Ferkans — Interactive Telecom Tutor

JSDM Revisited: The FDD Perspective

In Chapter 7, we introduced JSDM (Joint Spatial Division and Multiplexing) as a two-stage precoding framework that exploits the spatial covariance structure of user channels. The motivating application was MU-MIMO with reduced-dimension channel estimation. Now we revisit JSDM from a specifically FDD perspective: the pre-beamforming matrix $\mathbf{B}_g$ , designed from the long-term covariance $\mathbf{R}_k$ (which can be estimated from UL measurements even in FDD), reduces the effective channel dimension from $N_t$ to $r_g$ . This means both the DL pilot overhead and the feedback dimension scale with $r_g \ll N_t$ — a fundamental improvement over generic FDD massive MIMO.

Definition:
JSDM System Model for FDD

Recall the JSDM framework from Chapter 7. Users are partitioned into $G$ groups based on their spatial covariance: group $g$ contains users whose covariance matrices $\mathbf{R}_k$ share a common approximate eigenspace. For each group $g$ :

Covariance eigendecomposition: $\mathbf{R}_g \approx \mathbf{U}_g \mathbf{\Lambda}_g \mathbf{U}_g^H$ where $\mathbf{U}_g \in \mathbb{C}^{N_t \times r_g}$ contains the $r_g$ dominant eigenvectors.
Pre-beamforming: $\mathbf{B}_g = \mathbf{U}_g \in \mathbb{C}^{N_t \times r_g}$ .
Effective channel: For user $k$ in group $g$ , $\mathbf{H}_{\text{eff},k} = \mathbf{B}_g^H \mathbf{H}_{k} \in \mathbb{C}^{r_g}$ .

The two-stage precoder is $\mathbf{v}_{k} = \mathbf{B}_g \mathbf{v}_{\text{eff},k}$ where $\mathbf{v}_{\text{eff},k} \in \mathbb{C}^{r_g}$ is designed from $\mathbf{H}_{\text{eff},k}$ .

FDD implications: The effective channel $\mathbf{H}_{\text{eff},k}$ is $r_g$ -dimensional. The DL pilot overhead is $\tau_d = r_g$ (not $N_t$ ). The feedback dimension is $r_g$ (not $N_t$ ). Both overheads are reduced by a factor of $N_t/r_g$ .

The covariance $\mathbf{R}_k$ depends on angles and array geometry — not on the carrier frequency. This is the "partial reciprocity" property: the UL covariance is approximately equal to the DL covariance. Hence $\mathbf{B}_g$ can be designed from UL measurements alone, even in FDD. Only the instantaneous effective channel $\mathbf{H}_{\text{eff},k}$ must be estimated from DL pilots and fed back.

Theorem: JSDM Overhead Reduction in FDD

Under JSDM with group rank $r_g$ and $K_g$ users per group, the FDD overhead satisfies:

DL pilot overhead: $\tau_d^{\text{JSDM}} = r_g$ per group (vs. $\tau_d = N_t$ without JSDM).
Feedback bits: $B_{\text{fb}}^{\text{JSDM}} = 2 b r_g$ per user (vs. $2 b N_t$ ).
Data efficiency: $\eta_{\text{data}}^{\text{JSDM}} = 1 - \frac{r_g + K_g}{T_c},$ which is independent of $N_t$ and scales only with the group rank $r_g$ and the number of users $K_g$ .

The overhead reduction factor is $N_t / r_g$ . For a typical macro-cell scenario with $N_t = 128$ and $r_g = 8$ , this is a $16\times$ reduction.

JSDM confines each user's effective channel to an $r_g$ -dimensional subspace determined by its spatial covariance. The BS only needs to train and receive feedback for this subspace — not for the full $N_t$ -dimensional antenna space. The pre-beamforming matrix $\mathbf{B}_g$ "projects away" the irrelevant dimensions before any DL training or feedback occurs.

Proof

Effective pilot design

After pre-beamforming, the effective channel is $\mathbf{H}_{\text{eff},k} = \mathbf{B}_g^H \mathbf{H}_{k} \in \mathbb{C}^{r_g}$ . The BS transmits $r_g$ pilot vectors through the pre-beamformer $\mathbf{B}_g$ : the DL pilot matrix is $\mathbf{P} = \mathbf{B}_g \mathbf{\Psi}$ where $\mathbf{\Psi} \in \mathbb{C}^{r_g \times r_g}$ is unitary. User $k$ observes $\mathbf{y}_k^{\text{pilot}} = \mathbf{\Psi}^H \mathbf{H}_{\text{eff},k} + \mathbf{w}_{k}$ and estimates $\mathbf{H}_{\text{eff},k}$ from $r_g$ observations.

