From Local Network Motifs to Global Invariants Victor M. Preciado and Ali Jadbabaie Department of Electrical and Systems Engineering University of Pennsylvania Third Year Review, October 29, 2010 ONR MURI: NexGeNetSci Preciado Local Motifs and Global Invariants Theory • First principles • Rigorous math • Algorithms • Proofs Data Analysis • Correct statistics • Only as good as underlying data Lab Numerical Experiments Experiments • Simulation • Synthetic, clean data • Stylized • Controlled • Clean, real-world data Field Exercises Real-World Operations • SemiControlled • Messy, real-world data • Unpredictable • After action reports in lieu of data Outline • • • • • Motivation and context The role of local structural information Spectral analysis from local structural information Bounds on spectral properties via optimization Implications in dynamical processes ONR MURI: NexGeNetSci Complex Network: Properties • Generic features: – Large number of nodes – Sparse connectivity – Lack of regularity • Examples: – Comm networks (e.g. Internet) – Social networks (e.g. Facebook) – Biological networks • We assume limited structural information: ?? ?? – Privacy and/or security concerns – Storage/computing limitations ONR MURI: NexGeNetSci Complex Networks: Some Challenges • Challenges when only local structural information is available: – Estimation: How could we aggregate local measurements to infer global properties of the network? – Inference: What could we say about the behavior of a dynamical process in the network from local measurements? – Actuation: How could we modify the structure of a network to induce a desired global behavior? Actuation Estimation Inference Usual Approach in “Network Science” • Overly focused on random graph models and degree distributions, but we can have very different networks with the same degree distribution [Li et al., 2005]: • Main drawbacks: 1. Degree distributions are a zero-th order approximation of the network structure, by far not enough 2. Random models are difficult, if not impossible, to justify from an engineering perspective ONR MURI: NexGeNetSci More Structured Random Models • We also find random graph models capturing increasingly richer structural properties [Mahadevan et al., 2006] Average degree Degree distribution Distribution of Triangles • Joint Degree Distribution Original HOT model Main drawbacks: – Visual inspection is clearly not enough to measure similarity – What structural measurements are relevant in the behavior of dynamical processes in networks? ONR MURI: NexGeNetSci Networks = Graphs + Dynamics • Our framework: We consider the dynamical behavior of networks • Since the eigenvalues and eigenvectors are closely related with the network dynamical behavior, spectral graph theory is a convenient framework to study network dynamics 800 0 500 600 1000 400 1500 r(A) 200 2000 0 500 1000 1500 nz = 45572 2000 0 -40 -20 0 20 40 60 • Some relationships between spectra and dynamics are: – Spreading processes – Synchronization – Diffusion/Consensus adjacency spectral radius (combinatorial) Laplacian eigenratio (normalized) Laplacian eigenvalues ONR MURI: NexGeNetSci 80 Inference from Local Measurements Our problems: 1. - What measurements are most relevant in the behavior of dynamical processes? 2. - How can we aggregate local measurements to say something about the global dynamical behavior? Aggregation of local structural measurements • Dynamical Implications We study those problems in the framework of spectral graph theory and convex optimization, without making any assumption on the global network structure (i.e., no random models) ONR MURI: NexGeNetSci Structure of our Approach Use algebraic graph theory to relate the frequency of certain small subgraphs, or motifs, with the so-called spectral moments of the network II. Propose a distributed technique to compute the frequency of subgraphs from the distribution of local network measurements III. Use convex optimization to extract relevant spectral information from a sequence of spectral moments IV. Study implications on dynamical processes I. ONR MURI: NexGeNetSci I. From Subgraphs to Spectral Moments • Algebraic graph theory allows us to compute spectral moments from local structural information. We use the following result: 1 n k 1 mk ( A) i # closed walks of length k n i 1 n • Low-order moments: For k≤3 we have the following expressions k=2 k=3 i i m2 1 2e d i n n i m3 1 6 2 t i