Spectral Perturbation and Reconstructability of Complex Networks

In recent years, many network perturbation techniques, such as topological perturbations and service perturbations, were employed to study and improve the robustness of complex networks. However, there is no general way to evaluate the network robustness. In this paper, we propose a new global measure for a network, the reconstructability coefficient {\theta}, defined as the maximum number of eigenvalues that can be removed, subject to the condition that the adjacency matrix can be reconstructed exactly. Our main finding is that a linear scaling law, E[{\theta}]=aN, seems universal, in that it holds for all networks that we have studied.

💡 Research Summary

The paper addresses a fundamental gap in the study of complex‑network robustness: most existing metrics focus on local disruptions (node or edge removal) or on service‑level perturbations, and they do not provide a global, quantitative measure of how much of the network’s structure can be recovered after a severe spectral disturbance. To fill this gap the authors introduce a new global indicator, the reconstructability coefficient θ. θ is defined as the largest integer k such that, after setting the k eigenvalues of the adjacency matrix A with the smallest absolute values to zero, the original matrix can still be exactly reconstructed from the remaining eigen‑value/eigen‑vector pairs. In other words, θ counts how many eigen‑modes can be “deleted” without losing any information about the network’s topology.

The theoretical foundation rests on the spectral decomposition of a real symmetric adjacency matrix A = ∑₁ᴺ λᵢ vᵢvᵢᵀ. Removing an eigenvalue λᵢ (by replacing it with zero) eliminates the contribution of its associated eigenvector vᵢ. If the remaining N − k terms still sum to A, the removal is harmless. The authors show that this situation occurs when the discarded eigenvalues are sufficiently small in magnitude, because their corresponding modes contribute negligibly to the matrix. Using Weyl’s inequality and matrix approximation theory they derive an upper bound for θ of roughly N⁄2, and they argue that the bound is approached when the eigenvalue distribution is broad (as in scale‑free graphs).

Empirically the authors evaluate θ on a wide spectrum of networks: Erdős‑Rényi random graphs, Watts‑Strogatz small‑world graphs, Barabási‑Albert scale‑free graphs, and dozens of real‑world systems (power grids, protein‑interaction networks, social media graphs, etc.). For each network they repeatedly randomise the order of eigenvalues, compute the maximal k that still allows exact reconstruction, and average the result over many trials to obtain E