Ferkans — Interactive Telecom Tutor

From Beamforming to Super-Resolution

Classical beamforming (Telecom Ch 07) resolves sources separated by at least the Rayleigh limit $\Delta\theta \approx \lambda/(N_a d)$ . This is the diffraction limit of the array aperture. For imaging applications where targets are closely spaced or where the array is small, we need methods that go beyond this limit.

Subspace methods (MUSIC, ESPRIT) exploit the eigen-structure of the data covariance to achieve super-resolution — resolving sources separated by much less than the Rayleigh limit. These methods connect directly to the spectral analysis of $\mathbf{A}^{H}\mathbf{A}$ in the imaging context (Ch 07.2, Ch 13.4).

Definition:
Classical (Bartlett) Beamformer

For a ULA with $N_a$ elements at spacing $d$ , receiving $K$ narrowband sources, the classical beamformer scans the steering vector $\mathbf{a}(\theta)$ across angles:

$P_{\text{BF}}(\theta) = \mathbf{a}^{H}(\theta)\,\hat{\mathbf{R}}\,\mathbf{a}(\theta),$

where $\hat{\mathbf{R}} = \frac{1}{L}\sum_{\ell=1}^{L}\mathbf{x}_\ell\mathbf{x}_\ell^H$ is the sample covariance from $L$ snapshots.

The resolution is limited by the array factor: $\Delta\theta \approx 0.886\lambda/(N_a d)$ (Rayleigh criterion for a uniform taper).

Definition:
Capon (MVDR) Beamformer

The Capon beamformer minimizes the output power while maintaining unity gain in the look direction:

$P_{\text{Capon}}(\theta) = \frac{1}{\mathbf{a}^{H}(\theta)\,\hat{\mathbf{R}}^{-1}\,\mathbf{a}(\theta)}.$

Compared to the classical beamformer, Capon provides:

Narrower mainlobe (data-dependent resolution improvement).
Better sidelobe suppression (adapts to the interference).
Reduced dynamic range for closely spaced sources.

However, Capon is biased: it underestimates the power of strong sources due to self-nulling, and it requires accurate covariance estimation ( $L \geq N_a$ snapshots).

Theorem: MUSIC (MUltiple SIgnal Classification) Algorithm

Given $K$ narrowband sources ( $K < N_a$ ) and the eigendecomposition of the covariance matrix:

$\mathbf{R} = \sum_{i=1}^{N_a} \lambda_i \mathbf{e}_i \mathbf{e}_i^H = \mathbf{E}_s \boldsymbol{\Lambda}_s \mathbf{E}_s^H + \sigma^2\mathbf{E}_n\mathbf{E}_n^H,$

where $\mathbf{E}_s \in \mathbb{C}^{N_a \times K}$ spans the signal subspace and $\mathbf{E}_n \in \mathbb{C}^{N_a \times (N_a - K)}$ spans the noise subspace, the MUSIC pseudo-spectrum is:

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

$P_{\text{MUSIC}}(\theta) \to \infty$ when $\mathbf{a}(\theta)$ lies in the signal subspace — i.e., when $\theta$ equals a source direction. The DOA estimates are the $K$ peaks of the pseudo-spectrum.

MUSIC exploits the orthogonality between the signal and noise subspaces. The steering vector of a true source is (approximately) orthogonal to the noise subspace, making the denominator small. The sharpness of the peaks is not limited by the array aperture — hence super-resolution.

Proof

Signal model

$\mathbf{x} = \mathbf{A}\mathbf{s} + \mathbf{w}$ , where $\mathbf{A} = [\mathbf{a}(\theta_1), \ldots, \mathbf{a}(\theta_K)]$ and $\mathbf{R} = \mathbf{A}\mathbf{R}_s\mathbf{A}^H + \sigma^2\mathbf{I}$ .

Subspace decomposition

The $K$ largest eigenvalues correspond to signal + noise; the remaining $N_a - K$ eigenvalues equal $\sigma^2$ . The noise eigenvectors $\mathbf{E}_n$ satisfy $\mathbf{A}^H\mathbf{E}_n = \mathbf{0}$ (orthogonality).

Pseudo-spectrum peaks

For a true source angle $\theta_k$ : $\mathbf{a}^{H}(\theta_k)\mathbf{E}_n = \mathbf{0}$ , so $P_{\text{MUSIC}}(\theta_k) \to \infty$ . At non-source angles, the denominator is bounded away from zero.

Definition:
ESPRIT Algorithm

ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques) exploits the shift-invariance structure of a ULA to estimate DOAs without a spectral search.

Partition the ULA into two overlapping subarrays of $N_a - 1$ elements (elements $1, \ldots, N_a - 1$ and $2, \ldots, N_a$ ). The signal subspaces of the two subarrays satisfy:

$\mathbf{E}_{s,2} = \mathbf{E}_{s,1}\,\boldsymbol{\Phi},$

where $\boldsymbol{\Phi} = \text{diag}(e^{j2\pi d\sin\theta_1/\lambda}, \ldots, e^{j2\pi d\sin\theta_K/\lambda})$ .

