Source-linked AI summary
SR-ANM: A Square-Root Atomic Norm Minimization Estimator for Noise-Robust Direction-of-Arrival Estimation
Youval Klioui
TL;DR
In SMV DoA estimation, ANM's tuning parameter depends on noise level, which is difficult to estimate from a single measurement vector. The paper proposes SR-ANM, using an L2-norm residual, and reports asymptotic efficiency without requiring prior knowledge of the noise standard deviation.
Problem
ANM's optimal tuning parameter typically depends on noise level, which is difficult to estimate from a single measurement vector in SMV DoA estimation.
Source: §I · p. 1Method
Results
SR-ANM achieves the same asymptotic efficiency for (ρ, β) = (1, 0.5) and (0.1, 0.5), corresponding to κ = (log M_tilde)^(1/2)/2 without prior knowledge of σ.
Source: B. Results Discussion · p. 5Takeaways & Limitations
Experiments confirm SR-ANM is asymptotically efficient without prior knowledge of the noise standard deviation.
Source: §V · p. 5
Abstract
from arXiv · showhide
This paper proposes a Square-root Atomic Norm Minimization (SR-ANM) estimator for direction-of-arrival (DoA) estimation in the single measurement vector (SMV) setting. This new framework, inspired by the Square-root Least Absolute Shrinkage and Selection Operator (Square-root LASSO), relies on a residual term based on the $L_{2}$-norm rather than the usual squared $L_{2}$-norm so as to obtain an estimator whereby the hyperparameter of the atomic norm in the unconstrained convex optimization problem exhibits robustness to the noise level compared to classical ANM. Experimental verification across a wide signal-to-noise ratio (SNR) range from 0 dB to 40 dB using the the probability of detection and the angular root-mean-squared error (RMSE) confirms the increased noise robustness of the proposed framework compared to ANM.
I. INTRODUCTION
ANM avoids grid mismatch in SMV direction-of-arrival estimation, but its tuning parameter depends on noise level, which is difficult to estimate from a single measurement vector. The paper proposes SR-ANM, an L2-residual extension intended to reduce that dependence, and evaluates its noise robustness against ANM.
I. INTRODUCTION
- ANM addresses grid mismatch, but its tuning parameter typically depends on noise level, which is difficult to estimate from one measurement vector.For sparse arrays, the stated per-iteration ANM SDP complexity is at least O(M_tilde^3), making cross-validation prohibitively expensive.Source: §I · p. 1
- Square-root LASSO motivates replacing the squared L2 residual with an L2 residual to weaken the tuning parameter's dependence on noise.The paper extends this approach from LASSO to atomic norm minimization.Source: §I · p. 1
- SR-ANM uses an L2-norm residual to make the atomic-norm hyperparameter more robust to noise level than in ANM.The paper also characterizes the optimality conditions satisfied by the global solution.Source: §I · p. 1
- Experiments evaluate detection probability and angular RMSE from 0 dB to 40 dB SNR to compare SR-ANM's noise robustness with ANM.The paper also provides closed-form ADMM updates for solving the SR-ANM optimization problem.Source: §I · p. 1
II. ANM AND SQUARE-ROOT LASSO
This section reviews DoA estimation for a subsampled array and contrasts grid-based LASSO with continuous-domain ANM. It explains why the noise-level dependence of regularization motivates the Square-root LASSO and describes ANM’s atomic-norm formulation.
II. ANM AND SQUARE-ROOT LASSO
- The DoA setup uses an M-element array subsampled from a ˜M-element ULA and models a single measurement vector with Gaussian noise.LASSO represents candidate angles on a discrete grid, with a sparse coefficient vector whose support estimates the source angles.Source: §II · p. 2
- LASSO’s regularization strength κ depends directly on the noise level σ, complicating its selection when σ is unknown.Under pure noise, the minimum κ for a null solution is σ-dependent.Source: §II · p. 2
- For ANM under pure noise, the minimum κ threshold remains dependent on σ.The passage characterizes this dependence as linear under the stated condition that the relevant design-matrix quantity is independent of σ.Source: §II · p. 2
- Square-root LASSO replaces the squared L2 residual with an L2 residual, making the pure-noise threshold for κ independent of σ.The threshold instead depends primarily on the design matrix and the realization of the standard complex Gaussian noise vector.Source: §II · p. 2
- ANM uses a continuous atomic set to avoid grid mismatch, and its atomic norm has an equivalent semidefinite-program formulation.For a subsampled array, the SDP uses a Toeplitz-Hermitian matrix; the ANM hyperparameter scales linearly with σ, with κ = σ(M log ˜M)^1/2 given as an asymptotically efficient setting.Source: §II · pp. 2–3
III. SQUARE-ROOT ANM
SR-ANM replaces the squared L2 residual with an L2 residual in atomic norm minimization. Its pure-noise null-solution threshold for κ is independent of σ, unlike ANM’s, and it is solved with ADMM.
