Hyperbolic Geometry of Complex Network

Hyperbolic Geometry of Complex Network

Duality between static and dynamic networks Dmitri Krioukov Massimo Ostilli CAIDA/UCSD NIST-Bell Labs Workshop on Large-Scale Networks October 2013 Motivation • Real networks are growing, but static network models exist and are widely used – Simpler – More tractable • In a recent work, a minimalistic growing extension of a “very static” model describes remarkably well the growth dynamics of some real networks (F. Papadopoulos et al., Nature, 489:537, 2012) • Informally: – One camp: there cannot be any connection – Another camp: there should be some connection • Preferential attachment versus configuration model: what is the difference??? Graph ensemble • Set of graphs 풢 • With probability measure 푃 퐺 , 푃(퐺) = 1 퐺∈풢 Random graphs with hidden variables • Given N nodes • Assign to each node a 휅푠 휅푡 random variable 휅 푝푁(휅푠, 휅푡) drawn from distribution 휌푁 휅 • Connect each node pair 푠, 푡 with probability 푝푁 휅푠, 휅푡 Example: Classical random graphs • Given N nodes • Hidden variables: None 휅 휅 • Connection 푠 푝 푡 probability: 푝푁 휅푠, 휅푡 = 푝 • Results: – Degree distribution: Poisson (휆 = 푝푁) – Clustering: Weak Example: Soft configuration model • Given N nodes • Hidden variables: Expected degrees 휅 휅푠휅푡 −훾 휅 푝 ~ 휅 휌 휅 ~ 휅 푠 푠푡 푁 푡 • Connection probability: 휅 휅 푝 휅 , 휅 = 푠 푡 푁 푠 푡 푘 푁 • Results: – Degree distribution: Power law (푃 푘 ~ 푘−훾) – Clustering: Weak Example: Random geometric graphs • Given N nodes • Hidden variables: Node coordinates 휅 in a space 휌 휅 : uniform in the space 푁 휅푠 휅푡 • Connection probability: 푑 휅푠, 휅푡 < 푅 푝푁 휅푠, 휅푡 = Θ 푅 − 푑푠푡 • Results: – Clustering: Strong – Degree distribution: • Euclidean or spherical space: Poisson • Hyperbolic or de Sitter spaces: Power law Probability measure • Let – 퐺 = 푎푠푡 the adjacency matrix – 휅 = 휅1, 휅2, … , 휅푁 a hidden variable assignment – 푝푠푡 = 푝 휅푠, 휅푡 the connection probability matrix • Then 푎푠푡 1−푎푠푡 – 푃 퐺 휅 = 푠<푡 푝푠푡 1 − 푝푠푡 푁 – 휌푁 휅 = 푡=1 휌푁 휅푡 – 푃 퐺 = ∫ 푃 퐺 휅 휌푁 휅 푑휅 Static versus dynamic graphs with hidden variables • Static ensemble 풢푆 • Dynamic ensemble 풢퐷 – Given N nodes – For each new node 푡 = 1,2, … – Assign to each node – Assign to 푡 a random 푡 = 1,2, … , 푁 a random variable 휅푡 drawn from variable 휅푡 drawn from distribution 휌푡 휅푡 distribution 휌푁 휅푡 – Connect each node pair – Connect 푡 to each existing 푠, 푡 with probability node 푠 with probability 푝푁 휅푠, 휅푡 푝푡 휅푠, 휅푡 Equivalence • Let dynamic 푡 = 1,2, … , 푁 • Two ensembles 풢푆 and 풢퐷 are identical (풢푆 = 풢퐷) if 푃푆 퐺 = 푃퐷 퐺 for all 퐺 ∈ 풢 • Distributions 푃푆 퐺 = 푃퐷 퐺 if – 휌푆 휅 = 휌퐷 휅 and – 푝푆 휅푠, 휅푡 = 푝퐷 휅푠, 휅푡 • The difference is – 휌푁 휅푡 (static) versus 휌푡 휅푡 (dynamic) – 푝푁 휅푠, 휅푡 (static) versus 푝푡 휅푠, 휅푡 (dynamic) Weak equivalence • Ensembles 풢푆 and 풢퐷 are weakly equivalent if 풢푆 = 풢퐷 for some 푁 • The simplest example – Dynamic 휌푡 휅푡 is equal to static 휌푁 휅푡 – Dynamic 푝푡 휅푠, 휅푡 is equal to static 푝푁 휅푠, 휅푡 – The difference is only in node labeling Strong equivalence • Ensembles 풢푆 and 풢퐷 are strongly equivalent if 풢푆 = 풢퐷 for any 푁 • In this case the connection probability cannot depend on graph size 푁 – 푝푁 휅푠, 휅푡 = 푝푡 휅푠, 휅푡 = 푝 휅푠, 휅푡 • The simplest example – 휌푁 휅푡 = 휌푡 휅푡 = 휌 휅푡 – Graphs are dense because average degree 푘 = 푁 ∬ 휌 휅 푝 휅, 휅′ 휌 휅′ 푑휅 푑휅′ = 푂 푁 • In sparse strongly equivalent ensembles, the connection probability does not depend on graph size, but the