Non-Uniform Sparse Arrays: Principles & Designs

• NSAs are defined by irregular sensor positions on a subset of a uniform grid, leveraging co-array theory to create high virtual degrees of freedom with fewer elements.
• They enable enhanced resolution and target detection by synthesizing virtual apertures through methods like compressive sensing, convex optimization, and gridless recovery.
• NSA designs reduce hardware complexity and power consumption, making them ideal for advanced radar, sonar, medical imaging, and MIMO communications applications.

A non-uniform sparse array (NSA) is an antenna or sensor array in which the physical positions of the elements form a strict subset of a uniform grid, leading to irregular (non-uniform) inter-sensor spacings. By exploiting the properties of co-arrays, NSAs can emulate the spatial degrees of freedom of much denser arrays with dramatically fewer elements. This allows for reduced hardware complexity, lower power consumption, and expanded virtual apertures, enabling resolution and target capacity that would otherwise be unattainable. NSAs form a foundation for modern radar, sonar, medical imaging, and communications systems, and their design incorporates discrete mathematics, convex optimization, and statistical array processing.

1. Mathematical Foundations and Co-Array Theory

Let the physical sensor locations be $$\mathcal D = \{ d_n \in \mathbb{Z}^D : n=1,\ldots,N \}$$, a subset of the complete $D$-dimensional $\lambda/2$ grid. An NSA is defined by $$\mathcal D \subset \{0 : L_1\} \times \cdots \times \{0 : L_D\}$$ with $|\mathcal D| \ll \prod_i (L_i+1)$. NSAs are characterized by their co-arrays:

• Difference co-array: $$\mathcal C_\Delta = \{ d_m - d_n : d_m, d_n \in \mathcal D \}$$
• Sum co-array: $$\mathcal C_\Sigma = \{ d_m + d_n : d_m, d_n \in \mathcal D \}$$

The difference co-array defines the lags at which data covariance samples are observed, underpinning virtual aperture, spatial DOF, and beamforming resolution. If $\mathcal C_\Delta$ is contiguous (e.g., $\{ -L : L \}^D$), the virtual array is fully filled, and maximum DOF are attained per physical aperture.

In active sensing or MIMO configurations, the sum co-array $\mathcal C_\Sigma$ encodes virtual Tx-Rx pairs, allowing $M$ transmitters and $N$ receivers to simulate up to $MN$ virtual channels, vastly outperforming $N$-DOF phased arrays of the same size (1803.02219, Rajamäki et al., 21 Jan 2026).

2. NSA Geometries and Design Principles

Several key NSA classes have emerged, each with analytic placement rules and distinct co-array properties:

• Minimum Redundancy Arrays (MRA): For given aperture, minimize element count such that $\mathcal D + \mathcal D$ is contiguous. MRAs are optimal in redundancy but require intractable combinatorial search for large dimensions.
• Boundary and Concentric Rectangular Arrays (CRA): Perimeter or interleaved perimeters (CRA) in 2D, such as $\mathcal D_{CRA}$ for $L \times L$ grids, guarantee a contiguous co-array with constant or near-constant number of unit spacings, directly controlling mutual-coupling (1803.02219).
• Nested, Co-prime, and Semi-Coprime Arrays: Analytic constructions using nested or interleaved subarrays or coprimality theory, e.g., $\mathcal S_{NA}$, $\mathcal S_{CP}$, $\mathcal S_{SCA}$, achieve filled or near-filled co-array lags, quadratic virtual aperture scaling, and favorable trade-offs between sensor count and DOF (Daher, 2018, Adhikari, 2018).
• Non-uniform Planar and Irregular Arrays: Placement optimized via off-grid compressive sensing or iterative spatial refinement to match specific beampattern or sum-rate targets, with constraints on inter-sensor spacing (Yang et al., 2022, Chen et al., 2024).

Design Metrics

Array Class	No. Sensors $N$	Virtual DOF	Unit Spacings $U$	Redundancy $R_\infty$
URA	$L^2$	$(2L+1)^2$	$2L^2+2L$	$\infty$
MRA	$\leq 2L$	$(2L+1)^2$	$\geq 8$	$1.19$–$2$
BA/CRA	$2\cdot 2L$	$(2L+1)^2$	$4L$/~16	$(\rho+1)^2/(2\rho)$

$U$, the count of unit and near-unit spacings, dominates mutual-coupling, which is reduced in perimeter/interleaved designs (BA/CRA) (1803.02219).

3. Array Synthesis, Optimization, and Signal Processing Algorithms

NSA synthesis comprises both closed-form designs and algorithmic element placement:

• Compressive Sensing (CS) Synthesis: Grid-based or off-grid sparse synthesis methods select locations and excitations to achieve desired beampattern fitting, subject to minimum-element spacing for physical realization (Yang et al., 2022). Off-grid OMP/LAOMP iteratively refines element positions using Taylor expansion of the array manifold, offering improved beampattern fidelity and lower computational load than dense grid CS.
• Convex/Successive Approximation: In multiuser near-field MIMO communications, positions are optimized to maximize sum-rate or minimize interference using quadratic surrogate functions and convex programming subject to geometric constraints, e.g., minimum spacing and aperture bounds (Chen et al., 2024).
• Gridless Methods: Atomic-norm minimization with array manifold separation and accelerated proximal gradient solvers delivers efficient super-resolution DOA estimation for arbitrary NSA geometries, sidestepping grid artifacts and permitting effective use in automotive MIMO radars (Gao et al., 2023).
• Bayesian/Deep Learning Approaches: One-bit DOA estimation with off-grid Bayesian inference, and learned surrogate algorithms (unrolled networks), can robustly recover source locations and off-grid offsets, addressing quantization artifacts and manifold irregularities imposed by NSA layouts (Hu et al., 2024).

