Professional Documents
Culture Documents
a -
z
"
~
"a..
'"
"
~~
"'",
"""
u"
3:~
~"
G~
;c~
~~
~~
The brain is a network of nerve cells con- number of links and no node is typical of network theory is that, despite the ran-
nected by axons, and cells themselves are the others. These networks also behave in dom placement of links, the resulting sys-
networks of molecules connected by bio- certain predictable ways; for example, tem will be deeply democratic: most
chemical reactions. Societies, too, are net- they are remarkably resistant to acciden- nodes will have approximately the same
works of people linked by friendships, tal failures but extremely vulnerable to number of links. Indeed, in a random net-
familial relationships and professional coordinated attacks. work the nodes follow a Poisson distrib-
ties. On a larger scale, food webs and eco- Such discoveries have dramatically ution with a bell shape, and it is extreme-
systems can be represented as networks changed what we thought we knew about ly rare to find nodes that have significant-
of species. And networks pervade tech- the complex interconnected world around ly more or fewer links than the average.
nology: the Internet, power grids and us. Unexplained by previous network the~ Random networks are also called expo-
transportation systems are but a few ex" ories, hubs offer convincing proof that nential, because the probability that a
amples. Even the language we are using various complex systems have a strict ar- node is connected to k other sites de-
to convey these thoughts to you is a net- chitecture, ruled by fundamental laws- creases exponentially for large k.
work, made up of words connected by laws that appear to apply equally to cells, So in 1998, when we, together with
syntactic relationships, col11puters, languages and society. Fur- Hawoong Jeong and Reka Albert of the
Yet despite the importance and per- thermore, these organizing principles have University of Notre Dame, embarked on
vasiveness of networks, scientists have significant implications for developing a project to map the World Wide Web,
had little understanding of their structure better drugs, defending the Internet from we expected to find a random network.
and properties. How do the interactions hackers, and halting the spread of deadly Here's why: people follow their unique
of several malfunctioning llodes in a com- epidemics, among other applicati()ns. interests when deciding what sites to link
plex genetic network resl).lt in cancer? their Web documents to, and given the di-
How does diffusion occur so rapidly.in Networks without Scale versity of everyone's .interests and the
certain social and communications sys- FOR MORE THAN 40 YEARS, science tremendous number of pages they can
tems, leading to epidemics ()fdiseases and treated all complex net\'\(orks as being choose from, the resulting pattern of con-
computer viruses? How do ~ome net- completely randoIIl. This paradigIIl has its nections should appear fairly random.
works continue to function even after the roots in the work of two H1ir1garianmath~ The measurements, however, defied
vast majority of their nodes have failed? ematicians, the inimitable Paul Erdos and that expectation. Software designed for
Recent research has begun to ansWer his close collaborator Alfred Renyi. In this project hopped from one Web page
such questions. Over the past feW years, 1959, aiming to describe networks seen in to another and collected all the links it
investigators from a variety of fields have communications and the lif~ sciences, could. Although this virtu:alrobot reached
discovered that many .networks-froIIl Erdos and Renyi suggested that suc:h sys- only a tiny fraction of the entire Web, the
the World Wide Web to a cell's metabol- tems could be effectively l110deledby con- map it assembled revealed something
ic system to actors in Hollywood-are necting their nodes withrandornlyplaced quite surprising: a few highly connected
dominated by a relatively sm.all number links. The simplicitYof their approach and pages are essentially holding the World
of nodes that ani c:oll1l~ctedto many oth" the elegance of so~e of their related theo- Wide Web together. More than 80 per-
er sites. Networks containing suchim- rems revitalized graph theory, leading to cent ()fthe pages on the map had feWer
portant nodes, or hubs, tend to be what the emergence of a field in mathematics than four links, but a small minority, less
we call "scale-free," in the sense that that focuses on random networks. than 0.01 percent of all nodes, had more
some hubs have a seemingly unlimited An important prediction of random- than 1,000. (A subsequent Web survey
would uncover one document that had
been referenced by more than two million
other pages!)
