How Small Are Building Blocks of Complex Networks

Almerima Jamakovic,1 Priya Mahadevan,2 Amin Vahdat,3 Marian Bogu

na,4 and Dmitri Krioukov5

arXiv:0908.1143v1 [physics.soc-ph] 8 Aug 2009

1
TNO Information and Communication Technology,
Netherlands Organisation for Applied Scientific Research,
P.O. Box 5050, 2600 GB Delft, The Netherlands
2
HP Labs, 1501 Page Mill Rd, Palo Alto, California, 94304, USA
3
Department of Computer Science and Engineering, University of California,
San Diego (UCSD), 9500 Gilman Drive, La Jolla, California 92093, USA
4
Departament de Fsica Fonamental, Universitat de Barcelona, Mart i Franqu`es 1, 08028 Barcelona, Spain
5
Cooperative Association for Internet Data Analysis (CAIDA), University of California,
San Diego (UCSD), 9500 Gilman Drive, La Jolla, California 92093, USA

Network motifs are small building blocks of complex networks, such as gene regulatory networks.
The frequent appearance of a motif may be an indication of some network-specific utility for that
motif, such as speeding up the response times of gene circuits. However, the precise nature of
the connection between motifs and the global structure and function of networks remains unclear.
Here we show that the global structure of some real networks is statistically determined by the
distributions of local motifs of size at most 3, once we augment motifs to include node degree
information. That is, remarkably, the global properties of these networks are fixed by the probability
of the presence of links between node triples, once this probability accounts for the degree of the
individual nodes. We consider a social web of trust, protein interactions, scientific collaborations,
air transportation, the Internet, and a power grid. In all cases except the power grid, random
networks that maintain the degree-enriched connectivity profiles for node triples in the original
network reproduce all its local and global properties. This finding provides an alternative statistical
explanation for motif significance. It also impacts research on network topology modeling and
generation. Such models and generators are guaranteed to reproduce essential local and global
network properties as soon as they reproduce their 3-node connectivity statistics.

I.

INTRODUCTION

A promising direction in the studies of the structure

and function of complex networks is to identify their
building blocks, or motifs [13], which are small subgraphs in a real network. A great deal of researchin
particular, research on gene regulatory networksshows
that specific motifs perform specific functions, such as
speeding up response times of regulatory networks [4, 5].
However, motifs have also raised many questions [613],
including continuing debates on whether and how motif statistical profiles are related to the global structure,
function, and evolution of certain networks.
Our recent work [14] introduces dK-series, see Section II. The dK-series, with analogy to the Taylor or
Fourier series, is the first systematic and complete basis
for characterizing network structure. The dK-series is a
generalization of known degree-based statistical characteristics of complex networks. The zero-th element of the
dK-series, the 0K-distribution, is the average degree in
a given network. The first element, the 1K-distribution,
is the networks degree distribution, or the number of
nodessubgraphs of size 1of degree k. The second element, the 2K-distribution, is the joint degree distribution, the number of subgraphs of size 2linksinvolving
nodes of degrees k1 and k2 . For d = 3, the subgraphs
are triangles and wedges, composed of nodes of degrees
k1 , k2 , and k3 . Generalizing, the dK-distribution is the
numbers of different subgraphs of size d involving nodes
of degrees k1 , k2 , . . . , kd .

The dK-series is systematic and complete because it

is inclusive and converging. Inclusiveness results from
the fact that the (d + 1)K-distribution contains the same
information about the network as the dK-distribution,
plus some additional information. That is, by increasing
d, we provide increasingly more detail about the network structure. As d increases toward the network size,
we fully specify the entire network structure, which explains the second convergence property of dK-seriesit
converges to the given network in the limit of large d.
Does this convergence happen only at d equal to the
network size, or much sooner, at smaller d? In other
words, how much local information, i.e., information
about concentrations of degree-labeled subgraphs of what
size, is needed to fully capture global network structure?
To answer these questions, we must compare a real
network with typical random networks defined by its dKdistribution. If there is no difference between such dKrandom networks and the real network, then the latter
is fixed by its dK-distribution. To obtain a dK-random
version of the real network, we dK-randomize it as illustrated in Fig. 1(a)we randomly rewire (pairs of) links
preserving the dK-distribution in the network, generalizing known network randomization techniques [17, 18]
used to compute motif statistical significance. The result of this dK-randomization procedure are random networks that have the same dK-distribution as the original
real network, but that are maximally random in all other
respects.
Our question thus becomes what is the minimum value
of d such that there is no difference between a real net-

FIG. 1: The dK-randomization null models for d = 0, 1, 2, 3. a) Illustration of dK-randomizing rewiring. All nodes
are labeled by their degrees, and a dK-rewiring preserves the graphs dK-distribution, and consequently its d0 K-distributions
for all d0 < d, but randomizes the d00 K-distributions for d00 > d. The 0K-randomization involves rewiring of a link to any pair
of disconnected nodes, which preserves the average degree only. The 1K-randomization preserves the degree distribution, too,
by rewiring a pair of links as shown. The 2K-distribution preserves the joint degree distribution as well, because at least two
nodes adjacent to the rewired links are of the same degree. The 3K-randomization preserves the number of degree-labeled
wedges and triangles. As d increases, the rewiring becomes increasingly more constrained since fewer links can be rewired
without altering the dK-distribution. To dK-randomize a network, we randomly select a pair of links, and rewire them if they
can be dK-rewired, or, if they cannot be rewired, select another random pair. This process is repeated for a sufficient number
of successful rewirings, i.e., until all network properties stop changing, at which point we say that the graph has converged to
its dK-randomization. b) Visualization of the social web of trust (PGP network [15]) and its dK-randomizations.
We use the LaNet-vi tool [16] for visualization, which encodes the node coreness in color, see the right legends. The coreness is
a measure of node centrality, i.e., how deeply in the network core the node lies [16]. Nodes with larger coreness are also placed
closer to the circle centers. The quick convergence of the dK-randomizations to the original PGP network, and the similarity
between it and its 3K-randomization are remarkable.

work and its dK-randomizations? It seems at first that

the answer to this question should strongly depend on
the specific networks we consider.
We consider a variety of social, biological, transportation, communication, and technological networks, see
Section III. Although the dK-series applies to directed
and even annotated networks [19], here we report results
for undirected networks. The dK-distributions for directed or annotated networks contain more information
than for undirected networks. Therefore, dK-series converges faster in the former case [19]. Below we show
the results for the well-studied social web of trust relationships extracted from Pretty Good Privacy (PGP)
data [15]. The results for all other networks, except the
power grid, are similar, cf. Section IV, where we also discuss possible reasons for why the power grid appears as
an exception.

