A Probabilistic Spectral Analysis of Multivariate Real-Valued Nonstationary Signals

Bruno Scalzo, Ljubiša Stanković, Danilo P. MandicarXiv 2020 · DOI · arXiv

Abstract

A class of multivariate spectral representations for real-valued nonstationary random variables is introduced, which is characterised by a general complex Gaussian distribution. In this way, the temporal signal properties—harmonicity, wide-sense stationarity, and cyclostationarity—are designated respectively by the mean, Hermitian variance, and pseudo-variance of the associated time-frequency representation (TFR). For rigour, the estimators of the TFR distribution parameters are derived within a maximum likelihood framework and are shown to be statistically consistent, owing to the statistical identifiability of the proposed distribution parametrization. By virtue of the assumed probabilistic model, a generalised likelihood ratio test (GLRT) for nonstationarity detection is also proposed. Intuitive examples demonstrate the utility of the derived probabilistic framework for spectral analysis in low-SNR environments.

Introduction

Traditional spectral analysis is often built upon two foundational assumptions that frequently fail in practical applications: stationarity and magnitude-only modeling. Many real-world signals, such as those found in radar, biomedical engineering, and speech processing, exhibit statistical properties that evolve over time. Furthermore, conventional methods often discard phase information, treating it as uniform noise. This "maximum entropy" approach ignores vital structural data that can distinguish between deterministic components and stochastic noise.

This paper addresses these limitations by introducing a rigorous probabilistic framework for multivariate nonstationary signals. By treating spectral representations as complex Gaussian random variables, the authors provide a unified mathematical language to detect and distinguish between different signal behaviors—specifically harmonics and cyclostationarity—even in environments with high noise levels.

The Probabilistic Framework

The core innovation of this work is the shift from time-domain analysis to a multivariate Time-Frequency Representation (TFR) space modeled by augmented complex statistics.

Augmented Complex Statistics

Instead of relying solely on the standard Hermitian variance ($E[|X|^2]$), which characterizes Wide-Sense Stationary (WSS) components, the framework utilizes two additional parameters:

  1. The Spectral Mean ($E[X]$): This parameter captures deterministic harmonic components. Unlike legacy power spectra that lump harmonics and noise together, the spectral mean isolates these steady signals.
  2. The Pseudo-Variance ($E[X^2]$): This captures "noncircularity" or phase-coupling, which serves as a distinct signature of cyclostationarity.
Representation of a harmonic signal and its corresponding spectral mean, illustrating how deterministic components are isolated in the TFR space.

By modeling the "ellipticity" of the distribution in the complex plane via the pseudo-variance, the framework captures time-symmetry and phase-coupling that magnitude-only models miss entirely.

Estimation and Detection

The authors derive statistically consistent Maximum Likelihood (ML) estimators for the spectral mean and covariance matrices. Using this probabilistic model, they develop a suite of Generalized Likelihood Ratio Tests (GLRT) to detect specific signal properties. For instance, testing for a non-zero spectral mean identifies harmonics, while testing for non-zero off-diagonal elements in the spectral covariance identifies cyclostationarity.

Experimental Results and Validation

The framework was validated using synthetic multivariate Gaussian processes, including signals with 10 sensors containing harmonics and varying levels of cyclostationary noise. Monte Carlo simulations with 1,000 independent realizations confirmed the statistical consistency of the estimators: as the sample size increases, the estimation variance approaches zero.

Superior Performance in Low SNR

One of the most significant findings is the robustness of the GLRT for harmonics, which remains highly effective at Signal-to-Noise Ratios (SNR) as low as -20 dB. This performance is attributed to the "centered" moments approach, which prevents the entanglement of harmonic signals and stochastic noise.

Component Identification

The utility of the framework is most clearly seen in its ability to separate a complex signal into its constituent parts. While legacy absolute power spectra often present a single, indistinguishable blur when multiple signal types are present, the proposed method can perfectly isolate harmonic, WSS, and cyclostationary components.

The proposed framework allows for the clear separation of a nonstationary signal into its constituent harmonic, WSS, and cyclostationary parts, as seen in the comparison between centered moments and legacy absolute moments.

Discussion and Conclusion

The presented multivariate probabilistic framework offers a powerful tool for the analysis of nonstationary signals. By leveraging augmented complex statistics and the ergodicity of TFRs in the first and second moments, researchers can perform deep statistical analysis even with a single signal realization.

While the current framework assumes a complex Gaussian distribution and involves computational costs associated with iterative matrix inversions, it sets a new standard for spectral analysis. Future work may involve extending this model to non-Gaussian "heavy-tailed" noise environments and applying it to real-world datasets such as EEG or mechanical vibration data.

Summary updated

Abstract

A class of multivariate spectral representations for real-valued nonstationary random variables is introduced, which is characterised by a general complex Gaussian distribution. In this way, the temporal signal properties -- harmonicity, wide-sense stationarity and cyclostationarity -- are designated respectively by the mean, Hermitian variance and pseudo-variance of the associated time-frequency representation (TFR). For rigour, the estimators of the TFR distribution parameters are derived within a maximum likelihood framework and are shown to be statistically consistent, owing to the statistical identifiability of the proposed distribution parametrization. By virtue of the assumed probabilistic model, a generalised likelihood ratio test (GLRT) for nonstationarity detection is also proposed. Intuitive examples demonstrate the utility of the derived probabilistic framework for spectral analysis in low-SNR environments.
BibTeX
@misc{scalzo2020probabilistic,
  author = {Bruno Scalzo and Ljubiša Stanković and Danilo P. Mandic},
  title = {{A Probabilistic Spectral Analysis of Multivariate Real-Valued Nonstationary Signals}},
  year = {2020},
  eprint = {2007.13855},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.13855},
}
Figure