Google matrix of the citation network of Physical Review

We study the statistical properties of spectrum and eigenstates of the Google matrix of the citation network of Physical Review for the period 1893 - 2009. The main fraction of complex eigenvalues with largest modulus is determined numerically by different methods based on high precision computations with up to $p=16384$ binary digits that allows to resolve hard numerical problems for small eigenvalues. The nearly nilpotent matrix structure allows to obtain a semi-analytical computation of eigenvalues. We find that the spectrum is characterized by the fractal Weyl law with a fractal dimension $d_f \approx 1$. It is found that the majority of eigenvectors are located in a localized phase. The statistical distribution of articles in the PageRank-CheiRank plane is established providing a better understanding of information flows on the network. The concept of ImpactRank is proposed to determine an influence domain of a given article. We also discuss the properties of random matrix models of Perron-Frobenius operators.

💡 Research Summary

This paper investigates the spectral properties and eigenstates of the Google matrix constructed from the citation network of Physical Review (CNPR) covering the period 1893–2009. The network contains 46 3348 articles and 4 691 015 citation links, and its adjacency matrix is almost triangular because citations generally point backward in time. The authors first build the stochastic matrix S by column‑normalizing the adjacency matrix and then form the Google matrix G = αS + (1 − α)/N with the usual damping factor α = 0.85.

A major technical challenge arises from the near‑nilpotent structure of S: the matrix contains large Jordan blocks that make standard eigenvalue algorithms unstable, especially for eigenvalues with modulus |λ| < 0.3–0.4. To overcome this, the authors employ two complementary strategies. (1) They perform high‑precision (arbitrary‑precision) Arnoldi iterations using up to p = 16384 binary digits, which allows them to resolve very small eigenvalues that would otherwise be lost in round‑off noise. (2) They decompose S into S = S₀ + E/N, where S₀ contains the almost‑triangular part (nearly nilpotent) and E accounts for dangling nodes (columns of zeros) that are replaced by uniform transitions. This decomposition enables a semi‑analytical treatment: the eigenvalues of S₀ can be obtained analytically (or with negligible numerical error), and the effect of the uniform term E/N is treated as a perturbation. The combination of high‑precision numerics and semi‑analytical theory yields a reliable spectrum for the full Google matrix.

The resulting eigenvalue distribution follows a fractal Weyl law with a fractal dimension d_f ≈ 1, indicating that the density of eigenvalues scales as N^{d_f/2} and that the underlying dynamics are effectively one‑dimensional, reflecting the temporal ordering of citations. Most eigenvectors are found to be localized, i.e., they have significant weight on a limited set of articles, typically clustered in specific years or sub‑fields. This localization contrasts with the delocalized eigenvectors seen in many random or undirected networks.

Beyond spectral analysis, the paper explores two‑dimensional ranking. PageRank (the right eigenvector at λ = 1) measures the stationary probability of a random surfer on each article, while CheiRank, obtained from the transpose of the stochastic matrix, quantifies the propensity of articles to cite others. Plotting articles in the (PageRank, CheiRank) plane reveals that top PageRank papers rarely cite each other, whereas top CheiRank papers act as major “receivers” of citations. This 2D ranking provides a richer picture of information flow than PageRank alone.

The authors also introduce ImpactRank, a novel metric that quantifies the influence domain of a given article. Starting from a unit vector on a selected article, they propagate probabilities through G for a fixed number of steps and measure the set of articles that receive a non‑negligible probability. ImpactRank captures indirect influence that is not reflected in raw citation counts, highlighting papers that serve as bridges between different research areas.

Finally, the paper compares the empirical spectrum with that of random matrix models designed to mimic Perron–Frobenius operators. Random models produce eigenvalues clustered near the origin and lack the fractal Weyl scaling and localization observed in the real citation network, underscoring the special, highly structured nature of scientific citation graphs.

In summary, the study combines ultra‑high‑precision numerical diagonalization with a semi‑analytical treatment of the near‑nilpotent structure to obtain an accurate spectrum of a large, directed citation network. It demonstrates that the spectrum obeys a fractal Weyl law, that eigenvectors are predominantly localized, and that two‑dimensional ranking (PageRank vs. CheiRank) together with ImpactRank offers powerful tools for assessing scientific influence beyond traditional citation counts. The methodology and insights are likely applicable to other temporal or hierarchical networks, opening avenues for deeper understanding of information propagation in complex directed systems.