Compressive Sensing

The foundational principle: a signal must be sparse in some transform domain. For spread spectrum signals, sparsity exists in the frequency domain—only a few narrowband carriers are active within a wide monitored bandwidth. The sampling rate is proportional to the information rate, not the Nyquist rate. A signal with K active tones across bandwidth B requires only O(K log(B/K)) samples rather than 2B samples. This enables analog-to-information converters that digitize multi-GHz spans at sub-GHz sample rates.

Random measurement ensembles satisfy the Restricted Isometry Property with high probability, ensuring stable reconstruction. Common choices include:

• Gaussian random matrices: Each entry drawn i.i.d. from N(0, 1/M)
• Bernoulli matrices: Binary ±1 entries with equal probability
• Partial Fourier matrices: Randomly subsampled DFT rows, ideal for spectrum sensing These matrices are maximally incoherent with sparse Fourier dictionaries, guaranteeing that each measurement captures equal information about all possible signal components.

Recovery solves a convex optimization: minimize ||x||₁ subject to y = ΦΨx, where Φ is the measurement matrix and Ψ is the sparsifying dictionary. The ℓ₁ norm promotes sparsity while remaining computationally tractable. Algorithms include:

• Basis Pursuit: Exact ℓ₁ minimization via linear programming
• LASSO: ℓ₁-regularized least squares for noisy measurements
• OMP: Greedy iterative support selection for low-complexity recovery For spectrum sensing, the reconstructed vector directly reveals occupied frequency bins.

A practical hardware implementation for wideband spectrum sensing. The signal is multiplied by a pseudo-random chipping sequence alternating at the Nyquist rate, then integrated and sampled at a low rate. This spreads each frequency component across the entire measurement bandwidth. The measurement matrix becomes a partial Fourier ensemble with random sign flips. Commercial implementations achieve 2 GHz instantaneous bandwidth with only 200 MHz equivalent sampling, enabling real-time detection of frequency-hopping signals and burst transmissions.

An advanced multi-branch architecture where the input signal is split into m parallel channels, each mixed with a distinct periodic waveform and lowpass filtered before sub-Nyquist sampling. The periodic mixing function aliases all frequency bands into baseband. Reconstruction uses Continuous-to-Finite block formulation solved via Multiple Measurement Vector techniques. This enables blind spectrum reconstruction without prior knowledge of carrier frequencies—critical for intercepting unknown FHSS hop sets and DSSS signals with unknown chip rates.

When signal sparsity bases are unknown a priori, dictionary learning jointly estimates the sparse representation and the dictionary atoms from compressive measurements. For spread spectrum identification, this adapts to:

• Unknown chip pulse shapes in DSSS signals
• Arbitrary hop dwell patterns in FHSS
• Non-ideal power amplifier nonlinearities The K-SVD algorithm alternates between sparse coding and dictionary update steps, converging to a basis that compactly represents the intercepted signal class for subsequent classification.

The fundamental prerequisite for compressive sensing. A signal is sparse if it can be represented by only a small number of non-zero coefficients in a specific transform domain. Dictionary learning algorithms, such as K-SVD, automatically discover the optimal basis functions (atoms) from training data, enabling the representation of complex spread spectrum signals with minimal coefficients. Without sparsity, sub-Nyquist reconstruction is mathematically ill-posed.

The measurement process relies on sensing matrices that are incoherent with the sparsity basis. The Restricted Isometry Property (RIP) provides a formal guarantee that a random matrix (e.g., Gaussian or Bernoulli) will preserve the geometry of sparse signals during dimensionality reduction. In RF applications, this is physically realized through random demodulator architectures or non-uniform sampling clocks that spread aliasing into noise-like artifacts.

The core reconstruction algorithm. While finding the sparsest solution (L0-norm) is NP-hard, convex relaxation using the L1-norm provides a tractable proxy. Solvers like Basis Pursuit and LASSO iteratively search for the coefficient vector with the minimum absolute sum that still matches the compressed measurements. This transforms a combinatorial search into a convex optimization problem solvable in polynomial time.

A computationally efficient alternative to convex optimization. Algorithms like Orthogonal Matching Pursuit (OMP) and Compressive Sampling Matching Pursuit (CoSaMP) iteratively build a sparse support set by selecting dictionary atoms most correlated with the residual measurement error. While lacking the formal guarantees of L1-minimization, greedy methods offer significantly faster reconstruction for real-time spectrum monitoring applications.

The physical hardware implementation of compressive sensing for RF signals. An AIC bypasses the traditional Nyquist-rate ADC by modulating the input signal with a pseudo-random chipping sequence, integrating the product, and sampling at a low rate. The Modulated Wideband Converter (MWC) is a prominent multi-branch architecture that extends this concept to multiband signals, enabling direct sub-Nyquist digitization of gigahertz-wide spectrum.

A specialized application combining compressive sensing with cyclostationary feature extraction. Instead of reconstructing the full signal, this technique directly estimates the Spectral Correlation Density (SCD) from compressed measurements. By exploiting the fact that modulated signals exhibit unique cyclic frequencies, it enables blind modulation identification and spread spectrum detection at sampling rates far below the Nyquist limit.