10. Array, Multidimensional, and Graph Signal Processing

Introduction and Context

Classical digital signal processing usually treats a signal as a sequence indexed by time. This model is powerful, but it becomes inadequate when measurements are collected simultaneously at multiple locations, across several spatial dimensions, or over a network whose connections are irregular. A microphone array records pressure at many sensors; a radar antenna array observes electromagnetic wavefronts from different directions; an image is indexed by two spatial coordinates; and a smart grid or communication network produces measurements on vertices connected by edges. These signals are not merely time series. Their spatial arrangement and interconnection structure carry essential information.

Array, multidimensional, and graph signal processing extend DSP principles to these settings. The central concerns are spatial sampling, representation of structured measurements, filtering across dimensions or network links, estimation of source location, fusion of distributed observations, and learning from non-Euclidean data. The mathematical tools include steering vectors, covariance matrices, multidimensional transforms, tensor representations, graph Laplacians, graph Fourier transforms, and polynomial or spectral graph filters.

The topics are closely connected. An antenna array can be viewed as a spatial sampler of a propagating field. An image is a regularly sampled two-dimensional signal. A sensor network is a signal domain described by a graph rather than a rectangular coordinate system. In each case, the designer must determine what constitutes a sample, how neighboring samples are related, which frequencies or modes are meaningful, and how noise, aliasing, uncertainty, and computational cost affect the result.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain spatial sampling, array manifolds, steering vectors, and spatial aliasing.

2. Derive and interpret conventional beamformers and adaptive beamforming methods.

3. Compare major direction-of-arrival estimation techniques and identify their assumptions.

4. Formulate multidimensional filtering problems for images, video, and volumetric data.

5. Explain sensor fusion using statistical, geometric, and distributed approaches.

6. Define graph signals, graph shift operators, graph Laplacians, and graph spectral representations.

7. Design basic graph filters and assess their computational and statistical limitations.

8. Select appropriate methods for practical sensing, imaging, communication, and networked-data problems.

1. Spatial Sampling and Array Models

1.1 From temporal samples to spatial samples

Consider a uniform linear array (ULA) of M identical sensors separated by distance d. A narrowband plane wave of wavelength λ arriving from angle θ produces a phase difference between adjacent sensors. Under a far-field, narrowband model, the steering vector may be written as

\[

\mathbf a(\theta)= [1,e^{-jkd\sin\theta},e^{-j2kd\sin\theta},\ldots,e^{-j(M-1)kd\sin\theta}]^T,

\]

where k=2π/λ is the wavenumber. The measured array vector is commonly modeled as

\[

\mathbf x(t)=\mathbf A\mathbf s(t)+\mathbf n(t),

\]

where A contains steering vectors for the sources, s(t) contains source signals, and n(t) is sensor noise.

This model is the spatial counterpart of a sampled temporal signal. The phase progression across sensors encodes direction. Consequently, an array does not simply measure signal amplitude; it samples a wavefield at different spatial coordinates and preserves relative phase information.

1.2 Spatial sampling theorem and aliasing

For a ULA, the spatial frequency associated with a plane wave is proportional to d sin θ/λ. If d is greater than approximately λ/2, distinct arrival angles can produce indistinguishable phase progressions after spatial sampling. This phenomenon is spatial aliasing, often observed as grating lobes in the array beam pattern. Choosing d≤λ/2 over the intended angular range is a standard design rule, although practical bandwidth, scan angle, mutual coupling, and array geometry may require more careful analysis.

The analogy with temporal sampling must not be applied mechanically. In time sampling, frequency is typically unbounded before sampling; in an array, the physically relevant angular field and propagation model constrain the spatial spectrum. Nonuniform arrays can reduce coherent grating lobes but complicate calibration and processing. Sparse arrays deliberately use nonuniform positions to obtain a large effective aperture with fewer sensors, at the cost of more complicated ambiguity and reconstruction problems.

1.3 Array manifold and covariance

The collection of all possible steering vectors forms the array manifold. Its geometry depends on sensor locations, frequency, propagation medium, polarization, and calibration. For stochastic sources, the spatial covariance matrix is

\[

\mathbf R_x=E[\mathbf x\mathbf x^H]=\mathbf A\mathbf R_s\mathbf A^H+\mathbf R_n.

