9.4 Adaptive beamforming

Principles of adaptive beamforming

Adaptive beamforming dynamically adjusts the directivity pattern of a sensor array to enhance desired signals while suppressing interference and noise. Unlike fixed beamformers with static weight vectors, adaptive beamformers continuously update their weights based on the received data, making them essential for environments where signal and interference characteristics change over time.

Optimizing signal reception

The primary objective is to maximize the signal-to-noise ratio (SNR) of the desired signal by adjusting the complex weights applied to each sensor element. By coherently combining signals from multiple sensors, the beamformer steers the main lobe of the array pattern toward the desired signal's direction of arrival (DOA).

This spatial filtering focuses the array's sensitivity on the target while attenuating signals from other directions. The optimization process typically involves minimizing a cost function such as the mean squared error (MSE) between the beamformer output and a reference signal:

$J(\mathbf{w}) = E\left[|d(n) - \mathbf{w}^H \mathbf{x}(n)|^2\right]$

where $d(n)$ is the reference signal, $\mathbf{w}$ is the weight vector, and $\mathbf{x}(n)$ is the received signal vector.

Suppressing interference

Adaptive beamformers suppress interfering signals by placing nulls in the array pattern toward the directions of interfering sources. This is achieved by adjusting the complex weights so that signals arriving from those directions combine destructively at the array output.

The result is a high signal-to-interference ratio (SIR) that adapts as interference conditions change. A key constraint here: the number of independent nulls an $M$-element array can place is at most $M - 1$, so the array size limits how many interferers can be simultaneously suppressed.

Adapting to changing environments

In practice, the desired signal and interferers may move, multipath conditions may shift, and signal statistics may vary over time. Adaptive beamforming algorithms handle this by continuously updating the weight vector based on incoming signal samples.

By exploiting both spatial and temporal properties of the signals, adaptive beamformers track these variations and maintain near-optimal performance. The speed at which the weights can track environmental changes depends on the algorithm's convergence rate and the rate of change in the environment.

Adaptive beamforming algorithms

The choice of algorithm for computing optimal complex weights involves trade-offs between computational complexity, convergence speed, tracking ability, and robustness. The three most important algorithms are LMS, RLS, and CMA.

Least mean squares (LMS)

LMS is the simplest and most widely used adaptive beamforming algorithm. It updates the weight vector iteratively using the instantaneous gradient of the MSE cost function.

The weight update rule is:

$\mathbf{w}(n+1) = \mathbf{w}(n) + \mu \, e^*(n) \, \mathbf{x}(n)$

where:

• $\mathbf{w}(n)$ is the weight vector at iteration $n$
• $\mu$ is the step size (learning rate)
• $e(n) = d(n) - \mathbf{w}^H(n)\mathbf{x}(n)$ is the error signal
• $\mathbf{x}(n)$ is the input signal vector

The step size $\mu$ controls the trade-off between convergence speed and steady-state misadjustment. A larger $\mu$ converges faster but produces a noisier steady-state solution. For stability, $\mu$ must satisfy:

$0 < \mu < \frac{2}{\lambda_{\max}}$

where $\lambda_{\max}$ is the largest eigenvalue of the input covariance matrix $\mathbf{R}_{xx}$.

Computational complexity per iteration: $O(M)$, where $M$ is the number of array elements.

Recursive least squares (RLS)

RLS minimizes a weighted sum of squared errors over all past data, with an exponential forgetting factor $\lambda$ (typically $0.95 < \lambda < 1$) that gives more weight to recent samples. This makes it better at tracking non-stationary environments than LMS.

