Ferkans — Interactive Telecom Tutor

Why Spatial Correlation Is a Resource, Not a Nuisance

In classical MIMO theory, spatial correlation is viewed as harmful — it reduces the rank of the channel and degrades multiplexing gain. But in massive MIMO with many more antennas than users ( $N_t \gg K$ ), spatial correlation becomes a structural resource. Users whose signals arrive from similar angular directions share a common channel subspace, and this subspace can be estimated from long-term statistics alone — no instantaneous CSI is needed. The JSDM framework turns this observation into a practical system architecture.

Definition:
Spatial Covariance Matrix

Consider a base station equipped with $N_t$ antennas serving user $k$ . The downlink channel vector $\mathbf{h}_k \in \mathbb{C}^{N_t}$ is drawn from a distribution with zero mean and spatial covariance

$\mathbf{R}_k \triangleq \mathbb{E}[\mathbf{h}_k \mathbf{h}_k^H] \in \mathbb{C}^{N_t \times N_t}.$

The covariance $\mathbf{R}_k$ captures the long-term spatial statistics of user $k$ 's channel. It depends on the scattering environment, antenna geometry, and angular power spectrum, but it changes on a much slower timescale than the instantaneous channel realization $\mathbf{h}_k$ .

Spatial covariance matrix

The matrix $\mathbf{R}_k = \mathbb{E}[\mathbf{h}_k \mathbf{h}_k^H]$ that characterizes the second-order spatial statistics of user $k$ 's channel vector. It determines the subspace in which the channel lives and evolves on a slow timescale compared to fast fading.

Definition:
Covariance Eigenspace and Effective Rank

The eigendecomposition of $\mathbf{R}_k$ is

$\mathbf{R}_k = \mathbf{U}_k \boldsymbol{\Lambda}_k \mathbf{U}_k^H$

where $\mathbf{U}_k \in \mathbb{C}^{N_t \times N_t}$ is unitary and $\boldsymbol{\Lambda}_k = \text{diag}(\lambda_1^{(k)}, \ldots, \lambda_{N_t}^{(k)})$ with eigenvalues ordered $\lambda_1^{(k)} \geq \lambda_2^{(k)} \geq \cdots \geq 0$ . The effective rank of user $k$ 's channel is

$r_k \triangleq \min\left\{ r : \frac{\sum_{i=1}^{r} \lambda_i^{(k)}}{\sum_{i=1}^{N_t} \lambda_i^{(k)}} \geq 1 - \epsilon \right\}$

for a threshold $\epsilon > 0$ (typically $\epsilon \in [0.01, 0.05]$ ). The dominant $r_k$ eigenvectors $\mathbf{U}_k^{(r_k)} \in \mathbb{C}^{N_t \times r_k}$ span the covariance eigenspace — the subspace that captures essentially all the channel energy.

In massive MIMO with a ULA and limited angular spread, $r_k \ll N_t$ for each user. This dimensional collapse is the key enabler of JSDM.

Covariance eigenspace

The column space of the dominant eigenvectors of the spatial covariance matrix $\mathbf{R}_k$ . The channel vector $\mathbf{h}_k$ lies (approximately) within this subspace with probability close to one, and its dimension $r_k$ is typically much smaller than $N_t$ .

Example: Covariance Eigenspace for a ULA with Limited Angular Spread

Consider a ULA with $N_t = 64$ antennas and half-wavelength spacing. A user's scattering cluster subtends an angular interval $[\theta_{\min}, \theta_{\max}] = [20°, 35°]$ with uniform angular power spectrum. Compute the covariance matrix and determine the effective rank.

Solution

Covariance via integration

For a ULA with element spacing $d = \lambda/2$ , the $(m,n)$ entry of $\mathbf{R}_k$ is

$[\mathbf{R}_k]_{m,n} = \int_{\theta_{\min}}^{\theta_{\max}} p(\theta) \, e^{j2\pi (m-n) \frac{d}{\lambda} \sin\theta} \, d\theta$

where $p(\theta) = \frac{1}{\theta_{\max} - \theta_{\min}}$ for a uniform angular power spectrum.