\]

Covariance estimation is central to beamforming and subspace methods. In practice, R_x is replaced by a sample covariance based on a finite number of snapshots. This substitution introduces estimation error, particularly when the number of snapshots is comparable to the number of sensors or when sources are correlated.

2. Beamforming: Spatial Filtering with an Array

Beamforming combines sensor outputs using complex weights:

\[

y(t)=\mathbf w^H\mathbf x(t).

\]

The spatial response is

\[

B(\theta)=\mathbf w^H\mathbf a(\theta).

\]

Thus, the weight vector determines the directions that are passed, attenuated, or rejected.

2.1 Conventional delay-and-sum beamforming

The simplest beamformer compensates for the relative delays expected from a chosen look direction θ0 and sums the signals. In narrowband form, the weights are proportional to a(θ0). The beamformer has a main lobe centered near θ0 and sidelobes elsewhere. Increasing the aperture narrows the main lobe and improves angular resolution, while increasing the number of elements can improve sidelobe control.

Delay-and-sum beamforming is robust, transparent, and easy to implement. Its limitations are fixed sidelobe behavior, sensitivity to steering errors, and limited interference suppression. Broadband systems require frequency-dependent delays or filter-and-sum structures rather than a single narrowband phase vector.

2.2 Minimum-variance distortionless response

The MVDR or Capon beamformer minimizes output power while preserving unit response in the look direction:

\[

\min_{\mathbf w}\mathbf w^H\mathbf R_x\mathbf w\quad \text{subject to}\quad \mathbf w^H\mathbf a(\theta_0)=1.

\]

Its solution is

\[

\mathbf w_{MVDR}=\frac{\mathbf R_x^{-1}\mathbf a(\theta_0)}{\mathbf a^H(\theta_0)\mathbf R_x^{-1}\mathbf a(\theta_0)}.

\]

MVDR uses the covariance structure to place nulls toward interference. However, covariance estimation error, steering mismatch, coherent multipath, and matrix ill-conditioning can produce severe degradation. Diagonal loading replaces R_x with R_x+δI, improving robustness at the expense of reduced adaptivity.

2.3 Linearly constrained and broadband beamformers

The linearly constrained minimum-variance beamformer imposes several constraints, such as unity response over a set of frequencies or nulls in known interference directions. General constraints can be expressed as C^H w=f. Such designs are useful when the desired signal occupies a bandwidth or when environmental knowledge is available.

Broadband beamforming may be implemented by true time delays, frequency-domain processing, subband filter banks, or tapped-delay lines at each sensor. The choice involves a trade-off among latency, computational load, calibration accuracy, and beam squint. Beam squint occurs when phase-shift-only steering designed at one frequency points in different directions at other frequencies.

2.4 Performance measures and design trade-offs

Important measures include beamwidth, sidelobe level, array gain, white-noise gain, interference suppression, robustness, and output signal-to-interference-plus-noise ratio. A narrow beam is not automatically superior: it may be fragile under position errors, source motion, or reverberation. In communications, beamforming also affects spatial multiplexing and user separation. In radar, waveform design and beam scheduling interact with the spatial filter.

3. Direction-of-Arrival Estimation

Beam scanning estimates source direction by evaluating beamformer output over a grid of angles. It is simple and works well at moderate signal-to-noise ratios, but its resolution is limited by the beamwidth and grid selection.

3.1 Conventional and Capon spectra

The conventional spatial spectrum is commonly defined as

\[

P_B(\theta)=\mathbf a^H(\theta)\mathbf R_x\mathbf a(\theta).

\]

The Capon spectrum is

\[

P_C(\theta)=\frac{1}{\mathbf a^H(\theta)\mathbf R_x^{-1}\mathbf a(\theta)}.

\]

Peaks in these spectra indicate likely directions. Capon processing can produce sharper peaks but is more sensitive to covariance errors.

3.2 MUSIC and subspace separation

MUSIC exploits the eigendecomposition of the covariance matrix. If the noise is spatially white, the eigenvectors associated with the smallest eigenvalues span a noise subspace En. For a true source direction, the steering vector is orthogonal to this subspace. The MUSIC pseudospectrum is

\[

P_{MUSIC}(\theta)=\frac{1}{\mathbf a^H(\theta)\mathbf E_n\mathbf E_n^H\mathbf a(\theta)}.

