Ferkans — Interactive Telecom Tutor

The Inter-User Orthogonality Miracle

In conventional MIMO, users' channels are generally correlated, and this correlation limits performance. Interference between users is a fundamental bottleneck. Favorable propagation is the second key property of massive MIMO: as $N_t \to \infty$ , the channel vectors of different users become asymptotically orthogonal — their inner product normalized by $N_t$ vanishes. When this holds, intra-cell interference disappears even with the simplest linear processing, and the multi-user channel decouples into $K$ independent parallel channels.

Definition:
Favorable Propagation

The channel vectors $\{\mathbf{h}_1, \ldots, \mathbf{h}_{K}\}$ satisfy favorable propagation if for any two users $k \neq j$ :

$\frac{1}{N_t}\mathbf{h}_k^H\mathbf{h}_j \xrightarrow{N_t \to \infty} 0,$

and simultaneously:

$\frac{1}{N_t}\mathbf{h}_k^H\mathbf{h}_k \xrightarrow{N_t \to \infty} \beta_{k}.$

In matrix form: $\frac{1}{N_t}\mathbf{H}^{H}\mathbf{H} \xrightarrow{N_t \to \infty} \text{diag}(\beta_{1}, \ldots, \beta_{\ntn{nusers}})$ .

Favorable propagation combines channel hardening (diagonal terms) with inter-user orthogonality (off-diagonal terms vanishing). The two conditions together imply that the Gram matrix $\frac{1}{N_t}\mathbf{H}^{H}\mathbf{H}$ becomes diagonal — the multi-user channel is asymptotically equivalent to $K$ parallel AWGN channels.

,

Favorable Propagation

The property that channel vectors of different users become asymptotically orthogonal as the number of BS antennas $N_t \to \infty$ : $\frac{1}{N_t}\mathbf{h}_k^H \mathbf{h}_j \to 0$ for $k \neq j$ . Enables simple linear processing to achieve near-optimal sum-rate.

Theorem: Favorable Propagation for i.i.d. Rayleigh Fading

Let $\mathbf{h}_k \sim \mathcal{CN}(\mathbf{0}, \beta_{k}\mathbf{I})$ and $\mathbf{h}_j \sim \mathcal{CN}(\mathbf{0}, \beta_{j}\mathbf{I})$ be independent for $k \neq j$ . Then:

$\mathbb{E}\!\left[\frac{1}{N_t}\mathbf{h}_k^H\mathbf{h}_j\right] = 0,$ $\mathbb{E}\!\left[\left|\frac{\mathbf{h}_k^H\mathbf{h}_j}{N_t}\right|^2\right] = \frac{\beta_{k}\,\beta_{j}}{N_t} \to 0.$

Therefore $\frac{1}{N_t}\mathbf{h}_k^H\mathbf{h}_j \xrightarrow{P} 0$ , and i.i.d. Rayleigh fading satisfies favorable propagation.

The inner product $\mathbf{h}_k^H\mathbf{h}_j = \sum_{i=1}^{N_t} h_{k,i}^*h_{j,i}$ is a sum of $N_t$ independent zero-mean random variables (since $\mathbf{h}_k$ and $\mathbf{h}_j$ are independent). The sum divided by $N_t$ converges to zero by the law of large numbers.

Show Hint

Compute $\mathbb{E}[h_{k,i}^* h_{j,i}]$ using independence and zero mean.

Compute $\mathbb{E}[|h_{k,i}^* h_{j,i}|^2]$ using $|ab|^2 = |a|^2|b|^2$ and independence.

Apply Markov or Chebyshev inequality to bound the probability that $|\mathbf{h}_k^H\mathbf{h}_j|/N_t$ exceeds $\epsilon$ .

Proof

Zero mean

Since $\mathbf{h}_k$ and $\mathbf{h}_j$ are independent, $\mathbb{E}[h_{k,i}^* h_{j,i}] = \mathbb{E}[h_{k,i}^*]\,\mathbb{E}[h_{j,i}] = 0$ (both have zero mean). By linearity, $\mathbb{E}[\mathbf{h}_k^H\mathbf{h}_j] = 0$ .

