Poisson convergence for the largest eigenvalues of heavy tailed random matrices (Q731725): Difference between revisions

The statistics of the largest eigenvalues of two important families of random matrices, i.e., of real symmetric and sample covariance matrices with heavy tailed entries are studied. Firstly, the case of random real symmetric matrices with independent and heavy tailed entries are considered. Let~\((a_{ij})\), \(1 \leq i \leq n\), \(1 \leq j \leq n\) be i.i.d.~random variables of matrix \(A_n\) such that \({\mathbb P}(|a_{ij}| > x)=L(x)x^{-\alpha}\), where \(\alpha > 0\) and \(L(x)\) is slowly varying function. For any \(\alpha > 0\) the point process of extreme values of the entries of \(A_n\) is asymptotically Poissonian. The largest eigenvalues of \(A_n\) behave as the largest entries of the matrix \(A_n\) when \(0 < \alpha < 4\). This result of \textit{A. Soshnikov} [Electron. Commun. Probab. 9, 82--91 (2004; Zbl 1060.60013)] for \(0 < \alpha < 2\) is here extended under the condition that the fourth moment is absent. Then, the behavior of the ensemble of large random covariance matrices is also investigated. A sample covariance matrix is given as \( X_n = \frac {1}{p}A_n {A_n}^t\), where \(A_n\) is an \(n \times p\) random matrix with i.i.d. centered entries \((a_{ij})\), \(1 \leq i \leq n\), \(1 \leq j \leq p\). The same result as for the case of random symmetric matrices is derived through an estimate of traces of high powers of random matrices properly truncated.