Optimal upper tail estimates for the edge eigenvalues of $β$-ensembles

Abstract

Hermite and Laguerre $β$-ensembles are important and well studied models in random matrix theory, with the special cases $β=1,2,4$ corresponding to classical matrix ensembles. Although sharp moderate-deviation estimates have been established for the largest eigenvalues, substantially less is known for other edge eigenvalues. Even for the limiting distributions $TW_β^{(k)}$, the leading constants in the tail asymptotics are not known for general $β$ and $k\ge2$. In this paper, for each fixed $k\ge2$, we prove matching upper and lower moderate deviation bounds for the right tail of the $k$-th largest eigenvalue, with the optimal exponential constant $2βk/3$, for general $β$. We also establish sharp lower tail estimates for the second largest eigenvalue of the Gaussian and Laguerre orthogonal ensembles, with exponential constant $1/24$, same as for the largest eigenvalue. For $TW_β^{(k)}$, we obtain matching right-tail asymptotics for $β\ge2/k$ and left-tail asymptotics for $0<β\le2$ for $k=2$. We also obtain new stochastic domination results for the edge eigenvalues. Our proofs combine variational arguments from tridiagonal matrix models, last passage percolation, stochastic comparisons of sums of edge eigenvalues, and superposition-decimation identities.