Fig. 1(b) visualizes the PGP network and its dKrandomizations. We observe that the dK-series converges
at d = 3. While the 0K-random network has little in
common with the real network, the 1K-random one is
somewhat more similar, even more so for 2K, and there
is very little difference between the real PGP network
and its 3K-random counterpart.
To provide a more detailed and insightful comparison
between the real network and its dK-randomizations, we
compute a variety of metrics for each. Some popular metrics, such as degree distribution, average nearest neighbor
connectivity, clustering, etc., are functions, sometimes
peculiar, of dK-distributions, and therefore it is not surprising that they are properly captured by dK-series, as
confirmed in Section IV A. We classify metrics that do
not explicitly depend on dK-distributions as microscopic,
mesoscopic, and macroscopic. We choose them to probe

Distribution of motifs for PGP Web of Trust .

0.8

0.6

0.4

10

10

10

0.2

1.E+06

1.E+05

PGP
0K-randomization
1K-randomization
2K-randomization
3K-randomization

10 0
10

Z-score for PGP Web of Trust

10

community size

PGP Web of Trust

3K - randomization
2K - randomization
1K - randomization
0K - randomization

3K - randomization
2K - randomization
1K - randomization
0K - randomization

1.E+04

1.E+03

10

rank

10

FIG. 3: Mesoscopic scale: community structure. We

compute communities in the PGP network using the Extremal
Optimization algorithm [20]. We then sort the found communities in the order of decreasing size. The size of a community
is the number of nodes in it. The rank of a community is its
position number in the size-ordered list. We then show the
community size distribution by plotting the community sizes
vs. their ranks.

1.E+02

1.E+01

1.E+00

FIG. 2: Microscopic scale: motifs. There are six different

graphs of size 4 shown on the x-axes. The top plot shows the
distribution of the numbers of these subgraphs in the PGP
network and its dK-randomizations, d = 0, 1, 2, 3. Each blue
bar, for example, is the number of the corresponding subgraph
occurrences in the PGP network divided by the total number
of subgraphs of size 4 in it. For dK-randomizations, the values are averaged, for each d, over several realizations of the
dK-randomized network. In the case of 0K-randomization,
the last two motifs did not occur in any randomized sample
of the network. The bottom plot shows the Z-scores for the
six subgraphs in the four dK-randomization null models. The
Z-score [1] of a subgraph is a measure of its statistical significance in a real network, compared to a randomization null
model. Specifically, the Z-score Z is the difference between
the number N of the occurrences of a subgraph in the real
of its occurrences in the
network and the average number N
corresponding randomized networks, divided by the standard
deviation of its occurrences in the randomized networks,
|/.
Z = |N N

the network structure at the local, medium, and global

scales.
The simplest microscopic, local-structure statistics,
which are not fixed by the dK-distributions with d 6 3,
are the frequencies of motifs of size 4 without degree information. We compute these frequencies in the real network and its dK-randomizations, and show the results in

Fig. 2. We find that the (relative) statistical significance

of the motifs strongly depends on d. More importantly,
no motif is statistically significant for d = 3.
At the mesoscopic scale, we consider the community
structure of the PGP network. A community is a subgraph with many internal connections, and a relatively
small number of connections external to the subgraph.
Fig. 3 shows that the community structure is indeed a
mesoscopic metric because the community sizes range
from a few nodes to thousands of nodes for largest communities. Fig. 3 shows that the community size distributions in the PGP network and its 3K-randomization are
very similar.
At the macroscopic scale, we consider the two most
popular and important statistics that depend on a networks global structure: the node betweenness centrality and the distribution of lengths of shortest paths in a
network. Fig. 4 once again shows that 3K is sufficient
to capture even global graph properties; the considered
metrics are approximately the same for the PGP network
and its 3K-randomization.
We call a given real network dK-random if all its
metrics, at all scales from local to global, are approximately the same as the corresponding metrics in its dKrandomizations. We see in Section IV that in agreement
with the results of Vazquez et al. [12], almost all networks that we collected data for are 3K-random at most
(some networks are 2K- or even 1K-random). That is,
surprisingly, the global structure of these networks is captured entirely by the distribution of node triples and their
degrees.
It is an open question why many different real net-

4
1.E+04
Mean Distance dK-random
Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

Macroscopic statistics for

PGP Web of Trust

1.E+03

1.E+02

1.E+01

1.E+00

1.E-01
0K-random

1.E-01

1K-random

2K-random

3K-random

Mean Betweenness dK-random

Mean Betweenness real-world network
StDev Betweenness dK-random

Macroscopic statistics for

PGP Web of Trust

StDev Betweenness real-world network

1.E-02

1.E-03

1.E-04
0K-random

1K-random

2K-random

3K-random

FIG. 4: Macroscopic scale: the distance and betweenness distributions. The top plot shows the metrics related to the hop length of shortest paths, or distances, between nodes in the PGP network and its dK-randomizations.
These metrics are the average and maximum distance between
nodes, the latter called the network diameter, and the standard deviation of the distance distribution. The bottom plot
shows the average betweenness and the standard deviation of
the betweenness distribution of nodes in the PGP network
and its dK-randomizations. The betweenness of a node is a
measure of its communication centrality [21]. It is equal to the
number of shortest paths passing through the node, divided
by the total number of shortest paths between the same source
and destination, summed over all source-destinations pairs.
In both plots the values for dK-randomizations are averaged,
for each d, over several realizations of the dK-randomized
network.

