Eigenvector decorrelation for random matrices
Giorgio Cipolloni, László Erdős, Joscha Henheik, Oleksii Kolupaiev
TL;DR
The paper analyzes how eigenvectors of two random, deformed Wigner matrices decorrelate under perturbations. It develops a two-resolvent local law for products of resolvents $G_1 A G_2$ with a deterministic approximation $M_{12}^A$, and introduces a stability framework with a time-dependent control parameter and a linear LT term to capture deformation- and energy-difference effects. It then proves two main results: (i) an ETH-type decomposition for regular observables and an overlap-based ETH for general observables, and (ii) an optimal bound on eigenvector overlaps that decays with the deformation distance $igra (D_1-D_2)^2ig a$ and the energy differences, leading to strong decorrelation when the deformations diverge or spectra separate. The analysis hinges on a zigzag strategy combining a global law, a characteristic-flow zig, and a Green function comparison zag, complemented by a rigorous stability analysis and careful handling of regular observables. The results extend ETH-type eigenvector statistics to pairs of spectral families and quantify decorrelation in a broad random-matrix setting with deterministic deformations, offering new insights into eigenvector behavior under perturbations relevant for quantum chaos and high-dimensional statistics.
Abstract
We study the sensitivity of the eigenvectors of random matrices, showing that even small perturbations make the eigenvectors almost orthogonal. More precisely, we consider two deformed Wigner matrices $W+D_1$, $W+D_2$ and show that their bulk eigenvectors become asymptotically orthogonal as soon as $\mathrm{Tr}(D_1-D_2)^2\gg 1$, or their respective energies are separated on a scale much bigger than the local eigenvalue spacing. Furthermore, we show that quadratic forms of eigenvectors of $W+D_1$, $W+D_2$ with any deterministic matrix $A\in\mathbf{C}^{N\times N}$ in a specific subspace of codimension one are of size $N^{-1/2}$. This proves a generalization of the Eigenstate Thermalization Hypothesis to eigenvectors belonging to two different spectral families.
