Sparse Methods for Direction-of-Arrival Estimation

Introduction

• DOA estimation retrieves directions of electromagnetic waves or sources from sensor-array outputs and is used in radar, sonar, and wireless communications.Source: §1 · p. 3
• Capon’s beamformer, subspace methods developed after Pisarenko including MUSIC and ESPRIT, and NLS have known limitations.Subspace methods and NLS require source-count knowledge; covariance-based methods need sufficient snapshots and can be sensitive to source correlations; NLS requires accurate initialization.Source: §1 · p. 3

Sparse DOA methods

• Sparse DOA methods address demanding cases with unknown source count, few or even one snapshot, and highly or completely correlated sources.Motivated by sparse representation and compressed sensing, these methods have attracted extensive study.Source: §1 · p. 3
• Because DOAs are continuous-valued and data depend nonlinearly on them, sparse DOA methods are classified as on-grid, off-grid, or gridless.On-grid methods constrain DOAs to fixed grid points; off-grid methods use a grid without restricting DOAs to it; gridless methods operate directly in the continuous domain.Source: §1 · p. 3

Introduction

• The article introduces the far-field narrowband data model, sparse representation foundations, and on-grid and off-grid methods before focusing on gridless methods and future directions.Gridless methods operate directly in the continuous domain and have strong theoretical guarantees.Source: §1 · pp. 3–4

2.1 Data Model

• At each snapshot, sensor outputs combine K source signals through steering vectors and additive noise; DOA estimation seeks source directions from data matrix Y.The source count K is usually unknown and is typically assumed smaller than the sensor count M.Source: §2.1 · p. 4

2.2 The Role of Array Geometry

• Array geometry determines steering vectors; for linear arrays, the phase depends on sensor position and direction, and a single snapshot becomes a line-spectral estimation problem.For a linear array, the steering map can be parameterized by spatial frequency.Source: §2.2 · p. 5
• Two-dimensional arrays can resolve DOAs over 360°, whereas linear arrays resolve 180°; grid-based sparse methods apply to arbitrary arrays, but gridless methods typically target ULAs or SLAs.Source: §2.2 · p. 6

2.3 Parameter Identifiability

• Theorem 2.1 gives a necessary and sufficient uniqueness condition; more snapshots can permit more sources, and for a ULA the condition simplifies to K < M + rank(Y).The additional-snapshots conclusion excludes the case where source signals are identical up to scaling.Source: §2.3 · pp. 6–7
• With fixed DOAs and randomly drawn source signals, gives a probability-one uniqueness guarantee under (14), while notes finite-SNR cases can still produce false DOAs far from the true directions.The passages also state that condition (15) is necessary.Source: §2.3 · p. 7

3.1.1 Problem Formulation

• Sparse recovery seeks a sparse coefficient vector that fits the observations; the ℓ0 formulation expresses this goal but is NP-hard, and its uniqueness condition parallels single-snapshot DOA identifiability.The model uses a dictionary with M ≪ N atoms and represents the observation using only K ≪ N nonzero coefficients.Source: §3.1.1 · pp. 7–8

3.1.2 Convex Relaxation

• Basis pursuit replaces ℓ0 minimization with convex ℓ1 optimization; mutual-coherence and RIP conditions guarantee exact recovery, while under similar RIP conditions noisy recovery error is proportional to noise.Computing mutual coherence is straightforward, whereas computing the restricted isometry constant may become dramatically more complex as sparsity increases.Source: §3.1.2 · pp. 8–9

3.1.3 ℓq Optimization

• Because ℓq optimization more closely approximates ℓ0 than ℓ1, it can recover sparse signals under weaker RIP conditions than basis pursuit, but practical methods guarantee only local optimality.FOCUSS is an iterative reweighted least-squares method; tuning its regularization parameter is difficult, while SLIM iteratively updates the noise parameter and is hyper-parameter free once q is given.Source: §3.1.3 · pp. 10–11

3.1.4 Maximum Likelihood Estimation (MLE)

• MLE does not require known noise or sparsity levels, promotes sparsity as most estimated power entries approach zero, and in the noiseless limit its global optimizer coincides with ℓ0 optimization.Its main difficulty is the nonconvex log-determinant term in the objective.Source: §3.1.4 · p. 11

