MUSIC Algorithm Explained (with MATLAB + Simulator)

Practical Implementation of the MUSIC Algorithm

The focus is on how the algorithm works computationally, not just theory, and it explains the denominator (a^H E_n E_n^H a) mathematically and intuitively.

1. Introduction

The MUSIC (Multiple Signal Classification) algorithm is a high-resolution method used in signal processing and array processing to estimate the Direction of Arrival (DOA) of signals received by a sensor array.

Unlike classical beamforming methods, MUSIC uses eigenvector decomposition of the covariance matrix to separate the signal subspace and noise subspace, allowing it to achieve much higher angular resolution.

In practical implementations, MUSIC works by:

• Simulating or collecting array signals
• Computing the covariance matrix
• Performing eigenvalue decomposition
• Separating signal and noise subspaces
• Scanning possible angles using a steering vector
• Constructing a pseudo-spectrum where peaks indicate signal directions

2. Signal Model

Two complex exponential signals are generated as examples:

s1 = exp(1j*2*pi*0.05*t);
s2 = exp(1j*2*pi*0.1*t);

These represent two narrowband sources.

The signals received by the antenna array follow the standard array model:

X = A*S + N

Where:

• X → received signal matrix
• A → steering matrix
• S → source signals
• N → noise

Each column of A corresponds to a steering vector for one source direction.

3. Steering Vector

For a Uniform Linear Array (ULA), the steering vector is:

steering = exp(-1j*2*pi*d*(0:M-1)'*sin(theta));

Mathematically:

a(θ) = [
1
e^(-j2πd sin(θ))
e^(-j2π 2d sin(θ))
⋮
e^(-j2π (M-1) d sin(θ))
]

This vector represents the phase delays across sensors for a signal arriving from angle θ. Each element corresponds to the response of a specific antenna element.

4. Covariance Matrix

R = (X*X')/N;

Mathematically:

R = E[ X X^H ]

The covariance matrix contains information about:

• signal correlations
• noise characteristics
• spatial structure of the received signals

5. Eigenvalue Decomposition

R = E Λ E^H

Where:

• E → eigenvectors
• Λ → eigenvalues

Large eigenvalues correspond to signal components, while smaller ones correspond to noise.

6. Signal and Noise Subspaces

Es = Evec_sorted(:,1:K);
En = Evec_sorted(:,K+1:end);

Key property:

E_n^H a(θ_i) = 0

for true signal directions. This means the steering vector is orthogonal to the noise subspace.

7. MUSIC Spectrum Computation

The pseudo-spectrum used in MUSIC is:

P(θ) = 1 / (a^H(θ) E_n E_n^H a(θ))

Implementation:

Pmusic(i) = 1/(steering'*(En*En')*steering);

8. Meaning of the Denominator

The denominator a^H E_n E_n^H a measures the energy of the steering vector projected onto the noise subspace. Step by step:

1. Projection matrix: E_n E_n^H projects any vector onto the noise subspace.
2. Apply to steering vector: E_n^H a(θ) measures how much the steering vector lies in noise. If θ corresponds to a true signal, E_n^H a(θ) ≈ 0.
3. Energy in noise subspace: a^H E_n E_n^H a = ||P_n a(θ)||² = ||E_n^H a(θ)||².

9. Why Peaks Appear

If the tested angle equals a true signal direction:

E_n^H a(θ) ≈ 0
a^H E_n E_n^H a ≈ 0
P(θ) = 1 / 0 → large peak

These peaks indicate the Direction of Arrival (DOA).

10. Practical Interpretation

The denominator measures how much the steering vector looks like noise:

• If it looks like noise → denominator large → spectrum small
• If it looks like signal → denominator small → spectrum peak

MUSIC finds angles where the steering vector is orthogonal to the noise subspace.

11. Angle Scanning

Scan angles (e.g., -90° to 90°). For each angle:

1. Generate steering vector
2. Project onto noise subspace
3. Compute pseudo-spectrum

The result is a spatial spectrum where peaks correspond to signal directions.

12. Final Result

The final plot shows:

• x-axis → angle
• y-axis → MUSIC spectrum

Sharp peaks appear at the true source angles.

13. Key Practical Insight

The MUSIC algorithm works because:

• Signals occupy a low-dimensional subspace
• Noise occupies the remaining orthogonal subspace
• Steering vectors of real signals lie orthogonal to the noise space

The denominator a^H E_n E_n^H a acts as a test measuring whether a steering vector belongs to the signal space or noise space.

Summary

The denominator a^H E_n E_n^H a measures the projection energy of the steering vector onto the noise subspace, and MUSIC finds angles where this energy is nearly zero, indicating the presence of a signal.

Why Do We Need the Covariance Matrix and Eigen Vectors?

Received signal model:

For an array of M sensors and K narrowband sources:

X = A*S + N

Where:

• X — M×N received data matrix (N snapshots)
• A = [a(θ₁), …, a(θ_K)] — M×K steering matrix
• S — K×N source signal matrix
• N — noise, assumed white Gaussian

Covariance Matrix

We compute the sample covariance:

R = (1/N) * X * X^H

Why compute R?

• R contains all correlations between sensors.
• It separates the signal subspace (strong eigenvalues) from the noise subspace (small eigenvalues).
• MUSIC requires eigen decomposition of R to identify the noise subspace.
• Without R, we cannot identify the direction-independent noise subspace, which is crucial.

2. Eigen-decomposition

Decompose R:

R = E * Λ * E^H

Where:

• E — eigenvectors
• Λ — eigenvalues

Sort eigenvalues in descending order:

• First K eigenvectors → signal subspace E_s
• Remaining M−K eigenvectors → noise subspace E_n

3. Why the Steering Vector is Orthogonal to Noise Eigenvectors

Key property: For a true signal from angle θ_i, the steering vector lies exactly in the signal subspace:

a(θ_i) ∈ span(E_s)

Since signal and noise subspaces are orthogonal:

E_s^H * E_n = 0

Thus:

E_n^H * a(θ_i) = 0

Mathematically:

• E_n * E_n^H — projection onto noise subspace
• a^H * E_n * E_n^H * a — energy of steering vector projected onto noise

If θ = θ_i (true signal angle):

E_n^H * a(θ_i) = 0

So the denominator of the MUSIC pseudo-spectrum goes to zero, producing a peak.

 

Further Reading

1. MATLAB Code for MUSIC 
2. MUSIC Online Simulator