Scale free networks by preferential depletion

We show that not only preferential attachment but also preferential depletion leads to scale-free networks. The resulting degree distribution exponents is typically less than two (5/3) as opposed to the case of the growth models studied before where the exponents are larger. Our approach applies in particular to biological networks where in fact we find interesting agreement with experimental measurements. We investigate the most important properties characterizing these networks, as the cluster size distribution, the average shortest path and the clustering coefficient.

💡 Research Summary

The paper introduces a novel mechanism for generating scale‑free networks that does not rely on the traditional growth and preferential attachment paradigm. Instead, the authors propose “preferential depletion,” a process in which edges are removed rather than added, while the number of nodes remains constant. The algorithm starts from a fully connected graph of N nodes. At each iteration a node i is chosen at random, and one of its incident edges eij is selected for removal with probability

 pij = pj / Σl∈Ni pl,  pj = k_j^(-α) for k_j > k_min, otherwise 0,

where k_j is the degree of neighbor j, α > 0 controls the strength of the depletion bias, and k_min sets a lower bound on node degree. The removal continues until the total number of edges M reaches N·k_min.

The authors explore the impact of the two parameters α and k_min through extensive numerical simulations. They find that when α ≈ 2 the resulting degree distribution follows a power law p(k) ∼ k^(-γ) with γ ≈ 5/3, a value significantly smaller than the γ > 2 typical of growth‑based models. This exponent matches empirical measurements on several biological networks, such as the Escherichia coli metabolic network (γ ≈ 1.7) and various gene‑interaction networks (γ ≈ 1.5–1.6). For α < 2 the power‑law regime is limited by an exponential cutoff at intermediate k; for α > 2 the power‑law region shrinks, the effective exponent increases, and a small bump appears at moderate degrees. The minimal degree parameter k_min mainly influences the connectivity of the final graph: with k_min = 1 the system fragments into many small clusters plus a large tree‑like component, whereas k_min ≥ 2 yields a single connected component.

Cluster size statistics are also examined. The frequency f(S) of clusters containing S nodes follows a power law f(S) ∼ S^(-β) with β ≈ 3.5 for node‑based clusters and β ≈ 4 for edge‑based clusters, indicating a broad distribution of component sizes when fragmentation occurs.

Structural properties such as clustering coefficient C and average shortest path ℓ are measured. The clustering coefficient decays with system size as C ∼ N^(-c0), where c0 ≈ 0.7 for k_min = 1 and c0 ≈ 0.24 for k_min ≥ 2. This decay is slower than that of random graphs (c0 = 1) and Barabási‑Albert networks (c0 ≈ 0.75), implying relatively high local cohesion. The average shortest path exhibits ultra‑small‑world behavior, scaling as ℓ ∝ ln ln N for all k_min > 1, which is slower than the logarithmic scaling of classic small‑world networks and characteristic of ultra‑small networks.

The sensitivity of the final topology to the initial configuration is investigated. Starting from (i) a fully connected graph, (ii) a random graph with a high average degree, or (iii) a lattice with uniformly distributed degrees, the authors find that only (i) and (ii) lead to scale‑free degree distributions after depletion. A uniformly random degree lattice fails to generate a power‑law tail, suggesting that a dense and homogeneous initial network is required for the depletion process to produce the desired heterogeneity.

In the discussion, the authors connect the preferential depletion mechanism to biological processes such as synaptic pruning during neural development, where weaker connections are eliminated to refine circuitry. They argue that, unlike growth‑based models, depletion can naturally produce the small γ values observed in many biological networks and can also generate disconnected structures, a feature often seen in real biological systems.

The paper concludes that preferential depletion offers a plausible alternative to growth for the emergence of scale‑free topologies, especially in biological contexts. It reproduces key empirical exponents, yields ultra‑small‑world distances, and maintains relatively high clustering, thereby providing a comprehensive framework for understanding the formation of complex networks where edge removal, rather than addition, dominates the evolutionary dynamics.