Delocalized Spiked Matrix Models

Updated 5 January 2026

Delocalized spiked matrix models are defined by low-rank perturbations added to noise matrices, where the signal vector is uniformly spread across entries.
They reveal precise entrywise Gaussian fluctuations beyond a critical spike threshold, highlighting phase transitions and universal behavior in high dimensions.
These models provide rigorous asymptotic predictions that enhance spectral algorithms for inference in statistics, physics, and machine learning.

Entrywise eigenvector fluctuations refer to the asymptotic distribution and universality properties of the individual coordinates of eigenvectors associated with spiked (low-rank perturbations) random matrix models, particularly in the high-dimensional regime and in the presence of delocalized signal vectors. This theme is central in modern random matrix theory, with significant implications for high-dimensional statistics, statistical physics, and spectral algorithms in inference and learning. The precise behavior of entrywise fluctuations, especially under varying localization regimes and noise laws, encodes both signal detectability and the universality class of the eigenvector entries.

1. Delocalized Spikes and Model Setup

Spiked random matrix models typically take the form

$H = \theta v v^* + W$

where $H$ is an $n \times n$ Hermitian (or real symmetric) matrix, $\theta$ quantifies the spike (signal) strength, $v$ is a unit-norm signal vector (“spike”), and $W$ is a noise matrix—often a generalized Wigner matrix with independent, centered entries and variances typically of order $1/n$ (Chen et al., 12 Dec 2025).

In the delocalized regime—defined by the condition

$\|v\|_\infty \le C_v n^{-1/2 + \alpha_v}, \qquad \alpha_v \in (0,1/20)$

—no entry of $v$ carries a nonnegligible fraction of its mass. This regime is standard in high-dimensional PCA with i.i.d. or spherical priors on $v$ ; it ensures isotropy and typicality of fluctuations (Chen et al., 12 Dec 2025, Perry et al., 2018, Miolane, 2018).

The spiked covariance, separable covariance, and multiplicative random matrix models generalize this framework to cases with several spikes, nontrivial base distributions, or multiplicative (rather than additive) perturbations (Ding et al., 2019, Ding et al., 2023).

2. Statistical Phase Transitions and Eigenvector Alignment

A core aspect of entrywise eigenvector fluctuations is the presence of spectral phase transitions, particularly the BBP transition, which demarcates a regime where the principal eigenvector becomes informative about the spike.

In the spiked Wigner model, for $H$ 0 the top eigenvector of $H$ 1 is asymptotically orthogonal to $H$ 2; for $H$ 3 (the supercritical regime), an eigenvalue outlier emerges and the corresponding eigenvector exhibits nontrivial overlap with $H$ 4:

$H$ 5

Here, $H$ 6 denotes the empirical principal eigenvector (Chen et al., 12 Dec 2025, Miolane, 2018).

In multiplicative or separable covariance settings, spikes above model-dependent thresholds (depending on subordination functions or edge spectral locations) generate similar outliers and eigenvector “cone concentration” onto the planted directions (Ding et al., 2019, Ding et al., 2023).
When the spike is sufficiently strong (supercritical), precise asymptotics for the entrywise behavior of $H$ 7 are accessible and universal (Chen et al., 12 Dec 2025).

3. Universality of Entrywise Fluctuations

In the supercritical and delocalized regime, the distribution of individual entries of the top eigenvector $H$ 8 is universal, i.e., depends only on the first two moments of the noise distribution, not on its detailed form.

If $H$ 9 is a generalized Wigner, and $n \times n$ 0 is delocalized, then for each $n \times n$ 1:

$n \times n$ 2

with $n \times n$ 3 and $n \times n$ 4 (Chen et al., 12 Dec 2025).

Universality theorem (Theorem 2.4 in (Chen et al., 12 Dec 2025)): If two generalized Wigner matrices $n \times n$ 5, $n \times n$ 6 have matching second moments, then all entrywise statistics of their respective spiked eigenvectors are asymptotically equivalent up to $n \times n$ 7, provided the spike is delocalized and $n \times n$ 8.
For localized spikes (e.g., $n \times n$ 9), the fluctuations are model-dependent and not universal, as originally exhibited by Capitaine–Donati-Martin (2018).
In the presence of non-Gaussian noise, the same entrywise Gaussian fluctuation holds as long as the variance profile is sufficiently regular and the delocalization condition is met (Chen et al., 12 Dec 2025, Perry et al., 2018).

