Hybrid Beamforming

The defining characteristic of hybrid beamforming is the cascade of digital and analog precoding. The baseband processor performs low-dimensional digital precoding (interference cancellation, multi-stream multiplexing) on a small number of RF chains. The analog beamformer—implemented via a network of phase shifters or switches—then maps these signals to a large antenna array. This decomposition is typically expressed as F = F_RF * F_BB, where F_RF is the analog precoder (constant modulus, phase-only) and F_BB is the digital baseband precoder. The analog stage provides array gain through directional beamforming, while the digital stage handles multi-user MIMO and fine-grained spatial processing.

The primary motivation for hybrid beamforming is the drastic reduction in RF chains—the expensive, power-hungry components comprising DACs/ADCs, mixers, and amplifiers. In a fully digital architecture, each antenna requires a dedicated RF chain, making massive MIMO (e.g., 256 antennas) economically infeasible. Hybrid beamforming enables N_RF << N_antennas, where a system with 256 elements might use only 4-8 RF chains. This reduces hardware cost, power consumption, and baseband processing complexity by an order of magnitude while retaining most of the beamforming gain. The trade-off is constrained precoding flexibility due to the analog network's phase-only limitations.

The analog stage is implemented using one of several physical architectures:

• Phase Shifter Network: Each antenna element connects to each RF chain through a programmable phase shifter, forming a fully-connected structure. This provides maximum beamforming flexibility but requires N_RF × N_antennas phase shifters.
• Switched Beam Network: A fixed set of pre-defined beams (e.g., via a Butler matrix or lens array) is available, and the system selects the best beam. Lower complexity but reduced adaptability.
• Sub-Connected Architecture: Each RF chain connects to a disjoint subset of antennas, reducing the number of phase shifters to N_antennas. This trades spatial resolution for lower hardware complexity. The constant modulus constraint on phase shifters makes the joint optimization of F_RF and F_BB a non-convex problem.

The non-convex, constrained optimization of hybrid precoders has driven adoption of deep reinforcement learning (DRL) and deep unfolding techniques. A DRL agent can learn to map channel state information (CSI) directly to hybrid precoder matrices, treating the analog constraints as part of the environment. Alternatively, model-driven unfolding unrolls iterative algorithms like manifold optimization into trainable neural network layers, learning optimal step sizes and projections. These learned approaches often outperform classical methods like orthogonal matching pursuit (OMP) for hybrid precoding, especially in highly correlated or rapidly changing channels where real-time adaptation is critical.

Hybrid beamforming introduces unique channel estimation difficulties because the baseband processor cannot directly observe the full-dimensional received signal at each antenna—it only sees the low-dimensional output after analog combining. This necessitates compressed channel estimation techniques. The system must estimate a high-dimensional mmWave channel (characterized by sparse multipath components in the angular domain) from a small number of RF chain observations. Compressive sensing algorithms and deep learning-based estimators (e.g., learned denoising or generative models) are employed to recover the full channel matrix from these compressed measurements, often exploiting the inherent sparsity of mmWave propagation.

Hybrid beamforming is particularly critical for millimeter-wave (mmWave) communications (30-300 GHz), where short wavelengths enable packing hundreds of antennas into small form factors. At these frequencies, the high path loss necessitates extreme directional beamforming gain, which the analog stage provides. Simultaneously, the digital stage enables spatial multiplexing of multiple data streams to different users. This synergy makes hybrid beamforming the de facto architecture for 5G NR base stations and 802.11ad/ay Wi-Fi systems operating in the 28 GHz, 39 GHz, and 60 GHz bands. The architecture balances the competing demands of high gain, multi-stream transmission, and hardware feasibility.

A planar metasurface composed of many passive or semi-passive elements that can dynamically tune the phase, amplitude, or polarization of impinging electromagnetic waves. Neural networks optimize the reflection coefficients for beamforming.

• Extends coverage in dead zones without active RF chains
• Complements hybrid architectures by providing an additional passive beamforming layer
• Controlled via a low-rate digital interface from the base station

A beamforming architecture that employs the attention mechanism from transformers to dynamically weigh the importance of different propagation paths or antenna elements. This enables robust beam prediction in highly scattering environments.

• Learns to focus on dominant multipath components
• Naturally handles variable numbers of users and paths
• Can be integrated into both digital and hybrid precoding stages

A beamforming approach that models the wireless network as a graph, where nodes represent transmitters or users and edges represent interference links. A Graph Neural Network (GNN) learns distributed and scalable precoding policies.

• Naturally captures interference topology
• Scales to large networks without retraining
• Enables decentralized coordination between base stations

The use of a reinforcement learning agent to dynamically select the optimal Modulation and Coding Scheme (MCS) based on real-time channel quality feedback. This maximizes throughput while maintaining a target block error rate.

• Adapts to non-stationary channel conditions
• Learns optimal policies without explicit channel models
• Often paired with hybrid beamforming for joint optimization of beam and MCS selection