Spectrum of large random reversible Markov chains: heavy-tailed weights on the complete graph (Q717890)

The main object of study is the asymptotic behaviour of the spectral distribution of certain random matrices. The model under consideration depends on a single probability measure \(U\) supported on the non-negative real numbers. The authors consider first a Wigner (i.e., with i.i.d. entries) \(N\times N\) matrix with elements distributed according to \(U\) and then normalize the rows so that the sum along each row is \(1\). The resulting random matrix \(K\) can be also viewed as a (random) Markov kernel of a random walk on the complete graph with \(N\) vertices. The authors consider the empirical spectral distribution \(\mu_K\) of the matrix \(K\) (i.e., the measure with atoms at eigenvalues of \(K\)) and its limit behavior (perhaps, after some rescaling) as \(N\to\infty\). If the second moment of \(U\) exists, then in the limit one uncovers the familiar Wigner semi-circle law, but the situation changes dramatically when \(U\) has heavy tails. In the paper the case when \(U\) has a tail of index \(\alpha\), that is, when \(U([t,+\infty))\) decays like \(t^{-\alpha}\) up to some slowly changing function, is studied. The authors prove that when \(\alpha\in(1,2)\), then \(\mu_K\) converges to a non-random limit law which turns out to be the same as for Wigner (i.e., not normalized) random matrices with \(U\)-distributed entries. The intuitive reason for this fact is the law of large numbers for the sums of \(U\)-distributed random variables which means that the normalization constant in each row is approximately the same and does not change the shape of the limit distribution. This is no longer true when \(\alpha<1\) and, indeed, the authors prove that there is a phase transition at \(\alpha=1\). For \(\alpha\in(0,1)\), the limit of \(\mu_K\) still exists but no longer coincides with the one appearing in Wigner matrices. The relation of the limiting law with Poisson weighted infinite trees and the Poisson-Dirichlet distribution is also explained in the article.