Feedback dimension

The UE quantizes $\mathbf{H}_{\text{eff},k} \in \mathbb{C}^{r_g}$ to $B_{\text{fb}} = 2 b r_g$ bits and transmits them to the BS. The feedback dimension is $r_g$ , independent of $N_t$ .

Data efficiency

The coherence block allocates $r_g$ symbols for DL pilots (per group) and $K_g$ for UL pilots. The data efficiency is $\eta_{\text{data}}^{\text{JSDM}} = 1 - (r_g + K_g) / T_c$ . For $r_g = 8$ , $K_g = 8$ , $T_c = 200$ : $\eta = 1 - 16/200 = 92\%$ , compared to $\eta = 1 - (128+8)/200 = 32\%$ without JSDM. $\blacksquare$

🎓CommIT Contribution(2013)

JSDM as a Structured FDD Massive MIMO Solution

A. Adhikary, J. Nam, J.-Y. Ahn, G. Caire — IEEE Transactions on Information Theory, vol. 59, no. 10

The JSDM framework, introduced by Adhikary, Nam, Ahn, and Caire, provides a principled approach to FDD massive MIMO by exploiting the low-rank structure of user channel covariances. The key insight is that in realistic propagation environments, the channel covariance $\mathbf{R}_k$ has rank $r_k \ll N_t$ because scatterers occupy a limited angular spread. By grouping users with similar covariance eigenspaces and designing per-group pre-beamformers $\mathbf{B}_g$ , JSDM:

Reduces DL training overhead from $N_t$ to $r_g$ per group,
Reduces feedback from $N_t$ to $r_g$ dimensions per user,
Provides inter-group interference suppression through orthogonal pre-beamformers,
Enables standard MU-MIMO precoding (ZF, MMSE) on the reduced-dimension effective channels.

The paper proves that JSDM achieves the same multiplexing gain as full-dimensional precoding with perfect CSI, while requiring only $r_g$ -dimensional CSI — a result that makes FDD massive MIMO practically viable in macro-cell deployments.

jsdmfdd-massive-mimospatial-covariancetwo-stage-precodingView Paper →

Example: JSDM FDD Overhead in a Macro-Cell Deployment

A macro-cell BS with $N_t = 128$ ULA antennas at 2 GHz FDD serves 3 user groups, each with angular spread $\Delta\theta_g = 10°$ . The coherence block has $T_c = 200$ symbols. Each group has $K_g = 4$ users. Compare FDD overhead with and without JSDM.

Solution

Group rank

Each group's covariance has rank $r_g \approx \lceilN_t \cdot \Delta\theta / \pi\rceil = \lceil 128 \times 10/180 \rceil = 8$ . (With guard bins: $r_g \approx 10$ .)

Without JSDM

DL pilots: $\tau_d = N_t = 128$ .
UL pilots: $\tau_u = 3 \times 4 = 12$ .
Data efficiency: $\eta = 1 - (128 + 12)/200 = 30\%$ .
Feedback per user: $2 \times 5 \times 128 = 1280$ bits.
Total feedback: $12 \times 1280 = 15{,}360$ bits.

With JSDM

DL pilots per group: $\tau_d^{(g)} = r_g = 10$ . Total: $3 \times 10 = 30$ (or $10$ if groups are trained sequentially and share pilot dimensions).
UL pilots: $\tau_u = 12$ .
Data efficiency: $\eta = 1 - (30 + 12)/200 = 79\%$ .
Feedback per user: $2 \times 5 \times 10 = 100$ bits.
Total feedback: $12 \times 100 = 1{,}200$ bits.

Improvement: Data efficiency from 30% to 79% ( $2.6\times$ ). Total feedback from 15,360 to 1,200 bits ( $12.8\times$ ). The overhead becomes manageable for FDD operation.

JSDM Effective Dimension vs. Angular Spread

Explore how the JSDM group rank $r_g$ varies with the angular spread $\Delta\theta$ for different array sizes. Narrower angular spread (typical of elevated macro-cell BS) yields smaller $r_g$ and greater overhead reduction. The plot also shows the corresponding feedback reduction factor $N_t/r_g$ .

