Network Growth with Feedback

Existing models of network growth typically have one or two parameters or strategies which are fixed for all times. We introduce a general framework where feedback on the current state of a network is used to dynamically alter the values of such parameters. A specific model is analyzed where limited resources are shared amongst arriving nodes, all vying to connect close to the root. We show that tunable feedback leads to growth of larger, more efficient networks. Exact results show that linear scaling of resources with system size yields crossover to a trivial condensed state, which can be considerably delayed with sublinear scaling.

💡 Research Summary

The paper introduces a novel framework for network growth that incorporates feedback from the evolving network state to dynamically adjust growth parameters. Traditional models of network evolution—preferential attachment, copying, or optimization—typically fix their governing parameters at the outset, which can be suboptimal as the network scales from a few seed nodes to thousands or millions. The authors propose a feedback loop that measures a global fitness function of the network at regular intervals and uses the result to increase or decrease the arrival rate of new nodes, thereby indirectly controlling how much resource is allocated per new node for optimizing its attachment.

The concrete model studied is a hierarchical tree. At each discrete time step t, λ(t) ≥ 1 new nodes arrive. A fixed fraction c (0 < c ≤ 1) of the total system resources is devoted to growth; this amount is shared equally among the λ(t) newcomers. Consequently each newcomer can examine only k(t) = ⌈c·N(t)^α / λ(t)⌉ candidate parent nodes, where N(t) is the current number of nodes and α determines how resources scale with system size (α = 1 corresponds to linear scaling). The newcomer selects the candidate closest to the root (the “optimal” parent), breaking ties at random.

Network fitness is quantified by the characteristic time τ_c of a weighted random walk that models message propagation. The walk’s transition matrix P has entries P_ij = 1/d_i for i ≠ j (if i and j are adjacent) and P_ii = 1 – 1/d_i, where d_i is the degree of node i. The second‑largest eigenvalue r_2 of P determines τ_c = –1 / ln|r_2|; smaller τ_c means faster information flow and higher fitness. Empirically the authors find ⟨τ_c(N)⟩ ∼ ln N^β, with β depending on c and λ. The fitness function is defined as F(G) = –τ_c / ln N^β (the negative sign reflects that larger τ_c is worse).

Feedback operates as follows: every δ time steps the current fitness is compared with the fitness δ steps earlier. If fitness has improved, λ(t+1) = λ(t) + 1; if it has deteriorated, λ(t+1) = max