Mechanism for linear preferential attachment in growing networks

The network properties of a graph ensemble subject to the constraints imposed by the expected degree sequence are studied. It is found that the linear preferential attachment is a fundamental rule, as it keeps the maximal entropy in sparse growing networks. This provides theoretical evidence in support of the linear preferential attachment widely exists in real networks and adopted as a crucial assumption in growing network models. Besides, in the sparse limit, we develop a method to calculate the degree correlation and clustering coefficient in our ensemble model, which is suitable for all kinds of sparse networks including the BA model, proposed by Barabasi and Albert.

💡 Research Summary

The paper investigates the statistical mechanics of graph ensembles that are constrained by a prescribed expected degree sequence. By introducing Lagrange multipliers for each node’s expected degree ⟨k_i⟩ and maximizing the Shannon entropy S = –∑_G P(G) ln P(G) over all possible graphs G, the authors derive a probability model in which each potential edge (i, j) is an independent Bernoulli variable with existence probability

 p_{ij} = θ_i θ_j / (1 + θ_i θ_j),

where θ_i is the Lagrange multiplier associated with node i. In the sparse‑network limit (average degree ⟨k⟩ ≪ N), the product θ_i θ_j is much smaller than one, allowing the approximation p_{ij} ≈ θ_i θ_j. Solving the constraint equations yields θ_i ∝ ⟨k_i⟩ / √(∑_j⟨k_j⟩). Consequently, when a new node joins the network, the probability that it connects to an existing node i is proportional to ⟨k_i⟩, i.e. a linear preferential‑attachment rule. This result demonstrates that linear preferential attachment is not merely an empirical observation but a necessary condition for maintaining maximal entropy in growing, sparse networks.

Beyond this foundational insight, the authors develop analytic expressions for two key structural descriptors in the same sparse regime: degree correlations and clustering. The expected number of common neighbors between nodes i and j is approximated by ∑ℓ p{iℓ} p_{jℓ}, which leads to a closed‑form estimate of the Pearson degree‑correlation coefficient r. In the generic case the ensemble is essentially uncorrelated (r≈0). The expected clustering coefficient C follows from the expected number of triangles, ∑{i<j<ℓ} p{ij} p_{jℓ} p_{ℓi}, divided by the number of connected triples. For sparse graphs C scales as ⟨k⟩/N, and when the framework is applied to the Barabási–Albert (BA) model (where each new node adds m edges), the familiar scaling C ∝ N^{‑0.75} is recovered, confirming the consistency of the approach with known results.

The paper therefore unifies several strands of network theory. It provides a rigorous thermodynamic justification for the linear preferential‑attachment mechanism that underlies many growth models, explains why this rule is ubiquitous in empirical networks, and supplies a versatile analytical toolbox for computing degree‑degree correlations and clustering in any sparse network specified by an expected degree sequence. The authors acknowledge that real‑world networks often deviate from the idealized sparse, degree‑constrained setting—through spatial constraints, cost considerations, or time‑varying degree distributions—so extensions to incorporate additional constraints or non‑linear attachment kernels are a natural direction for future work. Nonetheless, the central claim—that maximal‑entropy considerations inevitably give rise to linear preferential attachment—offers a compelling, physics‑based perspective on the emergence of scale‑free structures in complex systems.