Evolving networks with disadvantaged long-range connections
Extracted source text ยท arXiv cond-mat/0205136
\documentstyle[aps,psfig,multicol]{revtex}
\begin{document}
\setlength{\topmargin}{-1.3cm}
\draft
\title{Evolving networks with disadvantaged long-range connections}
\author{R. Xulvi-Brunet$^1$ and I.M. Sokolov$^{1,2}$}
\address{$^1$Institut f\"{u}r Physik, Humboldt Universit\"{a}t zu Berlin,
Invalidenstra\ss e 110, D-10115 Berlin, Germany}
\address{$^2$Theoretische Polymerphysik, Universit\"at Freiburg, Hermann Herder
Str. 3, D-79104 Freiburg, Germany}
\date{\today}
\maketitle
\bigskip
\begin{abstract}
{\bf Abstract.} We consider a growing network, whose growth algorithm is based
on the preferential attachment typical for scale-free constructions, but where
the long-range bonds are disadvantaged. Thus, the probability to get connected
to a site at distance $d$ is proportional to $d^{-\alpha }$, where $\alpha $
is a tunable parameter of the model. We show that the properties of the
networks grown with $\alpha <1$ are close to those of the genuine scale-free
construction, while for $\alpha >1$ the structure of the network is vastly
different. Thus, in this regime, the node degree distribution is no more a
power law, and it is well-represented by a stretched exponential. On the
other hand, the small-world property of the growing networks is preserved at
all values of $\alpha $.
\end{abstract}
\pacs{PACS numbers: 89.75.-k, 05.50.+q, 89.75.Hc}
\begin{multicols}{2}
Complex weblike structures (the small-world or scale-free networks) have
recently become an object of extensive investigation, and in the last years a
great success in understanding the properties of these structures was achieved
(see Ref.\cite{review} as a review). Apart from appealing mathematics,
this recent interest is due to the fact that many natural and technological
systems, like polymer networks \cite{Sokolov}, the science collaboration network
\cite{New1,New2,New3}, or networks of chemical reactions in a living cell
\cite{Oltvai,Fell,Mason} seem to be organized according to some internal
principles. Thus, the Internet \cite{Faloutsos}, the network of human sexual
contacts \cite{Amaral} or the WWW \cite{Albert} possess a similar structure, e.g.
are they all based on the preferential attachment of the newly introduced nodes
to the highly connected old ones. All these networks show the small-world
property: the typical distance (in terms of the number of intermediate
connections) between two nodes grows logarithmically with the web's size.
One of the prominent examples of a mathematical model of such a growing network
is the scale-free (SF) construction of Barab\'{a}si and Albert
\cite{review,scalefr}; and one of its most interesting properties is the
very specific form of the probability distribution of the degree of
nodes (i.e. of the numbers of bonds connecting any given node $i$ with other
ones in the network): $P(k)\propto k^{-3}$ \cite{review,scalefr,Do2,Kull,Kra}.
Many models have been presented, based on the same two most important
ingredients: growth and preferential attachment. Examples are models with an
accelerated growth of the network \cite{Do4,Bar3}, models with a nonlinear
preferential attachment \cite{Kra}, with nodes provided by a initial
attractiveness \cite{Do2,Je1}, with growth constraints as aging and cost
\cite{Amar,Do1}, models that have a competitive aspect of the nodes
\cite{Bianconi}, or models of networks that incorporate local events as the
addition, rewiring or removal of nodes or edges \cite{Alb}.
The SF-construction may be a reasonable approximation for such
world-spanning networks like one of the Internet's information transmission
channels or one of the formal links of WWW. On the other hand, in many
situations (like in a network of human sexual contacts) a connection
means a physical contact, i.e. means that the contacting individuals,
representing the nodes of the network, have to occur at the same site and at
the same time, thus introducing a clear geographical aspect. In what follows
we present a simple model taking into account this metrical (''geographical'')
aspect, where the probability to connect two nodes depends both on the number
of connections that the nodes already have (as in the genuine SF-construction),
and on the distance between them. That is, we treat an emerging network in a
metric space. In this emerging network the probability that a newly introduced
node $n$ is connected to a previously existing node $i$ is proportional to the
number $k_{i}$ of the already existing connections of node $i$ (preferential
attachment prescription), but on the other hand the too long bonds are
disadvantaged, because this probability depends on the Euclidean distance
$d_{in}$ between the nodes $n$ and $i$ as $d_{in}^{-\alpha }$, (clearly, a
''scale-free'' function), with $\alpha >0$.
