Non-Hermitian Rank-One Perturbations

• The paper demonstrates that adding a rank-one non-Hermitian perturbation to a Hermitian operator yields new spectral phenomena, analyzed using resolvent identities like the Sherman–Morrison formula.
• It reveals how eigenvalue outliers and resonance poles emerge via determinant conditions in both deterministic and random settings, impacting quantum resonance theory and ensemble statistics.
• Practical applications span quantum physics, neural network dynamics, and renormalization models, highlighting shifts in spectral measures and non-orthogonal eigenvector structures.

A non-Hermitian perturbation of rank one is an operation in which a fixed linear operator, typically a normal or Hermitian matrix/operator $A$ (or $H_0$), is perturbed by an operator $V$ of the form $V = u v^*$ (matrix case) or $V = \kappa |\phi\rangle\langle\psi|$ (in operator-theoretic context), where $u, v$ (resp. $\phi, \psi$) are fixed vectors and $\kappa$ is a coupling constant that may be complex. This generates an operator $A + V$ (or $H_0 + V$) which is, in general, non-Hermitian whenever $H_0$0 or $H_0$1, and introduces fundamentally new spectral phenomena in both deterministic and random settings. Non-Hermitian rank-one perturbations are central to a variety of domains, including quantum resonance theory, random matrix theory, neural network dynamics, and operator theory, owing to their ability to induce spectral instabilities, eigenvalue outliers, resonance poles, and nontrivial eigenvector overlap structure (Bourget et al., 2017, Dubach et al., 6 Jan 2026, Dubach et al., 2021, Kozhan, 2015, Dereziński, 2016, Amara et al., 2022, Alpan et al., 2021).

1. Operator-Theoretic Formulation and Resolvent Identities

Let $H_0$2 be a densely defined self-adjoint operator on a Hilbert space $H_0$3, and $H_0$4 a rank-one perturbation, with $H_0$5 and $H_0$6. The perturbed operator is $H_0$7, which, except in the Hermitian case ($H_0$8, $H_0$9), is non-self-adjoint. The Sherman–Morrison formula (matrix-inverse lemma) and Kreĭn–Birman resolvent identity provide the central tools for analyzing the perturbed resolvent:

$V$0

where $V$1. This highlights that new poles (i.e., spectral points of $V$2 not present in $V$3) correspond to zeros of the scalar denominator, leading to the resonance equation $V$4 (Bourget et al., 2017, Dubach et al., 6 Jan 2026).

In random matrix settings, for $V$5 and $V$6, the analogous formula holds:

$V$7

with implications for both the spectrum and eigenvector statistics (Dubach et al., 6 Jan 2026, Kozhan, 2015, Alpan et al., 2021).

2. Spectral Outliers and Resonance Poles

A hallmark of non-Hermitian rank-one perturbations is the emergence of eigenvalue outliers and resonance poles that detach from the bulk spectrum. For random matrices, the master equation $V$8 determines the location of outliers. In classic ensembles (Ginibre, GUE, unitary), detailed phase transitions occur:

Ensemble	Rank-One Perturbation	Outlier Condition
Ginibre (complex)	$V$9	$V = u v^*$0
GUE (Hermitian)	$V = u v^*$1	$V = u v^*$2
Haar unitary (CUE)	Multiplicative spike	$V = u v^*$3

For the anti-Hermitian spike of GUE, there exists a sharp BBP-type threshold: No outlier for $V = u v^*$4, and for $V = u v^*$5, a unique outlier appears near $V = u v^*$6, with high precision for large $V = u v^*$7 (Dubach et al., 2021, Dubach et al., 6 Jan 2026). This mechanism generalizes to operator-theoretic resonances: Embedded eigenvalues in the continuous spectrum are destabilized, disappearing as true eigenvalues and replaced by resonance poles $V = u v^*$8 satisfying $V = u v^*$9, located in the lower half-plane. The perturbed spectral measure exhibits a Lorentzian-type singularity near the original eigenvalue (Bourget et al., 2017).

