The MUSIC-like direction-of-arrival (DOA) estimator attains MUSIC-level resolution without estimating the number of sources by solving, in each look direction, a quadratically constrained power minimization whose single tuning parameter beta was bounded and heuristically chosen under idealized assumptions (spatially white noise, uncorrelated equal-power sources, exact covariance, ideal steering vectors). Since Reddy, Ng and Khong's 2013 analysis, computational-cost and known-model correlated-noise sub-cases received follow-up within the same research group, but no quantitative statistical performance theory under realistic deviations, no tightness or calibration-robustness analysis of the admissible beta interval, and no beta-selection rule justified by such an analysis exist. The surviving open core is formulated here as one coherent objective: replace the idealized analysis and heuristic beta tuning with a realistic-condition theory and principled beta design.

Background

Direction-of-arrival (DOA) estimation infers the arrival angles of MM far-field narrowband sources from the snapshots of an LL-sensor array, most commonly a uniform linear array (ULA). The classical high-resolution route is MUSIC (Multiple Signal Classification, Schmidt 1986): the sensor covariance matrix R\mathbf{R} is eigendecomposed, and the noise subspace — the eigenvectors associated with the L−ML-M smallest eigenvalues — is used to build a spatial spectrum that peaks where the steering vectors a(θ)\mathbf{a}(\theta) (the array's phase-response vectors for direction θ\theta) are orthogonal to it. This requires the number of sources MM (the model order) to be estimated first, e.g. by the Akaike information criterion (AIC) or the minimum description length (MDL) criterion; errors in that estimate degrade MUSIC sharply, especially at low signal-to-noise ratio (SNR). The beamforming alternative, Capon's minimum-variance distortionless response (MVDR) estimator (Capon 1969), avoids model-order estimation but resolves closely spaced sources far worse. The MUSIC-like algorithm of Ying and Ng (2010) was proposed to combine the two virtues: for each look direction it solves the quadratic minimization of the steered output power wθHRwθ\mathbf{w}_\theta^H \mathbf{R}\mathbf{w}_\theta subject to the quadratic constraint wθH(a(θ)aH(θ)+βI)wθ=c\mathbf{w}_\theta^H(\mathbf{a}(\theta)\mathbf{a}^H(\theta)+\beta\mathbf{I})\mathbf{w}_\theta=c, whose solution is the minimum-eigenvalue eigenvector of the generalized eigenvalue problem (a matrix pair Aw=λBw\mathbf{A}\mathbf{w}=\lambda\mathbf{B}\mathbf{w}) for the pair (R, a(θ)aH(θ)+βI)(\mathbf{R},\,\mathbf{a}(\theta)\mathbf{a}^H(\theta)+\beta\mathbf{I}), where c>0c>0 is an arbitrary constant fixing only the normalization of wθ\mathbf{w}_\theta; the resulting spatial spectrum is P(θ)=10log⁡10(1/∣wθHa(θ)∣2)P(\theta)=10\log_{10}\big(1/|\mathbf{w}_\theta^H\mathbf{a}(\theta)|^2\big), and it achieves MUSIC-like resolution without ever segregating signal and noise subspaces. The scalar β\beta is the sole tuning parameter. Reddy, Ng and Khong (2013) analyzed this optimization: under spatially white sensor noise (R=σv2IL\mathbf{R}=\sigma_v^2\mathbf{I}_L plus signal terms) and uncorrelated equal-power sources with an exact covariance, the objective drives the weight toward the noise subspace, which explains the MUSIC-like behavior and its distinction from Capon; using the rank-one structure of a(θ)aH(θ)\mathbf{a}(\theta)\mathbf{a}^H(\theta) they derived the admissible bounds βmin⁡=max⁡θ∈ΘλR,min⁡ aH(θ)R−1a(θ)\beta_{\min}=\max_{\theta\in\Theta}\lambda_{\mathbf{R},\min}\,\mathbf{a}^H(\theta)\mathbf{R}^{-1}\mathbf{a}(\theta) and βmax⁡=min⁡θ∉ΘλR,min⁡ aH(θ)R−1a(θ)\beta_{\max}=\min_{\theta\notin\Theta}\lambda_{\mathbf{R},\min}\,\mathbf{a}^H(\theta)\mathbf{R}^{-1}\mathbf{a}(\theta), where λR,min⁡\lambda_{\mathbf{R},\min} is the smallest eigenvalue of R\mathbf{R}, aH(θ)R−1a(θ)\mathbf{a}^H(\theta)\mathbf{R}^{-1}\mathbf{a}(\theta) is the reciprocal of the Capon spectrum 1/(aH(θ)R−1a(θ))1/\big(\mathbf{a}^H(\theta)\mathbf{R}^{-1}\mathbf{a}(\theta)\big), and Θ\Theta is the set of true source directions; the admissible interval is therefore an analytical quantity, defined by a case split over source (θ∈Θ\theta\in\Theta) and non-source (θ∉Θ\theta\notin\Theta) look directions under the exact covariance of the idealized analysis, rather than a function computable from the data at a single look direction. Their selection rule β=(1−α)βmin⁡+αβmax⁡\beta=(1-\alpha)\beta_{\min}+\alpha\beta_{\max} with α=βmin⁡/βmax⁡\alpha=\beta_{\min}/\beta_{\max} was motivated only by an observed inverse dependence of the best β\beta on SNR, and gives one data-dependent value that is fixed across all look directions. The analysis and its Monte Carlo simulations, however, all rest on the idealized assumptions above; the author's own thesis records that with finite snapshots (a sample covariance built from finitely many data records) the noise covariance is no longer diagonal and the estimator's error no longer approaches MUSIC's, and lists steering-vector gain/phase calibration errors as untreated. What remains unresolved is a quantitative theory of this estimator once those idealizations are dropped, and a principled rule for choosing β\beta.

