Zhan 2010 Distributions
Zhan 2010 Distributions
Zhan 2010 Distributions
Physica A
journal homepage: www.elsevier.com/locate/physa
article
info
Article history:
Received 13 April 2009
Received in revised form 23 October 2009
Available online 2 January 2010
Keywords:
Complex network
Laplacian matrix
Eigenvalue
Graph theory
Node-degree
Random-graph network
Small-world network
Scale-free network
abstract
In this paper, the important issue of Laplacian eigenvalue distributions is investigated
through theory-guided extensive numerical simulations, for four typical complex network
models, namely, the ER random-graph networks, WS and NW small-world networks, and
BA scale-free networks. It is found that these four types of complex networks share some
common features, particularly similarities between the Laplacian eigenvalue distributions
and the node degree distributions.
2009 Elsevier B.V. All rights reserved.
1. Introduction
Recently, the notion of complex networks has become more and more popular for research in physics, engineering,
biology and even social sciences [17].
A network can be viewed as a graph consisting of nodes connected by edges according to certain rules, in which the
nodes and edges typically represent some physical, biological or social entities and their relationships, respectively [8]. One
of the greatest challenges in understanding the nature and essence of various complex networks today is how the network
structures affect their dynamical behaviors [4,5,9]. To better understand and to deal with various complex networks, it is
important and even necessary to study and comprehend the structural characteristics of such networks. From a graphtheoretic perspective, the Laplacian matrix of a network is essential, because it contains most information about the
underlying network, such as the node-degree distribution and edge connectivity, among others. On the other hand, for a
constant matrix, the spectrum of its eigenvalues plays a fundamental role in its behavior and functioning. Based on the study
and comparison of the eigenvalues of the Laplacian matrices of different networks, called Laplacian eigenvalues hereafter,
one can gain a basic understanding of the structural properties of such networks, as well as their similarities and differences,
which are very useful for many relevant applications such as network synchronization [10,9].
Among the many representative network models, the classical ErdsRnyi (ER) random-graph networks [11] and the
relatively new WattsStrogatz (WS) and NewmanWatts (NW) small-world networks [12,13] and BarabsiAlbert (BA)
scale-free networks [14] are particularly significant and important. The question to be addressed in this paper is: Are there
any similar statistical properties of the Laplacian eigenvalues with respect to the topological features in ER, WS, NW and
BA network models? Noticeably, this question has already been addressed from different points of view in the literature.
Corresponding author. Tel.: +852 2788 7922; fax: +852 2788 7791.
E-mail address: [email protected] (G. Chen).
0378-4371/$ see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2009.12.005
1780
For example, it was shown in Ref. [15] that the largest eigenvalues of the Laplacian and adjacency matrices of a scale-free
network are determined mainly by the largest-degree node of the network, while the smallest nonzero eigenvalue of the
Laplacian matrix depends on the way the nodes are connected to each other. Also, based on the random matrix theory it
was shown in Ref. [16] that despite the differences between the spectral densities of some different types of networks, their
eigenvalue fluctuations are about the same and they all follow the Gaussian orthogonal ensemble distribution of the random
matrices. From a somewhat different standpoint and approach, this paper performs a careful numerical analysis based on
intensive computer simulations, to study the eigenvalue distributions of the unweighted symmetrical Laplacian matrices of
four representative complex network models: the ER random-graph networks, WS small-world networks, NW small-world
networks, and BA scale-free networks. It is well known that for large-sized networks, their possible ensembles or realizations
are very large. For instance, the set of all possible ER random-graph networks with 2500 nodes has roughly O(225002499/2 )
different network realizations, implying that there are totally about O(225002499/2 ) different Laplacian matrices, which is
literally intractable given the currently available computing power. Therefore, following the common practice this paper
carries out a theory-guided numerical study of the subject in interest.
The main finding of this paper is that the Laplacian eigenvalues of the aforementioned four types of complex networks
have very different properties in general and yet meanwhile they also share some common features. In particular, based
on the spectral theory of graphs [17,18], the results obtained in this paper reveal that the distributions of the Laplacian
eigenvalues are very similar to the distributions of the node-degree distributions, which further verify some similar findings
in Refs. [15,16]. More precisely, by plotting the Laplacian eigenvalues of the four network models against the node-degree
indices in increasing order, we obtain various eigenvalue curves and found that in different realizations of the same
type of network topology of the same size, the relative deviations of their Laplacian eigenvalue curves are very small
with only up to 5% errors. This phenomenon shows that, for a certain network with fixed size and given node-degree
distribution, the eigenvalue distributions of these four very different types of complex networks are all insensitive to the
network connectivity characteristics, namely the locations of the 1s and 0s in their coupling matrices, where 1 means
connected and 0 means unconnected.
