Ferkans — Interactive Telecom Tutor

The CSI Feedback Bottleneck in FDD Massive MIMO

In FDD massive MIMO, the base station cannot exploit TDD reciprocity to learn the downlink channel. Instead, it must send $N_t$ downlink pilots and each user must feed back a $N_t$ -dimensional channel estimate. When $N_t$ is large (64, 128, or more), this overhead can consume a significant fraction of the coherence interval, leaving little room for data transmission. JSDM's dimensionality reduction is the mechanism that breaks this bottleneck: by projecting onto the $r_g$ -dimensional covariance eigenspace, the feedback cost per user drops from $N_t$ to $r_g$ .

Theorem: CSI Overhead Reduction via Pre-Beamforming

With JSDM two-stage precoding, the total CSI feedback per coherence interval is

$C_{\text{JSDM}} = \sum_{g=1}^{G} |\mathcal{S}_g| \cdot r_g$

complex scalars, compared to

$C_{\text{full}} = K \cdot N_t$

for full-dimensional precoding. The feedback reduction ratio is

$\rho = \frac{C_{\text{JSDM}}}{C_{\text{full}}} = \frac{\sum_{g=1}^{G} |\mathcal{S}_g| \cdot r_g}{K \cdot N_t} \leq \frac{\max_g r_g}{N_t}.$

For typical massive MIMO parameters with limited angular spread, $r_g / N_t \ll 1$ , yielding an order-of-magnitude feedback reduction.

Each user only needs to estimate and feed back a $r_g$ -dimensional effective channel instead of a $N_t$ -dimensional full channel. Since $r_g$ is controlled by the angular spread of the scattering environment (not by $N_t$ ), the overhead becomes independent of the array size — a dramatic advantage as arrays grow.

Proof

Pilot overhead

In standard FDD, the base station transmits $N_t$ orthogonal downlink pilots so each user can estimate $\mathbf{h}_k \in \mathbb{C}^{N_t}$ . With JSDM, the base station transmits group-specific pilots through $\mathbf{B}_g$ : for group $g$ , only $r_g$ pilots are needed (transmitted as $\mathbf{B}_g \boldsymbol{\Phi}_g$ where $\boldsymbol{\Phi}_g \in \mathbb{C}^{r_g \times r_g}$ is a pilot matrix).

Feedback overhead

Each user $k \in \mathcal{S}_g$ estimates $\tilde{\mathbf{h}}_k = \mathbf{B}_g^H \mathbf{h}_k \in \mathbb{C}^{r_g}$ and feeds back $r_g$ complex scalars (or a quantized version). The total across all users is $\sum_g |\mathcal{S}_g| \cdot r_g$ .

Reduction ratio

Dividing: $\rho = \frac{\sum_g |\mathcal{S}_g| r_g}{K N_t}$ . Since $|\mathcal{S}_g| r_g \leq |\mathcal{S}_g| \max_g r_g$ and $\sum_g |\mathcal{S}_g| = K$ , we get $\rho \leq \frac{\max_g r_g}{N_t}$ . $\blacksquare$

Definition:
DFT-Based Pre-Beamformer

For a ULA with $N_t$ antennas and half-wavelength spacing, a computationally efficient choice for $\mathbf{B}_g$ uses columns of the $N_t$ -point DFT matrix. Define the DFT matrix $\mathbf{F} \in \mathbb{C}^{N_t \times N_t}$ with entries

$[\mathbf{F}]_{m,n} = \frac{1}{\sqrt{N_t}} e^{-j2\pi mn / N_t}, \quad m,n = 0, \ldots, N_t-1.$

The DFT pre-beamformer for group $g$ selects the $r_g$ columns of $\mathbf{F}$ corresponding to the angular region of group $g$ :

$\mathbf{B}_g = \mathbf{F}(:, \mathcal{I}_g)$

where $\mathcal{I}_g \subset \{0, \ldots, N_t-1\}$ is the set of $r_g$ DFT beam indices covering group $g$ 's angular support.

The DFT pre-beamformer is appealing because (i) it requires no eigendecomposition, (ii) it can be implemented efficiently using the FFT, and (iii) it provides a natural grid of angular beams that partitions the spatial domain. However, it is optimal only for the ULA with half-wavelength spacing; for other array geometries, the covariance eigenvectors are preferred.

Example: DFT Beam Selection for Two Groups

A ULA with $N_t = 32$ antennas serves two groups. Group 1 arrives from $\theta \in [-30°, -10°]$ and Group 2 from $\theta \in [10°, 30°]$ . Using the DFT pre-beamformer, determine the beam indices $\mathcal{I}_1$ and $\mathcal{I}_2$ .

Solution

Map angle to DFT index

For a ULA with half-wavelength spacing, angle $\theta$ maps to spatial frequency $\nu = \frac{1}{2}\sin\theta$ . DFT bin $n$ corresponds to $\nu_n = n/N_t$ (modulo 1). Thus angle $\theta$ maps to DFT index $n(\theta) = \lfloor N_t \cdot \frac{1}{2}\sin\theta \rfloor \mod N_t$ .

Group 1 beams

$\sin(-30°) = -0.5$ , $\sin(-10°) \approx -0.174$ . Spatial frequencies: $\nu \in [-0.25, -0.087]$ . DFT indices (wrapped): $n \in \{24, 25, 26, 27, 28, 29\}$ , giving $r_1 = 6$ beams.

Group 2 beams

$\sin(10°) \approx 0.174$ , $\sin(30°) = 0.5$ . Spatial frequencies: $\nu \in [0.087, 0.25]$ . DFT indices: $n \in \{3, 4, 5, 6, 7, 8\}$ , giving $r_2 = 6$ beams.