3. Statistical Properties and Joint Eigenvalue Distributions

The distributional laws for eigenvalues under rank-one non-Hermitian perturbations are explicit in both Gaussian and Laguerre $V = \kappa |\phi\rangle\langle\psi|$0-ensembles. Given a Hermitian tridiagonal $V = \kappa |\phi\rangle\langle\psi|$1 and perturbation $V = \kappa |\phi\rangle\langle\psi|$2, the joint eigenvalue density $V = \kappa |\phi\rangle\langle\psi|$3 for the imaginary perturbation is:

$V = \kappa |\phi\rangle\langle\psi|$4

where $V = \kappa |\phi\rangle\langle\psi|$5 is the law of the perturbation norm and $V = \kappa |\phi\rangle\langle\psi|$6 is a normalization (Kozhan, 2015).

For chiral Gaussian $V = \kappa |\phi\rangle\langle\psi|$7-ensembles perturbed by an anti-Hermitian rank-one term, the joint PDF in the upper half-plane is governed by the Vandermonde determinant and an additional chiral repulsion $V = \kappa |\phi\rangle\langle\psi|$8, encoding the pairing of complex conjugate eigenvalues and nontrivial correlations (Alpan et al., 2021).

4. Dynamical, Resonance, and Time-Decay Effects

Non-Hermitian rank-one perturbations induce explicit dynamical effects in both deterministic and random settings. In quantum systems, the Feshbach–Livšic reduction and Aronszajn–Donoghue theory show that the perturbed resolvent develops a resonance pole $V = \kappa |\phi\rangle\langle\psi|$9, with decay width governed by the Fermi golden rule $u, v$0. The time-dependent survival amplitude $u, v$1 exhibits almost exponential decay with rate $u, v$2 as established for quasi-Lorentzian Fourier kernels (Bourget et al., 2017).

In random matrix models with time-dependent anti-Hermitian perturbations $u, v$3, the spectral evolution tracks the emergence and trajectory of the outlier eigenvalue, governed by explicit ODEs and a critical phase transition at $u, v$4 (Dubach et al., 2021).

5. Complex Eigenvector Structure and C-Normality

Non-Hermitian rank-one perturbations generate non-orthogonal eigenvectors, biorthogonal systems, and spectral instabilities. Operator-theoretic advances classify perturbed operators $u, v$5 as $u, v$6-normal (admitting a conjugation $u, v$7 such that $u, v$8) if and only if the joint spectral measures of $u, v$9 and $\phi, \psi$0 for $\phi, \psi$1 coincide on all Borel sets. This geometric matching provides a complete characterization and constructive approach for conjugations and defines a broad class strictly between normal and arbitrary operators, including truncated Toeplitz and block models (Amara et al., 2022).

6. Applications in Physics, Mathematics, and Neural Networks

Non-Hermitian rank-one perturbations have significant impact in:

• Quantum resonance theory: Open quantum Hamiltonians $\phi, \psi$2 with $\phi, \psi$3 finite rank, modeling decay and resonance statistics (Bourget et al., 2017, Dubach et al., 6 Jan 2026).
• Random matrix and statistical physics: Spiked models underlie the location and statistics of resonance poles and outlier eigenvalues in Ginibre, GUE, CUE, and chiral ensembles (Kozhan, 2015, Alpan et al., 2021, Dubach et al., 6 Jan 2026).
• Neural networks: The addition of rank-one excitatory/inhibitory weights impacts stability and dynamics; LoRA adapters in LLMs are realized as low-rank perturbations, and their influence on the Jacobian spectrum affects trainability and generalization (Dubach et al., 6 Jan 2026).
• Renormalization-group toy models: The scaling flows of homogeneous rank-one perturbations model toy RG fixed points and spectral transitions, with exact analytic formulation (Dereziński, 2016).

7. Historical Development and Open Problems

The matrix-determinant lemma (Sylvester’s identity) forms the mathematical foundation, while the BBP phase transition underscored the richness of low-rank Hermitian deformations. Non-Hermitian cases were rigorously analyzed by Tao (Ginibre), Fyodorov–Khoruzhenko, O'Rourke–Wood, and others. Current open problems include the fluctuation laws of spectral outliers at finite $\phi, \psi$4, full eigenvalue-eigenvector flow dynamics, and the role of left/right eigenvector overlap in pseudospectral instability (Dubach et al., 6 Jan 2026). Additionally, extensions to higher-rank perturbations, functional calculus under C-normality, and connections with PT-symmetry remain active topics (Amara et al., 2022).

Non-Hermitian rank-one perturbations, despite their apparent simplicity, reveal deep universal structures in spectral theory and continue to drive discovery across mathematics, physics, and data science.