9.6 Multiple signal classification (MUSIC) algorithm

Overview of MUSIC algorithm

The MUSIC (Multiple Signal Classification) algorithm is a high-resolution, subspace-based method for estimating parameters of multiple signals, with direction of arrival (DOA) being the most common application. It works by decomposing the input covariance matrix into signal and noise subspaces, then exploiting the orthogonality between them to pinpoint signal parameters with accuracy that far exceeds classical beamforming.

MUSIC can resolve closely spaced signals below the Rayleigh limit, which makes it a go-to technique in radar, sonar, and wireless communications. The tradeoff is higher computational cost and sensitivity to modeling errors, but several variants (Root-MUSIC, Beamspace MUSIC, Cyclic MUSIC) exist to address these issues.

Key Assumptions

MUSIC relies on a specific set of assumptions. Violating any of them can degrade performance significantly:

• Narrowband signals: The signals must be narrowband relative to the array bandwidth, so a single steering vector per direction is valid.
• Uncorrelated signals and noise: The source signals are uncorrelated with each other and with the noise. Correlated or coherent sources require preprocessing (e.g., spatial smoothing).
• White Gaussian noise: The noise is additive, zero-mean, white, and Gaussian with variance $\sigma^2$, meaning it contributes equally across all eigenvalues of the covariance matrix.
• Fewer signals than sensors: The number of signals $d$ must satisfy $d < M$, where $M$ is the number of array elements. Otherwise, the signal and noise subspaces can't be separated.
• Known array geometry: The array manifold (the set of all steering vectors) must be accurately modeled. Calibration errors directly corrupt the subspace decomposition.

Signal and Noise Subspaces

Eigendecomposition and Subspace Separation

The core of MUSIC is the eigendecomposition of the array covariance matrix $\mathbf{R}_{xx}$. For $d$ signals impinging on an $M$-element array, $\mathbf{R}_{xx}$ is an $M \times M$ matrix with $M$ eigenvalue-eigenvector pairs.

The eigenvalues split into two groups:

• Signal eigenvalues: The $d$ largest eigenvalues, each greater than $\sigma^2$. Their corresponding eigenvectors span the signal subspace $\mathbf{E}_s$.
• Noise eigenvalues: The remaining $M - d$ eigenvalues, all equal to $\sigma^2$ (in the ideal case). Their corresponding eigenvectors span the noise subspace $\mathbf{E}_n$.

These two subspaces are orthogonal to each other. Critically, the steering vectors of the true signal directions lie entirely within the signal subspace, which means they are orthogonal to the noise subspace. This orthogonality is the property MUSIC exploits.

Estimating the Number of Signals

You can estimate $d$ by inspecting the eigenvalue spread. In practice, noise eigenvalues won't be exactly equal due to finite snapshots, so you need a principled criterion:

• Akaike Information Criterion (AIC): Tends to overestimate $d$ at high SNR but works well in many scenarios.
• Minimum Description Length (MDL): More conservative; generally preferred because it's consistent (converges to the true $d$ as snapshots increase).

Both criteria balance model fit against model complexity to select the most likely number of sources.

Pseudospectrum Estimation

Constructing the MUSIC Pseudospectrum

Once you have the noise subspace $\mathbf{E}_n$, the MUSIC pseudospectrum is computed by scanning over all candidate directions $\theta$:

$P_{\text{MUSIC}}(\theta) = \frac{1}{\mathbf{a}^H(\theta) \, \mathbf{E}_n \, \mathbf{E}_n^H \, \mathbf{a}(\theta)}$

where $\mathbf{a}(\theta)$ is the steering vector for direction $\theta$. The steering vector encodes the phase shifts across the array for a plane wave arriving from $\theta$.

The procedure step by step:

1. Estimate the sample covariance matrix $\hat{\mathbf{R}}_{xx} = \frac{1}{N} \sum_{n=1}^{N} \mathbf{x}(n)\mathbf{x}^H(n)$ from $N$ snapshots.
2. Perform eigendecomposition of $\hat{\mathbf{R}}_{xx}$.
3. Estimate $d$ (using AIC, MDL, or prior knowledge) and partition eigenvectors into signal and noise subspaces.
4. For each candidate angle $\theta$, compute the steering vector $\mathbf{a}(\theta)$ and evaluate $P_{\text{MUSIC}}(\theta)$.
5. Identify the $d$ highest peaks in the pseudospectrum. These correspond to the estimated DOAs.

Interpreting the Peaks

When $\mathbf{a}(\theta)$ aligns with a true signal direction, it's orthogonal to $\mathbf{E}_n$, so the denominator approaches zero and the pseudospectrum produces a sharp peak. For directions with no signal, the projection onto the noise subspace is nonzero, keeping the pseudospectrum value low.

The term "pseudospectrum" is used deliberately: unlike a true power spectral density, the peak heights don't directly represent signal power. They indicate how well a candidate direction matches the signal subspace.

Estimating Signal Parameters

Direction of Arrival (DOA)

DOA estimates come from the peak locations in the MUSIC pseudospectrum. The angular resolution depends on:

• SNR: Higher SNR sharpens the peaks and improves separation of closely spaced sources.
• Number of snapshots $N$: More snapshots yield a better estimate of $\mathbf{R}_{xx}$, which tightens the peaks.
• Array geometry: Larger aperture and more elements improve resolution.

