We study sample covariance matrices arising from multi-level components of variance. Thus, let $ B_n=\frac{ 1}{ N}\sum_{ j=1}^NT_{ j}^{ 1/2}x_jx_j^TT_{ j}^{ 1/2}$, where $x_j\in R^n$ are i.i.d. standard Gaussian, and $T_{ j}=\sum_{ r=1}^kl_{ jr}^2\Sigma_{ r}$ are $n\times n$ real symmetric matrices with bounded spectral norm, corresponding to $k$ levels of variation. As the matrix dimensions $n$ and $N$ increase proportionally, we show that the linear spectral statistics (LSS) of $B_n$ have Gaussian limits. The CLT is expressed as the convergence of a set of LSS to a standard multivariate Gaussian after centering by a mean vector $\Gamma_n$ and a covariance matrix $\Lambda_n$ which depend on $n$ and $N$ and may be evaluated numerically. Our work is motivated by the estimation of high-dimensional covariance matrices between phenotypic traits in quantitative genetics, particularly within nested linear random-effects models with up to $k$ levels of randomness. Our proof builds on the Bai-Silverstein \cite{ baisilverstein2004} martingale method with some innovation to handle the multi-level setting.
Categories: math.PR math.ST stat.TH