Enhancing the network synchronizability

Ming ZHAO , Tao ZHOU , Guan-rong CHEN , Bing-hong WANG

Front. Phys. ›› 2007, Vol. 2 ›› Issue (4) : 460 -468.

PDF (1563KB)
Front. Phys. ›› 2007, Vol. 2 ›› Issue (4) :460 -468. DOI: 10.1007/s11467-007-0058-8
REVIEW ARTICLE
Enhancing the network synchronizability
Author information +
History +
PDF (1563KB)

Abstract

The structural and dynamical properties, particularly the small-world effect and scale-free feature, of complex networks have attracted tremendous interest and attention in recent years. This article offers a brief review of one focal issue concerning the structural and dynamical behaviors of complex network synchronization. In the presentation, the notions of synchronization of dynamical systems on networks, stability of dynamical networks, and relationships between network structure and synchronizability, will be first introduced. Then, various technical methods for enhancing the network synchronizability will be discussed, which are roughly divided into two classes: Structural Modification and Coupling-Pattern Regulation, where the former includes three typical methods--dividing hub nodes, shortening average distances, and deleting overload edges, while the latter mainly is a method of strengthening the hub-nodes' influence on the network.

Graphical abstract

Keywords

complex networks / synchronization / structural modification / coupling pattern

Cite this article

Download citation ▾
Ming ZHAO, Tao ZHOU, Guan-rong CHEN, Bing-hong WANG. Enhancing the network synchronizability. Front. Phys., 2007, 2 (4) : 460-468 DOI:10.1007/s11467-007-0058-8

登录浏览全文

4963

注册一个新账户 忘记密码

1 Introduction

To study the collective behaviors of complex systems, a common way is to represent each individual system by a node and the interaction between two individuals by an edge, thereby abstracting the coupled systems as a network. Then, the study is correspondingly converted to investigating the dynamical behaviors of the resultant network. Within the context of general networks, random networks have been the most well-studied ones in the recent half-century. In recent years, the speedy development of computer science and the rapid evolvement of the Internet have provided researchers with super-computing power and huge databases for much more efficient network studies, leading to the discoveries of small-world effect and scale-free property of networks, and further to the construction of some new network models which are more realistic and in many aspects closer to real-world networks in social, biological and technological systems. These new models, particularly the small-world and scale-free network models [1, 2], characterize various complex networks by such measures as average distance, clustering coefficient, power-law degree distribution, and so on [3−6].

A particular subject in network research is concerned with the synchronization of coupled dynamical systems (as nodes) on a complex network, and this important subject has been extensively investigated in the past few years [7−15]. In retrospect, the investigation of synchronization of dynamical systems on complex networks has gone through several stages. At the very beginning, scientists eagerly studied the synchronizability of two typical kinds of network models and found that coupled dynamical oscillators on small-world and scale-free network models are much easier to synchronize than on regular lattices, and the synchronizability of a small-world network model is even better than that of random networks in some cases [16−19]. It was then intuitively believed, that the strong synchronizability of such networks is due to their short average distances, however, it was lately found [20] that in some complex networks, with the increasing of heterogeneity of degree and betweenness distributions, although the average distance becomes shorter, the network synchronizability actually becomes worse. The subsequent works on possible relationships between network structures and the synchronizability have shown that a combination of shorter average distance and more homogeneous degree distribution can make the network to achieve synchronization easier [20−24]. Thereafter, great efforts have been devoted to seeking ways to enhance the network synchronizability, and this remains one of the most active research activities in the field today.

This article offers a brief review of the focal issue concerning the enhancement of network synchronizability. The basic notions of synchronization of dynamical systems on networks, stability of dynamical networks, and relationships between network structure and synchronizability, will be first introduced. Then, some technical methods for enhancing the network synchronizability will be discussed; they are divided into two classes: Structural Modification, which includes three typical methods–dividing hub nodes, shortening average distances, and deleting overload edges, and Coupling-Pattern Regulation, which is mainly a method of strengthening the hub-nodes’ influence on the network. Finally in this article, some open questions will be posed, which are deemed interesting and significant thus might further enlighten the readers’ interests in this attractive research subject.

2 Network models and the master stability function

In this section, a generic model of coupled dynamical systems on a network is considered, and the master stability function approach is taken to study the stability of the fully synchronized states of the network [25−28].