Based on extensive numerical simulations of a one-dimensional situation, we
show that even if the length penalties are mild, the model exhibits
properties which differ strongly from those of the usual scale-free networks.
Thus, the corresponding degree distribution function $P(k)$ depends strongly
on $\alpha $. We show, in particular, that for $\alpha <1$ the behavior of
$P(k)$ is similar to the behavior of the SF model without penalties, so that
asymptotically $P(k)\propto k^{-3}$, (which distribution possesses a mean, but
no dispersion, and corresponds to strong, universal fluctuations). On the other
hand, for $\alpha >1$ the behavior of $P(k)$ is well-described by a
stretched-exponential $P(k)\propto \exp (-bk^{\gamma })$, with the power $%
\gamma $ depending on $\alpha $, so that the fluctuations in $k$ are rather
weak. We discuss the reasons for such a dramatic change, being rooted in the
probability of connection between the nodes as function of the distance, and
the overall structure of the emerging network, preserving its small-world
nature even at large (probably at all) $\alpha $-values.
\end{multicols}
\begin{figure}[t]
\centerline{\hspace{-1cm}\psfig{figure=figure1.ps,width=9in}}
\caption{Networks generated using the simulation prescription, Eq.(2), with
different values of $\alpha$: (a) $\alpha=0$, (b) $\alpha=1.5$
and (c) $\alpha=15$. All three examples have $300$ edges, $L=10^6$,
$N=102$, and $m=3$. Note the change in the appearence of the
networks. The network (a) is a genuine SF construction while (c)
strongly resembles the Watts and Strogatz's small-world network.}
\label{fig:5}
\end{figure}
\bigskip
\begin{multicols}{2}
We start from a one-dimensional lattice of $L$ sites, spaced by a unit
distance and apply cyclic boundary conditions. On this structure we will let
our network grow, so that each lattice site will be a possible location of a
network's node. We denote by $n_{i}$ the position in the lattice of a node $i$.
The distance $d_{ij}$ between any two nodes $i$ and $j$ is defined as:
\begin{equation}
d_{ij}=min\{|n_{i}-n_{j}|,(L-|n_{i}-n_{j}|)\}.
\end{equation}
Let us now construct the network. First, we choose randomly an even number $%
m_{0}$ of sites from the lattice and we bind them in pairs with one bond each.
This will be our initial condition. That is, at $t=0$, our network will
consist from $m_{0}$ nodes connected in pairs. As in the SF model we will
add at every time step a new node to our network (linear growth). We proceed
according to the following rule: at every time step we choose at random a
free site of our lattice, and pose the new node there. This new node is then
connected through $m$ edges ($m\le m_{0}$) with $m$ different nodes already
present in the network. After $t$ time steps the algorithm results
in a network with $t+m_{0}$ nodes and $mt+m_{0}/2$ edges. In contrast with
the SF model, the probability $\Pi $ for the new node $n$ to be connected to
an old one $i$ will depends not only on the number of edges $k_{i}$ which $i$
already possesses, but also on the distance $d_{in}$ between them:
\begin{equation}
\Pi (k_{i},d_{in},\alpha )=\frac{k_{i}\cdot d_{in}^{-\alpha }}{%
\sum_{j}k_{j}\cdot d_{jn}^{-\alpha }}. \label{Apriori}
\end{equation}
Here the sum in the denominator goes over all nodes in the system except the
newly introduced one and $\alpha $ is a real non-negative parameter
describing the distance penalties. For large $\alpha $, the probability of
connection between two distant nodes is very small. On the other hand,
for a very small $\alpha $ the probability is almost independent from the
distance. In the case $\alpha =0$ our model reduces to the genuine
scale-free one. Note that our model is to some extent also scale-free: the
connection probabilities depend only on the {\it relative} distances.
Our initial condition is slightly different from one of Barab\'{a}si and
Albert, where the initial $m_{0}$ nodes are not connected: in our case all
nodes introduced at $t=0$ have exactly one edge, which allows to use Eq.(\ref
{Apriori}) from the very beginning. This simplifies the algorithm, since we
do not have to distinguish between the initial and the further steps. The
only difference with the genuine SF construction is that at time $t$ one has
$mt+m_{0}/2$ (instead of $mt$) edges present; hence, the asymptotic behavior of
both models for $t\rightarrow \infty $ is the same.
Three examples of the evolving networks of such a kind are given en Fig.1. Here
is $m=3$, $L=10^6$, $N=102$ and $m_0=6$, (so that all three networks have
exactly $300$ edges). Three different values of $\alpha$ were used:
$\alpha=0.0$ (scale-free model), $\alpha=1.5$ and $\alpha=15.0$. Note that
increasing $\alpha$ leads to marked changes in the topology of the network.
