Answers are grounded in this paper and cite their sources.
We study the outlying eigenvalues of a sample covariance matrix whose population covariance is itself random. The population covariance is Σₚ=Aₚ+OₚBₚOₚ^*+σ²Iₚ, where Aₚ and Bₚ are deterministic positive semidefinite matrices of fixed ranks r and s, and where Oₚ is Haar distributed on the orthogonal group in the real case, and on the unitary group in the complex case. Conditionally on Oₚ, we observe n independent centered Gaussian vectors with covariance Σₚ, and we build the sample covariance matrix Σₚ from these vectors. The fluctuations of the spike eigenvalues of Σₚ come from two different sources: the Haar matrix Oₚ, which displaces the population spikes at the scale $1/p$, and the sampling, which acts at the scale n-1/2. Comparing these two scales amounts to comparing p² with n, so the ratio which governs the fluctuations is p²/n, while in the classical spiked model it is the aspect ratio $p/n$. We give the limiting law of each spike in four regimes where n,p→ ∞. When p/n→ c>0, and when p/n→0 with p²/n→∞, the sampling term dominates, the limit does not depend on Oₚ and is Gaussian, as in the classical spiked model. When p²/n→0, the Haar randomness dominates, and the limit is a weighted sum of independent χ²₁ variables in the real case, and of independent exponential variables in the complex case. When p²/n→τ∈(0,∞), the two terms are of the same order, and the limit is the convolution of Gaussian and χ²₁ or Exponential variables. We then apply these results to a linear mixed model with a continuous batch effect, in which the Haar assumption is what makes the separation of the signal and of the batch possible, and we build a test for the value of a single spike.
No takes yet. Share an insight, caveat, or question.
Sakuma et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: