arXiv · 2611.15927
Eigenvalue Crossings and Discontinuous BBP Transitions in Spiked Euclidean Random Matrices
Abstract
We study the recovery of a rank-one signal embedded in a symmetric Euclidean random matrix, whose dense and geometrically correlated entries generally lack rotational invariance. We show that signal recovery through the leading eigenvector is governed by a competition between the principal directions of the noise and the signal. In particular, the signal may already be encoded in a subleading eigendirection, while remaining undetected by the leading principal component. As the signal strength increases, a crossing between competing eigenvalues can change their ordering, causing the leading eigenvector to switch abruptly from a noise-dominated to a signal-bearing direction. This mechanism gives rise to a discontinuous BBP-type transition, in which the overlap with the signal jumps from zero to a finite value. Using the replica method, we reduce the problem to a Fredholm eigenvalue equation and derive general criteria for the existence of a positive recovery threshold, expressed in terms of the unperturbed spectral edge, the spike mean and the spatial averages of the principal eigenfunctions. We obtain analytically the limiting largest eigenvalue, eigenvector-component distribution and signal overlap, and illustrate the results with explicit solutions for quadratic distance and rank-two kernels, finding excellent agreement with numerical diagonalisation.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Pasquale Casaburi, Pierpaolo Vivo. 2026-11-11. Eigenvalue Crossings and Discontinuous BBP Transitions in Spiked Euclidean Random Matrices. https://arxiv.science/abs/2611.15927
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.