Paperlane

no. 17 · Wednesday 18 November 2026

Spectral Classification of Random Skew-Kernel Matrices

Guangqian Zhao

math.PR · 26 pp

Abstract

We study the unscaled empirical singular spectra of deterministic skew Hilbert–Schmidt kernels evaluated on independent isotropic rows satisfying a uniform quadratic-form variance bound. For arbitrary varying kernels at spatial dimension $d$, tightness and collapse to zero are characterized by boundedness and vanishing, respectively, of the rescaled quadratic energy remaining after deletion of $\lfloor\varepsilon d\rfloor$ leading skew pairs, for every $\varepsilon>0$. A regularized positive-covariance detector proves the necessary direction without a spectral envelope and retains diffuse energy across arbitrarily many small modes. On this exact tightness class, a weighted deterministic source characterizes all probability limits of the empirical spectra through the classical canonical parametrization of symmetric freely infinitely divisible laws. For a fixed kernel, convergence of one bounded resolvent response to a positive deterministic limit, convergence of one normalized logarithmic determinant, and full empirical spectral convergence are each equivalent to a critical Weyl law; the limit is necessarily a possibly degenerate half-Cauchy law. Finally, we give a necessary and sufficient criterion for polygonal Brownian Lévy-area spectra to recover the continuous-time bulk under arbitrary deterministic time partitions, even when the normalized mean-square matrix error diverges.

comments
26 pages
fields
Probability
pdf ↗ page arXiv:2611.23095 ↗