works are 3K-random. A trivial answer would be that

d = 3 is just constraining enough. There may only be
a few possible rewirings preserving the 3K-distribution.
But why exactly is d = 3 sufficient for real networks?
There are many classes of synthetic graphs, such as latices, for which no d substantially smaller than the graph
size is constraining enough. Perhaps the answer can be
obtained by studying the hidden metric spaces underlying real networks [22]. The distances in such spaces abstract intrinsic similarities between nodes. If these spaces
are metricand there is empirical evidence that they are
indeed such [23]then the triangle inequality naturally
yields and explains network clustering, which the 3K-

distribution captures by definition.

Whatever the actual explanation, our results have diverse implications. First, our dK-randomization basis
makes it clear that there is no preferred null model for
network randomization. To tell how statistically important a given motif is, it is necessary to compare its frequency in the real network with the same frequency in a
network randomization, a null model. But one can dKrandomize any network for any d. Therefore choosing
any specific value of d, or more generally, any specific
null model to compute motif significance requires some
non-trivial justification.
The second implication concerns the difference between motifs and dK-series. This difference is small but
crucial. Motifs are subgraphs whose nodes can have any
degree in the original network, while dK-series preserves
the information about these degrees. This difference is
crucial because a motif-based series cannot be inclusive.
Node degrees are necessary to make the series inclusive
and thus systematic, see Section V.
Our finding that many networks are 3K-random can
assist our understanding of how functions of an evolving network shape its structure. Indeed, one can potentially simplify such explanations to how the observed
3K-distribution has emerged in the network. As soon
as one explains the emergence of the 3K-distribution, all
other network structural properties follow.
Finally, our work very practically impacts the design
of network topology models and generators. For simulation experiments, hypothesis testing, etc., network researches in many sciences, including biology [9, 2426]
and computer science [2730], must model real networks
in laboratory settings, and generate random graphs that
reproduce important properties of the real network. Our
results show that it is sufficient to generate 3K-random
graphs for such purposes. But even if these graphs do not
capture some important property not previously considered, the dK-series will remain applicable given its convergence property and a sufficient increase in d.
We conclude this introduction with a reference to [19]
for a detailed discussion of various graph generation techniques based on dK-series and extensions to generate random graphs with rich semantic, structural, or functional
annotations of nodes and links.

II.

THE dK-SERIES ILLUSTRATED

In Fig. 5(a) we illustrate dK-series for a graph of size

4. The 4K-distribution is the graph itself. The 3Kdistribution consists of its three subgraphs of size 3: one
triangle connecting nodes of degrees 2, 2, and 3, and two
wedges connecting nodes of degrees 2, 3, and 1. The
2K-distribution is the joint degree distribution in the
graph. It specifies the number of links (subgraphs of size
2) connecting nodes of different degrees: one link connects nodes of degrees 2 and 2, two links connect nodes
of degrees 2 and 3, and one link connects nodes of degree

FIG. 5: The dK-series illustrated: a) dK-distributions for a graph of size 4; b) convergence and inclusiveness of dK-series.

3 and 1. The 1K-distribution is the degree distribution in

the graph. It lists the number of nodes (subgraphs of size
1) of different degree: one node of degree 1, two nodes of
degree 2, and one node of degree 3. The 0K-distribution
is just the average degree in the graph, which is 2.
Fig. 5(b) illustrates the inclusiveness and convergence
of dK-series by showing the hierarchy of dK-graphs,
which are graphs that have the same dK-distribution as
some graph G of size n. The black circles schematically
shows the sets of dK-graphs.
The set of 0K-graphs is largest: the number of different graphs that have the same average degree as G
is enormous. These graphs may have a structure drastically different from Gs. The set of 1K-graphs is a subset
of 0K-graphs, because each graph with the same degree
distribution as in G has also the same average degree as
G, but not vice versa. As a consequence, typical (maximally random) 1K-graphs tend to be more similar to
G than 0K-graphs. The set of 2K-graphs is a subset of
1K-graphs, also containing G.
As d increases, the circles become smaller because the
number of different dK-graphs decreases. Since all the
dK-graph sets contain G, the circles zoom-in on it,
and while their number decreases, dK-graphs become in-

creasingly more similar to G. In the d = n limit, the set

of nK-graphs consists of only one element, G itself.

III.

THE REAL NETWORKS CONSIDERED

We collected data for a number of real networks. We

wanted the set of considered networks to be representative, in the sense that it should contain networks of different nature, coming from different domains, thus showing
the universality of our dK-basis. The considered networks include social, biological, transportation, and technological networks. Specifically, we report results for:
The social web of trust relationships among people.
The trust relationships are inferred using the data
from the Pretty Good Privacy (PGP) encryption
algorithm [15]. We extract the strongly connected
component from this network. The nodes are people, and there is a link between two people if they
trust each other.
The social network of scientific collaborations
extracted from the arXiv condensed-matter

6
1.

TABLE I: The considered networks and their abbreviations.

Network

Abbreviation

PGP Web of Trust [15]

Scientific collaboration network [31]
Protein interaction network [32]
US air transportation network [33]
Internet at the level of ASs [34]
Power grid in the western US [35]

PGP
Collab.
Protein
Air
Internet
Power

database [31]. The nodes are authors, and there is

a link between two authors if they co-authored a
paper.
The biological network of protein interactions in the
yeast Saccharomyces cerevisiae collected from the
database of interacting proteins [32]. The nodes are
proteins, and there is a link between two proteins
if they interact.
The US air transportation network [33]. The nodes
are airports, and there is a link between two airports if there is a direct flight between them.
The topology of the Internet at the level of Autonomous Systems (ASs) [34]. The nodes are ASs,
i.e., organizations owing parts of the Internet infrastructure, and there is a link between two ASs if
they are physically connected.
The electrical power grid in the western US [35].
The nodes are generators, transformers, or substations, two of which are linked if there is a highvoltage transmission line between them.
Table I lists these networks and their abbreviations used
in the subsequent figures and tables.

