Cascades on correlated and modular networks James P. Gleeson Department of Mathematics and Statistics, University of Limerick, Ireland www.ul.ie/gleesonj Collaborators and funding Sergey Melnik, UL Diarmuid Cahalane, UCC (now Cornell) Rich Braun, University of Delaware Donal Gallagher, DEPFA Bank SFI Investigator Award MACSI (SFI Maths Initiative) IRCSET Embark studentship Some areas of interest Noise effects on oscillators Applications: Microelectronic circuit design Diffusion in microfluidic devices Applications: Sorting and mixing devices Complex systems Agent-based modelling Dynamics on complex networks Applications: Pricing financial derivatives Some areas of interest Noise effects on oscillators Applications: Microelectronic circuit design Diffusion in microfluidic devices Applications: Sorting and mixing devices Complex systems Agent-based modelling Dynamics on complex networks Applications: Pricing financial derivatives Overview Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). Overview Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). What is a network? A collection of N “nodes” or “vertices” which can be labelled i… …connected by links or “edges”, {i,j}. Examples: • World wide web • Internet • Social networks • Networks of neurons • Coupled dynamical systems Examples of network structure The Erdós-Rényi random graph Consider all possible links, create any link with a given probability p. Degree distribution is Poisson with mean z: e z z k pk k! z k pk k 0 Examples of network structure The Small World network Start with a regular ring having links to k nearest neighbours. Then visit every link and rewire it with probability p. [Watts & Strogatz, 1998] Examples of network structure Scale-free networks Many real-world networks (social, internet, WWW) are found to have scale-free degree distributions. “Scale-free” refers to the power law form: p k ~ k Examples [Newman, SIAM Review 2003] Overview Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). Dynamics on networks • Binary-valued nodes: • Epidemic models (SIS, SIR) • Threshold dynamics (Ising model, Watts) • ODEs at nodes: • Coupled dynamical systems • Coupled phase oscillators (Kuramoto model) Global Cascades and Complex Networks Initially small localized effects can propagate over the whole network, causing a global cascade Examples of global cascades: • Epidemics, computer viruses • Spread of fads and innovations • Cascading failures in infrastructure (e.g. power grid) networks Similarity: initial failures increase the likelihood of subsequent failures Cascade dynamics depends strongly on: • Network topology (degree distribution, degree-degree correlations, community structure, clustering) • Resilience of individual nodes (node response function) Structures and dynamics review see: • M.E.J. Newman, SIAM Review 45, 167 (2003). • S.N. Dorogovtsev et al., arXiv:0705.0010 (2007) Watts` model D.J. Watts, Proc. Nat. Acad. Sci. 99, 5766 (2002). Threshold dynamics The network: • aij is the adjacency matrix (N ×N) • un-weighted • undirected aij {0,1} aij a ji The nodes: • are labelled i , i from 1 to N; • have a state ; vi (t ) {0,1} • and a threshold ri from some distribution. Threshold dynamics Node i has state vi (t ) {0,1} and threshold ri Neighbourhood average: r i Updating: 1 ki a v ij j if ri ri 1 vi unchanged otherwise The fraction of nodes in state vi=1 is r(t): j Watts` model D.J. Watts, Proc. Nat. Acad. Sci. 99, 5766 (2002). Watts` model z k e z pk k! P( r ) ( r R ) R Cascade condition: k ( k 1) pk F k1 1 z k 1 r Thresholds CDF: F ( r ) P( s) ds Watts` model Watts: initially activate single node (of N), determine if r at steady state. Us: 1 initially activate a fraction r 0 of the nodes, and determine the steady state value of r . Conditions for global cascades (and dependence on the size of the seed fraction) follow… Main result Our result: k k k 1 m 0 r r0 (1 r0 ) pk qm (1 q )k m F mk m with qn1 r0 (1 r0 )G( qn ), q0 r0 , and k k 1 k 1 m k 1m G( q) pk q (1 q ) F mk k 1 z m 0 m Derivation: Generalizing zero-temperature random-field Ising model results from Bethe lattices (D. Dhar, P. Shukla, J.P. Sethna, J. Phys. A 30, 5259 (1997)) to arbitrary-degree random networks. Results R 0.18 N 105 z k e z pk k! P( r ) ( r R ) r0 103 r 0 5 103 r 0 102 Results r 0 104 z k e z pk k! P( r ) ( r R ) r 0 102 Main result Our result: k k k 1 m 0 r r0 (1 r0 ) pk qm (1 q )k m F mk m with qn1 r0 (1 r0 )G( qn ), q0 r0 , and k k 1 k 1 m k 1m G( q) pk q (1 q ) F mk k 1 z m 0 m 1.5 Cascade condition 1.25 qn1 r0 (1 r0 )G( qn ), q0 r0 , G(q) 1 0.75 0.5 0.25 0.25 0.5 q 0.75 1 1.25 1.5 Cascade condition 1.25 qn1 r0 (1 r0 )G( qn ), q0 r0 , G(q) 1 0.75 0.5 0.25 0.25 0.5 q 0.75 1 1.25 Simple cascade condition qn 1 First-order cascade condition: using qn1 r0 (1 r0 )G( qn ), slope>1 q0 r0 , slope=1 demand qn (1 r0 )G(0) 1 (slope>1) for global cascades to be possible. This yields the condition k (k 1) 1 1 p F F (0) , k k z 1 r0 k 1 reproducing Watts’ percolation result when r0 0 and F (0) 0. Simple cascade condition r 0 104 z k e z pk k! P( r ) ( r R ) r 0 102 Extended cascade condition qn 1 Second-order cascade condition: expand qn1 r0 (1 r0 )G( qn ), q0 r0 , above slope=1 to second order and demand no positive zeros of the quadratic aq2 bq c 0 for global cascades to be possible. The extension is, to first order in r 0 : G(0) 12 2G(0)G(0) 2r0 G(0) G(0)2 G(0) 2G(0)G(0) 0. qn Extended cascade condition r 0 104 R z k e z pk k! P( r ) ( r R ) r 0 102 Gaussian threshold distribution r0 0 0.05 z k e z pk k! ( r R)2 P( r ) exp 2 2 2 2 1 0.2 Gaussian threshold distribution R 0.2 0.05 R 0.362 R 0.38 N 105 r0 0 z k e z pk k! ( r R)2 P( r ) exp 2 2 2 2 1 0.2 Bifurcation analysis qn1 r0 (1 r0 )G( qn ), R 0.35 r0 0; 0.2 q G ( q) 0 R 0.371 R 0.375 Results: Scale-free networks r 0 103 r 0 10 2 P( r ) ( r R ) z 10.5 pk k exp( k ) 100 0.2 0.45 ( r R)2 P( r ) exp 2 2 2 2 r0 0 1 z6 Results: Scale-free networks z k e z pk k! r 0 102 P( r ) ( r R ) pk k exp( k ) 100 Overview Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). Watts` model of global cascades Consider undirected unweighted network of N nodes (N is large) defined by degree distribution pk Each node i has: vi {0,1} • fixed threshold ri • binary state given by thresholds CDF r F (r) (probability that a node has threshold < r) P( s) ds Initially activate fraction ρ0<<1 of N nodes. mi 1 , if ri Updating: node i becomes active if the vi ki active fraction of its neighbours exceeds unchanged otherwise its threshold The average fraction of active nodes r (t ) 1 N N v (t ) i 1 i Derivation of result Derivation: Generalizing zero-temperature random-field Ising model results from Bethe lattices (D. Dhar, P. Shukla, J.P. Sethna, J. Phys. A 30, 5259 (1997)) to arbitrary-degree random networks. Derivation of result Main idea: pick a node A at random and calculate its probability of becoming active. This will give ρ(∞). A Derivation of result Main idea: pick a node A at random and calculate its probability of becoming active. This will give ρ(∞). Re-arrange the network in the form of a tree with A being the root. ∞ A … ……………… n+2 … … … n+1 n ………………….. … qn 1 qn q:nprobability that a node on level n is active, conditioned on its parent (on level n+1) being inactive. Derivation of result Main idea: pick a node A at random and calculate its probability of becoming active. This will give ρ(∞). Re-arrange the network in the form of a tree with A being the root. ∞ qn 1 r 0 A … ……………… n+2 … (1 r 0 ) (initially inactive) … … n+1 n ………………….. (initially active) … qn 1 qn q:nprobability that a node on level n is active, conditioned on its parent (on level n+1) being inactive. Derivation of result Main idea: pick a node A at random and calculate its probability of becoming active. This will give