In the network, each node is located with a dynamical system such as an oscillator, and an edge connecting two nodes means that there exits a bi-directional coupling between them. Denote the state of the ith node by xi, i = 1, ···, N, and the set of dynamical equations for the whole network are obtained as

(1)x˙i=F(xi)−σ∑j=1NGijH(xj), i=1,⋯,N

where x˙i = F(xi) describes the dynamics of each individual node, σ is the coupling strength, and H(xj) is the output function. In this model, Gij are the elements of the coupling matrix G, i, j = 1, ···, N, which is determined by the coupling pattern among the nodes; for example, when the nodes are coupled symmetrically, G has the same form as the network Laplacian L, i.e., G = L, where

(2)Lij={ki,for i=j−1,  for j∈ Λi0,  otherwise

in which ki is the degree of node i and Λi is the set of node i’s neighbors.

Because of the irreducible and diffusive coupling structure and the symmetry and positive semi-definiteness of L, it has zero row sums and all eigenvalues are nonnegative real and the smallest eigenvalue (denoted by λ1) is always zero with multiplicity 1. Thus, the eigenvalues can be ranked as 0 = λ1 < λ2 ≤ λ3 ≤ ··· ≤ λN. In this setting of the network, the synchronization manifold is invariant, on which the fully synchronized state satisfies x1 = x2 = ··· = xN = s, with s˙=F(s).

For some dynamical systems, the master stability function is negative when all the products of the coupling strength σ and the nonzero eigenvalues of the Laplacian L are larger than a certain constant α1 > 0, so that the largest Lyapunov exponent of the connected network is negative [25, 26]. Therefore, the network is synchronizable when the eigenvalue λ2 > α1/σ. In this case, the larger the λ2, the better the network synchronizability. For some other dynamical systems, however, the master stability function is negative only in a finite interval (α1, α2) [29], over which the largest Lyapunov exponent of the network is negative, so that the network is synchronizable for some σ when the eigenratio R = λN/λ2 satisfies R < α2/α1 [27, 28]. In this case, the smaller the R, the better the network synchronizability. The former corresponds to the case that the stability zone is unbounded, and the latter, bounded. As a matter of fact, these two cases can also co-exist [30−33]. Notice, moreover, that the right-hand of the two inequalities depend only on the dynamics of the individual system and the output function, while the eigenvalue λ2 and the eigenratio R depend only on the Laplacian L. Consequently, the problem of synchronization can be divided into two parts: choosing suitable parameters of node dynamics and analyzing the eigenvalues of the Laplacian. The stability analysis is also valid for some more complicated coupling patterns [25−28, 34], including the case when G is nondiagonalizable [35].

3 Network structures and the synchronizability

As mentioned above, when the average distance keeps small, the more heterogenous the degree or the betweenness distribution, the more difficult the networks achieves synchronization. Nishikawa et al. [20] gave an explanation of this phenomenon: when the nodes are coupled symmetrically, the hubs (nodes of large degrees) receive too many signals (or too much information) at the same time, so the signals may cancel one another, resulting in effectively no communications passing through the hubs. They also pointed out that edges with larger loads may encounter the same situation. Later, in comparison to the results of [20], Hong et al. [21] concluded that the node with the maximal betweenness is the bottleneck for network synchronization, so it can be used as an indicator for the network synchronizability: the smaller the maximal betweenness, the better the network synchronizability. Although later it is found that the maximal betweenness is not a suitable indicator for network synchronizability [23, 33, 36], this claim may still be valid to some extent.

Based on the knowledge about the relations between the network structure and its synchronizability, it is easy to conclude that to improve the network synchronizability, the heterogeneity of degree or betweenness distribution, or the average distance, or the maximal betweenness (load) of a network should be reduced. This ideas leads to the classification of two types of methods for enhancing the network synchronizability: one is to modify the network structure and the other is to regulate the coupling pattern. In the subsequent section, four specific methods of these two types will be discussed respectively.

4 Some methods for enhancing network synchronizability

In this section, two specific methods are introduced, which can enhance the network synchronizability effectively: modifying the network structure and regulating the coupling pattern.

