Large deviation theorem for zeros of polynomials and Hermitian random matrices

Random matrices (algebraic aspects) (15B52) Potentials and capacity, harmonic measure, extremal length and related notions in two dimensions (31A15) Random matrices (probabilistic aspects) (60B20) Large deviations (60F10)

Abstract: We give abstract versions of the large deviation theorem for the distribution of zeros of polynomials and apply them to the characteristic polynomials of Hermitian random matrices. We obtain new estimates related to the local semi-circular law for the empirical spectral distribution of these matrices when the 4th moments of their entries are controlled.

Review text

Let \(\mu_{0}\) be a given probability measure whose support is contained in a smooth compact non-closed curve \(K\) in the complex plane and denote by \(u_{0}\) its logarithmic potential that is assumed to be Hölder continuous on \(K.\) If \(Q\) is a polynomial of degree \(n\) and \(a_{k}=1, 2, \ldots, n,\) are their zeros, then the counting measure of \(Q\) is defined as \(\mu_{Q}= \frac{1}{n} \sum_{k=1}^{n} \delta_{a_{k}},\) i.e., it is the probability measure equidistributed on the zero set of \(Q.\) For two probability measures \(\mu, \nu\) with compact support on the complex plane we denote \(d(\mu, \nu) = ||u - v||_{L^1(\bar{\mathbb{C}})}\), where the standard Hermitian metric given by the Fubini-Study form is considered in the natural compactification of the complex plane. Here, \(u\) (resp. \(v\)) is the logarithmic potential of \(\mu\) (resp. \(\nu\)). A sufficient condition for the equidistribution of zeros of a polynomial \(Q\) with respect to a measure \(\mu_{0}\) as above is the first main result of this contribution. Indeed, if you assume that there exists a real number \(c_{n} \geq 1\) such that the expectation of \(|Q(z)|^{\alpha}, \alpha >0,\) is bounded above by \(c_{n} e^{\alpha n u_{0} (z)}\) for all \(z\in K,\) then for every \(\delta >0\) \[\mathrm{Prob}\{ Q, deg Q=n, \mathrm{dist} (\mu_{Q}, \mu_{0}) \geq \delta\} \leq A_{1} n^{A_{1}} e^{- A_{2} \delta n}.\] Here, \(A_{j}= A_{j}( K, \mu_{0}, \alpha), j= 1,2,\) are positive constants independent on \(n, c_{n}, \delta.\) This means that only with an upper bound for polynomials, the zeros of these polynomials are almost distributed with respect to the measure \(\mu_{0}.\) Notice that in the above result the zeros of the polynomial \(Q\) do not necessarily lie in the curve \(K.\) Under some extra conditions the author obtains an uniform estimate of the number of zeros of \(Q\) assuming that they satisfy the condition \(|\mu_{Q} (I)- \mu_{0} (I)| \geq \delta^{1/2}\) for at least one arc \(I\) in a closed subset \(L\) of a smooth curve \(l\) containing \(K,\) \(\mu_{0}\) has a bounded density with respect to the Lebesgue measure on \(l\) and the zeros of almost polynomial \(Q\) lie in \(l\). In a second step, these results are discussed in the framework of characteristic polynomials of Hermitian random matrices. In such a case, the natural choice is \(I=L= \mathbb{R}\) since the eigenvalues of such matrices are real numbers. For the empirical spectral distribution of the Wigner Hermitian matrix \(W_{n}= n^{-1/2}M_{n}\) associated with a Hermitan random matrix (denoted \(\mu_{W_{n}}\)) it is well known that it converges to the semicircle distribution \(\mu_{sc}\) in the interval \(K= [-2,2]\) almost surely once you fix \(\alpha, \beta\) or they are not too big (see, among others, [\textit{G. W. Anderson} et al., An introduction to random matrices. Cambridge: Cambridge University Press (2010; Zbl 1184.15023)]). The entries of \(M_{n}\) are of zero mean, the diagonal entries have variance bounded by a nonnegative real number \(\alpha,\) the off diagonal entries have variance 1 and their fourth moments are bounded by a positive real number \(\beta.\) Under such conditions, there are universal positive constants \(A_{j}\), \(j=1,2\), i.e., independent on \(n, \alpha, \beta, \delta,\) such that for \(\delta >0\) there exists a set \(\Omega_{n}(\delta) \) of \(n \times n\) Hermitian matrices satisfying \(\mathrm{Prob}( M_{n}\in \Omega_{n}(\delta) ) \leq A_{1} n^{A_{1}}e^{\alpha + \beta/2} e^{- \delta n A_{2}}.\) Moreover, if \(M_{n}\) does not belong to \(\Omega_{n}(\delta) \) and \(I\) is an interval of the real line, then dist \((\mu_{W_{n}}, \mu_{sc}) \leq \delta\) and \(|\mu_{W_{n}}(I) - \mu_{sc}(I) | \leq \delta^{1/2}.\) An application of this result in the case of adjacency matrices of the Erdős- Rényi random graph is presented. Notice that estimates on the rate of convergence and the probability to have good rate of convergence are known in the literature (see, in particular, [\textit{L. Erdős} et al., Ann. Probab. 41, No. 3B, 2279--2375 (2013; Zbl 1272.05111)]).

Recommendations

Cites work

Cited in

(5)

This page was built for publication: Large deviation theorem for zeros of polynomials and Hermitian random matrices

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2187686)