Emergence of Scaling in Random Networks

Author(s): Albert-László Barabási and Réka Albert

Source: Science, New Series, Vol. 286, No. 5439 (Oct. 15, 1999), pp. 509-512
Published by: American Association for the Advancement of Science
Stable URL: http://www.jstor.org/stable/2899318 .
Accessed: 06/02/2014 19:43

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Association for the Advancement of Science is collaborating with JSTOR to digitize, preserve and
extend access to Science.

http://www.jstor.org

This content downloaded from 18.7.29.240 on Thu, 6 Feb 2014 19:43:52 PM

All use subject to JSTOR Terms and Conditions

A thatproposedto explain,in termsof sp-dbond-
Co/STO/LSMO
-50 1X . P.
-40
-30 wi
-20
> and a positive one when thereare only s and p
z z
-10
0 tion should also dependon the position of the
-2-1.5 -1 -0.5 0 0.5 1 1.5 2
Bias voltage (V)
B Co/ALO/STO/LSMO
20 _
~16-
12
8 6. M. B. Stearns,J. Magn. Magn. Mater. 5, 187 (1997).
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
Bias voltage (V)
Fig. 3. Bias dependence of the TMRratio in (A)
Co/STO/LSMO and (B) Co/ALO/STO/LSMO
tunnel junctions.
level of LSMO is situated above the Fermi
level of Co and a maximum of inverse TMR
is expectedwhen the Fermi level of LSMO is
approximatelyat the maximumof the spin I
DOS of Co. This is consistent with the max-
imum of inverse TMR observed at -0.4 V
for Co/STO/LSMOjunctions (Fig. 3A). For a
positive bias, the TMR is expected to change
sign and become normalabove 1 V when the
Fermi level of LSMO goes down into the
energy range of the majority spin d-band of
Co. This is also observed in Fig. 3A.
ForALO andALO/STObarriers,a predom-
inanttunnelingof s-characterelectrons(see ar-
row in Fig. 2B) is the usual explanationof the
positive polarization(6-8). The rapid drop
with bias (Fig. 3B) is similarto what has been
observedin mostjunctionswith ALO barriers,
andcompletelydifferentfromwhat is obtained
whenthe tunnelingis predominantly by d-char-
acter electrons (Fig. 3A). The origin of this
rapiddecreaseof the TMR at relativelysmall
bias has never been clearly explained.This is
roughlyconsistentwith the energydependence
of the DOS inducedby sp-d bondingeffects on
the firstatomiclayerof ALO in the calculation
of Nguyen-Mahnet al. (8) for the Co-ALO
interface.ButZhanget al. (13) havealso shown
that a large part of the TMR drop can be
attributedto the excitationof spin waves.
The experimentsreportedhere and in sev-
eral recentpublications(3, 4) demonstratethe
importantrole of the electronicstructureof the
metal-oxideinterfacein determiningthe spin
polarizationof the tunnelingelectrons.The neg-
ativepolarizationfor the Co-STOinterfacehas
been ascribedto d-d bonding effects between
Al and Ti (4). This interpretation is similarto
www.sciencemag.org
This content downloaded from 18.7.29.240 on Thu, 6 Feb 2014 19:43:52 PM
REPORTS
down to about 5% at 300 K in Co/STO/
ing, the positive polarizationat the Co-ALO LSMO (4). However, othertypes of oxides of
interface(8). However,thereis no generalthe- the double-perovskite family (for example,
ory predictingthe trend of the experimental Sr2FeMoO6) combine electronic properties
resultsfor Co-that is, a negativepolarization similar to those of manganites with a defi-
with oxides of d elements(STO, CLO,Ta205) nitely higher Curie temperature(15). Their
use in magnetic tunneljunctions is promising
states(ALO). It is likely thatthe spin polariza- for a new generationof tunneljunctions with
very high magnetoresistancefor room-tem-
Fermilevel with respectto the electroniclevels peratureapplications.
of each characterabove and below the gap of
the insulator.In addition, as an evanescent References and Notes
1. J. S. Moodera,L.R. Kinder,T. M. Wong, R. Meservey,
wave in an insulatoris a Bloch wave with an Phys. Rev. Lett. 74, 3273 (1995).
imaginarywave vector, one can expect differ- 2. J. De Boeck,Science 281, 357 (1998).
ent decay lengthsfor Bloch waves of different 3. M. Sharma,S. X. Wang, J. H. Nickel, Phys. Rev. Lett.
82, 616 (1999).
character.This meansthatthe finalpolarization 4. J. M. De Teresa et at., ibid., p. 4288.
could also dependon the thicknessof the bar- 5. P. M. Tedrow and R. Meservey, Phys. Rev. B 7, 318
rier, as illustratedby the calculationsof Mac- (1973); R. Meservey and P. M. Tedrow, Phys. Rep.
Larenet al. for Fe/ZnSe/Fejunctions(14). 238, 173 (1994).
The influence of the barrieron the spin 7. J. A. Hertz and K.Aoi, Phys. Rev. B 8, 3252 (1973).
polarizationopens new ways to shape and op- 8. D. Nguyen-Mahnet at., Mat. Res. Soc. Symp. Proc.
timize the TMR. Interestingbias dependencies 492, 319 (1998).
9. J. H. Parket at., Nature 392, 794 (1998).
can be obtainedwith barriersselecting the d 10. J. Nassar,M. Hehn,A. Vaures,F. Petroff,A. Fert,Appl.
electronsand probingthe fine structureof the Phys. Lett. 73, 698 (1998).
11. J. S. Mooderaet at., ibid. 81, 1953 (1998).
d-DOS,as in Fig. 3A. The DOS of a d-bandcan