Counting how many Web pages have
ex'1stly k links showed that the distribu-
tion followed a so-called power law: the
probability that any node was connected
to k other nodes was proportional to Ykn.
The value of n for incoming links was ap-
proximately 2, so, for instance, any node
was roughly four times as likely to have
just half the number of incoming links as
another node. Power laws are quite dif-
ferent from the bell-shaped distributions
that characterize random networks.
RANDOMNETWORKS,
which resemble the U.S.highway system nodes with a very high number of links. In such networks, the
(simplified in left map), consist of nodes with randomly placed distribution of node linkages follows a power law [center graph)
connections. In such systems, a plot of the distribution of node in that most nodes have just a few connections and some have
linkages will follow a bell-shaped curve (left graph), with most a tremendous number of links. In that sense, the system has no
nodes having approximately the same number of links. "scale." The defining characteristic of such networks is that the
In contrast, scale-free networks, which resemble,the U.S. distribution of links, if plotted on a double-logarithmic scale
airline system (simplified in right map). contain hubs [red)- [right graph), results in a straight line.
L ~
If)
QJ
-c
0
~ . ~~
Z 0 0 QJ
'0 Z zCij
""'0
0 0 If)
c;;
..c ~ ~ 011 '
E ~ ~~
:::J E E~
Z :::J. :::J
Z Z
Specifically, a power law does not have a Scale-Free Networks Abound some social networks are scale-free. A col-
peak, as a bell curve does, but is instead de- OVER THE PAST several years, re- laboration between scientists from Boston
scribed by a continuously decreasing func- searchers have uncovered scale-free struc" University and Stockholm University, for
tion. When plotted on a double-logarith- tures in a stunning range of systems. instance, has shown that a netWork of
mic scale, a power law is a straight line When we studied the World Wide Web, sexual relationships among people in
[see illustration above]. In contrast to the we looked at the virtual network of Web Sweden followed a poWer law: although
democratic distribution of links seen in pages connected to one another by hy- most individuals had only a few sexual
random networks, power laws describe perlinks. In contrast, .ty1ichalisFaloutsos partners during their lifetime, a few (the
systems in which a few hubs, such as Ya- of the University of California at River- hubs) had hundreds. A recent study led
hoo and Google, dominate. side, Petros Falotitsos of the University of by Stefan Bornholdt of the University of
Hubs are simply forbidden in random Toronto arid Christos Faloutsos of Car- Kiel in Germany concluded that the net-
networks. When we began to map the negie MelloQ Uq~versity .analyzed tbe work of people connected bye-mail is
Web, we expected the nodes to follow a physical structure of the Internet. These likewise scahfree. Sidney Redner of
bell-shaped distribution, as do people's three computer-scientist brothers investi- Boston University demonstrated that the
heights. Instead we discovered certain gated the routers connected by optical or network of scientific papers, connected
nodes that defied explanation, almost as other communications lines and found by citations, follows a power law as well.
~
~ if we had stumbled on a significant num- that the topology of that network, too, is And Mark Newman of the University of
;;:
'" ber of people who were 100 feet tall, thus scale-free. Michigan at Ann Arbor examined col-
~ prompting us to coin the term" scale-free." Researchers have also discovered that laborations among scientists in several
.---- -1
~ ~ ~
connected actors are more likely to be why scale-free networks are so ubiquitous
~~
an existing node that has twice as many
chosen for new roles. On the Internet the in the real world. connections), one hub will tend to run
more connected routers, which typically Growth and preferential attachment away with the lion's share of connections.
have greater bandwidth, are more desir- can even help explicate the presence of In such "winner take all" scenarios, the
able for new users. In the U.S. biotech in- scale-free networks in biological systems. network eventually assumes a star topol-
dustry, well-established companies such as Andreas Wagner of the University of ogy with a central hub.
