Ferkans — Interactive Telecom Tutor

Why Physical Models Matter

The i.i.d. Rayleigh model ( $\mathbf{H}$ with independent $\mathcal{CN}(0,1)$ entries) is analytically convenient but unrealistic: real antenna arrays have finite spacing, real scattering environments have limited angular spread, and real channels exhibit spatial correlation. This section develops physically motivated channel models that capture these effects while remaining tractable for analysis.

Understanding correlation is crucial for system design: it determines whether a MIMO deployment achieves the promised capacity gains or falls far short.

Definition:
Spatial Correlation

The spatial correlation of a MIMO channel describes the statistical dependence between different entries of $\mathbf{H}$ . For a channel with zero-mean entries, the full correlation is captured by

$\mathbf{R}_{\ntn{ch}} = \mathbb{E}[\mathrm{vec}(\mathbf{H})\, \mathrm{vec}(\mathbf{H})^H] \in \mathbb{C}^{n_r n_t \times n_r n_t}$

where $\mathrm{vec}(\mathbf{H})$ stacks the columns of $\mathbf{H}$ into a single vector. When $\mathbf{R}_{\ntn{ch}} = \mathbf{I}_{n_r n_t}$ , the channel entries are uncorrelated (i.i.d. model).

The full correlation matrix has $n_r^2 n_t^2$ entries, making it unwieldy for large arrays. Structured models like Kronecker and Weichselberger reduce this to a manageable number of parameters.

Definition:
Kronecker Correlation Model

The Kronecker model assumes that transmit and receive correlations are separable:

$\mathbf{H} = \mathbf{R}_{r}^{1/2}\, \mathbf{H}_{w}\, \mathbf{R}_{t}^{1/2}$

where:

$\mathbf{H}_{w} \in \mathbb{C}^{n_r \times n_t}$ has i.i.d. $\mathcal{CN}(0,1)$ entries
$\mathbf{R}_{r} \in \mathbb{C}^{n_r \times n_r}$ is the receive correlation matrix, $[\mathbf{R}_{r}]_{ij} = \mathbb{E}[h_{ik} h_{jk}^*]$
$\mathbf{R}_{t} \in \mathbb{C}^{n_t \times n_t}$ is the transmit correlation matrix, $[\mathbf{R}_{t}]_{ij} = \mathbb{E}[h_{ki} h_{kj}^*]$

The full correlation becomes $\mathbf{R}_{\ntn{ch}} = \mathbf{R}_{t}^{T} \otimes \mathbf{R}_{r}$ , which is the Kronecker product of the two marginal correlations.

The Kronecker assumption is that the scattering environment "seen" from the transmit side is independent of that seen from the receive side. This holds well for moderate array sizes and rich scattering but can fail for keyhole channels or very large arrays.

,

Definition:
Weichselberger Model

The Weichselberger model generalises the Kronecker model by allowing coupling between transmit and receive eigenmodes:

$\mathbf{H} = \mathbf{U}_r (\boldsymbol{\Omega} \odot \mathbf{G}) \mathbf{U}_t^H$

where:

$\mathbf{U}_r$ and $\mathbf{U}_t$ are the eigenvector matrices of $\mathbf{R}_{r}$ and $\mathbf{R}_{t}$ respectively
$\mathbf{G}$ has i.i.d. $\mathcal{CN}(0,1)$ entries
$\boldsymbol{\Omega} \in \mathbb{R}_{\geq 0}^{n_r \times n_t}$ is the coupling matrix with $[\boldsymbol{\Omega}]_{ij}$ specifying the average power coupling between receive eigenmode $i$ and transmit eigenmode $j$
$\odot$ denotes element-wise (Hadamard) product

The Kronecker model is the special case where $\boldsymbol{\Omega} = \boldsymbol{\lambda}_r \boldsymbol{\lambda}_t^T$ (rank-1 coupling matrix).

Definition:
Keyhole Channel

A keyhole (pinhole) channel arises when all propagation passes through a single spatial bottleneck (e.g., a tunnel, a narrow gap between buildings). The channel has the structure

$\mathbf{H} = \mathbf{a}_r \mathbf{a}_t^H$

where $\mathbf{a}_r \in \mathbb{C}^{n_r}$ and $\mathbf{a}_t \in \mathbb{C}^{n_t}$ are random vectors with possibly rich distributions.

Despite potentially rich scattering at both the transmitter and receiver, the channel has rank 1 because all energy funnels through a single spatial mode. This means $\mathrm{rank}(\mathbf{H}) = 1$ regardless of the number of antennas, and spatial multiplexing is impossible.

Keyhole channels are pathological for spatial multiplexing but still benefit from array gain ( $n_t + n_r$ in dB) and diversity gain. They demonstrate that spatial multiplexing requires both multiple antennas AND a rich scattering environment.

Theorem: Capacity Loss from Spatial Correlation (Kronecker Model)

Under the Kronecker model, the ergodic capacity satisfies

$C_{\mathrm{corr}} \leq C_{\mathrm{iid}}$

with equality if and only if $\mathbf{R}_{t} = \mathbf{I}_{n_t}$ and $\mathbf{R}_{r} = \mathbf{I}_{n_r}$ .