Second moment

$\mathbb{E}\!\left[|h_{k,i}^* h_{j,i}|^2\right] = \mathbb{E}[|h_{k,i}|^2]\,\mathbb{E}[|h_{j,i}|^2] = \beta_{k}\,\beta_{j}.$ $The entries of$ \mathbf{h}_k^H\mathbf{h}_j $are uncorrelated (different$ i $values), so$ \mathbb{E}[|\mathbf{h}k^H\mathbf{h}j|^2] = N_t,\beta{k},\beta{j} $. Dividing by$ N_t^{2} $:$ \mathbb{E}[|\mathbf{h}k^H\mathbf{h}j / N_t|^2] = \beta{k}\beta{j}/N_t \to 0$.

Convergence in probability

Apply Markov's inequality to $|\mathbf{h}_k^H\mathbf{h}_j/N_t|^2$ : $\Pr\!\left[\left|\frac{\mathbf{h}_k^H\mathbf{h}_j}{N_t}\right| > \epsilon\right] = \Pr\!\left[\left|\frac{\mathbf{h}_k^H\mathbf{h}_j}{N_t}\right|^2 > \epsilon^2\right] \leq \frac{\beta_{k}\beta_{j}}{N_t\,\epsilon^2} \to 0. \quad\blacksquare$

,

Favorable Propagation Under Spatial Correlation

For spatially correlated channels $\mathbf{h}_k \sim \mathcal{CN}(\mathbf{0}, \mathbf{R}_k)$ , the asymptotic cross-correlation is

$\frac{1}{N_t}\mathbb{E}[\mathbf{h}_k^H\mathbf{h}_j] = 0$

(still zero, since $\mathbf{h}_k$ and $\mathbf{h}_j$ are independent), but the normalized variance of the cross-correlation is now

$\mathbb{E}\!\left[\left|\frac{\mathbf{h}_k^H\mathbf{h}_j}{N_t}\right|^2\right] = \frac{\text{tr}(\mathbf{R}_k\mathbf{R}_j)}{N_t^{2}}.$

This vanishes (favorable propagation holds) if and only if $\text{tr}(\mathbf{R}_k\mathbf{R}_j) / N_t^{2} \to 0$ , which requires $\mathbf{R}_k$ and $\mathbf{R}_j$ to have asymptotically orthogonal eigenspaces. Users served by the same BS must occupy different angular sectors. If two users have nearly identical spatial covariance matrices, their channels remain correlated even with unlimited antennas — a key motivation for intelligent user scheduling (Chapter 5).

Example: Favorable Propagation: Numerical Check for Two Users

Two users have i.i.d. Rayleigh fading channels with $\beta_{1} = \beta_{2} = 1$ . For $N_t = \{4, 16, 64, 256\}$ , compute the expected value of $|\mathbf{h}_1^H\mathbf{h}_2|^2 / N_t^{2}$ and verify that it decreases as $1/N_t$ .

Solution

Apply Theorem <a href="#thm-iid-favorable-propagation" class="ferkans-ref" title="Theorem: Favorable Propagation for i.i.d. Rayleigh Fading" data-ref-type="theorem"><span class="ferkans-ref-badge">T</span>Favorable Propagation for i.i.d. Rayleigh Fading</a>

From the theorem, $\mathbb{E}[|\mathbf{h}_1^H\mathbf{h}_2|^2/N_t^{2}] = \beta_{1}\beta_{2}/N_t = 1/N_t$ .