\]

Sharp peaks identify estimated directions. MUSIC can resolve sources more closely than conventional beam scanning under favorable conditions, but it requires knowledge or estimation of the number of sources, adequate snapshots, a valid array model, and sufficiently uncorrelated or appropriately processed sources. Spatial smoothing can restore rank in some coherent-source situations, especially for uniform linear arrays, although it reduces effective aperture.

3.3 ESPRIT and model-based estimation

ESPRIT uses rotational invariance between two overlapping subarrays. It avoids an angular grid and can estimate directions algebraically. Its assumptions include a suitable array structure and accurate sensor relationships. Maximum-likelihood methods can achieve strong statistical performance but are computationally demanding and may involve nonconvex optimization. Sparse and Bayesian methods formulate DOA estimation as a structured inverse problem and are attractive for sparse sources, though their performance depends on model selection, regularization, and computational resources.

A professional choice among these methods should consider not only nominal resolution but also calibration, source coherence, snapshot availability, computational latency, and the consequences of false peaks.

4. Multidimensional Signal Processing

4.1 Images, video, and multidimensional spectra

An image f[m,n] is a two-dimensional discrete signal. A video may be represented as f[m,n,k], with two spatial and one temporal coordinate. Medical imaging, hyperspectral sensing, synthetic aperture radar, and seismic volumes may require three or more dimensions.

For a separable two-dimensional discrete Fourier transform,

\[

F[k,l]=\sum_m\sum_n f[m,n]e^{-j2\pi(km/N+l n/M)}.

\]

The frequency pair (ω1,ω2) describes variation along two coordinate axes. Low spatial frequencies correspond broadly to slowly varying regions; high spatial frequencies often represent edges, texture, or noise. Unlike one-dimensional filtering, multidimensional frequency responses can be circular, elliptical, rectangular, directional, or otherwise shaped in frequency space.

4.2 Two-dimensional filtering

A linear shift-invariant two-dimensional filter is described by

\[

y[m,n]=\sum_k\sum_l h[k,l]x[m-k,n-l].

\]

Separable filters satisfy h[k,l]=h1[k]h2[l], reducing implementation cost because one-dimensional filtering is performed along rows and columns. Gaussian smoothing, Sobel edge operators, Laplacian sharpening, and frequency-domain low-pass filters are familiar examples. Nonseparable filters may offer better directional selectivity but require more computation or specialized hardware.

Boundary handling is not a minor implementation detail. Zero padding can create artificial edges, periodic extension can cause wraparound artifacts, and reflection or replication may better preserve local structure. The appropriate method depends on the physical meaning of the boundary.

4.3 Sampling, interpolation, and aliasing in images

Reducing image resolution without low-pass prefiltering causes aliasing: fine textures may appear as false patterns, and diagonal edges may become jagged. Interpolation reconstructs values between samples, but it cannot recover information already lost through undersampling. Bicubic and windowed-sinc methods generally provide better quality than nearest-neighbor interpolation, although the best choice depends on computational and application constraints.

Multirate image systems use analysis filters, downsampling, upsampling, and synthesis filters. Wavelet transforms provide localized scale information and support denoising, compression, and feature extraction. In video, temporal sampling introduces motion-related aliasing, while joint space-time filtering can distinguish moving structures from static backgrounds.

4.4 Tensor and separable representations

High-dimensional data are naturally represented as tensors. Tensor decompositions, such as canonical polyadic and Tucker models, can reduce storage and expose latent structure. Low-rank matrix or tensor models are useful for background subtraction, hyperspectral denoising, and missing-data completion. Their success depends on assumptions about rank, incoherence, noise, and sampling patterns; a low-rank approximation may suppress physically meaningful rare events.

5. Sensor Fusion and Distributed Processing

Sensor fusion combines measurements that differ in location, modality, resolution, reliability, or sampling rate. Fusion can occur at the raw-data level, feature level, or decision level. Early fusion preserves information but requires synchronization and compatible calibration. Decision-level fusion is more flexible but may discard useful uncertainty information.

