Random degree-degree correlated networks

Correlations may affect propagation processes on complex networks. To analyze their effect, it is useful to build ensembles of networks constrained to have a given value of a structural measure, such as the degree-degree correlation $r$, being random in other aspects and preserving the degree distribution. This can be done through Monte Carlo optimization procedures. Meanwhile, when tuning $r$, other network properties may concomitantly change. Then, in this work we analyze, for the $r$-ensembles, the impact of $r$ on properties such as transitivity, branching and characteristic lengths, that are relevant when investigating spreading phenomena on these networks. The present analysis is performed for networks with degree distributions of two main types: either localized around a typical degree (with exponentially bounded asymptotic decay) or broadly distributed (with power-law decay). Correlation bounds and size effects are also investigated.

💡 Research Summary

The paper addresses how degree‑degree correlations, quantified by the assortativity coefficient r, influence structural properties that are crucial for spreading processes on complex networks. The authors introduce the concept of an “r‑ensemble”: a family of networks that share a prescribed degree distribution P(k) while being random in all other respects, but constrained to a target value of r. To generate such ensembles they employ a Monte Carlo rewiring algorithm that preserves the exact degree sequence. Starting from a configuration‑model graph, two edges are selected at random and swapped; the swap is accepted according to a Metropolis criterion that drives the measured r toward the desired target. Because the degree sequence is invariant, basic statistics such as the mean degree, degree variance, and the existence of a giant component remain unchanged, allowing a clean isolation of the effect of r.

Two representative degree distributions are examined. The first is an exponential‑type distribution (e.g., Poisson‑like) with a well‑defined typical degree and a rapidly decaying tail. The second is a power‑law (scale‑free) distribution, characterized by a heavy tail and a small number of hubs. For each case the authors systematically vary r from strongly disassortative (negative) to strongly assortative (positive) values and measure four key structural indicators: (i) the global clustering coefficient C (transitivity), (ii) the average excess degree ⟨k_ex⟩ (a proxy for branching in the early stage of a contagion), (iii) the average shortest‑path length ⟨ℓ⟩, and (iv) the network diameter D.

The results reveal consistent, yet distribution‑dependent, trends. In both families, increasing r (more assortative mixing) raises the clustering coefficient. For exponential networks C roughly doubles when r is pushed from 0 to +0.4, whereas for scale‑free networks C can increase threefold when r reaches +0.6. The intuition is that high‑degree nodes preferentially connect to each other, forming dense “core” subgraphs that generate many triangles. Conversely, disassortative mixing suppresses triangles because hubs are linked mainly to low‑degree nodes.

Branching behavior, captured by ⟨k_ex⟩, shows an opposite monotonicity: assortative networks exhibit a reduced excess degree, reflecting the fact that once a contagion reaches the high‑degree core it quickly exhausts new high‑degree contacts. Disassortative networks display larger ⟨k_ex⟩ because hubs are scattered throughout the graph, continually providing fresh high‑degree contacts to the spreading front.

Path‑length metrics respond non‑linearly to r. Positive r shortens both ⟨ℓ⟩ and D, sometimes by as much as 25 % for exponential graphs, due to the emergence of a compact core that acts as a shortcut hub. However, beyond a certain assortativity threshold the reduction saturates; overly dense cores can even fragment the network if connectivity constraints are not enforced. Negative r lengthens paths modestly (≈15 % increase for scale‑free graphs) but the presence of hubs as bridges prevents catastrophic growth of distances.

A crucial part of the study concerns the feasible range of r for a given degree sequence. The authors confirm analytically derived bounds (Newman’s assortativity limits) and show that exponential networks can only achieve |r|≈0.3–0.4, whereas scale‑free networks can reach |r|≈0.6 before the rewiring process either fails to converge or threatens connectivity. Finite‑size effects are also quantified: smaller networks (N≈10³) can attain target r values more easily, but as N grows to 10⁵ the realized r converges toward the theoretical bound, and the Monte Carlo acceptance rate diminishes. This scaling behavior underscores that in the thermodynamic limit the degree sequence alone dictates the maximal attainable assortativity.

The discussion links these structural findings to dynamical processes. In epidemic models (SIR, SIS) assortative networks tend to confine outbreaks within the high‑degree core, reducing overall prevalence, while disassortative networks facilitate rapid, widespread infection because hubs act as long‑range conduits. Similar implications hold for information diffusion, opinion formation, and robustness of infrastructural networks. By providing a systematic method to generate r‑constrained ensembles, the paper offers a valuable experimental platform for disentangling the causal impact of degree correlations from other network features.

In summary, the work demonstrates that degree‑degree correlation is a powerful lever that simultaneously reshapes clustering, branching potential, and characteristic path lengths. These structural changes, in turn, have predictable consequences for any process that propagates along edges. The r‑ensemble framework thus equips researchers with a controlled setting to probe, validate, and ultimately design networks with desired spreading characteristics.