Precoding Matrix Indicator (PMI) in 5G NR

• PMI is the foundational signaling parameter in 5G NR that indexes a precoding matrix from a standardized codebook to adapt MIMO downlink transmission.
• It involves UE channel measurement and performance metric optimization (e.g., post-precoding SNR, mutual information) to select the best matrix from Type I or Type II codebooks.
• Recent advances integrate AI-driven methods and deep learning architectures to reduce feedback overhead and improve multi-user spectral efficiency in dynamic environments.

The Precoding Matrix Indicator (PMI) is the foundational signaling parameter in closed-loop multiple-input multiple-output (MIMO) downlink systems standardized in 5G New Radio (NR). PMI enables user equipment (UE) to convey, with only a few bits, its preferred precoding matrix from a codebook to the base station (gNB), enabling near-optimal adaptation to the instantaneous channel while minimizing uplink feedback overhead. PMI feedback is central to achieving the spectral efficiency, reliability, and latency objectives of contemporary 5G wireless systems through flexible spatial transmission and efficient resource utilization.

1. Formal Definition, Purpose, and Signaling Flow

In 5G NR, PMI denotes the index $i$ in a finite, standardized set (the codebook $\mathcal{W}$), such that the corresponding matrix $W_i$ best matches the channel $H\in\mathbb{C}^{N_R\times N_T}$ as measured via downlink reference signals (CSI-RS). The UE identifies $W_i$ by maximizing a performance metric (typically post-precoding SNR or mutual information):

$$i^* = \arg\max_{W\in\mathcal{W}} \text{Metric}(H, W)$$

Feedback reporting comprises the PMI, the rank indicator (RI), and a channel quality indicator (CQI). At the gNB, $W_i$ is mapped from the codebook structure (fully specified in 3GPP TS 38.214 or further clarified in (Ning et al., 8 Jan 2026)), and applied to the downlink data symbols prior to OFDM modulation (Díaz-Ruiz et al., 2024, Ning et al., 8 Jan 2026, Ntassah et al., 20 Apr 2025).

The end-to-end PMI feedback flow consists of:

• Configuration: gNB assigns CSI-RS and feedback structure to the UE.
• Measurement: UE performs channel estimation and computes selection metrics.
• Reporting: UE feeds back the PMI/RI/CQI tuple over PUCCH/PUSCH.
• Reconstruction: gNB recovers the beamforming matrix from the index and applies it to transmitted symbols (Ning et al., 8 Jan 2026).

2. Structure and Construction of 5G NR Codebooks

Two principal codebook families are defined in 3GPP Releases 15–18: regular codebooks (Type I) and enhanced or multi-beam codebooks (Type II), with subsequent enhancements for frequency and Doppler domain selectivity (Ning et al., 8 Jan 2026, Díaz-Ruiz et al., 2024).

Type I (Regular Codebooks)

• Employ separable Kronecker-structured horizontal and vertical DFT beamformers ($W_h$, $W_v$), with low feedback overhead.
• Each precoder entry is parameterized by oversampling factors $(O_1, O_2)$ and the number of antenna elements $(N_1, N_2)$:

$W_i = W_h(i_{1,1}, i_{1,2}) \otimes W_v(i_{1,3})$

• Overhead: PMI bits distributed among selection indices for spatial and layer parameters (see Eq. (7) in (Díaz-Ruiz et al., 2024)).
• Scalability: Supports up to 4 layers in Release 15.

Type II (Multi-Beam Codebooks)

• UEs select several quantized beams and combine with per-beam phase/amplitude weights.
• Codebook basis constructed as combinations of DFT/steering vectors, parameterized by number of beams $B$, PSK order $N_\text{PSK}$, layers $L$.
• Overhead: Feedback grows rapidly with $B$, $N_\text{PSK}$, and number of selected beams (see Eq. (8) in (Díaz-Ruiz et al., 2024)).
• Type II offers superior spatial granularity but is layer-limited to $L\leq 2$ as per Release 15.

Later releases introduce port-selection codebooks (free port selection), spectral compression (selection of dominant delay taps), and temporal compression (selection of dominant Doppler bases), each with associated definitions for PMI bit fields and physical interpretation of codebook entries (Ning et al., 8 Jan 2026).

The following table summarizes key variables and their physical meanings from (Ning et al., 8 Jan 2026):

Symbol	Meaning
$N_H, N_V$	# horizontal/vertical logical antenna elements
$O_H, O_V$	Horizontal/vertical oversampling factors
$\mathbf{v}_{l,m}$	2-D DFT beam (Kronecker of DFT vectors)
$L$	# spatial bases per layer
$M_v$	# delay taps (spectral compression)
$Q$	# temporal/Doppler bases (temporal compression)

3. Mathematical Criteria for PMI Selection

PMI selection at the UE exhaustively evaluates each candidate precoding matrix according to a link performance metric:

• Mutual Information: $$I(H, W) = \log_2 \det \left( I_L + \frac{P_t}{N_t \sigma_n^2} H W W^H H^H \right)$$
• Post-Precoding SNR (per layer): $\| HW \|_F^2 / \sigma_n^2$
• Single-layer (beamforming): Maximize $w^H H^H H w$ for $w\in\mathbb{C}^{N_T \times 1}$

The index maximizing the selected metric, under block-error-rate (BLER) and other constraints, is reported as the PMI (Díaz-Ruiz et al., 2024, Ntassah et al., 20 Apr 2025). In multi-user or interference-limited scenarios, PMI selection becomes a multi-objective optimization to jointly maximize overall spectral efficiency and minimize interference, and may utilize distributed or centralized algorithms, including AI/ML-based policies (Ntassah et al., 20 Apr 2025).