A basic probabilistic model combines a prior p(x) with sensor likelihoods p(zi|x):

\[

p(x|z_1,\ldots,z_N)\propto p(x)\prod_i p(z_i|x),

\]

when conditional independence is justified. Kalman filtering is appropriate for linear Gaussian state-space models; extended and unscented variants address nonlinear dynamics. Particle filters represent non-Gaussian posterior distributions using weighted samples.

Distributed fusion is required when communication bandwidth, privacy, or energy prevents transmission of all measurements to a central processor. Consensus algorithms allow nodes to iteratively approach a common estimate. Distributed least squares, diffusion adaptation, and federated learning extend this principle to estimation and learning. The network topology affects convergence speed, resilience, and the propagation of erroneous information.

Fusion systems must address time synchronization, coordinate registration, sensor bias, missing data, correlated errors, and confidence calibration. Treating correlated measurements as independent leads to overconfident estimates. In safety-critical systems, uncertainty and provenance should accompany each fused result.

6. Graph Signal Processing

6.1 Signals on irregular domains

Let G=(V,E,W) be a weighted graph with N vertices. A graph signal is a vector x∈R^N or C^N whose value x_i is attached to vertex i. Edges encode relationships such as physical connectivity, communication links, geographic proximity, similarity, or functional dependence. The graph does not necessarily represent physical space; it represents a structure over which smoothness or information flow is defined.

The weighted adjacency matrix W contains edge weights. The degree matrix D is diagonal with D_ii=Σ_j W_ij. The combinatorial Laplacian is

\[

L=D-W.

\]

For an undirected graph, L is symmetric positive semidefinite. Its quadratic form is

\[

x^TLx=\frac12\sum_{i,j}W_{ij}(x_i-x_j)^2.

\]

This expression measures disagreement across connected vertices. A small value indicates graph smoothness under the chosen edge weights.

The normalized Laplacian, L_norm=D^{-1/2}LD^{-1/2}, is often useful when vertex degrees vary substantially. Directed or dynamic graphs require more careful definitions because the adjacency or Laplacian may be nonsymmetric and may not possess an orthogonal eigenbasis.

6.2 Graph Fourier transform and graph frequencies

For an undirected graph, let L=UΛU^T, where U contains orthonormal eigenvectors and Λ contains eigenvalues. The graph Fourier transform is

\[

\hat x=U^Tx,

\]

and the inverse transform is x=U\hat x. Eigenvectors associated with small eigenvalues vary slowly across strongly connected vertices and are interpreted as low graph frequencies. Large eigenvalues correspond to rapid disagreement.

This interpretation generalizes the ordinary Fourier transform, but it is not unique. Different graph shift operators produce different spectral bases. The graph itself therefore embodies modeling assumptions, and graph construction can be as consequential as filter design.

6.3 Graph filters

A spectral graph filter applies a response h(λ) to graph frequencies:

\[

y=Uh(Λ)U^Tx.

\]

A low-pass graph filter preserves signals smooth over the graph; a high-pass filter emphasizes local disagreement or anomalies. Direct eigendecomposition is expensive for large graphs. Polynomial filters approximate h(L) using

\[

y=\sum_{k=0}^{K}a_kL^kx.

\]

Because L is often sparse, this implementation requires local exchanges over K-hop neighborhoods and avoids explicit eigenvectors. Chebyshev polynomial approximations improve numerical control and permit localized filters. Rational filters and graph wavelets provide additional flexibility but may require iterative solvers.

6.4 Sampling and reconstruction on graphs

Graph sampling selects a subset of vertices from which a graph-bandlimited or approximately bandlimited signal is reconstructed. The problem is analogous to temporal sampling, but there is no universal uniform grid. Sampling quality depends on the graph spectrum, selected vertices, and signal model. Greedy methods, leverage-score strategies, and optimization-based designs seek informative node sets.

Graph signal reconstruction may minimize a smoothness objective, enforce observed samples, or use a learned generative model. If the graph is wrong, smoothness-based reconstruction can spread errors along incorrect edges. This is a central distinction from classical DSP: the domain geometry may itself be uncertain.