3.2 Sparse Representation and DOA Estimation: the Link and the Gap

• DOA estimation can be cast as sparse representation of each snapshot, with each direction corresponding to a steering-vector atom and sparsity arising from fewer sources than sensors.Some special arrays and methods can detect more sources than sensors.Source: §3.2 · p. 12
• Unlike finite-dictionary sparse recovery, DOA has continuously parameterized atoms and typically multiple snapshots, whose temporal redundancy is important when antenna counts are limited.The section identifies continuous-versus-discrete atoms as the first gap addressed by the following three sections, which also examine sparsity, temporal redundancy, and theoretical guarantees.Source: §3.2 · p. 12
• Sparse-recovery guarantees based on incoherence are difficult to apply to DOA's coherent atoms, but DOA success is measured by support error, where small support errors can be acceptable.In sparse recovery, success is measured by signal reconstruction error, and even slight support errors can cause large estimation errors.Source: §3.2 · p. 12

4.1 Data Model

• On-grid methods discretize continuous DOAs onto N ≫ M candidate angles, making snapshot vectors jointly row-sparse with source locations encoded in their support, but incurring grid mismatch.The grid must be dense enough to keep true DOAs close to grid points; the resulting model is a zeroth-order approximation, with approximation error included in the noise term.Source: §4.1 · pp. 13–14

4.2 ℓ2,0 Optimization

• Theorem 4.1 establishes unique recovery under its stated condition; more snapshots can increase recoverable DOAs through rank(Y), except when snapshots differ only by scaling, while ℓ2,0 remains NP-hard.Source: §4.2 · p. 14

4.3.1 ℓ2,1 Optimization

• ℓ2,1 performance can improve with more measurement vectors when source rows are randomly drawn at general positions, but scaled-identical snapshots add no information and noisy formulations require difficult λ tuning.The cited theoretical guarantee depends on the source-signal assumption; given noise variance, λ-selection results are available for ULA and SLA.Source: §4.3.1 · pp. 14–15

4.3.2 Dimensionality Reduction via ℓ2,1-SVD

• ℓ2,1-SVD reduces snapshots to the K-dimensional signal subspace, is argued to improve noise robustness, tolerates an estimate of K, but remains difficult to tune and compare.Its regularization parameters cannot use the noise-level tuning solutions available for standard ℓ2,1 optimization because the data structure changes.Source: §4.3.2 · pp. 15–16

4.3.3 Another Dimensionality Reduction Technique

• An alternative reduction compresses L snapshots to M while preserving ℓ2,1 solutions and row-power spectra; it enables M×M covariance decomposition and retains original parameter-tuning guidance.It keeps both signal and noise subspaces, unlike ℓ2,1-SVD, which retains only the K-dimensional signal subspace.Source: §4.3.3 · p. 16

4.4 ℓ2,q Optimization

• For 0 < q < 1, ℓ2,q is a nonconvex relaxation; M-FOCUSS iteratively solves weighted least squares, while SLIM uses a Gaussian-noise MAP formulation to update X and variance without λ tuning.The SLIM formulation assumes i.i.d. Gaussian noise.Source: §4.4 · p. 17

4.5.1 Generalized Least Squares

• Generalized least squares estimates source powers and noise from sample covariance and can yield a large-snapshot ML estimator under conditions, but its criterion is nonconvex; a convex alternative requires L ≥ M and nonsingular sample covariance.Source: §4.5.1 · p. 18

4.5.2 SPICE

• SPICE uses convex covariance-fitting criteria for nonsingular and singular sample covariance cases, then alternates C and power/noise updates; the monotonically decreasing objective supports expected convergence to a global minimum.The convex formulations are difficult to solve directly in practice because of the problem's high dimensionality.Source: §4.5.2 · pp. 18–20
• SPICE promotes row sparsity through a weighted ℓ2,1 term and joint sparsity through shared row variances, but its p and σ estimates are generally nonunique.In the single-snapshot case, SPICE is equivalent to square-root LASSO with τ = 1.Source: §4.5.2 · pp. 20–21

4.6 Maximum Likelihood Estimation

• Multiple-snapshot MLE estimates p and σ from a Gaussian likelihood with covariance R = AP Aᴴ + σI; related approaches include LIKES, SBL, and EM.The snapshots are assumed i.i.d. Gaussian with zero mean, and all share an identical sparse prior in the cited SBL approach.Source: §4.6 · p. 21

