Numerical Performance Of A Tensor Music Algorithm Based On Hosvd For A Mixture Of Polarized Sources

Mélanie Boizard, Guillaume Ginolhac, Frédéric Pascal, Sébastian Miron, Philippe Forster2013 · DOI

TL;DR

This paper introduces a novel Tensor MUSIC algorithm that leverages the Higher Order Singular Value Decomposition (HOSVD) to improve the estimation of Direction of Arrival (DOA) and polarimetric parameters for multiple polarized sources. By preserving the multidimensional structure of the data, this tensor-based approach significantly outperforms traditional vector-based MUSIC methods, especially in challenging scenarios.

What Problem Does This Paper Address?

Modern signal processing applications, such as polarimetric radar and seismic source localization, increasingly deal with multidimensional data. Traditional algorithms often flatten this data into matrices, losing valuable structural information. This paper tackles the challenge of accurately estimating parameters like the Direction of Arrival (DOA) and polarization characteristics of multiple sources when the received signals possess this inherent multidimensionality. Existing methods that attempt to handle multidimensional data either struggle with computational complexity, convergence issues, or fail to fully exploit the data's structure. The authors aim to bridge this gap by developing a tensor-based MUSIC algorithm that preserves and utilizes this multidimensional information for enhanced performance.

Methods

The core of this work is the development of a Tensor MUSIC (TMUSIC) algorithm. Unlike traditional MUSIC which operates on covariance matrices, TMUSIC works with a covariance tensor. The key innovation lies in using the Higher Order Singular Value Decomposition (HOSVD) to decompose this covariance tensor. HOSVD is chosen for its ability to preserve the orthogonality properties of the tensor, which is crucial for separating signal and noise subspaces in a multidimensional setting.

The algorithm proceeds in three main steps:

  1. HOSVD Decomposition: The covariance tensor is decomposed using HOSVD into a core tensor and a set of orthogonal matrices for each mode (dimension).
  2. Subspace Estimation: The signal and noise subspaces are estimated from the HOSVD components, specifically by truncating the orthogonal matrices corresponding to the signal subspace.
  3. Parameter Estimation: The DOA and polarimetric parameters are then estimated by maximizing a specific criterion derived from the steering matrix and the estimated noise subspace.

This approach is contrasted with existing methods:

  • Vector MUSIC (VMUSIC): This is the baseline, where multidimensional data is flattened into a matrix, and standard MUSIC is applied. It does not explicitly handle polarization.
  • Long Vector MUSIC (LVMUSIC): This method is mentioned as an improvement over VMUSIC but is noted to not fully exploit the multidimensional structure of the covariance tensor, making it equivalent to the vector approach in some aspects.
  • CP-based Tensor MUSIC: While CP decomposition is another tensor decomposition method, the authors note potential convergence issues and identifiability problems, opting for HOSVD instead.

Datasets

Not discussed in the paper. The paper relies on simulated data for its numerical experiments.

Key Insights and Ideas

  • Preserving Multidimensional Structure is Key: The paper strongly advocates for and demonstrates the benefit of maintaining the multidimensional structure of the signal data, rather than flattening it into a matrix. This allows for a more accurate representation and exploitation of the underlying physics.
  • HOSVD's Suitability for Subspace Estimation: The choice of HOSVD is justified by its orthogonality properties, which are well-suited for estimating signal and noise subspaces in a multidimensional context, avoiding some pitfalls of other tensor decompositions like CP.
  • Polarization Information Enhances Resolution: When polarimetric properties of sources are different, incorporating this information through the tensor approach significantly improves the resolution and accuracy of DOA estimation, especially at lower Signal-to-Noise Ratios (SNRs).
  • TMUSIC Outperforms Vector Approaches: The simulations consistently show that TMUSIC provides superior performance in terms of Mean Squared Error (MSE) for parameter estimation compared to both standard VMUSIC and LVMUSIC.

Key Results

  • Improved Parameter Estimation: TMUSIC demonstrates a lower Mean Squared Error (MSE) for estimating all five parameters (DOA and polarimetric parameters) compared to VMUSIC, particularly in scenarios with a single source.
  • Enhanced DOA Resolution: For two closely spaced sources, TMUSIC and LVMUSIC show significantly better pseudo-spectra and DOA separation compared to VMUSIC, especially when the sources have different polarimetric properties.
  • Polarization Matters (When Different): When sources have distinct polarimetric characteristics, LVMUSIC shows improved resolution over VMUSIC. However, TMUSIC consistently outperforms all other methods, regardless of whether the polarimetric properties are the same or different.
  • Robustness to Rank Selection: The performance of TMUSIC is shown to be relatively insensitive to the choice of the rank ($r_2$) for the polarimetric dimension, with similar results for $r_2=1$ and $r_2=2$.

Limitations and Future Work

  • Computational Complexity: While TMUSIC offers performance gains, the paper acknowledges that estimating all five parameters requires a significant amount of computational data, which limited the number of realizations for some simulations.
  • Adaptation of CP Algorithm: The authors suggest adapting the CP algorithm and comparing it with their HOSVD-based TMUSIC, along with deriving the Cramér-Rao bound for their model.
  • Theoretical Performance Analysis: Future work will involve theoretically analyzing the performance of TMUSIC using a perturbation method.
  • Further Study of Rank Selection: An exhaustive study on the selection of the rank ($r_2$) for the polarimetric dimension and its link to polarimetric properties is planned for a future publication.

Summary updated

Abstract

Publication in the conference proceedings of EUSIPCO, Marrakech, Morocco, 2013
BibTeX
@misc{boizard2013numerical,
  author = {Mélanie Boizard and Guillaume Ginolhac and Frédéric Pascal and Sébastian Miron and Philippe Forster},
  title = {{Numerical Performance Of A Tensor Music Algorithm Based On Hosvd For A Mixture Of Polarized Sources}},
  year = {2013},
  doi = {10.5281/zenodo.43466},
}
Figure