4. Fluctuation Expansions and Asymptotic Theory

A full asymptotic theory involves first and higher-order expansions of both eigenvalues and eigenvectors for general noise and signal scaling (Fan et al., 2019):

Under mild regularity and for diverging spikes, let $\theta$ 0 be an arbitrary unit vector. Then:

$\theta$ 1

with the leading term $\theta$ 2 given by explicit resolvent expansions. For $\theta$ 3 aligned with the spike, second-order expansions are required:

$\theta$ 4

(Fan et al., 2019). These results generalize to bilinear forms and projections onto subspaces, enabling CLTs for general entrywise linear maps of the eigenvectors.

For non-outlier eigenvectors (those not associated with spikes above the threshold), one obtains delocalization bounds:

$\theta$ 5

uniformly over coordinates, as in the bulk regime of Wigner or separable covariance models (Roy et al., 3 Jul 2025, Ding et al., 2023, Ding et al., 2019).

5. Nonlinear Models and Entrywise Transformations

Entrywise eigenvector fluctuations are altered in nonlinear spiked models and under entrywise pre-transformations:

In non-linear Wigner spiked models, the critical scaling for the signal-to-noise ratio is dictated by the first nonzero generalized information coefficient ( $\theta$ 6) of the nonlinearity. The effective spike becomes $\theta$ 7 (entrywise $\theta$ 8th power), and outlier and eigenvector behavior track this transformed spike (Guionnet et al., 2023).
Under non-Gaussian noise, spectral PCA can become sub-optimal for detection, but pre-transforming the entries with the Fisher-score function $\theta$ 9 restores universality and optimal entrywise fluctuations (Perry et al., 2018).
In these settings, the entrywise law for the leading eigenvector realigns with the transformed signal, and universality can often be shown for statistics of the properly normalized entries (Chen et al., 12 Dec 2025, Guionnet et al., 2023).

6. Implications for Statistical Inference and Algorithms

Universal entrywise fluctuations underpin the analysis of spectral algorithms followed by nonlinear or discrete rounding, yield single-letter formulae for estimation losses, and determine statistical limits under delocalized signal priors (Chen et al., 12 Dec 2025, Miolane, 2018, Perry et al., 2018).

For spectral methods in angular synchronization, stochastic block models, and related group-estimation settings, averaging entrywise loss functions over the limiting eigenvector law gives tight predictions for post-processing (e.g., MSE or misclassification rate after rounding) (Chen et al., 12 Dec 2025).
In the supercritical regime, the limiting spectral estimator's error can be computed as an explicit integral involving the entrywise limiting law:

$v$ 0

with expectations over the asymptotic fluctuations of normalized eigenvector entries (Chen et al., 12 Dec 2025).

The detection and estimation thresholds for spikes coincide for uniformly delocalized priors, simplifying phase transition characterizations (Perry et al., 2018, Miolane, 2018), whereas for localized signals or disproportionate noise, non-universal effects dominate.

7. Extensions, Limitations, and Regimes of Delocalization

Models with infinitely many accumulating spikes (growing with matrix size) do not yield persistent outlier eigenvalues or eigenvector localization; all eigenvectors become fully delocalized, reflected in vanishing overlaps with any given spike direction (Thompson, 2019).
In separable and multiplicative models, outlier eigenvectors exhibit cone concentration onto the population spike direction with explicit cone angle and optimal convergence, whereas the bulk eigenvectors remain delocalized (order $v$ 1 for each coordinate) (Ding et al., 2019, Ding et al., 2023).
At the BBP critical threshold (spike strength near the phase transition), eigenvectors lose alignment and become delocalized at the $v$ 2 scale, with corresponding Tracy–Widom fluctuation behavior for outlier locations (Ding et al., 2023).

Table: Regimes and Entrywise Fluctuation Behavior

Spike Regime	Entrywise Fluctuations	Universality
Delocalized, $v$ 3 threshold	Gaussian CLT: $v$ 4	Yes (Chen et al., 12 Dec 2025)
Delocalized, $v$ 5 threshold	Isotropic, $v$ 6	Yes
Localized spike	Non-Gaussian, model-dependent	No
Growing spikes	Delocalized, all overlaps vanish	Yes (delocalized)

Here, universality means independence from higher moments of $v$ 7, contingent on delocalization.

In summary, the entrywise fluctuations of eigenvectors in spiked random matrix models, under delocalized signal regimes and sufficiently regular noise, exhibit robust universality: the limiting distribution is Gaussian and controlled by the rotation-invariant structure and first two moments of the underlying noise. This underpins the performance and precise asymptotic analysis of spectral methods across a wide array of high-dimensional inference problems (Chen et al., 12 Dec 2025, Perry et al., 2018, Miolane, 2018, Fan et al., 2019, Ding et al., 2019, Ding et al., 2023, Thompson, 2019, Guionnet et al., 2023).