4.7 Remarks on Grid Selection

• Grid selection affects DOA accuracy, computational speed, and theoretical analysis, while dense grids increase coherence and can cause complexity, slow convergence, and numerical instability.Discrete grid points inevitably mismatch continuous DOAs, making a fully satisfactory on-grid solution difficult.Source: §4.7 · pp. 21–22
• Coherence-inhibiting methods discourage nearby atoms, while grid refinement and atom-similarity criteria offer grid-selection heuristics.The rank criterion treats a grid as dense enough when rank(Cn) ≈ 1, but it can only be evaluated heuristically.Source: §4.7 · pp. 21–22

5.1.1 Data Model

• A first-order off-grid model jointly estimates the row-sparse signal X and grid offsets β, then shifts X’s row support by β to obtain DOAs.Unlike the on-grid zeroth-order approximation, estimating β can partially compensate for grid mismatch.Source: §5.1.1 · p. 23

5.1.2 ℓ1 Optimization

• Under a suitable RIP condition on [A, B], the BPDN-like method stably recovers x and β with error proportional to η, and exactly when η = 0.The result is difficult to apply to dense grids; other ℓ1 methods include nonconvex alternating formulations with parameter-tuning challenges and a convex method [105] that permits global polynomial-time solution but cannot exploit prior information on β.Source: §5.1.2 · pp. 23–24

5.1.3 Sparse Bayesian Learning

• OGSBI uses an EM algorithm for multiple snapshots, updating β alongside signal-power and noise-variance parameters while the likelihood increases monotonically.Its expectation step computes the posterior of the row-sparse signal as in standard SBL.Source: §5.1.3 · p. 24

5.2.1 Data Model

• With a dynamic grid, jointly estimated grid points can take continuous DOA values, avoiding grid mismatch, but grid initialization and size still affect accuracy and computational speed.The joint estimation is difficult because the steering-vector mapping is nonlinear.Source: §5.2.1 · p. 24, §5.2.2 · p. 25

5.2.2 Algorithms

• Among presented off-grid sparse methods, variational EM is typically exploited; strong nonlinearity in the θ update during parameter estimation leaves no closed-form solutions, so numerical approaches are required.Most presented algorithms involve nonconvex optimization and guarantee only local convergence (except the algorithm in [105]); few theoretical guarantees are available for most algorithms (see, however,).Source: §5.2.2 · pp. 24–25

• These methods avoid direction-domain gridding, operate continuously to eliminate grid mismatch, and are convex with strong theoretical guarantees.Source: §6 · p. 25
• So far, gridless sparse methods apply only to uniform or sparse linear arrays, so the section treats DOA estimation as frequency estimation.Source: §6 · p. 25
• The section revisits ULA and SLA data models, introduces Vandermonde decomposition of Toeplitz covariance matrices, then discusses single- and multiple-snapshot methods, highlighting atomic norm and gridless SPICE.Source: §6 · p. 25

6.1 Data Model

• For an M-element ULA, the array manifold contains the steering vectors for K frequencies, each defined by complex exponentials across the array elements.The frequencies have a one-to-one relationship with the DOAs.Source: §6.1 · p. 26
• An M-element SLA retains antennas indexed by Ω from an N-element virtual ULA, where N ≥ M, and its data model follows from that virtual array.The ULA is the special case M = N with Ω = {1, . . . , N}.Source: §6.1 · p. 26
• Given the array data, the objective is to estimate the frequencies, while the number of sources K is usually unknown.For a single snapshot, this problem coincides with line spectral estimation; multiple snapshots can be handled by exploiting their joint sparsity.Source: §6.1 · p. 26

6.2 Vandermonde Decomposition of Toeplitz Covariance Matrices

• The decomposition, discovered by Carathéodory and Fejér and later used by Pisarenko for frequency retrieval [2], became a basis for subspace-based frequency and DOA methods.Source: §6.2 · p. 26
• Any PSD Toeplitz matrix of rank r admits an r-atom Vandermonde decomposition with distinct frequencies, unique when r < N.For a noiseless covariance matrix with uncorrelated sources and K < N, this decomposition permits exact frequency retrieval.Source: §6.2 · pp. 26–27
• A computational route factors T = VV^H, then uses a matrix pencil and its eigenvalues to recover frequencies and source powers, similarly to ESPRIT.The eigenvalues yield the frequencies, while the normalized eigenvectors and factor V provide the powers.Source: §6.2 · p. 28
• With homoscedastic noise, a PSD Toeplitz matrix decomposes into an atomic signal term plus σI, uniquely when σ is its smallest eigenvalue.If 0 ≤ σ < λmin(T), the decomposition is not unique.Source: §6.2 · p. 28
• Gridless sparse methods estimate a PSD Toeplitz matrix encoding the frequencies, then recover them by Vandermonde decomposition.They exploit Toeplitz structure, low rank, and positive semidefiniteness; the matrix need not satisfy the statistical assumptions required for it to be the data covariance matrix.Source: §6.2 · p. 28