6.5 Graph neural networks and learning-based processing

Graph convolutional networks implement localized aggregation and learned nonlinear transformations. A typical layer combines a node's features with those of its neighbors through a graph-dependent operator. Such models are useful for traffic prediction, anomaly detection, recommendation, molecular analysis, and communication networks. They should not be treated as replacements for graph signal processing in every setting. Interpretability, oversmoothing, distribution shift, adversarial edges, and the quality of labels remain important concerns.

7. An Integrated Design Workflow

A reliable solution begins by defining the domain and the physical meaning of adjacency. For an array, specify sensor positions, carrier frequency, bandwidth, propagation assumptions, and calibration. For an image, identify spatial resolution, boundary conditions, noise statistics, and desired invariances. For a network, define vertices, edge semantics, weights, directionality, and graph dynamics.

Next, determine the signal model and the objective: localization, suppression, reconstruction, denoising, classification, prediction, or anomaly detection. Select a representation—array snapshots, covariance matrices, multidimensional transforms, tensors, or graph spectra—whose assumptions match the data. Then assess sampling adequacy, estimate uncertainty, design the filter or estimator, and validate using data not used for tuning.

Validation should include stress tests. Vary source angles, SNR, sensor failures, graph perturbations, missing samples, synchronization errors, and model mismatch. Report not only average accuracy but also resolution, false-alarm rate, latency, energy consumption, robustness, and confidence calibration.

8. Illustrative Case Study: Urban Acoustic Event Localization

Background

Consider a city authority deploying microphone arrays at several intersections to detect emergency sirens and estimate their locations. Each site contains a small planar array, while the sites are connected through a wireless network. The objective is to produce a location estimate without continuously transmitting raw audio.

Situation and problem

Traffic noise, reflections from buildings, moving vehicles, and intermittent packet loss make simple energy detection unreliable. Sirens are broadband, and their direction changes as the source moves. The system must operate with limited bandwidth and should avoid retaining unnecessary personal audio.

Analysis

Each array first performs calibrated broadband beamforming in subbands. Directional spectra from multiple subbands are combined to obtain candidate bearings. Because reverberation produces correlated arrivals, a single narrowband MUSIC estimate is not sufficient. Local confidence measures are calculated from peak separation, SNR, and covariance stability.

The intersection sites form a graph whose vertices are array locations and whose edge weights reflect geographic distance, line-of-sight reliability, and communication quality. Each site transmits bearing estimates, time stamps, and uncertainty rather than raw recordings. A central estimator or distributed consensus process intersects the bearing constraints to estimate the source position. A temporal state-space model then tracks the moving event.

Approach and solution

The processing chain uses adaptive noise reduction, robust broadband beamforming, outlier rejection, and probabilistic fusion. Measurements with inconsistent time stamps or low confidence receive reduced weight. Graph-based smoothing is applied to the sequence of site-level estimates, not blindly to raw audio, preserving privacy and reducing communication load.

Outcome

The design illustrates how several DSP domains cooperate: arrays provide spatial filtering, multidimensional or subband processing handles broadband data, probabilistic fusion combines heterogeneous estimates, and graph methods exploit relationships among sites. Performance is evaluated under sensor dropout, reverberation, and incorrect bearings rather than only under ideal simulations.

Learning points

The principal lesson is architectural. No single algorithm solves the entire problem. Physical modeling, calibration, uncertainty representation, communication constraints, privacy, and validation determine whether an apparently accurate algorithm is operationally useful.

9. Challenges, Risks, and Professional Considerations

Array methods are vulnerable to mutual coupling, sensor position errors, phase mismatch, finite snapshots, coherent multipath, and nonstationary interference. DOA estimates may be overconfident when the source-count assumption is wrong. Adaptive beamformers can accidentally suppress the desired source if the steering vector is inaccurate.

Multidimensional filters may introduce ringing, blur edges, or create boundary artifacts. Compression and low-rank models can remove rare but important structures. In sensor fusion, timestamp errors and unmodeled correlation are frequent causes of failure. In graph processing, an arbitrary or biased graph can make a mathematically smooth result physically misleading.

Ethical and governance issues arise when sensing systems capture speech, images, location, or behavioral patterns. Data minimization, encryption, access control, retention limits, auditability, and informed governance are essential. In safety-critical or public-facing applications, decisions should be explainable enough to support contestability and human oversight. Training data and graph construction can encode geographic or social bias; performance must therefore be assessed across relevant environments and populations.