Fig. 1(a) corresponds to a genuine scale-free construction and exhibits a
lot of long bonds connecting distance sites. On the other hand, only few such
bonds are present in Fig. 1(c).
In our further simulations we use a lattice of $L=2\cdot 10^{7}$
sites; the maximal number of the introduced nodes is $N=2\cdot 10^{5}$. All
simulation results are based on the average of 10 realizations of this
structure. The error bars on Figs. 3-5 correspond just to this ensemble average.
The simulations are done for several values of $\alpha $ and for two values of
$m$, the number of the outgoing bonds: $m=1$ and $m=3$; $m_0=2m$.
\begin{figure}[htbp]
\centerline{\hspace{0.4cm}\psfig{figure=figure2.ps,width=3.9in}}
\caption{The degree distribution $P(k)$ for different values of $\alpha$ and
for $m=1$ (a) and $m=3$ (b). The values of $\alpha$ are
$\alpha=0$ (squares), $\alpha=0.8$ (crosses), $\alpha=1.5$
(triangles), $\alpha=2$ (filled circles), $\alpha=5$ (plusses)
and $\alpha=45$ (diamonds). The dashed lines correspond to the
theoretical curve for the scale-free model, (Ref. [1,12])}.
\label{fig:5}
\end{figure}
\bigskip
One of the prominent features of the scale free-model is that the
distribution of the degrees of the nodes decays as a power law, i.e. like $%
P(k)\sim k^{-\gamma }$, with $\gamma =3$. This corresponds to the fact that
the mean number of connections per site exists, but its dispersion diverges.
Let us now discuss, how this distribution changes if the long-range
connections are penalized. In Fig. 2 we plot the probability distribution of
$k$ for different values of $\alpha $ on double logarithmic scales. One
readily infers that for all $0<\alpha <1$ no important differences with the
scale free model ($\alpha =0$) can be detected: in any case the
asymptotic behavior of $P(k)$ is well-described by $P(k)\sim k^{-3}$. The
distributions seem to be almost identical; however, small but statistically
significant deviations can be detected for small $k$-values. At $\alpha
\simeq 1$ the degree distribution shows a pronounced change in its behavior
and ceases being a power law; now the behavior of the model with distance
penalties is quite different.
Let us concentrate on the case $\alpha >1$ and try to describe the shape
of the degree distribution under such conditions. The analysis of the
simulations suggests that the corresponding mathematical expression could be
a stretched-exponential function of the form:
\begin{equation}
P(k)=a\exp (-b\,k^{\gamma }) , \label{Fit}
\end{equation}
where the parameters $a$, $b$ and $\gamma $ depend on $\alpha $ and $m$. To
obtain the values of these parameter and to analyze the goodness of this
fitting function we have fitted the data to Eq.(\ref{Fit}) using the nonlinear
least-squares Levenberg-Marquardt algorithm \cite{Marq}, taking into
consideration the error bars as coming out of $10$ realizations of each
situation. The data is replotted together with the outcomes of the fits in
Fig. 4 on the scales in which the fitting function, Eq.(\ref{Fit}), is
represented by a straight line. One namely takes $k^{\gamma }$ as the abscissa
and $\ln P(k)$ as the ordinata of the graph. Fig. 4 shows that such a fit
(straight line) is surprisingly good!
The values of the exponent $\gamma $ are
shown as a function of $\alpha $ ($\alpha >1$) in Fig. 3, for the two different
situations corresponding to $m=1$ and $m=3$. We see that $\gamma $
monotonously grows with $\alpha $, and that the dependences $m=1$ and $m=3$
differ, i.e. that the $\gamma (\alpha )$ dependence is nonuniversal.