12. K.Wang, thesis, New YorkUniversity (1999).
also be easily tailoredby alloying(for example, 13. S. Zhang,P. M. Levy,A. C. Marley,S. S. P. Parkin,Phys.
by introductionof virtualboundstates)to pro- Rev. Lett. 79, 3744 (1997).
duce specificbias dependencies.Althoughhere 14. J. M. MacLaren,X. G. Zhang, W. H. Butler,X. Wang,
Phys. Rev. 8 59, 5470 (1999).
we concentratedon the problem of the spin 15. K. I. Kobayashi,T. Kimura,H. Sawada, K.Terakura,Y.
polarizationof the Co electrodeand regarded Tokura,Nature 395, 677 (1998).
the strongly spin-polarizedLSMO only as a 16. We thank the group of A. Revcoleschi for providing
useful spin analyzer,the largeTMR ratiosob- LSMOtargets, R. Lyonnet and A. Vaures for their
experimental help and J. Nassar for fruitful discus-
tainedby combiningCo and LSMO electrodes sions. Supportedby the EuropeanUnion and the New
(50%with a STObarrier)arealso an interesting Energyand IndustrialTechnology Development Or-
result. The drawback arising from the low ganization of Japan.
Curie temperatureof LSMO (-350 K) is the
reduction of the TMR at room temperature, 1 July 1999; accepted 2 September 1999
Emergence of Scaling in
Random Networks
ALbert-LaszL6 Barabasi* and Reka Albert
Systems as diverse as genetic networks or the World Wide Web are best
described as networks with complex topology. A common property of many
large networks is that the vertex connectivities follow a scale-free power-law
distribution.This feature was found to be a consequence of two generic mech-
anisms: (i) networks expand continuously by the addition of new vertices, and
(ii) new vertices attach preferentially to sites that are already well connected.
A model based on these two ingredients reproduces the observed stationary
scale-free distributions,which indicates that the development of large networks
is governed by robust self-organizingphenomena that go beyond the particulars
of the individualsystems.
The inability of contemporaryscience to de- actions currently limits advances in many
scribe systems composed of nonidentical el- disciplines, ranging from molecular biology
ements that have diverse and nonlocal inter- to computer science (1). The difficulty of
describing these systems lies partly in their
Departmentof Physics,Universityof Notre Dame, topology: Many of them form rathercomplex
Notre Dame,IN 46556, USA. networks whose vertices are the elements of
*To whom correspondenceshould be addressed.E- the system and whose edges represent the
mail:alb@nd.edu interactionsbetween them. For example, liv-
SCIENCE VOL 286 15 OCTOBER 1999 509
All use subject to JSTOR Terms and Conditions
 REPORTS
ing systems form a huge genetic network these two ingredients,we show that they are
whose vertices are proteins and genes, the responsible for the power-law scaling ob-
chemical interactions between them repre- served in real networks. Finally, we argue
sentingedges (2). At a differentorganization- that these ingredientsplay an easily identifi-
al level, a large network is formed by the able and importantrole in the formation of
nervoussystem, whose vertices are the nerve many complex systems, which implies that
cells, connected by axons (3). But equally our results are relevant to a large class of
complex networks occur in social science, networks observed in nature.
where vertices are individuals or organiza- Although there are many systems that
tions and the edges are the social interactions form complex networks, detailed topological
between them (4), or in the WorldWide Web data is available for only a few. The collab-
(WWW), whose vertices are HTML docu- oration graph of movie actors represents a
ments connected by links pointing from one well-documented example of a social net-
page to another(5, 6). Because of their large work. Each actor is representedby a vertex,
size and the complexity of their interactions, two actors being connected if they were cast
the topology of these networks is largely together in the same movie. The probability
unknown. that an actorhas k links (characterizinghis or
Traditionally,networksof complex topol- her popularity)has a power-law tail for large
ogy have been described with the random k, following P(k) - k- Yactor,where yactor
graph theory of Erd6s and Renyi (ER) (7), 2.3 ? 0.1 (Fig. IA). A more complex net-
but in the absence of data on large networks, work with over 800 million vertices (8) is the
the predictions of the ER theory were rarely WWW, where a vertex is a documentand the
tested in the real world. However, driven by edges are the links pointing from one docu-
the computerizationof data acquisition,such ment to another. The topology of this graph
topological informationis increasinglyavail- determinesthe Web's connectivity and, con-
able, raising the possibility of understanding sequently, our effectiveness in locating infor- each vertex being connected to its two
the dynamical and topological stability of mation on the WWW (5). Informationabout
