Finite-Rank Normal Deformations

• Finite-rank normal deformations are perturbations of normal operators by finite-rank spikes, leading to the emergence of outlier eigenvalues and phase transitions.
• They involve a master equation that governs the location of outliers and predict Gaussian fluctuations when the spike exceeds a critical threshold.
• These deformations unify concepts in random matrix theory, operator theory, and high-dimensional statistics, with applications in signal detection and invariant subspace analysis.

A finite-rank normal deformation refers to the perturbation of a normal (possibly random) operator by another operator of finite rank, typically normal or diagonalizable, and encompasses both infinite- and large-finite-dimensional settings as well as both Hermitian and non-Hermitian ensembles. The paradigm unites the spectral and eigenvector phenomena associated with the appearance of so-called outlier eigenvalues, phase transitions, central limit theorems for spectral statistics, and the generation of nontrivial invariant (and hyperinvariant) subspaces. These mechanisms, which generalize the Baik–Ben Arous–Péché (BBP) transition, have broad impact in random matrix theory, operator theory, and high-dimensional statistics (Bousseyroux et al., 15 Jan 2026, &&&1&&&, Gallardo-Gutiérrez et al., 2024).

1. Definition and Basic Models

Let $A$ be a large normal matrix, or more generally a bounded normal operator $N$, with an absolutely continuous (or decomposable) spectral measure. A finite-rank normal deformation is an operator of the form $M = A + T$, where $T$ is normal and finite-rank, and is typically parameterized by its spectral data ("spike strengths" and "spike directions"). In random matrix settings, $A$ often possesses rotational or unitary invariance, leading to deterministic limiting spectral distributions (e.g., semicircle or circular laws).

A canonical example is $T = \sum_{i=1}^r \theta_i u_i u_i^*$, where $\{u_i\}$ forms an orthonormal family and $\theta_i$ are the fixed deformation (“spike”) amplitudes; then $T$ is normal and diagonalizable in a suitably extended orthonormal basis. In infinite-dimensional operator theory, finite-rank deformations take the form $T = D_\Lambda + F$ with $D_\Lambda$ diagonal and $F$ finite-rank (Gallardo-Gutiérrez et al., 2024, Putinar et al., 2019).

2. Outlier Eigenvalues, Phase Transitions, and Master Equations

Under broad conditions (rotational invariance, compact spectrum, resolvent control), the spectrum of $M = A + T$ exhibits both a bulk (persisting from $A$) and possible outliers. The location of outliers is governed by a universal deterministic equation known as the outlier equation or master equation. For $z \notin \mathrm{supp}(\mu_A)$, the outlier equation is

$m_A(z) = \frac{1}{\theta_i}$

where $$m_A(z) = \lim_{N \rightarrow \infty} \frac{1}{N} \mathrm{Tr}\,(z-A)^{-1}$$ is the limiting Stieltjes or Cauchy transform (Bousseyroux et al., 15 Jan 2026, Benaych-Georges et al., 2010, Pizzo et al., 2011).

The existence of an outlier — that is, a solution $z$ outside the bulk spectrum $S = \mathrm{supp}(\mu_A)$ — arises if and only if the spike magnitude $|\theta_i|$ exceeds a threshold $\theta_c$, often determined by the spectral edge. For rotationally invariant random matrices with a circular bulk of radius $R$, the condition is $|\theta_i| > R$ (Bousseyroux et al., 15 Jan 2026, Renfrew et al., 2012). These results specialize to the BBP threshold in the case of Hermitian ensembles: $$|\theta_i| > \sigma \Longrightarrow \text{outlier at } z_\mathrm{out} = \theta_i + \frac{\sigma^2}{\theta_i}$$ for a semicircle law of variance $\sigma^2$ (Pizzo et al., 2011, Benaych-Georges et al., 2010).

In high-dimensional statistical models, analogous phase transitions in spectral outliers occur, e.g., the detection threshold for canonical correlations in CCA (Bao et al., 2014) is

$$r_c = \frac{c_1c_2 + \sqrt{c_1c_2(1-c_1)(1-c_2)}}{(1-c_1)(1-c_2)+\sqrt{c_1c_2(1-c_1)(1-c_2)}}$$

for aspect ratios $p/n \to c_1$, $q/n \to c_2$.

3. Fluctuation Theory for Outliers

When a spike is strictly supercritical, the associated outlier eigenvalue detaches from the spectral bulk and exhibits Gaussian fluctuations at scale $O(N^{-1/2})$ in the large-$N$ limit. For spiked random matrices,

$$\sqrt{N} \left( \lambda_{\max}^{(i)} - z_{\mathrm{out},i} \right) \xrightarrow{d} \mathcal N(0, \sigma_i^2)$$

with $\sigma_i^2$ specified by derivatives of the matrix R-transform associated to $A$ or the limiting law, for both Hermitian and non-Hermitian settings (Bousseyroux et al., 15 Jan 2026, Renfrew et al., 2012, Benaych-Georges et al., 2010).

