⚡ electrical engineering

Spatial Power Estimation via Riemannian Covariance Matching

This paper introduces SERCOM, a novel spatial power estimation method that leverages the Riemannian geometry of Hermitian positive definite matrices via the efficient Jensen-Bregman LogDet divergence to outperform conventional Euclidean-based approaches, particularly in challenging scenarios involving low SNR, limited snapshots, and correlated sources.

Original authors: Or Cohen, Alon Amar, Ronen Talmon

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Or Cohen, Alon Amar, Ronen Talmon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a large, echoey room with a group of friends, each holding a microphone. Several people in the room are talking at once. Your goal is to figure out exactly where each person is standing and how loud they are speaking, even if the room is noisy or if some people are whispering.

In the world of signal processing, this is called Spatial Power Estimation. The "microphones" are sensors on an array, and the "voices" are signals coming from different directions.

The Old Way: Flattening a Balloon

For a long time, scientists tried to solve this by treating the data from these microphones like a simple, flat list of numbers (a vector in Euclidean space). They would measure the "distance" between what they heard and what they expected to hear using standard math, like measuring the distance between two points on a piece of paper.

The paper argues that this is like trying to measure the distance between two points on a balloon by flattening the balloon onto a table first. When you flatten a balloon, you stretch and distort the rubber. Similarly, treating complex signal data as simple flat numbers distorts the true relationships between the signals, especially when the data is messy (low volume) or scarce (few snapshots).

The New Way: Navigating the Curved Surface

The authors, Or Cohen, Alon Amar, and Ronen Talmon, propose a new method called SERCOM. Instead of flattening the data, they treat it as if it lives on a curved surface (a Riemannian manifold).

Think of it this way:

The Old Way: If you want to walk from New York to London, you might draw a straight line on a flat map. But because the Earth is round, that line is actually wrong; you'd end up in the ocean.
The New Way (SERCOM): This method understands that the Earth is round. It calculates the path along the curve of the planet (the "geodesic"), which is the true shortest distance.

In their method, they use a specific mathematical tool called JBLD divergence. You can think of this as a specialized GPS that knows the terrain is curved. It allows them to compare the "noisy" data they collected with their "ideal" model without stretching or distorting the information.

Why This Matters (The "Superpowers")

The paper claims that by using this "curved surface" approach, SERCOM is much better at finding the sources in difficult situations:

Low Volume (Low SNR): When the speakers are whispering and the room is noisy, the old methods get confused and point in the wrong direction. SERCOM cuts through the noise like a good ear in a quiet room.
Few Snapshots: If you only have a split second to listen (few data points), the old methods struggle to make a clear picture. SERCOM can still figure it out with very little data.
Correlated Sources: If two people are speaking in perfect unison (like a choir), the old methods often think they are just one person. SERCOM can tell them apart.

The Efficiency Trick

Usually, calculating paths on a curved surface is very heavy on the computer (like trying to solve a complex puzzle for every single step). The authors found a clever shortcut using the JBLD tool. It gives them the benefits of the "curved surface" math without the heavy computer cost. It's like finding a scenic hiking trail that is just as beautiful as the difficult mountain climb but takes half the time to walk.

The Results

In their tests, they simulated various scenarios (different numbers of microphones, different noise levels, and different types of speakers).

Accuracy: SERCOM consistently found the correct locations and loudness levels better than the current top methods (called SPICE and SAMV).
Speed: It was faster than other "curved surface" methods because it didn't need to do heavy calculations for every step.

Summary

The paper introduces SERCOM, a smarter way to listen to multiple signals at once. Instead of squashing complex data into a flat, distorted shape, it respects the natural "curved" shape of the data. This allows it to find hidden voices in noisy, crowded, or data-scarce environments more accurately and efficiently than previous methods.

Technical Summary: Spatial Power Estimation via Riemannian Covariance Matching

Problem Statement
Spatial power spectrum estimation is a fundamental challenge in array signal processing, closely linked to Direction-of-Arrival (DOA) estimation. While classical methods (e.g., Bartlett, MVDR) and subspace techniques (e.g., MUSIC) exist, and deep learning approaches have shown promise, many model-based methods rely on "on-grid" covariance matching. In these approaches, the sample covariance matrix ( $\hat{R}$ ) is matched to a model covariance matrix ( $R$ ) parameterized by a spatial power spectrum.

The core issue identified is that conventional covariance matching formulations (such as Asymptotic Minimum Variance (AMV), SPICE, and SAMV) treat covariance matrices as vectors in Euclidean space. This ignores the intrinsic geometry of Hermitian Positive Definite (HPD) matrices, which reside on a Riemannian manifold. Consequently, Euclidean-based methods often suffer from performance degradation when dealing with ill-conditioned sample covariance matrices, low Signal-to-Noise Ratio (SNR), limited snapshots, or correlated sources.

