The mutual affinity of random measures (Q1882120)

The paper considers a set of probability measures defined on a finite event space \(\Omega\), and defines the notion of mutual affinity among these measures in terms of spectrum of the associated Gram matrix. The aim of the paper is to study, in Section 2 of the paper, the Gram matrix for independently and randomly chosen probability measures, when the cardinality \(n=\text{Card}(\Omega)\) of \(\Omega\), and the number \(k(n)\) of measures become large, i.e.\ in the limit \(n\), \(k\to \infty\), provided that the ratio \(k/n=\alpha\) (or \(k/n\to \alpha)\), with \(\alpha\) a fixed positive number. The authors prove, on the basis of Stieltjes transform, that the empirical eigenvalue distribution of the Gram matrix of the i.i.d.\ (independent and identically distributed) probability measure converges in expectation to the Marchenko-Pastur distribution. For the spectrum of the Gram matrix \(G\) it is pointed out the presence of one eigenvalue which is much larger than the others. This eingenvalue is the norm of the Gram matrix \(G\), and \(G\) is positive definite, it grows linearly with the value \(n=\text{Card}(\Omega)\). The remaining eigenvalues are shown to typically concentrate on an interval close to zero. Moreover, Section 3 strengthens these results for the almost sure convergence of the empirical eigenvalue distributions.