Compressive Sensing

The fundamental prerequisite: the signal of interest must be sparse in some known transform domain. A wideband spectrum is sparse when only a few narrowband carriers are active within a large frequency range. The discrete Fourier transform (DFT) or wavelet basis typically provides this sparse representation. Without sparsity, reconstruction from sub-Nyquist samples is impossible. The sparsity level K—the number of active frequencies—directly determines the minimum number of measurements required.

Instead of uniform sampling, compressive sensing uses randomized measurement matrices to acquire compressed samples. Each measurement is an inner product between the signal and a pseudo-random sensing waveform. This process ensures the sampling basis is incoherent with the sparsity basis—a mathematical condition guaranteeing that information about each sparse component is spread across every measurement. Common hardware implementations include the random demodulator and randomly interleaved ADC architectures.

Recovery is not a linear inversion but a convex optimization problem. The core algorithm minimizes the L1-norm of the sparse coefficient vector subject to measurement consistency constraints. This replaces the computationally intractable L0-norm (counting non-zero elements) with a tractable surrogate. Solvers include:

• Basis Pursuit (BP)
• LASSO (Least Absolute Shrinkage and Selection Operator)
• Orthogonal Matching Pursuit (OMP) for greedy, faster approximations

The Restricted Isometry Property is the mathematical guarantee that a measurement matrix preserves the geometry of all K-sparse signals. A matrix satisfies RIP when it acts as a near-isometry on sparse vectors—no two distinct sparse signals map to the same compressed measurement. Verifying RIP for a given matrix is NP-hard, but random Gaussian and Bernoulli matrices provably satisfy it with high probability when the number of measurements M = O(K log(N/K)).

In cognitive radio, compressive sensing enables direct wideband digitization without a bank of narrowband tuners. A single compressive receiver captures the entire multi-GHz band at a rate proportional to the aggregate bandwidth of active signals, not the total span. This enables:

• Real-time spectrum occupancy mapping across bands that would require multi-GS/s ADCs conventionally
• Detection of frequency-hopping signals that hop faster than traditional scanning receivers can track
• Reduced power consumption in battery-operated spectrum monitoring nodes

The hardware paradigm shift from Analog-to-Digital Conversion (ADC) to Analog-to-Information Conversion (AIC). An AIC directly outputs compressed digital samples at the information rate rather than the Nyquist rate. Architectures include:

• Random Demodulator: mixes the signal with a pseudo-random chipping sequence, integrates, and samples at a low rate
• Modulated Wideband Converter (MWC): uses a bank of parallel mixers with periodic waveforms to achieve multi-band sensing
• Random Convolution: convolves the signal with a random pulse before uniform sub-Nyquist sampling

The core application of CS in RFML is the direct digitization of multi-GHz swaths of spectrum using sampling rates orders of magnitude below the Nyquist rate. This is achieved by exploiting the inherent sparsity of the spectrum—where only a few narrowband signals are active within a wide band. The technique replaces high-speed ADCs with analog front-ends that perform pseudo-random linear projections, capturing the essential information needed for reconstruction.

• Mechanism: Uses a random demodulator or modulated wideband converter (MWC) to mix the signal with a pseudo-random sequence before low-rate sampling.
• Benefit: Eliminates the need for prohibitively expensive and power-hungry giga-sample-per-second ADCs in cognitive radio front-ends.
• Trade-off: Shifts complexity from hardware to the software-based sparse recovery algorithm.

Reconstructing the original wideband signal from compressed measurements relies on solving an underdetermined linear system using convex optimization or greedy iterative methods. These algorithms search for the sparsest solution in a known transform domain (e.g., Fourier or wavelet).

• Basis Pursuit Denoising (BPDN): An l1-norm minimization technique that is robust to noisy measurements, widely used for spectral reconstruction.
• Orthogonal Matching Pursuit (OMP): A computationally lighter, greedy alternative that iteratively identifies the active frequency components one by one.
• Deep Unfolding: A modern approach where iterative optimization steps are mapped to neural network layers, dramatically accelerating convergence for real-time sensing.

The MWC is a practical hardware architecture that implements compressive sensing for multiband signal acquisition. It consists of a bank of parallel channels, each mixing the input signal with a distinct periodic waveform before low-pass filtering and low-rate sampling. This spreads the spectrum so that each sample contains a weighted combination of all active bands.

• Key Feature: The mixing function is a periodic square wave, easily generated by a shift register, making it practical for silicon implementation.
• CTF Block: The Continuous-to-Finite (CTF) block processes the compressed samples to construct a finite-dimensional frame from which the signal's spectral support is recovered.
• Application: Enables a single device to simultaneously monitor multiple non-contiguous frequency bands, such as LTE, Wi-Fi, and radar bands.

In a network of distributed sensors, compressive sensing extends to distributed compressive sensing (DCS) , which exploits both intra-signal and inter-signal correlation structures. Multiple low-cost sensor nodes take independent compressed measurements of the same wideband scene and transmit them to a fusion center.

• Joint Sparsity Models (JSM): Frameworks that model the common support (active frequencies) shared across all sensors, plus any individual innovations unique to a sensor's location.
• Simultaneous OMP (SOMP): A multi-measurement vector extension of OMP that recovers the common spectral support with significantly fewer measurements per node than independent recovery.
• Benefit: Dramatically reduces the data rate required for reporting from each sensor in a spectrum monitoring network, enabling dense, low-cost deployment.

Spectrum cartography constructs a power spectral density (PSD) map over a geographic area. Compressive sensing enables this by treating the spatial-spectral field as a sparse matrix, where only a few transmitters contribute to the aggregate power at any location. This allows reconstruction of a complete map from a small subset of sensor measurements.

• Matrix Completion: A technique that fills in the missing entries of a low-rank PSD matrix, where the rank corresponds to the number of active transmitters.
• Kriging with Sparsity: Combines geostatistical interpolation (Kriging) with CS principles to estimate spectrum occupancy at unobserved locations by leveraging the spatial smoothness of path loss.
• Application: Enables regulatory agencies to visualize spectrum usage across a city using only a handful of monitoring stations.

Traditional iterative recovery algorithms are often too slow for real-time cognitive radio. Deep learning models are now trained to perform the inverse mapping from compressed measurements directly to the reconstructed spectrum in a single feed-forward pass.

• Convolutional Sparse Autoencoders: Trained end-to-end to map low-dimensional measurement vectors to high-dimensional sparse spectrograms, learning the signal structure from data rather than relying on hand-crafted sparsity priors.
• Recurrent Reconstruction Networks: Use LSTMs or GRUs to process sequential compressed measurements, exploiting temporal correlation in spectrum occupancy for more accurate dynamic recovery.
• Advantage: Achieves reconstruction speeds suitable for microsecond-level spectrum agility, a critical requirement for dynamic spectrum access in 5G and future 6G networks.