Asymptotic distribution of eigenvalues of weakly dilute Wishart matrices (Q1586062): Difference between revisions

Nonlinear equations are deduced for limits of eigenvalue distribution for dilute Gaussian and Wishart matrices of increasing dimension. The dilution means that matrices have only a small portion \(q\) of nonzero elements per each row. Suppose \(A\) is an \(N\times N\) matrix that has entries of the form \(N^{-1/2}a(x, y)\pi(x,y)\), where \(\{a(x,y)\}\) have a joint Gaussian distribution, \(x\leq y\), and expectation \[ {\mathbf E}a(x,y) a(s,t)= V(x- s)V(y- t)+ V(x- t)(y- s), \] where \(|V(x)|\) are uniformly bounded (the case of ``weakly correlated variables''). The independent random \(\pi_{xy}= \pi_{yx}\) equals \(q^{-1/2}\) with probability \(q\), otherwise \(0\). Theorem 1. The portion of eigenvalues of \(A\) less or equal \(u\) weakly converges as \(N\to\infty\) and tends to a function \(\sigma_1(u, V,q)\) such that its Stieltjes transform is uniquely determined by \(V(\cdot)\) and \(q\) by a nonlinear functional equation. Suppose now that \[ B= N^{-1} \sum^m_{\mu=1} \xi_\mu(x) \xi_\mu(y)\pi(x,y), \] where \(m>1\), \(\xi\mu(x)\) are independent \(N\)-dimensional vectors having standard normal distribution, and \(\pi(x,y)\) are the same independent random variables as before. Theorem 2. The portion of eigenvalues of \(B\) less or equal \(u\) weakly converges as \(N\to\infty\) and \(m/N\to c\) and tends to a function \(\sigma_2(u,c,q)\) such that its Stieltjes transform is uniquely determined by \(c\) and \(q\) by a nonlinear functional equation. It is noteworthy that even a weak dilution removes the well-known singularity at the origin produced by a finite contribution of zero eigenvalues.