n n i ONR MURI: NexGeNetSci From Subgraphs to Moments • Higher-order moments: As we increase the order of the moments, a variety of more and more complicated subgraphs come into the picture. For k=4, we have the following types of closed walks: i i i i Moments from local structural measurements j m4 1 2qi n i m4 1 ( 8Q n (d i 1)d i (d i 1)d i 4 di 8e Moments from subgraphs frequencies • In the first expression, we observe that local measurements can be aggregated via distributed consensus to compute spectral moments ONR MURI: NexGeNetSci Moments from Subgraphs Frequencies • Key observation: The spectral moments are linear combinations of subgraphs embedding frequencies [Preciado, Jadbabaie, 2010]. The coefficients for all nonisomorphic connected subgraphs with 4 or less nodes are mk k=4 2 4 - - - - 8 - - k=5 - - 30 - - 10 - - - k=6 2 12 24 12 6 - 48 36 - k=7 - - 126 - - 84 - 112 - k=8 2 28 168 72 32 64 264 464 528 • For example, the fifth moment can be computed as: 1 m5 ( 30 10 30 e n ONR MURI: NexGeNetSci II. Distributed Computation of Moments • We propose a distributed technique to compute subgraph frequencies. Note that each subgraph can be ’discovered’ by a number of its nodes. For example, for 1-hop neighborhoods: 1 i i n n 3 1 1 i i i i (d i 2) n n n Q 1 ??? n n i • In general, if each node have access to its r-hops neighborhood, we can discover all subgraphs involved in moments of order up to 2r+1 (and part of the subgraphs involved in moments of higher order) ONR MURI: NexGeNetSci Empirical Example • Subgraph with 2,404 nodes and 22,786 edges obtained from crawling the Facebook graph in a breadth-first search around a particular node Node Degree Distribution -1 Distribution of Triangles -1 10 4 Degree-Triangle Scatter Plot 10 10 3 10 -2 -2 10 t,Triangles P(t) P(k) 10 -3 -3 2 10 10 10 1 10 -4 -4 10 0 10 1 2 10 10 k,Degree 3 10 10 0 0 10 1 10 2 10 t,Triangles 3 10 • We can compute the relevant quantities which allow us to compute moments ONR MURI: NexGeNetSci 4 10 10 0 10 1 2 10 10 k,Degree 3 10 III. Extracting Spectral Info from Moments • So far, we have Counting subgraph frequencies Computing spectral moments ? • We now present an SDP-based approach to extract information from the spectral moments that 1. It is agnostic, in the sense that it does not make any assumption on the global network structure (no random model) 2. It allows to study the effect of arbitrarily complicated structural measurements in the network spectral properties ONR MURI: NexGeNetSci The Classical Moment Problem • How can we extract information from spectral moments? The following problem, called the classical moment problem, is closely related to ours: Given a sequence of moments (m1,…,mk), and Borel measurable sets T W R, we are interested in computing where m in M(W), M(W) being the set of positive Borel measures supported by W. • Generalization of Markov and Chebyshev´s inequalities from probability theory, when a sequence of moments is available ONR MURI: NexGeNetSci Moment Problems, SOS and SDP • Using duality theory, we obtain the following formulation [Bertsimas, 2005]: This dual problem is a sum-of-squares program (SOSP) and can be formulated as a semidefinite program [Parrilo, 2006]. • We define the spectral distribution of a graph as and define the r.v. XmG. Hence, using SOS, we can compute optimal bounds on Pr(X T)=#{i T}/n when we have access to a sequence of spectral moments ONR MURI: NexGeNetSci Numerical Results • From the set of spectral moments, we compute optimal bounds on #{i [a,b]}/n, and #{i [-c,c]}/n B b 1 a* a C c • Notice that only those intervals [a,b] in region B and [-c,c] in C are able to support the whole set of eigenvalues. Hence, – We have a lower bound on the spectral radius r(A)>a* – We can also compute a bound on the Laplacian eigenratio from the Laplacian spectral moments ONR MURI: NexGeNetSci IV Dynamical Implications: Spreading Processes We study a stochastic dynamical model of viral dissemination: - Each node has two possible b b d states: 0. Susceptible (blue) 1. Infected (red) - Spreading parameters: b probability of contagion d probability of recovering Spectral results [Draief et al., 2008]: - r(A)>d/b is a necessary condition for a small infection to infect a significant part of the network - The larger r(A), the better a network disseminate a virus/rumor ONR MURI: NexGeNetSci Spreading Processes: Simulations Counting subgraph frequencies Computing