The circular law for sparse non-Hermitian matrices (Q2327948)

Let \(\lambda_1,\dots, \lambda_n\) the eigenvalues of a \(n\times n\) matrix \(B\); its empirical spectral distribution (ESD) is defined by \(L_B:=\frac{1}{n}\,\sum_{i=1}^n \delta_{\lambda_i}\), where \(\delta_x\) is the Dirac measure concentrated at \(x\). The sub-Gausssian norm of a random variable \(\xi\) is defined by \[ \|\xi\|_{\psi_2}:=\sup_{k\ge 1} k^{-1/2}\,\mathbb{E}^{1/k}(|\xi|^k)\,. \] The main result of this important paper is the following theorem, which extends previous results quoted in the introduction. Theorem. Let \(A_n\) be an \(n\times n\) matrix with i.i.d. entries \(a_{i,j}=\delta_{i,j}\,\xi_{i,j}\), where the \(\delta_{i,j}\) are independent Bernoulli random variables taking the value \(1\) with probability \(p_n\in\left]0,1\right]\) and \(\xi_{i,j}\) are real-valued i.i.d. sub-Gaussian centred random variables with unit variance. \begin{itemize} \item[(i)] If \(p_n\) is such that \(np_n=\omega(\log^2 n)\), then as \(n\to\infty\) the ESD of \(A_n/\sqrt{n\,p_n}\) converges weakly in probability to the circular law. \item[(ii)] There exists a constant \(c\), which depends only on the sub-Gaussian norm of \(\{\xi_{i,j}\}\), such that if \(p_n\) satisfies the inequality \(np_n>\exp(c\,\sqrt{\log n})\), then the conclusion of (i) holds almost surely. \end{itemize} Here, if \((a_n)\) and \((b_n)\) are two sequences of positive reals, one writes \(a_n=\omega(b_n)\) if \(b_n=o(a_n)\), \(a_n=O(b_n)\) and \(\limsup_{n\to\infty} a_n/b_n<\infty\).