For finer DOA estimates beyond the grid resolution used in the pseudospectrum scan, you can apply interpolation around the peak or switch to Root-MUSIC (described below), which avoids the grid entirely.

Number of Signals

As discussed above, the eigenvalue structure reveals the number of sources. In the ideal (infinite snapshot, exact model) case, the $M - d$ smallest eigenvalues are exactly $\sigma^2$. With real data, you rely on AIC or MDL to make this determination robustly.

Advantages Over Other Methods

High Resolution

MUSIC provides angular resolution that is not fundamentally limited by the array aperture the way classical methods are. Traditional beamformers like Bartlett (conventional) and even Capon (MVDR) are constrained by the beamwidth of the array. MUSIC breaks through the Rayleigh resolution limit because it relies on subspace orthogonality rather than beamwidth.

Resolving Closely Spaced Signals

Two signals separated by less than the Rayleigh limit will merge into a single peak under conventional beamforming. MUSIC can still resolve them as long as:

• The signals are uncorrelated (or spatial smoothing is applied for correlated sources).
• The SNR is sufficient.
• Enough snapshots are available for an accurate covariance estimate.

This capability comes directly from the subspace decomposition: even closely spaced steering vectors project differently onto the noise subspace, producing distinct peaks.

Limitations and Drawbacks

Sensitivity to Array Imperfections

MUSIC assumes the array manifold is perfectly known. In practice, gain/phase mismatches, element position errors, and mutual coupling distort the steering vectors. This causes the true signal steering vectors to no longer be perfectly orthogonal to the estimated noise subspace, leading to biased DOA estimates or spurious peaks.

Mitigation strategies include:

• Array calibration: Measuring and compensating for element-level errors.
• Autocalibration: Jointly estimating DOAs and calibration parameters.
• Array interpolation: Mapping an imperfect array response onto an ideal virtual array.

Computational Complexity

The two main computational costs are:

1. Eigendecomposition of the $M \times M$ covariance matrix: $O(M^3)$.
2. Pseudospectrum evaluation over a fine angular grid: cost scales with the number of grid points and $M$.

For large arrays or real-time applications, this can be prohibitive. Beamspace MUSIC and subspace tracking algorithms reduce the dimensionality and update cost, respectively.

Correlated Sources

Standard MUSIC fails when sources are correlated or coherent (e.g., multipath). The signal subspace dimension effectively shrinks, and the algorithm can't distinguish the sources. Spatial smoothing (subdividing the array into overlapping subarrays and averaging their covariance matrices) is the standard fix, at the cost of reduced effective aperture.

Variants and Extensions

Root-MUSIC

Root-MUSIC reformulates the pseudospectrum peak search as a polynomial rooting problem. For a uniform linear array (ULA), the steering vector has a Vandermonde structure, which lets you express the MUSIC null-spectrum denominator as a polynomial in $z = e^{j\phi}$. The DOAs are estimated from the roots of this polynomial that lie closest to the unit circle.

• Advantage: No angular grid is needed, so you avoid grid-limited resolution and reduce computation.
• Limitation: Directly applicable only to ULAs. Extension to arbitrary geometries requires array interpolation to map onto a virtual ULA.

Cyclic MUSIC

Cyclic MUSIC replaces the standard covariance matrix with the cyclic autocorrelation matrix at a specific cycle frequency $\alpha$. Many communication signals exhibit cyclostationarity (periodic statistical properties due to carrier frequency, symbol rate, etc.), and exploiting this structure provides two benefits:

• Signals with different cycle frequencies can be separated even if they overlap spatially.
• Stationary noise and interference that lack cyclostationarity are suppressed.

This variant is particularly useful in communication environments with modulated signals.

Beamspace MUSIC

Beamspace MUSIC first projects the array data through a beamforming matrix $\mathbf{B}$ (typically a DFT-based or Butler matrix) into a lower-dimensional space focused on the angular sector of interest.

• Reduced dimensionality: The eigendecomposition operates on a smaller matrix, cutting computation.
• Improved robustness: By discarding out-of-sector energy, noise and interference outside the region of interest are attenuated.
• Best suited for: Large arrays where the number of signals is much smaller than the number of elements.

Applications of MUSIC

Radar and Sonar

MUSIC is widely used for target localization and tracking in both radar and sonar. It can estimate the DOA of multiple targets in the presence of clutter and interference, and it's often combined with Doppler processing to jointly estimate angle and velocity.

Wireless Communications

In cellular and massive MIMO systems, MUSIC enables DOA estimation of multiple users, supporting spatial multiplexing and interference management. Smart antenna systems use MUSIC-based DOA estimates to steer beams toward desired users and place nulls toward interferers.

Seismology and Geophysics

Seismic arrays use MUSIC to locate earthquake sources and estimate focal mechanisms. In exploration geophysics, MUSIC-based processing of seismic reflection and refraction data helps image subsurface layers and detect geological features like oil and gas reservoirs.