Ricci Curvature of Internet Topology
- 1. Ricci Curvature of
Internet Topology
Chien-Chun Ni
Stony Brook University
Join work with: Yu-Yao Lin1, Jie Gao1, David Gu1, Emil Saucan2
1: Stony Brook University, 2: Technion, Israel Institute of Technology.
- 2. What is Internet Topology?
• Topology Graph: A node & edge relationship
INFOCOM 2015 2
P2P Network Router Network AS Network
- 3. Why Internet Topology?
• Structural property:
• Robustness
• Vulnerability
• Connectivity
• Information flow:
• Congestion control
• Virus diffusion
• Network Evolution
INFOCOM 2015 3
- 4. Analysis of Internet Topology?
• Graph Structure:
• Degree distribution
• Graph diameter
• Mean shortest path length
• Graph Geometry:
• Curvature
INFOCOM 2015 4
- 5. Curvature?
• A geometric property: Flatness of an object
INFOCOM 2015 5
=π
2D Plane
Zero Curvature
3D Sphere
Positive Curvature
3D Saddle
Negative Curvature
- 6. Internet Curvature, Prior Discovery
• Internet has negative curvature[1-2]
• Def. by Gromov’s δ-hyperbolicity (thin triangle property)
[Def.] For any triple a, b, c, the min distance from one
shortest path to the other two is no greater than δ
INFOCOM 2015 6
[1]: Narayan, O., & Saniee, I. (2011). Large-scale curvature of networks. Physical Review E.
[2]: Shavitt, Y., & Tankel, T. (2004). On the Curvature of the Internet and its usage for Overlay Construction and Distance
Estimation. Infocom 2004.
- 7. Internet is Negatively Curved
A variety of data sets, both AS-level and router level
topologies, show δ-hyperbolicity for small constant δ.
• Properties:
• Tree-Like
• There is a ‘core’ in which all shortest path visit
• Congestion inside the core is high
• Preprocessing can speed up shortest path queries
INFOCOM 2015 7
- 8. Global v.s. Local Curvature
• Gromov δ-hyperbolicity: global measure
• Define one parameter on whole network graph
• No information provided on each vertex and edge
• Local Curvature?
• Which edges are negatively curved?
• Local curvature v.s. network congestion?
• Local curvature v.s. node centrality?
INFOCOM 2015 8
- 10. Sectional Curvature
Consider a tangent vector v = xy. Take another tangent vector
wx and transport it along v to be a tangent vector wy at y.
If |x’y’| < |xy| the sectional curvature is positive.
INFOCOM 2015 10
• Ricci Curvature: averaging over all directions w
- 11. Discrete Ricci Curvature
[Take the analog]: For an edge xy, consider the
distances from x’s neighbors to y’s neighbors and
compare it with the length of xy.
INFOCOM 2015 11
- 12. Discrete Ricci Curvature
• Issue: how to match x’s neighbors to y’s neighbors?
• Assign uniform distribution μ1 , μ2 on x’ and y’s neighbors.
• Use optimal transportation distance (earth-mover
distance) from μ1 to μ2: the matching that minimize the
total transport distance.
INFOCOM 2015 12
- 13. Discrete Ricci Curvature[1,2]
INFOCOM 2015 13
[1]: Ollivier, Y. (2007, January 31). Ricci curvature of Markov chains on metric spaces. arXiv.org.
[2]: Lin, Y., Lu, L., & Yau, S.-T. (2011). Ricci curvature of graphs. Tohoku Mathematical Journal, 63(4), 605–627.
- 24. Data Sets: Model Network
• Erdos-Renyi random graph (G(n,p) graph)
• Watt-Strogatz graph
• Random regular
• Preferential attachment
• Configuration model
• H(3,7) grid (Hyperbolic grid)
- 27. Conclusion and Discussion
• Another measure to understand the Internet topology
• Limitation of data sets
• Network embedding: Euclidean, hyperbolic, hybrid?
• Network evolution: Why?
INFOCOM 2015 27
Editor's Notes
- Social: node-> ppl, edge-> relation
In different resolution: the meaning of edge and node changes
- How easy to cut a tree into half, min max flow
Retweet or like, predictmulticast: windows update,
Infrastructure upgrade priority
How internet evolve
- Result like power law, short diameter
- Draw a tree
Imaging high way
Euclidean =>inf
Hyperbolic plain => constant
- Delta => hopcoun
reft value grow slower
- How do we know it’s negative?
Two graph, are they identical
- Local Curvauture
Hit equater
- Negatively curved edges are like “backbones”, maintaining the connectivity of clusters, in which edges are mostly positively curved.
- Core!
Robustness
Backbone effect
Negatively curved edges are well connected. Adding edges with
increasing/decreasing curvature: few/many connected components.
- Removing edges with increasing curvature: size of largest connected
component drops quickly.
- Positive edge => cluster happen
- Core of network H37 => hyperbolic core
backbone effect
Small world property is not relevant;
graph hyperbolicity and power law degree distribution appear to be more relevant.
- No edge weight
Different network => different curvature distribution
Negatively curved edges:
Tend to act as backbones of graph
Related to robustness of graph
Positively curved edges:
Act as cluster or boundary leaves
- X