6.3 The Single Snapshot Case

• The section covers deterministic atomic-norm and Hankel-based nuclear-norm methods alongside a gridless covariance-fitting version of SPICE.The deterministic methods treat the signal as deterministic and seek the sparsest candidate under a prescribed sparse metric.Source: §6.3 · p. 29

6.3.1 A General Framework for Deterministic Methods

• Deterministic methods seek a sparse signal representation without statistical assumptions, using bounded-noise constrained or regularized optimization formulations.The sparse metric is chosen to reduce the number of atoms used to express the sinusoidal signal and thereby obtain frequency estimates.Source: §6.3.1 · p. 29

6.3.2 Atomic ℓ0 Norm

• The atomic ℓ0 metric counts the fewest atoms synthesizing a signal and equals a rank-minimization objective over a PSD Toeplitz matrix.The rank problem is difficult to solve; its Toeplitz matrix encodes the frequencies, recoverable through Vandermonde decomposition.Source: §6.3.2 · pp. 29–30

6.3.3 Atomic Norm

• ANM replaces the intractable rank objective with a nuclear-norm convex relaxation, formulates the problem as an SDP, and retrieves frequencies from Toeplitz-matrix Vandermonde decomposition.The Toeplitz matrix is covariance-like, but statistical assumptions are not required.Source: §6.3.3 · p. 32
• ANM exactly recovers noiseless frequencies under a sufficient separation condition of 4/N, relaxed to 2.52/N in.The guarantee also assumes N ≥ 257, a technical requirement the authors say should pose little practical problem.Source: §6.3.3 · p. 32

6.3.4 Hankel-based Nuclear Norm

• EMaC uses the Hankel-matrix nuclear norm as a convex relaxation of rank minimization, enabling SDP formulation and recovery under a coherence condition.In the SLA case, gives exact recovery without noise and stable recovery under bounded noise when measurements exceed a constant times the number of sinusoids up to a polylog factor.Source: §6.3.4 · p. 34

6.3.5 Connection between ANM and EMaC

• ANM captures constant-modulus signal structure and directly yields frequencies, while EMaC may achieve higher resolution, including without a resolution limit in the noiseless ULA case.ANM may better capture signal structure, but theoretically requires frequency separation for successful operation; EMaC's coherence condition may be weaker than ANM's separation condition.Source: §6.3.4 · p. 34, §6.3.5 · p. 35

6.3.6 Covariance Fitting Method: Gridless SPICE (GLS)

• Introduced as gridless SPICE, GLS reparameterizes covariance with a PSD Toeplitz matrix linear in u, avoiding SPICE’s gridding for ULAs or SLAs.GLS otherwise makes the same assumptions as SPICE.Source: §6.3.6 · p. 35
• GLS covariance fitting becomes a polynomial-time SDP, and DOA parameters are recovered from the estimated full Toeplitz covariance by its Vandermonde decomposition.For an SLA, parameters are recovered from the full-data covariance estimate as in the ULA case.Source: §6.3.6 · p. 36
• GLS guarantees at most N-1 sources but can overestimate K in noise because it assumes neither the source count nor the noise variance.An approach in estimates source number from covariance-estimate eigenvalues using SORTE [148], then can refine frequencies with MUSIC.Source: §6.3.6 · p. 36

6.3.7 Connection between ANM and GLS

• GLS is equivalent to atomic-norm minimization for noisy ULA data up to scaling, and to data-consistency-constrained ANM for SLA data as if y were noiseless.A related connection in the discrete setting was also provided.Source: §6.3.7 · p. 37
• GLS requires no noise-level knowledge; unlike ANM, its frequency retrieval uses Corollary 6.1 because the Toeplitz covariance encodes both frequencies and noise variance.ANM uses Theorem 6.1 for frequency retrieval.Source: §6.3.7 · p. 37