Problem Statement

For the MUSIC-like DOA estimator — the per-look-direction constrained minimization of wθHRwθ\mathbf{w}_\theta^H\mathbf{R}\mathbf{w}_\theta subject to wθH(a(θ)aH(θ)+βI)wθ=c\mathbf{w}_\theta^H(\mathbf{a}(\theta)\mathbf{a}^H(\theta)+\beta\mathbf{I})\mathbf{w}_\theta=c analyzed by Reddy, Ng and Khong (2013) — establish the theory that the 2013 analysis could only develop under idealized assumptions, as one coherent objective of replacing heuristic β\beta tuning with a realistic-condition account: (1) characterize the spatial spectrum P(θ)P(\theta) and the resulting DOA estimates (bias, resolution capability, root-mean-square error (RMSE)) under each of the following deviation classes, characterized one at a time with the remaining idealizations (spatially white noise, uncorrelated equal-power sources, exact covariance, ideal steering vectors) retained: (1a) sensor noise spatially correlated beyond the case of a known noise-correlation model; (1b) source powers that are unequal; (1c) sources that are mutually correlated but not fully coherent; (1d) fully coherent sources (perfectly correlated, as produced by multipath propagation); and (1e) R\mathbf{R} estimated from finitely many snapshots; (2) characterize how tight the admissible interval [βmin⁡,βmax⁡][\beta_{\min},\beta_{\max}] is and how it shifts or shrinks under covariance-estimation error and steering-vector errors (calibration and grid mismatch), so that the interval's practical reliability is quantified rather than assumed; and (3) derive a β\beta-selection rule (fixed, SNR-adaptive, or direction-adaptive) justified by the quantitative performance analysis of items (1) and (2), replacing the empirically motivated rule β=(1−α)βmin⁡+αβmax⁡\beta=(1-\alpha)\beta_{\min}+\alpha\beta_{\max}, α=βmin⁡/βmax⁡\alpha=\beta_{\min}/\beta_{\max}. The requested theory concerns the estimator stated above — the per-look-direction generalized eigensolution of the pair (R, a(θ)aH(θ)+βI)(\mathbf{R},\,\mathbf{a}(\theta)\mathbf{a}^H(\theta)+\beta\mathbf{I}) — rather than a redesigned algorithm. All three numbered items are required: the theory is complete only when, in addition to items (2) and (3), every deviation class (1a)–(1e) has been characterized for each of the three metrics named in item (1).