The DOA estimates are obtained from the eigenvalues $\phi_k$ of the matrix $\boldsymbol{\Phi} = (\mathbf{E}_{s,1}^\dagger\mathbf{E}_{s,2})$ :

$\hat{\theta}_k = \arcsin\!\left(\frac{\lambda}{2\pi d}\,\text{arg}(\phi_k)\right).$

ESPRIT avoids the grid search of MUSIC and has better statistical efficiency for closely spaced sources.

Theorem: Cramer-Rao Bound for DOA Estimation

For a single source at angle $\theta_0$ observed by a ULA with $N_a$ elements, spacing $d$ , $L$ snapshots, and $\text{SNR}$ per element, the CRB for the DOA estimate is:

$\text{CRB}(\theta_0) = \frac{6\lambda^{2}}{(2\pi d \cos\theta_0)^2 \cdot 2L \cdot \text{SNR} \cdot N_a(N_a^2 - 1)}.$

The key scaling is:

CRB $\propto N_a^{-3}$ (cubic improvement with array size).
CRB $\propto 1/(L \cdot \text{SNR})$ (standard estimation scaling).
CRB $\propto 1/\cos^2\theta_0$ (degradation at endfire).

The $N_a^{-3}$ scaling reflects three sources of information: $N_a$ elements each contribute one phase measurement, the phase sensitivity scales with element index ( $\propto N_a$ ), and the effective aperture scales with $N_a$ .

Proof

Signal model

$\mathbf{x}_\ell = \alpha_\ell \mathbf{a}(\theta_0) + \mathbf{w}_\ell$ , $\mathbf{w}_\ell \sim \mathcal{CN}(\mathbf{0}, \sigma^2\mathbf{I})$ .

Fisher information

$\mathbf{J}(\theta_0) = \frac{2L \cdot \text{SNR}}{\sigma^2} \left\|\frac{\partial \mathbf{a}(\theta_0)}{\partial \theta_0}\right\|^2$ where $\frac{\partial a_n}{\partial\theta_0} = j\frac{2\pi d n \cos\theta_0}{\lambda} a_n$ .

Sum over elements

$\sum_{n=0}^{N_a - 1} n^2 = N_a(N_a - 1)(2N_a - 1)/6 \approx N_a^3/3$ . Therefore $\text{CRB} = 1/\mathbf{J} \propto N_a^{-3}$ .

Example: MUSIC DOA Estimation with Two Sources

A ULA with $N_a = 16$ , $d = \lambda/2$ , receives two equal-power uncorrelated sources at $\theta_1 = -5°$ and $\theta_2 = 5°$ with per-element $\text{SNR} = 10$ dB. Using $L = 100$ snapshots: (a) Can the classical beamformer resolve these sources? (b) Can MUSIC resolve them? (c) What is the CRB for each DOA estimate?

Solution

Rayleigh limit

$\Delta\theta_R = 0.886\lambda/(N_a d) = 0.886/(16 \times 0.5) = 0.111$ rad $= 6.3°$ . The source separation is $10° > 6.3°$ , so the classical beamformer can resolve them (barely).

MUSIC resolution

MUSIC resolution depends on SNR and snapshots, not the Rayleigh limit. At $\text{SNR} = 10$ dB with $L = 100$ , MUSIC can resolve sources separated by $\sim 1°$ - $2°$ — well within the $10°$ separation here.

CRB

$\text{CRB} = 6/(4\pi^2 \cdot 0.25 \cdot \cos^2(5°) \cdot 2 \times 100 \times 10 \cdot 16 \times 255)$ $= 6/(4\pi^2 \cdot 0.249 \cdot 2000 \cdot 4080) = 2.97 \times 10^{-7}$ rad $^2$ . RMSE $\geq \sqrt{2.97 \times 10^{-7}} = 5.5 \times 10^{-4}$ rad $= 0.031°$ .

MUSIC Pseudo-Spectrum vs. Snapshots and SNR

Compare the classical beamformer, Capon, and MUSIC spectra for two closely spaced sources. Adjust the source separation, SNR, and number of snapshots to see when each method resolves the sources.

