Non-Uniform Sparse Arrays Overview

• Non-uniform sparse arrays are sensor configurations using irregular placements to create rich difference coarrays, maximizing virtual aperture and degrees of freedom.
• They employ advanced signal processing methods such as MUSIC, ESPRIT, and sparse recovery to estimate more sources than sensors, often achieving O(N²) virtual apertures.
• Design strategies include analytical constructs like nested arrays and optimization-based synthesis to balance sensor spacing, reduce mutual coupling, and enhance radar and MIMO performance.

Non-uniform sparse arrays are sensor array configurations in which sensor elements are positioned non-uniformly across an aperture, with many possible spacings often exceeding the classical half-wavelength criterion. These arrays are engineered to maximize spatial degrees of freedom (DoF), virtual aperture, and parameter-estimation performance while minimizing the number of physical elements. Unlike uniform linear or planar arrays (ULAs, UPAs), non-uniform sparse arrays (NSAs) exploit irregular spacing and combinatorial coarray effects to achieve resolutions and source-number capabilities far beyond their sensor count, though at the expense of increased design and signal-processing complexity. Their central theoretical mechanism is the difference coarray, which determines virtual sensor locations available for super-resolutive processing, typically using subspace, sparse-recovery, or gridless convex optimization techniques. Non-uniform sparse arrays underpin advances in direction finding, radar, MIMO communications, near-field focusing, and sparse array synthesis.

1. Array Geometries, Coarray, and Sparsity Metrics

A non-uniform sparse array is formally specified by a set of sensor positions $\mathcal{S} = \{d_1, \dots, d_N\}$ (for linear arrays, $d_n \in \mathbb{R}$ or $\mathbb{Z}$). The key analytic object is the difference coarray: $\mathcal{D} = \{d_i - d_j : 1 \le i,j \le N\}$ which collects all pairwise inter-element displacements and underpins the array's effective aperture and achievable DoF. The set of unique lags, $\mathcal{D}_{\text{unique}}$, is the foundation for virtual array modeling: any direction-finding or imaging algorithm operating on $\mathcal{D}_{\text{unique}}$ leverages the virtual sensor manifold, $$\mathbf{a}_D(\theta) = [e^{j2\pi c_m \sin\theta}]_{m=1}^{|\mathcal{D}_{\text{unique}}|}$$, for $c_m \in \mathcal{D}_{\text{unique}}$ (Leite et al., 2021).

Popular sparse non-uniform geometries include:

• Minimum-Redundancy Arrays (MRAs): Sensor placements that maximize the length of the largest contiguous hole-free ULA segment in $\mathcal{D}$, but require computational search for large $N$ (Daher, 2018).
• Minimum-Hole Arrays (MHAs): Configurations that minimize the number of "holes" in the difference coarray.
• Nested Arrays: Concatenate a dense sub-ULA and a sparser, offset sub-ULA to ensure a contiguous virtual coarray with $O(N^2)$ unique lags for $N$ sensors (Sarangi et al., 2023).
• Coprime, Extended Coprime, Thinned Coprime Arrays: Constructs interleaving two or more sub-ULAs with coprime or semi-coprime undersampling factors. Thinned designs further reduce sensor count by eliminating redundant interior sensors (Raza et al., 2017, Adhikari, 2018).
• Concentric Rectangular Arrays (CRAs): For 2D apertures, interleaved sparse rectangles achieve rectangular difference coarrays with minimal unit-spacing pairs, thus reducing mutual coupling (1803.02219).

The sparseness metrics include total element count, minimum/mean/maximum inter-element spacings, number of unit-spaced element pairs $S(1)$ (which quantifies nearest-neighbor coupling propensity), and (especially in 2D) the asymptotic redundancy ratio $R_\infty$. For example, a rectangular CRA achieves $S_\text{CRA}(1)=16$, constant for large $L_x, L_y$, compared to $S_\text{BA}(1)=2(L_x+L_y)$ for a boundary array (1803.02219).

