Optimal tail estimates in $\beta$-ensembles and applications to last passage percolation

Abstract

Hermite and Laguerre $\beta$-ensembles are important and well studied models in random matrix theory with special cases $\beta=1,2,4$ corresponding to eigenvalues of classical random matrix ensembles. It is well known that the largest eigenvalues in these, under appropriate scaling, converge weakly to the Tracy-Widom $\beta$ distribution whose distribution function $F_{\beta}$ has asymptotics given by $1-F_{\beta}(x)=\exp\left(-\frac{2\beta}{3}(1+o(1))x^{3/2}\right)$ as $x\to \infty$ and $F_{\beta}(x)=\exp\left(-\frac{\beta}{24}(1+o(1))|x|^3\right)$ as $x\to -\infty$. Although tail estimates for the largest eigenvalues with correct exponents have been proved for the pre-limiting models, estimates with matching constants had not so far been established for general $\beta$; even in the exactly solvable cases, some of the bounds were missing. In this paper, we prove upper and lower moderate deviation estimates for both tails with matching constants. We illustrate the usefulness of these estimates by considering certain questions in planar exponential last passage percolation (LPP), a well-studied model in the KPZ universality class in which certain statistics have same distributions as largest eigenvalues in Laguerre $\beta$-ensembles (for $\beta=1,2,4$). Using our estimates in conjunction with a combination of old and new results on the LPP geometry, we obtain three laws of iterated logarithm including one which settles a conjecture of Ledoux (J. Theor. Probab., 2018). We expect that the sharp moderate deviation estimates will find many further applications in LPP problems and beyond.