Fractal Weyl law for Linux Kernel Architecture

We study the properties of spectrum and eigenstates of the Google matrix of a directed network formed by the procedure calls in the Linux Kernel. Our results obtained for various versions of the Linux Kernel show that the spectrum is characterized by the fractal Weyl law established recently for systems of quantum chaotic scattering and the Perron-Frobenius operators of dynamical maps. The fractal Weyl exponent is found to be $\nu \approx 0.63$ that corresponds to the fractal dimension of the network $d \approx 1.2$. The eigenmodes of the Google matrix of Linux Kernel are localized on certain principal nodes. We argue that the fractal Weyl law should be generic for directed networks with the fractal dimension $d<2$.

💡 Research Summary

The paper investigates the spectral properties of the Google matrix constructed from the directed network of function‑call relationships within the Linux kernel source code. By extracting call graphs from several kernel releases (spanning from early 2.6 versions to recent 5.x releases), the authors build a directed graph where each node represents a function and each edge denotes a call from one function to another. The resulting networks contain on the order of 10⁴–10⁵ nodes and a few times that number of edges, exhibiting a sparse but highly heterogeneous connectivity pattern.

Using the standard Google matrix formulation G = αS + (1 − α)evᵀ with the damping factor α = 0.85, they compute the full spectrum of G (all eigenvalues λ) and the associated right eigenvectors ψ. The eigenvalues lie inside the unit disk of the complex plane, with a pronounced cluster near the unit circle (|λ|≈1). By defining γ = −ln|λ|, the authors analyze the density ρ(γ) and find a power‑law scaling ρ(γ) ∝ γ^{ν‑1}. The fitted exponent ν≈0.63 is remarkably stable across kernel versions and across variations of α (0.5–0.95).

The exponent ν is interpreted through the fractal Weyl law, originally derived for open quantum chaotic systems and for Perron‑Frobenius operators of chaotic maps. According to that law, the number of eigenvalues with decay rate less than γ scales as N(γ) ∼ γ^{‑ν}, and ν is related to the fractal (Hausdorff) dimension d of the underlying phase space by ν = d/2. To test this relationship, the authors compute the fractal dimension of the call‑graph using a box‑counting method, obtaining d≈1.2. The relation ν≈d/2 holds within numerical uncertainty, confirming that the Linux‑kernel call network obeys the fractal Weyl law.

Beyond the global scaling, the paper examines the structure of individual eigenvectors. Eigenvectors associated with eigenvalues close to 1 are strongly localized on a small set of “principal” functions—those involved in memory allocation, scheduler initialization, and core file‑system setup. These nodes correspond to the highest PageRank scores and act as bottlenecks or hubs for information flow in the call graph. In contrast, eigenvectors linked to eigenvalues deep inside the unit disk are delocalized, reflecting global diffusion modes of the network.

The authors also compare different kernel releases. While the absolute number of nodes and edges grows with the kernel’s evolution, the fractal dimension d and the Weyl exponent ν remain essentially unchanged. This suggests that the fractal geometry of the call network is a robust, scale‑invariant property of the kernel’s architecture, persisting despite code additions and refactorings. Moreover, the stability of ν under variations of the damping factor indicates that the observed scaling is intrinsic to the network topology rather than an artifact of the Google‑matrix construction.

Finally, the paper argues that any directed network with a fractal dimension d < 2 should exhibit a similar fractal Weyl scaling of its Google‑matrix spectrum. Examples include social‑media follower graphs, the World‑Wide‑Web hyperlink network, and metabolic or gene‑regulation networks. The authors propose that the fractal Weyl law provides a unifying framework for understanding spectral gaps, relaxation times, and the localization of eigenmodes in complex directed systems. This insight could be leveraged for practical tasks such as identifying critical components in software, assessing network robustness, or designing efficient navigation algorithms on large‑scale directed graphs.