Spectral MUSIC for curve-constrained source localization

Abstract Multiple Signal Classification (MUSIC) can degrade under model-order mismatch and low signal-to-noise ratio, producing deterministic ghost artifacts and clutter that masks weak sources. We propose a spectral MUSIC variant that incorporates eigenvalue information by weighting eigenvector projections with fractional powers of the associated MSR eigenvalues. We prove that this spectral weighting preserves MUSIC’s asymptotic quadratic-order resolution, while improving robustness and detection contrast. Numerically, we test circle, ellipse, and non-convex kite geometries. In the kite experiment with intentional model-order underestimation, spectral MUSIC suppresses spurious boundary detections and lowers the mean boundary-noise floor under our curve-based noise-floor metric. In a high-noise, weak-source sensitivity test, spectral MUSIC reduces spurious local maxima above the weak-source sensitivity threshold, and a simple voting rule across independent Monte Carlo realizations further suppresses residual clutter.