III. SQUARE-ROOT ANM
- SR-ANM uses an L2-norm residual in a convex atomic norm minimization problem, extending the Square-root LASSO approach.The paper gives optimality conditions for solutions to this problem.Source: §III · p. 3
- Under pure noise, SR-ANM’s minimum κ for the null solution is independent of σ, while ANM’s threshold depends on σ.The SR-ANM threshold instead depends on atomic-set geometry, subsampling, and the noise realization.Source: §III · p. 3
- The ANM threshold that guarantees asymptotic performance is derived from an upper bound on the atomic-norm dual norm.The supplied passage does not state the bound’s value or its assumptions.Source: §III · p. 3
- ADMM updates for SR-ANM have O(M̃^3) computational complexity, identical to ANM, with cost mostly from the PSD projection.The formulation decouples the Toeplitz-Hermitian constraint from the positive-semidefinite constraint.Source: §III · pp. 3–4
A. Experimental Setup
The experiments compare ANM and SR-ANM on subsampled-array DoA estimation over varied SNR and regularization settings. Detection rate and angular RMSE are evaluated across Monte Carlo trials.
A. Experimental Setup
- The experiments use M = 20 elements subsampled from a ULA with M̃ = 31 elements and generate K = 5 random source angles.The angles are drawn from U(−90°, 90°) with minimum normalized-frequency separation 1/M; source amplitudes are randomized.Source: A. Experimental Setup · p. 4
- For each trial, Root-MUSIC estimates angles from the ANM- or SR-ANM-derived Toeplitz-Hermitian matrix, using the known model order K.A true angle counts as detected when its estimate lies within 10°; angular RMSE is computed only for trials with at least one detection.Source: A. Experimental Setup · p. 4
- SNR ranges from 0 dB to 40 dB in 5 dB steps, with R = 1000 Monte Carlo trials per SNR.ANM uses κ = βσ(M log M̃)^1/2, while SR-ANM uses κ = β(log M̃)^1/2.Source: A. Experimental Setup · p. 4
- Figure 1 presents detection rate Pd against κ, ADMM parameter ρ, and SNR for ANM and SR-ANM.The panels show ANM in (a) and SR-ANM in (b).Source: Figure 1 · p. 5
B. Results Discussion
Both ANM and SR-ANM attain the reported asymptotic efficiency under tested parameter settings. SR-ANM does so with a κ choice that does not require prior knowledge of σ.
B. Results Discussion
- At (ρ, β) = (1, 0.5) and (0.1, 0.5), SR-ANM achieves the same asymptotic efficiency with κ = (log M̃)^1/2/2, without prior knowledge of σ.The result is shown in Fig. 1b and Fig. 2b.Source: B. Results Discussion · p. 5
- ANM achieves asymptotic efficiency at ρ = 0.1 and β = 1, corresponding to κ = σ(M log M̃)^1/2.The result is shown in Fig. 1a and Fig. 2a.Source: B. Results Discussion · p. 5
- ANM also reaches the same efficiency at β = 0.5, so its stated minimum regularizer for guaranteeing efficiency is sufficient but not necessary.This conclusion is reported alongside the Fig. 1a and Fig. 2a results.Source: B. Results Discussion · p. 5
V. CONCLUSION
The conclusion presents SR-ANM as a gridless DoA estimator that combines atomic norm regularization with an L2 residual. Experiments confirm asymptotic efficiency without prior knowledge of the noise standard deviation.
V. CONCLUSION
- SR-ANM combines atomic norm regularization with an L2-norm residual for gridless sparse recovery in DoA estimation.The paper characterizes the approach as increasing the regularization parameter’s robustness to noise level.Source: §V · p. 5
- Experimental verification confirms that SR-ANM is asymptotically efficient without prior knowledge of the noise standard deviation.Source: §V · p. 5
- The conclusion frames noise-robust regularization and asymptotic efficiency as the supported scope of the proposed estimator.Source: §V · p. 5