Ferkans — Interactive Telecom Tutor

Why Massive MIMO Makes Things Simple

In conventional multi-user MIMO, inter-user interference requires sophisticated nonlinear processing (dirty paper coding, successive interference cancellation) to approach capacity. Massive MIMO renders these techniques unnecessary. When $M \gg K$ , the user channel vectors become nearly orthogonal — a property called favourable propagation. Under this condition, the simplest linear receiver (matched filter / MR combining) nearly eliminates inter-user interference, and the performance gap between MR and the optimal joint decoder vanishes. This dramatic simplification is the central engineering appeal of massive MIMO.

Definition:
Favorable Propagation

The channel matrix $\mathbf{H} = [\mathbf{h}_1, \ldots, \mathbf{h}_K]$ satisfies the favourable propagation condition if:

$\frac{\mathbf{h}_i^H \mathbf{h}_j}{\|\mathbf{h}_i\|\,\|\mathbf{h}_j\|} \xrightarrow{M \to \infty} 0 \quad \forall\, i \neq j$

Equivalently, in matrix form:

$\frac{1}{M}\mathbf{H}^{H}\mathbf{H} \xrightarrow{M \to \infty} \mathbf{D}_\beta = \text{diag}(\beta_1, \ldots, \beta_K)$

When favourable propagation holds, the Gram matrix $\mathbf{H}^{H}\mathbf{H}$ becomes asymptotically diagonal, meaning user channels are asymptotically mutually orthogonal. The off-diagonal elements (inter-user interference) vanish relative to the diagonal elements (desired signal power).

,

Theorem: Asymptotic Orthogonality under i.i.d. Rayleigh Fading

Let $\mathbf{h}_k = \sqrt{\beta_k}\,\mathbf{g}_k$ with $\mathbf{g}_k \sim \mathcal{CN}(\mathbf{0}, \mathbf{I}_M)$ independently for $k = 1, \ldots, K$ . Then for $i \neq j$ :

$\frac{\mathbf{h}_i^H \mathbf{h}_j}{\|\mathbf{h}_i\|\,\|\mathbf{h}_j\|} \xrightarrow{\text{a.s.}} 0 \quad \text{as } M \to \infty$

and the convergence rate is:

$\mathbb{E}\!\left[\left|\frac{\mathbf{h}_i^H \mathbf{h}_j}{M}\right|^2\right] = \frac{\beta_i \beta_j}{M}$

The inner product $\mathbf{h}_i^H\mathbf{h}_j = \sqrt{\beta_i\beta_j}\sum_{m=1}^{M}g_{mi}^* g_{mj}$ is a sum of $M$ i.i.d. zero-mean random variables. By the law of large numbers, $\mathbf{h}_i^H\mathbf{h}_j/M \to 0$ while $\|\mathbf{h}_i\|^2/M \to \beta_i > 0$ . Thus the normalised inner product (the cosine of the angle between channels) vanishes: the channel vectors become orthogonal in high-dimensional space. This is a manifestation of the concentration of measure phenomenon in high dimensions.

Proof

Computing the mean

For $i \neq j$ , the independence of $\mathbf{g}_i$ and $\mathbf{g}_j$ gives:

$\mathbb{E}\!\left[\frac{\mathbf{h}_i^H\mathbf{h}_j}{M}\right] = \frac{\sqrt{\beta_i\beta_j}}{M}\sum_{m=1}^{M} \mathbb{E}[g_{mi}^* g_{mj}] = 0$

since $g_{mi}^*$ and $g_{mj}$ are independent zero-mean.

Computing the second moment

$\mathbb{E}\!\left[\left|\frac{\mathbf{h}_i^H\mathbf{h}_j}{M}\right|^2\right] = \frac{\beta_i\beta_j}{M^2}\sum_{m=1}^{M}\sum_{n=1}^{M} \mathbb{E}[g_{mi}^* g_{mj} g_{ni} g_{nj}^*]KATEXPLACEHOLDER0END\mathbb{E}[|g_{mi}|^2|g_{mj}|^2] = 1KATEXPLACEHOLDER1END\mathbb{E}\!\left[\left|\frac{\mathbf{h}_i^H\mathbf{h}_j}{M}\right|^2\right] = \frac{\beta_i\beta_j}{M^2} \cdot M = \frac{\beta_i\beta_j}{M} \to 0$ $

Almost sure convergence

Define $Z_M = \mathbf{h}_i^H\mathbf{h}_j/M$ . We showed $\mathbb{E}[Z_M] = 0$ and $\mathbb{E}[|Z_M|^2] = \beta_i\beta_j/M$ .

By the strong law of large numbers applied to $Z_M = \sqrt{\beta_i\beta_j}\cdot\frac{1}{M}\sum_{m=1}^{M}g_{mi}^* g_{mj}$ (a mean of i.i.d. zero-mean, unit-variance terms):

$Z_M \xrightarrow{\text{a.s.}} 0$

Combined with $\|\mathbf{h}_i\|/\sqrt{M} \xrightarrow{\text{a.s.}} \sqrt{\beta_i}$ (Theorem 18.1), the normalised inner product converges to zero almost surely. $\blacksquare$

,

Favorable Propagation Visualisation

Explore how the inner product between user channels $|\mathbf{h}_i^H\mathbf{h}_j|/(M\sqrt{\beta_i\beta_j})$ concentrates around zero as $M$ increases. Increase $K$ to see how the number of users affects the degree of inter-user interference.

Parameters

K

(number of users)5

Example: Verifying Favorable Propagation Numerically

Consider a massive MIMO system with $M = 64$ antennas and $K = 4$ users with equal large-scale fading $\beta_k = 1$ . Under i.i.d. Rayleigh fading, compute: (a) the expected squared magnitude of the normalised inter-user interference $\mathbb{E}[|\mathbf{h}_i^H\mathbf{h}_j/M|^2]$ , (b) the expected signal-to-interference ratio from favourable propagation alone (no noise).