Parameters

N_t

128

Number of BS antennas

\Delta\theta

(degrees)10

Angular spread per group

Energy threshold0.95

Fraction of covariance energy captured by dominant eigenmodes

Sum Rate: Perfect CSI vs. Quantized vs. JSDM

Compare the achievable sum rate under three CSI regimes: (1) perfect CSI at the BS (genie-aided), (2) codebook-based feedback with $B$ bits per user, (3) JSDM with pre-beamforming and $r_g$ -dimensional feedback. JSDM approaches the perfect-CSI rate more closely than codebook feedback at the same total feedback budget.

Parameters

N_t

64

Number of BS antennas

K

8

Number of users

B_{\\text{fb}}

10

Feedback bits per user

\Delta\theta

(degrees)10

Angular spread per group

Theorem: JSDM Preserves Multiplexing Gain

Under JSDM with $G$ groups, group ranks $\{r_g\}$ , and $K_g$ users per group (with $K_g \leq r_g$ ), the sum multiplexing gain satisfies

$d_{\text{sum}}^{\text{JSDM}} = \sum_{g=1}^{G} \min(K_g, r_g) = \sum_{g=1}^{G} K_g = K,$

which equals the multiplexing gain with perfect CSI, provided the groups have non-overlapping angular supports.

Each group's pre-beamformer $\mathbf{B}_g$ "opens" $r_g$ spatial dimensions for group $g$ . Within these dimensions, $K_g \leq r_g$ users can be spatially multiplexed using ZF or MMSE precoding. If the groups have disjoint angular supports, the pre-beamformers are approximately orthogonal, so inter-group interference is automatically suppressed. The total multiplexing gain equals the total number of simultaneously served users — the same as with full-dimensional perfect CSI.

Proof

Per-group degrees of freedom

Group $g$ has $r_g$ spatial dimensions (columns of $\mathbf{B}_g$ ). With $K_g$ users and ZF precoding on $\mathbf{H}_{\text{eff},k} \in \mathbb{C}^{r_g}$ , the per-group multiplexing gain is $\min(K_g, r_g) = K_g$ (since $K_g \leq r_g$ by design).

Inter-group orthogonality

If the angular supports of groups $g$ and $g'$ are disjoint (no overlapping beams), then $\mathbf{B}_g^H \mathbf{R}_{g'} \mathbf{B}_g \approx \mathbf{0}$ . The pre-beamformers null out inter-group interference without using any group- $g$ degrees of freedom.

Sum multiplexing gain

Summing over groups: $d_{\text{sum}} = \sum_{g=1}^G K_g = K$ . This equals the perfect-CSI multiplexing gain $\min(K, N_t) = K$ (when $K \leq N_t$ ). $\blacksquare$

⚠️Engineering Note

Practical JSDM Implementation Considerations

Deploying JSDM in a real FDD system requires addressing several practical issues:

Covariance estimation: $\mathbf{R}_k$ changes slowly (seconds to minutes) and can be estimated from UL SRS using partial reciprocity. The estimation requires averaging over many coherence blocks, introducing a latency-accuracy tradeoff.
User grouping: Grouping users by covariance similarity is a clustering problem. K-means on the dominant eigenvectors of $\mathbf{R}_k$ works well in practice. The number of groups $G$ must balance inter-group interference (fewer groups = more overlap) against scheduling flexibility (more groups = fewer users per group).
Pre-beamformer update rate: $\mathbf{B}_g$ depends on $\mathbf{R}_k$ , which changes slowly. Updating $\mathbf{B}_g$ every 100–1000 ms suffices, compared to per-slot updates for the inner precoder $\mathbf{v}_{\text{eff},k}$ .
Angular overlap: When groups have overlapping angular supports, JSDM's inter-group interference suppression degrades. Solutions include: generalized block diagonalization, regularized pre-beamformers, and joint optimization of $\mathbf{B}_g$ across groups.