Orthogonality check

Since $\mathcal{I}_1 \cap \mathcal{I}_2 = \emptyset$ , the DFT beams of the two groups are exactly orthogonal: $\mathbf{B}_1^H \mathbf{B}_2 = \mathbf{0}$ . Inter-group interference is zero at the pre-beamformer level.

CSI Overhead: Full CSI vs. JSDM

Compare the CSI feedback overhead (in complex scalars per coherence interval) between full-dimensional precoding and JSDM as functions of the number of antennas $N_t$ . JSDM's overhead is controlled by the effective rank $r_g$ , which depends on the angular spread rather than $N_t$ .

Parameters

N_t^{\max}

128

Maximum number of antennas

K

8

G

4

\Delta\theta

(deg)10

🚨Critical Engineering Note

Pilot Overhead in FDD Massive MIMO

In a practical FDD system, the downlink pilot overhead scales as $\tau_p^{\text{DL}} / T_c$ where $\tau_p^{\text{DL}}$ is the number of pilot symbols and $T_c$ is the coherence interval in symbols. Without JSDM, $\tau_p^{\text{DL}} = N_t$ ; with JSDM, $\tau_p^{\text{DL}} = \sum_g r_g$ (if groups are trained sequentially) or $\max_g r_g$ (if groups are trained in parallel with orthogonal resources). For $N_t = 128$ and $r_g = 6$ with $G = 4$ groups, sequential training requires $\tau_p^{\text{DL}} = 24$ pilots vs. $128$ — a $5.3\times$ reduction. In a typical 5G NR deployment with $T_c \approx 200$ symbols (14 OFDM symbols per slot at 30 kHz SCS, coherence time $\sim$ 1 ms), this difference between $64\%$ overhead (128/200) and $12\%$ overhead (24/200) is the difference between an infeasible and a viable system.

Practical Constraints

•
Pilot overhead must remain below ~20% of the coherence interval for acceptable spectral efficiency
•
Group-specific pilots require synchronization between pre-beamformer updates and pilot scheduling
•
Quantization of the effective channel feedback adds additional overhead not captured in the ideal analysis

Theorem: Inter-Group Pilot Reuse

If the covariance eigenspaces of groups $g$ and $g'$ are orthogonal ( $\mathbf{U}_g^H \mathbf{U}_{g'} = \mathbf{0}$ ), then groups $g$ and $g'$ can reuse the same pilot sequences without contamination. In this case, the total pilot overhead is

$\tau_p^{\text{DL}} = \max_g r_g$

rather than $\sum_g r_g$ , because orthogonal groups do not interfere with each other's channel estimation.

When a user in group $g$ projects the received pilot signal through $\mathbf{B}_g^H$ , pilots transmitted by group $g'$ with $\mathbf{B}_{g'}$ are annihilated if the eigenspaces are orthogonal. This is the spatial analog of frequency-domain orthogonality in OFDM — different groups occupy non-overlapping "spatial frequencies."

Proof

Pilot signal model

Group $g$ transmits pilots $\mathbf{B}_g \boldsymbol{\Phi}_g$ where $\boldsymbol{\Phi}_g \in \mathbb{C}^{r_g \times \tau_p}$ . User $k \in \mathcal{S}_g$ receives $\mathbf{y}_k^{\text{pilot}} = \mathbf{h}_k^H \sum_{g'=1}^{G} \mathbf{B}_{g'} \boldsymbol{\Phi}_{g'} + \mathbf{w}_{k}$ .

Pre-beamforming projection

Projecting through $\mathbf{B}_g$ : $\mathbf{B}_g^H \mathbf{y}_k^{\text{pilot}} = \tilde{\mathbf{h}}_k^H \boldsymbol{\Phi}_g + \sum_{g' \neq g} \underbrace{\mathbf{B}_g^H \mathbf{h}_k \cdot \mathbf{B}_{g'}}_{\approx \mathbf{0}} \boldsymbol{\Phi}_{g'} + \mathbf{B}_g^H \mathbf{w}_{k}$ .

Orthogonality eliminates cross-group contamination

When $\mathbf{U}_g^H \mathbf{U}_{g'} = \mathbf{0}$ , we have $\mathbf{B}_g^H \mathbf{B}_{g'} = \mathbf{0}$ and thus $\mathbf{h}_k^H \mathbf{B}_{g'} \approx \mathbf{0}$ (the approximation tightens as $N_t \to \infty$ ). The cross-group pilot interference vanishes, and all $G$ groups can share the same $\max_g r_g$ pilot resources. $\blacksquare$

Quick Check

A base station with $N_t = 128$ antennas serves $K = 16$ users in $G = 4$ groups with effective rank $r_g = 8$ per group. What is the CSI feedback reduction ratio $\rho = C_{\text{JSDM}} / C_{\text{full}}$ ?

$\rho = 1/2$

$\rho = 1/16$

$\rho = 1/4$

$\rho = 1/64$

Correction:

\rho = 1/16

$C_{\text{JSDM}} = 16 \times 8 = 128$ , $C_{\text{full}} = 16 \times 128 = 2048$ . So $\rho = 128/2048 = 1/16 = 0.0625$ .

Common Mistake: Choosing the Effective Rank $r_g$ Too Aggressively

Mistake:

Setting $r_g$ to the minimum value that captures $90\%$ of the covariance trace, then finding that the sum rate degrades severely because the pre-beamformer discards signal energy in the tail eigenvectors.