10. Current Developments and Best Practices

Recent research emphasizes hybrid model-based and learning-based processing, reconfigurable intelligent surfaces, massive MIMO, integrated sensing and communications, sparse and movable arrays, self-calibrating sensor networks, graph transformers, dynamic graph neural networks, and privacy-preserving distributed learning. These developments increase capability but also intensify dependence on data quality, computational infrastructure, and reliable uncertainty estimates.

Best practice includes:

• calibrate sensor gain, phase, position, timing, and orientation before estimating spatial structure;
• match array spacing and graph construction to the physical or functional domain;
• use diagonal loading, regularization, or robust optimization when covariance estimates are uncertain;
• distinguish training performance from performance under environmental or graph shift;
• preserve uncertainty, provenance, and quality indicators through the processing chain;
• test failure modes such as missing sensors, adversarial measurements, packet loss, and model mismatch;
• choose local polynomial graph filters when scalable and interpretable computation is more important than exact spectral selectivity;
• treat privacy, security, and governance as design requirements rather than post-processing additions.

Chapter Summary

Array processing extends DSP into the spatial domain by modeling phase differences across sensors. Steering vectors and covariance matrices support beamforming, interference suppression, and direction-of-arrival estimation. Conventional, MVDR, MUSIC, ESPRIT, and sparse methods offer different balances of resolution, robustness, assumptions, and computation. Multidimensional DSP represents images, video, and volumetric measurements using multidimensional transforms, spatial filters, tensors, and multirate operations. Sensor fusion combines heterogeneous observations while accounting for uncertainty, synchronization, bias, and communication constraints.

Graph signal processing generalizes filtering and spectral analysis to irregular domains. The graph Laplacian defines a notion of smoothness, its eigenvectors provide graph-frequency modes, and polynomial filters enable scalable localized processing. Across all these domains, performance depends on the quality of sampling, geometry or graph construction, calibration, model validity, and uncertainty management. Contemporary systems increasingly combine physical models, distributed computation, and machine learning, but professional practice requires robust validation, privacy protection, and awareness of failure modes.

Key Terms

Array manifold: The set of steering vectors corresponding to possible source parameters.

Beamforming: Weighted combination of array measurements to control spatial response.

Beam squint: Frequency-dependent pointing error caused by phase-shift steering in broadband arrays.

Covariance matrix: Matrix describing second-order relationships among sensor measurements.

Direction of arrival: The estimated direction from which a wavefront reaches an array.

Grating lobe: An unwanted beam maximum caused by spatial undersampling.

Graph signal: Values assigned to the vertices of a graph.

Graph Fourier transform: Expansion of a graph signal in eigenvectors of a graph shift operator, commonly the Laplacian.

Graph Laplacian: Matrix L=D−W that quantifies disagreement across graph edges.

Graph smoothness: Degree to which connected vertices have similar signal values.

Graph filter: Operator that selectively modifies graph-frequency components.

MUSIC: A subspace-based DOA method using orthogonality between steering vectors and the noise subspace.

MVDR: An adaptive beamformer that minimizes output power subject to a distortionless response.

Spatial aliasing: Ambiguity created when spatial sampling cannot distinguish different wave directions.

Spatial smoothing: Covariance-processing technique used to mitigate coherent-source rank deficiency.

Steering vector: Complex sensor response predicted for a particular source direction or parameter.

Sensor fusion: Integration of measurements from multiple sensors or sources.

Tensor representation: Multidimensional array representation of structured data.

White-noise gain: A measure of beamformer sensitivity to spatially white noise.

Review Questions

Short Answer Questions

1. What information does a steering vector represent in a narrowband array model?

2. Why does sensor spacing greater than approximately half a wavelength create grating lobes?

3. State the optimization objective and constraint of the MVDR beamformer.

4. What is meant by graph smoothness?

5. Distinguish early fusion from decision-level fusion.

Long Answer Questions

1. Derive the MVDR beamformer and discuss the effects of covariance estimation error and diagonal loading.

2. Compare conventional beam scanning, Capon, MUSIC, ESPRIT, and maximum-likelihood DOA estimation.

3. Explain two-dimensional Fourier filtering, separable filters, boundary handling, and image aliasing.

4. Develop the graph Fourier transform from the eigendecomposition of the graph Laplacian and explain graph frequency.

5. Discuss the design of a distributed sensor-fusion system, including synchronization, uncertainty, communication, and privacy.