Genzyme tend to attract more alliances, New Mexico and David A. Fell of Oxford
which further increases their desirability Brookes University in England have AnAchilles' Heel
for future partnerships. Likewise, the found, for instance, that the most-con- AS HUMANITY BECOMES increasing-
most cited articles in the scientific litera- nected molecules in the E. coli metabolic ly dependent on power grids and com-
ture stimulate even more researchers to network tend to have an early evolution- munications webs, a much-voiced con-
read and cite them, a phenomenon that ary history: some are believed to be rem- cern arises: Exactly how reliable are these
noted sociologist Robert K. Merton nants of the so-called RNA world (the types of networks? The good news is that
called the Matthew effect, after a passage evolutionary step before the eInergence of complex systems can be amazingly re-
in the New Testament: "For unto every DNA), and others are coiJ;].poneptsohre silient against accidental failures. In fact,
one that hath shall be given, and he shall most ancient metabolic pathways. although hundreds of routers routinely
have abundance." Interestingly, the mechanism of pref- malfunction on the Internet at any mo-
These two mechanisms-growth and erential attachment tends to be linear. In ment, the network rarely suffers major
preferential attachment-help to explain other words, a new node is twice as Hke- disruptions. A similar degree of robust-
the existence of hubs: as new nodes ap- ly to link to an existing node that has ness characterizes. living systems: people
pear, they tend to connect to the more twice as many connections as its neigh- rarely notice the consequences of thou-
connected sites, and these popular loca- bor. Redner and his colleagues at Boston sands of errors in their cells, ranging from
tions thus acquire more links over time University and elsewhere have investigat- mutations to misfolded proteins. What is
than their less connected neighbors. And ed different types of preferential attach- the origin of this robustness?
this "rich get richer" process will gener- ment and have learned that if the mecha- Intuition tells us that the breakdown
ally favor the early nodes, which are more nism is faster than linear (for example, a ofa substantial number of nodes will re-
likely to eventually become hubs. new node is four times as likely to link to sult in a network's inevitable fragmenta-
Along with Reka Albert, we have used
computer simulations and calculations to ALBERT'L4SZL.6BARABASI. and ERIC BONABEAU study the behavior and characteristics of
show that a growing network with pref- myriad complex systems, ranging from the Internet to insect colonies. Barabasi is Emil T.
erential attachment will indeed become Hofman. Professor of Physics at the. University of Notre Dame, where he directs research
scale-free, with its distribution of nodes on complex networks. He is author of Linked: The New Science of Networks. Bonabeau is
following a power law. Although this the- chief scientist at Icosystem, a consulting firm based in Cambridge, Mass., that applies the
~
::;: oretical model is simplistic and needs to
~ tools of complexity science to the discovery of business opportunities. He is co-author of
;;:
be adapted to specific situations, it does SwarmIntelligence: From Natural to ArtificialSystems. Thisis Bonabeau's second article
~
'" appear to confirm our explanation for for Scientific American.
THEACCIDENTAL
FAILUREof a number of nodes in a random more robust in the face of such failures (middle panels).
network (top panels) can fracture the system into non- But they are highly vulnerable to a coordinated attack against
communicating islands. In contrast, scale-free networks are their hubs (bottom panels).
Failed node
0 .
Before After
0
~
0
p 0
'*
Before After
.
0
Before After
then spread the virus throughout the en- traditionalpublichealth approach of ran- thoggh, what if doctors targeted the hubs,
tire system. dom immunization could easily fail be- or the most connected individuals? Re-
The fact that biological viruses spread cause it would very likely neglect a num- search in scale-free networks indicates
in social networks, which in many cases ber ofthe hubs. In fact, nearly everyone that this alternative approach could beef-
appear to be scale-free, suggests that sci- would have to be treated to ensure that fective even if the immunizations reached
entists should take a second look at the the hubs were not missed. A vaccination only a small fraction of the overall popu-
:::i
~
volumes of research written on the inter- for measles, for instance, must reach 90 lation, provided that the fraction con-
"- tained the hubs.
