Gaussian fluctuations of eigenvalues of random Hermitian matrices associated with fixed and varying weights (Q2822699)

The probability density function for the set of eigenvalues \(\lambda_{1}\leq \lambda_{2} \leq \cdots \leq \lambda_{n}\) of an \(n\times n\) Hermitian matrix takes the form \(\mathcal{P}^{(n)}( \lambda_{1}, \lambda_{2}, \dots, \lambda_{n})= \frac{1}{Z_{n}} (\prod_{1\leq i< j\leq n} (\lambda_{i}- \lambda_{j})^{2})) (\prod_{j=1}^{n} \mu'_{n} (\lambda_{j}))\), where \(Z_{n}\) is the so-called partition function and \(\mu_{n}\) is an absolutely continuous measure supported on the real line. In many cases \(\mu'_{n}(x)= \exp (-n Q_{n}(x)),\) where \(Q_{n}\) is a given function.NEWLINENEWLINEThe sequence of (varying) orthonormal polynomials \(\{p_{n,m}(x)\}_{m\geq0}\) associated with the measure \(\mu_{n}\) yields the \(n\)th reproducing kernel for \(\mu_{n}\) defined as \(K_{n}(x,y)= \sum_{k=0}^{n-1} p_{n,k}(x) p_{n,k}(y)\). The basic formula for the probability density function defined as above is NEWLINE\[NEWLINE\mathcal{P}^{(n)}( \lambda_{1}, \lambda_{2}, \dots, \lambda_{n})= \frac{1}{n!}\det [K_{n}(\lambda_{i}, \lambda_{j}) (\mu' (\lambda_{i}) \mu' (\lambda_{j}))^{1/2} ]_{1\leq i,j\leq n}.NEWLINE\]NEWLINE In the paper under review, the author deals with the fluctuations of the eigenvalues of such matrices in order to determine the expected number of them in an interval when the measure is either a fixed weight on a compact interval, or an exponential weight on a possibly unbounded interval, or varying exponential weights. Notice that in the case \(\mu'_{n} (x)=\exp (-n x^{2})\), the Gaussian unitary ensemble, this problem was studied in [\textit{J. Gustavsson}, Ann. Inst. Henri Poincaré, Probab. Stat. 41, No. 2, 151--178 (2005; Zbl 1073.60020)]. For weights \(\mu'_{n}(x)=\exp( -2n Q(x)),\) where \(Q\) is a polynomial, the fluctuation of eigenvalues is studied in [\textit{D. Zhang}, Acta Math. Sin., Engl. Ser. 31, No. 9, 1487--1500 (2015; Zbl 1328.15053)].NEWLINENEWLINESuch fluctuations are given by lim \(\frac{\lambda_{j}- x_{j,n}}{\sqrt{\frac{\log n}{2 \pi^{2}}} \frac{1} {n \sigma_{n}^{*}(x_{j,n})}}= N(0,1)\) in a distribution sense. Here, \(x_{1,n}< x_{2,n} <\cdots< x_{n,n}\) are the zeros of the polynomials \(p_{n,n}\) and \(\sigma_{n}^{*}\) is the contracted density of an equilibrium distribution for the external field \(Q\). In other words, the normalized difference between the \(j\)th eigenvalue in the bulk and the \(j\)th zero of the polynomial \(p_{n,n}(x)\) is normally distributed in the bulk.NEWLINENEWLINEThe novelty of the methods of proof with respect to previous contributions is the fact that some much weaker asymptotics for the orthonormal polynomials and related quantities are required. More precisely, the author assumes some special asymptotic behavior of leading coefficients of such orthonormal polynomials, asymptotic spacing of their zeros, asymptotics and bounds for their reproducing kernels as well as pointwise asymptotics of orthonormal polynomials with a rate. These assumptions are not independent of one another.NEWLINENEWLINEThis paper constitutes a very nice example of the interplay between the theory of random matrices and the theory of orthogonal polynomials.