In massive MIMO systems, the choice between digital and hybrid beamforming architectures dictates performance, complexity, and cost. Digital beamforming, with a dedicated RF chain for each antenna element, offers maximum degrees of freedom. This allows for per-element phase and amplitude control, enabling fully flexible multi-user MIMO and sophisticated null-steering. The precoding matrix $W$ can be directly calculated to approximate the inverse of the channel matrix $H$, such as in zero-forcing precoding where $W = H^H(HH^H)^{-1}$.

Hybrid architectures, however, are often a practical necessity, especially in FR2 deployments. They partition the precoding into a digital baseband precoder ($W_{BB}$) and an analogue RF precoder ($W_{RF}$), where $W = W_{RF}W_{BB}$. The analogue stage, typically implemented with phase shifters, imposes a constant-modulus constraint on the elements of $W_{RF}$, significantly reducing the solution space for the precoding matrix. This necessitates algorithms like orthogonal matching pursuit (OMP) to iteratively find a sparse approximation of the optimal, fully digital precoding matrix. The trade-off is clear: reduced hardware cost and power consumption versus a performance hit due to quantization lobes and a constrained beam pattern.

The evolution from Type I to Type II codebooks in 3GPP Release 16 marked a significant shift in channel state information (CSI) feedback. Type I codebooks rely on a predefined grid of beams, which the UE selects from. This is efficient but fundamentally limited by the grid's granularity. Type II codebooks introduce a more flexible, component-based structure. The UE reports a set of coefficients and basis vectors, allowing the gNB to reconstruct a much more accurate representation of the channel's spatial characteristics.

This structure is based on a two-dimensional discrete Fourier transform (DFT) basis, allowing for the selection of dominant beams in both the angular and delay domains. The UE reports parameters ($i_{1,1}$, $i_{1,2}$, etc.) that indicate the selected basis patterns and their corresponding complex-valued coefficients. The gNB then reconstructs the precoding matrix using these components. This provides a substantial gain in spectral efficiency, particularly in channels with rich multipath, at the cost of increased UE complexity and feedback overhead.

Release 18 and the push towards 5G-Advanced introduce AI/ML-based CSI compression to manage this overhead. Instead of transmitting raw coefficients, the UE employs a trained neural network encoder to create a compact, latent-space representation of the CSI. The gNB uses a corresponding decoder to reconstruct the full CSI. This approach can significantly reduce the bits required on the PUCCH for feedback, allowing for even more detailed channel reports without saturating the uplink control channel. The standardisation focuses on the interface and decoder, leaving the encoder/decoder implementation details to vendors, fostering innovation.

Multi-TRP (Transmission/Reception Point) operation is critical for delivering the reliability and throughput promises of 5G, particularly for URLLC and edge eMBB scenarios. By coordinating transmissions from multiple, geographically separated TRPs to a single UE, the network can mitigate blockage and improve signal quality. The most advanced form is coherent joint transmission, where multiple TRPs transmit the same data stream using phase-aligned signals. This requires precise time and phase synchronisation between TRPs, often on the order of nanoseconds, to ensure the signals combine constructively at the UE.

This tight synchronisation allows the TRPs to act as a single, distributed antenna array, creating a macrodiversity gain that is highly resilient to blockage. For non-coherent joint transmission, each TRP can transmit different data streams (or redundancy versions), which is simpler to implement but offers less peak performance.

Efficiently managing beams in this environment, especially in FR2, is handled by the unified TCI framework introduced in Release 16. The Transmission Configuration Indicator (TCI) state associates a logical channel (like a PDSCH) with a specific spatial configuration, typically a quasi-co-location (QCL) relationship with a CSI-RS or SSB. The UE is configured with a list of TCI states and can be dynamically instructed via MAC-CE or DCI to switch between them. This framework decouples the beam indication from the actual data scheduling, allowing the gNB to update the UE's beam information proactively, minimising latency when a beam switch is required due to mobility or blockage.

For high-mobility scenarios, the challenge is maintaining a stable link as the optimal beam changes rapidly. Here, the expansion of DMRS and SRS ports to support up to 24 and 32 ports respectively (for certain configurations) becomes vital. More SRS ports allow the UE to sound a wider range of spatial directions simultaneously, giving the gNB a more comprehensive view of the channel to predict the next best beam. Similarly, expanded DMRS ports enable the gNB to transmit demodulation reference signals on more layers or beams, allowing the UE to track channel changes with finer granularity across the massive MIMO array.