BBP transition and the leading eigenvector of the spiked Wigner model with inhomogeneous noise

Abstract: The spiked Wigner ensemble is a prototypical model for high-dimensional inference. We study the spectral properties of an inhomogeneous rank-one spiked Wigner model in which the variance of each entry of the noise matrix is itself a random variable. In the high-dimensional limit, we derive exact equations for the spectral edges, the outlier eigenvalue, and the distribution of the components of the outlier eigenvector. These equations determine the BBP transition line that separates the gapped phase, where the signal is detectable, from the gapless phase. In the gapped regime, the distribution of the outlier eigenvector provides a natural estimator of the spike. We solve the equations for a noise matrix whose variances are generated from a truncated power-law distribution. In this case, the BBP transition line is non-monotonic, showing that an inhomogeneous noise can enhance signal detectability.

• The paper introduces a novel extension of the spiked Wigner model by incorporating inhomogeneous noise via random variance profiles, deriving exact equations for spectral edges and BBP transition thresholds.
• The paper employs resolvent and cavity methods to derive self-consistent equations for eigenvalue distributions and eigenvector statistics, quantifying recovery error using mean squared error metrics.
• The paper reveals non-monotonic detectability behavior, where variance heterogeneity alternately enhances or suppresses signal recovery depending on noise profile parameters.

Spectral Inference in Spiked Wigner Models with Inhomogeneous Noise

Introduction

The paper "BBP transition and the leading eigenvector of the spiked Wigner model with inhomogeneous noise" (2604.18523) systematically investigates the spectral properties and inference limits of the spiked Wigner ensemble generalized to allow inhomogeneous noise, where the variance profile of the noise matrix is itself random. This approach extends classical models by encoding heterogeneity akin to random graphs with arbitrary degree distributions directly into the noise matrix, offering analytic tractability in the high-dimensional limit.

The primary focus is on understanding the conditions for signal detectability via spectral methods—specifically, the Baik-Ben Arous-Péché (BBP) phase transition—when the structure of the noise matrix is governed by an arbitrary variance profile, including randomly drawn variances from power-law distributions. The authors derive exact equations for spectral edges, outlier eigenvalues, and the distribution of the outlier eigenvector components, elucidating how signal-to-noise ratio thresholds and recovery properties fundamentally depend on noise characteristics.

Model Definition and Motivation

The observation matrix $\boldsymbol{A}_n$ is modeled as a sum of a rank-one signal and an inhomogeneous random noise matrix:

$$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$

where $\boldsymbol{W}_n$ is a symmetric matrix with i.i.d. entries of zero mean and unit variance, and $\boldsymbol{S}_n$ encodes entry-wise variance via $S_{ij} = S_i S_j$, with $\{S_i\}$ drawn from some distribution $P_S$. The spike vector $\boldsymbol{X}_n$ is generated from a prior $P_X$, and $\gamma$ quantifies the signal-to-noise ratio. The structure makes entry variances heterogeneous and random, generalizing model assumptions beyond the homogeneous Wigner setting and aligning the ensemble with random graph adjacency matrices under high connectivity, where degree distributions inform variance profiles.

By relaxing the i.i.d. noise assumption, the model allows exploration of how variance disorder impacts the fundamental spectral phase transition controlling detectability, the BBP transition, with practical relevance to inference in structured data regimes such as network clustering, recommendation systems, and genomics.

Spectral Properties and Analytical Methods

The analysis is anchored in random matrix theory, leveraging the resolvent method and cavity approaches to derive self-consistent equations for the spectral density, bulk edges, outlier location, and eigenvector statistics for arbitrary spike and variance distributions. Key results center around exact distributional equations for the diagonal resolvent entries $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$0, yielding the spectral density $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$1:

$$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$2

leading to an implicit characterization of the spectral edge $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$3. The outlier eigenvalue $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$4, associated with detectability and governed by a rank-one perturbation, is given through an integral equation as a function of $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$5 and $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$6, generalized to arbitrary $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$7.

Additionally, the distribution $$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$8 of the outlier eigenvector components is derived, resulting in a weighted superposition of Gaussians determined by both spike and variance profiles:

$$\boldsymbol{A}_n = \frac{1}{\sqrt{n}}(\boldsymbol{S}_n^{\odot 1/2} \odot \boldsymbol{W}_n) + \frac{\gamma}{n} \boldsymbol{X}_n \boldsymbol{X}_n^T,$$9

where the mean and variance are explicit functions of $\boldsymbol{W}_n$0, spike moments, and $\boldsymbol{W}_n$1.

Figure 1: Spectrum of the homogeneous spiked Wigner model; an outlier eigenvalue (red dot) detaches for $\boldsymbol{W}_n$2, with critical edge $\boldsymbol{W}_n$3.

