Extreme eigenvalues of log-concave ensemble (Q7010937)

In the context of random matrices, an important model is the so-called Whishart ensemble \(H\), which can be described as follows. Let \(X:=X_{m,n}\) be a \(m\times n\) random matrix, where all the entries are independent normal random variables. Then define \(H:= X X^{*}\), where \(X^{*}\) is the tranpose of \(X\). It turns out that \(H\) can be written as \N\[\NH=\frac{1}{n}\sum_{i=1}^{n}q_{i}q_{i}^{*},\N\]\Nwhere \(q_{i}\) is the \(i\)-th column of \(X\) multiplied by the factor \(\sqrt{n}\).\N\NAn important problem is to know the asymptotic behavior of the eigenvalues of \(H\), \(\lambda_{1}\geq \lambda_{2}\geq \ldots \geq \lambda_{m}\). For instance, within the study of the spectral empirical distribution (SED), a well-known result is that the SED converges to the so-called Marchenko-Pastur law. Moreover, after some rescaling, the limits of the extreme eigenvalues, \(\lambda_{1}\) and \(\lambda_{m}\), obey the so-called Trady-Widom distribution.\N\NIn this paper, the authors study the asymptotic behavior of the extreme eigenvalues of \(H\) when each \(q_{i}\) is an isotropic log-concave random vector, and all of these vectors are independent. This means that \(E(q_{i}q_{i}^{*})\) is the identity matrix, and that the density on \(\mathbb{R}^{m}\) of \(q_{i}\) is of the form \(e^{-V}\), with \(V\) convex. It is found the extreme eigenvalues also obeys the Trady-Widom distribution for this model.\N\NAnother ensemble studied here is called spiked model, which is of the form \(TXX^{*}T^{*}\), with \(T\) being some kind of perturbation of a unitary matrix, using some values which are outliers values. It is shown that the asymptotic behavior of the extreme eigenvalues is governed by a Gaussian distribution.\N\NAn important tool used in the analysis is the resolvent, \(G(z):= (H-zI)^{-1}\).