Optimality of Gradient-MUSIC for Spectral Estimation (2504.06842v1)

Abstract: The goal of spectral estimation is to estimate the frequencies and amplitudes of a nonharmonic Fourier sum given noisy time samples. This paper introduces the Gradient-MUSIC algorithm, which is a novel nonconvex optimization reformulation of the classical MUSIC algorithm. Under the assumption that $m\Delta\geq 8\pi$, where $\pi/m$ is the Nyquist rate and $\Delta$ is the minimum separation of the frequencies normalized to be in $[0,2\pi)$, we provide a thorough geometric analysis of the objective functions generated by the algorithm. Gradient-MUSIC thresholds the objective function on a set that is as coarse as possible and locates a set of suitable initialization for gradient descent. Although the objective function is nonconvex, gradient descent converges exponentially fast to the desired local minima, which are the estimated frequencies of the signal. For deterministic $\ell^p$ perturbations and any $p\in [1,\infty]$, Gradient-MUSIC estimates the frequencies and amplitudes at the minimax optimal rate in terms of the noise level and $m$. For example, if the noise has $\ell^\infty$ norm at most $\epsilon$, then the frequencies and amplitudes are recovered up to error at most $C\epsilon/m$ and $C\epsilon$, respectively, which are optimal in $\epsilon$ and $m$. Aside from logarithmic factors, Gradient-MUSIC is optimal for white noise and matches the rate achieved by nonlinear least squares for various families of nonstationary independent Gaussian noise. Our results show that classical MUSIC is equally optimal, but it requires an expensive search on a thin grid, whereas Gradient-MUSIC is always computationally more efficient, especially for small noise. As a consequence of this paper, for sufficiently well separated frequencies, both Gradient-MUSIC and classical MUSIC are the first provably optimal and computationally tractable algorithms for deterministic $\ell^p$ perturbations.