Spectral Analysis and Optimal Synchronizability of Complex Networks 复杂网络谱分析及最优同步性能研究 Guanrong (Ron) Chen City Univesity of Hong Kong Synchronization Synchrony can be essential 2009 “Clock synchronization is a critical component in the operation of wireless sensor networks, as it provides a common time frame to different nodes.” IEEE Signal Processing Magazine (2012) Contents Spectra of network Laplacian matrices Spectra and synchronizability of some typical complex networks Networks with best synchronizability A General Dynamical Network Model A linearly and diffusively coupled network: N xi f ( xi ) c aij H ( x j ) j 1 f (.) – Lipschiz xi R n i 1,2,..., N Coupling strength c > 0 A = [aij] – Adjacency matrix H – Coupling matrix function If there is a connection between node i and node j (j ≠ i), then aij = aji = 1 ; otherwise, aij = aji = 0 and aii = 0, i = 1, … , N Laplacian matrix L = D – A where D = diag{ d1 , … , dN } For connected networks: 0 1 2 N Network Synchronization N xi f ( xi ) c aij H ( x j ) xi R n i 1,2,..., N j 1 Complete state synchronization: lim || xi (t ) x j (t ) || 2 0, t i, j 1,2, , N Linearized at equilibrium s : N xi [ Df ( xi )] xi s c aij [ DH ]xi s xi j 1 f (.) Lipschitz, or assume: || f ( s) || M Only c[aij ] or {ci } is important Network Synchronization: Criteria N xi [ Df ( xi )]s c aij [ DH ]s xi j 1 0 1 2 N Master stability equation: (L.M. Pecora and T. Carroll, 1998) yi [[ Df ]s [ DH ]s ] yi , : c2 A (conservative) sufficient condition: Maximum Lyapunov exponent Lmax 0 which is a function of Lmax S max is called a synchronization region Synchronizing: if 0 1 or if 0 2 3 Network Synchronization: Criteria Recall: Eigenvalues of L D A synchronizing if Case I: No sync 0 1 2 N 0 1 c2 or if 0 2 Case II: Sync region Case III: Sync region S1 (1 , ) S2 ( 2 , 3 ) 2 bigger is better 2 bigger is better N 2 3 N Case IV: Union of intervals A Brief History Synchronizability characterized by Laplacian eigenvalues: 1. unbounded region (X.F. Wang and G.R. Chen, 2002) 2 , 0 1 2 N S1 (1 , ) 2. bounded region (M. Banahona and L.M. Pecora, 2002) 2 / N , 0 1 2 N S2 ( 2 , 3 ) 3. union of several disconnected regions S m (1 , 2 ) ( 3 , 4 ) ( m , ) (A. Stefanski, P. Perlikowski, and T. Kapitaniak, 2007) (Z.S. Duan, C. Liu, G.R. Chen, and L. Huang, 2007 - 2009) Spectra of Networks Some theoretical results Relation with network topology Role in network synchronizability Spectrum Theoretical Bounds of Laplacian Eigenvalues d max , d min , d avg - maximum, minimum, average degree D – Diameter of the graph N N 2 d min d max N 2d max N 1 N 1 d min 2 d max N 1 2 ... N N ND d avg N ( A) Graph Theory Textbooks F.M. Atay, T. Biyikoglu, J. Jost, Network synchronization: Spectral versus statistical properties, Physica D, 224 (2006) 35–41. Theoretical Bounds of Laplacian Eigenvalues d (d1 , d 2 ,..., d N )T 1 , 2 ,..., N T (both in increasing order) || d ||1 || d ||2 N || d ||2 || d ||2 || d ||1 Distribution: (interlacing property) For any node degree d i there exists a * j | j 1,2,..., N such that d i d i * d i d i C. J. Zhan, G. Chen and L. F. Yeung, Physica A (2010) 1 , 2 ,..., N T || d ||1 || d ||2 N || d ||2 || d ||2 || d ||1 d (d1 , d 2 ,..., d N )T Lemma (Hoffman-Wielanelt, 1953) For matrices B – C = A: 2 2 | ( B ) ( C ) | || A || i1 i i F n Frobenius Norm For Laplacian L = D – A: 2 2 | ( L ) d | || A || i1 i i F n || A ||2F i 1 j 1| aij |2 i 1 di n 2 || d ||2 || d || 2 d d i 2 i || d ||1 di N N || d ||2 d 2 / N d i || d ||1 i Cauchy inequality n n Difference between Eigenvalue and Degree Sequences 1000 d1000 42 ER random-graph networks N=2500, average of 2000 runs BA scale-free networks N=2500, average of 2000 runs Above: relative variances below: relative errors C. J. Zhan, G. Chen and L. F. Yeung, Physica A (2010) Upper Bounds for Largest Eigenvalues N N - for any graph N max{ du dv | edges (u, v) } N max{ d u d u } d u - average degree of all neighbors of node u N max{ du dv | Nu Nv | | nodes u v in G } Nu N v - total number of common neighbors of u and v W.N. Anderson, T.D. Morley, Eigenvalues of the Laplucian of a graph, Linear and Multilinear Algebra, 18 (1985) 141-145. R. Merris, A note on Laplacian graph eigenvalues, Linear Algebra and its Applications, 285 (1998) 33-35. O. Rojo, R. Sojo, H. Rojo, An always nontrivial upper bound for Laplacian graph eigenvalues, Linear Algebra and its Applications, 312 (2000) 155-159. Lower Bounds for Smallest Eigenvalues 2 2[cos( / N ) cos(2 / N )] 2 cos( / N )(1 cos( / N ))d max 2 2e(1 cos( / N )) 4e sin 2 ( / 2 N ) e v 2 1/( ND) e – edge connectivity (number of edges whose removal will disconnect the graph e.g., for a chain, e = 1 ; for a triangle, e = 2) v – node connectivity D – diameter Spectra of Typical Complex Networks Regularly-connected networks Complete graph, ring, chain, star … Random-graph networks For every pair of nodes, connect them with a certain probability. Poisson node distribution Small-world networks Similar to random graph, but with short average distance and large clustering, with near-Poisson node distribution Scale-free networks Heterogeneous, with power-law degree distribution Piet Van Mieghem, Graph Spectra for Complex Networks (2011) Regularly-Connected Networks Complete graphs: 0, 2 ... N N 2K-rings: i 4sin 2 , i 1, 2,..., N 1 K=1: N N i N - even: 0, 4 sin 2 , i 1,2,..., K≠1: Chains: Stars: 2 / N 1 1, 4, 2 N K 2 (i 1)l i 2 K 2 cos , i 2,3,..., N N l 1 N 1,. .. , N 1 2 i 0, 4 sin 2 , i 1,2, .. , N 1 2 N 0, 2 ... N 1 1 N N 2 / N 1 / N Rectangular Random Networks E. Estrada and G. Chen (2015) RRN: N nodes are randomly uniformly and independently distributed in a unit rectangle [a, b]2 R 2 with a b 1 (It can be generalized to higher-dimensional setting) Two nodes are connected by an edge if they are inside a ball of radius r > 0 Example: N =250 a=1 r = 0.15 Theorem: For RRN, eigenvalues are bounded by 1 2 8(ar ) 2 2 log 2 N 2 4 ( N 1) N N a 1 Lower bound The worst case: all nodes are located on the diagonal 2 1 1 ND N ( N 1) and N N E. Estrada and G. Chen (2015) 1 2 8(ar ) 2 2 log 2 N 2 4 ( N 1) N N a 1 Upper bound Lemma 1: Diameter D = diagonal length / r a4 1 D ar Lemma 2: Based on a result of Alon-Milman (1985) 2 2 8d max log 2 N 2 D E. Estrada and G. Chen (2015) Spectral Role in Network Synchronization Topology affects Eigenvalues affects Sync 2 2 S / S • Recall: d min 2 d max N S a iS jG S ij S - any subgraph, with 0 < |S| N/2 (bigger is better) Implication: small local changes affect eigenratio, hence network synchronizability, which however do not affect much the network statistical properties in general: degree distribution, clustering, average distance, … Statistical properties depend on H, yet 