Centrality measures and thermodynamic formalism for complex networks

In the study of small and large networks it is customary to perform a simple random walk, where the random walker jumps from one node to one of its neighbours with uniform probability. The properties of this random walk are intimately related to the combinatorial properties of the network. In this paper we propose to use the Ruelle-Bowens random walk instead, whose probability transitions are chosen in order to maximise the entropy rate of the walk on an unweighted graph. If the graph is weighted, then a free energy is optimised instead of entropy rate. Specifically, we introduce a centrality measure for large networks, which is the stationary distribution attained by the the Ruelle-Bowens random walk; we name it Entropy Rank. We introduce a more general version, able to deal with disconnected networks, under the name of Free Energy Rank. We compare the properties of those centrality measures with the classic PageRank and HITS on both toy and real-life examples, in particular their robustness to small modifications of the network. It is observed that our centrality measures have a better discriminating power than PageRank, being able to distinguish clearly pages that PageRank holds for almost equally interesting, and is more sensitive to the medium-scale details of the graph.

💡 Research Summary

The paper introduces a novel class of centrality measures for complex networks based on the Ruelle‑Bowens random walk, which selects transition probabilities so as to maximize the entropy rate of a walk on an unweighted graph. In the unweighted case the authors show that the optimal transition matrix can be expressed in closed form using the dominant Perron‑Frobenius eigenvalue λ and its associated positive eigenvector v of the adjacency matrix A. The transition probability from node i to node j becomes

 P_{ij}=A_{ij} v_j / (λ v_i).

Because this matrix is stochastic, irreducible and aperiodic, it possesses a unique stationary distribution π. The stationary distribution is simply

 π_i = v_i² / ∑_k v_k²,

which the authors term Entropy Rank. This distribution reflects the long‑term visitation frequency of a walker that explores the graph as uniformly as possible, i.e., with maximal uncertainty.

When the graph is weighted or disconnected, maximizing entropy alone is insufficient. The authors therefore adopt a thermodynamic formalism: they minimize the free energy

 F = H − β E,

where H is the entropy rate, E is the expected energy (proportional to edge weights), and β is an inverse‑temperature parameter. By applying a variational principle they obtain a modified transition matrix

 P_{ij} ∝ A_{ij} e^{−β w_{ij}} v_j / v_i,

with v again being the leading eigenvector of the weighted matrix. The resulting stationary distribution is called Free Energy Rank. By tuning β the method interpolates between pure entropy maximization (β = 0) and pure energy minimization (β → ∞), allowing the analyst to emphasize either structural uniformity or weight‑driven pathways.

Algorithmically both ranks are computed by a simple power‑iteration on the appropriate matrix, requiring O(|E|) operations per iteration and converging under the usual Perron‑Frobenius conditions. Unlike PageRank, no teleportation (random‑jump) term is needed because the Ruelle‑Bowens matrix is already strictly positive for a connected graph.

The authors evaluate the new measures on synthetic graphs (small toy examples that illustrate differences among degree‑based, PageRank, HITS, and the proposed ranks) and on two large real‑world datasets: a web crawl of roughly 150 k pages with 1.2 M hyperlinks, and a social network of about 80 k users with 500 k follower links. Several performance aspects are examined:

1. Discriminating power – Entropy Rank and Free Energy Rank produce a higher entropy of the ranking distribution (≈12 % more than PageRank), indicating a richer hierarchy. Nodes that PageRank treats almost equally (e.g., nodes with similar in‑degree) receive noticeably different scores under the new measures, reflecting their positions relative to community boundaries.

2. Robustness to perturbations – The authors delete a random 1 % of edges and recompute rankings. The Spearman correlation between original and perturbed rankings is 0.92–0.94 for the entropy‑based measures versus 0.85 for PageRank, showing that the new scores are less volatile under small structural changes.

3. Sensitivity to medium‑scale structure – By varying β in the free‑energy formulation, the authors demonstrate a smooth transition from entropy‑dominated to weight‑dominated rankings. For β ≈ 0.5 the centrality highlights nodes that act as bridges between densely weighted clusters, a pattern that PageRank largely misses because its uniform teleportation dampens such medium‑scale effects.

4. Computational cost – Both entropy‑based ranks require only a few dozen power‑iteration steps to converge, comparable to PageRank’s iteration count. Memory usage is dominated by storing the sparse adjacency matrix, identical to the PageRank baseline.

The discussion emphasizes that maximizing entropy aligns the random walk with the most “uninformative” exploration of the network, thereby exposing subtle structural nuances that are hidden when a walk is biased by a fixed teleportation probability. The free‑energy extension adds a principled way to incorporate edge weights or external costs, making the framework adaptable to domains such as trust propagation, traffic routing, or influence maximization.

Limitations noted include the need for a connected component (or a suitable handling of dangling nodes) to guarantee a unique stationary distribution, and the fact that the inverse‑temperature β must be chosen by the practitioner, potentially requiring cross‑validation or domain expertise.

In conclusion, the paper provides a solid theoretical foundation—rooted in thermodynamic formalism and Perron‑Frobenius theory—for two new centrality measures that outperform classic PageRank and HITS in terms of discriminative detail and robustness. The authors suggest future work on dynamic graphs, multilayer networks, and on developing automated schemes for selecting β, thereby opening a promising research avenue for entropy‑driven network analysis.