Solution

Inter-user interference power

From Theorem 18.2, for $\beta_i = \beta_j = 1$ :

$\mathbb{E}\!\left[\left|\frac{\mathbf{h}_i^H\mathbf{h}_j}{M}\right|^2\right] = \frac{1}{M} = \frac{1}{64} = 0.0156$

The normalised interference is 18 dB below the normalised signal power ( $\beta_k = 1$ ).

Signal-to-interference ratio

With MR combining $\hat{x}_k = \mathbf{h}_k^H\mathbf{y}/M$ , the desired signal power scales as $|\mathbf{h}_k^H\mathbf{h}_k/M|^2 \approx \beta_k^2 = 1$ , while each interferer contributes $|\mathbf{h}_k^H\mathbf{h}_j/M|^2 \approx 1/M$ .

The signal-to-interference ratio (ignoring noise) is:

$\text{SIR} \approx \frac{1}{(K-1)/M} = \frac{M}{K-1} = \frac{64}{3} \approx 21.3 = 13.3\;\text{dB}$

This $M/(K-1)$ scaling demonstrates that massive MIMO naturally suppresses inter-user interference through favourable propagation alone, even with the simplest MR combining. $\blacksquare$

Quick Check

In a massive MIMO system with $M = 128$ antennas and $K = 8$ users under i.i.d. Rayleigh fading, what is the approximate SIR (in dB) achieved by matched-filter combining due to favourable propagation alone?

6 dB

12.6 dB

16 dB

21 dB

Correction:

12.6 dB

$\text{SIR} \approx M/(K-1) = 128/7 \approx 18.3$ , which is $10\log_{10}(18.3) \approx 12.6$ dB. The matched filter alone provides substantial interference suppression.

Common Mistake: Favorable Propagation Breaks with Correlated Channels

Mistake:

Assuming that favourable propagation always holds in massive MIMO, regardless of the propagation environment.

Correction:

Favourable propagation relies on the channel vectors being statistically independent across users, which holds for i.i.d. Rayleigh fading but can fail in practice. Specifically:

Spatially correlated channels: When the BS array has limited angular spread or users share similar angles of arrival, $\mathbf{R}_{i} \approx \mathbf{R}_{j}$ , and the channels do not become orthogonal even as $M \to \infty$ .
Line-of-sight (LoS): If two users have similar angles, $\mathbf{a}(\theta_i)^H\mathbf{a}(\theta_j)$ does not vanish.
Keyhole channels: Rank-deficient channels prevent asymptotic orthogonality.

In correlated scenarios, the off-diagonal terms of $\frac{1}{M}\mathbf{H}^{H}\mathbf{H}$ converge to nonzero values $\frac{1}{M}\text{tr}(\mathbf{R}_{i}\mathbf{R}_{j}) \neq 0$ , degrading the SIR. This motivates spatial filtering (ZF, MMSE) even in massive MIMO.

🎓CommIT Contribution(2018)

Massive MIMO Has Unlimited Capacity

G. Caire — IEEE Transactions on Wireless Communications

Caire showed that when the user channel covariance matrices $\mathbf{R}_{k}$ have different eigenspaces (as they do in spatially correlated channels with distinct angular spreads), the pilot contamination bottleneck can be overcome. By exploiting the low-rank structure of $\mathbf{R}_{k}$ in channel estimation, the contaminating channels from other-cell users can be projected out, and the rate grows without bound as $M \to \infty$ . This overturned the conventional belief that pilot contamination imposes a fundamental capacity ceiling, showing that the ceiling is an artifact of the i.i.d. Rayleigh model where all covariance matrices are identical ( $\mathbf{R}_{k} = \beta_k\mathbf{I}$ ).

massive-mimospatial-correlationpilot-contaminationCommIT

🎓CommIT Contribution(2013)

Joint Spatial Division and Multiplexing (JSDM)

A. Adhikary, J. Nam, J.-Y. Ahn, G. Caire — IEEE Transactions on Information Theory

JSDM addresses the CSI acquisition bottleneck in FDD massive MIMO (where channel reciprocity is unavailable). The key idea is to group users whose channel covariance matrices share similar eigenspaces and apply a two-stage precoding:

Pre-beamforming (based on long-term statistics): project onto the dominant eigenspace of each user group using the covariance eigenvectors.
Multi-user MIMO precoding (based on reduced-dimension effective channels): standard ZF/MMSE within each group.

This reduces the CSI feedback dimension from $M$ (full channel) to $r_k$ (effective rank of the user group's covariance), making FDD massive MIMO practical. JSDM achieves rates close to the perfect-CSI benchmark while requiring orders of magnitude less feedback than naive approaches.

massive-mimofddjsdmspatial-correlationCommITView Paper →

Favorable Propagation

The property that user channel vectors become asymptotically mutually orthogonal as $M \to \infty$ : $\mathbf{h}_i^H\mathbf{h}_j/(\|\mathbf{h}_i\|\|\mathbf{h}_j\|) \to 0$ for $i \neq j$ . This enables simple linear combining to effectively suppress inter-user interference.

Gram Matrix

The matrix $\mathbf{G} = \mathbf{H}^{H}\mathbf{H} \in \mathbb{C}^{K \times K}$ whose $(i,j)$ entry is $\mathbf{h}_i^H\mathbf{h}_j$ . In massive MIMO, $\frac{1}{M}\mathbf{G} \to \mathbf{D}_\beta$ under favourable propagation.

Favorable Propagation and Asymptotic Orthogonality