⚡ electrical engineering

Decoupled Azimuth Elevation AoA Estimation Exploiting Kronecker Separable Steering Matrices

This paper proposes an economical subspace decoupling framework for 2D AoA estimation in Kronecker-separable arrays that extracts joint signal subspaces to enable independent 1D processing with pairing, achieving higher accuracy and spectral efficiency than state-of-the-art methods for medium- and large-scale arrays.

Original authors: Faizan A. Khattak, Ian K. Proudler, Stephan Weiss, Fazal-E Asim

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Faizan A. Khattak, Ian K. Proudler, Stephan Weiss, Fazal-E Asim

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, open field with a group of friends holding microphones. You want to figure out exactly where a sound is coming from. In a 3D world, you need to know two things: the horizontal direction (left or right, called azimuth) and the vertical direction (up or down, called elevation).

Usually, figuring out these two directions at the same time with a grid of microphones is like trying to solve a giant, tangled knot. It requires massive amounts of computer power and takes a long time, especially if the microphones aren't perfectly spaced out.

This paper introduces a clever new way to untangle that knot. Here is the simple breakdown:

1. The "Lego" Trick (The Core Idea)

The authors noticed that for certain types of microphone grids (specifically rectangular ones, whether they are perfectly uniform or have a specific, structured irregularity), the math describing the sound direction works like Lego bricks.

Instead of one giant, complicated block representing the whole 3D direction, the math shows that the direction is actually just two smaller, separate blocks snapped together:

One block handles the Left/Right (Azimuth) info.
One block handles the Up/Down (Elevation) info.

The paper calls this a "Kronecker product," but think of it as realizing that a complex 3D puzzle is actually just two simple 2D puzzles stuck together.

2. The Old Way vs. The New Way

The Old Way (The "Search Every Corner" Method):
Traditional methods treat the whole grid as one giant, messy system. To find the sound, the computer has to check every possible combination of Left/Right and Up/Down simultaneously.

Analogy: Imagine trying to find a lost key in a city. The old method forces you to check every single street and every single building floor at the same time. If the city has 100 streets and 100 floors, you have to check 10,000 combinations. It's slow and exhausting.

The New Way (The "Split and Conquer" Method):
The authors' method first separates the data. It uses a mathematical trick to pull the "Left/Right" information out of the "Up/Down" information, creating two separate, clean lists.

Analogy: Now, instead of checking the whole city at once, you first find the correct street (checking only 100 options). Once you know the street, you only need to find the correct floor on that specific street (checking another 100 options). You only did 200 checks instead of 10,000.

3. Why This Matters

The paper claims this "Split and Conquer" approach offers three main benefits:

It's Much Faster: Because the computer solves two small problems instead of one giant one, it saves a huge amount of time. The paper notes that for a specific setup, the number of calculations dropped by over 90 times.
It Works Better with Less Data: Usually, to get a clear picture, you need many "snapshots" (samples) of the sound. This new method can get a very accurate result even with fewer snapshots. It's like being able to guess the flavor of a soup with just one spoonful, whereas other methods need a whole bowl.
It Handles "Messy" Arrays: It works not just on perfect grids, but also on "structured non-uniform" arrays (where the microphones aren't perfectly spaced but follow a pattern). This gives engineers more flexibility in how they build their sensor arrays.

4. The "Pairing" Step

There is one small catch. Once the computer finds the best "Left/Right" angle and the best "Up/Down" angle separately, it has to make sure they belong to the same sound source.

Analogy: If you found the correct street and the correct floor, you still need to make sure they match the same apartment. The paper uses a standard, efficient method to "pair" these two answers together to ensure they describe the same source.

Summary

In short, this paper presents a mathematical shortcut. Instead of brute-forcing a complex 3D direction-finding problem, it breaks the problem into two easy 1D problems. The result is a system that is faster, more accurate in noisy conditions, and requires less data to work, without needing the heavy computing power usually associated with 3D sound tracking.

Technical Summary: Decoupled Azimuth–Elevation AoA Estimation

Problem Statement
While linear arrays have been extensively studied for Angle-of-Arrival (AoA) estimation, planar array geometries—specifically Uniform Rectangular Arrays (URAs) and structured Non-Uniform Rectangular Arrays (NURAs)—have received comparatively less attention despite their advantages in modern applications like MIMO and 3-D localization. Existing methods for planar arrays, such as 2-D MUSIC or 2-D ESPRIT, often suffer from high computational complexity, particularly when requiring 2-D spectral searches or tensor decompositions. Furthermore, computationally efficient methods for structured NURAs (which offer greater deployment flexibility than URAs) remain scarce. The core challenge addressed is the development of an economical framework for 2-D AoA estimation that reduces computational load while maintaining or improving accuracy, especially under conditions of limited snapshots or low Signal-to-Noise Ratio (SNR).

