New model order selection in large dimension regime for complex elliptically symmetric noise

Eugénie Terreaux, Jean‐Philippe Ovarlez, Frédéric Pascal2017 · DOI

TL;DR

This paper introduces a novel method for determining the number of underlying signals (model order selection) in complex data, particularly when the noise is not well-behaved (complex elliptically symmetric) and the data dimensions are large. Their approach combines a robust covariance matrix estimation technique with Random Matrix Theory to accurately distinguish signals from noise, outperforming traditional methods like AIC and MDL.

What Problem Does This Paper Address?

In many signal processing applications, such as source localization or subspace estimation, a crucial step is to determine how many distinct signals are present within noisy observations. This is known as model order selection. Traditional methods often assume that the noise is simple (e.g., white Gaussian) and that the data dimensions are manageable. However, real-world data, especially from fields like hyperspectral imaging, often exhibit more complex noise characteristics (like Complex Elliptically Symmetric or CES noise) and can have very high dimensions (many sensors or features relative to the number of observations). When these assumptions are violated, standard model order selection techniques like the Akaike Information Criterion (AIC) and Minimum Description Length (MDL) can perform poorly, leading to incorrect estimations of the number of signals. This paper tackles this challenge by developing a method that is robust to CES noise and effective in the large-dimensional regime.

Methods

The proposed method consists of two main steps:

  1. Robust Covariance Matrix Estimation: The authors first address the issue of correlated and non-Gaussian noise by proposing a consistent estimator for the background covariance matrix. This involves estimating the Toeplitz structure of the covariance matrix, which is a common characteristic of correlated noise. They achieve this by averaging elements along diagonals of the sample covariance matrix, creating a "toeplitzified" estimator. This step is crucial for "whitening" the data, meaning they transform the data so that the noise appears more like simple white noise.
  2. Model Order Selection using Random Matrix Theory (RMT): After whitening the data, the authors leverage the properties of Random Matrix Theory. RMT provides tools to analyze the distribution of eigenvalues of large matrices. They apply a Maronna's M-estimator to the whitened observations to obtain a scatter matrix. The key insight is that under the null hypothesis (no signals present), the eigenvalues of this scatter matrix follow a predictable distribution (related to the Marchenko-Pastur law). When signals are present, they tend to manifest as eigenvalues that deviate significantly from this expected distribution. By setting a threshold on these eigenvalues, they can effectively distinguish between signal-induced eigenvalues and noise-induced eigenvalues, thereby estimating the number of signals.

Datasets

The paper evaluates its proposed method using:

  • Simulated CES noise: The authors generate synthetic data with correlated CES noise (specifically, Student-t distributed noise) to test their method under controlled conditions.
  • Real hyperspectral images: The method is also validated on publicly available hyperspectral datasets: Indian Pines, SalinasA (from AVIRIS), and PaviaU (from ROSIS). These datasets are relevant because hyperspectral imaging often involves high-dimensional data with complex noise characteristics.

Key Insights and Ideas

  • CES noise is a significant challenge: The paper highlights that standard model order selection methods fail when dealing with CES noise, underscoring the need for more robust techniques.
  • Toeplitz structure estimation is key for whitening: By estimating and exploiting the Toeplitz structure of the covariance matrix, the authors can effectively whiten the data, making subsequent signal detection more reliable.
  • RMT is powerful for large-dimensional data: Random Matrix Theory provides the theoretical framework to analyze eigenvalue distributions in high-dimensional settings, enabling a principled way to separate signal from noise.
  • Maronna's M-estimators offer robustness: Using M-estimators provides robustness to the non-Gaussian nature of CES noise, further enhancing the method's performance.
  • The proposed method outperforms AIC and MDL: The simulations and real-data experiments demonstrate a clear superiority of their technique over established methods, especially in challenging scenarios.

Key Results

  • The proposed covariance matrix estimator is shown to be consistent, meaning it converges to the true covariance matrix as the data size increases, outperforming the standard Sample Covariance Matrix (SCM) in the presence of correlated CES noise.
  • In simulations with CES noise, the proposed method accurately estimates the number of sources across a wide range of Signal-to-Noise Ratios (SNRs), even at low SNRs where AIC and other methods overestimate the number of sources.
  • When applied to real hyperspectral images, the proposed method consistently yields better estimates of the number of sources compared to the non-whitened fixed-point estimator and the AIC method, as summarized in Table I.

Limitations and Future Work

  • Computational complexity: While not explicitly detailed as a limitation, the process of estimating the Toeplitz structure and applying M-estimators can be computationally intensive, especially for very large dimensions.
  • Generalizability of M-estimator: The paper mentions that "any M-estimator can then be used," suggesting that the choice of M-estimator might influence performance, and further investigation into optimal M-estimator selection for specific CES distributions could be beneficial.
  • Future work: The authors suggest that their general technique can be applied to a broader range of model order selection problems beyond source detection, including radar clutter rank estimation, source localization, and various hyperspectral imaging techniques like anomaly detection and unmixing. They also imply that further analysis of the theoretical properties of their estimator in different regimes could be explored.

Summary updated

Abstract

This paper presents a new model order selection technique for signal processing applications related to source localization or subspace orthogonal projection techniques in large dimensional regime (Random Matrix Theory) when the noise environment is Complex Elliptically Symmetric (CES) distributed, with unknown scatter matrix. The proposed method consists first in estimating the Toeplitz structure of the background covariance matrix. In a second step, after a whitening process, the eigenvalues distribution of any Maronna's M-estimators is exploited, leading to the order selection. Simulations made on different kinds of CES noise as well as analysis of real hyperspectral images demonstrate the superiority of the proposed technique compared to those of Akaike Information Criterion and the Minimum Description Length.
BibTeX
@misc{terreaux2017new,
  author = {Eugénie Terreaux and Jean‐Philippe Ovarlez and Frédéric Pascal},
  title = {{New model order selection in large dimension regime for complex elliptically symmetric noise}},
  year = {2017},
  doi = {10.23919/eusipco.2017.8081376},
}
Figure