This paper studies how a certain integral over pairs of vectors behaves for very large non-Hermitian matrices. In plain terms, the authors fix the lengths of two vectors and their mutual scalar product, and then average an exponential weight over all such vector pairs. They show that, as the matrix size N grows, the dominant contribution to that integral can change suddenly. The change, or phase transition, is visible at the saddle point used in large-N asymptotics and is linked to a matrix quantity called the largest singular value.

More concretely, the authors work with a shifted matrix B_z = M - z, where M is the given (possibly random) complex matrix and z is a point in the complex plane. Singular values are the non-negative numbers that measure the action of a matrix on directions; each has a left and a right singular vector, the input and output directions that realize that action. In the regime they call “delocalized,” the integral’s asymptotics are governed by two non-Hermitian transforms R_1 and R_2 introduced in earlier work by Bousseyroux and Potters. At a critical value the regular saddle point hits the largest singular value, and beyond that the integral’s mass concentrates, or “localizes,” on the associated left and right singular vectors. The authors state this result as their main saddle-point transition (Result: thm:main_transition).

Why this matters: in the Hermitian (symmetric) random-matrix world, a similar saddle-point sticking is the mechanism behind large deviations of extreme eigenvalues. There are two standard ways to study such rare events: one uses spherical integrals and saddle points, the other uses the Coulomb-gas picture of interacting charges (the eigenvalues). For many non-Hermitian ensembles the eigenvalue law in the plane is not available in a Coulomb-gas form, so extending the spherical-integral route is valuable. The authors use their non-Hermitian spherical integral analysis to formulate conjectures about one-eigenvalue large deviations (the cost to place a single eigenvalue at a prescribed point) and about fluctuations near the spectral boundary.