Z play of network topology and epidemics. percent of the population to be effective.
:J
'" Specifically, in a scale-free network, the Instead of random immunizations, But identifyingthe hubs in a social
--~--
congestion along specific links is a major
consideration: too much traffic on a par-
ticular link can cause it to break down,
leading to the potential failure of other
links that must then handle the spillover.
And the nodes themselves might not be
homogeneous-certain Web pages have
more interesting content, for instance-
which could greatly alter the preferential-
attachment mechanism.
Because of these and other factors, sci-
entists have only begun to uncover the be-
havior of scale-free systems. Immunizing
hubs, for instance, might not be sufficient
to stop the spread of a disease; a more ef-
fective solution might be found by con-
sidering not just the number of connec-
tions a person has but also the frequency
and duration of contact for those links.
In essence, we have studied complex
networks first by ignoring the details of
their individual links and nodes. By dis-
tancing ourselves from those particulars,
we have been able to better glimpse some
of the organiziIlg principles behind these
seemingly incomprehensible systems. At
the very least, knowledge from this en-
MAPOFINTERACTING PROTEINSin yeast highlights tl'\!!discoverytha.~high1ylinked, or hub, proteins
deavor has led to the rethinking of many
tend to be crucial for a cell's survival. Red denotes essential proteir"js(their removal will cause the cell
to die). Orange represents proteins of some importance [their removal will slow cell growth). Green
basic assumptions. In the past, for exam-
and yellow represent proteins of less.er or unknoWn significance, respectively. ple, researchers modeled the Internet as a
random network to test how a new rout-
seen in materials science. In a crystallat- b~havior of such systems. There might be ing protocol might affect system conges-
tice, for instance, atoms have the same steep~osts, for instance, with the addition tion. But we now know that the Internet is
number of links to their neighbors. With of each link to a given node that could a scale-free system with behavior that is
other networks, the data are inconclusive. prevent certain networks (such as the U.S. dramatically different from a random net-
The relatively small size of food webs, highway system) from becoming scale- work's. Consequently, investigators such
whiclI §how predator-prey relationships, free. In food Ghains, some prey are easier as John W. Byers and his colleagues at
has prevented scielltists from reaching a to catch than others, and that fact has a Boston University are revamping the com-
clear conclusion regarding that network's profound effe.cton the overall ecosystem. puter models they have been using to sim-
type. And the absenc~ oflarge-scale con- With social networks, ties among house- ulate the Internet. Similarly, knowledge of
nectivitymaps of the brain has kept re- hold members are much stronger than the properties of scale-free networks will
searchers from knowing the nature of connections to c~sual acquaintances, so be valuable in a number of other fields, es-
that important network as well. diseases (and information) are more like- pecially as we move beyond network to-
Determining whetl~er a network is ly to spread through such linkages. For pologies to probe the intricate and often
scale-free is important in understanding transportation, transmission and commu- subtle dynamics taking place within those
the system's behavior, but other signifi- nications systems (such as the Internet), complex systems. Ii!I1
'"
cantparameters merit attention, too.
One such characteristic is the diameter, MORE: TO E:XPLORE: .I
:z
or path length, of a network: the largest All the World's a Net, David Cohen in New Scientist, Vol. 174, No. 2338, pages 24-29; April 13, 2002.
~
'"
:z number'gf hops required to get from one Statistical Mechanics IIfComplex Networks. Reka Albert and Albert-Laszlo Barabasi in Reviews
0
3< node to another by following the shortest of Modern Physics, Vol.74, pages 47-97; January 2002.
::I: route possible [seebox on opposite page]. Linked: The New Science of Networks. Albert-Laszlo Barabasi. Perseus Publishing, 2002.
~
0
0
u derstanding the overall characteristics and Find lillks tQ paper.!>on sca.le-free networks at www.nd.edu/-networks