large networks. P(k) can be obtained using robots (6), indi-
Here we reporton the existence of a high cating that the probabilitythat k documents
degree of self-organizationcharacterizingthe point to a certainWeb page follows a power
large-scale properties of complex networks. law, with ywww= 2.1 ? 0.1 (Fig. iB) (9). A
Exploring several large databasesdescribing network whose topology reflects the histori-
the topology of large networks that span cal patternsof urbanand industrialdevelop-
fields as diverse as the WWW or citation ment is the electricalpower grid of the west-
patterns in science, we show that, indepen- ern United States, the vertices being genera-
dent of the system and the identity of its tors, transformers,and substations and the
constituents,the probabilityP(k) that a ver- edges being to the high-voltage transmission
tex in the network interacts with k other lines between them (10). Because of the rel-
vertices decays as a power law, following atively modest size of the network, contain-
P(k) - k-Y. This result indicates that large ing only 4941 vertices, the scaling region is
networksself-organize into a scale-free state, less prominent but is nevertheless approxi-
a feature unpredictedby all existing random mated by a power law with an exponent
_
networkmodels. To explain the origin of this 'Ypower 4 (Fig. IC). Finally, a ratherlarge
scale invariance,we show that existing net- complex network is formed by the citation
work models fail to incorporategrowth and patternsof the scientific publications,the ver- P(k) for the networks studied indicates that
preferentialattachment,two key features of tices being paperspublished in refereedjour-
real networks. Using a model incorporating nals and the edges being links to the articles
10'
es ~~~~~~~
0 0
~10 \10 6
3. X A t' B * , C els. First, both models assume that we start
10
6~~~~~~~~~~10l-
105 1093 10
Ne~~~~~~~~ 107 i o2
6~~~~~~~~~~~~~~ 60
- 10' -3
10 102 1n' 102 103 104 100
100 10' 103 100 101
k actors to the system, the WWW grows expo-
Fig.1. The distributionfunctionof connectivitiesforvariouslargenetworks.(A) Actorcollaboration nentially over
graph with N = 212,250 vertices and average connectivity (k) = 28.78. (B) WWW,N =
325,729, (k) = 5.46 (6). (C) Powergrid data, N = 4941, (k) = 2.67. The dashed lines have constantly grows by the publication of new
slopes (A) )Yactor = 2.3, (B) -ywW = 2.1 and (C) 'Ypower = 4
510 15 OCTOBER 1999 VOL 286 SCIENCE www.sciencemag.org
This content downloaded from 18.7.29.240 on Thu, 6 Feb 2014 19:43:52 PM
All use subject to JSTOR Terms and Conditions
cited in a paper. Recently Redner (11) has
shown that the probability that a paper is
cited k times (representingthe connectivityof
a paper within the network) follows a power
law with exponent ycite = 3.
The above examples (12) demonstratethat
many large randomnetworks share the com-
mon featurethatthe distributionof their local
connectivityis fiee of scale, following a power
law for large k with an exponent oybetween
2.1 and 4, which is unexpected within the
framework of the existing network models.
The randomgraphmodel of ER (7) assumes
thatwe startwith N vertices and connect each
pair of vertices with probabilityp. In the
model, the probability that a vertex has k
edges follows a Poisson distributionP(k) =
e- Xk/k!, where
A=Nz I -k
jpk( 1-p)N-
\k/
In the small-world model recently intro-
duced by Watts and Strogatz (WS) (10), N
vertices form a one-dimensional lattice,
nearest and next-nearest neighbors. With
probabilityp, each edge is reconnected to a
vertex chosen at random. The long-range
connections generated by this process de-
crease the distance between the vertices,
leading to a small-world phenomenon (13),
often referred to as six degrees of separa-
tion (14). Forp = 0, the probability distri-
bution of the connectivities is P(k) = 6(k -
z), where z is the coordination number in
the lattice; whereas for finite p, P(k) still
peaks around z, but it gets broader (15). A
common feature of the ER and WS models
is that the probability of finding a highly
connected vertex (that is, a large k) decreas-
es exponentially with k; thus, vertices with
large connectivity are practically absent. In
contrast, the power-law tail characterizing
highly connected (large k) vertices have a
large chance of occurring, dominating the
connectivity.
There are two generic aspects of real net-
works that are not incorporatedin these mod-
with a fixed number(N) of vertices that are
then randomlyconnected (ER model), or re-
connected (WS model), without modifying
N. In contrast,most real world networks are
open and they form by the continuous addi-
tion of new vertices to the system, thus the
number of vertices N increases throughout
the lifetime of the network.For example, the
actor network grows by the addition of new
time by the addition of new
Web pages (8), and the research literature
papers. Consequently, a common feature of
 REPORTS

