Compressive Spectrum Sensing

The foundational principle enabling sub-Nyquist operation. A wideband spectrum is considered sparse because only a small fraction of frequencies are actively occupied by primary users at any given moment. Compressive sensing leverages this by acquiring incoherent linear measurements via random demodulation or non-uniform sampling. The sampling rate is dictated by the information rate (the number of active signals) rather than the total bandwidth, often achieving a compression ratio of 5:1 to 10:1 in practical implementations.

The hardware architecture that physically implements compressive acquisition. An AIC front-end modulates the wideband analog signal with a pseudo-random chipping sequence operating at the Nyquist rate, integrates the product, and samples the output at a low rate. Architectures include:

• Random Demodulator (RD): Multiplies the signal by a ±1 sequence before integration.
• Modulated Wideband Converter (MWC): Uses a periodic waveform and multiple parallel channels to handle multiband signals.
• Random Triggering: Non-uniform time-domain sampling using a randomized clock.

The computational engine that recovers the sparse spectral estimate from the compressed measurements. Unlike convex optimization, greedy methods iteratively identify the dominant frequency components. Key algorithms include:

• Orthogonal Matching Pursuit (OMP): Iteratively selects the dictionary atom most correlated with the residual signal, then projects the measurement onto the span of all selected atoms.
• Compressive Sampling Matching Pursuit (CoSaMP): A more robust variant that selects multiple atoms per iteration and prunes the support set, offering stronger theoretical guarantees.
• Iterative Hard Thresholding (IHT): A gradient-based approach that applies a hard thresholding operator at each step to enforce sparsity.

The mathematical condition that guarantees robust signal recovery. A sensing matrix A satisfies the RIP of order K if it approximately preserves the Euclidean length of all K-sparse vectors. Formally, there exists a restricted isometry constant δ_K ∈ (0,1) such that the energy of the compressed measurements is bounded relative to the original signal energy. Matrices formed by random Gaussian, Bernoulli, or partial Fourier ensembles satisfy the RIP with high probability, ensuring that the reconstruction problem is well-posed and stable in the presence of measurement noise.

A direct operational benefit of compressive sensing for cognitive radio. Traditional wideband scanning requires sequentially tuning a narrowband receiver across hundreds of channels, incurring a scanning delay proportional to the number of channels. Compressive sensing collapses this to a single acquisition window, reducing the minimum sensing time from milliseconds to microseconds. This enables the detection of frequency-agile emitters and short-duration transmissions that would be missed by sequential sweep architectures, directly improving the sensing-throughput tradeoff.

An extension of compressive sensing that exploits correlation across multiple sensing nodes in a cooperative network. When multiple cognitive radios observe the same wideband spectrum, their spectral occupancy vectors share a common support set (the indices of occupied frequencies) due to spatial proximity. Distributed Compressive Sensing (DCS) frameworks, such as the Simultaneous Orthogonal Matching Pursuit (SOMP) algorithm, jointly reconstruct the spectrum by fusing compressed measurements from all nodes, achieving a higher probability of detection at a given sampling rate than independent recovery.

The foundational principle of sampling a signal at a rate significantly below the Nyquist-Shannon criterion (twice the maximum frequency). This is feasible because the wideband spectrum is sparse—only a few narrowband signals are active simultaneously. Analog-to-Information Converters (AICs) implement this in hardware, capturing compressed measurements directly rather than full-rate samples. The sampling rate is proportional to the information content (sparsity level), not the total bandwidth.

A mathematical domain where the wideband signal has a concise, non-zero representation. The signal x is sparse if x = Ψs, where s is a vector with only K non-zero coefficients. Common bases include:

• Fourier basis: For signals sparse in frequency (tonal components).
• Wavelet basis: For piecewise smooth signals.
• Gabor frames: For time-frequency localized signals. The choice of basis directly impacts reconstruction accuracy and the required number of measurements.

The construction of the sensing matrix Φ (M × N, where M << N) used to acquire compressed samples y = Φx. The matrix must satisfy the Restricted Isometry Property (RIP) to ensure stable reconstruction. Common constructions include:

• Random Gaussian/Bernoulli matrices: Universal, RIP-compliant with high probability.
• Random Demodulator: Multiplies the signal by a pseudo-random ±1 sequence, integrates, and samples—efficient for analog implementation.
• Modulated Wideband Converter (MWC): A multi-branch architecture using periodic waveforms to mix and alias the spectrum into baseband.

The core convex optimization algorithm for reconstructing the sparse signal s from compressed measurements y. Instead of solving the NP-hard L0-norm problem (min ||s||₀), Basis Pursuit (BP) solves: min ||s||₁ subject to y = ΦΨs. This is a linear program that provably recovers the exact K-sparse solution under RIP conditions. LASSO extends this for noisy measurements by minimizing ||y - ΦΨs||₂² + λ||s||₁, balancing data fidelity with sparsity.

Iterative, low-complexity alternatives to convex optimization for sparse signal recovery. These algorithms build the support set (non-zero indices) incrementally:

• Orthogonal Matching Pursuit (OMP): Selects the column of ΦΨ most correlated with the residual, then projects the measurement onto the selected subspace.
• Compressive Sampling Matching Pursuit (CoSaMP): Selects multiple candidates per iteration and prunes back, offering stronger theoretical guarantees.
• Iterative Hard Thresholding (IHT): A gradient-descent approach with a hard-thresholding step to enforce sparsity. Preferred for real-time applications due to speed.

A probabilistic framework that models the sparse signal using a hierarchical prior, typically a spike-and-slab or Laplace prior. Instead of a point estimate, it computes the full posterior distribution of the signal, providing uncertainty quantification alongside the reconstruction. Sparse Bayesian Learning (SBL) uses an expectation-maximization algorithm to estimate hyperparameters, automatically determining the sparsity level K without requiring it as an input. This is particularly robust to coherent dictionaries where L1 methods struggle.