Ferkans — Interactive Telecom Tutor

From Theory to Deployment

The theoretical analysis of JSDM assumes known covariance matrices, perfect eigenspace estimation, and idealized grouping. In practice, these must be learned from data, quantized for feedback, and updated as the environment changes. This section addresses the key practical challenges and connects JSDM to the 5G NR framework.

Definition:
Group Formation Algorithm

Given estimated covariances $\{\hat{\mathbf{R}}_k\}_{k=1}^{K}$ , a group formation algorithm partitions users into groups $\{\mathcal{S}_g\}$ that minimize inter-group interference while maintaining sufficient multiplexing gain within each group. A practical approach uses hierarchical clustering:

Compute the dominant $r_k$ -dimensional eigenspace $\hat{\mathbf{U}}_k^{(r_k)}$ for each user.
Define a pairwise distance: $d(k, j) = d_c(\hat{\mathbf{U}}_k^{(r_k)}, \hat{\mathbf{U}}_j^{(r_j)})$ (chordal distance).
Apply agglomerative clustering with a threshold $d_{\max}$ to form groups.
For each group $g$ , compute $\bar{\mathbf{R}}_g$ and $\mathbf{B}_g$ .

The grouping threshold $d_{\max}$ controls the tradeoff: a small $d_{\max}$ produces many small groups (low intra-group interference but limited multiplexing), while a large $d_{\max}$ produces fewer, larger groups (more multiplexing but higher intra-group heterogeneity).

Hierarchical User Grouping for JSDM

Complexity:

O(K^{2} N_t^{2})

for pairwise distance computation;

O(K^{2} \log K)

for clustering.

Input: Covariance estimates

\{\hat{\mathbf{R}}_k\}_{k=1}^{K}

, threshold

d_{\max}

, energy capture

1-\epsilon

Output: Groups

\{\mathcal{S}_g\}_{g=1}^G

, pre-beamformers

\{\mathbf{B}_g\}_{g=1}^G

1. for

k = 1, \ldots, K

do

2.

\quad

Eigendecompose

\hat{\mathbf{R}}_k

, determine

r_k

at threshold

1-\epsilon

3.

\quad

Extract

\hat{\mathbf{U}}_k^{(r_k)}

4. end for

5. Compute pairwise chordal distance matrix

\mathbf{D}

with

D_{kj} = d_c(\hat{\mathbf{U}}_k, \hat{\mathbf{U}}_j)

6. Apply agglomerative clustering on

\mathbf{D}

with linkage "complete" and cutoff

d_{\max}

7. for each cluster

g = 1, \ldots, G

do

8.

\quad

\bar{\mathbf{R}}_g = \frac{1}{|\mathcal{S}_g|} \sum_{k \in \mathcal{S}_g} \hat{\mathbf{R}}_k

9.

\quad

Eigendecompose

\bar{\mathbf{R}}_g

, set

\mathbf{B}_g = \hat{\mathbf{U}}_g^{(r_g)}

10. end for

In practice, the covariance estimation and grouping are performed on a slow timescale (every 100–1000 ms), so the $O(K^{2} N_t^{2})$ cost is amortized.

Definition:
Beam-Domain CSI

Instead of feeding back the effective channel vector $\tilde{\mathbf{h}}_k \in \mathbb{C}^{r_g}$ directly, a user can report beam-domain CSI: the indices and complex gains of the strongest beams from a predefined codebook (e.g., the DFT beams). This provides a compressed representation:

$\tilde{\mathbf{h}}_k \approx \sum_{i=1}^{L} \alpha_{k,i} \, \mathbf{b}_{n_i}$

where $\{n_1, \ldots, n_L\}$ are the indices of the $L$ strongest beams and $\alpha_{k,i}$ are the corresponding complex gains. The feedback consists of $L$ indices + $L$ complex scalars, with $L \leq r_g$ .

Beam-domain CSI

A compressed representation of the channel in terms of indices and gains of the strongest beams from a predefined spatial codebook. It is the natural CSI format for JSDM and forms the conceptual basis for 5G NR Type II codebook feedback.

Related: {{Ref:Gloss Pre Beamformer}}

Example: Connection to 5G NR Type II Codebook

Describe how the 5G NR Type II CSI codebook implements a JSDM-like two-stage structure, and identify the correspondence between JSDM components and the 3GPP parameters.

Solution

Wideband component (analogous to $\mathbf{B}_g$)

The UE selects $L$ beams from a DFT-based oversampled codebook (the "wideband" component). These $L$ beam indices are reported and remain fixed across the bandwidth. This corresponds to selecting the DFT columns for the pre-beamformer $\mathbf{B}_g$ .

Subband component (analogous to $\mathbf{P}_g$)

For each subband (a group of contiguous subcarriers), the UE reports amplitude and phase coefficients for the selected beams. This subband-specific information is analogous to the inner precoder $\mathbf{P}_g$ that adapts to the instantaneous effective channel.

Parameter mapping

JSDM	5G NR Type II
$\mathbf{B}_g$ (eigenspace projection)	Wideband beam selection ( $L$ DFT beams)
$\mathbf{P}_g$ (inner MU-MIMO)	Subband amplitude/phase coefficients
$r_g$ (effective rank)	Number of selected beams $L$ ( $L \in \{2, 3, 4\}$ )
Group angular region	Beam group restriction set

Key difference

In 5G NR, the UE performs beam selection (choosing which DFT beams to report), whereas in the original JSDM, the base station determines the pre-beamformer from covariance knowledge. The practical effect is similar: both reduce the effective channel dimension from $N_t$ to $L$ .