these systems is that the networkcontinuous-
ly expands by the addition of new vertices
that are connected to the vertices already
present in the system.
Second, the random network models as-
sume thatthe probabilitythattwo vertices are
connected is random and uniform. In con-
trast, most real networks exhibit preferential
connectivity. For example, a new actor is
most likely to be cast in a supportingrole
with more established and better-knownac-
tors. Consequently,the probabilitythata new
actor will be cast with an established one is
much higher than that the new actor will be
cast with other less-known actors. Similarly,
a newly createdWeb page will be more likely
to include links to well-known popular doc-
uments with already-highcoinectivity, and a
new manuscriptis more likely to cite a well-
known and thus much-cited paper than its
less-cited and consequentlyless-known peer.
These examples indicate that the probability
with which a new vertex connects to the
existing vertices is not uniform; there is a
higher probabilitythat it will be linked to a
vertex that already has a large number of
connections.
We next show thata model based on these
two ingredients naturally leads to the ob-
served scale-invariantdistribution.To incor-
porate the growing characterof the network,
startingwith a smallnumber(io0) of vertices,
at every time step we add a new vertex with
/n(?m<770)edges that link the new vertex to in
different vertices alreadypresent in the sys-
tem. To incorporatepreferentialattachment,
we assume that the probabilityH7that a new
vertex will be connectedto vertex i depends
on the connectivityki of that vertex, so that
fl(ki) = kill kj. After t time steps, the
model leads to a random network with t +
m 0 vertices and 711t edges. This network
evolves into a scale-invariantstate with the
probabilitythat a vertex has k edges, follow-
ing a power law with an exponent Ymode
2.9 + 0.1 (Fig. 2A). Because the power law
obselved for real networksdescribes systems
of ratherdifferentsizes at different stages of
their development, it is expected that a cor-
that a new vertex is connected with equal
probabilityto any vertex in the system [that
is, fl(k') = couist = 1/1(i0 + t - 1)]. Such
a model (Fig. 2B) leads to P(k) -
exp(- k), indicating that the absence of
preferentialattachmenteliminates the scale-
free feature of the distribution.In model B,
we start with N vertices and no edges. At
each time step, we randomly select a vertex
and connect it with probabilityfl(k1) = kil
Ijkj to vertex i in the system. Although at
early times the model exhibits power-law
scaling, P(k) is not stationary:because N is
constant and the number of edges increases
with time, after T N2 time steps the system
reaches a state in which all vertices are con-
nected. The failure of models A and B indi-
cates thatboth ingredients growthand pref-
erential attachment are needed for the de-
velopment of the stationarypower-law distri-
bution observed in Fig. 1.
Because of the preferentialattachment,a
vertex that acquires more connections than
anotherone will increase its connectivity at a
higher rate; thus, an initial difference in the
connectivity between two vertices will in-
crease furtheras the networkgrows. The rate
at which a vertex acquires edges is aki/lt =
ki/2t, which gives k1i(t)= mII(t/ti)0.5, where
ti is the time at which vertex i was added to
the system (see Fig. 2C), a scaling property
that could be directly tested once time-re-
solved dataon networkconnectivitybecomes
available. Thus older (with smaller t,) verti-
ces increase their connectivity at the expense
of the younger (with larger ti) ones, leading
over time to some vertices that are highly
colunected, a "rich-get-richer"phenomenon
that can be easily detected in real networks.
Furthermore,this property can be used to
calculate y analytically. The probabilitythat
a vertex i has a connectivity smaller than k,
P[k,(t) < k], can be written as P(t1 >
III2tl/k2).Assunmingthat we add the vertices
to the system at equal time intervals, we
obtain P(ti > ;112tl/k2) = 1 - P(ti '
100 0
i11'tlk 2)
= 1 - ;112tl/k2(t + ino). The prob-
ability density P(k) can be obtained from
P(k) = aP[k1(t) < k]/ak, which over long
time periods leads to the stationarysolution
2i112
P(k) = k3
giving y = 3, independentof rn. Although it
reproduces the observed scale-free distribu-
tion, the proposedmodel cannot be expected
to account for all aspects of the studied net-
works. For that, we need to model these
systems in more detail. For example, in the
model we assumed linear preferentialattach-
ment; that is, H(k) k However, although
k.
in general H(k) could have an arbitrarynon-
linear folrn f(k) - ka, simulations indicate
that scaling is present only for a = 1. Fur-
thermore,the exponents obtainedfor the dif-
ferentnetworksare scatteredbetween 2.1 and
4. However, it is easy to modify our model to
account for exponents different from y = 3.
For example, if we assume thata fractionp of
the links is directed,we obtain y(p) = 3 -
p, which is supportedby numerical simula-
tions (16). Finally, some networksevolve not
only by adding new vertices but by adding
(and sometimes removing) connections be-
tween established vertices. Although these
and other system-specific features could
modify the exponent y, our model offers,the
first successful mechanismaccountingfor the
scale-invariantnatureof real networks.
Growth and preferential attachment are
mechanisms common to a number of com-
plex systems, including business networks
(17, 18), social networks(describingindivid-
uals or organizations), transportationnet-
works (19), and so on. Consequently, we
expect that the scale-invariantstate observed
in all systems for which detailed data has
been available to us is a generic propertyof
many complex networks, with applicability
reaching far beyond the quoted examples. A
better description of these systems would
help in understandingother complex systems
3