More precisely, for equal power allocation at high SNR:

$C_{\mathrm{corr}} \approx C_{\mathrm{iid}} - \log_2 \det(\mathbf{R}_{t}) - \log_2 \det(\mathbf{R}_{r}) + \text{const}$

The capacity loss is governed by the "effective rank" of the correlation matrices: the more singular the correlation matrices, the greater the loss.

Spatial correlation reduces the effective number of independent channel coefficients. In the extreme case of full correlation ( $\mathbf{R}_{t}$ or $\mathbf{R}_{r}$ rank-1), the MIMO channel degenerates to a single spatial stream regardless of the number of antennas.

Proof

Jensen inequality argument

The ergodic capacity involves $\mathbb{E}[\log\det(\cdot)]$ . By the concavity of $\log\det$ and Jensen's inequality applied to the Kronecker structure, correlation cannot increase $\mathbb{E}[\log\det(\mathbf{I} + \frac{\text{SNR}}{n_t}\mathbf{H}\mathbf{H}^{H})]$ compared to the i.i.d. case.

Hadamard inequality

For positive definite $\mathbf{A}$ , $\det(\mathbf{A}) \leq \prod_i [\mathbf{A}]_{ii}$ with equality iff $\mathbf{A}$ is diagonal. Correlation introduces off-diagonal structure in $\mathbf{H}\mathbf{H}^{H}$ , reducing $\det(\mathbf{I} + \frac{\text{SNR}}{n_t}\mathbf{H}\mathbf{H}^{H})$ on average. $\blacksquare$

Kronecker Correlated MIMO Channel

Visualise how transmit and receive spatial correlation affect the MIMO channel eigenvalue distribution and capacity. Compare with the i.i.d. baseline at various correlation levels.

Parameters

n_t

4

Number of transmit antennas

n_r

4

Number of receive antennas

\\rho_t

0.5

Transmit correlation coefficient (exponential model)

\\rho_r

0.5

Receive correlation coefficient (exponential model)

SNR (dB)15

Keyhole Channel vs. Full-Rank Channel

Compare the eigenvalue spread and capacity of a keyhole (rank-1) channel with a full-rank i.i.d. Rayleigh channel. Observe how the keyhole channel collapses all spatial degrees of freedom into a single stream.

Parameters

n_t

4

n_r

4

SNR (dB)20

Example: Exponential Correlation Model

A $4 \times 4$ MIMO system uses a uniform linear array (ULA) at both transmitter and receiver. The spatial correlation follows the exponential model:

$[\mathbf{R}_{t}]_{ij} = \rho_t^{|i-j|}, \qquad [\mathbf{R}_{r}]_{ij} = \rho_r^{|i-j|}$

with $\rho_t = 0.7$ and $\rho_r = 0.3$ .

(a) Compute the eigenvalues of $\mathbf{R}_{t}$ and $\mathbf{R}_{r}$ .

(b) How many "effective" spatial modes does each side contribute?

(c) Estimate the capacity loss relative to i.i.d. at SNR = 20 dB.

Solution

Build the correlation matrices

$\mathbf{R}_{t} = \begin{bmatrix} 1 & 0.7 & 0.49 & 0.343 \\ 0.7 & 1 & 0.7 & 0.49 \\ 0.49 & 0.7 & 1 & 0.7 \\ 0.343 & 0.49 & 0.7 & 1 \end{bmatrix}KATEXPLACEHOLDER0END\mathbf{R}_{r} = \begin{bmatrix} 1 & 0.3 & 0.09 & 0.027 \\ 0.3 & 1 & 0.3 & 0.09 \\ 0.09 & 0.3 & 1 & 0.3 \\ 0.027 & 0.09 & 0.3 & 1 \end{bmatrix}$ $

Eigenvalue analysis

Numerical eigenvalues of $\mathbf{R}_{t}$ : $\lambda_1 \approx 2.67$ , $\lambda_2 \approx 0.86$ , $\lambda_3 \approx 0.33$ , $\lambda_4 \approx 0.14$ .

Eigenvalues of $\mathbf{R}_{r}$ : $\lambda_1 \approx 1.42$ , $\lambda_2 \approx 1.08$ , $\lambda_3 \approx 0.92$ , $\lambda_4 \approx 0.58$ .

The transmit correlation ( $\rho_t = 0.7$ ) has a dominant eigenvalue, reducing the effective rank to roughly 2. The receive correlation ( $\rho_r = 0.3$ ) has relatively flat eigenvalues, preserving nearly 4 effective modes.

Capacity loss estimate

The capacity loss is approximately

$\Delta C \approx -\log_2 \det(\mathbf{R}_{t}) \approx -\log_2(0.067) \approx 3.9 \;\text{bits/s/Hz}$

The receive side contributes a smaller loss since its eigenvalues are more uniform. The total loss of roughly 4--5 bits/s/Hz (relative to i.i.d. capacity of about 22 bits/s/Hz at 20 dB for $4\times 4$ ) represents a significant $\sim 20\%$ reduction. $\blacksquare$

Quick Check

In the Kronecker MIMO channel model $\mathbf{H} = \mathbf{R}_{r}^{1/2}\, \mathbf{H}_{w}\, \mathbf{R}_{t}^{1/2}$ , what happens to the channel when $\mathbf{R}_{t}$ has rank 1?

