Ferkans — Interactive Telecom Tutor

From Analog Compression to Digital Codebooks

Compressed feedback (Section 2) treats the CSI as a continuous-valued signal and compresses it via random projection followed by scalar quantization. An alternative — and the one adopted in every cellular standard since LTE — is to directly quantize the channel direction using a discrete codebook: the UE selects the best matching codeword from a finite set and feeds back only the index. This reduces the feedback to $\log_2 |\mathcal{C}|$ bits, with the quality determined by how well the codebook tessellates the channel space.

Definition:
Beamforming Codebook and PMI

A beamforming codebook is a finite set of unit-norm vectors

$\mathcal{C} = \{\mathbf{v}_{1}, \mathbf{v}_{2}, \ldots, \mathbf{v}_{2^B}\} \subset \mathbb{C}^{N_t}, \quad \|\mathbf{v}_{i}\| = 1 \; \forall i,$

where $B$ is the number of feedback bits. Given its channel estimate $\hat{\mathbf{H}}_k$ , user $k$ selects the precoding matrix indicator (PMI):

$i_k^* = \arg\max_{i \in \{1, \ldots, 2^B\}} \frac{|\hat{\mathbf{H}}_k^H \mathbf{v}_{i}|^2}{\|\hat{\mathbf{H}}_k\|^2},$

and feeds back the $B$ -bit index $i_k^*$ . The BS uses $\mathbf{v}_{i_k^*}$ as the precoding vector for user $k$ .

The selection criterion maximizes the normalized beamforming gain — equivalently, it minimizes the angle between the channel direction and the codeword.

In practice, the UE also reports a channel quality indicator (CQI) and a rank indicator (RI), in addition to the PMI. The PMI selects the precoding direction, the RI selects the number of spatial layers, and the CQI selects the MCS level.

,

Precoding Matrix Indicator (PMI)

An index into a predefined codebook of beamforming/precoding matrices. The UE computes the PMI by selecting the codebook entry that maximizes the expected throughput (or equivalently, the beamforming gain) for its estimated channel. In 5G NR, the PMI is reported alongside the RI and CQI as part of the CSI report.

Definition:
Chordal Distance and Quantization Error

The chordal distance between two unit-norm vectors $\mathbf{v}_{1}, \mathbf{v}_{2} \in \mathbb{C}^{N_t}$ is

$d_c(\mathbf{v}_{1}, \mathbf{v}_{2}) = \sqrt{1 - |\mathbf{v}_{1}^{H} \mathbf{v}_{2}|^2}.$

This measures the distance between the one-dimensional subspaces (lines) spanned by $\mathbf{v}_{1}$ and $\mathbf{v}_{2}$ on the Grassmann manifold $\mathcal{G}(1, N_t)$ .

The quantization error of codebook $\mathcal{C}$ for a random channel direction $\bar{\mathbf{H}}_k = \mathbf{H}_{k} / \|\mathbf{H}_{k}\|$ is

$\epsilon(\mathcal{C}) = \mathbb{E}\!\left[\min_{\mathbf{v} \in \mathcal{C}} d_c^2(\bar{\mathbf{H}}_k, \mathbf{v})\right] = \mathbb{E}\!\left[1 - \max_{\mathbf{v} \in \mathcal{C}} |\bar{\mathbf{H}}_k^H \mathbf{v}|^2\right].$

Smaller $\epsilon(\mathcal{C})$ means better codebook resolution.

Theorem: Rate Loss from Codebook Quantization

For a single-user MISO system with $N_t$ antennas, transmit power $P_t$ , and codebook $\mathcal{C}$ with $B$ feedback bits, the rate loss due to quantization is bounded by

$\Delta R = R_{\text{perfect}} - R_{\text{codebook}} \leq \log_2\!\left(1 + \frac{P_t}{\sigma^2} \epsilon(\mathcal{C})\right).$

For an isotropic channel $\mathbf{H}_{k} \sim \mathcal{CN}(\mathbf{0}, \mathbf{I}_{N_t})$ , the average quantization error of a random codebook with $2^B$ entries satisfies

$\epsilon(\mathcal{C}) \leq c \cdot 2^{-B/(N_t-1)}$

for a constant $c$ depending on $N_t$ . Hence the rate loss decays exponentially in $B/(N_t-1)$ : each additional bit per antenna dimension halves the gap.