Effects of Inhomogeneous Noise: Truncated Power-Law Variance Profile

To concretely investigate the impact of noise inhomogeneity, the authors study variance profiles $\boldsymbol{W}_n$4 defined by truncated power-law distributions, parameterized by width $\boldsymbol{W}_n$5 and exponent $\boldsymbol{W}_n$6, encompassing regimes from highly concentrated to broadly fluctuating variances. This family interpolates between homogeneous (delta-like) and heavy-tailed structures, mirroring real-world network systems.

The spectral density, as computed numerically and analytically, reveals non-trivial behavior: the bulk support shrinks or stretches depending on $\boldsymbol{W}_n$7 and $\boldsymbol{W}_n$8, with sharp spectral singularities linked to localized eigenvector states at low variance sites. The analytic approach matches numerical diagonalizations, confirming theoretical predictions.

Figure 2: Empirical spectral density $\boldsymbol{W}_n$9 of the observation matrix for various power-law variance profiles, showing deviations from semicircular law and singularities at $\boldsymbol{S}_n$0 for broad, heavy-tailed distributions.

BBP Transition Line: Detectability Thresholds and Phase Behavior

A fundamental result is the characterization of the BBP transition line in terms of the critical signal-to-noise ratio $\boldsymbol{S}_n$1, marking the boundary between the gapped and gapless spectral regimes. In the gapped phase ($\boldsymbol{S}_n$2), detectability via the leading eigenvector is feasible; in the gapless phase, recovery is impossible via spectral methods.

Notably, the critical $\boldsymbol{S}_n$3 exhibits non-monotonic dependence on noise profile parameters: while increased variance heterogeneity can suppress detectability for $\boldsymbol{S}_n$4, with $\boldsymbol{S}_n$5 increasing as $\boldsymbol{S}_n$6 grows, it can also enhance detectability for $\boldsymbol{S}_n$7 near $\boldsymbol{S}_n$8, where many entries have vanishing variance. Thus, inhomogeneous noise is shown to sometimes reduce and sometimes boost signal recovery, contradicting intuitions from homogeneous ensembles.

Figure 3: Critical value $\boldsymbol{S}_n$9 of signal-to-noise ratio as a function of variance profile parameters, marking the BBP transition; non-monotonic behavior evident with certain parameter choices.

Figure 4: Shape parameter $S_{ij} = S_i S_j$0 marking the BBP transition boundary in the $S_{ij} = S_i S_j$1 plane, demonstrating how variance profile broadening can alternately suppress or enable signal detectability.

Leading Eigenvector Distribution and Recovery Quality

Exact characterization of the leading eigenvector distribution outside the bulk provides both a practical estimator of the spike and a tool for evaluating recovery error. The mean squared error (MSE) between the spike and the outlier eigenvector is derived, quantifying recovery quality:

$S_{ij} = S_i S_j$2

Results demonstrate that, as $S_{ij} = S_i S_j$3, MSE vanishes, while near the BBP threshold it diverges, confirming the transition in recoverability.

Figure 5: Distribution of eigenvector components outside the bulk, with pronounced deviations from Gaussianity for inhomogeneous noise.

Figure 6: Mean squared error between spike and outlier eigenvector as a function of signal-to-noise ratio, diverging at the critical threshold $S_{ij} = S_i S_j$4.

Implications and Outlook

The theoretical findings advance the understanding of inference in structured high-dimensional random matrix ensembles, bridging random graph theory and spectral analysis. The non-monotonic BBP transition lines imply that the detectability gap for spectral estimators is sensitive to fine-grained noise structure, a fact with both practical ramifications (for clustering, bioinformatics, recommendation systems) and theoretical significance.

On a methodological level, the analytic framework—cavity method combined with distributional equations for resolvent and eigenvector statistics—affords direct generalization to other structured matrix models, including spiked Wishart ensembles and sparse random matrices, and opens avenues for studying inference in non-Hermitian settings.

Conclusion

The paper rigorously extends the spiked Wigner ensemble paradigm to inhomogeneous noise governed by random variance profiles, deriving exact spectral and eigenvector statistics in the high-dimensional limit. The main results include explicit equations for the spectral density, spectral edges, outlier eigenvalue, BBP transition threshold, and eigenvector component distribution for arbitrary spike and noise distributions. The analysis reveals rich, non-monotonic phase behavior in signal detectability, with variance heterogeneity alternately hindering or enhancing recovery depending on parameter regimes. These insights both broaden the theoretical landscape of spectral inference in random matrices and inform practical considerations in high-dimensional data analysis, signaling directions for future work on structured matrix ensembles and spectral inference algorithms.