Eigenvalue distribution

The eigenvalues of $\mathbf{R}_k$ exhibit a sharp drop-off. For $N_t = 64$ and an angular spread of $15°$ , roughly $r_k \approx \lceil N_t \cdot \frac{\sin\theta_{\max} - \sin\theta_{\min}}{2} \rceil + 1 \approx 6$ eigenvalues capture $99\%$ of the trace.

Interpretation

The channel $\mathbf{h}_k$ is effectively $6$ -dimensional despite living in $\mathbb{C}^{64}$ . The pre-beamformer in JSDM will project onto this $6$ -dimensional subspace, and CSI feedback scales with $r_k = 6$ rather than $N_t = 64$ .

Covariance Eigenvalue Spectrum vs. Angular Spread

Explore how the eigenvalue distribution of the spatial covariance matrix changes with the angular spread of the scattering cluster. Narrow angular spreads produce a low-rank covariance (few dominant eigenvalues), which is the regime where JSDM provides the largest gains.

Parameters

N_t

64

Number of ULA antennas

\theta_c

(deg)30

Center angle of arrival

\Delta\theta

(deg)15

Angular spread of scattering cluster

1 - \epsilon

0.99

Energy capture threshold for effective rank

Definition:
User Grouping by Spatial Signature

The $K$ users are partitioned into $G$ groups $\mathcal{S}_1, \ldots, \mathcal{S}_G$ such that users within the same group share approximately the same covariance eigenspace. Formally, group $g$ is characterized by a representative covariance

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

with eigendecomposition $\bar{\mathbf{R}}_g = \mathbf{U}_g \boldsymbol{\Lambda}_g \mathbf{U}_g^H$ and effective rank $r_g$ . The grouping criterion is that the chordal distance between any two users' eigenspaces within a group is small:

$d_c(\mathbf{U}_k^{(r_k)}, \mathbf{U}_g^{(r_g)}) \triangleq \frac{1}{\sqrt{2}} \left\| \mathbf{U}_k^{(r_k)} (\mathbf{U}_k^{(r_k)})^H - \mathbf{U}_g^{(r_g)} (\mathbf{U}_g^{(r_g)})^H \right\|_F$

is below a design threshold.

In practice, users arriving from similar angles of arrival cluster naturally into groups. With a ULA, the angular domain provides a natural partition: each group corresponds to a contiguous range of angles.

Angular spread

The range of angles of arrival over which a user's multipath components are distributed. A small angular spread means the channel is concentrated in a low-dimensional subspace of the array response, which is the key condition for JSDM to work effectively.

Historical Note: From Nuisance to Resource: The Evolution of Spatial Correlation in MIMO

2002–2013

Early MIMO capacity analyses (Telatar 1999, Foschini 1996) assumed i.i.d. Rayleigh fading, under which spatial correlation only reduces capacity. The Kronecker model (Weichselberger et al., 2006) and the virtual channel representation (Sayeed 2002) provided structured alternatives. But the decisive shift came with massive MIMO: when $N_t$ grows large, the eigenspace structure of $\mathbf{R}_k$ becomes sharp and stable, turning correlation from a capacity penalty into a tool for dimensionality reduction and interference management. JSDM (Adhikary et al., 2013) was the first framework to systematically exploit this structure for FDD operation.

Common Mistake: Assuming i.i.d. Channels in Massive MIMO

Mistake:

Treating the channel vectors $\mathbf{h}_k$ as i.i.d. $\mathcal{CN}(\mathbf{0}, \mathbf{I}_{N_t})$ in a massive MIMO analysis and concluding that all users have equivalent spatial signatures.

Correction:

In any realistic deployment with $N_t \geq 32$ , the channel covariance $\mathbf{R}_k$ is far from $\mathbf{I}$ . The angular resolution of a large array means different users occupy distinct, low-rank subspaces. Ignoring this structure wastes the opportunity for dimensionality reduction and leads to pessimistic CSI overhead estimates.

Quick Check

A ULA with $N_t = 128$ antennas serves a user whose scattering cluster spans an angular range of $5°$ . What is the approximate effective rank $r_k$ of the channel covariance?

$r_k \approx 128$

$r_k \approx 64$

$r_k \\approx 4$

$r_k \\approx 1$