IV.

TOPOLOGIES OF REAL NETWORKS AND

THEIR dK-RANDOMIZATIONS

In this section we compare the real networks to their

dK-randomizations across a number of topological metrics.

1K: degree distribution

Fig. 6 shows the distributions P (k) of node degrees k:

P (k) =

N (k)
,
N

(1)

where N (k) is the number of nodes of degree k in the

network, and N is the total
P number of nodes in it, so
that P (k) is normalized, k P (k) = 1 (we do not consider nodes of degree k = 0). The 1K-distribution fully
defines the 0K-distribution, i.e., the average degree k in
the network, by
X
k =
kP (k),
(2)
k

but not vice versa.

We observe in Fig. 6 that while 0K-randomizations are
off, the 1K-random graphs reproduce the degree distributions in the real networks exactly, which is by dentition:
the 1K-distribution is the degree distribution, and 1Krandomization does not alter it. The dK-randomizations
with d > 1 do not alter the 1K-distribution either, therefore they also match the degree distributions in the real
networks exactly (not shown).
2.

2K: average neighbor degree

Fig. 7 shows the average degree knn (k) of neighbors

of nodes of degree k. This function is a commonly used
projection of the joint degree distribution (JDD) P (k, k 0 ),
i.e., the 2K-distribution. The JDD is defined as
P (k, k 0 ) = (k, k 0 )

N (k, k 0 )
,
2M

(3)

where N (k, k 0 ) = N (k 0 , k) is the number of links between

nodes of degrees k and k 0 in the network, M is the total
number of links in it, and
(
2 if k = k 0 ,
0
(k, k ) =
(4)
1 otherwise,
P
so that P (k, k 0 ) is normalized, k,k0 P (k, k 0 ) = 1. The
2K-distribution fully defines the 1K-distribution by
P (k) =

k X
P (k, k 0 ),
k 0

(5)

A.

Metrics defined by dK-distributions

We first consider the most basic metrics, which are defined by the appropriate dK-distributions. Therefore it
is not surprising that dK-random graphs with appropriate d have the values of these metrics equal exactly to
those in the real networks. Nevertheless, we report these
results for consistency and illustration purposes.

but not vice versa. The average neighbor degree knn (k)
is a projection of the 2K-distribution P (k, k 0 ) via
P 0
0
k X 0
0 k P (k, k )
knn (k) =
k P (k, k 0 ) = Pk
. (6)
0
kP (k) 0
k0 P (k, k )
k

We observe in Fig. 7 that while 0K-randomizations are

way off, the 1K-randomization are much closer to the

degree distribution

10

10
PGP
0K
1K

10

10

10

10

10

10

10

10

degree distribution

10

10

10

10

10

10

10

10
Collab.
0K
1K

10

Internet
0K
1K

10

10
4

10

10

10

10

10

10

10

10

degree distribution

Protein
0K
1K

10

10

10

10

10

10

10
Air
0K
1K

10

10

10

Power
0K
1K

10

10

10

10

10

10

10
degree

degree

10

FIG. 6: The degree distribution in the real networks and their dK-randomizations.

aver. neigb. degree

10

10

10
0
10

PGP
0K
1K
2K

10

10

10

Protein
0K
1K
2K

10

10

10
0
10

10

10

10

10

aver. neigb. degree

10

10

Collab.
0K
1K
2K

10

10

10
0
10

10

10

10

aver. neigb. degree

10
0
10

10

10

10

10

10

10

Air
0K
1K
2K

10

10
0
10

Internet
0K
1K
2K

10

10

10
degree

Power
0K
1K
2K

10

10
0
10

10
degree

10

FIG. 7: The average neighbour degree in the real networks and their dK-randomizations.

aver. clustering coeff.

10

10
PGP
0K
1K
2K
3K

10

10

10

Protein
0K
1K
2K
3K

10

10

10

10

10

10

10

10

10

aver. clustering coeff.

10

10

Collab.
0K
1K
2K
3K

10

10

Internet
0K
1K
2K
3K

10

10

10

10

10

10

10

10

10

10

aver. clustering coeff.

10

10

Air
0K
1K
2K
3K

10

10

10

Power
0K
1K
2K
3K

10

10

10
degree

10

10

10

10
degree

10

FIG. 8: The degree-dependent clustering in the real networks and their dK-randomizations.

9
real networks, whereas the 2K-randomizations have exactly the same average neighbor degrees as the real networks, which is again by definition: 2K-randomization
does not change P (k, k 0 ). In the Internet case, even 1Krandomization does not noticeably affect knn (k). The
dK-randomizations with d > 2 do not alter P (k, k 0 ) and
consequently knn (k) at all, therefore they reproduce the
latter exactly as well for all the networks (not shown).
3.

