Limit of infinite band width for product of two random matrices (Q4396199)

The authors consider \(N\times N\) random matrices \(H_{N,R}\) with entries \(H_{N,R}(x,y)={1\over{NR}}\sum_{j=1}^{m}\xi_{j}(x)\) \(\xi_{j}(y) \varphi_{R}(x-y),\;x,y=1,\dots,N\), where \(\xi_{j}(x)\) are jointly independent Gaussian random variables with \(E\xi_{j}(x)=0, E[\xi_{j}(x)]^2=v^2\). Let \(\varphi(t),t\geq 0\) be a continuous function that satisfies the conditions \(\varphi(0)=0,| \varphi(t)| \leq\phi_{0},\int_{0}^{\infty}\varphi(t)dt= \phi_{1}<\infty\) and let \(\varphi_{R}(x)=\varphi(| x| R^{-1})\). Then the normalized eigenvalue counting function \(\sigma(\lambda;H_{N,R})=\sharp \{\lambda_{j}^{(N)}\leq \lambda\}/N\), (\(\lambda_{j}^{(N)}\) are eigenvalues of \(H_{N,R}\)) weakly converges in probability as \(m,N,R\to\infty\) to a nonrandom function \(\sigma_{c,b}(\lambda)\) depending on parameters \(c=\lim_{m,N\to\infty} {m/N}\) and \(b=\lim_{N,R\to\infty}{R/N}\). The Stieltjes transform \(g(z)=\int(\lambda-z)^{-1}d\sigma_{c,b}(\lambda)\) can be found from the relation \(g(z)=\int_{0}^{1}f(t,z)dt\), where \(f(t,z)\) is a solution of the system of equations \[ -zf(t)=1+v^2f(t)p(t,t'), \] \[ p(t,t')={cv^2\over b}\int_{0}^{1}\varphi \left({(t-u)/b}\right)f(u)\varphi\left((u-t')/b \right)du- {1\over{\sqrt{b}}}\int_{0}^{1}p(t,u)f(u)\varphi\left((u-t')/ b \right)du. \]