2 depends on S Statistics Determines Synchronizability? Answer - Not necessarily Example: Graph G1 Graph G2 • They have same structural characteristics: Graphs G1 and G2 have same degree sequence {3,3,3,3,3,3} • So, all nodes have degree 3 ; same average distance 7/5 ; and same node betweenness 2 ; same …… Z. S. Duan, G. R. Chen and L. Huang, Phys. Rev. E, 2007 G1 G2 But they have different synchronizabilities: Eigenvalues of G1 are 0, 3, 3, 3, 3 and 6 Eigenvalues of G2 are 0, 2, 3, 3, 5 and 5 2 (G1 ) 3 2 (G2 ) 2 Eigenratio r 2 / N satisfies: r (G1 ) 0.5 r (G2 ) 0.4 Topology Determines Synchronizability ? Answer: Not necessarily • Lemma 1: For any given connected undirected graph G , all its nonzero eigenvalues will not decrease with the number of added edges; namely, by adding any edge e , one has i (G e) i (G), i 1,2,.., N • Note: 2 2 N 2 N Z. S. Duan, G. R. Chen and L. Huang, Phys. Rev. E, 2007 What Topology Good Synchronizability ? Example Given Laplacian: 2 0 0 0 0 2 1 1 0 1 3 0 L 0 1 0 3 1 0 1 1 1 0 1 1 1 1 0 0 1 1 1 1 4 1 1 4 Q: How to replace 0 and -1 while keeping the connectivity (and all row-sums = 0), such that 2 N = maximum ? Answer: 3 1 0 1 0 1 1 3 1 0 1 0 0 1 3 1 0 1 * L 1 0 1 3 1 0 0 1 0 1 3 1 1 0 1 0 1 3 2 N = maximum Observation: Homogeneous + Symmetrical Topology Problem With the same numbers of node and edges, while keeping the connectivity, what kind of network has the best possible synchronizability? Max LL* 2 N L * - set of N N Laplacian matrices Such that “row-sum = 0” and (Laplacian) 2 0 (connected) Computationally, this is NP-hard Objective: Find analytic solutions Progress • Nishikawa et al., Phys. Rev. Lett. 91, 014101(2003), regular networks with uniform small (node or edge) betweenness [we found: edge betweenness is more important than node betweenness] • Donetti et al., Phys. Rev. Lett. 95, 188701(2005) Entangled networks have biggest eigenratios [we found: not necessarily] • Donetti et al., J. Stat. Mech.: Theory and Experiment 8, 1742(2006), algorithm based on algebraic graph theory; big spectral gap big eigenratio [we found: the opposite] • Zhou et al., Eur. Phys. J. B. 60, 89(2007), algorithm based on smallest clustering coefficient • Hui, Ann Oper. Res., July (2009), algorithm based on entropy • Xuan et al., Physica A 388, 1257(2009), algorithm based on short average path length • Mishkovski et al., ISCAS, 681(2010), fast generating algorithm • Shi, Chen, Thong, Yan., IEEE CAS Magazine, 13, 1(2013), 3regular optimal graphs Our Approach • Homogeneity + Symmetry • Same node degree d1 dN • Minimize average path length l1 • Minimize path-sum • Maximize girth g1 li j i lij gN D. Shi, G. Chen, W. W. K. Thong and X. Yan, IEEE Circ. Syst. Magazine, (2013) 13(1) 66-75 lN Non-Convex Optimization Illustration: White:Optimal solution location Grey:networks with same numbers of nodes and edges Green:degree-homogeneous networks Blue:networks with maximum girths Pink:possible optimal networks Red:near homogenous networks Optimal 3-Regular Networks Optimal 3-Regular Networks Summary Optimal network topology (in the sense of having best possible synchronizability) should have o Homogeneity o Symmetry o Shortest average path-length o Shortest path-sum o Longest girth o Anything else ? Open Problem Looking for optimal solutions: 1 d1 1 0 1 d 0 2 Given: 0 0 1 0 0 d N Move which -1 1 d1 1 0 1 d 0 2 0 0 1 0 0 d N 1 d1 1 0 1 d 1 0 2 0 1 0 1 0 0 d N can maximize 2 ? N can maximize 2 ? N Where to delete -1 …………………… can maximize 2 ? N Where to add -1 And so on …… ??? Thank You ! Homogeneity Symmetry References [1] F.Chung, Spectral Graph Theory, AMS (1992, 1997) [2] T.Nishikawa, A.E.Motter, Y.