The channel direction lives on the $(N_t-1)$ -dimensional complex unit sphere (a $(2N_t-2)$ -dimensional real manifold). A codebook with $2^B$ entries tessellates this manifold; the quantization error is the average distance to the nearest codeword. As $B$ grows, the tessellation becomes finer and the error decays exponentially — but the exponent is $B/(N_t-1)$ , so keeping the error constant as $N_t$ grows requires $B \propto N_t$ .

Proof

Rate decomposition

With perfect CSI, the BS uses MRT: $\mathbf{v}_{k} = \bar{\mathbf{H}}_k$ , achieving $R_{\text{perfect}} = \log_2(1 + P_t \|\mathbf{H}_{k}\|^2 / \sigma^2)$ . With codebook feedback, $\mathbf{v}_{k} = \mathbf{v}_{i_k^*}$ , achieving $R_{\text{codebook}} = \log_2(1 + P_t |\mathbf{H}_{k}^{H} \mathbf{v}_{i_k^*}|^2 / \sigma^2)$ .

Bound the gap

Since $|\mathbf{H}_{k}^{H} \mathbf{v}_{i_k^*}|^2 \geq \|\mathbf{H}_{k}\|^2 (1 - d_c^2(\bar{\mathbf{H}}_k, \mathbf{v}_{i_k^*}))$ , the rate loss satisfies $\Delta R \leq \log_2\!\left(\frac{1 + P_t \|\mathbf{H}_{k}\|^2/\sigma^2}{1 + P_t \|\mathbf{H}_{k}\|^2 (1 - \epsilon_k)/\sigma^2}\right) \leq \log_2(1 + \text{SNR} \cdot \epsilon_k).$

Quantization error bound

For i.i.d. Rayleigh fading and a random codebook, the sphere-covering bound gives $\epsilon(\mathcal{C}) \leq c \cdot 2^{-B/(N_t-1)}$ . This is tight for Grassmannian codebooks (optimal packing on the Grassmannian). The rate loss becomes $\Delta R \leq \log_2(1 + c \cdot \text{SNR} \cdot 2^{-B/(N_t-1)})$ . $\blacksquare$

,

Definition:
Grassmannian Codebook

A Grassmannian codebook is a codebook $\mathcal{C} = \{\mathbf{v}_{1}, \ldots, \mathbf{v}_{2^B}\}$ that maximizes the minimum chordal distance between any pair of codewords:

$\mathcal{C}^* = \arg\max_{\mathcal{C}} \min_{i \neq j} d_c(\mathbf{v}_{i}, \mathbf{v}_{j}).$

This is the optimal packing problem on the Grassmann manifold $\mathcal{G}(1, N_t)$ . Grassmannian codebooks minimize the worst-case quantization error and achieve the sphere-covering lower bound on $\epsilon(\mathcal{C})$ asymptotically.

For small $N_t$ and $B$ , Grassmannian codebooks can be computed numerically (e.g., by alternating projection). For large $N_t$ , they are impractical to store and search — motivating structured codebooks like DFT and 5G NR Type I/II.

Definition:
DFT Codebook

The DFT codebook for a ULA with $N_t$ antennas and $B$ -bit feedback is

$\mathcal{C}_{\text{DFT}} = \left\{\mathbf{f}_i = \frac{1}{\sqrt{N_t}} \begin{bmatrix} 1 \\ e^{j 2\pi i / 2^B} \\ \vdots \\ e^{j 2\pi i (N_t-1) / 2^B} \end{bmatrix} : i = 0, 1, \ldots, 2^B - 1 \right\}.$

Each codeword is a steering vector pointing in a quantized direction. The DFT codebook uniformly samples the angular domain with resolution $\Delta\phi = 2\pi / 2^B$ .

Advantages: Simple to generate, store, and search ( $O(N_t)$ per codeword comparison). No explicit storage needed — codewords are computed on the fly.

Limitation: The DFT codebook is designed for ULA geometry only. It quantizes only the beam direction, not the beam shape, so it is suboptimal for correlated channels with multiple scattering clusters.

Codebook Quantization Error vs. Feedback Bits

Explore how the quantization error $\epsilon(\mathcal{C})$ decreases as the number of feedback bits $B$ increases, for different codebook types (random, DFT, Grassmannian bound) and antenna counts. The exponential decay rate $B/(N_t-1)$ reveals the per-antenna cost of achieving a target error.

