arXiv:2609.22079
Universal Eigenvector Statistics of Non-Hermitian Random Matrices
Ze Chen, Zhenyu Xiao, Shinsei Ryu
Abstract
Eigenvector overlaps quantify nonorthogonality and govern the response and dynamics of non-Hermitian systems. We extend the universality of non-Hermitian random matrices to the statistics of these overlaps. We obtain analytical expressions for eigenvector overlaps in the spectral bulk and near the origin, covering ten symmetry classes in the limit of large matrix size. Using fermionic replica nonlinear $σ$ models, we relate these overlaps to Hermitian level statistics, symmetry class by symmetry class and topological sector by topological sector. Numerical calculations in various physical models support the universality of the normalized overlaps in the regimes studied. Our work establishes a duality between Hermitian level statistics and non-Hermitian eigenvector overlaps.
AI Summary
The paper asks how symmetry and topology organize eigenvector nonorthogonality in non-Hermitian random matrices, and whether its statistics can be related to Hermitian spectral correlations. It studies density-weighted diagonal and off-diagonal spectral-projector overlaps across the ten AZ$_0$ classes, whose defining constraints involve transposition and sublattice symmetry rather than complex or Hermitian conjugation. In the large-matrix limit, normalized bulk off-diagonal overlaps fall into three families selected solely by transposition-based time-reversal symmetry, with $\beta=1,2,4$. They are expressed through the Poisson transform $\mathcal P_2$ of the corresponding Hermitian pair correlation: $2\pi\mathcal O_2/\mathcal O_1=-1+\partial_s[s\mathcal P_2(s)]$, where $s=|z_1-z_2|^2$ in unfolded units. All three have the mesoscopic tail $\mathcal O_2\sim-2\mathcal O_1/(\beta\pi|z_1-z_2|^4)$. Near the origin, diagonal overlaps satisfy $\pi\mathcal O_1=\beta[1+\mathcal P_1(|z|^2)]/4$, with the Hermitian hard-edge density determining class and topological-sector dependence. Protected zero modes induce $|z|^{-2}$ contributions to the overlap density of nonzero eigenvalues. The relations follow from fermionic replica nonlinear $\sigma$ models, reducing overlap-source insertions to derivatives of spectral partition functions. Numerical comparisons cover Gaussian ensembles, non-Hermitian SYK and Anderson models, and dissipative bipartite fermions. The predictions concern normalized profiles, not universal absolute amplitudes, and apply in specified bulk and origin scaling regimes, not generally at outer edges or symmetry axes. Physical hard-edge comparisons calibrate spectral scale and overlap amplitude separately by sector. The zero-mode overlap weight itself is excluded; extension beyond random-matrix statistics, including localization transitions, remains open.
Error found: The claimed analytical hard-edge coverage of seven symmetry classes is incompletely justified: the overlap–Poisson-transform identity is derived explicitly for class D but asserted without the corresponding source reductions for the other six classes.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.