2. Signal Processing and Virtual Arrays

Non-uniform sparse arrays are fundamentally distinguished by their non-Vandermonde steering matrices and the leveraging of virtual coarrays for super-resolution.

Virtual Aperture and Effective DoF: By vectorizing and redundancy-averaging the observed array covariance, one constructs a virtual covariance matrix indexed by $\mathcal{D}_{\text{unique}}$. The length of the largest hole-free portion of $\mathcal{D}$ sets the coarray DoF; for nested arrays, this grows as $O(N^2)$, whereas for ULAs it is always $N$ (Sarangi et al., 2023, Daher, 2018).

Coarray Domain Processing: Methods such as MUSIC, ESPRIT, and spatial smoothing adapted to the coarray domain enable estimating more sources than sensors by operating on the virtual uniform manifold (Daher, 2018, Leite et al., 2021). For closely spaced or coherent sources, and in passive and active array regimes, sparse reconstruction or atomic-norm minimization (ANM) on the virtual coarray offer gridless and robust estimation, as developed for automotive radar and communications (Gao et al., 2023).

Mutual Coupling: Sparse arrays often exhibit reduced mutual coupling compared to filled ULAs, particularly for designs with constant or small $S(1)$. For example, the CRA achieves a $\approx 25\%$ lower RMSE under mutual coupling versus filled uniform or boundary arrays of the same aperture (1803.02219).

3. Synthesis and Design Algorithms

Non-uniform sparse array design can be approached analytically or via optimization.

• Block-Partitioned and Coprime Designs: Analytical, closed-form constructions leverage number-theoretic properties (e.g., coprimality or block partitioning) to produce coarrays with desirable DOF and minimal grating lobes (Anarakifirooz et al., 2022, Raza et al., 2017, Adhikari, 2018). For example, semi-coprime arrays interleave three ULAs, yielding virtual aperture and sidelobe behavior equivalent to a much larger ULA at a fraction of physical sensor count (Adhikari, 2018).
• Optimization-Based Array Synthesis: Compressive sensing (CS) approaches recast array synthesis as an $\ell_0$- or mixed-norm minimization, often subject to pattern-matching constraints and physical limitations such as minimum inter-element spacing (Hawes et al., 2016, Yang et al., 2022). Off-grid OMP and variants extend this to continuous element position optimization (Yang et al., 2022), while low-rank Hankel matrix completion (via log-det heuristics and Matrix Pencil Method for extraction) offers an alternative relaxation for achieving a desired pattern with a minimum number of elements (Zhang, 2022).
• Specialized Patterns and Near-Field Focusing: For near-field tasks such as constant-distance beam focusing, analytic closed-form spacing (e.g., $d_\text{opt}\simeq \sqrt{\lambda L/N}$) ensures suppression of focal shift and equalized beam width across scan positions (Li et al., 12 May 2025).

4. Advanced Direction-Finding and Channel Estimation

Finite Snapshot and Low-SNR Regimes: Contrary to earlier assumptions, sparse arrays do not require significantly more snapshots than filled ULAs for reliable parameter estimation. In the regime of $S=O(1)$ sources, nested arrays guarantee matching distance error $\leq\epsilon$ with high probability using only $O(\ln P/\epsilon^2)$ snapshots, where $P$ is the physical sensor count. The aperture, and hence resolution, scales as $O(1/P^2)$, breaking the $O(1/P)$ barrier of ULAs (Sarangi et al., 2023).

Gridless Sparse Recovery: For arbitrary array geometry, manifold separation or array interpolation transforms the non-uniform steering manifold into an equivalent virtual Vandermonde representation, enabling gridless atomic norm minimization (ANM). The fast ANM-based DoA estimator (FNLANM) utilizes this to match ULA-level estimation performance with nonuniform arrays efficiently, critical for MIMO radar and automotive applications (Gao et al., 2023).

List-Based Sparse Recovery: Enhanced Difference Coarray Transformation Models (EDCTM) remove finite-sample errors in the virtual coarray, while List-Based Maximum Likelihood OMP (LBML-OMP) integrates fast greedy support selection and maximum likelihood cost reduction to improve support recovery over OMP, CoSaMP, or SS-MUSIC (Leite et al., 2021).