spectral moments Bound on r(A)>45.0 0.2% initial infection d/b=65 > r(A)=60 300 300 250 250 # of infections # of infections 0.05% initial infection d/b=35 < 45 < r(A)=60 200 150 100 50 0 t 0 20 40 60 80 100 120 140 Implications on Spreading 160 ONR MURI: NexGeNetSci 200 150 100 50 0 t 0 10 20 30 40 50 Decentralized Network Design • Ongoing work: Design incentives for each individual to modify their local neighborhood in order to achieve a particular global spectral property. Some preliminaries results [Preciado et al., 2010]: ONR MURI: NexGeNetSci Future Work • Adapt our framework to: 1. Evolving networks: Tracking the evolution of subgraphs frequencies and model their interactions 2. Links with weights and directions: Since the eigenvalues become complex, we have to work with 2D support 3. Nodes with attributes • Incentive design: How can we drive nodes to take local actions that improve the global dynamical behavior of the network? ONR MURI: NexGeNetSci Conclusions • Our work is devoted to study local structural properties and dynamical processes in large-scale complex networks • There is a direct relationship between many dynamical processes in networks and the eigenvalues of the underlying graph • There is plenty of information about the eigenvalue spectra from the distribution of local network measurements • Our approach is agnostic, in which we do not assume any global structure (no random graphs) • Our results can be of interest to analyze and design largescale networks from a spectral point of view ONR MURI: NexGeNetSci Some References • • • • • • • • • • F. Chung, L. Lu, and V. Vu, "The Spectra of Random Graphs with Given Expected Degrees," Internet Mathematics, vol. 1, pp. 257-275, 2003. M. Draief, A. Ganesh, and L. Massoulié, "Thresholds for Virus Spread on Networks," Annals of Applied Probability, vol. 18, pp. 359-378, 2008. L. Li, D. Alderson, J.C. Doyle, and W. Willinger, " Towards a Theory of Scale-Free Graphs,“ Interner Math, vol. 2, pp. 431-523, 2005. P. Parrilo, Algebraic Techniques and Semidefinite Optimization, Massachusetts Institute of Technology: MIT OpenCourseWare, Spring 2006. L.M. Pecora, and T.L. Carroll, "Master Stability Functions for Synchronized Coupled Systems," Physical Review Letters, vol. 80(10), pp. 2109-2112, 1998. I. Popescu and D. Bertsimas, "An SDP Approach to Optimal Moment Bounds for Convex Classes of Distributions," Mathematics of Operation Research, vol. 50, pp. 632-657, 2005. V.M. Preciado, and G.C. Verghese, "Synchronization in Generalized Erdös-Rényi Networks of Nonlinear Oscillators," IEEE Conference on Decision and Control, pp. 4628-4633, 2005. V.M. Preciado, and A. Jadbabaie, "Spectral Analysis of Virus Spreading in Random Geometric Networks," IEEE Conference on Decision and Control, pp. 4802-4807, 2009. V.M. Preciado, M.M. Zavlanos, A. Jadbabaie, and G.J. Pappas, “Distributed Control of the Laplacian Spectral Moments of a Network,” American Control Conference, 2010. V.M. Preciado and A. Jadbabaie, " From Local Measurements to Network Spectral Properties: Beyond Degree Distributions, " IEEE Conference on Decision and Control, 2010. ONR MURI: NexGeNetSci • QUESTIONS? ONR MURI: NexGeNetSci IV.b Dynamical Implications: Synchronization • We study a collection of resistively coupled nonlinear oscillators g g Network dynamics: g g g Question: What values of g do make the network synchronize? Spectral results [Pecora and Carrol]: L=D-A, and ki are its eigenvalues For stability of the synchronous state we need the Laplacian eigenration kn/k2 < t , where t depends on the individual oscillator dynamics ONR MURI: NexGeNetSci Laplacian Moments • • • Using our SDP approach, we can also bound the Laplacian eigenratio from local structural properties via the Laplacian moments In the Laplacian moments, not only the frequencies of subgraphs are important, but also the degrees of the nodes involved. For example, for the 4th moment the following substructures are involved The Laplacian moments are functions of the frequencies of these structures and the degrees of the nodes involved: m4 ( L) ( 1 2 2 8P4c 6P (k1 k 2 k 3 4P 4PE k1 k 2 k1k 2 1 n k1 ,k2 ,k3 ,k4 k1 , k 2 , k3 k1 , k 2 , k3 k1 , k 2 , k3 ONR MURI: NexGeNetSci Synchronization: Simulations • We simulate a network of 200 resistively coupled Rossler oscillators Counting frequencies of substructures Computing Laplacian moments Bound on kn/k2 ONR MURI: NexGeNetSci Implications on Synchronization