A Two-Spectrum Deterministic Equivalent for Random Skew-Kernel Matrices
Abstract
Rank-one perturbations of the mean characteristic polynomial of a random skew-kernel matrix assign a positive spatial weight to each root.
For independent centered isotropic rows, the resulting response tracks the sample almost surely against each deterministic spatial test, uniformly over positive scales and arbitrary inputs, exactly when row energies concentrate on common projections of proportional rank.
Directional absolute first moments determine this condition.
The determinant Jensen gap controls the spatial response in every real orthonormal basis.
For identity spatial covariance, logarithmic derivatives of the mean determinant give the temporal weights of the row energies it detects.
Flat kernels of one prescribed rank suffice to test the row condition.
A scalar equation in the two input spectra approximates the matrix response of the roots within $4/d$ in normalized trace norm.
A varying ordinary response and the same-scale overlap of independent samples recover both input laws up to reciprocal rescaling, with Lipschitz stability of the relative spatial parameter and an optimal logarithmic modulus for the compactified laws on intervals of positive variation.
Brownian area gives explicit spatial half-Cauchy mixtures.