Paperlane

no. 12 · Tuesday 10 November 2026

Staircase phase transitions for the extreme eigenvalues of heavy-tailed sample correlation matrices

Yanpeng Li, Zeqin Lin, Yiming Liu, Jiahui Xie and Haozhu Zhao

math.PR · v2

Abstract

We establish phase transitions in the first-order asymptotics of the largest and smallest nonzero eigenvalues of high-dimensional sample correlation matrices formed from data with i.i.d. centered real entries of finite variance, allowing an infinite fourth moment. Because of self-normalization, spectral outliers are driven by collisions of large entries in distinct rows of a common column rather than by individual large entries. The resulting phase transitions depend jointly on the aspect ratio and the entry tail, producing staircase boundaries in the phase diagram. We prove convergence to the corresponding Marčenko–Pastur edges in subcritical regimes and to deterministic outlier levels in supercritical regimes. In critical regimes, Poisson limits for the point processes of spectral outliers and nondegenerate limiting laws for the extreme eigenvalues are established.

comments
This version merges the original manuscript with arXiv:2609.15731v1 to give a unified treatment of both spectral edges
fields
Probability
pdf ↗ page arXiv:2610.06731 ↗