\begin{figure}[htbp]
\centerline{\hspace{-0.4cm}\psfig{figure=figure4.ps,width=4.75in}}
\caption{The parameter $\gamma$ as a function of $\alpha$. The upper
dependence corresponds to $m=1$, and the lower one $m=3$. The
lines are drawn as a guide for eyes.}
\label{fig:5}
\end{figure}
\bigskip
We note that in related models of growing networks another form of
degree distribution appears: an exponentially demped power-law \cite{Newman},
\end{multicols}
\begin{figure}[t]
\centerline{\hspace{-0.2cm}\psfig{figure=figure3.ps,width=10in}}
\caption{Shown is $\ln{P(k)}$ as a function of $k^{\gamma}$, where $\gamma$
is the output of the fit, Eq.(3). (See text for details). The
parameters are: (a): $m=1$, $\alpha=1.5$, $\gamma=0.37$. (b):
$m=3$, $\alpha=1.5$, $\gamma=0.33$. (c): $m=1$, $\alpha=3$,
$\gamma=0.69$. (d): $m=3$, $\alpha=3$, $\gamma=0.64$. (e): $m=1$,
$\alpha=10$, $\gamma=1.07$. (f): $m=3$, $\alpha=10$, $\gamma=0.96$.}
\label{fig:5}
\end{figure}
\bigskip
\begin{multicols}{2}
\begin{equation}
P(k)=ak^{\gamma }\exp (-b\,k). \label{Cut}
\end{equation}
We tested also this fit function and found out that it gives a good fit for
larger $\alpha $-values, but is definitely inferior to our fit, Eq.(\ref{Fit}),
for $1<\alpha <3$.
A growing network with disadvantaged long bonds is a very interesting
hierarchical construction. Thus, for large $\alpha $, the strong correlation
between the age of the connection and its length exists. The old
connections, made when the nodes were sparce, are typically long, while the
younger connections get shorter and shorter, since more sites in the
immediate vicinity of a newly introduced site can be found. The simulations
show that for $\alpha $ large, the nodes are almost surely connected to
their nearest neighbors. On the other hand, the old, long-range connections
are of great importance for the overall topology of the lattice, since
they guarantee that for any $\alpha $ the network is a small-world one.
\begin{figure}[htbp]
\centerline{\hspace{-0.3cm}\psfig{figure=figure5.ps,width=7.7in}}
\caption{The diameter of a network as a function of the number of sites
$N$. Panel (a) corresponds to $\alpha=1.5$ and panel (b)
corresponds to $\alpha=5$. The upper lines in each panel are
those for $m=1$, the lower lines correspond to $m=3$. Note the
logarithmic scale.}
\label{fig:5}
\end{figure}
\bigskip
In Fig. 5 we plot the mean number of connections between each two nodes of
the network for two different values of $\alpha $ ($\alpha =1.5$ and
$\alpha=5$) and for the two values $m=1$ and $m=3$ as a function of the network
size $N$. The algorithm here is trivial: starting from a node (labeled 0) we
pass to all nodes connected to it (nodes of the first generation, labeled 1),
then to nodes of the second generation (labeled 2), etc; untill all nodes are
labeled. The mean distance between this node (labeled 0) and any other given
node of the network is then the sum of all values of these labels divided by
$N-1$. This procedure is repeated for each node, and the overall mean value,
the so-called path diameter of the network ($l$), is evaluated. The error bars of
the figure correspond to the average of the mean diameters over 10 realizations
of the network. Fig.5 shows that the mean diameter of the network grows linearly
in $\ln N$, i.e. it shows the typical small-world behavior. This behavior is
preserved for all tested values of $\alpha ;$ the largest value tested was $%
\alpha =45,$ which, for $m=1$, corresponds to a practically sure connection
of a newly introduced node to its nearest neighbor. The high-$\alpha $
networks resemble closely the simple small-world constructions \cite{Watts}.
Let us summarize our findings. We considered a growing network, whose growth
algorithm is based, as in the SF construction, on a preferential attachment
of the newly introduced nodes to the highly connected old ones. However,
here the too long connections are disadvantaged by introducing penalties.
Thus, the probability to connect two nodes separated by a distance $d$ is
proportional to $d^{-\alpha }$, where $\alpha $ is a variable parameter. We
found out that for $\alpha <1$ the degree distribution $P(k)$ decays, as in the
SF model, like $P(k)\sim k^{3}$, whereas for $\alpha >1$ a stretched exponential
form $P(k)=a\exp (-bk^{\gamma })$ gives an extremely good description of this
distribution. On the other hand, the small-world property is preserved at all
checked values of $\alpha $.
The authors are grateful to Mr. F. Jasch for fruitful discussions and to Mr.
M. Obermayr for the valuable technical assistence. Financial support by the
Fonds der Chemischen Industrie is gratefully acknowledged.
