Field Guide · algorithm

Also known as: MUSIC, MUltiple SIgnal Classification

MUSIC (MUltiple SIgnal Classification) is a subspace method that estimates the directions of arrival of several signals impinging on an antenna array by eigen-decomposing the array’s covariance matrix and exploiting the fact that the true arrival directions are orthogonal to the noise subspace.1 Because it works with subspaces rather than a beam pattern, MUSIC resolves sources far closer together than the array’s physical beamwidth would allow — it is a super-resolution technique.

array covarianceR = E[xxᴴ] eigen-decompose signal subspacenoise subspace Eₙ angle θ → P(θ)=1/(aᴴEₙEₙᴴa)
MUSIC builds the array covariance, splits it into signal and noise subspaces, then scans a steering vector so its null-spectrum spikes at each true arrival angle.

How it works

Model M narrowband sources arriving at an N-element array (N > M). Each source adds a rank-one term to the spatial covariance matrix R = E[xxᴴ], so R has M large eigenvalues (signal + noise) and N − M small ones (noise only). The eigenvectors split into two orthogonal blocks:

  • Signal subspace — spanned by the eigenvectors of the large eigenvalues; it contains the true array-response (steering) vectors.
  • Noise subspace Eₙ — spanned by the remaining eigenvectors; crucially, every true steering vector is orthogonal to it.

MUSIC then sweeps a candidate steering vector a(θ) across all angles and plots the pseudo-spectrum P(θ) = 1 / (aᴴEₙEₙᴴa). When θ matches a real source, the denominator collapses toward zero and P(θ) shoots up as a sharp, narrow peak. The M tallest peaks are the arrival angles. The same math applied to a time-shift array estimates closely spaced frequencies (spectral MUSIC).

In practice

MUSIC needs to know (or estimate) the number of sources M, and it needs a reasonably accurate covariance estimate, which means enough snapshots and a calibrated array — element gain/phase errors and mutual coupling smear the peaks. Coherent sources (e.g. a signal and its multipath copy) collapse the signal subspace rank and must be decorrelated by spatial smoothing first. It is more expensive than beamforming because of the eigendecomposition plus the angle search, and its resolution advantage fades at low SNR.

Relevance to SDR

MUSIC is a workhorse of direction finding, radar, sonar, and channel sounding, and spectral MUSIC is used for high-resolution frequency estimation. Its main rival, ESPRIT, reaches the same subspace but skips the angle search. Both assume a multi-element phased array with coherent per-channel sampling — hardware GopherTrunk does not have: GT is a single-front-end trunking receiver, so it does no DOA estimation. MUSIC is included here as the canonical super-resolution array algorithm in the broader RF world, the direct counterpart to the beamforming that shares the same array data.

Sources

  1. MUSIC (algorithm) — Wikipedia, on subspace DOA estimation via the noise-subspace null spectrum. ↩

See also