On the norm and eigenvalue distribution of large random matrices (Q1807210)

This investigation is an extension of the paper by the second author [Mat. Fiz. Anal. Geom. 3, No. 1/2, 80-101 (1996; Zbl 0881.15025)]. Symmetric random matrices \(H=\{H_{ij}\}\) of size \(N\times N\) are considered whose entries have a joint Gaussian distribution for each \(N\) as \(N\to\infty\). It is assumed that \[ EH_{ij}=0,\;EH_{ij}H_{kl}=N^{-1}(V_{ik}V_{jl}+ V_{il} V_{jk}), \quad i,j,k,l=1, \dots,N, \] where the real symmetric matrices \(V= \{V_{ij}\}\) have uniformly bounded spectral norms not greater than \(a\) and the quantities \(\lambda=N^{-1} \text{tr} V\) have the upper bounds \(v_2\) and lower bounds \(v_1>0\) for \(N=1,2,\dots\) Additionally, it is supposed that the limit ``counting function'' \(\nu(u)\) exists that presents the limit value of a portion of eigenvalues of \(V\) that does not exceed \(u\), \(u\geq 0\). These assumptions are satisfied for the matrices \(H\) with Gaussian entries weakly correlated for large distances as measured by shifts of the indexes \(i\) and \(j\). Similarly to Khorunzhy (loc. cit.), the following results are obtained: (1) Spectral norms of \(H\) have an upper limit with probability 1 as \(N\to\infty\) that is not greater than \(2\sqrt{v_2a}\). (2) There exists almost surely the limit counting function \(\mu(u)\) for eigenvalues of \(H\). (3) The Stieltjes transform \(f(z) \) of \(d\mu(u)\) can be found from the relations \[ f(z)=-\int^\infty_0 \bigl(z+ \lambda g(z)\bigr)^{-1} d\nu(\lambda),\quad g(z)=-\int^\infty_0 \bigl(z+ \lambda g(z)\bigr)^{-1} \lambda d\nu (\lambda), \] where the equation for \(g(z)\) can be uniquely solved. (4) The support of \(\mu(u)\) is located on a segment defined by the support of \(\nu(u)\). (5) Under a supplementary condition, no eigenvalues of \(H\) lay outside the support of \(\mu(u)\) with probability 1. For diagonal matrices \(V\), a quadratic equation is obtained for \(f(z)\). To prove these statements, the asymptotic behaviour of expectation values of powers of the matrices \(H\) is considered. For these matrices, the counting function convergence is proved to be much faster than the strong self-averaging predicts.