-C.Lai, F.C.Hoppensteadt, Heterogeneity in oscillator networks: Are smaller worlds easier to synchronize? PRL 91 (2003) 014101 [3] H.Hong, B.J.Kim, M.Y.Choi, H.Park, Factors that predict better synchronizability on complex networks, PRE 65 (2002) 067105 [4] M.diBernardo, F.Garofalo, F.Sorrentino, Effects of degree correlation on the synchronization of networks of oscillators, IJBC 17 (2007) 3499-3506 [5] M.Barahona, L.M.Pecora, Synchronization in small-world systems, PRL 89 (5) (2002)054101 [6] H.Hong, M.Y.Choi, B.J.Kim, Synchronization on small-world networks, PRE 69 (2004) 026139 [7] F.M. Atay, T.Biyikoglu, J.Jost, Network synchronization: Spectral versus statistical properties, Physica D 224 (2006) 35–41 [8] A.Arenas, A.Diaz-Guilera, C.J.Perez-Vicente, Synchronization reveals topological scales in complex networks, PRL 96(11) (2006 )114102 [9] A.Arenas, A.Diaz-Guilerab, C.J.Perez-Vicente, Synchronization processes in complex networks, Physica D 224 (2006) 27–34 [10] B.Mohar, Graph Laplacians, in: Topics in Algebraic Graph Theory, Cambridge University Press (2004) 113–136 [11] F.M.Atay, T.Biyikoglu, Graph operations and synchronization of complex networks, PRE 72 (2005) 016217 [12] T.Nishikawa, A.E.Motter, Maximum performance at minimum cost in network synchronization, Physica D 224 (2006) 77–89 [13] J.Gomez-Gardenes, Y.Moreno, A.Arenas, Paths to synchronization on complex networks, PRL 98 (2007) 034101 [14] C.J.Zhan, G.R.Chen, L.F.Yeung, On the distributions of Laplacian eigenvalues versus node degrees in complex networks, Physica A 389 (2010) 17791788 [15] T.G.Lewis, Network Science –Theory and Applications, Wiley 2009 [16] T.M.Fiedler, Algebraic Connectivity of Graphs, Czech Math J 23 (1973) 298305 [17] W.N.Anderson, T.D.Morley, Eigenvalues of the Laplacian of a graph, Linear and Multilinear Algebra 18 (1985) 141-145 [18] R.Merris, A note on Laplacian graph eigenvalues, Linear Algebra and its Applications 285 (1998) 33-35 [19] O.Rojo, R.Sojo, H.Rojo, An always nontrivial upper bound for Laplacian graph eigenvalues, Linear Algebra and its Applications 312 (2000) 155-159 [20] G.Chen, Z.Duan, Network synchronizability analysis: A graph-theoretic approach, CHAOS 18 (2008) 037102 [21] T.Nishikawa, A.E.Motter, Network synchronization landscape reveals compensatory structures, quantization, and the positive effects of negative interactions, PNAS 8 (2010)10342 [22] P.V.Mieghem, Graph Spectra for Complex Networks (2011) [23] D. Shi, G. Chen, W. W. K. Thong and X. Yan, Searching for optimal network topology with best possible synchronizability, IEEE Circ. Syst. Magazine, (2013) 13(1) 66-75 Acknowledgements Prof Ernesto Estrada, University of Strathclyde, UK Prof Xiaofan Wang, Shanghai Jiao Tong University Prof Zhi-sheng Duan, Peking University Prof Jun-an Lu, Wuhan University Prof Dinghua Shi, Shanghai University Dr Juan Chen, Wuhan University Dr Choujun Zhan, City University of Hong Kong Dr Wilson W K Thong, City University of Hong Kong Dr Xiaoyong Yan, Beijing Normal University