Parameters

N_t

16

Number of BS antennas

B_{\\max}

12

Maximum feedback bits

Codebook type

Codebook Types for CSI Feedback

Property	Grassmannian	DFT	5G NR Type I	5G NR Type II
Design criterion	Max min distance on $\mathcal{G}(1, N_t)$	Uniform angular sampling	Beam selection from DFT grid	Beam combination with amplitude/phase
Feedback bits	$B$	$B$	4–8 bits (wideband)	8–22 bits (wideband + subband)
Complexity (UE)	High: search $2^B$ entries	Low: $O(N_t \log N_t)$ via FFT	Moderate: beam sweeping	High: subband evaluation
Storage	$2^B \times N_t$ complex	None (compute on fly)	Predefined beams	Predefined + combination coefficients
Suited for	Small $N_t$ , theoretical analysis	ULA, LOS/single-cluster	Wideband PMI, FDD massive MIMO	High-resolution FDD massive MIMO
Standard adoption	None (theoretical)	LTE Release 8 (partial)	5G NR Release 15	5G NR Release 15

,

Definition:
5G NR Type I CSI Reporting

Type I CSI is a single-beam codebook. The UE reports:

RI (Rank Indicator): number of spatial layers $\nu \in \{1, 2, \ldots, 8\}$ .
PMI (Precoding Matrix Indicator): selects one or two beams from a DFT-based grid.
CQI (Channel Quality Indicator): suggested MCS level.

For a dual-polarized antenna panel with $N_1 \times N_2$ antenna ports per polarization, the codebook is parameterized by beam indices $(l, m)$ sampling the 2D angular domain:

$\mathbf{v}_{l,m} = \frac{1}{\sqrt{2 N_1 N_2}} \begin{bmatrix} \mathbf{u}_l \otimes \mathbf{v}_m \\ \varphi \, \mathbf{u}_l \otimes \mathbf{v}_m \end{bmatrix},$

where $\mathbf{u}_l \in \mathbb{C}^{N_1}$ and $\mathbf{v}_m \in \mathbb{C}^{N_2}$ are oversampled DFT vectors and $\varphi \in \{1, j, -1, -j\}$ is a co-phasing factor between polarizations. The oversampling factors $(O_1, O_2)$ determine the angular resolution.

Feedback overhead: For rank 1, the PMI requires $\lceil\log_2(O_1 N_1)\rceil + \lceil\log_2(O_2 N_2)\rceil + 2$ bits (beam indices + co-phasing).

Type I CSI is a wideband report — the same PMI applies to all subbands. It is efficient for LOS and single-cluster channels but cannot capture frequency-selective beam patterns.

Definition:
5G NR Type II CSI Reporting

Type II CSI represents the precoding vector as a linear combination of $L$ beams from the DFT grid, with per-subband amplitude and phase coefficients:

$\mathbf{v}_{k}^{(\text{sb})} = \sum_{i=1}^{L} c_{k,i} \, \mathbf{v}_{l_i, m_i},$

where the beam indices $(l_i, m_i)$ are reported wideband (same for all subbands) and the combination coefficients $c_{k,i} = p_{k,i} e^{j\phi_{k,i}}$ are reported per subband:

Amplitude $p_{k,i}$ : quantized to 1 or 3 bits (wideband) + 1 bit (differential subband).
Phase $\phi_{k,i}$ : quantized to 2 or 3 bits (QPSK or 8-PSK per subband).

The number of beams $L \in \{2, 3, 4\}$ is configured by the BS. The wideband beam selection reduces the search space; the subband coefficients capture frequency selectivity.

Feedback overhead: For $L = 4$ beams, rank 1, 13 subbands, 3-bit phase: $\approx 4 \times 2 + 4 \times 13 \times (1+3) \approx 216$ bits per report.

Type II provides significantly better CSI quality than Type I (3–5 dB SNR gain in MU-MIMO), at the cost of higher feedback overhead and UE complexity. It is the primary CSI mechanism for FDD massive MIMO in 5G NR deployments.

Example: Type II CSI Overhead Calculation

A 5G NR FDD system uses a $8 \times 4$ dual-polarized panel ( $N_1 = 8, N_2 = 4$ , total $N_t = 64$ antenna ports). Type II CSI is configured with $L = 4$ beams, rank $\nu = 2$ , oversampling $(O_1, O_2) = (4, 4)$ , 3-bit subband phase quantization, and 13 subbands. Compute the total feedback bits and compare with Type I.