Parameters

Source separation (deg)5

SNR (dB)10

L

(snapshots)100

N_a

16

Comparison of DOA Estimation Methods

Property	Bartlett	Capon (MVDR)	MUSIC	ESPRIT
Resolution limit	Rayleigh ( $\sim \lambda/(N_a d)$ )	Better than Rayleigh	Super-resolution	Super-resolution
Requires $K$ known?	No	No	Yes	Yes
Computation	$O(N_a^2)$ per angle	$O(N_a^3)$ once	$O(N_a^3)$ + search	$O(N_a^3)$ (no search)
Snapshot requirement	Low	$L \geq N_a$	$L \geq N_a$	$L \geq N_a$
Correlated sources	OK	Degraded	Fails (need spatial smoothing)	Fails
Calibration errors	Robust	Moderate	Sensitive	Less sensitive
Amplitude estimation	Biased (high)	Biased (low)	No	No

Historical Note: The Subspace Revolution

1979-1989

MUSIC was introduced by Ralph Schmidt in 1979 (published 1986) at the University of Cincinnati, revolutionizing direction finding by breaking the Rayleigh resolution barrier. ESPRIT followed in 1989 from Roy and Kailath at Stanford, eliminating the need for a spectral search. These algorithms emerged from the signal processing community but quickly influenced radar, sonar, and telecommunications.

The subspace idea — decomposing the data covariance into signal and noise components — is the same principle that underlies principal component analysis, sparse recovery, and the low-rank structure exploited in modern imaging algorithms.

Common Mistake: Wrong Number of Sources in MUSIC

Mistake:

Applying MUSIC with an incorrect estimate of the number of sources $K$ , expecting correct DOA estimates.

Correction:

If $\hat{K} < K$ , some sources are invisible (projected onto the noise subspace). If $\hat{K} > K$ , spurious peaks appear. Use information-theoretic criteria (AIC, MDL/BIC) or sequential hypothesis testing to estimate $K$ from the eigenvalue spectrum. MDL is preferred in radar as it is consistent (selects the correct $K$ as $L \to \infty$ ).

MUSIC

MUltiple SIgnal Classification — a subspace-based DOA estimation algorithm that achieves super-resolution by projecting the steering vector onto the noise subspace. The pseudo-spectrum peaks at true source directions.

Quick Check

The CRB for DOA estimation scales as $N_a^{-3}$ . If a radar doubles its array from 8 to 16 elements (all else equal), by how many dB does the minimum DOA estimation error (RMSE) decrease?

4.5 dB

3 dB

9 dB

6 dB

Correction:

4.5 dB

CRB $\propto N_a^{-3}$ , so RMSE $\propto N_a^{-3/2}$ . Doubling $N_a$ : RMSE decreases by $2^{3/2} = 2.83$ , or $20\log_{10}(2.83) = 9.0$ dB in power, $4.5$ dB in amplitude.

🎓CommIT Contribution(2026)

Caire's Unified Illumination and Sensing Model

G. Caire — Internal note, TU Berlin CommIT group

Caire's note unifies two communities: the imaging/diffraction tomography view (where the forward model is a Fourier-domain restriction operator on the Ewald sphere) and the radar/wireless view (where the forward model is a product of array steering vectors and frequency responses). Both views lead to the same sensing matrix $\mathbf{A}$ with Kronecker structure.

The radar signal processing tools of this chapter — matched filtering (the adjoint $\mathbf{A}^{H}$ ), range-Doppler processing (the 2D FFT of the Kronecker components), and STAP (the MVDR applied column-by-column to $\mathbf{A}$ ) — are the building blocks that Caire's model connects to the imaging algorithms of Part IV.

The key insight: the point-spread function of $\mathbf{A}^{H}\mathbf{A}$ for a physical (structured) sensing matrix has dramatically different sidelobe structure than for a random matrix. This structure determines which reconstruction algorithms succeed and which fail.

rf-imagingforward-modelsensing-operator

Key Takeaway

MUSIC and ESPRIT achieve super-resolution DOA estimation by exploiting the signal/noise subspace decomposition of the data covariance. The CRB scales as $N_a^{-3}$ — much faster than the Rayleigh limit improvement with aperture. These subspace methods generalize directly to the imaging context, where they become adaptive beamforming for image reconstruction (Ch 13.4).

Direction Finding and Angular Resolution

From Beamforming to Super-Resolution

Definition: Classical (Bartlett) Beamformer

Definition: Capon (MVDR) Beamformer

Theorem: MUSIC (MUltiple SIgnal Classification) Algorithm

Signal model

Subspace decomposition

Pseudo-spectrum peaks

Definition: ESPRIT Algorithm

Theorem: Cramer-Rao Bound for DOA Estimation

Signal model

Fisher information

Sum over elements

Example: MUSIC DOA Estimation with Two Sources

Rayleigh limit

MUSIC resolution

CRB

MUSIC Pseudo-Spectrum vs. Snapshots and SNR

Parameters

Comparison of DOA Estimation Methods

Historical Note: The Subspace Revolution

Common Mistake: Wrong Number of Sources in MUSIC

MUSIC

Quick Check

Caire's Unified Illumination and Sensing Model

Key Takeaway

Definition:
Classical (Bartlett) Beamformer

Definition:
Capon (MVDR) Beamformer

Definition:
ESPRIT Algorithm