rect model should provide a distribution

whose main featuresare independentof time.
Indeed, as Fig. 2A demonstrates, P(k) is
independentof time (and subsequentlyinde- I 0~~~~~~b 0
pendentof the system size in0 + t), indicat- 10-2 ~0 :\102
1o 0 0 80
ing that despite its continuous growth, the
system organizes itself into a scale-free sta- -6
0~~~~~~1
tionaly state. 2 1 0 1 2 3 4 5
The development of the power-law scal- 10 ~ 0 0 4061030 0 01 1 1
ing in the model indicates that growth and
preferentialattachmentplay an importantrole k k t
in network development.To verify that both
Fig. 2. (A) The power-law connectivity distribution at t = 1 50,000 (0) and t = 200,000 (L) as
ingredients are necessary, we investigated obtained from the model, using mo = m = 5. The slope of the dashed line is y = 2.9. (B) The
two variantsof the model. Model A keeps the exponential connectivity distribution for model A, in the case of mo = m = 1 (O), MO = m =
growing characterof the network,but prefer- 3 (0), MO= m = 5 (O), and mo= m = 7 (A). (C) Time evolution of the connectivity for two
ential attachmentis eliminated by assuming vertices added to the system at t1 5 and t2 = 95. The dashed line has slope 0.5.