Solution

Wideband beam selection

The beam grid has $O_1 N_1 \times O_2 N_2 = 32 \times 16 = 512$ directions. Selecting $L = 4$ beams requires $\lceil\log_2\binom{512}{4}\rceil \approx 34$ bits (in practice, NR uses a structured selection with fewer bits).

Wideband amplitude and co-phasing

Per layer: $L = 4$ beam amplitudes (3 bits each) + 1 co-phasing factor (2 bits). For $\nu = 2$ layers: $2 \times (4 \times 3 + 2) = 28$ bits.

Subband coefficients

Per layer per subband: $L = 4$ phase coefficients (3 bits each) + $L - 1 = 3$ differential amplitudes (1 bit each). Per layer: $13 \times (4 \times 3 + 3) = 195$ bits. For $\nu = 2$ layers: $2 \times 195 = 390$ bits.

Total and comparison

Type II total: $34 + 28 + 390 \approx 452$ bits.

Type I total (rank 2): $2 \times (\lceil\log_2(32)\rceil + \lceil\log_2(16)\rceil + 2) = 2 \times (5 + 4 + 2) = 22$ bits.

Type II uses $\approx 20\times$ more feedback bits than Type I, but provides subband-level beam combination, achieving 3–5 dB better MU-MIMO performance in frequency-selective channels.

⚠️Engineering Note

5G NR CSI Framework: Practical Considerations

The 5G NR CSI framework (Release 15+) defines the complete pipeline for FDD massive MIMO:

CSI-RS (CSI Reference Signal): The BS transmits up to 32 antenna ports of CSI-RS (beamformed or non-precoded). The UE uses these to estimate the DL channel.
CSI report: The UE reports RI + PMI + CQI. The report can be periodic (on PUCCH), semi-persistent, or aperiodic (on PUSCH, triggered by DCI).
Codebook restriction: The BS can restrict the codebook search space (e.g., disable certain beams) to reduce UE complexity and steer the feedback.
PMI subband size: The bandwidth is divided into subbands of 4–16 PRBs for subband reporting. Larger subbands = fewer bits but coarser frequency resolution.

Timing: The CSI report latency (from CSI-RS transmission to precoder application) is typically 5–10 ms, which can be significant in high-Doppler scenarios ( $f_D > 100$ Hz).

Practical Constraints

•
Maximum 32 CSI-RS ports (NR Release 15); up to 64 in Release 17
•
Type II feedback: up to ~500 bits per report (PUSCH-based, aperiodic)
•
PUCCH can carry only up to ~100 bits → Type I only for periodic reporting
•
CSI report delay: 5–10 ms → CSI aging in vehicular scenarios

📋 Ref: 3GPP TS 38.214, §5.2

Common Mistake: Codebook Design Assumes Isotropic Channels

Mistake:

Designing or analyzing codebooks under the assumption that $\mathbf{H}_{k}$ is isotropically distributed (i.i.d. Rayleigh), then deploying them in environments with strong spatial correlation. The optimal codebook for a correlated channel concentrates codewords in the dominant angular directions, not uniformly on the sphere.

Correction:

For correlated channels with covariance $\mathbf{R}_k$ , the effective channel direction $\bar{\mathbf{H}}_k$ is not uniformly distributed — it concentrates near the dominant eigenvectors of $\mathbf{R}_k$ . A DFT codebook with oversampling in the relevant angular range (as in NR Type I/II) partially addresses this. Full adaptation would require user-specific codebooks conditioned on $\mathbf{R}_k$ — which is what JSDM achieves (Section 5) by pre-beamforming to the dominant subspace.

Codebook

A finite set of precoding vectors (or matrices) known to both the BS and UE. The UE selects the best-matching entry and reports its index (PMI) to the BS. In 5G NR, codebooks are DFT-based and parameterized by beam direction, co-phasing, and (for Type II) beam combination coefficients.

Channel Quality Indicator (CQI)

A quantized measure of the channel quality reported by the UE to the BS. The CQI maps to a recommended modulation and coding scheme (MCS) index. In 5G NR, CQI is reported alongside RI and PMI as part of the CSI report.

Quick Check

Which statement best describes the key difference between 5G NR Type I and Type II CSI?

Type I uses DFT beams; Type II uses Grassmannian codewords

Type I selects a single beam; Type II linearly combines multiple beams with subband coefficients

Type I is for TDD; Type II is for FDD

Type I uses more feedback bits than Type II