For an unweighted symmetrical Laplacian matrix, the smallest eigenvalue is always zero and the others are strictly
positive, denoted as 0 = 1 < 2 3 N [28,10]. In many applications, the largest and the smallest
nonzero Laplacian eigenvalues play a very important role, therefore they have attracted more attention than the other
eigenvalues [19,20]. One contribution of this paper is to further reveal and confirm, from a somewhat different viewpoint,
the similarities between the Laplacian eigenvalue distributions and the node-degree distributions of the aforementioned
four types of complex networks, supported by theory-guided numerical simulations. By plotting the node degrees against
the node-degree indices in increasing order, various node-degree curves are obtained, showing that (i) for the ER randomgraph networks, the eigenvalue curves are similar to the node-degree curves, and all these curves are located nearby, mostly
with up to 5% errors; (ii) for the BA scale-free networks, the eigenvalue curves are also similar to the node-degree curves
and their differences are also located nearby, mostly with up to 7.5% errors; (iii) for the WS and NW small-world networks,
however, the eigenvalue curves are quite different from their corresponding node-degree distributions in some small-index
regions, although they are still similar in other regions.
The rest of this paper is organized as follows. Section 2 provides a brief summary of some notations and generating
algorithms for the ER random-graph networks, WS and NW small-world networks and BA scale-free networks, useful
throughout the paper. Section 3 presents some theory-guided numerical analysis of the Laplacian eigenvalue and nodedegree distributions, reporting the main results of the paper. Section 4 gives more detailed analysis about the Laplacian
eigenvalue properties of the four complex network models, respectively. Finally, conclusions are drawn in Section 5.
2. Preliminaries
2.1. Notations
N : the number of nodes in a random graph model;
p : the probability parameter;
ki : the degree of node i, defined to be the number of edges directly connecting node i;
k = {k1 , k2 , . . . kN } : the set of node degrees arranged in the non-decreasing order: k1 k2 kN ;
IN : N N identity matrix;
A : N N adjacency matrix, defined by
1
0
aij =
if i 6= j, i adjacent j
otherwise;
(
lij =
1
0
if i = j
if i 6= j, i adjacent j
otherwise;
1781
1q
kx kq =
i =
ki
n
P
kj
j =1
Here, n = m0 + t 1 is the total number of existing nodes at the (t 1)th step of the node-adding process.
3. Distributions of Laplacian eigenvalues
To start with, some numerical results of extensive simulations are shown, to have a sense of different distributions of
Laplacian eigenvalues for the above-reviewed four typical network models, namely the ER random-graph model, the WS
small-world network model, the NW small-world model, and the BA scale-free network model, respectively.
The results are summarized in Fig. 1(a), where the number of simulation runs on each network model is 1000, therefore
totally 4000 network realizations were generated by each algorithm and a total of 2500 4000 eigenvalues were plotted
altogether in the same figure (they are not the averages of the Laplacian eigenvalues of the 1000 simulations). Fig. 1(b) shows
the averages of the Laplacian eigenvalues of the four typical models plotted in Fig. 1(a). More precisely:
1782
300
250
250
200
200
150
MeanEigenvalue
Eigenvalue
300
ll
100
lll
V
150
Vl
100
Vll
50
50
lV
Vlll
0
0
0
500
1000
1500
2000
Index of Eigenvalue
2500
500
1000
1500
2000
Index of Eigenvalue
2500
Fig. 1. (a) Laplacian eigenvalues of four typical network models: I. ER model; II. NW model; III. WS model; IV. BA model. (b) Mean Laplacian eigenvalues
of four typical network models: V. ER model; VI. NW model; VII. WS model; VIII. BA model.
1.
2.
3.
4.