www.sciencemag.org SCIENCE VOL 286 15 OCTOBER 1999 511

This content downloaded from 18.7.29.240 on Thu, 6 Feb 2014 19:43:52 PM

All use subject to JSTOR Terms and Conditions
 REPORTS

as well, for which less topological informa-
tion is currently available, including such
importantexamples as genetic or signaling
networksin biological systems. We often do
not think of biological systems as open or
growing, because their featuresare genetical-
ly coded. However, possible scale-free fea-
tures of genetic and signaling networkscould
reflect the networks' evolutionaiy history,
dominatedby growth and aggregationof dif-
ferent constituents,leading from simple mol-
ecules to complex organisms. With the fast
advances being made in mappingout genetic
networks, answers to these questions might
not be too far away. Similar mechanisms
could explain the origin of the social and
economic disparities goveming competitive
systems, because the scale-free inhomogene-
ities are the inevitable consequence of self-
organizationdue to the local decisions made
by the individualvertices, based on informa-
tion that is biased toward the more visible
(richer) vertices, irrespective of the nature
and origin of this visibility.
References and Notes
1. R. Gallagher and T. Appenzeller, Science 284, 79
(1999); R. F. Service,ibid., p. 80.
2. G. Weng, U. S. Bhalta,R. Iyengar,ibid., p. 92.
3. C. Kochand G. Laurent,ibid., p. 96.
4. S. Wasserman and K. Faust, Social Network Analysis
(CambridgeUniv. Press,Cambridge,1994).
5. Members of the Clever project, Sci. Am. 280, 54
(June 1999).
6. R. Albert, H. Jeong, A.-L. Barabasi,Nature 401, 130
(1999); A.-L.Barabasi,R. Albert, H. Jeong, PhysicaA
272, 173 (1999); see also http://www.nd.edu/-net-
works.
7. P. Erd6sand A. Renyi,Publ. Math. Inst. Hung. Acad.
Sci. 5, 17 (1960); B. Bollobas,RandomGraphs(Aca-
demic Press, London,1985).
8. S. Lawrenceand C. L. Giles, Science 280, 98 (1998);
Nature 400, 107 (1999).
9. In addition to the distributionof incoming links,the
WWW displays a numberof other scale-free features
characterizing the organization of the Web pages
within a domain [B. A. Hubermanand L.A. Adamic,
Nature 401, 131 (1999)], the distributionof searches
[B. A. Huberman, P. L. T. Pirotli,J. E. Pitkow, R. J.
Lukose,Science 280, 95 (1998)], or the number of
links per Web page (6).
10. D. J. Watts and S. H. Strogatz, Nature 393, 440
(1998).
11. S. Redner,Eur.Phys.J. B 4, 131 (1998).
12. We also studied the neural network of the worm
Caenorhabditiselegans (3, 10) and the benchmark
diagram of a computer chip (see http://visicad.cs.
ucla.edu/-cheese/ispd98.html). We found that P(k)
for both was consistent with power-law tails, despite
the fact that for C. elegans the relativelysmall size of
the system (306 vertices) severely limits the data
quality, whereas for the wiring diagramof the chips,
vertices with over 200 edges have been eliminated
from the database.
13. S. Milgram,Psychol. Today 2, 60 (1967); M. Kochen,
ed., The Small World(Ablex, Norwood, NJ, 1989).
14. J. Guare,Six Degrees of Separation:A Play (Vintage
Books, New York,1990).
15. M. Barthelemyand L. A. N. Amaral,Phys. Rev. Lett.
82, 15 (1999).
16. For most networks, the connectivity m of the newly
added vertices is not constant. However, choosing m
randomly will not change the exponent y (Y. Tu,
personal communication).
17. W. B. Arthur,Science 284, 107 (1999).
18. Preferential attachment was also used to model
evolving networks (L. A. N. Amaral and M. Bar-
thelemy, personal communication).
19. J. R. Banavar,A. Maritan,A. Rinaldo,Nature 399, 130
(1999).
20. We thank D. J. Watts for providingthe C. elegans and
power grid data, B. C. Tjadenfor supplyingthe actor
data, H. Jeong for collecting the data on the WWW,
and L.A. N. Amaralfor helpful discussions.This work
was partiallysupported by NSF CareerAwaOdDMR-
9710998.
24 June 1999; accepted 2 September 1999