4.1 Dividing hub nodes

In Ref. [21], Hong et al. calculated the eigenratio of the Laplacian before and after the removal of the node with the maximal betweenness. It turns out that after the removal, the eigenratio is sharply reduced, indicating an improvement of the network synchronizability. It is an effective way to test the role of the hub-node, but is not practical because the hub-node in a real-world network is very important and cannot be removed freely. Nevertheless, this may be got around by, for instance, letting other nodes to share the load of a hub.

Inspired by the aforementioned idea, a method named structural perturbations was developed in Ref. [37] to improve the network synchronizability. Specifically, for a hub x0, m−1 auxiliary nodes, labelled as x1, x2, ···, xm−1, are added around it and let them be fully connected. Then, each edge of x0 will be relinked to a randomly picked node xi, i = 0, 1, ···, m−1. After this process, the betweenness of x0 is divided into m almost equal parts associating with these m nodes. As a result, the load on the hub x0 has been almost uniformly redistributed among the m nodes. The process is called m-division for short. A sketch map of a 3-division process on node x0 is shown in Fig. 1.

Due to the huge size of many real-life networks, it is usually impossible to obtain the nodes’ betweenness. Fortunately, previous studies show that there exists strongly positive correlation between degree and betweenness in Barabási-Albert (BA) networks and some other real heterogeneous networks [38, 39]; that is, the nodes with larger degrees will statistically have larger betweenness. Therefore, for practical reason, it can be assumed that a node with higher betweenness also has a larger degree in BA networks.

In order to enhance the network synchronizability, a few nodes with the highest degrees will be divided by the above-described m-division process. Rank all nodes of a given network G according to their degrees in decreasing order, so the node has the highest degree is arranged in the top of the queue. Then, a new network G(ρ, m) can be obtained in the following Nρ steps. Firstly, carry out m-division on the top node in G, yielding a new network G(1N,m). Secondly, calculate all nodes’ degrees in G(1N,m), and rank them again in the same way. Thirdly, generate network G(2N,m) by dividing the top node in G(1N,m) in the same way. Finally, repeat this process Nρ times, during which totally Nρ nodes will be divided, thus network G(ρ, m) is obtained. Since the edges are relinked to the m nodes at random in the dividing process, G(ρ, m) is generally not unique, but this is not important.

To explore how the structural perturbations affect the network synchronizability, the eigenratios r before and after the dividing processes are now compared. A BA network of size N = 2000 with average degree 〈k〉 = 12 is used for simulation. In Fig. 2, the ratio R = r′/r is plotted against the number of nodes that are divided, where r is the eigenratio of the original network and r′, the one after the process. Here, m = 3. With the probability ρ or the number of divided nodes increasing, the ratio R is decreasing, indicating the enhancement of the synchronizability. In Fig. 2, it can be seen that dividing only a few nodes is already able to enhance the network synchronizability prominently. For example, when R decreases to 0.7 while only 5 nodes are divided, the ratio will drop to half after only 0.6% nodes (i.e., 12 nodes) are divided.

To better understand the underlying mechanism of synchronization and the reason why these structural perturbations can significantly enhance the network synchronizability, the behaviors of two extensively studied quantities, the average distance L and the maximal degree kmax, are further examined. In a BA network, the node with the maximal degree is most likely the only node having the maximal betweenness. As illustrated by Fig. 3, L will increase with ρ, while kmax will decrease. This results provide some evidences for the way the two factors affect the synchronization of the network. The maximal degree (and, the maximal betweenness) may play a major role in determining the network synchronizability. It is worthwhile to emphasize that from the simulation results, one cannot say anything about how the average distance affects the network synchronizability. When L is varied slightly, the probably has insignificant influence on the synchronization as compared with the change of kmax. This suggests that reducing the maximal betweenness of a network is a practical and effective approach to enhancing the network synchronizability.

4.2 Shortening the average distance

In a network, if the degree distribution keeps fixed, shortening the average distance will enhance the network synchronizability. Based on this observation, Zhou et al. [40] investigated the synchronizability of a network model named crossed double cycles (CDCs), not only testifying the relationship between the average distance and the network synchronizability but also providing a possible way to make network more synchronizable.

