a smaller value of the
structural entropy. When the value of the entropy is minimal,
␤ → 1 the degree distribution 共91兲 has a large tail with an
S
i␯N = − ␤ . 共85兲 exponent ␥ → 2.
2具k典2

These equations always have a solution for sparse networks VII. CONCLUSIONS
with L = O共N兲 provided that ␤ ⬎ 1 and 具k典 ⬎ 1. The marginal
In conclusion we have shown that there is a wide set of
probability that Lk = kNk is given by
network ensembles that can be naturally described by statis-
tical mechanics methods. We have provided the theoretical
1 2 Z k共L,kN k,N兲
P共Lk = kNk兲 = e−␤Nk关ln共k!兲+i␭k+i␮+i␯k 兴 , evaluation of the entropy of these ensembles, quantifying the
共kNk兲! Z共L兲 cardinality of the network ensembles. We believe that the
共86兲 entropy of randomized ensembles constructed from a given
real network will have wide applications for inference prob-
with lems defined on technological social and biological net-

036114-9
GINESTRA BIANCONI
works. In this paper we have focused on some theoretical
problems that can be approached with the use of this quan-
tity. First we have formulated a series of canonical or hidden-
variable models that can be used for generating networks
with community structure and spatial embedding. Second we
have focused on the degree distribution of networks. The
degree distributions are not all equivalent. In fact the struc-
tural entropy depends strongly on the degree distribution of
the network ensemble. In particular, the power-law degree
distribution with exponent ␥ and fixed average degree are
associated with a structural entropy that decreases with ␥.
关1兴 S. N. Dorogovtsev, A. Goltsev, and J. F. F. Mendes, Rev. Mod.
Phys. 80, 1275 共2008兲.
关2兴 S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D. U.
Hwang, Phys. Rep. 424, 175 共2006兲.
关3兴 A.-L. Barabási and R. Albert, Science 286, 509 共1999兲.
关4兴 R. Pastor-Satorras, A. Vázquez, and A. Vespignani, Phys. Rev.
Lett. 87, 258701 共2001兲.
关5兴 S. Maslov and K. Sneppen, Science 296, 910 共2002兲.
关6兴 J. Berg and M. Lassig, Phys. Rev. Lett. 89, 228701 共2002兲.
关7兴 D. J. Watts and S. H. Strogatz, Nature 共London兲 4, 393 共1998兲.
关8兴 E. Ravasz, A. L. Somera, A. D. Mongru, Z. N. Oltvai, and
A.-L. Barabási, Science 297, 1551 共2002兲.
关9兴 S. Carmi, S. Havlin, S. Kirkpatrick, S. Shavitt, and E. Shir,
Proc. Natl. Acad. Sci. U.S.A. 104, 11150 共2007兲.
关10兴 S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, Phys.
Rev. Lett. 96, 040601 共2006兲.
关11兴 J. I. Alvarez-Hamelin, L. Dall’Asta, A. Barrat, and A. Vespig-
nani, cs.Ni/0511007.
关12兴 M. Girvan and M. E. J. Newman, Proc. Natl. Acad. Sci. U.S.A.
99, 7821 共2002兲.
关13兴 L. Danon, A. Díaz-Guilera, J. Duch, and A. Arenas, J. Stat.
Mech.: Theory Exp. 共2005兲 P09008.
关14兴 M. E. J. Newman and E. A. Leich, Proc. Natl. Acad. Sci.
U.S.A. 104, 9564 共2007兲.
关15兴 B. Blasius and L. Stone, Nature 共London兲 406, 846 共2000兲.
关16兴 S. N. Dorogovtsev, P. L. Krapivsky, and J. F. F. Mendes, Eu-
rophys. Lett. 81, 30004 共2008兲.
关17兴 M. Boguña, R. Pastor-Satorras, A. Díaz-Guilera, and A. Are-
nas, Phys. Rev. E 70, 056122 共2004兲.
关18兴 A. Barrat, M. Barthélemy, R. Pastor-Satorras, and A. Vespig-
nani, Proc. Natl. Acad. Sci. U.S.A. 101, 3747 共2004兲.
关19兴 G. Bianconi, N. Gulbahce, and A. E. Motter, Phys. Rev. Lett.
100, 118701 共2008兲.
036114-10
PHYSICAL REVIEW E 79, 036114 共2009兲
Nevertheless we have shown that power-law degree distribu-
tions are the more likely distributions associated with small
structural entropy. This sheds light on the evidence that
power-law networks constitute a large universality class in
complex networks with a nontrivial level of organization.
ACKNOWLEDGMENT
This work was supported by IST STREP GENNETEC
Contract No. 034952.
关20兴 Z. Burda, J. D. Correia, and A. Krzywicki, Phys. Rev. E 64,
046118 共2001兲.
关21兴 S. N. Dorogovstev, J. F. F. Mendes, and A. N. Samukhin, Nucl.
Phys. B 666, 396 共2003兲.
关22兴 J. Park and M. E. J. Newman, Phys. Rev. E 70, 066146
共2004兲.
关23兴 B. Soderberg, Phys. Rev. E 66, 066121 共2002兲.
关24兴 F. Chung and L. Lu, Ann. Comb. 6, 125 共2002兲.
关25兴 D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J.
Watts, Phys. Rev. Lett. 85, 5468 共2000兲.
关26兴 M. Boguñá and R. Pastor-Satorras, Phys. Rev. E 68, 036112
共2003兲.
关27兴 D. Garlaschelli and M. I. Loffredo, Phys. Rev. E 78,
015101共R兲 共2008兲.
关28兴 R. F. i Cancho and R. Solé, e-print arXiv:cond-mat/0111222.
关29兴 L. Bogacz, Z. Burda, and B. Waclaw, Physica A 366, 587
共2006兲.
关30兴 M. Bauer and D. Bernard, preprint, arXiv:cond-mat/0206150.
关31兴 G. Bianconi, Europhys. Lett. 81, 28005 共2008兲.
关32兴 G. Bianconi, A. C. C. Coolen, and C. J. Perez Vicente, Phys.
Rev. E 78, 016114 共2008兲.
关33兴 J. Gomez-Gardenes and V. Latora, Phys. Rev. E 78, 065102
共2008兲.
关34兴 H. Kim, Z. Toroczkai, I. Miklos, P. L. Erdös, and L. A. Székely
共unpublished兲.
关35兴 M. Molloy and B. A. Reed, Random Struct. Algorithms 6, 161
共1995兲.
关36兴 G. Bianconi, Chaos 17, 026114 共2007兲.
关37兴 E. Bender and E. Rodney Canfield, J. Comb. Theory, Ser. A
24, 296 共1978兲.
关38兴 B. D. McKay, J. Parallel Distrib. Comput. 19A, 15 共1985兲.
关39兴 F. Ritort, Phys. Rev. Lett. 75, 1190 共1995兲.