3K: clustering

Fig. 8 shows degree-dependent clustering c(k). Clustering of node i is the number of triangles 4i it forms, or
equivalently the number of links among its neighbors, divided by the maximum such number, which is k(k 1)/2,
where k is is degree, deg(i) = k. Averaging over all
nodes of degree k, the degree-dependent clustering is
X
24(k)
4i . (7)
c(k) =
, where 4(k) =
k(k 1)N (k)
i: deg(i)=k

The degree-dependent clustering is a commonly

used projection of the 3K-distribution [38]. The 3Kdistribution is actually two distributions characterizing
the concentrations of the two non-isomorphic degreelabeled subgraphs of size 3, wedges and triangles:
k

k'

k''

00

3T + W + M
= k2 .
N

(12)

The degree-dependent clustering coefficient c(k) is the

following projection of the 3K-distribution
P
0 00
6T k0 ,k00 P4 (k, k , k )
c(k) =
.
(13)
N
k(k 1)P (k)
We observe in Fig. 8 that clustering in the real networks and their dK-randomizations with d = 3 is exactly
the same, which is again by definition. For d < 3, clustering differs drastically in many cases, except for the
air transportation network and especially the Internet.
Therefore we can say that the Internet is very close to
being 1K-random, i.e., fully defined by its degree distribution, as far as the dK-based metrics are concerned.
Neither 3K-, 2K-, nor even 1K-randomization alter its
dK-based (projection) metrics noticeably.

B.

Motifs and their Z-scores

k'

k''

.
Let N (k , k, k ) = N (k , k, k ) be the number wedges
involving nodes of degrees k, k 0 , and k 00 , where k is the
central node degree, and let N4 (k, k 0 , k 00 ) be the number
of triangles consisting of nodes of degrees k, k 0 , and k 00 ,
where N4 (k, k 0 , k 00 ) is assumed to be symmetric with
respect to all permutations of its arguments. Then the
two components of the 3K-distribution are
0

The normalization of 2K- and 3K-distributions implies

the following identity between the numbers of triangles,
wedges, edges, nodes, and
P the second moment of the degree distribution k2 = k k 2 P (k):

00

N (k 0 , k, k 00 )
,
(8)
2W
0 00
N4 (k, k , k )
P4 (k, k 0 , k 00 ) = (k, k 0 , k 00 )
,
(9)
6T
where T and W are the total numbers of triangles and
wedges in the network, and

0
00

6 if k = k = k ,
(k, k 0 , k 00 ) = 1 if k 6= k 0 6= k 00 ,
(10)

2 otherwise,
P (k 0 , k, k 00 ) = (k 0 , k 00 )

0
00
0 00
so that
normalP both P (k ,0 k, k 00) andPP4 (k, k , k ) are
ized, k,k0 ,k00 P (k , k, k ) = k,k0 ,k00 P4 (k, k 0 , k 00 ) = 1.
The 3K-distribution defines the 2K-distribution (but not
vice versa), by
X  6T
1
0
P (k, k ) =
P4 (k, k 0 , k 00 )
k + k 0 2 00 M
k

W
0
00
0 00
+
[P (k , k, k ) + P (k, k , k )] . (11)
M

There are six non-isomorphic motifs of size 4, shown

as the x-axes in Figs. 9,10. For each network and for
each d = 0, 1, 2, 3, we obtain several dK-randomized samples of the network, and then for each motif we compute
its distribution (normalized to the total number of subgraphs of size 4) in the real network, and its average
distribution in the dK-randomized samples of the network. The results are in Fig. 9. Fig. 10 reports the
corresponding Z-scores. In certain cases, often for 0Krandomizations, some motifs do not occur at all in any
randomized samples, which explains the absence of some
bars in the figures.
The key observation is that when the randomization
null model is 3K, the distributions of all motifs in the
randomizations of all the networks except the power grid,
are close to those in the real networks. The corresponding
Z-scores are either low or zero. In other words, all motifs
are statistically non-significant.

C.

Distance and betweenness distributions

Fig. 11 shows the distance distribution in the real networks and in their dK-randomizations. The distance distribution is the distribution of hop-lengths of shortest
paths between nodes in a network. Formally, if N (h) is
the number of node pairs located at hop distance h from
each other, then the distance distribution (h) is
(h) =

2N (h)
,
N (N 1)

(14)

10

0.4

0.2

0.6

0.4

0.2

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2

0.8

0.6

0.4

0.2

Air Transportation
3K - randomization
2K - randomization
1K - randomization
0K - randomization

Distribution of motifs for Air Transportation

Internet AS-level
3K - randomization
2K - randomization
1K - randomization
0K - randomization

Distribution of motifs for Internet AS-level

Distribution of motifs for Scientific Collaborations

0.6

0.8

Scientific Collaborations
3K - randomization
2K - randomization
1K - randomization
0K - randomization

0.8

0.6

0.4

0.2

Power Grid
3K - randomization
2K - randomization
1K - randomization
0K - randomization

Distribution of motifs for Power Grid

Distribution of motifs for PGP Web of Trust

0.8

Protein Interactions
3K - randomization
2K - randomization
1K - randomization
0K - randomization

Distribution of motifs for Protein Interactions

PGP Web of Trust

3K - randomization
2K - randomization
1K - randomization
0K - randomization

0.8

0.6

0.4

0.2

0.1
0

FIG. 9: The motif distributions in the real networks and their dK-randomizations.

1.E+06

1.E+05

Z-score for Protein Interactions

Z-score for PGP Web of Trust

1.E+05

1.E+04

1.E+03

1.E+02

1.E+04

1.E+03

1.E+02

1.E+01

1.E+01

1.E+00

1.E+00

1.E+09
1.E+08

3K - randomization
2K - randomization
1K - randomization
0K - randomization

1.E+06

1.E+06
1.E+05
1.E+04
1.E+03
1.E+02

Z-score for Air Transportation

Z-score for Internet AS-level

1.E+05
1.E+07

3K - randomization
2K - randomization
1K - randomization
0K - randomization

1.E+06

Z-score for Scientific Collaborations

3K - randomization
2K - randomization
1K - randomization
0K - randomization

3K - randomizations
2K - randomization
1K - randomization
0K - randomization

1.E+05

1.E+04

1.E+03

1.E+02

1.E+01

1.E+00

3K - randomization
2K - randomization
1K - randomization
0K - randomization

1.E+04

1.E+03

1.E+02

3K - randomization
2K - randomization
1K - randomization
0K - randomization

1.E+03

Z-score for Power Grid

1.E+06

1.E+02

1.E+01

1.E+01
1.E+01
1.E+00

1.E+00

1.E+00

FIG. 10: The motif Z-scores in the real networks and their dK-randomizations.