6.4.1 Gridless SPICE (GLS)

• For L < M, GLS reduces the semidefinite matrix dimension from 2M × 2M to (M + L) × (M + L), a potentially significant saving with few snapshots.Source: §6.4.1 · p. 39
• GLS yields at most N - 1 sources and is snapshot-consistent for redundancy arrays when K ≤ N - 1, while suitable arrays can let it estimate more sources than antennas [92].With Gaussian sources and noise, GLS is asymptotic ML for redundancy arrays, and its resolution limit vanishes as the snapshot count grows [92].Source: §6.4.1 · p. 39

6.4.2 SMV-based Atomic Norm Minimization (ANM-SMV)

• ANM-SMV maps covariance entries to a virtual (2N - 1)-element ULA and applies atomic-norm recovery to estimate sufficiently separated frequencies stably.The probability of stable estimation increases with snapshot count; the stated η ∝ σ error bound assumes i.i.d. Gaussian sources and noise, sufficiently many snapshots, and a co-prime array [149].Source: §6.4.2 · p. 40

6.4.3 Nuclear Norm Minimization followed by MUSIC (NNM-MUSIC)

• NNM-MUSIC estimates a low-rank covariance matrix before applying MUSIC, but its denoising step omits Toeplitz structure and requires choosing η.With L → ∞, the method can exactly recover the covariance matrix; PSD preservation holds when the estimate T(ũ) is PSD, which is ensured with sufficiently many snapshots.Source: §6.4.3 · pp. 40–41

6.4.4 Comparison of GLS, ANM-SMV and NNM-MUSIC

• Unlike ANM-SMV and NNM-MUSIC, GLS is hyperparameter-free, usable with one snapshot, asymptotic-ML, and exactly recovers noiseless frequencies under mild separation [92,149,150].GLS and NNM-MUSIC are statistically consistent, whereas ANM-SMV is not; ANM-SMV retains a resolution limit even with infinitely many snapshots and exact virtual-array data [92,149,150].Source: §6.4.4 · p. 41

6.5.1 A General Framework

• For multiple snapshots, deterministic methods minimize a sparse metric to reduce the number of frequencies, using noiseless or bounded-noise formulations without statistical source assumptions.The bounded-noise framework uses a constraint on the noise or a regularized objective.Source: §6.5 · p. 41, §6.5.1 · p. 42

6.5.2 Atomic ℓ0 Norm

• In the noiseless case, atomic ℓ0 uniquely determines frequencies/DOAs exactly when they are identifiable; rank(Y_Ω)>1 can increase recoverable frequencies over SMV.The identifiability result also extends to general array geometries and suitably defined parameter-estimation problems.Source: §6.5.2 · p. 43

6.5.3 Atomic Norm

• The multiple-snapshot atomic norm is a convex SDP relaxation with exact frequency recovery under appropriate conditions.Its worst-case guarantees do not improve over the single-snapshot case because coherent sources are allowed, although simulations show significantly better recovery at general source positions.Source: §6.5.3 · pp. 43–44

6.5.4 Hankel-based Nuclear Norm

• M-EMaC concatenates per-snapshot Hankel matrices and uses nuclear-norm recovery, exploiting low rank when K < min(m,nL).Choosing m is challenging because the Hankel rank depends on the unknown source signals; tuning when K ≥ N/2 remains open.Source: §6.5.4 · p. 45

6.6.1 A Smooth Surrogate for ∥Z∥A,0

• As ε→0, Mε(Z) approaches the atomic ℓ0 norm and its optimizer’s smallest N−∥Z∥A,0 eigenvalues are zero or approach zero as fast as ε; as ε→∞, it approaches the atomic norm.The log-det heuristic provides a smooth rank surrogate formulated in the re-parameterized u domain. Together these limits make Mε(Z) bridge the sparsity of the atomic ℓ0 norm and the convexity of the atomic norm.Source: §6.6.1 · pp. 46–47

6.6.2 A Locally Convergent Iterative Algorithm

• The nonconvex log-det objective is majorized by its tangent plane at each iterate, yielding an SDP whose repeated solutions monotonically decrease the objective and converge to a local minimum.The tangent-plane subproblem can be globally solved at each iteration.Source: §6.6.2 · pp. 47–48

6.6.3 Interpretation as RAM

• RAM is exactly weighted atomic norm minimization: its first iteration is standard ANM, while solution-based reweighting makes current frequencies more preferred and enhances sparsity and resolution.A practical implementation can start with ANM and gradually decrease ε during the iterations.Source: §6.6.3 · pp. 48–49

