TL;DR
This paper proposes a novel direction-of-arrival (DoA) estimation algorithm for coprime arrays that reconstructs the covariance matrix using rank minimization. By filling in missing "holes" in the virtual array data and reformulating the problem as a multi-convex optimization, the method achieves enhanced DoA estimation accuracy and resolution compared to existing techniques, especially in challenging low-SNR scenarios.
What Problem Does This Paper Address?
Direction-of-arrival (DoA) estimation is crucial in many applications, and sparse arrays like coprime arrays offer advantages in terms of increased degrees-of-freedom (DoFs) and resolving more sources than physical sensors. However, a key challenge with coprime arrays is the non-uniform spacing of their "virtual sensors" (derived from signal correlations), which leads to missing data points or "holes." This model mismatch degrades the performance of standard DoA estimation algorithms. Existing solutions either require hardware modifications or rely on interpolation methods that can introduce approximation errors. This paper aims to address this by developing a robust interpolation and reconstruction method that fully utilizes the available information from the coprime array.
Methods
The core of the proposed method lies in reconstructing the covariance matrix of a virtual uniform linear array (ULA) from the sparse and incomplete data of the coprime array.
- Virtual Array Interpolation: The algorithm first constructs a virtual array signal from the cross-correlations of physical sensors. To handle the missing data points (holes) in this virtual array, it employs an interpolation strategy.
- Atomic Norm Minimization (ANM) Reformulation: The interpolated virtual array signal is then framed as an Atomic Norm Minimization (ANM) problem. This is a standard technique for sparse recovery.
- Rank Minimization Conversion: Crucially, the ANM problem is reformulated into an equivalent rank minimization problem of a Hermitian and Toeplitz matrix. This is a key innovation, as directly solving rank minimization can be NP-hard.
- Multi-Convex Optimization: To overcome the complexity of rank minimization, the paper introduces a dual-variable cyclic rank minimization approach. This involves a multi-convex optimization problem that can be solved iteratively by alternately optimizing two variables (related to the signal and noise subspaces). This approach avoids the approximation loss often associated with relaxing rank to nuclear norm.
- DoA Estimation: Once the covariance matrix of the virtual ULA is accurately reconstructed, conventional subspace-based algorithms like MUSIC can be applied to estimate the DoAs.
Datasets
Not discussed in the paper. The evaluation is performed using simulated data with controlled parameters (SNR, number of snapshots, source locations).
Key Insights and Ideas
- Bridging the Gap: The paper effectively bridges the gap between sparse array advantages (high DoFs) and their inherent data sparsity issues (holes) by proposing a principled reconstruction method.
- Rank Minimization for Covariance: Reformulating the DoA problem as a rank minimization of a Toeplitz matrix is a clever way to leverage the structure of covariance matrices and avoid direct ANM relaxation errors.
- Multi-Convexity for Robustness: The use of a multi-convex optimization framework allows for simultaneous optimization of signal and noise subspaces, leading to more accurate covariance matrix estimation and, consequently, better DoA performance.
- Avoiding Nuclear Norm Approximation: By using a specific rank function reformulation, the method avoids the approximation errors inherent in nuclear norm minimization, which is a common relaxation for rank minimization.
Key Results
- Enhanced DoF Capability: The proposed algorithm accurately estimates the DoAs of nine sources using only seven physical sensors, demonstrating the high DoF capability of coprime arrays when properly utilized.
- Improved Resolution: The method successfully resolves two closely spaced sources (0.5° apart) with distinct peaks in the spatial spectrum.
- Superior RMSE Performance: In simulations, the proposed algorithm consistently achieves lower Root Mean Square Error (RMSE) compared to state-of-the-art methods (SSR, NNM, ANM, etc.), especially at low SNRs. It closely approaches the theoretical Cramér-Rao Bound (CRB).
- Robustness to SNR and Snapshots: The performance advantage is maintained across a wide range of SNRs and numbers of snapshots.
- Computational Trade-off: While slightly more computationally intensive than ANM (which only optimizes the signal subspace), the proposed method's complexity is manageable and justified by its improved accuracy.
Limitations and Future Work
- Computational Complexity: While efficient, the proposed method is noted to be slightly more computationally complex than ANM because it optimizes both signal and noise subspaces.
- Future Work: The paper implicitly suggests that further optimization of the cyclic iteration process or exploring alternative convex relaxations could potentially reduce computational load. Investigating its performance with correlated sources or in more complex environments (e.g., 2D DoA) could also be future directions.
Summary updated