where N (N 1)/2 is the total number of nodes pairs in

the network.

statistics are the average distance

X
=
h
h(h),

(15)

To provide a clearer view of how close the distance distributions in dK-randomizations are to the real networks,
we show in Fig. 12 some scalar summary statistics of the
distance distribution as functions of d. These summary

and the standard deviation of the distance distribution

(h). In addition we show in Fig. 12 the network diameter, i.e., the maximum hop-wise distance between nodes
in the network, which is an extremal statistics of the dis-

distance distribution

distance distribution

distance distribution

11

0.4

0.2

10

15

20

0.2
0

25
Collab.
0K
1K
2K
3K

0.4
0.2

10

15

10

15

Internet
0K
1K

0.6
0.4
0.2
0

20

0.8

0.4
Air
0K
1K
2K
3K

0.6
0.4
0.2
0

Protein
0K
1K
2K
3K

0.4

0.8

0.6

0.6

PGP
0K
1K
2K
3K

Power
0K
1K
2K
3K

0.3
0.2
0.1
0

10

10

10

distance

15
distance

20

25

30

FIG. 11: The distance distribution in the real networks and their dK-randomizations.

1.E+03
Mean Distance dK-random
Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

Macroscopic statistics for

PGP Web of Trust

1.E+03

1.E+02

1.E+01

Macroscopic statistics for

Protein Interactions

1.E+04

Mean Distance dK-random

Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

1.E+02

1.E+01

1.E+00

1.E+00

1.E-01
0K-random

1K-random

2K-random

1.E-01
0K-random

3K-random

1.E+04

2K-random

3K-random

1.E+02

1.E+01

Macroscopic statistics for

Internet AS-level

1.E+03
Mean Distance dK-random
Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

1.E+03

Macroscopic statistics for

Scientific Collaborations

1K-random

Mean Distance dK-random

Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

1.E+02

1.E+01

1.E+00
1.E+00

1.E-01
0K-random

1K-random

2K-random

1.E-01
0K-random

3K-random

2K-random

3K-random

1.E+04

1.E+03
Mean Distance dK-random
Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

1.E+02

1.E+01

1.E+00

Mean Distance dK-random

Mean Distance real-world network
StDev Distance dK-random
StDev Distance real-world network
Diameter dK-random
Diameter real-world network

1.E+03

Macroscopic statistics for

Power Grid

Macroscopic statistics for

Air Transportation

1K-random

1.E+02

1.E+01

1.E+00

1.E-01

1.E-01
0K-random

1K-random

2K-random

3K-random

1.E-02
0K-random

1K-random

2K-random

3K-random

FIG. 12: The average distance, the standard deviation of the distance distribution, and the network diameter as functions of
d for dK-randomisations of the real networks. The corresponding values for the real networks are shown by dashed lines.

average betweenness

12

10

10

10

10

average betweenness

10

10

10
10

10

10

10

10

10

10

10
Collab.
0K
1K
2K
3K

10

6
0

10

10

10

Internet
0K
1K
2K
3K

10

10

10

10

10

10

10

10

10

5
0

10

10

10

Power
0K
1K
2K
3K

Air
0K
1K
2K
3K

10

10

10
2

10

10

Protein
0K
1K
2K
3K

10
average betweenness

PGP
0K
1K
2K
3K

10

10

10

10

10

10

degree

10
degree

10

FIG. 13: The average betweenness of nodes of a given degree in the real networks and their dK-randomizations.

1.E-01

1.E-01

Mean Betweenness dK-random

Mean Betweenness dK-random

Mean Betweenness real-world network

Mean Betweenness real-world network

StDev Betweenness dK-random

StDev Betweenness dK-random

StDev Betweenness real-world network

Macroscopic statistics for

Protein Interactions

Macroscopic statistics for

PGP Web of Trust

StDev Betweenness real-world network

1.E-02

1.E-03

1.E-04
0K-random

1K-random

2K-random

3K-random

1.E-01

1.E-02

1.E-03

1.E-04
0K-random

1K-random

Mean Betweenness dK-random

Mean Betweenness real-world network

Mean Betweenness real-world network

StDev Betweenness dK-random

StDev Betweenness dK-random

StDev Betweenness real-world network

Macroscopic statistics for

Internet AS-level

Macroscopic statistics for

Scientific Collaborations

1.E-02

1.E-03

1K-random

2K-random

1.E-02

1.E-03

1.E-04
0K-random

3K-random

1.E+00

1K-random

3K-random

Mean Betweenness dK-random

Mean Betweenness dK-random

Mean Betweenness real-world network

Mean Betweenness real-world network

StDev Betweenness dK-random

StDev Betweenness dK-random

StDev Betweenness real-world network

Macroscopic statistics for

Power Grid

Macroscopic statistics for

Air Transportation

2K-random

1.E-01

StDev Betweenness real-world network

1.E-01

1.E-02

1.E-03
0K-random

3K-random

Mean Betweenness dK-random

StDev Betweenness real-world network

1.E-04
0K-random

2K-random

1.E-01

1K-random

2K-random

3K-random

1.E-02

1.E-03

1.E-04
0K-random

1K-random

2K-random

3K-random

FIG. 14: The average betweenness and the standard deviation of the betweenness distribution as functions of d for dKrandomisations of the real networks. The corresponding values for the real networks are shown by dashed lines.