It is clear, from a comparison of Fig. 1(a) with (b), that the two figures look quite similar; namely, the actual values of
each eigenvalue in all individual runs and the averages of the Laplacian eigenvalues are not much different. This implies that
in various realizations of the same type of network topology of the same size, the variation of their Laplacian eigenvalues is
quite small. And this is roughly the same for all four different types of network models, a phenomenon that needs further
investigation in the future.
Another observation is that the distributions of Laplacian eigenvalues of the four types of network models are all strongly
related to their node-degree distributions. The finding is: for the BA scale-free networks and ER random-graph networks,
the eigenvalue curves are similar to their node-degree curves; for the WS and NW small-world networks, the eigenvalue
curves are different from their corresponding node-degree distributions in some small-index regions, although they are
quite similar in other regions, as shown in Fig. 2 (ER random-graph networks with N = 2500 and p = 0.2), Fig. 4 (WS
small-world networks with N = 2500, m0 = 50 and p = 0.05), Fig. 6 (NW small-world networks with N = 2500, m0 = 40
and p = 0.03), and Fig. 8 (BA scale-free networks with N = 2500, m0 = 20 and m = 10), respectively.
To quantify the variability of a Laplacian eigenvalue, define the relative deviation as
i =
i hi i
,
hi i
(1)
where hi i is the mean of i , and hii 6= 0 for all i, i = 2, 3, . . . , N. For two real vectors x, y RN 1 , with y 6= 0, define their
relative average deviation as
kx y k2
.
kyk2
(2)
Note that small values of i and imply a small variability of i and a small difference between x and y, respectively.
Lemma 1 (WielaneltHoffman Theorem [16]). Suppose C = A + B, where A, B, C RN N are symmetric matrices, and let the
eigenvalue sets (B) and (C ) be arranged in the non-decreasing order. Then,
v
u N
uX
t
|i (C ) i (B)|2 kAkF .
(3)
i=1
Theorem 1. Let G be a graph of N nodes with Laplacian matrix L. Arrange the node-degree set k = (k1 , k2 , . . . , kN )T and the
eigenvalue set (L) = (1 , 2 , . . . , N )T in the non-decreasing order, respectively. Then,
k(L) kk2
=
kkk2
kkk1
kkk22
kkk1
(4)
1783
Proof. By definition,
L = A + K ,
(5)
where A is the adjacency matrix and K = diag{k1 , k2 , . . . , kN }. Note that L, A and K are all symmetric matrices. It then
follows from Lemma 1 that
N
X
(i ki )2 kAk2F .
(6)
i =1
Note also that A is the adjacency matrix, of which the diagonal elements are all 0 and the off-diagonal elements are either 0
or 1 with the sum of column (row) i being equal to ki . This implies that
N X
N
X
kAk2F =
|ai,j |2 =
i=1 j=1
N
X
ki .
(7)
i=1
!
N
X
ki
i=1
!
.
k2i
(8)
i=1
!2
ki
N
X
i =1
!
k2i
(9)
i=1
N
2
!
X
N
ki
(10)
i =1
Theorem 1 shows that for the Laplacian matrix L, the differences between Laplacian eigenvalues (L) and node-degrees k
are bounded by kkk1 /kkk22 . Thus, for a large-scale complex network with a huge number of connections, kkk1 /kkk22 typically
has a small value, implying that the distribution of the Laplacian eigenvalues is indeed similar to the distribution of the node
degrees.
Theorem 2. Let j be the eigenvalue of the Laplacian matrix L of a graph with N nodes, and ki be the
degree of node i for
i = 1, . . . , N. Then, in every interval [(ki ki ), (ki + ki )], there is at least one eigenvalue j |j = 1, 2, . . . , N of
L, i.e.,
(ki
ki ) (ki +
ki );
i = 1, 2, . . . , N .
ki ), (ki +
(11)
Proof. Recalling the following well-known result from matrix theory: for any matrix M RN N and x RN , one has
(12)
Since the Laplacian matrix L is symmetric and positive semidefinite, one has
(13)
| ki | = min |1 ki |, |2 ki |, . . . , |j ki |, . . . , |N ki | .
j
(14)
200
5
0
-5
150
100
50
500
1000
1500
2000
Relative difference(%)
250
Relative variance(%)
1784
2500
Index
500
1000
500
1000
1500
2000
2500
1500
2000
2500
15
10
5
0
Index
Fig. 2. Distributions of Laplacian eigenvalues and node degrees of the ER network model: (a) Laplacian eigenvalues are shown by black curves, node
degrees are shown by gray curves, (b) relative deviations of the Laplacian eigenvalues, (c) relative differences between eigenvalues and node-degrees.