In the language of graph theory [41−43], the cycle CN denotes a network consisting of N vertices, x1, x2, ···, xN. These N vertices are arranged as a ring, where the nearest two vertices are connected together. Hence, CN has N edges connecting the vertices x1x2, x2x3, ···, xN−1xN, xN x1. The CDCs, denoted by G(N, m), can be constructed by adding two edges, called crossed edges, to each vertex in CN. The two vertices connecting by a crossed edge have distance m in CN. For example, network G(N, 3) can be constructed from CN by connecting x1x4, x2x5, ···, xN−1x2, xN x3. And network G(N, 2) is isomorphic to a one-dimensional lattice with periodic boundary conditions and with coordination number z = 2 [44]. A network G is isomorphic to a network H if there exists a bijective map θ: V (G) → V (H), satisfying the adjacency-preserving condition (xy) ∈ E(G)⇔(θ(x)θ(y)) ∈ E(H), where V(·) and E(·) denote the sets of vertices and edges, respectively. Roughly speaking, two networks are isomorphic if they have the same structure.

A sketch map of G(20, 4) is shown in Fig. 4. Clearly, all the vertices in G(N, m) are of degree 4, thus the variance of degree is equal to 0. Furthermore, G(N, m) is vertex-transitivity; that is, for any two vertices x and y in G(N, m), there exists an automorphism map θ∶V(G) → V(G) such that y = θ(x). The vertex-transitivity networks are completely symmetrical, which are of particular practicability in the design of topological structures of data memory allocation and multiple processor systems [45].

Figure 5 shows how the average distance L affects the network synchronizability (again, measured by the eigenratio R) when the nodes are coupled symmetrically. One can see clearly that the network synchronizability is very sensitive to the average distance; as L increases, the eigenratio R sharply spans more than three magnitudes. And the network synchronizability is remarkably enhanced by simply reducing L. When the crossed length m is not too small or too large (compared with N), the networks with the same average distance have approximately the same synchronizability, no matter what these network sizes are. More interestingly, the numerical results indicate that the eigenratio R approximately obeys a power-law form as R ∼ L1.5 (see the inset of Fig. 5).

4.3 Deleting overload edges

In the synchronization process, not only the hub-nodes may be the bottlenecks but also some edges with larger loads may trouble the network synchronizability. Yin et al. [46] found that in scale-free networks, when the nodes are coupled symmetrically, if some overload edges (i.e., the edges having largest loads) are removed, the network will become more synchronizable. To get the load of each edge need the global information of the network and is of huge computational complexity. To avoid the heavy computational cost, they define the significance of an edge Sij by the product of the degrees of two nodes i and j at both sides of the edge, i.e., Sij = ki ×kj . After calculating the significance of all the edges, they rank the edges according to the values Sij. Subsequently, at each time step, an edge with the highest rank is cut, i.e., decouple the tow nodes at both sides of the edge.

Figure 6 shows the ratio of the eigenratios of the coupling matrix after and before the edges are deleted, as a function of the proportion of the cut-edges. It is clear that as the number of deleted edges increases, the synchronizability is improved, reflected by the decrease of the eigenratio. But the average degree seems has little influence on the results.

4.4 Strengthening the hub-nodes’ influence

From the discussion in Section 2, it is known that when the nodes are coupled symmetrically, it is usually hard for the coupling signals to pass through the hub-nodes. Thus, during the coupling process, the states of the hub-nodes are influenced insignificantly, and the nodes around the hub-nodes will approach the states of the hub-nodes, little by little, but eventually all the nodes are synchronized. Therefore, if the coupling strength from hub-nodes to other nodes are stronger than the reverse, i.e., the nodes are coupled asymmetrically, then the network may actually synchronize much easier. Based on this obseervation, Motter et al. [47−49] proposed a new coupling pattern, in which the coupling matrix has zero row-sums, and the off-diagonal elements are Gij = Lij/kiβ, where β is a tunable parameter and Lij are the elements of the Laplacian, i, j = 1, ···, N. The coupling is weighted when β ≠ 0, but unweighted when β = 0. It was proved that network nodes coupled in this way are much easier to synchronize than being coupled symmetrically, and β = 1 is always the best case.

The aforementioned coupling method only takes into account the degree difference between the nodes that are directly connected, where each node is coupled by its neighbors with the same coupling strength. However, in some real-world networks, different neighbors may have different influences, i.e., the coupling strengths from different neighbors may be different. Thus, based on the assumption that different nodes play different roles in a network, one may adjust the influence strength of each node, received from their neighbors, according to its neighbors’ degrees [50]. The coupling matrix so designed is as follows:

(3)Gij={−kjα/Siβ,for j∈ΛiSi/Siβ,for i=j0,otherwise

where Si = ∑j∈Λi kjα, i = 1, ···, N. When α = β = 0, this coupling pattern degenerates to the symmetric coupling pattern [27], where the case of α = 0 corresponds to the MZK pattern [47], and the case of β = 1 is equivalent to the one introduced in Ref. [49] [see Eq. (15) in Ref. [49] for more details].