hidden variable distribution does Example: Classical random graphs • Ensemble 풢푁,푝 is strongly equivalent and dense since 푘 = 푁푝 – The growing definition is: connect new nodes 푡 to existing nodes 푠 < 푡 with probability 푝 • An attempt to fix: ensemble 풢푁,푘 : 푘 – Static: connect all node pairs with probability 푝 = 푁 • The ensemble is 풢푁,푝 • The degree distribution is Poisson – Dynamic: connect new nodes 푡 to existing nodes 푠 < 푡 with 푘 probability 푝 = 푡 • The ensemble is not 풢푁,푝 • The degree distribution is exponential • The ensembles are very different because the connection probability depends on graph size Example: Soft configuration model • Strong equivalence in the dense case (graphons) • No strong equivalence in the sparse case since 휅 휅 the connection probability 푝 ~ 푠 푡 depends 푠푡 푁 on graph size Example: Preferential attachment • Weak equivalence in the soft formulation (M.Boguna et al., PRE 68:36112, 2003) • No strong equivalence since the probability that new node 푡 connects to existing node 푠 푘푠 푝푠푡~ depends on graph size 푠 푘푠 • Needless to say, 푃 퐺 is unknown Example: Random geometric graphs • Strong equivalence in sparse graphs! • Connection probability 푝푠푡 = Θ 푅 − 푑푠푡 does not depend on graph size • The distribution of new node coordinates 휌푡 휅푡 is such that 휌푆 휅 = 휌퐷 휅 because both static and dynamic constructions implement the same (Poisson) point process Poisson point processes behind random geometric graphs • Let 푟푡 be the radial coordinate of node 푡 • Then 푉푡 = 퐵 0, 푟푡 is its volume coordinate • Points 푉푡, 푡 = 1,2, … is a PPP on ℝ+ Two pairs of identical ensembles • Fixed-volume pair – Static ensemble 풢푆,푉 • Given 푉, sample 푁 from the Poisson distribution with mean 훿푉, where 훿 is the point density (PPP rate) • Sample 푁 points on [0, 푉] – Dynamic ensemble 풢퐷,푉 • Given Δ푉, sample 푁 from the Poisson distribution with mean 훿Δ푉, where 훿 is the point density (PPP rate) • Sample 푁 points on [푉, 푉 + Δ푉] • Fixed-size pair – Static ensemble 풢푆,푁 • Given 푁, sample 푉 from the Gamma distribution with rate 훿 and shape 푁 • Sample 푁 − 1 points on [0, 푉] – Dynamic ensemble 풢퐷,푁 • Given ΔN, sample Δ푉 from the Gamma distribution with rate 훿 and shape Δ푁 • Sample Δ푁 − 1 points on [푉, 푉 + Δ푉] Two pairs of identical ensembles Final example: de Sitter causal sets • Random geometric graphs in de Sitter spacetime – Angular coordinates 휃 are as in all spherically symmetric random geometric graphs, i.e., sampled from the uniform distribution on a unit sphere 핊푑 – “Radial” (conformal time) coordinates 휂 are sampled from 휌 휂 ∼ sec푑+1 휂 – Nodes are connected if Δ휂 > Δ휃 • The resulting graph ensembles are – Strongly equivalent – Sparse – Have power-law distributions of node degrees with exponent 훾 = 2 – Have strong clustering Conclusions • There does exist a very simple connection between equilibrium and non-equilibrium approaches to network analysis • Some real networks appear to be “close” to graph ensembles that admit dual, equilibrium and growing, description • This justifies applications of powerful equilibrium tools to the analysis of real networks • Does nature optimize? • If the least action principle applies, then what is the action? • D. Krioukov and M. Ostilli Duality between equilibrium and growing networks, PRE, v.88, p.022808, 2013 .

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    25 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us