ρ(∞). Re-arrange the network in the form of a tree with A being the root. ∞ qn 1 r 0 A … ……………… n+2 … (1 r 0 ) (initially inactive) … ………………….. p k 1 … n+1 n (initially active) … qn 1 qn k (has degree k; k-1 children) k 1 m qn 1 qn k 1 m m (m out of k-1 children active) k-1 children q:nprobability that a node on level n is active, conditioned on its parent (on level n+1) being inactive. Degree distribution of nearest neighbours: pk k pk . z Derivation of result Main idea: pick a node A at random and calculate its probability of becoming active. This will give ρ(∞). Re-arrange the network in the form of a tree with A being the root. ∞ qn 1 r 0 A … ……………… n+2 … (1 r 0 ) (initially inactive) … ………………….. p k 1 k 1 … n+1 n (initially active) … qn 1 qn k-1 children q:nprobability that a node on level n is active, conditioned on its parent (on level n+1) being inactive. k (has degree k; k-1 children) k 1 m m k 1 m q 1 q F n n m 0 m k (m out of k-1 children active) (activated by m active neighbours) Derivation of result q0 r 0 k 1 k 1 m k 1 m m ~ qn 1 r 0 (1 r 0 ) pk qn 1 qn F k k 1 m 0 m Our result for the average fraction of active nodes k m m r r 0 (1 r 0 ) pk q 1 q k m F k k 1 m 0 m k Valid when: (i) Network structure is locally tree-like (vanishing clustering coefficient). (ii) The state of each node is altered at most once. Conclusions • Demonstrated an analytical approach to determine the average avalanche size in Watts’ model of threshold dynamics. • Derived extended condition for global cascades to occur; noted strong dependence on seed size. • Results apply for arbitrary degree distribution, but zero clustering important. • Further work… Overview Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). Extensions • Generalized dynamics: • SIR-type epidemics • Percolation • K-core sizes • Degree-degree correlations • Modular networks • Asynchronous updating • Non-zero clustering Derivation of result q0 r 0 k 1 k 1 m k 1 m m ~ qn 1 r 0 (1 r 0 ) pk qn 1 qn F k k 1 m 0 m Our result for the average fraction of active nodes k m m r r 0 (1 r 0 ) pk q 1 q k m F k k 1 m 0 m k Generalization to other dynamical models q0 r 0 k 1 k 1 m k 1 m ~ qn 1 r 0 (1 r 0 ) pk qn 1 qn F m, k k 1 m 0 m Our result for the average fraction of active nodes k m r r0 (1 r0 ) pk q 1 q k m F m, k k 1 m 0 m Fraction of active neighbours (Watts): Absolute number of active neighbours: Bond percolation: Site percolation: k m F m, k CDFthr k F m, k CDFthr m F m, k 1 (1 p) m 0, if m 0 F m, k Qk , if m 0 Generalization to other dynamical models q0 r 0 k 1 k 1 m k 1 m ~ qn 1 r 0 (1 r 0 ) pk qn 1 qn F m, k k 1 m 0 m Our result for the average fraction of active nodes k m r r0 (1 r0 ) pk q 1 q k m F m, k k 1 m 0 m k K-core: the largest subgraph of a network whose nodes have degree at least K Initially activate (damage) fraction ρ0 of nodes. A node becomes active if it has fewer than K inactive neighbours: 0, if k m K F m, k 1, if k m K Final inactive fraction (1- ρ) of the total network gives the size of K-core K-core sizes on degree-degree correlated networks Theory vs Numerics: 7-cores in Poisson random graphs with z = 10 r = -0.5 r=0 r = 0.98 Initial damage ρ0 Case r = 0 considered in S.N. Dorogovtsev et al., PRL 96, 040601 (2006). Degree-degree correlated networks Adopt approach of M. Newman for percolation problems (PRE 67, 026126 (2003), PRL 89, 208701 (2002)). P(k,k’) – joint PDF that an edge connects vertices with degrees k, k’ ……………… … … Consider a k-degree node at level n+1: qn( k1) n+1 n ………………….. … – probability that a k-degree node is active (conditioned on its parent being inactive) qn( k ) ( k ) P ( k , k ) q k n k P (k , k ) – probability that a child of an inactive k-degree node is active Degree-degree correlated networks q0( k ) r 0( k ) qn( k ) q (k ) n 1 r (k ) n 1 ( k ) P ( k , k ) q k n r (k ) 0 r (k ) 0 k P (k , k ) (1 r k 1 ( k ) qn ) m 0 m k 1 (k ) 0 (1 