Although G is asymmetric for all α with β ≠ 0, it can be proved that all its eigenvalues are non-negative real with only one eigenvalue being equal to zero if the network is connected. To do so, first notice that, Eq. (3) can be written as

(4)G=DL′

where D = diag{k1−αS1−β,k2−αS2−β,⋯,kN−αSN−β} is a diagonal matrix, and L′ = (Lij') is a symmetric matrix with zero row-sums, whose off-diagonal elements are Lij' = kiαkjα,i,j=1,⋯,N. From the identity

(5)det(DL′−λI)=det(D12L′D12−λI)

which is valid for arbitrary λ, where “det” denotes the determinant and I is the N × N identity matrix, one can deduce that the spectrum of eigenvalues of matrix G is equal to the spectrum of a symmetric matrix defined by

(6)H=D12L′D12

As a result, the eigenvalues of matrix G are all nonnegative real and the smallest eigenvalue is always zero.

In this coupling method, for a given parameter β, and for any α > 0, the nodes with larger degrees have stronger influence, but for α < 0, the nodes with less edges are more influential. Parameter β is now exploited so as to eliminate the discrepancies between the coupling signals that each node receives: Given α, when β = 1, each node receives the equal quantum of signals; when β < 1, the nodes having a larger sum of neighbors’ degrees are being influenced more strongly; when β > 1, the contrary is true.

Figure 7(a) shows some numerical values of the eigenratio R on the parameter space (α, β) for BA networks [2]. To clearly show the effects of α and β on R, the values of R are plotted as a function of α for different values of β in Fig. 7(b). No matter what value the parameter β takes, there exists a region of α in which the eigenratio R is smaller than that of the case α = 0. That is, when proper parameters are chosen, this coupling method can be even better than the MZK method. Similarly to the results obtained by the MZK method, β = 1.0 corresponds to the case of best synchronizability. Hereinafter, the discussion will be focused on the case of β = 1.

Note that, in the limit α = +∞(−∞), each node is only influenced by a neighbor having the largest (smallest) degree. Similar situation as mentioned in Ref. [51] appears here: The original network approaches a new configuration connected by some effective directed edges. And the new network, named an equivalent network, may be connected or disconnected. In the disconnected case, the eigenratio R will approach infinity, while in the connected case, the eigenratio equals a constant larger than or equal to 2. Figure 8 illuminates three simple equivalent networks with α = +∞; the former two are connected, and the third one is disconnected. Their eigenratios are 2, 6.8284 and +∞, respectively. Figure 9 shows the changes of the eigenratio R with the parameter α, when β = 1, on different network configurations. When α > 0.4, the eigenratios for different configurations fall apart: some approach 2 or constants not much larger than 2, while some go to infinity, depending on whether the equivalent networks are connected or not. In addition, from the simulations, one can see that with the increasing of the network size, the proportion of networks being disconnected, when α =∞ (−∞), will increase sharply.

The effect of degree-degree correlation on the network synchronizability has also been investigated [52]. Some correlated networks are generated by an extended BA algorithm [53, 54]: starting from m0 fully connected nodes; then, at each time step, a new node is being added to the network and m (< m0) previously existing nodes are chosen to be connected to it, with probability

(7)pi∝ki+k0∑j(kj+k0)

where pi and ki denote the choosing probability and degree of node i, respectively, i = 1, 2, ···. By varying the free parameter k0 (> −m), one can obtain the desired scale-free networks with different assortative coefficients r [55, 56].

Figure 10 shows the relationship between the eigenratio R and the parameter α, for different assortative coefficients, when β = 1. Interestingly, there exists a unique cross point at αcro ≈ −0.25. When α < αcro, a stronger assortative value implies a better synchronizability, while when α > αcro, the opposite is true. It can be proved that in networks of different sizes, the unique cross point always exits, but its position may change with the network size.