The algorithm recursively updates the inverse covariance matrix $\mathbf{P}(n)$ using the matrix inversion lemma, avoiding direct matrix inversion at each step:

1. Compute the gain vector: $\mathbf{k}(n) = \frac{\mathbf{P}(n-1)\mathbf{x}(n)}{\lambda + \mathbf{x}^H(n)\mathbf{P}(n-1)\mathbf{x}(n)}$

2. Compute the error: $e(n) = d(n) - \mathbf{w}^H(n-1)\mathbf{x}(n)$

3. Update weights: $\mathbf{w}(n) = \mathbf{w}(n-1) + \mathbf{k}(n) e^*(n)$

4. Update inverse covariance: $\mathbf{P}(n) = \lambda^{-1}\left[\mathbf{P}(n-1) - \mathbf{k}(n)\mathbf{x}^H(n)\mathbf{P}(n-1)\right]$

RLS converges significantly faster than LMS (convergence rate is largely independent of the eigenvalue spread of $\mathbf{R}_{xx}$), but at a computational cost of $O(M^2)$ per iteration.

Constant modulus algorithm (CMA)

CMA is a blind adaptive beamforming technique that requires no reference signal. It exploits the constant envelope property of certain modulation schemes (PSK, FSK) by minimizing the cost function:

$J_{CM}(\mathbf{w}) = E\left[\left(|\mathbf{w}^H\mathbf{x}(n)|^2 - R_2\right)^2\right]$

where $R_2 = E[|s(n)|^4] / E[|s(n)|^2]$ is the dispersion constant determined by the signal constellation.

CMA is particularly useful when a training/reference signal is unavailable, such as in non-cooperative communications scenarios. However, it typically converges more slowly than LMS or RLS and can suffer from local minima in the cost surface.

Comparison of algorithms

Array geometry considerations

The physical arrangement of sensor elements directly affects the beamformer's spatial resolution, angular coverage, and ability to suppress interference. Different geometries suit different application requirements.

Linear arrays

Uniform linear arrays (ULAs) place sensor elements along a straight line with equal inter-element spacing $d$. They're the simplest geometry and the most commonly analyzed.

• Provide good directivity and resolution in the plane perpendicular to the array axis
• Angular resolution is approximately $\Delta\theta \approx 0.886\lambda / (Md\cos\theta)$ for an $M$-element array
• Suffer from left-right ambiguity: they cannot distinguish signals arriving from symmetric angles about the array axis
• Grating lobes appear when $d > \lambda/2$, creating spatial aliasing

Planar arrays

Planar arrays extend the concept to two dimensions, with elements arranged on a surface. Common configurations include rectangular, triangular, and hexagonal lattices.

• Enable beam steering in both azimuth and elevation, resolving signals in 3D space
• Provide improved angular coverage compared to linear arrays
• Require more sensors for equivalent resolution in each dimension
• Hexagonal lattices offer the most efficient spatial sampling (fewest elements for a given aperture and grating-lobe-free region)

Circular arrays

Circular arrays place elements uniformly along a circular contour, providing 360-degree azimuthal coverage.

• Exhibit uniform beamforming performance across all azimuthal directions (no preferred look direction)
• Well-suited for omnidirectional surveillance applications
• Limited elevation coverage without stacking multiple rings (cylindrical arrays)
• Require phase-mode excitation or more complex weight computation compared to linear arrays

Impact on beamforming performance

The optimal geometry depends on the application:

• Linear arrays are best when interference and signals arrive from a known plane and simplicity is valued
• Planar arrays are necessary for 3D spatial filtering but increase sensor count and computational load
• Circular arrays are ideal for scenarios requiring uniform azimuthal coverage

In all cases, the total number of elements $M$ determines the maximum number of independent nulls ($M - 1$) and the achievable spatial resolution.

Adaptive weight computation

Computing the optimal weight vector is the central task in adaptive beamforming. This process relies on accurate estimation of the steering vector and the covariance matrix, and its practical performance depends on convergence behavior and computational cost.