References

  1. R. O. Schmidt, "Multiple emitter location and signal parameter estimation", IEEE Transactions on Antennas and Propagation 34(3):276–280, 1986. DOI: 10.1109/TAP.1986.1143830, https://doi.org/10.1109/TAP.1986.1143830
  2. J. Capon, "High-resolution frequency-wavenumber spectrum analysis", Proceedings of the IEEE 57(8):1408–1418, 1969. DOI: 10.1109/PROC.1969.7278, https://doi.org/10.1109/PROC.1969.7278
  3. Ying Zhang and Boon Poh Ng, "MUSIC-Like DOA Estimation Without Estimating the Number of Sources", IEEE Transactions on Signal Processing 58(3):1668–1676, 2010. DOI: 10.1109/TSP.2009.2037074, https://doi.org/10.1109/TSP.2009.2037074
  4. Vinod V. Reddy, Boon Poh Ng, and Andy W. H. Khong, "Insights Into MUSIC-Like Algorithm", IEEE Transactions on Signal Processing 61(10):2551–2556, 2013. DOI: 10.1109/TSP.2013.2251337, https://doi.org/10.1109/TSP.2013.2251337
  5. Vinod Veera Reddy, "Narrowband and wideband DOA estimation with unknown number of sources", PhD thesis, Nanyang Technological University, 2013. DOI: 10.32657/10356/54656, https://doi.org/10.32657/10356/54656
  6. Vinod V. Reddy, Mohamed Mubeen, and Boon Poh Ng, "Reduced-Complexity Super-Resolution DOA Estimation with Unknown Number of Sources", IEEE Signal Processing Letters 22(6):772–776, 2015. DOI: 10.1109/LSP.2014.2368987, https://doi.org/10.1109/LSP.2014.2368987
  7. Hock Siong Lim and Boon Poh Ng, "Application of MUSIC-like processing for reduced-rank acoustic vector sensor array", 2015 IEEE International Conference on Digital Signal Processing (DSP), pp. 1176–1180, 2015. DOI: 10.1109/ICDSP.2015.7252065, https://doi.org/10.1109/ICDSP.2015.7252065
  8. Hock Siong Lim, Boon Poh Ng, and Vinod V. Reddy, "Generalized MUSIC-Like Array Processing for Underwater Environments", IEEE Journal of Oceanic Engineering, published online 2016. DOI: 10.1109/JOE.2016.2542644, https://doi.org/10.1109/JOE.2016.2542644
  9. Narong Borijindargoon and Boon Poh Ng, "Novel Directional Adaptive Relaxation Parameters for MUSIC-Like Algorithm", Progress In Electromagnetics Research Letters 81:21–28, 2019. DOI: 10.2528/pierl18112002, https://doi.org/10.2528/pierl18112002
  10. Narong Borijindargoon and Boon Poh Ng, "Directional Adaptive MUSIC-Like Algorithm Under Symmetric α-Stable Distributed Noise", Progress In Electromagnetics Research Letters 87:29–37, 2019. DOI: 10.2528/pierl19062605, https://doi.org/10.2528/pierl19062605
  11. Narong Borijindargoon, Boon Poh Ng, and Susanto Rahardja, "MUSIC-Like Algorithm for Source Localization in Electrical Impedance Tomography", IEEE Transactions on Industrial Electronics 66(6):4661–4671, 2019. DOI: 10.1109/TIE.2018.2863196, https://doi.org/10.1109/TIE.2018.2863196
  12. Sayed Zeeshan Asghar, "Estimation of the direction of arrival of unknown number of sources using compressive MMV and sparse array geometry", PhD thesis, Nanyang Technological University, 2017. DOI: 10.32657/10356/72492, https://doi.org/10.32657/10356/72492
  13. Feng-Gang Yan, Jun Wang, Shuai Liu, Bin Cao, and Ming Jin, "Computationally efficient direction of arrival estimation with unknown number of signals", Digital Signal Processing 78:175–184, 2018. DOI: 10.1016/J.DSP.2018.03.012, https://doi.org/10.1016/J.DSP.2018.03.012
  14. Vinod V. Reddy, B. P. Ng, Ying Zhang, and Andy W. H. Khong, "DOA estimation of wideband sources without estimating the number of sources", Signal Processing 92(4):1032–1043, 2012. DOI: 10.1016/j.sigpro.2011.10.015, https://doi.org/10.1016/j.sigpro.2011.10.015
  15. Juan Li, Ying Wu, and Changgang Lu, "Pipeline leak detection using the multiple signal classification-like method", Journal of Hydroinformatics 22(5):1321–1337, 2020. DOI: 10.2166/hydro.2020.194, https://doi.org/10.2166/hydro.2020.194