CSI-RS

CSI-RS is designed for extreme configurability to minimize overhead. It can be placed on any OFDM symbol within a slot and configured with a non-zero power (NZP) or zero power (ZP) pattern. The resource spans a configurable number of antenna ports ({1,2,4,8,12,16,24,32}) and can be mapped to every 1, 2, or 4 subcarriers in the frequency domain, allowing operators to trade off between channel estimation granularity and pilot density based on the deployment scenario.

In millimeter wave (FR2) deployments, CSI-RS is the primary tool for beam management. A set of CSI-RS resources can be configured to be transmitted on different downlink beams in a sweeping pattern. The UE measures the Layer 1 Reference Signal Received Power (L1-RSRP) for each beam and reports the best beam index back to the gNB. This procedure enables rapid beam refinement and recovery without relying on always-on synchronization signal blocks (SSBs).

A Zero-Power (ZP) CSI-RS resource represents a set of resource elements where the gNB transmits nothing, creating a silent period. The UE is configured to measure on these muted resources to accurately gauge inter-cell interference from neighboring base stations. This is critical for the UE to calculate the accurate Channel Quality Indicator (CQI) and Rank Indicator (RI), as it separates the serving cell's signal quality from the noise-plus-interference floor.

A specific CSI-RS configuration, known as the Tracking Reference Signal (TRS), is used for fine time and frequency synchronization. A TRS resource set consists of multiple periodic NZP CSI-RS resources spanning two consecutive slots. The UE uses the TRS to perform fine Doppler shift estimation and delay spread estimation, enabling precise symbol-level timing correction that is essential for high-order modulation schemes like 256QAM.

Multiple CSI-RS resources are grouped into a CSI-RS Resource Set. The 'repetition' flag within the set fundamentally changes UE behavior. When repetition is ON, the UE assumes the gNB uses the same spatial filter for all resources, allowing the UE to sweep its own receive beams to find the optimal one. When repetition is OFF, the gNB sweeps its transmit beams, and the UE reports the best one, enabling downlink transmit beam selection.

CSI-RS supports three time-domain behaviors. Periodic resources are RRC-configured and always on. Semi-persistent resources are activated and deactivated via MAC CE signaling, providing a balance between overhead and flexibility. Aperiodic resources are triggered dynamically by a single DCI message, allowing the gNB to schedule a burst of CSI-RS for a specific UE on-demand. This dynamic triggering is essential for serving bursty eMBB traffic efficiently.

The end product of CSI-RS measurement. CSI represents the known channel properties—scattering, fading, and power decay—that describe how a signal propagates from transmitter to receiver. In 5G NR, the UE computes CSI from CSI-RS and reports it back via:

• CQI (Channel Quality Indicator): Recommended modulation and coding scheme
• PMI (Precoding Matrix Indicator): Preferred precoding matrix from a codebook
• RI (Rank Indicator): Number of spatial layers the channel can support
• LI (Layer Indicator): Strongest layer for multi-layer transmission Accurate CSI enables closed-loop spatial multiplexing and link adaptation, directly determining peak throughput.

A deep learning model trained to infer CSI from received pilot signals (including CSI-RS) with higher accuracy than classical estimators. Architectures include:

• Convolutional Neural Networks (CNNs): Exploit spatial-frequency correlations in the channel grid
• Transformer Networks: Leverage self-attention to model long-range dependencies across antenna ports and subcarriers
• Complex-Valued Neural Networks: Operate natively on IQ samples, preserving magnitude and phase relationships without separating into real-valued channels
• Recurrent Networks (LSTM/GRU): Exploit CSI temporal correlation across slots for channel tracking These models consistently achieve lower NMSE than Least Squares (LS) and Minimum Mean Square Error (MMSE) estimators, especially in low pilot density regimes.

The primary use case driving CSI-RS design. Massive MIMO equips the gNB with 64, 128, or more antenna elements to serve multiple UEs simultaneously via spatial multiplexing. CSI-RS enables:

• Grid of Beams (GoB): The gNB sweeps a set of beams; the UE reports the best beam index
• Hybrid Beamforming: Combines analog beamforming (phase shifters) with digital precoding, requiring CSI-RS to sound the effective channel after analog combining
• MU-MIMO Precoding: The gNB uses CSI reports from multiple UEs to compute Zero-Forcing (ZF) or MMSE precoders that nullify inter-user interference
• Angular Domain Sparsity: The channel is sparse in the DFT domain, with multipath concentrated in few angles—a property exploited by compressed sensing and deep unfolding

The uplink counterpart to CSI-RS. While CSI-RS is a downlink pilot measured by the UE, the SRS is an uplink reference signal transmitted by the UE to allow the gNB to estimate the uplink channel. Critical distinctions:

• TDD Reciprocity: In Time Division Duplex systems, the uplink channel estimated from SRS is assumed identical to the downlink channel, eliminating the need for CSI feedback
• CSI-RS still required in TDD: For interference measurement (CSI-IM) and for UEs to measure non-reciprocal impairments
• SRS for Positioning: Used alongside CSI-RS for NR Positioning (NRPPa), enabling sub-meter accuracy via Time of Arrival (ToA) and Angle of Arrival (AoA) measurements
• SRS Resource Sets: Configured for codebook-based, non-codebook-based, or antenna-switched transmission

Standardized geometric channel models used for link-level simulation and AI training data generation. The Clustered Delay Line (CDL) model defines:

• Clusters: Groups of multipath rays with shared spatial parameters
• Angles of Arrival/Departure (AoA/AoD): Azimuth and zenith angles per cluster
• Delay Spread: Time dispersion of multipath components
• Doppler Spectrum: Frequency dispersion due to mobility CDL models (CDL-A through CDL-E) represent scenarios from indoor office to urban macro, while Tapped Delay Line (TDL) models simplify to single-antenna profiles. These models are essential for generating synthetic CSI-RS training data for neural channel estimators and validating CSI compression algorithms under standardized conditions.