In Hermitian Wigner deformations, the variance is determined through the Stieltjes transform and its derivative: $$\tau_j^2 = \frac{\sigma^4(1-\sigma^2 m_{\mathrm{sc}}'(\theta_j))}{(d_j^2-\sigma^2 m_{\mathrm{sc}}(\theta_j)^2)^2}$$ (Renfrew et al., 2012).

If the spike is subcritical, the would-be outlier "sticks" to the spectral edge and exhibits Tracy–Widom or similar edge scaling, with no Gaussian separation (Benaych-Georges et al., 2010, Pizzo et al., 2011, Bousseyroux et al., 15 Jan 2026).

4. Eigenvector Localization and Overlap

In finite-rank deformations, not only the eigenvalues but also the eigenvectors display deterministic limiting behavior. Specifically, for each outlier, the squared overlap between the corresponding eigenvector $\phi_i$ and the spike direction $u_i$ is given by: $$|\langle u_i, \phi_i \rangle|^2 \longrightarrow 1 - \frac{\partial_\alpha \mathcal R_{1, A}(0, 1/\theta_i)}{|\theta_i|^2}$$ in the general framework (Bousseyroux et al., 15 Jan 2026). This reduces to the BBP formula in Hermitian cases: $$f(\theta_i) = 1 - \frac{R'(1/\theta_i)}{\theta_i^2}$$ For Wigner-like matrices, delocalized spike directions (coordinatewise $O(N^{-1/2})$) guarantee this deterministic overlap and ensure the universal Gaussian fluctuation property for the outliers (Renfrew et al., 2012, Benaych-Georges et al., 2010).

5. Functional Model and Operator-Theoretic Framework

For a bounded normal operator $N$ on a separable Hilbert space and finite-rank perturbation $K$, the spectrum of $T = N + K$ is explicitly encoded by the holomorphic “characteristic function” $$\Theta(z) = I_m + \int\frac{v(t) u(t)^*}{t-z}\,d\mu(t)$$, or more generally by the perturbation determinant $\theta(z) = \det \Theta(z)$ (Putinar et al., 2019). The location and multiplicity of eigenvalues outside the essential spectrum are determined by zeros of $\theta(z)$, with corresponding algebraic multiplicity.

The functional model realizes $T$ as a multiplication operator on a quotient of a suitable Hardy space modulo the range of $\Theta(z)$, enabling explicit construction of Riesz projections, invariant subspaces, and criteria for decomposability. Geometric conditions on the spectral measure (e.g., dissectibility) and smoothness of $K$ (boundedness, H\"older continuity) facilitate spectral analysis and invariant subspace results (Putinar et al., 2019).

6. Invariant and Hyperinvariant Subspaces

For finite-rank perturbations of diagonal (or diagonalizable normal) operators in infinite dimensions, recent results provide sharp necessary and sufficient summability criteria for the existence of nontrivial invariant and hyperinvariant subspaces. For rank-one perturbations $T = D_\Lambda + u \otimes v$ on $\ell^2$ with Fourier coefficients $\{\alpha_n\}, \{\beta_n\}$ relative to the eigenbasis, the log-summability condition

$$\sum_{n=1}^\infty \left( |\alpha_n|^2 \log\frac{1}{|\alpha_n|} + |\beta_n|^2 \log\frac{1}{|\beta_n|} \right) < \infty$$

guarantees the existence of such subspaces except for degenerate (scalar) $T$ (Gallardo-Gutiérrez et al., 2024). Analogous results extend to higher ranks, and previous weaker criteria (e.g., $\ell^1$, $\ell^{2/3}$) are subsumed in this framework.

7. Applications, Special Cases, and Statistical Models

Finite-rank normal deformations unify a range of phenomena:

• In spiked random matrix models (Wigner, Ginibre, Wishart, MANOVA, "single-ring"), they provide a universal description for both bulk and outlier eigenvalues, clarify phase transitions, and enable precise fluctuation and eigenvector analyses (Pizzo et al., 2011, Renfrew et al., 2012, Bousseyroux et al., 15 Jan 2026).
• In high-dimensional statistics and signal detection, the BBP-type threshold manifests in CCA, sample correlation, and covariance estimation models by determining the detectability and estimation rate for population spikes (Bao et al., 2014).
• In operator theory, they solve longstanding questions on invariant subspaces for normal plus finite-rank operators, with functional-analytic and geometric tools (Gallardo-Gutiérrez et al., 2024, Putinar et al., 2019).

Setting	Outlier Equation	Threshold Condition
Hermitian RMT	$m_A(z) = 1/\theta$	$|\theta| > \theta_c$
Ginibre/Single-Ring	$|z| > r_+$, $z = \theta$	$|\theta| > r_+$
Infinite-Dim Operators	$\theta(z) = 0$ (char. func.)	As in log-summability

These results establish a unified and precise mathematical architecture that explains spectral outliers, their fluctuations, and the transition from nontrivial randomness to deterministic limits as finite-rank deformations perturb bulk spectral structures (Bousseyroux et al., 15 Jan 2026, Benaych-Georges et al., 2010, Gallardo-Gutiérrez et al., 2024, Putinar et al., 2019).