Riemannian Covariance Fitting for Direction-of-Arrival Estimation

Joseph S. Picard, Amitay Bar, Ronen TalmonarXiv 2024 · DOI · arXiv

Abstract

Covariance fitting (CF) is a comprehensive approach for direction of arrival (DoA) estimation. Standard practice uses Euclidean criteria, disregarding the intrinsic Hermitian positive-definite (HPD) geometry of spatial covariance matrices. This paper presents a study of Riemannian metrics, specifically the Log-Euclidean (LE) metric, for CF. We propose a new practical LE beamformer, derive its spatial characteristics analytically, and demonstrate its significant advantages in resolution, sidelobe attenuation, and robustness under challenging conditions like low SNR and small sample sizes.

Introduction: The Geometry of Covariance

In array signal processing, Direction-of-Arrival (DoA) estimation identifies the source directions of incoming signals. Classical algorithms like the Conventional Beamformer (CB) and Minimum Variance Distortionless Response (MVDR) can be framed as "covariance fitting" problems: finding the direction where a model best fits the observed sample covariance matrix $\hat{\mathbf{R}}$.

Traditionally, this fit is measured using Euclidean distance (the Frobenius norm). However, covariance matrices are Hermitian Positive-Definite (HPD) and reside on a curved Riemannian manifold. Treating them as flat Euclidean objects ignores their intrinsic geometry, forcing a trade-off between the robustness of CB and the high resolution of MVDR. MVDR, while sharp, often becomes unstable or produces spurious peaks when data samples are limited or noise is high.

The Log-Euclidean Solution

To overcome these limitations, we propose moving from Euclidean geometry to Riemannian geometry. While the Affine-Invariant (AI) metric is the theoretical "gold standard" for HPD matrices, its computational cost is high. We focus on the Log-Euclidean (LE) metric, which provides a tight, efficient approximation of the AI metric.

The resulting LE spatial spectrum has a simple closed-form expression: $$P_{LE}(\theta) = \exp(\mathbf{a}_\theta^H \log(\hat{\mathbf{R}}) \mathbf{a}_\theta) - 1$$ where $\mathbf{a}_\theta$ is the array steering vector. This formulation effectively unifies the strengths of classical methods.

Comparison of the spatial power spectrum for different beamformers. The LE beamformer provides a sharp peak at the true DoA without the spurious peaks seen in MVDR.

As shown in the spectrum comparison, the LE beamformer ($P_{LE}$) closely tracks the AI metric ($P_{AI}$), providing a sharp, unambiguous peak at the true DoA. In contrast, the Conventional Beamformer ($P_{CB}$) is overly broad, and MVDR ($P_{MV}$) exhibits multiple misleading sidelobes in this scenario.

Key Insights: Invariance and Regularization

The core advantage of the Riemannian approach lies in invariance. In Euclidean space, the distance between matrices changes upon inversion, which is why fitting $\mathbf{R}$ leads to CB while fitting $\mathbf{R}^{-1}$ leads to MVDR. In Riemannian space, distance is invariant to inversion, allowing the LE beamformer to act as a "bridge" between these two extremes.

Furthermore, the LE beamformer can be interpreted as a non-linear "shrinkage" of the covariance matrix's eigenvalues. This provides automatic regularization, preventing the "spiky" instability of MVDR without sacrificing the resolution needed to distinguish closely spaced sources.

Performance Analysis and Results

We evaluated the LE beamformer using analytical derivations and synthetic simulations with Uniform Linear Arrays (ULA). Key metrics included the Half-Power Beamwidth (HPBW) and the Peak-to-Sidelobe Ratio (PSLR).

Analytical comparison of (a) Half-power beamwidth and (b) Peak-to-sidelobe ratio across SNR levels.

Our results demonstrate several critical advantages:

  • Superior Resolution at Low SNR: For SNRs below 5 dB, the LE beamformer maintains a significantly thinner mainlobe than both CB and MVDR, making it a specialist for noisy environments.
  • Enhanced Sidelobe Suppression: The LE method consistently achieves a much higher PSLR than its counterparts, effectively mitigating "leakage" from strong interferers.
  • Robustness to Small Sample Sizes: In tests with as few as 10 snapshots, MVDR often fails by producing "fake" peaks, whereas the LE beamformer remains stable and accurate.
  • Multipath Performance: The LE beamformer remains functional in coherent multipath scenarios where MVDR typically fails, though it may slightly under-estimate signal power compared to the CB.

Conclusion

By respecting the intrinsic Riemannian geometry of covariance matrices, the Log-Euclidean beamformer resolves the long-standing trade-off between resolution and robustness in DoA estimation. Its closed-form efficiency and superior performance in low-SNR and small-sample regimes make it a powerful tool for modern array processing. Future work will explore further "robustification" and validation with real-world experimental data.

Summary updated

Abstract

Covariance fitting (CF) is a comprehensive approach for direction of arrival (DoA) estimation, consolidating many common solutions. Standard practice is to use Euclidean criteria for CF, disregarding the intrinsic Hermitian positive-definite (HPD) geometry of the spatial covariance matrices. We assert that this oversight leads to inherent limitations. In this paper, as a remedy, we present a comprehensive study of the use of various Riemannian metrics of HPD matrices in CF. We focus on the advantages of the Affine-Invariant (AI) and the Log-Euclidean (LE) Riemannian metrics. Consequently, we propose a new practical beamformer based on the LE metric and derive analytically its spatial characteristics, such as the beamwidth and sidelobe attenuation, under noisy conditions. Comparing these features to classical beamformers shows significant advantage. In addition, we demonstrate, both theoretically and experimentally, the LE beamformer's robustness in scenarios with small sample sizes and in the presence of noise, interference, and multipath channels.
BibTeX
@misc{picard2024riemannian,
  author = {Joseph S. Picard and Amitay Bar and Ronen Talmon},
  title = {{Riemannian Covariance Fitting for Direction-of-Arrival Estimation}},
  year = {2024},
  eprint = {2404.03401},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.03401},
}
Figure