Steering vector estimation

The steering vector $\mathbf{a}(\theta)$ encodes the relative phase shifts of the desired signal across the array elements for a given DOA $\theta$. For a ULA with spacing $d$:

$\mathbf{a}(\theta) = \left[1, \, e^{j2\pi d\sin\theta/\lambda}, \, \ldots, \, e^{j2\pi(M-1)d\sin\theta/\lambda}\right]^T$

Accurate steering vector knowledge is critical. Errors in $\mathbf{a}(\theta)$ cause the beamformer to partially cancel the desired signal (a problem called signal self-nulling). Estimation approaches include:

• Direct measurement using a known calibration signal from a known direction
• Blind estimation from signal subspace decomposition (e.g., eigendecomposition of $\mathbf{R}_{xx}$)
• Robust methods that model uncertainty in $\mathbf{a}(\theta)$ as a bounded set

Covariance matrix estimation

The spatial covariance matrix $\mathbf{R}_{xx} = E[\mathbf{x}(n)\mathbf{x}^H(n)]$ captures the second-order statistics of the received data. In practice, it's estimated from $N$ snapshots:

$\hat{\mathbf{R}}_{xx} = \frac{1}{N}\sum_{n=1}^{N}\mathbf{x}(n)\mathbf{x}^H(n)$

For a reliable estimate, you need $N \geq 2M$ snapshots (a common rule of thumb). With fewer snapshots, the sample covariance matrix becomes ill-conditioned. Techniques to improve estimation quality include:

• Diagonal loading: adding $\epsilon\mathbf{I}$ to $\hat{\mathbf{R}}_{xx}$ to improve conditioning
• Shrinkage estimators: blending the sample covariance with a structured target matrix
• Exponentially weighted moving average (EWMA): for recursive updates in real-time systems

Convergence properties

Convergence behavior determines how quickly the beamformer reaches a near-optimal weight solution after initialization or after an environmental change.

Key factors affecting convergence:

• Eigenvalue spread of $\mathbf{R}_{xx}$: a large spread (ratio $\lambda_{\max}/\lambda_{\min}$) slows LMS convergence significantly, while RLS is relatively unaffected
• Step size (LMS) or forgetting factor (RLS): controls the speed-vs-stability trade-off
• SNR: higher SNR generally improves convergence quality
• Number of elements $M$: more elements means more weights to estimate, potentially requiring more data

The MSE learning curve (plotting $E[|e(n)|^2]$ vs. iteration $n$) is the standard tool for evaluating convergence behavior experimentally.

Computational complexity

Real-time operation demands that weight updates complete within the sample interval. The per-iteration costs are:

For large arrays, techniques like QR decomposition, systolic array architectures, and reduced-rank processing can bring the complexity down to manageable levels.

Robustness and limitations

Adaptive beamformers can degrade significantly when their assumptions are violated. Understanding these failure modes is essential for designing reliable systems.

Sensitivity to array imperfections

Real arrays have sensor position errors, gain/phase mismatches, and mutual coupling. These imperfections cause the actual steering vector to deviate from the assumed model, leading to:

• Degraded interference suppression
• Increased sidelobe levels
• Partial cancellation of the desired signal (signal self-nulling)

Mitigation strategies:

• Diagonal loading adds a small constant $\epsilon$ to the covariance matrix diagonal, preventing the beamformer from placing deep nulls based on imprecise steering information
• Robust Capon beamformer (RCB) models the steering vector uncertainty as $\|\mathbf{a} - \hat{\mathbf{a}}\| \leq \epsilon$ and optimizes within that constraint
• Array calibration (offline or online) estimates and compensates for element-level imperfections

Dealing with coherent signals

Coherent signals (e.g., multipath reflections of the same source) cause the signal components of the covariance matrix to become rank-deficient. This degrades both DOA estimation and beamforming performance.

Decorrelation techniques include:

• Spatial smoothing: divides the array into overlapping subarrays and averages their covariance matrices, restoring full rank at the cost of reduced effective aperture
• Forward-backward averaging: exploits the conjugate symmetry of the covariance matrix for centro-symmetric arrays
• Diagonal loading: provides a simple but less targeted fix

Adapting to wideband signals

Standard adaptive beamforming assumes narrowband signals, where the signal bandwidth $B$ satisfies $B \ll f_c$ (center frequency). For wideband signals, the array response varies across the signal bandwidth, and a single set of narrowband weights is insufficient.