13
TABLE II: The scalar topological metrics of the real networks
and the minimum value of d such that the networks dKrandomizations approximately preserve all the metrics.
Metrics

k
r
c

h
b
dK

PGP

Collab. Protein

Air

Internet Power

4.6
6.4
6.4
11.9
6.3
4.7
0.238 0.157 -0.137 -0.268 -0.236 -0.273
0.27
0.65
0.09
0.62
0.46
0.68
7.5
6.6
4.2
3.0
3.1
2.0
6 104 4 104 7 104 4 103 2 104 2 104
3K
3K
3K
2K
1K
?

r is the assortativity coefficient,

X
hki2
kk 0 P (k, k 0 ) hk 2 i2
r=

kk0

hk 3 ihki hk 2 i2

(18)

which is nothing but the Pearson correlation coefficient of the 2K-distribution P (k, k 0 );
c is the average clustering
X
c =
c(k)P (k),

(19)

tance distribution.
Fig. 13 shows degree-dependent betweenness centrality
b(k) in the real networks and their dK-randomizations.
Betweenness b(i) of node i is a measure of how important i is in terms of the number of shortest paths passing
through it. Formally, if st (i) is the number of shortest
paths between nodes s 6= i and t 6= i that pass through
i, and st is the total number of shortest paths between
the two nodes s 6= t, then betweenness of i is
b(i) =

X s,t (i)
s,t

s,t

(16)

Averaging over all nodes of degree k, degree-dependent

betweenness b(k) is
b(k) =

X
i: deg(i)=k

b(i)
.
N (k)

(17)

We also compute the betweenness distribution, and

show its average and standard deviation in Fig. 14.
We observer similar trends with respect to both distance and betweenness metrics. The power grid cannot
be approximated even by its 3K-randomization. The Internet lies at the other extreme: even 1K-randomization
does not disturb its global metrics too much. The air
transportation network appears to come next, as its 2Krandomizations resemble it closely. But all the networks
other than the power grid are very similar to their 3Krandomizations.
D.

Scalar topological metrics and dK-randomness

of real networks

To conclude this section we show in Table II the most

important scalar topological metrics for the real networks. These metrics are coarse summary statistics of
the more informative and detailed metrics that we have
considered in this section. Specifically, these coarse summaries are:
k is the average degree in the network, Eq. (2),
which is both the 0K-distribution and a summary
statistics of the 1K-distribution in the dK-series
terminology;

which is a coarse summary statistics of the 3Kdistribution;

is the average distance, Eq. (15), which is unreh
lated to dK-distributions;
b is the average betweenness,
X
b =
b(k)P (k),

(20)

unrelated to dK-distributions as well.

In Table II we also show the minimum value of d such the
dK-randomization null model approximately reproduces
the real network with respect to all the metrics above.
The observation that the power grid cannot be approximated even by its 3K-randomization is instructive.
It shows that there are networks for which there is no
sufficiently small d capable of preserving the network
structure upon dK-randomizing. In case of the power
grid, the explanation why this network is not even 3Krandom may be related to the fact that it is carefully
designed and fully controlled by human engineers. Informally, we can think of it as rather non-random, designed, and thus bearing a number of constraints that
the dK-distributions with low d cannot capture. Informally, the higher d required to approximately preserve
the network structure upon dK-randomization, the less
random the network is. The commonly referred explanation that the power grid is an outlier because it
is spatially embedded, may be less relevant here because
two other networks that we have considered (the Internet
and air transportation) are also spatially embedded.
What is different between the power grid and the other
considered networks is that the latter are self-evolving.
They may be engineered to a certain degree, such as the
Internet, but their global structure and evolution are not
fully controlled by any single human or organization. In
the Internet case, for example, the global network topology is a cumulative effect of independent decisions made
by tens of thousands of separate organizations, roughly
corresponding to Autonomous Systems, i.e., nodes of the
Internet graph.
In that sense, self-evolving complex networks are
more random. However, why the level of their randomness is at d 6 3 remains an open question.

14
TABLE III: dK-series vs. d-series
d dK-statistics d-statistics

k
0
1
2

N (k)
N (k, k0 )
N (k, k0 , k00 )
3
N4 (k, k0 , k00 )

V.

N
M
W
T

MOTIF-BASED SERIES VS. dK-SERIES

In this section we compare dK-series with the series

based on motifs, and show that the latter cannot form a
systematic basis for topology analysis.
The difference between dK-series and motif-series,
which we can call d-series, is that the former is the series
of distributions of d-sized subgraphs labeled with node
degrees in a given network, while the d-series is the distributions of such subgraphs in which this degree information is ignored. This difference explains the mnemonic
names for these two series: d in dK refers to the subgraph size, while K signifies that they are labeled by
node degreesK is a standard notation for node degrees.
This difference between the dK-series and d-series is
crucial. The dK-series are inclusive, in the sense that the
(d + 1)K-distribution contains the full information about
the dK-distribution, plus some additional information,
which is not true for d-series.
To see this, let us consider the first few elements of both
series in Table III. In Section IV A we show explicitly how
the (d + 1)K-distributions define the dK-distribution for
d = 0, 1, 2. The key observation is that the d-series does
not have this property. The 0th element of d-series is
undefined. For d = 1 we have the number of subgraphs
of size 1, which is just N , the number of nodes in the
network. For d = 2, the corresponding statistics is M ,
the number of links, subgraphs of size 2. Clearly, M and
N are independent statistics, and the former does not
define the latter. For d = 3, the statistics are W and T ,
the total number of wedges and triangles, subgraphs of
size 3, in the network. These do not define the previous

[1] R. Milo, S. Shen-Orr, S. Itzkovic, N. Kashtan,

D. Chklovskii, and U. Alon, Science 298, 824 (2002).
[2] U. Alon, Nat Rev Genet 8, 450 (2007).
[3] U. Alon, An Introduction to Systems Biology: Design
Principles of Biological Circuits (Chapman & Hall/CRC,
Boca Raton, 2006).
[4] N. Rosenfeld, M. Elowitz, and U. Alon, J Mol Biol 323,
785 (2002).
[5] S. Mangana, S. Itzkovitz, A. Zaslaver, and U. Alon, J
Mol Biol 356, 1073 (2006).
[6] J. Knabe, C. Nehaniva, and M. Schilstra, Biosystems 94,
68 (2008).