Methodology
The proposed approach exploits the mathematical structure of steering vectors in specific array geometries. The paper considers URAs, structured NURAs, and parallelogram-shaped arrays (UPgA and NUPgA). For these configurations, the full steering matrix $A$ can be expressed as the Khatri–Rao product of two lower-dimensional steering matrices: one for azimuth ( $A_h$ ) and one for elevation ( $A_v$ ).

The methodology proceeds through the following steps:

Subspace Extraction: The joint signal subspace is first extracted from the spatial covariance matrix via eigenvalue decomposition.
Decoupling via Matrix Manipulation: Instead of using tensor-based approaches, the authors derive separate signal and noise subspace projection matrices for each spatial dimension. By reshaping the joint signal subspace eigenvectors and stacking them appropriately, the column spaces of the individual azimuth ( $A_h$ ) and elevation ( $A_v$ ) steering matrices are recovered. This involves constructing matrices $C$ and $B$ from the joint subspace and performing Singular Value Decomposition (SVD) to obtain orthonormal bases for the decoupled subspaces.
Independent 1-D Estimation: Once the decoupled subspaces are obtained, conventional 1-D algorithms (MUSIC, root-MUSIC, or ESPRIT) are applied independently to the azimuth and elevation dimensions.
- For uniform arrays, root-MUSIC and ESPRIT are applied.
- For structured non-uniform arrays, MUSIC or irregular root-MUSIC (based on Vandermonde decomposition) is used.
Angle Pairing: The estimated azimuth and elevation angles are paired using a 2-D spectral function (pseudospectrum) to ensure they correspond to the same source.

Key Contributions

Subspace-Decoupling Framework: The paper introduces a purely matrix-based technique to decouple 2-D AoA estimation into two independent 1-D problems. This avoids the computational expense of tensor-based methods (e.g., tensor-MUSIC or tensor-ESPRIT) and 2-D spectral searches.
Computational Efficiency: The decoupling reduces the computational complexity significantly. For example, in the case of structured NURAs, a 2-D grid search of size $n_h \times n_v$ is reduced to two 1-D searches of size $n_h + n_v$ . The authors cite a specific example where a search space of 32,761 points is reduced to 362, achieving a 90.5-fold reduction.
Applicability to Structured Arrays: The method is explicitly designed to handle structured NURAs and parallelogram arrays, which are more flexible for practical deployment than standard URAs but lack efficient dedicated algorithms.
Improved Performance in Challenging Regimes: The approach is shown to be effective in scenarios with medium-to-large scale arrays, limited snapshots, and low SNR.

Results
Monte Carlo simulations were conducted to evaluate the proposed methods (De-RMUSIC, De-ESPRIT, De-MUSIC, and De-MUSIC-Opt) against state-of-the-art benchmarks (2-D MUSIC, Reduced-Dimension MUSIC, and 2-D ESPRIT/ESPRIT-MIMO).

Accuracy: For URAs of sizes $10\times10$ , $20\times20$ , and $30\times30$ , the proposed De-RMUSIC consistently achieved the lowest Root Mean Square Error (RMSE) across various SNR levels, outperforming RD-MUSIC and often outperforming ESPRIT-MIMO, particularly at low SNRs.
Snapshot Efficiency: The proposed algorithms demonstrated superior accuracy compared to ESPRIT-MIMO when the number of snapshots was limited. While performance converged at higher snapshot counts, the proposed methods maintained an advantage in low-data regimes.
Structured NURAs: For structured NURAs, the decoupled MUSIC approach (De-MUSIC) achieved accuracy comparable to 2-D MUSIC but with significantly lower execution time. The optimized variant (De-MUSIC-Opt), which uses nonlinear optimization for fine search, was the most computationally efficient.
Computational Cost: While the proposed methods require additional SVDs and an angle pairing step (making them slightly more expensive than pairing-free ESPRIT-MIMO in some cases), they are substantially more efficient than 2-D MUSIC and RD-MUSIC. The execution time gap with ESPRIT-MIMO is noted as not significant for the tested array sizes but may increase with larger spatial dimensions.

Significance and Claims
The paper claims that the proposed decoupling strategy offers a practical and efficient alternative to existing 2-D AoA estimation techniques. By leveraging the Kronecker-separable structure of steering matrices, the method enables the use of well-established 1-D algorithms on planar arrays without the heavy computational burden of 2-D searches or tensor decompositions.

The authors emphasize that while the method has a limitation on the maximum number of detectable sources (specifically, it can detect up to $\min\{M, N\} - 1$ sources, which is fewer than the $MN-1$ limit of 2-D MUSIC), it provides a compelling trade-off. For large uniform arrays and structured non-uniform arrays, the approach delivers higher accuracy with fewer snapshots and improved spectral efficiency. The work is presented as a significant step toward enabling high-performance 3-D localization and MIMO radar applications using flexible array geometries where computationally efficient algorithms were previously scarce.