5 Conclusion and discussion

This brief review has introduced some recent works on the synchronization of complex networks, especially some methods aiming at improving the network synchronizability. These methods can be divided into two classes, one is to modify the network structure and the other is to regulate the coupling pattern. In the former case, the network structure is changed according to some given rules; in the latter case, the network structure is kept unchanged while the coupling pattern is regulated to strengthen the influence of the hub-nodes on the entire network.

The study on network synchronizability is a rapidly growing area of research, attracting many scientists of various fields, from physics to engineering, to mathematics, to biology alike. Despite the encouraging and promising recent progress, there are still many open questions for further study. For instance, in the case of static coupling patterns, Nishikawa and Motter [35] pointed out that the optimal global synchornizability, with eigenratio being equal to 1, can be obtained from a directed structure without any loops. Even if adding one loop of length 2 (in directed networks, two reverse edges between i and j can be considered as a loop of length 2 as i → j → i), the eigenratio will be doubled [50, 51]. Some further works in this direction will be helpful for the in-depth understanding about the role of loops in network synchronization. Moreover, the synchronization processes for networks coupled in different ways are still unclear. As found in the collective phenomena of biological swarms, a very few effective leaders can well organize the whole population [57], implying that better synchronizabilty may be achieved in the same way. Similarly, a recent study shows that partial coupling is more than enough to sustain the coherence of self-propelled particles [58]. Last but not least, it is recently found that the theory of complementary graphs may also be a key to better understanding and ultimately enhancing the network synchronizability [33]. Therefore, it is believed that how to synchronize a complex network in shorter time with lower cost is a theoretically challenging and practically important research issue, which should be further addressed in the near future.

References

[1]

Watts D. J. and Strogatz S. H., Nature, 1998, 393: 440

[2]

Barabási A. -L. and Albert R., Science, 1999, 286: 509

[3]

Albert R. and Barabási A. -L., Rev. Mod. Phys., 2002, 74: 47

[4]

Dorogovtsev S. N. and Mendes J. F. F., Adv. Phys., 2002, 51: 1079

[5]

Newman M. E. J., SIAM Review, 2003, 45: 167

[6]

Boccaletti S., Latora V., Moreno Y., Chavez M., and Hwang D. -U., Phys. Rep., 2006, 424: 175

[7]

Pastor-Satorras R. and Vespignani A., Phys. Rev, Lett., 2001, 86: 3200

[8]

Yan G., Zhou T., Wang J., Fu Z. -Q., and Wang B. -H., Chin. Phys. Lett., 2005, 22: 510

[9]

Zhou T., Yan G., and Wang B. -H., Phys. Rev. E, 2005, 71: 046141

[10]

Motter A. E. and Lai Y. -C., Phys. Rev. E, 2002, 66: 065102

[11]

Goh K. -I., Lee D. -S., Kahng B., and Kim D., Phys. Rev. Lett., 2003, 91: 148701

[12]

Zhou T. and Wang. B. -H., Chin. Phys. Lett., 2005, 22: 1072

[13]

Tadić B., Thurner S., and Rodgers G. J., Phys. Rev. E, 2004, 69: 036102

[14]

Zhao L., Lai Y. -C., Park K, and Ye N., Phys. Rev. E, 2005, 71: 026125

[15]

Yan G., Zhou T., Hu B., Fu Z. -Q., and Wang B. -H., Phys. Rev. E, 2006, 73: 046108

[16]

Lago-Fernández L. F., Huerta R., Corbacho F., and Sigüenza J. A., Phys. Rev. Lett., 2000, 84: 2758

[17]

Gade P. M. and Hu C. -K., Phys. Rev. E, 2000, 62: 6409

[18]

Hong H., Choi M. Y., and Kim B. J., Phys. Rev. E, 2002, 65: 026139

[19]

Barahona M. and Pecora L. M., Phys. Rev. Lett., 2002, 89: 054101

[20]

Nishikawa T., Motter A. E., Lai Y. -C., and Hoppensteadt F. C., Phys. Rev. Lett., 2003, 91: 014101

[21]

Hong H., Kim B. J., Choi M. Y., and Park H., Phys. Rev. E, 2004, 69: 067105

[22]

Donetti L., Hurtado P. I., and Mu˜noz M. A., Phys. Rev. Lett., 2005, 95: 188701

[23]

Zhao M., Zhou T., Wang B. -H., Yan G., Yang H. -J., and Bai W. -J., Physica A, 2006, 371: 773