\smallskip
\begin{references}
\bibitem{review} R. Albert and A.-L. Barab\'asi, Statistical mechanics of
complex networks, Rev. Mod. Phys. {\bf 74}, 47 (2002)
\bibitem{Sokolov} S. Jespersen, I.M. Sokolov and A. Blumen, Small-world rouse
networks as models of cross-linked polymers, J. Chem. Phys. {\bf 113}, 7652
(2000)
\bibitem{New1} M.E.J. Newman, The structure of scientific collaboration networks,
Proc. Natl. Acad. Sci. {\bf 98}, 404 (2001)
\bibitem{New2} M.E.J. Newman, Scientific collaboration networks. I.Network
construction and fundamental results, Phys. Rev. E {\bf 64}, 016131 (2001)
\bibitem{New3} M.E.J. Newman, Scientific collaboration networks. II.Shortest
paths, weighted networks, and centrality, Phys. Rev. E {\bf 64}, 016132 (2001)
\bibitem{Oltvai} H. Jeong, B. Tombor, R. Albert, Z.N. Oltvai, A.-L. Barab\'asi,
The large-scale organization of metabolic networks, Nature {\bf 407}, 651 (2000)
\bibitem{Fell} D.A. Fell and A. Wagner, The small world of metamolism, Nat.
Biotechnol. {\bf 18}, 1121 (2000)
\bibitem{Mason} H. Jeong, S.P. Mason, A.-L. Barab\'asi, Z.N. Oltvai, Lethality
and centrality in protein networks, Nature {\bf 411}, 41 (2001)
\bibitem{Faloutsos} M. Faloutsos, P. Faloutsos and C. Faloutsos, On power law
relationships of the internet topology, Comput. Commun. Rev. {\bf 29}, 251
(1999)
\bibitem{Amaral} F. Liljeros, C. R. Edling, L. A. Nunes Amaral , H. Stanley,
and Y. Aberg, The web of human sexual contacts, Nature {\bf 411}, 907 (2001)
\bibitem{Albert} R. Albert, H. Jeong, A.-L. Barab\'asi, Diameter of the World-Wide
Web, Nature {\bf 401}, 130 (1999)
\bibitem{scalefr} A.-L. Barab\'asi and R. Albert, Emergence of scaling in
random networks, Science {\bf 286}, 509 (1999)
\bibitem{Do2} Dorogovtsev S.N., J.F.F. Mendes and A.N. Samuhkin, Structure of
growing networks with preferential linking, Phys. Rev. Lett. {\bf 85}, 4633
(2000)
\bibitem{Kull} L. Kullmann and J. Kert\'esz, Preferential growth: exact solution
of the time-dependent distributions, Phys. Rev. E {\bf 63}, 051112 (2001)
\bibitem{Kra} Krapivsky P.L., S. Redner and F. Leyvraz, Connectivity of growing
random networks, Phys. Rev. Lett. {\bf 85}, 4629 (2000)
\bibitem{Do4} Dorogovtsev S.N. and J.F.F. Mendes, Effect of the accelerating
growth of communications networks on their structure, Phys. Rev. E {\bf 63},
025101 (2001)
\bibitem{Bar3} Barab\'asi A.-L., H. Jeong, E. Ravasz, Z. N\'eda, A. Schubert
and T. Vicsek, Evolution of the social network of scientific collaborations,
cond-mat/0104162 (2001)
\bibitem{Je1} Jeong H., Z. N\'eda and A.-L. Barab\'asi, Measuring preferential
attachment for evolving networks, cond-mat/0104131 (2001)
\bibitem{Amar} Amaral L.A.N., A. Scala, M. Barth\'el\'emy and H.E. Stanley,
Classes of small-world networks, Proc. Nat. Acad. Sci. USA {\bf 97}, 11149 (2000)
\bibitem{Do1} Dorogovtsev S.N. and J.F.F. Mendes, Evolution of networks with
aging of sites, Phys. Rev. E {\bf 62}, 1842 (2000)
\bibitem{Bianconi} G. Bianconi and A.-L. Barab\'asi, Competion and multiscaling
in evolving networks, Europhys. Lett. {\bf 54}, 436 (2001)
\bibitem{Alb} Albert R. and A.-L. Barab\'asi, Topology of evolving networks:
local events and universality, Phys. Rev. Lett. {\bf 85}, 5234 (2000)
\bibitem{Marq} B.P. Flannery, S. A. Teukolsky, W.T. Vetterling, {\it Numerical
Recipes: The Art of Scientific Computing} (Cambridge Univ. Press, 1985) p.675
\bibitem{Newman} M.E.J. Newman, Exact solutions of epidemic models on networks,
cond-math/0201433 (2002)
\bibitem{Watts} D.J. Watts and S.H. Strogatz, Collective dynamics of 'small-world'
networks, Nature {\bf 393}, 440 (1998)
\end{references}
\end{multicols}
\end{document}