element M either. Indeed, consider the following two

networks of size N the chain and the star:
1

N-1
1

1
1

There are no triangles in either network, T = 0. In the

chain network, the number of wedges is W = N 2, and
in the star W = (N 1)(N 2)/2. We see that even
though W (d = 3) scales completely differently with N
in the two networks, the number of edges M = N 1
(d = 2) is the same.
In summary, d-series is not inclusive. For each d,
the corresponding element of the series reflects a differen kind of statistical information about the network
topology, unrelated or only loosely related to the information conveyed by the preceding elements. At the
same time, similar to dK-series, the d-series is also converging since at d = N it specifies the whole network
topology. However, this convergence is much slower
that in the dK-series case. In the two networks considered above, for example, neither W = N 2, T = 0
nor W = (N 1)(N 2)/2, T = 0, fix the network
topology as there are many non-isomorphic graphs with
the same (W, T ) counts, whereas the 3K-distributions
N (1, 2, 2) = 2, N (2, 2, 2) = N 4 and N (1, N 1, 1) =
(N 1)(N 2)/2 define the chain and star topologies exactly.
The node degrees thus provide necessary information
about subgraph locations in the original network, which
improves convergence, and makes the dK-series basis inclusive and systematic.

Acknowledgments

We thank Alex Arenas and Alessandro Vespignani for

useful comments and discussions, and Connie Lyu and
Bradley Huffaker for their help with Figs. 1,5. This
work was supported in part by DGES grant FIS200766485-C02-02, by NSF CNS-0434996 and CNS-0722070,
by DHS N66001-08-C-2029, and by Cisco Systems.

[7] P. Ingram, M. Stumpf, and J. Stark, BMC Genomics 7,

108 (2006).
[8] O. Cordero and P. Hogeweg, Mol Biol Evol 23, 1931
(2006).
[9] P. Kuo, W. Banzhaf, and A. Leier, Biosystems 85, 177
(2006).
[10] A. Mazurie, S. Bottani, and M. Vergassola, Genome Biol
6, R35 (2005).
[11] S. Sakata, Y. Komatsu, and T. Yamamori, Neurosci Res
51, 309 (2005).
[12] A. V
azquez, R. Dobrin, D. Sergi, J.-P. Eckmann, Z. N.
Oltvai, and A.-L. Barab
asi, Proc Natl Acad Sci USA 101,

15
17940 (2004).
[13] Y. Artzy-Randrup, S. Fleishman, N. Ben-Tal, and
L. Stone, Science 305, 1107 (2004).
[14] P. Mahadevan, D. Krioukov, K. Fall, and A. Vahdat,
Comput Commun Rev 36, 135 (2006).
[15] M. Bogu
na
, R. Pastor-Satorras, A. Daz-Guilera, and
A. Arenas, Phys Rev E 70, 056122 (2004).
[16] J. I. Alvarez-Hamelin, L. DallAsta, A. Barrat, and
A. Vespignani, in Advances in Neural Information Processing Systems 18, edited by Y. Weiss, B. Sch
olkopf, and
J. Platt (MIT Press, Boston, 2006), pp. 4150.
[17] S. Maslov and K. Sneppen, Science 296, 910 (2002).
[18] S. Maslov, K. Sneppen, and U. Alon, Handbook of Graphs
and Networks (Wiley-VCH, Berlin, 2003), chap. Correlations Profiles and Motifs in Complex Networks.
[19] X. Dimitropoulos, D. Krioukov, G. Riley, and A. Vahdat,
ACM Transactions on Modeling and Computer Simulation (to appear) (2009), arXiv:0708.3879.
[20] J. Duch and A. Arenas, Phys Rev E 72, 027104 (2005).
[21] L. Freeman, Sociometry 40, 35 (1977).
[22] M. Bogu
na
, D. Krioukov, and kc claffy, Nature Physics
5, 74 (2009).
Serrano, D. Krioukov, and M. Bogu
[23] M. A.
na
, Phys Rev
Lett 100, 078701 (2008).
[24] T. V. den Bulcke, K. V. Leemput, B. Naudts, P. van Remortel, H. Ma, A. Verschoren, B. D. Moor, and K. Marchal, BMC Bioinformatics 7 (2006).
[25] J. Knabe, C. Nehaniva, and M. Schilstra, Artif Life 14,

135148 (2008).
[26] S. Roy, M. Werner-Washburne, and T. Lane, Bioinformatics 24, 13181320 (2008).
[27] B. M. Waxman, IEEE J Sel Area Comm 6, 1617 (1988).
[28] E. Zegura, K. Calvert, and S. Bhattacharjee, in Proc INFOCOM (1996), vol. 2, pp. 594602.
[29] A. Medina, A. Lakhina, I. Matta, and J. Byers, in Proc
MASCOTS (2001), pp. 346353.
[30] J. Winick and S. Jamin, Technical Report UM-CSE-TR456-02, University of Michigan (2002).
[31] M. E. J. Newman, Proc Natl Acad Sci USA 98, 404
(2001).
Serrano, and A. Vespig[32] V. Colizza, A. Flammini, M. A.
nani, Nat Phys 2, 110 (2006).
[33] V. Colizza, R. Pastor-Satorras, and A. Vespignani, Nat
Phys 3, 276 (2007).
[34] P. Mahadevan, D. Krioukov, M. Fomenkov, B. Huffaker,
X. Dimitropoulos, kc claffy, and A. Vahdat, Comput
Commun Rev 36, 17 (2006).
[35] D. J. Watts and S. H. Strogatz, Nature 393, 440 (1998).
Serrano and M. Bogu
[36] M. A.
na
, Phys Rev E 74, 056114
(2006).
[37] M. A. Serrano and M. Bogu
na
, Phys Rev E 74, 056115
(2006).
[38] See [36, 37] for an alternative formalism involving three
point correlations