The MIMO channel degenerates to rank 1 regardless of $\mathbf{R}_{r}$ or $n_r$

The channel is unaffected because $\mathbf{R}_{r}$ compensates

The channel becomes full-rank because the noise fills in the missing modes

The capacity doubles because all power is focused on one strong mode

Correction:

The MIMO channel degenerates to rank 1 regardless of

\mathbf{R}_{r}

or

n_r

If $\mathbf{R}_{t}$ has rank 1, then $\mathbf{R}_{t}^{1/2}$ maps all columns of $\mathbf{H}_{w}$ to a single direction, making $\mathbf{H}$ rank-1. This is the beamforming-only regime: all transmit energy goes through one spatial mode.

Common Mistake: The Kronecker Model Can Underestimate or Overestimate Capacity

Mistake:

Blindly applying the Kronecker model to all MIMO scenarios, especially with large angular spreads or when the transmitter and receiver "see" overlapping clusters of scatterers.

Correction:

The Kronecker model assumes separability of transmit and receive correlation, which fails when the coupling between transmit and receive eigenmodes is not rank-1. For example, measured indoor channels at 5.2 GHz show 2--4 dB capacity overestimation by the Kronecker model. Use the Weichselberger model or full correlation $\mathbf{R}_{\ntn{ch}}$ when accuracy matters. Always validate against measurements for your deployment scenario.

Why This Matters: Spatial Correlation in Massive MIMO

In 5G NR massive MIMO (64--256 antenna elements at the base station), spatial correlation is not a bug but a feature. The large array aperture creates highly directional beams, and the channel becomes low-rank in the angular domain. This is exploited by:

Codebook-based precoding (Type I/II in 5G NR) that matches the dominant eigenmodes
FDD channel estimation that exploits low-rank structure to reduce feedback overhead from $\mathcal{O}(n_t)$ to $\mathcal{O}(r)$
Pilot contamination mitigation that uses angle-of-arrival separation between users

The Weichselberger model is particularly useful for massive MIMO because the coupling matrix $\boldsymbol{\Omega}$ directly captures the angular power spectrum.

See full treatment in Chapter 16

Comparison of MIMO Channel Models

Property	i.i.d. Rayleigh	Kronecker	Weichselberger	Keyhole
Correlation structure	None (uncorrelated)	Separable Tx/Rx	Full eigenmode coupling	Rank-1
Parameters	0	$n_t^2 + n_r^2$	$n_t^2 + n_r^2 + n_t n_r$	$n_t + n_r$
Typical rank	$\min(n_t, n_r)$	$\leq \min(n_t, n_r)$	$\leq \min(n_t, n_r)$	1
Best for	Analysis baseline	Moderate correlation	Measured channels	Tunnel/corridor
Capacity accuracy	Upper bound	Good ( $\pm 2$ dB)	Excellent	Exact for rank-1

Kronecker correlation model

A MIMO channel model that separates transmit and receive correlation: $\mathbf{H} = \mathbf{R}_{r}^{1/2} \mathbf{H}_{w} \mathbf{R}_{t}^{1/2}$ , assuming the correlation at one end is independent of the other.

Spatial correlation

The statistical dependence between channel gains at different antenna elements, arising from limited angular spread, antenna spacing, and scatterer geometry.

Weichselberger model

A generalisation of the Kronecker model that captures coupling between transmit and receive eigenmodes via a non-negative coupling matrix $\boldsymbol{\Omega}$ .

Related: Kronecker correlation model

Keyhole channel

A MIMO channel where all propagation passes through a single spatial bottleneck, resulting in rank-1 channel matrix $\mathbf{H} = \mathbf{a}_r \mathbf{a}_t^H$ regardless of the number of antennas.

Physical MIMO Channel Modeling

Why Physical Models Matter

Definition: Spatial Correlation

Definition: Kronecker Correlation Model

Definition: Weichselberger Model

Definition: Keyhole Channel

Theorem: Capacity Loss from Spatial Correlation (Kronecker Model)

Jensen inequality argument

Hadamard inequality

Kronecker Correlated MIMO Channel

Parameters

Keyhole Channel vs. Full-Rank Channel

Parameters

Example: Exponential Correlation Model

Build the correlation matrices

Eigenvalue analysis

Capacity loss estimate

Quick Check

Common Mistake: The Kronecker Model Can Underestimate or Overestimate Capacity

Why This Matters: Spatial Correlation in Massive MIMO

Comparison of MIMO Channel Models

Kronecker correlation model

Spatial correlation

Weichselberger model

Keyhole channel

Definition:
Spatial Correlation

Definition:
Kronecker Correlation Model

Definition:
Weichselberger Model

Definition:
Keyhole Channel