Sub-Nyquist Sampling

The core principle enabling sub-Nyquist rates is signal sparsity in a known domain. While a wideband signal may occupy a large total bandwidth, the actual number of active transmissions within it is often small.

• Spectral Sparsity: Only a few narrowband carriers are active within a wide surveillance band.
• Time-Frequency Sparsity: Signals appear as isolated islands in a spectrogram.
• Dictionary Learning: The signal is represented as a linear combination of a few atoms from a predefined basis.

By exploiting this structure, the sampling rate is proportional to the information content, not the total bandwidth.

Sub-Nyquist sampling is mathematically formalized through Compressed Sensing (CS). Instead of uniform sampling, CS acquires a small number of non-adaptive linear projections of the signal.

• Measurement Matrix: The signal x is multiplied by a sensing matrix A to produce a compressed measurement vector y = Ax.
• Incoherence: The sensing matrix must be incoherent with the sparsity basis, a property satisfied by random matrices.
• Nonlinear Reconstruction: The original signal is recovered by solving an l1-norm minimization problem, which promotes sparse solutions.

This shifts the computational burden from the analog-to-digital converter (ADC) to the digital reconstruction backend.

The Modulated Wideband Converter is a practical hardware architecture implementing sub-Nyquist sampling for multiband signals. It uses a bank of parallel channels to mix the input with periodic waveforms.

• Principle: The input signal is split into m channels, each mixed with a distinct pseudo-random T_p-periodic sequence p_i(t).
• Aliasing: Mixing aliases the entire wideband spectrum into baseband, where it is low-pass filtered and sampled at a low rate f_s = 1/T_p.
• Reconstruction: The resulting samples form a compressed measurement vector. The original signal support and values are recovered using Continuous-to-Finite (CTF) algorithms.
• Hardware Efficiency: The number of channels m is proportional to the number of active bands, not the Nyquist rate.

Xampling unifies sub-Nyquist sampling and processing into a single framework. It combines analog compression with direct digital processing of the compressed samples.

• Analog Preprocessing: The signal is compressed in the analog domain before the ADC, reducing the sampling rate.
• Digital Processing: Parameter estimation, detection, and classification are performed directly on the compressed samples without full signal reconstruction.
• Task-Based Sampling: The sampling mechanism can be optimized for a specific inference task, such as carrier frequency estimation or modulation classification.

This bypasses the computationally expensive reconstruction step entirely for many cognitive radio applications.

FRI sampling is a sub-Nyquist framework for signals characterized by a finite number of parameters per unit time. It applies to specific parametric signal classes.

• Parametric Signals: Includes streams of Dirac pulses, piecewise polynomials, and pulse-amplitude modulation (PAM) signals.
• Sampling Kernels: The signal is filtered with a specific kernel (e.g., a sinc or sum-of-sincs) that captures the signal's degrees of freedom.
• Spectral Estimation: The parameters (e.g., time delays and amplitudes of pulses) are recovered from the low-rate samples using annihilating filter or subspace methods.
• Radar Applications: FRI is highly effective for radar target detection, where the received signal is a sparse stream of delayed and attenuated pulses.

Multi-Coset Sampling is a non-uniform sub-Nyquist scheme that selects specific samples from a uniform grid at the Nyquist rate. It is a deterministic alternative to random demodulation.

• Pattern Selection: A set of p cosets (sample indices) is chosen from a block of L uniform Nyquist-rate samples.
• Universal Pattern: A well-designed pattern can guarantee recovery for any sparse multiband signal, regardless of the active band locations.
• Minimal Rate: The average sampling rate is p/L times the Nyquist rate, where p must exceed the number of active bands.
• Reconstruction: Signal recovery is performed using spectrum-blind reconstruction algorithms that exploit the correlation between the sampled cosets.

This method is particularly suited for spectrum sensing where the carrier frequencies are unknown.

A wideband sensing technique that directly applies the principles of compressive sensing to the spectrum monitoring problem. It exploits the inherent sparsity of the spectrum—where only a few frequency bands are occupied at any given time—to reconstruct the full wideband signal from far fewer samples than the Nyquist rate requires.

• Enables the use of lower-rate, lower-power analog-to-digital converters (ADCs)
• Relies on sparse recovery algorithms like Basis Pursuit or Orthogonal Matching Pursuit
• Directly identifies spectrum holes without needing to scan individual channels sequentially

The foundational theorem that sub-Nyquist methods deliberately violate. It states that a bandlimited continuous-time signal can be perfectly reconstructed from its samples if the sampling rate is at least twice the maximum frequency present in the signal.

• The Nyquist rate is the theoretical minimum for alias-free reconstruction of arbitrary bandlimited signals
• Sub-Nyquist sampling succeeds by exploiting the additional structure of signal sparsity, not by violating information theory
• Violating the Nyquist criterion without a sparse signal model results in irreversible aliasing

The mathematical property that makes sub-Nyquist sampling possible. A signal is sparse in a given basis or dictionary if it can be expressed using only a small number of non-zero coefficients. In spectrum sensing, the wideband signal is sparse in the frequency domain because most channels are idle.

• Common sparsity bases: Fourier, Wavelet, Discrete Cosine Transform
• The sparsity level (K) determines the minimum number of measurements required for recovery
• Sparsity must be exploited in the sampling hardware itself, not just in post-processing

A specific hardware architecture for implementing sub-Nyquist sampling in practice. The MWC uses a bank of parallel channels, each mixing the wideband input with a unique periodic pseudo-random sequence before low-pass filtering and sampling at a low rate.

• The mixing operation aliases the entire spectrum into baseband in a controlled, invertible way
• The resulting samples form a compressed measurement vector from which the original sparse spectrum can be reconstructed
• Eliminates the need for a single high-speed ADC by using multiple parallel low-rate ADCs

A key theoretical condition on the measurement matrix that guarantees stable and accurate recovery of a sparse signal from compressed measurements. A matrix satisfies the RIP if it approximately preserves the Euclidean length of all sufficiently sparse vectors.

• Ensures that distinct sparse signals remain distinguishable after compression
• Random measurement matrices (e.g., Gaussian, Bernoulli) satisfy the RIP with high probability
• Provides the theoretical backbone for why random demodulation and MWC architectures work reliably

The computational methods used to reconstruct the original sparse signal from the sub-Nyquist measurements. These algorithms solve an underdetermined linear system by seeking the sparsest solution consistent with the data.

• Convex relaxation: Basis Pursuit (L1-minimization) provides strong theoretical guarantees
• Greedy methods: Orthogonal Matching Pursuit (OMP) and Compressive Sampling Matching Pursuit (CoSaMP) offer faster computation
• Bayesian methods: Sparse Bayesian Learning (SBL) incorporates prior distributions on sparsity for robust recovery in noisy environments