6.7 Connections between ANM and GLS

• The connection between multiple-snapshot GLS and ANM extends the single-snapshot relationship and also applies to discrete ℓ2,1 optimization and SPICE.For small snapshot counts, GLS is equivalent to ANM as if there were no noise; for large counts, it is equivalent to weighted ANM.Source: §6.7 · p. 49

6.7.1 The Case of L < M

• For L < M, GLS is equivalent to an ANM formulation on transformed snapshots, with joint sparsity preserved in the limiting noiseless case.In the absence of noise, frequency recovery under a separation condition is supported for ULA, but a similar guarantee for SLA is unavailable because the required phase assumption may not hold.Source: §6.7.1 · p. 50

6.7.2 The Case of L ≥M

• For L ≥ M, GLS corresponds to weighted ANM with a weighting function given by the square root of Capon’s spectrum.The analysis says this weighting can overcome standard ANM’s resolution limit; GLS is consistent and is a large-snapshot realization of the MLE.Source: §6.7.2 · p. 51
• GLS is generally robust to source correlations like ANM, but its performance can degrade when its stronger source-statistics assumptions fail, and its power estimates can be biased.Standard ANM applies to general source signals and its resolution limit is based on worst-case analysis, whereas GLS’s statistical properties rely on stronger assumptions.Source: §6.7.2 · p. 51

6.8.1 Dimensionality Reduction

• For large L, dimensionality reduction cuts ANM snapshots from L to M while preserving u, enabling frequency and power recovery by Vandermonde decomposition.The same result holds for weighted ANM; the information in the snapshots is preserved in the sample covariance, even with coherent sources.Source: §6.8.1 · pp. 51–52

6.8.2 Alternating Direction Method of Multipliers (ADMM)

• ADMM is a first-order algorithm that guarantees global optimality, and its separable quadratic updates for X, u, and Z can be solved in closed form.The Q update projects onto the semidefinite cone by eigen-decomposition and setting negative eigenvalues to zero.Source: §6.8.2 · pp. 52–53
• O((N + L)^3) per iteration is reducible to O((N + M)^3) when L > M through dimensionality reduction; moderate accuracy is typically sufficient.The eigen-decomposition for projecting Q onto the semidefinite cone drives the per-iteration complexity.Source: §6.8.2 · p. 53

Improving speed and accuracy

• Efficient SDP solvers are needed for costly gridless sparse methods, while ℓ0-based approaches may address resolution limits.For ULAs and SLAs, ℓ0 optimization corresponds to matrix-rank minimization; nonconvex rank-minimization techniques have been explored for DOA estimation [173] [174], with recent progress in [176].Source: §7 · p. 53

Automatic model order selection

• Sparse methods need automatic model-order estimation because they can produce small spurious sources despite not requiring the source number in advance.Model order can be estimated within or after sparse estimation; also introduces parameter refinement once the order is obtained. Related results are reported in.Source: §7 · pp. 53–54

Gridless sparse methods for arbitrary array geometry

• Gridless sparse methods should be extended to general array geometries, where the Hankel/Toeplitz structures used for ULAs and SLAs do not exist.Grid-based methods apply to arbitrary geometries but suffer from grid mismatch and other problems; gridless methods avoid grid selection through the special structures available for ULAs and SLAs.Source: §7 · p. 54

Gridless sparse parameter estimation and continuous compressed sensing

• Gridless sparse methods should extend to general parameter estimation, where the continuous model’s lack of Hankel/Toeplitz structure makes development challenging.With a small parameter count K, the model supports sparsity; estimating θ_k and x_k is continuous or infinite-dimensional compressed sensing, extending compressed sensing from discrete to continuous settings.Source: §7 · p. 54

8 Conclusions

• Sparse DOA methods are classified as on-grid, off-grid, or gridless according to how they address the difference between discrete systems and continuous parameters.The review also identifies single versus multiple snapshots as a key distinction.Source: §8 · p. 54
• On-grid and off-grid methods support arbitrary array geometries but may suffer grid mismatch and weak theoretical guarantees, while gridless methods avoid these drawbacks but apply only to ULAs and SLAs.Temporal redundancy among snapshots can be exploited to handle the single-versus-multiple-snapshot distinction.Source: §8 · p. 54
• Future research should address challenging problems and compare method performance within specific applications, where preferred methods may depend on data quality and quantity.The review notes that sparse DOA methods have applications across many fields.Source: §8 · p. 54