Analytical / Application Questions

1. A ULA operates at wavelength 0.1 m with element spacing 0.08 m. Analyze the risk of spatial aliasing and recommend a revised spacing for conventional wide-angle scanning.

2. An MVDR beamformer begins suppressing a desired source after the source moves slightly from the presumed look direction. Identify the likely causes and propose at least three robust design measures.

3. Construct a graph-based processing strategy for estimating temperature at uninstrumented nodes in an irregular sensor network. Specify the graph, signal model, sampling assumptions, reconstruction method, and validation tests.

22. Multiple Choice Questions with Answers

Question 1

For a uniform linear array, what is the principal purpose of the steering vector?

A. To describe the predicted phase and amplitude response across sensors for a specified direction

B. To remove all sensor noise without using covariance information

C. To determine the temporal bandwidth of an unrelated source

D. To replace the physical spacing between array elements

Correct Answer: A. To describe the predicted phase and amplitude response across sensors for a specified direction

Question 2

An array has element spacing greater than half the operating wavelength. Which problem is most directly expected?

A. Temporal oversampling

B. Grating lobes caused by spatial aliasing

C. Guaranteed improvement in angular resolution

D. Elimination of mutual coupling

Correct Answer: B. Grating lobes caused by spatial aliasing

Question 3

Which constraint defines the conventional MVDR beamformer in the look direction?

A. The output power must be maximized without constraints

B. The response to the look-direction steering vector must equal one

C. All sidelobes must be exactly zero

D. The covariance matrix must be diagonal

Correct Answer: B. The response to the look-direction steering vector must equal one

Question 4

MUSIC identifies a source direction primarily because the corresponding steering vector is

A. parallel to every noise eigenvector

B. orthogonal to the estimated noise subspace

C. independent of the array geometry

D. equal to the covariance matrix

Correct Answer: B. orthogonal to the estimated noise subspace

Question 5

Why can downsampling an image without low-pass prefiltering create false textures?

A. It causes spatial aliasing of insufficiently sampled high-frequency content

B. It increases the physical wavelength of light

C. It makes all image boundaries periodic by definition

D. It converts a two-dimensional signal into a graph signal

Correct Answer: A. It causes spatial aliasing of insufficiently sampled high-frequency content

Question 6

For an undirected weighted graph, a small value of x^T L x generally indicates that

A. connected vertices have large disagreements

B. the graph contains no edges

C. the signal is smooth across strongly weighted edges

D. all graph frequencies have equal energy

Correct Answer: C. the signal is smooth across strongly weighted edges

Question 7

What is a major computational advantage of a polynomial graph filter?

A. It requires no graph definition

B. It can avoid explicit eigendecomposition and use sparse local graph operations

C. It always produces a globally optimal nonlinear classifier

D. It works only on complete graphs

Correct Answer: B. It can avoid explicit eigendecomposition and use sparse local graph operations

Question 8

In probabilistic sensor fusion, treating strongly correlated measurements as conditionally independent is likely to produce

A. underconfident estimates in every case

B. overconfident estimates because information is effectively counted more than once

C. perfect synchronization

D. automatic removal of sensor bias

Correct Answer: B. overconfident estimates because information is effectively counted more than once

Question 9

A broadband array uses phase shifts designed at one frequency. What effect may occur at other frequencies?

A. Beam squint

B. Graph spectral folding

C. Tensor rank expansion by necessity

D. Guaranteed spatial whitening

Correct Answer: A. Beam squint

Question 10

Which design choice is most appropriate for a privacy-conscious networked acoustic localization system?

A. Transmit and retain all raw audio indefinitely

B. Transmit only estimates, timestamps, and uncertainty when sufficient for the task

C. Ignore calibration because graph processing corrects every error

D. Treat all site measurements as independent and equally reliable

Correct Answer: B. Transmit only estimates, timestamps, and uncertainty when sufficient for the task