Eigen-inference by Marchenko-Pastur inversion

Abstract: A new formula for Marchenko-Pastur inversion is derived and used for inference of population linear spectral statistics. The formula allows for fast and accurate estimation of the Stieltjes transform of the population spectral distribution $s_H(z)$, when $z$ is sufficiently far from the support of the population spectral distribution $H$. If the dimension $d$ and the sample size $n$ go to infinity simultaneously such that $\frac{d}{n} \rightarrow c>0$, the estimation error is shown to be asymptotically less than $\frac{n^{\varepsilon}}{n}$ for arbitrary $\varepsilon > 0$. By integrating along a curve around the support of $H$, estimators for population linear spectral statistics are constructed, which benefit from this convergence speed of $\frac{n^{\varepsilon}}{n}$. The new method of estimating the Stieltjes transforms $s_H(z)$ is also applied to the numerical construction of estimators for the population eigenvalues, which in a simulation study are demonstrated to outperform state-of-the-art Ledoit-Wolf estimators.