[Paper Review] Phase transitions in edge-weighted exponential random graphs

Mei Yin|arXiv (Cornell University)|Jul 14, 2016

Graph theory and applications26 references3 citations

TL;DR

This paper extends exponential random graph models to edge-weighted networks by introducing a generic edge weight distribution, enabling rigorous analysis of phase transitions. It establishes a first-order phase transition curve and a second-order critical point for uniformly distributed edge weights, resolving a key limitation in modeling real-world weighted networks.

ABSTRACT

The exponential family of random graphs represents an important and challenging class of network models. Despite their flexibility, conventionally used exponential random graphs have one shortcoming. They cannot directly model weighted networks as the underlying probability space consists of simple graphs only. Since many substantively important networks are weighted, this limitation is especially problematic. We extend the existing exponential framework by proposing a generic common distribution for the edge weights and rigorously analyze the associated phase transitions and critical phenomena. We then apply these general results to get concrete answers in exponential random graph models where the edge weights are uniformly distributed.

Motivation & Objective

• To address the limitation of conventional exponential random graphs, which model only simple (unweighted) graphs, by extending the framework to edge-weighted networks.
• To develop a general theory for edge-weighted exponential random graphs using a generic edge weight distribution.
• To rigorously analyze phase transitions and critical phenomena in the resulting models, particularly for uniformly distributed edge weights.
• To provide a foundation for understanding asymptotic structures and critical behavior in weighted network models.

Proposed method

• Uses graph limit theory and graphon representations to model large edge-weighted graphs as symmetric measurable functions on [0,1]².
• Applies variational principles to derive the limiting normalization constant and concentration of measure for the model.
• Introduces a large deviation principle for the edge-weighted model to analyze rare events and phase transitions.
• Employs homomorphism densities to define subgraph statistics (e.g., edge and 2-star densities) in the weighted setting.
• Analyzes the free energy functional l(u) to identify global maximizers and phase boundaries.
• Specializes to uniformly distributed edge weights and derives the phase transition curve β₂ = r(β₁) via analytical optimization.

Experimental results

Research questions

• RQ1 How do phase transitions emerge in edge-weighted exponential random graph models with a generic edge weight distribution?
• RQ2 What is the structure of the phase transition curve when edge weights are uniformly distributed?
• RQ3 Does the model exhibit a second-order critical point, and if so, under what conditions?
• RQ4 How does the asymptotic behavior of the normalization constant reflect critical phenomena in the model?
• RQ5 What happens to phase structure when the edge weight distribution has infinite support, such as the Gaussian case?

Key findings

• For uniformly distributed edge weights, a first-order phase transition curve β₂ = r(β₁) exists, with a second-order critical point at which the system undergoes a qualitative change in structure.
• The phase transition curve lies above the line β₂ = -β₁ for p ≥ 3, and coincides with it (for β₁ ≤ -3) when p = 2, indicating a symmetry-breaking transition.
• Along the phase transition curve, the free energy functional l(u) has two global maximizers u₁* and u₂*, symmetric about u = 1/2, confirming coexistence of distinct graph structures.
• For p = 2, the maximizers are symmetric and located at u₁* < 1/2 < u₂*, confirming the presence of a first-order transition.
• In the directed Gaussian model, the limiting normalization constant ψ∞^β = β₁² / [2(1 - 2β₂)] is smooth for β₂ < 1/2, indicating no phase transition occurs.
• The absence of a large deviation principle for non-finite-support edge weights limits the analysis, but the Gaussian case shows no singularities, implying no phase transition in that setting.