Practical Constraints

•
Covariance estimation delay: 100–1000 ms (limits adaptation to mobile users)
•
Angular overlap between groups degrades multiplexing gain by 10–30% in dense urban
•
Pre-beamformer storage: $G \times N_t \times r_g$ complex values at BS
•
Requires minimum 2 groups for meaningful overhead reduction

Comparison of FDD CSI Feedback Approaches

Approach	DL pilot overhead	Feedback bits/user	CSI quality	Requires
Naive (element-wise)	$N_t$	$2 b N_t$	Good (high $b$ )	Nothing extra
Compressed (CS)	$N_t$	$O(r_k \log N_t)$	Good (sparse channels)	Sparsity in angular domain
Codebook (Type I)	$N_t$	$\sim 10$ – $30$ bits	Moderate	Predefined DFT codebook
Codebook (Type II)	$N_t$	$\sim 200$ – $500$ bits	Good	Multi-beam combination
CsiNet (DL)	$N_t$	$2 \gamma N_\tau N_t$	Best at low $\gamma$	Training data, UE NN
JSDM	$r_g \ll N_t$	$2 b r_g$	Good	UL covariance estimation

, ,

JSDM + Deep Learning: A Natural Combination

JSDM and deep learning CSI compression are complementary, not competing. JSDM reduces the channel dimension from $N_t$ to $r_g$ via linear pre-beamforming; a CsiNet-style autoencoder can then further compress the $r_g$ -dimensional effective channel. The combined system has:

DL pilot overhead: $r_g$ (from JSDM),
Feedback: learned compression of $\mathbf{H}_{\text{eff},k} \in \mathbb{C}^{r_g}$ ,
BS reconstruction: JSDM pre-beamformer + CsiNet decoder.

This combination achieves the best of both worlds: model-based dimensionality reduction (JSDM) with data-driven compression (CsiNet). We expect this hybrid approach to play a central role in future FDD massive MIMO systems.

,

Key Takeaway

JSDM is the most principled approach to FDD massive MIMO because it reduces both the DL pilot overhead and the feedback dimension from $N_t$ to the group rank $r_g$ , while preserving the full multiplexing gain. The overhead reduction factor $N_t/r_g$ is large in macro-cell deployments (narrow angular spread). JSDM requires only partial reciprocity (UL covariance estimation), which is available in FDD. It can be combined with any per-user feedback scheme (codebook, CS, deep learning) for further compression.

Partial Reciprocity

The property that the spatial covariance matrix $\mathbf{R}_k = \mathbb{E}[\mathbf{H}_{k} \mathbf{H}_{k}^{H}]$ is approximately the same at the UL and DL frequencies, even though the instantaneous channel realizations differ. This holds because the covariance depends on angles of arrival/departure and array geometry (frequency-independent to first order), not on the carrier wavelength. Partial reciprocity enables JSDM's UL-based covariance estimation for FDD pre-beamforming.

Pre-Beamforming

The first stage of JSDM's two-stage precoding, implemented by the matrix $\mathbf{B}_g$ . Pre-beamforming projects the transmitted signal into the dominant eigenspace of group $g$ 's covariance, reducing the effective channel dimension and providing inter-group interference suppression. Pre-beamforming is designed from long-term statistics and updated infrequently (every 100–1000 ms).

Quick Check

What is the primary advantage of JSDM over generic codebook-based FDD feedback?

JSDM eliminates the need for any CSI feedback

JSDM reduces the DL pilot overhead from $N_t$ to $r_g$ and the feedback dimension from $N_t$ to $r_g$

JSDM works only in TDD mode

JSDM uses deep learning at the UE

Correction:

JSDM reduces the DL pilot overhead from

N_t

to

r_g

and the feedback dimension from

N_t

to

r_g

By pre-beamforming to the group eigenspace, JSDM confines the channel to $r_g$ dimensions. Both pilot and feedback overhead scale with $r_g \ll N_t$ instead of $N_t$ .

JSDM as a Structured FDD Solution

JSDM Revisited: The FDD Perspective

Definition: JSDM System Model for FDD

Theorem: JSDM Overhead Reduction in FDD

Effective pilot design

Feedback dimension

Data efficiency

JSDM as a Structured FDD Massive MIMO Solution

Example: JSDM FDD Overhead in a Macro-Cell Deployment

Group rank

Without JSDM

With JSDM

JSDM Effective Dimension vs. Angular Spread

Parameters

Sum Rate: Perfect CSI vs. Quantized vs. JSDM

Parameters

Theorem: JSDM Preserves Multiplexing Gain

Per-group degrees of freedom

Inter-group orthogonality

Sum multiplexing gain

Practical JSDM Implementation Considerations

Comparison of FDD CSI Feedback Approaches

JSDM + Deep Learning: A Natural Combination

Key Takeaway

Partial Reciprocity

Pre-Beamforming

Quick Check

Definition:
JSDM System Model for FDD