[24]

Wang B., Tang H. -W., Zhou T., and Xu Z. -L., arXiv: condmat/0512079

[25]

Wang X. F. and Chen G., Int. J. Bifurcation Chaos Appl. Sci. Eng., 2002, 12: 187

[26]

Wang X. F. and Chen G., IEEE Trans. Circuits and Systems I, 2002, 49: 54

[27]

Pecora L. M. and Carrol T. L., Phys. Rev. Lett., 1998, 80: 2109

[28]

Barahona M, and Pecora L. M., Phys. Rev. Lett., 2002, 89: 054101

[29]

Hu G., Yang J., and Liu W., Phys. Rev. E, 1998, 58: 4440

[30]

Stefański A., Perlikowski P., and Kapitaniak T., Phys. Rev. E, 2007, 75: 016210

[31]

Duan Z., Chen G., and Huang L., arXiv:0706.2899v1 [cs.NI] 20 Jun. 2007

[32]

Duan Z., Chen G., and Huang L., arXiv:0706.2990v1 [cs.NI] 20 Jun. 2007

[33]

Liu C., Duan Z., Chen G., and Huang L., Physica A, accepted, 2007

[34]

Pecora L. M. and Barahona M., Chaos Complexity Lett., 2005, 1: 61

[35]

Nishikawa T. and Motter A. E., Phys. Rev. E, 2006, 73: 065106

[36]

Fan Z. P., Complex Networks: From Topology to Dynamics, PhD Thesis, City University of Hong Kong, 2006

[37]

Zhao M., Zhou T., Wang B. -H., and Wang W. -X., Phys. Rev. E, 2005, 72: 057102

[38]

Barthélemy M., Eur. Phys. J. B, 2004, 38: 163

[39]

Goh K. -I., Kahng B., and Kim D., Phys. Rev. Lett., 2001, 87: 278701

[40]

Zhou T., Zhao M., and Wang B. -H., Phys. Rev. E, 2006, 73: 037101

[41]

Bondy J. A. and Murty U. S. R., Graph Theory with Applications, MacMillan Press LTD, London and Basingstoke, 1976

[42]

Bollobás B., Modern Graph Theory, New York: Springer-Verlag, 1998

[43]

Xu J. -M., Theory and Application of Graphs, Dordrecht: Kluwer Academic, 2003

[44]

Newman M. E. J. and Watts D. J., Phys. Rev. E, 1999, 60: 7332

[45]

Xu J. -M., Topological Structure and Analysis of Interconnection Networks, Dordrecht: Kluwer Academic Publishers, 2001

[46]

Yin C. -Y., Wang W. -X., Chen G., and Wang B. -H., Phys. Rev. E, 2006, 74: 047102

[47]

Motter A. E., Zhou C., and Kurths J., Phys. Rev. E, 2005, 71: 016116

[48]

Motter A. E., Zhou C., and Kurths J., Europhys. Lett., 2005 69: 334

[49]

Motter A. E., Zhou C., and Kurths J., AIP Conf. Proc., 2005, 776: 201

[50]

Zhao M., Zhou T., Wang B. -H., Ou Q., and Ren J., Eru. Phys. J. B, 2006, 53: 375

[51]

Chavez M., Hwang D. -U., Amann A., Hentschel H. G. E., and Boccaletti S., Phys. Rev. Lett., 2005, 94: 218701

[52]

di Bernardo M., Garofalo F., and Sorrentino F., Int. Bifurca J. & Chaos, to be published

[53]

Dorogovtsev S. N., Mendes J. F. F., and Samukhin A. N., Phys. Rev. Lett., 2000, 85: 4633

[54]

Krapivsky P. L. and Redner S., Phys. Rev. E, 2001, 63: 066123

[55]

Newman M. E. J., Phys. Rev. Lett., 2002, 89: 208701

[56]

Newman M. E. J., Phys. Rev. E, 2003, 67: 026126

[57]

Couzin L. D., Krause J., Franks N. R., and Levin S. A. Nature, 2005, 433: 513

[58]

Zhang H. -T., Chen M., and Zhou T., arXiv:0707.3402

Rights & permissions

Higher Education Press and Springer-Verlag 2007

PDF (1563KB)

2035

Accesses

0

Citation

Detail

Sections
Recommended

/

〈 〉