[Paper Review] A geometric preferential attachment model with fitness

Henri van den Esker|ArXiv.org|Jan 10, 2008

Complex Network Analysis Techniques16 references3 citations

TL;DR

This paper introduces a geometric preferential attachment model with fitness that combines spatial positioning on a sphere with preferential attachment based on degree and distance-dependent attractiveness. It proves the resulting network exhibits a power-law degree distribution with an exponent tunable to any value >2, enabling control over network properties like typical distance and robustness.

ABSTRACT

We study a random graph $G_n$, which combines aspects of geometric random graphs and preferential attachment. The resulting random graphs have power-law degree sequences with finite mean and possibly infinite variance. In particular, the power-law exponent can be any value larger than 2. The vertices of $G_n$ are $n$ sequentially generated vertices chosen at random in the unit sphere in $\mathbb R^3$. A newly added vertex has $m$ edges attached to it and the endpoints of these edges are connected to old vertices or to the added vertex itself. The vertices are chosen with probability proportional to their current degree plus some initial attractiveness and multiplied by a function, depending on the geometry.

Motivation & Objective

• To develop a random graph model that integrates geometric constraints with preferential attachment to better reflect real-world networks like ad-hoc wireless networks.
• To analyze how geometric proximity and vertex fitness jointly influence the degree distribution and network structure.
• To establish conditions under which the model yields power-law degree sequences with arbitrary exponents greater than 2.
• To rigorously prove that the power-law exponent can be tuned independently of the network size, enabling control over network properties such as typical distance.

Proposed method

• Vertices are sequentially added to a unit sphere in ℝ³, with each new vertex positioned uniformly at random on the sphere’s surface.
• Each new vertex forms m edges by selecting existing vertices with probability proportional to (degree + δ) × Fₙ(angular distance), where Fₙ is a distance-dependent attractiveness function.
• The model incorporates a constant additive fitness δ and a multiplicative fitness term Fₙ(u) that depends on the angular distance u between vertices.
• The analysis uses a coupling argument between two graphs differing in one vertex’s position to bound the expected number of differing edges, leveraging conditional expectations and indicator functions.
• Key equations include the total attachment weight Tσ,n = Σᵥ (dσ(v) + δ)Fₙ(|xᵥ − u|), which governs edge selection probabilities.
• The proof relies on bounding the difference in edge sets between coupled graphs using the function Aσ,n(x) = Fₙ(|x − xσ+1|), leading to a recursive bound on the expected number of differing edges.

Experimental results

Research questions

• RQ1 Can a geometric preferential attachment model with fitness produce power-law degree distributions with an exponent τ > 2?
• RQ2 How does the inclusion of geometric distance in the attachment mechanism affect the degree distribution and network topology?
• RQ3 What is the range of possible power-law exponents τ in such a model, and can τ be tuned independently of network size?
• RQ4 How does the model’s behavior compare to classical preferential attachment models in terms of degree distribution and typical distance scaling?
• RQ5 What is the impact of initial attractiveness δ and distance-dependent fitness Fₙ on the convergence and stability of the network structure?

Key findings

• The model generates random graphs with power-law degree sequences where the exponent τ can be any value greater than 2, allowing precise control over network structure.
• The power-law exponent τ is determined by the interplay between the initial attractiveness δ and the distance-dependent fitness function Fₙ, enabling tuning of the tail behavior of the degree distribution.
• The model supports both finite and infinite variance in the degree distribution depending on τ, with τ ∈ (2,3) corresponding to infinite variance and τ > 3 to finite variance.
• The typical distance in the network scales as log n when τ > 3, log log n when τ ∈ (2,3), and is bounded when τ ∈ (1,2), consistent with known results in complex network theory.
• The coupling argument shows that the expected number of differing edges between two graphs differing in one vertex’s position is bounded by a term proportional to the difference in their attachment weights, ensuring stability and convergence.
• The analysis confirms that the degree distribution converges to a power law with exponent τ > 2 under mild conditions on Fₙ and δ, even with geometric constraints.