r k (k ) ) qn m 0 m k (k ) 0 1 q ( k ) k 1 m n m 1 q m ( k ) k m n r n k pk r n(k ) (Also obtain a cascade condition in matrix form). F m, k F m, k Degree-degree correlated networks Theory (curves) vs Numerics (symbols): 7-cores in Poisson random graphs with z = 10 r = -0.5 (zero initial damage) r=0 r = 0.98 Pearson correlation r r k P(k , k ) k ,k kk P(k , k ) 2 k ,k kP(k , k ) kP(k , k ) k ,k k ,k Initial damage ρ0 2 2 Case r = 0 considered in S.N. Dorogovtsev et al., PRL 96, 040601 (2006). Correlated networks (105 nodes) generated using Gaussian copula. Predicting K-cores in CAIDA internet router network Internet router network structure from www.caida.org k pk k k Degree distribution pk Degree-degree correlation matrix P(k , k ) Predicting K-cores in CAIDA internet router network Internet router network structure from www.caida.org Us: Predicted from analysis of degree distribution and degree-degree correlation. Actual size Predicted from analysis of degree distribution only (see S.N. Dorogovtsev et al., PRL 96, 040601 (2006)). Modular networks; asynchronous updating Similar idea, but instead of P(k,k’) use the mixing matrix e, which quantifies connections between different communities. qn(i ) ( j) e q j ij n e j ij qn(i)1 g (i ) (qn(i ) ) r n(i)1 h (i ) (qn(i ) ) Asynchronous updating gives continuous time evolution: h q r q (i ) g (i ) q (i ) q (i ) r (i ) (i ) (i ) (i ) Modular networks example Summary Structure of complex networks Dynamics on complex networks Derivation of main result Extensions and applications • • J.P. Gleeson and D.J. Cahalane, Phys. Rev. E. 75, 056103 (2007). J.P. Gleeson, Phys. Rev. E. 77, 046117 (2008). Cascades on correlated and modular networks James P. Gleeson Department of Mathematics and Statistics, University of Limerick, Ireland www.ul.ie/gleesonj What is best “random” model for the Internet? Medusa model: Carmi et al., Proc. Nat. Acad. Sci. ‘07 Jellyfish model: Siganos et al., J. Comm. Networks ‘06 Internet structure using router data from CAIDA pk k Transmissibility (bond Occupation probability) DEPFA Bank collaboration: CDO pricing m1 m2 p1 p2 m3 p3 m4 p4 S1 0 to 5% Definitions mi S2 10% to 15% Notional of credit i pi S3 15% to 25% S4 25% to 35% Default probability of credit i, (derived from the CDS quote). Sq mN pN {m1, m2,…,mN} {p1, p2,…,pN} Correlation Structure S5 35% to 100% Fair price for protection against losses in tranche q Problem ? {S1, S2,…,Ss} Existing models fail to reproduce the prices (Sq) observed on the market. An external field Stochastic Dynamics on Networks Hysteresis: PRG Stochastic Dynamics on Networks Hysteresis: PRG Stochastic Dynamics on Networks Stochastic dynamics Aim: Fundamental understanding of the interactions between nonlinear dynamical systems and random fluctuations. External noise sources e.g. transistor noise, thermal noise. Heterogeneity within system e.g. agent-based models, large-scale networks. Tools: • Numerical simulations …guiding fundamental understanding via… • Asymptotic methods • Perturbation techniques • Exact solutions Noise in oscillators (Theme 1) Prof. M. P. Kennedy, Microelectronic Engineering, UCC New computational and asymptotic methods for the spectrum of an oscillator subject to white noise Stochastic perturbation methods for effects of coloured noise Papers: • SIAM J. Appl. Math. • IEEE TCAS Collaboration (Feely/Kennedy): Noise effects in digital phaselocked loops Microfluidic mixing and sorting (Theme 3) Experimentalists at Tyndall National Institute, Cork Analysis of MHD micromixing in annular geometries Modelling of micro-sorting methods Papers: • SIAM J. Appl. Math. • Phys. Rev E • Phys. Fluids Collaborations: (Lindenberg/Sancho) Noise-induced sorting techniques for microparticles Cascades on correlated and modular networks James P. Gleeson Department of Mathematics and Statistics, University of Limerick, Ireland www.ul.ie/gleesonj