Methodology
The paper proposes SERCOM (Spatial power Estimation via Riemannian COvariance Matching), a new algorithm that formulates the estimation of the spatial power spectrum as a minimization problem on the Riemannian manifold of HPD matrices.

Objective Function: Instead of minimizing Euclidean distances, SERCOM minimizes the Jensen–Bregman LogDet (JBLD) divergence between the model covariance $R(p)$ and the sample covariance $\hat{R}$ . The optimization problem is defined as:
$\hat{p} = \arg \min_{p \in \mathbb{R}^D_+} \left( \log \left| \frac{\hat{R} + R(p)}{2} \right| - \frac{1}{2} \log |R(p)| \right)$
(The constant term $-\frac{1}{2}\log|\hat{R}|$ is omitted).
Advantages of JBLD:
- Geometry-Aware: Unlike Euclidean metrics, JBLD respects the curved geometry of the HPD manifold.
- Computational Efficiency: Unlike the Affine-Invariant Riemannian Metric (AIRM) or Log-Euclidean (LE) distance, JBLD does not require computationally expensive matrix logarithms or eigen-decompositions. It can be efficiently and stably computed via Cholesky factorization.
- Robustness: Theoretical analysis shows JBLD is less sensitive to "outlier" eigenvalues (caused by noise or model mismatch) compared to AMV and SPICE, as it penalizes deviations logarithmically rather than quadratically or linearly.
Optimization: The non-convex objective is solved using a gradient-based approach with the Adam optimizer. The algorithm initializes the power vector using a delay-and-sum beamformer and iteratively updates the power spectrum, projecting the result onto the non-negative orthant at each step to ensure physical validity.

Key Contributions

Novel Formulation: Introduction of a multi-source Riemannian covariance matching framework for spatial power estimation, moving beyond single-source or unit-power models used in previous geometry-aware works.
Theoretical Unification: The paper establishes analytical connections between Euclidean criteria (AMV, SPICE) and Riemannian distances (AIRM, JBLD). It demonstrates that while these criteria are asymptotically equivalent under ideal conditions (large $N$ , perfect model), they diverge in challenging regimes.
Robustness Analysis: Theoretical proofs (Theorem 2) show that JBLD and AIRM assign lower relative weight to eigenvalue outliers compared to AMV and SPICE, explaining their superior robustness to spectral distortions.
Algorithmic Efficiency: The proposed SERCOM(JBLD) achieves the robustness of Riemannian methods while maintaining computational tractability, avoiding the heavy overhead of AIRM and LE implementations.

Experimental Results
Extensive simulations were conducted comparing SERCOM (using JBLD, AIRM, and LE variants) against SPICE and SAMV. Key findings include:

Low SNR & Limited Snapshots: SERCOM variants consistently outperform Euclidean baselines (SPICE, SAMV) in DOA and power estimation accuracy, particularly when $N$ (snapshots) is close to or less than $M$ (sensors). SERCOM(JBLD) approaches the Cramér–Rao Bound (CRB) earlier than competitors.
Correlated Sources: In scenarios with highly correlated or coherent sources, Euclidean methods degrade significantly. SERCOM(JBLD) and SERCOM(AIRM) maintain stable performance, with JBLD showing the lowest DOA RMSE at high correlation levels.
Off-Grid Sensitivity: All grid-based methods suffer from discretization mismatch, but the Riemannian approaches remain robust.
Computational Performance: While SERCOM variants generally require more iterations than SPICE/SAMV, SERCOM(JBLD) avoids the per-iteration eigen-decomposition required by AIRM and LE. Consequently, SERCOM(JBLD) scales favorably with the number of sensors ( $M$ ), exhibiting significantly lower runtime than AIRM/LE variants for large arrays (e.g., $M=120$ ).

Significance and Claims
The paper claims that SERCOM provides a computationally practical and robust framework for spatial power estimation. By leveraging the Riemannian geometry of covariance matrices via the JBLD divergence, the method overcomes the limitations of traditional Euclidean matching, particularly in non-ideal operating conditions (low SNR, few snapshots, correlated sources).

The authors position SERCOM as a superior alternative to existing state-of-the-art methods (SPICE, SAMV) for practical applications where signal conditions are non-ideal. They note that while the performance of different Riemannian metrics is comparable in terms of accuracy, the JBLD-based approach offers a distinct advantage in computational efficiency and stability. The work suggests that SERCOM should be considered a strong candidate for real-world deployment and outlines future work to extend this framework to gridless DOA estimation.