Effective Channel Dimension vs. Eigenvalue Threshold

Explore how the effective rank $r_g$ changes with the energy capture threshold and the angular spread. This plot directly shows the tradeoff between CSI overhead (proportional to $r_g$ ) and signal energy preserved by the pre-beamformer.

Parameters

N_t

64

\Delta\theta

(deg)15

\theta_c

(deg)0

⚠️Engineering Note

Covariance Estimation in Practice

Estimating $\mathbf{R}_k$ requires averaging instantaneous channel samples: $\hat{\mathbf{R}}_k = \frac{1}{N_s} \sum_{n=1}^{N_s} \mathbf{h}_k^{(n)} (\mathbf{h}_k^{(n)})^H$ . The number of samples $N_s$ required for a reliable estimate depends on the effective rank $r_k$ and the eigenvalue spread. A rule of thumb is $N_s \geq 5 r_k$ for $r_k$ -dimensional covariance estimation. In mobile environments, the covariance changes on the order of the stationarity interval of the scattering environment (typically 100 ms–1 s), so $N_s$ samples must be collected within this window. At 1 ms slot duration, $N_s = 30{-}100$ samples are available, which is sufficient for $r_k \leq 10{-}20$ .

Practical Constraints

•
Covariance stationarity interval limits the number of averaging samples
•
Mobile users at high speed ( $> 60$ km/h) may have covariance update rates comparable to the coherence time
•
In FDD, covariance must be estimated from uplink signals with frequency-domain extrapolation or from downlink CSI-RS feedback

Historical Note: The FDD Challenge in Massive MIMO

2010–2014

When Marzetta's seminal 2010 paper launched the massive MIMO era, it was explicitly built on TDD reciprocity — uplink pilots scale with $K$ , not $N_t$ . The consensus at the time was that massive MIMO was fundamentally a TDD technology. The FDD challenge appeared insurmountable: $N_t$ downlink pilots and $N_t$ -dimensional feedback per user seemed to preclude FDD operation with large arrays. JSDM (2013) and its beam-domain extension by Nam et al. (2014) showed that FDD massive MIMO is feasible after all, by exploiting the spatial structure that massive arrays themselves create. This insight profoundly influenced the design of 5G NR's CSI framework.

Common Mistake: Stale Covariance in Mobile Scenarios

Mistake:

Computing the pre-beamformer $\mathbf{B}_g$ from a covariance estimate that is several seconds old, then using it for inner precoding on the current channel. In high-mobility scenarios, the covariance eigenspace may have rotated significantly.

Correction:

Monitor the covariance stationarity by tracking the chordal distance between successive estimates: $d_c(\hat{\mathbf{U}}_k^{(t)}, \hat{\mathbf{U}}_k^{(t-\Delta t)})$ . If this distance exceeds a threshold, trigger a covariance update and re-group the users. In 5G NR, the CSI-RS periodicity and reporting triggers are designed to handle exactly this scenario — the network can configure more frequent wideband CSI reports for high-mobility users.

Inter-Group Angular Spectrum Overlap

Visualize the angular power spectra of two user groups and quantify the overlap that leads to inter-group interference. When the spectra are well-separated, JSDM provides near-optimal performance; when they overlap, the rate penalty grows.

Parameters

N_t

64

\theta_1

(deg)-15

\theta_2

(deg)15

\Delta\theta

(deg)15

Quick Check

In 5G NR Type II codebook feedback, the UE selects $L$ beams from a DFT-based codebook. Which JSDM component does this beam selection correspond to?

The inner MU-MIMO precoder $\mathbf{P}_g$

The pre-beamforming matrix $\mathbf{B}_g$

The group formation algorithm

The channel covariance $\mathbf{R}_k$

Correction:

The pre-beamforming matrix

\mathbf{B}_g

The selected DFT beams form the spatial basis for the effective channel, exactly as $\mathbf{B}_g$ selects the covariance eigenspace.

Chordal distance

A metric on the Grassmann manifold measuring the distance between two subspaces: $d_c(\mathbf{U}, \mathbf{V}) = \frac{1}{\sqrt{2}} \|\mathbf{U}\mathbf{U}^H - \mathbf{V}\mathbf{V}^H\|_F$ . It ranges from $0$ (identical subspaces) to $1$ (orthogonal subspaces) and is used in JSDM for user grouping and covariance tracking.

Stationarity interval

The time duration over which the second-order channel statistics (covariance, angular power spectrum) remain approximately constant. It is much longer than the coherence time and typically ranges from 100 ms to several seconds, depending on user mobility and the scattering environment.

Practical Considerations

From Theory to Deployment

Definition: Group Formation Algorithm

Hierarchical User Grouping for JSDM

Definition: Beam-Domain CSI

Beam-domain CSI

Example: Connection to 5G NR Type II Codebook

Wideband component (analogous to $\mathbf{B}_g$)

Subband component (analogous to $\mathbf{P}_g$)

Parameter mapping

Key difference

Effective Channel Dimension vs. Eigenvalue Threshold

Parameters

Covariance Estimation in Practice

Historical Note: The FDD Challenge in Massive MIMO

Common Mistake: Stale Covariance in Mobile Scenarios

Inter-Group Angular Spectrum Overlap

Parameters

Quick Check

Chordal distance

Stationarity interval

Definition:
Group Formation Algorithm

Definition:
Beam-Domain CSI