A cognitive radio that hopes to use idle spectrum must first find it, and the search has to cover a band far wider than any single transmission (Haykin, 2005). Sampling several gigahertz at the Nyquist rate calls for converters whose power consumption and effective resolution are poor, and it produces a flood of samples that mostly describe empty spectrum. The observation that licensed bands are sparsely occupied at any place and time suggested another route. If only a few sub-bands carry energy, the spectrum is a sparse vector, and compressed sensing shows that a sparse vector can be recovered from far fewer linear measurements than its length (Donoho, 2006; Candès et al., 2006). Tian and Giannakis (2007) made the connection to cognitive radio explicit, and a large literature on compressive wideband sensing followed.

The guarantees behind these methods are stated for measurement noise of bounded energy. In that setting a least-squares fidelity term is the natural choice, and it coincides with the maximum-likelihood criterion when the noise is Gaussian. Radio environments often break the assumption. Man-made interference from switching devices and atmospheric discharges produces bursts whose amplitude can exceed the background by orders of magnitude, a fact documented in the electromagnetic compatibility literature long before compressed sensing existed (Middleton, 1977). A squared error gives such samples enormous weight, and because each compressive measurement mixes the whole band, one corrupted sample spreads its damage across every recovered frequency bin.

This article reviews what has been published on recovering sparse wideband spectra from sub-Nyquist samples when the noise is impulsive. It is addressed to readers who know the basics of compressed sensing and want an organised account of the robust variants, the assumptions each makes, and the questions that are still unsettled.

The article is a narrative review of the published literature. It draws on the foundational papers of compressed sensing and sub-Nyquist sampling, the standard texts on robust statistics and heavy-tailed signal models, and the papers that first combined the two. No claim is made to exhaustive coverage. The aim is to explain the main lines of thought and how they relate, so that sources were chosen for their influence on later work and for the clarity with which they state a method. Hardware design is treated only to the extent that the acquisition architecture fixes the structure of the sensing matrix, since that structure decides which robust methods apply.

The theoretical floor for sampling a multiband signal was set by Landau (1967), who showed that the average sampling rate cannot fall below the total occupied bandwidth. Reaching that rate when the band positions are unknown requires a sampling pattern that does not alias the active bands onto one another in an unresolvable way. Feng and Bresler (1996) proposed periodic nonuniform sampling for this spectrum-blind problem and showed that reconstruction is possible at rates close to the Landau bound without knowing the band positions in advance. These results predate compressed sensing and already contain its central idea: a small set of unknown active cells, found by solving an underdetermined linear system with a sparsity constraint.

Impulsive noise has been modelled in three main ways. Middleton's Class A and Class B models derive the amplitude distribution from a physical picture of many independent emitters with random positions and emission times (Middleton, 1977). The Class A density is an infinite mixture of Gaussians with Poisson weights, often truncated to two terms and then called a Bernoulli-Gaussian or contaminated-Gaussian model. Symmetric α-stable distributions offer a second description. They arise from the generalised central limit theorem, have algebraic tails, and for a characteristic exponent below two have infinite variance, which makes the mean-squared error undefined as a performance measure (Nikias and Shao, 1995). The third model is deterministic: an unknown but sparse vector of arbitrary large errors added to the measurements, with no distribution assumed.

The models lead to different algorithms, so the choice matters. Under a Gaussian mixture, the impulses are rare events that can be detected and discarded. Under an α-stable law there is no clean division between background and impulse, and the estimator must tolerate a continuum of large values. The sparse-error model suits wideband receivers in one respect, because a strong pulse arriving at the antenna corrupts a short run of consecutive samples, and it fails in another, since the model ignores the small dense noise that is always present as well. Robust statistics provides the common vocabulary for comparing estimators across these models, and Zoubir et al. (2012) review it for a signal processing readership.

The published methods fall into four families, which Figure 1 arranges as a tree and Table 1 compares by noise model and cost. The families are not exclusive. Several practical algorithms combine a screening stage with a robust loss, and the explicit outlier model can be read as a robust loss in disguise.

Four gaps stand out. The first concerns theory. Most robust recovery guarantees are proved for sensing matrices with independent random entries, while sub-Nyquist receivers produce structured matrices and, for multiband signals, jointly sparse unknowns. Results that hold for partial Fourier and sign-modulated matrices under heavy-tailed noise are scarce, and for α-stable noise even the right error metric is open to debate because second moments do not exist.

The second gap lies between the noise model and the hardware. Analyses place the impulse after acquisition, as an additive term on the digital measurements. A physical impulse arrives at the antenna, passes through mixers and filters that may saturate, and reaches the converter as a waveform shaped by the front end. Whether it remains sparse, and in which domain, depends on the architecture, and saturation makes the corruption signal-dependent. Joint modelling of front-end nonlinearity and impulsive interference has received little attention.

Third, sensing is a detection problem. The user of a spectrum sensor wants to know which bands are occupied, with controlled false-alarm probability, and not the best reconstruction of the waveform. Robust recovery papers mostly report reconstruction error. Detection thresholds that remain valid under heavy tails, and methods that estimate the support directly from robust covariance statistics, deserve more work. The fourth gap is practical: there is no agreed set of test conditions. Studies differ in the impulse model, its parameters and how the signal-to-noise ratio is defined when the noise variance is infinite, so that published comparisons are hard to reconcile. A shared, openly documented set of scenarios would allow methods to be compared on equal terms.

Sub-Nyquist spectrum sensing and robust estimation grew up separately, and the work reviewed here shows that they combine without great difficulty at the level of algorithms. A robust loss, an explicit outlier term, a screening stage or a heavy-tailed likelihood can each be attached to the standard sparse recovery model, and each protects against impulses under the noise model it assumes. What the literature does not yet provide is the same confidence at the level of systems: guarantees for the structured matrices that real receivers implement, models that follow an impulse through the analogue front end, and comparisons made under common conditions. Progress on those points would turn a collection of promising algorithms into design guidance for receivers that have to work in electrically noisy places.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.