Wideband beamforming approaches:

1. Frequency-domain beamforming: decompose the wideband signal via FFT into narrowband bins, apply independent narrowband beamforming to each bin, then combine
2. Tapped delay-line (TDL) beamforming: use FIR filters behind each element instead of scalar weights, providing frequency-dependent spatial filtering in the time domain
3. Wideband adaptive algorithms: extensions of LMS/RLS that operate on the TDL structure, adapting filter coefficients across the full bandwidth

Performance in non-stationary environments

When the signal environment changes faster than the algorithm can track, performance degrades. This is especially problematic for LMS with a small step size or RLS with a forgetting factor close to 1.

Techniques for improving tracking:

• Variable step size algorithms that increase $\mu$ when large errors are detected and decrease it near convergence
• Adjustable forgetting factor in RLS to weight recent data more heavily during rapid changes
• Multi-rate processing: use a fast adaptation rate for coarse tracking and a slower rate for fine optimization
• Robust beamformers with diagonal loading or norm constraints, which sacrifice some optimality for stability during transients

Applications of adaptive beamforming

Wireless communications

Adaptive beamforming is central to modern wireless systems. In cellular networks, smart antennas steer beams toward intended users while nulling co-channel interference, directly increasing system capacity and coverage.

In MIMO systems, adaptive beamforming enables spatial multiplexing (transmitting independent data streams on different spatial channels) and diversity combining. 5G massive MIMO base stations with 64+ antenna elements rely heavily on adaptive beamforming for simultaneous multi-user service.

Radar and sonar

In radar, adaptive beamforming suppresses clutter, jamming, and multipath to improve target detection. Space-time adaptive processing (STAP) jointly filters in the spatial and Doppler domains, which is particularly effective against ground clutter in airborne radar.

In sonar, adaptive beamformers localize and track underwater acoustic sources in the presence of ambient ocean noise and reverberation. The propagation environment underwater is highly variable, making adaptivity essential.

Acoustic signal processing

Microphone arrays with adaptive beamforming enhance speech in noisy environments. Applications include hands-free communication, teleconferencing systems, hearing aids, and voice-controlled smart devices.

Adaptive beamformers in these systems track a moving speaker, suppress competing talkers and background noise, and reduce reverberation effects. The Generalized Sidelobe Canceller (GSC) structure is commonly used in these applications.

Biomedical imaging

In ultrasound imaging, adaptive beamforming (e.g., minimum variance beamforming) improves lateral resolution and contrast compared to conventional delay-and-sum processing. It compensates for phase aberrations caused by tissue inhomogeneities.

In microwave imaging for breast cancer detection, adaptive beamforming exploits dielectric contrast between healthy and malignant tissue to localize tumors, offering a potential complement to mammography.

Advanced topics in adaptive beamforming

Robust adaptive beamforming

Robust methods explicitly account for model uncertainties. The two main approaches are:

• Worst-case optimization: assumes the steering vector lies within a known uncertainty set and optimizes for the worst case within that set (e.g., RCB)
• Bayesian approaches: model uncertainties probabilistically and optimize expected performance

The trade-off is always between robustness and adaptivity. More robustness means less sensitivity to model errors, but also less ability to fully exploit the spatial degrees of freedom.

Sparse array processing

Sparse arrays use non-uniform element placement to achieve more spatial degrees of freedom than the number of physical elements. Two important configurations:

• Nested arrays: combine two ULAs with different spacings to create $O(M^2)$ virtual elements from $M$ physical sensors
• Coprime arrays: use two subarrays with coprime numbers of elements and spacings to achieve similar virtual aperture extension

These configurations enable resolution of more sources than physical elements, though they require covariance-domain processing (e.g., spatial smoothing of the virtual array covariance) rather than direct snapshot-based beamforming. Sparse array beamforming is an active research area with applications in spectrum sensing, radar, and communications where physical array size or sensor count is constrained.