Eigenvalue spectral tails and localisation properties of asymmetric networks

Pietro Valigi, Joseph W. Baron, Izaak Neri, Giulio Biroli, Chiara CammarotaJournal of Physics A Mathematical and Theoretical 2025 · DOI

TL;DR

This paper develops a novel analytical framework to understand the "spectral tails" (outlying eigenvalues) and eigenvector localization in large, asymmetric random networks. The authors show how these tails are linked to highly connected "hubs" and provide a general method to predict their behavior, offering insights into the stability and dynamics of complex systems.

What Problem Does This Paper Address?

Many real-world complex systems, from ecosystems to neural networks, can be modeled as large, sparse networks with directed connections. While the behavior of eigenvalues (which dictate system stability) in dense or symmetric networks is well-understood, the "spectral tails" – outlying eigenvalues that can significantly impact system dynamics – in asymmetric sparse networks have remained mathematically challenging. These tails are crucial because they often contain the leading eigenvalue that determines linear stability and the corresponding eigenvector that describes dominant disturbance modes. The authors aim to bridge this gap by providing a quantitative, analytical theory for these spectral tails in asymmetric networks.

Methods

The core of the paper's methodology lies in developing an analytical approach to characterize the spectral tails. They leverage the observation that eigenvalues in these tails correspond to eigenvectors that are exponentially localized around highly connected "hubs" in the network.

Key techniques include:

  • Single-defect approximation: This approach focuses on the local network structure around these highly connected hubs, treating them as isolated "defects."
  • Series expansion: They approximate the localized eigenvectors using a series expansion in the inverse connectivity of the hub, systematically accounting for contributions from neighboring nodes.
  • Cavity equations and 1/w expansion: They employ cavity equations (a technique for analyzing disordered systems) and perform an asymptotic expansion in 1/w (where w represents the eigenvalue) to derive expressions for the eigenvalue density and the inverse participation ratio (IPR).
  • Saddle-point approximation: This is used to handle integrals and sums over network configurations, particularly for large eigenvalues.

The novelty lies in extending these techniques to asymmetric matrices, which requires significant analytical development beyond existing methods for symmetric cases.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis and uses simulated random graphs for numerical verification.

Key Insights and Ideas

  • Hubs drive spectral tails: The paper provides a strong analytical basis for the intuition that highly connected nodes ("hubs") are responsible for the outlying eigenvalues (spectral tails) in asymmetric networks.
  • Localization is key: Eigenvectors associated with these tail eigenvalues are exponentially localized around these hubs, meaning their influence is concentrated. This localization is a direct consequence of the network structure and asymmetry.
  • Asymmetry's dual role: The degree of asymmetry (the balance between reinforcing and antagonistic links) significantly impacts the spectral tails. Increasing asymmetry generally leads to faster decaying tails, potentially stabilizing the system.
  • Network structure matters: The paper demonstrates how different degree distributions (e.g., geometric, power-law, Poisson) lead to distinct behaviors in the spectral tails and eigenvector localization.
  • IPR as a measure of localization: The Inverse Participation Ratio (IPR) is used as a quantitative measure of how localized an eigenvector is. The authors show how IPR is minimized when there's a balance between reinforcing and antagonistic links.

Key Results

The paper provides closed-form analytical expressions for:

  • Eigenvalue density in the tails: They classify various asymptotic behaviors (e.g., power-law vs. exponential decay) depending on network structure and asymmetry.
  • Leading eigenvalue: They derive scalings for the largest eigenvalue, which is crucial for system stability.
  • Inverse Participation Ratio (IPR): They quantify the degree of eigenvector localization, showing it's minimized at a balance of link types.

Numerical comparisons with simulated random graphs (Erdős-Rényi, geometric, power-law) show excellent agreement between their theoretical predictions and empirical estimates for eigenvalue densities.

Limitations and Future Work

  • Single-defect approximation breakdown: The authors acknowledge that their "single-defect" approximation can break down in specific scenarios:
  • When there's a balance between antagonistic and reinforcing links (e.g., € ≈ 1/2), leading to less localized eigenvectors.
  • When the degree distribution decays very slowly, causing significant correlations between highly connected nodes.
  • When edge weight distributions are heavy-tailed, leading to localization on single edges rather than hubs.
  • Future Directions: The authors suggest that their framework could be extended to understand more complex nonlinear models with sparse interactions, noting that nonlinearities can also affect Lifshitz tails. Investigating how these findings apply to non-linear models is a promising avenue. They also highlight the need for further theoretical investigation using population dynamics algorithms for a more complete picture of the eigenvalue spectrum.

Summary updated

Abstract

Abstract In contrast to the neatly bounded spectra of densely populated large random matrices, sparse random matrices often exhibit unbounded eigenvalue tails on the real and imaginary axis, called Lifshitz tails. In the case of asymmetric matrices, concise mathematical results have proved elusive. In this work, we present an analytical approach to characterising these tails. We exploit the fact that eigenvalues in the tail region have corresponding eigenvectors that are exponentially localised on highly-connected hubs of the network associated to the random matrix. We approximate these eigenvectors using a series expansion in the inverse connectivity of the hub, where successive terms in the series take into account further sets of next-nearest neighbours. By considering the ensemble of such hubs, we are able to characterise the eigenvalue density and the extent of localisation in the tails of the spectrum in a general fashion. As such, we classify a number of different asymptotic behaviours in the Lifshitz tails, as well as the leading eigenvalue and the inverse participation ratio. We demonstrate how an interplay between matrix asymmetry, network structure, and the edge-weight distribution leads to the variety of observed behaviours.
BibTeX
@article{valigi2025eigenvalue,
  author = {Pietro Valigi and Joseph W. Baron and Izaak Neri and Giulio Biroli and Chiara Cammarota},
  title = {{Eigenvalue spectral tails and localisation properties of asymmetric networks}},
  journal = {Journal of Physics A Mathematical and Theoretical},
  year = {2025},
  doi = {10.1088/1751-8121/ae16ec},
}
Figure