A refinement of Wigner's semicircle law in a neighborhood of the spectrum edge for random symmetric matrices (Q1281515)

The paper concerns the spectral properties of the Wigner ensemble of random symmetric \(N\times N\) matrices \(W_N\) whose elements \(\{w_{ ij}N^{-1/2}, i\leq j\}\) are jointly independent random variables with zero mean value and variance \(1/4N\). The main purpose is to study the asymptotic behaviour of the averaged traces \(M_{2p}(N) =\mathbb E \text{Tr} W_N^{2p}\) in the limit \(N,p\to \infty\). The principal result obtained is that if \(1\ll p\ll N^{2/3}\), then \[ M_{2p}(N)= {N\over \sqrt{\pi p^3}} \bigl(1+o(1) \bigr)\tag{1} \] provided \(V_k=\mathbb E [w_{ij}]^{2k} \leq(\text{const.} k)^k\). Expression (1) coincides with the leading term of the asymptotics of \(M_{2p} (N)\), \(N\to\infty\) derived for fixed but large enough \(p\) by \textit{E. Wigner} [Ann. Math. (2) 62, 548--564 (1955; Zbl 0067.08403)]. This means that the eigenvalue distribution of \(W_N\) near the edge of the spectrum is governed by Wigner's semicircle law not only on the ``macroscopic'' scale studied by E. Wigner, but on the ``mesoscopic'' one as well. It is also shown that on this scale the distribution of random variables \(\text{Tr } W_N^{2p} -M_{2p}(N)\) converges to the Gaussian one. The parameters of the latter do not depend on \(V_k\). To prove (1), the authors develop deep modification of the approach originally suggested by E. Wigner. The principal idea here is to consider the averaged trace \[ \mathbb{E}\sum_i [W_N^{2p}]_{ii} ={1\over N} \sum_{i,\{j\}} \mathbb E w_{ij_1} w_{j_1j_2} \cdots w_{j_{2p-1}i} \tag{2} \] as the weighted sum over paths \(\eta= (i,j_1, \dots, j_{2p-1},i)\). The leading contribution to (2) gives the classes of paths \(\eta\) that correspond to trajectories of simple positive random \(2p\)-step walks. Consideration of the limit \(N,p\to \infty\) in (1) requires knowledge of the statistics self-intersections in the paths. This problem is studied in detail on the basis of combinatorial analysis.