4. Statistical Performance, Resolution, and Trade-offs

The statistical characteristics of NSA-based systems are shaped by three fundamental trade-offs:

• DOF and Sample Complexity: NSAs can offer $O(N^2)$ virtual DOF with $O(N)$ sensors. Super-resolution in DOA estimation, with minimum separation scaling as $O(1/N^2)$ (nested arrays), is feasible even in sample-starved regimes provided the number of snapshots $L$ satisfies $L = \Omega(\ln N / \epsilon^2)$, under moderate SNR and dynamic range (Sarangi et al., 2023).
• SNR Penalty of Sparsity: The compression of a ULA to an NSA incurs an SNR loss of $10 \log_{10}(C)$ dB, where $C$ is the ULA-to-NSA element number ratio. Cramér–Rao bounds rise accordingly, but estimation variance and bias above threshold SNR can match the ULA with proper orthogonal-subspace methods (Pakrooh et al., 2015).
• Bias-Variance-Resolution Trade-off: The expectation and covariance of product-processor outputs for NSA/apodized arrays are formally described as convolution of the true spatial PSD with the Fourier transform of the difference coarray taper. Taper design on the coarray yields direct control over main-lobe width (resolution), sidelobe level (bias), and estimator variance for spectral and DOA estimation (Sartori et al., 2021).

5. Mutual Coupling, Implementation, and Application Domains

Practical performance of NSAs is often limited by mutual coupling between closely spaced elements, especially in planar or active arrays:

• Mutual Coupling Mitigation: Designs such as the CRA explicitly minimize the number of unit inter-element spacings $U$ to bound coupling effects while maintaining full co-array coverage. Mirror-symmetric layouts and multi-perimeter interleaving (CRA) reduce worst-case coupling by an order of magnitude over naive layouts with similar element counts (1803.02219).
• Coupling Modeling and Compensation: NSA performance under coupling is modeled using coupling matrices, e.g., $C$, with iterative or joint optimization methods (e.g., CMA-ES augmented with block coordinate descent) restoring sub-degree accuracy in high-resolution DOA estimation (Daher, 2018).
• Applications: NSAs are now core to:
  • Active radar/sonar: high angular resolution, target detection beyond physical aperture limits (1803.02219, Daher, 2018, Rajamäki et al., 21 Jan 2026)
  • Medical ultrasound: downsized probe arrays with full imaging DOF (1803.02219, Rajamäki et al., 21 Jan 2026)
  • Wireless communications: massive-MIMO and near-field channel estimation, with sum-rate maximization via NSA placement (Chen et al., 2024)
  • Automotive MIMO radar: enhanced resolution and gridless DoA estimation with array manifold separation (Gao et al., 2023)

6. Contemporary Challenges and Future Directions

• Computational Intractability: Combinatorial MRA/NRA design remains practical only for small arrays; scalable approximations or convex relaxations are necessary for large NSA synthesis (1803.02219, Yang et al., 2022).
• Non-asymptotic and Sample-Starved Regimes: Modern analyses rigorously demonstrate that sample complexity and SNR requirements of NSAs are comparable—or superior—to dense arrays when leveraging co-array scaling, especially under well-separated source configurations (Sarangi et al., 2023).
• Extension to Planar/Volumetric and Near-Field Regimes: Recent optimization and beamforming approaches now handle 2D/3D placements, near-field focusing (with focal shift minimization), and multiuser MIMO with non-uniform spacing constraints (Li et al., 12 May 2025, Yang et al., 2022, Chen et al., 2024).
• Integration with Deep Learning: Unrolled optimization and data-driven surrogate algorithms deliver robust, hyperparameter-free, low-complexity solutions to nonlinear measurement and off-grid errors unique to NSA systems (Hu et al., 2024).

7. Summary Table: NSA Classes and Properties

Array Class	Construction/Design Rule	Virtual DOF	Notable Properties	Reference
Uniform Linear (ULA)	All positions filled	$N$	No DOF gain, full sampling	(Sarangi et al., 2023)
MRA	Minimize redundancy, contiguous co-array	Up to $N^2$	Intractable for large $N$	(1803.02219)
Nested	Short uniform + spaced subarray	$N_1N_2$	Simple closed-form, quadratic gain	(Daher, 2018)
Co-prime	Interleaved coprime subarrays	$\sim MN$	Mild holes, analytic geometry	(Daher, 2018)
Semi-Coprime	3-path ULA interleaving	$\sim PQMN$	Extreme savings, PSLR matched	(Adhikari, 2018)
CRA	Interleaved concentric rectangles	$(2L+1)^2$ (2D)	Minimized unit spacings $U$	(1803.02219)
Off-grid optimized	Iterative convex/CS refinement	User specified	Flexible, enforces spacing	(Yang et al., 2022)

NSAs represent a mature, deeply optimized family of array architectures enabling large-aperture, high-resolution sensing and communications with severe constraints on the number, placement, and mutual coupling of physical sensors. Their design continues to be advanced by rigorous statistical characterization, scalable optimization, and innovative signal processing tailored for practical deployment.