Near-Field Multiuser MIMO: In near-field multiuser communications, non-uniform sparse arrays can be optimized (using successive convex approximation) to maximize sum rate, suppress multiuser interference due to grating lobes, and harness SD-A domain sparsity with on-grid and off-grid super-resolution channel estimation algorithms (Chen et al., 2024).

5. Trade-offs, Practicalities, and Extensions

Aperture, Side Lobes, and Grating Lobe Mitigation: Non-uniform sparse designs achieve large aperture and low mutual coupling but may introduce irregular side-lobe structure and grating lobes. Closed-form designs (e.g., SCA, block-partitioned NULAs) explicitly suppress grating lobes for favorable propagation in massive MIMO, restoring the channel orthogonality lost in large-ULA deployments (Anarakifirooz et al., 2022, Adhikari, 2018).

Tapering and Multidimensionality: Non-uniform tapers (e.g., Hann, Hamming) on sparse arrays directly manipulate the spatial response, trading main-lobe width for sidelobe suppression. Product-processor output statistics for general sparse geometries and tapers extend bias-variance trade-off intuition from classical spectral estimation to multi-dimensional non-uniform apertures (Sartori et al., 2021).

Robustness to Mutual Coupling and Calibration: Both pattern-based and optimization-based approaches accommodate explicit mutual coupling modeling. Compensation techniques include calibration, iterative error minimization, and joint channel–coupling matrix estimation (Daher, 2018).

Extension to Planar/2D Arrays: Designs such as CRAs for rectangular apertures show how principles generalize to 2D, achieving fully filled rectangular difference coarrays with minimal unit-spacing counts (1803.02219). Synthesis via off-grid OMP and Hankel matrix completion extends to planar layouts, with explicit spacing and pattern constraints (Yang et al., 2022, Zhang, 2022).

6. Comparative Advantages and Limitations

Geometry/Method	Sensor Count	Virtual DOF	Analytical Design	Implementation Limitation
Nested Array	$O(N)$	$O(N^2)$	Yes	Discrete element positions
Thinned Coprime Array	$O(N)$	$O(N^2)$	Yes	Aperture holes (unique lag drop-offs)
MRA/MHA	$O(N)$	Maximal	No (search)	Computation-intensive for large $N$
CRA (rectangular 2D)	$2(L_x+L_y)$	$(2L_x+1)\times(2L_y+1)$	Yes	Even $L_x,L_y$ for exact construction
SCA (semi-coprime)	$\ll$ ULA	$PQMN$	Yes	Parameter selection for overlap/nulls
CS-based/Off-grid OMP	Application-dependent	Application-dependent	Yes/No	Requires grid refinement, high comp.
Block-partitioned NULA	$N_b N$	$N_bN$	Yes	Trade-off between block size/sparsity

Non-uniform sparse arrays consistently offer an improvement in the number of resolvable sources, spatial resolution, and mutual coupling robustness for a given footprint. Limitations include complexity in design and calibration, solver runtime for large-scale optimization-based synthesis, main-lobe/sidelobe irregularities in unconstrained designs, and possibly a higher sensitivity to calibration errors due to irregular element spacing. Extensions to multi-dimensional, conformal, or adaptive arrays are ongoing research themes.

7. Concluding Observations and Outlook

Non-uniform sparse arrays—via analytical, combinatorial, and optimization-based designs—enable spatial sensing and communications systems to surpass the Rayleigh and redundancy limits of conventional uniform arrays. Their central enabling mechanism is the exploitation of virtual coarrays, permitting super-resolution, high DoF, and robust direction finding with sparse sensing hardware. As research progresses, domains such as near-field multiuser MIMO, radar imaging, massive MIMO for wireless, and energy-efficient spatial sensing will continue to be driven by both new sparse array architectures and advances in sparse signal processing, gridless convex optimization, and mutual coupling modeling (1803.02219, Daher, 2018, Sarangi et al., 2023, Gao et al., 2023, Chen et al., 2024).