Next, setting xi = (0, 0, . . . , 0, 1i , 0, . . . 0)T RN . Then, since the Laplacian matrix L has zero row-sum and zero columnsum, one has
k(L ki IN )xi k2 =
ki .
(15)
| ki |
ki ,
(16)
i =
i ki
i
(17)
(3) From the simulation, it is found that 2 is always located in k1 k1 , k1 + k1 , which shows that 2 is strongly
related to the smallest node-degree k1 .
(4) Fig. 2(c) shows the relative differences between eigenvalues and node-degrees are mostly located nearby with 6%
errors, which shows that the Laplacian eigenvalues are strongly related to the node-degrees. Thus, an eigenvalue can be
roughly approximated by the corresponding node-degree, e.g., 1250 k1250 = 200.
1785
2500
2000
Eigenvalue
1500
1000
500
0
-500
500
1000
1500
2000
2500
Index of Eigenvalue
140
120
100
80
Relative difference(%)
Relative difference(%)
Relative variance(%)
Fig. 3. The eigenvalue curves of the ER random-graph network under different generation conditions. From bottom up: I. N = 2500 and p = 0.08; II.
N = 2500 and p = 0.2; III. N = 2500 and p = 0.4; IV. N = 2500 and p = 0.8.
2
0
-2
-4
500
1000
1500
2000
2500
500
1000
1500
2000
2500
500
1000
1500
Index
2000
2500
800
600
400
60
200
40
20
0
-20
500
1000
1500
Index
2000
2500
5
0
-5
-10
Fig. 4. (a) Distributions of Laplacian eigenvalues and node degrees of the WS small-world network model: Laplacian eigenvalues are shown by black curves,
node degrees are shown by gray curves, (b) relative deviations of Laplacian eigenvalues, (c) relative differences between eigenvalue and node-degree, (d)
same data as (c) are displaced in an expanded scale.
However, it is well known from algebraic graph theory [15] that for any connected undirected network, the largest
eigenvalue of its Laplacian matrix can never be larger than the size of the network, namely, N N. In the present case, if
N = 2500 is fixed but p is increased, then the eigenvalue curves will move upwards, as shown in Fig. 3, but can never go
beyond 2500. In the extreme case where p = 1, the ER random-graph network model becomes a fully-connected regular
network, with 1 = 0 and 2 = 3 = = 2500 = 2500 = N.
4.2. WS small-world networks
Similar to the analysis of ER complex networks, Fig. 4 shows the plots of the curves of eigenvalues, node-degrees, as well
as relative deviations and relative differences between eigenvalues and node-degrees.
The simulation results reveal the following observations:
(1) The eigenvalue curves of WS small-world models are not symmetric, with long tails. Fig. 4(b) shows that the relative
deviations of the eigenvalues are mostly located nearby, with up to 1.5% errors. This phenomenon shows that the
Laplacian eigenvalues of WS complex networks are insensitivity to the network connectivity characteristics, namely,
in different realizations with the same size N, parameter m0 , and rewiring probability p, the eigenvalues curves are
almost the same (as shown in Figs. 4 and 5). Thus, for a WS-type complex network, in order to estimate its eigenvalue
distribution one does not need to know the detailed information about its specific connectivity topology but only need
to know the values of N, m0 and p.
1786
a 140
120
Eigenvalue
100
80
140
120
60
100
80
40
60
40
20
20
0
-20
-20
500
500
1500
2000
1000
1500
Index of Eigenvalue
1000
2000
2500
2500
b
200
Relative difference(%)
Relative variance(%)
Relative difference(%)
Fig. 5. (Color online) (a) The eigenvalue curves of the WS small-world network under different generation conditions. From bottom up: I. Yellow curves
for N = 2500, m0 = 50 and p = 0.10; II. Red curves for N = 2500, m0 = 50 and p = 0.20; III. Blue curves for N = 2500, m0 = 50 and p = 0.30; IV. Green
curves for N = 2500, m0 = 50 and p = 0.50; V. Black curves for N = 2500, m0 = 50 and p = 0.80, (b) displays the same curve as (a), but just two curves:
I. Yellow curves for N = 2500, m0 = 50 and p = 0.10 and black curves for N = 2500, m0 = 50 and p = 0.80.
4
2
0
-2
150
500
1000
1500
2000
2500
500
1000
1500
2000
2500
1500
Index
2000
2500
80
60
40
100
20
50
0
0
500
1000
1500
Index
2000
2500
0
5
0
-5
500
1000
Fig. 6. (a) Distributions of Laplacian eigenvalues and node degrees of the NW small-world network model: Laplacian eigenvalues are shown by black
curves, node degrees are shown by gray curves, (b) relative deviations of the Laplacian eigenvalue, (c) relative differences between eigenvalues and nodedegrees, (d) same data as (c) are displaced on an expanded scale.
(3) No 2 are located in k1 k1 , k1 + k1 , implying that 2 is not only affected by k1 but also by some other factors.
(4) Fig. 4(c) and (d) show that the relative differences between eigenvalues and node-degrees are large in the small-index
region (index smaller than 51) but small in other regions. Fig. 4(d) shows that the relative differences are mostly located
nearby, within 6% errors, in the region from index 51 to index 2500. This means that the Laplacian eigenvalues are
strongly related to the node-degrees in those not-small-index regions.
(5) If N = 2500 and m0 are both fixed but the rewiring probability p is increased, then 2 is increased and the eigenvalue
curve has some gradual change in shape, as shown in Fig. 5. This phenomenon shows that the rewiring probability p
can be used to control the regular and random levels of the WS small-world network model, which is consistent
with our understanding of the model [12]: when p is small, which means only a few edges are rewired, the network
generated by the WS algorithm is more like a regular network; but as p is increased, more edges will be rewired, so
the generated network becomes more random; when p = 1, every edge will be rewired therefore the WS small-world
network model literally becomes an ER random-graph network. Thus, by tuning the value of p, one can obtain a transition
from a completely regular network to a completely random network, through a long spectrum of small-world networks.
1787
500
Eigenvalue
400
300
200
100
0
-100
500
1000
1500
Index of Eigenvalue
2000
2500
500
450
400
350
300
10
5
0
-5
-10
250
200
150
100
50
0
-50
500
1000
1500
Index
2000
2500
Relative difference(%)
Relative variance(%)
Fig. 7. The eigenvalue curves of the NW small-world network under different generation conditions. From bottom up: I. N = 2500, m0 = 40 and p = 0.01;
II. N = 2500, m0 = 40 and p = 0.03; III. N = 2500, m0 = 40 and p = 0.07; IV. N = 2500, m0 = 40 and p = 0.10; V. N = 2500, m0 = 40 and p = 0.13.
500
1000
1500
2000
2500
500
1000
1500
Index
2000
2500
10
5
0
Fig. 8. Distributions of Laplacian eigenvalues and node degrees of the BA scale-free network model: Laplacian eigenvalues are shown by black curves,
node degrees are shown by grey curves, (b) relative deviations of the WS model, (c) relative differences between eigenvalues and node degrees.
1788
(3) From simulations, it is found that 2 is always located in k1 k1 , k1 + k1 , which shows that 2 is strongly related
to the smallest node-degree k1 .
(4) Fig. 8(c) shows that the relative differences between eigenvalues and node-degrees are mostly located nearby, within
5% errors, implying that the Laplacian eigenvalues are strongly related to the node degrees. Thus, the node-degree
distribution can give a good approximation to the eigenvalue distribution, for example, 1000 k1000 = 42.
5. Conclusions
In this paper, the distributions of Laplacian eigenvalues of four representative complex network models have been
carefully investigated through extensive theory-guided numerical simulations, including the ER random-graph networks,
WS small-world networks, NW small-world networks, and BA scale-free networks. The results show that although the
Laplacian eigenvalues of these four types of complex networks have many different properties in general, they nevertheless
share some common features, particularly some similarities between the Laplacian eigenvalue distributions and the nodedegree distributions. New findings notwithstanding, there are other important issues remaining for future investigations, for
instance the relations between Laplacian eigenvalues and other network parameters such as average path length, clustering
coefficient, betweenness centrality, and so on [21,22].
Acknowledgements
This research was supported by the Hong Kong Research Grants Council under the GRF Grants CityU 111708 and CityU
122906.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]