4. Feedback Overhead, Performance, and Trade-offs

The feedback overhead of PMI depends on:

• Codebook size (determined by antenna configuration, oversampling factors, beams, phase/amplitude quantization)
• Reporting granularity (wideband vs. subband)
• Number of spatial, frequency, and temporal bases selected

For example, Release 15 Type I codebooks use as little as 2 bits per report, while Type II and compressed codebooks scale up to 60 bits or more depending on the number of beams and subbands reported. Newer standards introduce advanced compression and selection methods, reducing the required feedback while maintaining or improving SE (Ning et al., 8 Jan 2026).

Simulation studies reveal:

• In low SNR (${\leq}7$ dB), both codebook types perform similarly, as noise dominates (Díaz-Ruiz et al., 2024).
• In medium SNR, Type II outperforms Type I by exploiting finer beam resolution (up to 15% SE gain).
• In high SNR ($>26$ dB), Type I can regain SE advantage due to higher supported stream/rank (L) (Díaz-Ruiz et al., 2024).
• Type II incurs up to $3\times$ the uplink overhead of Type I in 8$\times$4 configurations, with RI and CQI distributions shifting toward higher RI in high SNR (Díaz-Ruiz et al., 2024).

5. Extensions: AI-Driven PMI Selection and Channel Reconstruction

Recent research addresses PMI optimization under network-wide constraints, including interference mitigation in dense deployments:

• The Advantage Actor-Critic (A2C) framework jointly selects PMI indices across UEs to maximize cell-level spectral efficiency and penalize inter-cell interference. This is realized in O-RAN xApp deployments, integrating PMI selection with closed-loop radio-resource management at 1 ms TTI granularity (Ntassah et al., 20 Apr 2025).
• Deep learning-based implicit feedback architectures, such as ImCsiNet and bi-ImCsiNet, replace the conventional codebook-based PMI mapping with neural encoder/decoder pairs. These learn environment-specific, compact PMI representations, achieving 25–48% feedback overhead reduction vs. standard codebooks (Type I/II), with full backward compatibility for the 5G PMI signaling chain (Chen et al., 2021).
• CSI sensing and constrained phase retrieval leverage the inequalities induced by PMI/CQI feedback to reconstruct the downlink channel at the BS, exploiting spatial consistency and dimension reduction for overhead efficiency in heterogeneous feedback scenarios (Li et al., 2022).

6. Evolution of PMI Codebooks Across 3GPP Releases

PMI codebooks have evolved through 3GPP Releases 15–18 to address growing MIMO array sizes, channel complexity, and use case diversity:

• Release 15: Type I (single-beam, low-complexity) and Type II (multi-beam, high-precision) codebooks.
• Release 16: Addition of spectral compression for wideband operation.
• Release 17: Port-selection (flexible, full-connect PEB) and further feedback reduction.
• Release 18: Temporal compression (Doppler basis) enabling predicted PMI for high-mobility scenarios.

Each evolutionary step is motivated by specific trade-offs between precision, feedback overhead, and targeted deployment scenarios (e.g., IoT, mmWave, wideband, vehicular) (Ning et al., 8 Jan 2026).

Example mapping of scenario to codebook family (cf. (Ning et al., 8 Jan 2026)):

Release	Codebook	Overhead	Use Cases
R15 Type I	2 bits	very low	IoT, legacy UE, low-complexity devices
R15 Type II	~60 bits	high	mmWave eMBB, large arrays
R16/17/18	20–25 bits	medium	URLLC, NTN, high-mobility

7. Implementation, Limitations, and Open Problems

Implementation of PMI-based adaptive precoding must address:

• Feedback efficiency: Compression and learning-based methods reduce uplink burden without protocol changes (Chen et al., 2021, Li et al., 2022).
• Computational complexity: Exhaustive codebook search may be intractable for large codebooks; practical UEs may use hierarchical or heuristic search algorithms (Díaz-Ruiz et al., 2024).
• Layer/rank constraints: Standard-imposed limits on streams/rank restrict achievable multiplexing gains.
• Interference mitigation: Naïve per-UE PMI selection can increase multi-user interference; centralized or coordinated approaches are needed in dense, multi-cell environments (Ntassah et al., 20 Apr 2025).
• Adaptation to channel non-stationarity: Site-specific codebooks and online learning are active research directions (Chen et al., 2021, Ning et al., 8 Jan 2026).

Open research areas include codebook design for near-field and ultra-massive MIMO, hierarchical and analog codebooks, AI-parameterized codebooks, joint communication-sensing (ISAC) codebooks, distributed MIMO feedback, and codebooks for reconfigurable intelligent surfaces (RIS) (Ning et al., 8 Jan 2026).

• (Díaz-Ruiz et al., 2024) Optimizing MIMO Efficiency in 5G through Precoding Matrix Techniques
• (Ntassah et al., 20 Apr 2025) Interference-Aware PMI selection for MIMO systems in an O-RAN scenario
• (Li et al., 2022) CSI Sensing from Heterogeneous User Feedbacks: A Constrained Phase Retrieval Approach
• (Chen et al., 2021) Deep Learning-based Implicit CSI Feedback in Massive MIMO
• (Ning et al., 8 Jan 2026) Precoding Matrix Indicator in the 5G NR Protocol: A Tutorial on 3GPP Beamforming Codebooks