Angular Domain Sparsity

The transformation from the antenna domain to the angular domain is achieved using a 2D Discrete Fourier Transform (DFT). When a base station employs a uniform linear array (ULA), the spatial channel matrix H is pre-multiplied and post-multiplied by DFT matrices. This operation maps the physical propagation paths to discrete angular bins corresponding to quantized angles of arrival (AoA) and angles of departure (AoD). Because there are typically only a few dominant scatterers in the environment, the resulting angular channel matrix H_a has only a small number of significant non-zero elements, with the remaining entries being near-zero noise.

In a typical macro-cellular environment, a massive MIMO base station with 64 or 128 antennas may communicate with a user equipment that has only 1-4 antennas. Despite the large dimensionality of the channel matrix, the actual number of resolvable multipath components is physically limited by the scattering environment. A channel might consist of only 3-6 dominant clusters, each with a specific AoA and AoD. This physical constraint is the root cause of angular sparsity: the number of significant paths is much smaller than the number of antenna elements, leading to a low-rank channel matrix in the angular domain.

Angular domain sparsity is the critical enabler for compressed sensing (CS) techniques in channel estimation. Classical methods like Least Squares (LS) require pilot overhead proportional to the number of antennas. However, by exploiting the sparsity of H_a, CS algorithms such as Orthogonal Matching Pursuit (OMP) or Iterative Hard Thresholding (IHT) can accurately recover the full channel matrix from a drastically reduced number of pilot symbols. This directly translates to reduced pilot overhead and increased spectral efficiency, as the number of required measurements scales with the sparsity level rather than the antenna count.

Traditional iterative sparse recovery algorithms can be enhanced through deep unfolding, a model-driven deep learning technique. An algorithm like the Iterative Shrinkage-Thresholding Algorithm (ISTA) is unrolled into a neural network where each layer corresponds to one iteration. The shrinkage thresholds and step sizes become learnable parameters optimized via backpropagation. This approach, known as Learned ISTA (LISTA), converges to an accurate sparse solution in significantly fewer iterations than the classical algorithm, making it suitable for the low-latency requirements of 5G channel estimation.

The angular domain can be rigorously interpreted as the spatial frequency domain. Just as a time-domain signal is decomposed into frequency components via the Fourier transform, the spatial signal across a ULA is decomposed into spatial frequency components. Each angular bin corresponds to a specific spatial frequency, which is directly related to the physical angle via the relationship f_s = (d/λ) sin(θ), where d is the antenna spacing and λ is the wavelength. This duality allows signal processing engineers to apply intuition from classical spectral estimation to the spatial domain.

A practical challenge in angular domain processing is power leakage caused by basis mismatch. The DFT basis assumes that the true AoAs and AoDs fall exactly on the discrete grid points. When a physical path arrives at an angle that lies between two grid points, its energy leaks into multiple adjacent angular bins, reducing the effective sparsity. This phenomenon is mitigated by using oversampled DFT dictionaries or gridless compressed sensing techniques like Atomic Norm Minimization (ANM), which operate directly in the continuous angular parameter space.

The mathematical mechanism that reveals angular domain sparsity by transforming the spatial channel matrix into the virtual angular domain. In a uniform linear array, the DFT matrix serves as the unitary transformation matrix, mapping antenna elements to discrete angles of arrival (AoA) and angles of departure (AoD). When the number of base station antennas is large, the DFT basis vectors approximate the array steering vectors, causing multipath components to concentrate in a few dominant bins. This sparsifying basis is the prerequisite for all compressed sensing and deep learning-based CSI recovery techniques.

A signal processing framework that exploits angular domain sparsity to reconstruct the full channel matrix from severely undersampled measurements. Algorithms such as Orthogonal Matching Pursuit (OMP) and Basis Pursuit iteratively identify the non-zero angular components, requiring far fewer pilot symbols than the antenna count. The key insight is that the number of dominant multipath clusters is much smaller than the number of antennas, making the channel compressible. This directly reduces pilot overhead and CSI feedback payload in FDD massive MIMO systems.

A model-driven deep learning architecture that maps the iterative steps of the Iterative Shrinkage-Thresholding Algorithm (ISTA) into neural network layers. Each layer performs a gradient descent step followed by a learnable soft-thresholding nonlinearity that promotes sparsity in the angular domain. Unlike black-box neural networks, deep unfolded networks have a structure that explicitly mirrors sparse recovery optimization, enabling:

• Fewer iterations for convergence
• Interpretable parameters tied to physical channel properties
• Robust generalization to varying sparsity levels

A fixed coordinate transformation that expresses the physical MIMO channel as a sparse matrix in the angular domain. The virtual representation uses fixed, pre-defined spatial basis functions—typically DFT beams—to decompose the channel into resolvable angular bins. Each non-zero entry corresponds to a physical multipath component connecting a specific AoD at the transmitter to a specific AoA at the receiver. This representation is deterministic and does not require estimating path parameters, making it ideal for grid-based compressive sensing and dictionary learning approaches.

A critical distinction in massive MIMO channel modeling. Spatial sparsity refers to the limited number of physical scatterers in the propagation environment, resulting in a low-rank channel matrix. Angular domain sparsity is the manifestation of this physical sparsity after a DFT transformation, where the channel energy concentrates in a small number of angular bins. While related, angular sparsity is the computationally exploitable property for:

• CSI compression via autoencoders
• Beamspace processing for millimeter-wave systems
• Pilot decontamination through angular user separation

A processing paradigm that operates directly in the angular domain by applying a lens antenna array or a spatial DFT to transform the received signal into the beamspace. In this domain, the channel becomes a sparse matrix where only a few beam indices contain significant energy. Channel estimation reduces to identifying these dominant beams, dramatically lowering the effective dimensionality. This approach is particularly effective for millimeter-wave massive MIMO, where the hybrid analog-digital architecture naturally aligns with beamspace processing and angular sparsity.