Osmium Isotope Constraints on suprasubductionxenolith locality, the Tubaf

seamount in the Lihir island group of the
Tabar-Lihir-Tanga-Feniisland arc in Papua
Ore Metal Recycling in New Guinea(Fig. 1). This xenolith locality is
important for the following reasons: (i) it
Subduction Zones contains samples that represent a complete
section of oceanic lithosphereat an intraoce-
Brent 1. A. McInnes,'* Jannene S. McBride,2 Noreen J. Evans,' anic convergent margin, (ii) it is located ad-
David D. Lambert,2 Anita S. Andrew3 jacent to one of the world's largest and
youngest volcano-hosted Au deposits, and
Veined peridotite xenoliths from the mantle beneath the giant Ladolamgold (iii) it containsmetasomatizedmantleperido-
deposit on LihirIsland, Papua New Guinea, are 2 to 800 times more enriched tite xenoliths with Au-enrichedvein minerals
in copper, gold, platinum,and palladiumthan surroundingdepleted arc mantle. that ciystallized in the mantle from oxidizing,
Gold ores have osmium isotope compositions similarto those of the underlying alkali- and sulfur-richhydrous fluids.
subduction-modified mantle peridotite source region, indicating that the pri- Duringthe oceanographicinvestigationof
mary origin of the metals was the mantle. Because the mantle is relatively submarine hydrothermalsystems in Papua
depleted in gold, copper, and palladium, tectonic processes that enhance the New Guinea (8), a submarine cinder cone
advective transport and concentration of these fluid soluble metals may be a (Tubaf volcano, 1280 m below sea level; 3?
prerequisite for generating porphyry-epithermalcopper-gold deposits. 15.25' S, 1520 32.50' E) was discovered 14
km southwestof the giant Ladolamgold mine
The tectonic relationshipbetween subduction- to 1870s) is a tracerof metallogenicprocessesat (>40 million oz contained Au) on Lihir Is-
relatedmagmatismat convergentmarginsand convergent margins because both elements land. Dredge and video-grab sampling of the
poiphyiy copper-gold(Cu-Au) ore fomiation have geochemicalpropertiessimilarto metals 1-km-diametervolcanic cone returned 130
has long been recognized (1). However, the that occur in poiphyryore deposits(2, 3). Be- ultramafic,mafic, and sediinentaiyxenoliths.
physical and chemical processes that govern cause Re is highly concentratedin crustalrocks The study of these samples has provided an
Cu-Au metallogenyand the ultimatesource(s) and Os is concentratedin the mantle(4, 5), this unprecedentedview of the source region of
of the metals in these ore deposits are poorly isotopic system is particularlyusefil for quan- an island arc magmaticsystem with a propen-
understood.The rhenium-osmium(Re-Os) iso- tifying the flux of ore elements in island arc sity to producegiant porphyry-epitheirmal ore
topic system (basedon the - decay of '87Re settingswhere the two principalreservoirsfor deposits. The xenolith assemblage includes
metals are subductedcruLst and mantle wedge spinel lherzolite, harzburgite,websterite, or-
1Commonwealth Scientific and Industrial Research peridotite. thopyroxenite, clinopyroxenite, syenite, ser-
Organization Exploration and Mining, Post Office Box Os isotope studies in subductionzones are pentinite, gabbro, hornblende gabbro, dia-
136, North Ryde, New South Wales 1670, Australia; currently limited because of the rarity of base, basalt, pelagic deep-sea sediment, and
2Department of Earth Sciences, Monash University, deep-seated suprasubductionsamples (rocks shallow-watervolcaniclasticsedimentas well
Clayton, Victoria 3168, Australia; 3Commonwealth
Scientific and Industrial Research Organization Petro-
overlying a subductedslab). Previous sttudies as coralline and coralgal limestone. These
leum Resources, Post Office Box 136, North Ryde, (6, 7) have demonstratedthat radiogenic Os lithologiesrepresenta cross section of the su-
New South Wales 1670, Australia. is introducedinto the subarc mantle by hy- prasubduction assemblageand can be reassem-
*To whom correspondence should be addressed. E- drous, oxidizing fluids derived during slab bled into an "ophiolite-type" model of oceanic
mail: Brent.Mclnnes@dem.csiro.au dehydration. We report on results from a lithosphere(Fig. 1) (9).

512 15 OCTOBER 1999 VOL 286 SCIENCE www.sciencemag.org

This content downloaded from 18.7.29.240 on Thu, 6 Feb 2014 19:43:52 PM

All use subject to JSTOR Terms and Conditions