Tabulate values

$N_t$	$\mathbb{E}[\|\mathbf{h}_1^H\mathbf{h}_2\|^2/N_t^{2}]$
4	0.250
16	0.0625
64	0.0156
256	0.0039

The residual interference power decreases as $1/N_t$ : doubling the antennas halves the inter-user interference-to-noise ratio. With $N_t = 256$ , the normalized cross-correlation is about 0.06 in standard deviation — already negligible for most practical purposes. $\blacksquare$

Favorable Propagation: $|\mathbf{h}_k^H\mathbf{h}_j|^2 / N_t^2$ vs $N_t$

Visualize the decay of the normalized inter-user interference $|\mathbf{h}_k^H\mathbf{h}_j|^2/N_t^{2}$ as $N_t$ increases. Use the correlation slider to explore how spatial correlation slows the decay.

Parameters

Spatial correlation

\rho

0

Show

1/N_t

theoretical line

Common Mistake: Favorable Propagation Is Not the Same as Orthogonal Channels

Mistake:

Conflating favorable propagation (asymptotic) with orthogonality: assuming that in a system with $N_t = 64$ and $K = 4$ , the channel vectors are exactly orthogonal.

Correction:

Favorable propagation is an asymptotic property; at finite $N_t$ , the off-diagonal entries of $\frac{1}{N_t}\mathbf{H}^{H}\mathbf{H}$ are small but nonzero. The residual inter-user interference decreases as $1/N_t$ (from Theorem TFavorable Propagation for i.i.d. Rayleigh Fading). For realistic system analysis at finite $N_t$ , this residual must be accounted for — see Chapter 4 on achievable rate analysis.

Channel Hardening vs. Favorable Propagation

Property	Channel Hardening	Favorable Propagation
What converges?	$\\|\mathbf{h}_k\\|^2/N_t \to \beta_k$ (diagonal)	$\mathbf{h}_k^H\mathbf{h}_j/N_t \to 0$ (off-diagonal)
Mathematical mechanism	LLN applied to $\sum \|h_{k,i}\|^2$	LLN applied to $\sum h_{k,i}^* h_{j,i}$ (zero mean)
Rate of convergence (i.i.d.)	$\text{Var} = \beta_k^2 / N_t$	$\mathbb{E}[\|\cdot\|^2] = \beta_k\beta_j/N_t$
Fails for	Rank-1 covariance (pure LoS)	Users with identical angular support
Implication for design	Open-loop scheduling, stable QoS	MRC achieves near-optimal SINR

Quick Check

Under what condition does favorable propagation fail even as $N_t \to \infty$ ?

When the SNR is too low

When the two users have covariance matrices with overlapping eigenspaces

When the number of users exceeds the number of antennas

When the path-loss coefficients are different

Correction:

When the two users have covariance matrices with overlapping eigenspaces

From the correlated case analysis, $\mathbb{E}[|\mathbf{h}_k^H\mathbf{h}_j|^2/N_t^2] = \text{tr}(\mathbf{R}_k\mathbf{R}_j)/N_t^2$ . This vanishes only if $\text{tr}(\mathbf{R}_k\mathbf{R}_j)/N_t^2 \to 0$ , i.e., the two covariance matrices have asymptotically orthogonal eigenspaces. If users share the same angular support, $\text{tr}(\mathbf{R}_k\mathbf{R}_j)$ grows with $N_t$ and favorable propagation fails.

Favorable Propagation

The Inter-User Orthogonality Miracle

Definition: Favorable Propagation

Favorable Propagation

Theorem: Favorable Propagation for i.i.d. Rayleigh Fading

Zero mean

Second moment

Convergence in probability

Favorable Propagation Under Spatial Correlation

Example: Favorable Propagation: Numerical Check for Two Users

Apply Theorem <a href="#thm-iid-favorable-propagation" class="ferkans-ref" title="Theorem: Favorable Propagation for i.i.d. Rayleigh Fading" data-ref-type="theorem"><span class="ferkans-ref-badge">T</span>Favorable Propagation for i.i.d. Rayleigh Fading</a>

Tabulate values

Favorable Propagation: ∣hkHhj∣2/Nt2|\mathbf{h}_k^H\mathbf{h}_j|^2 / N_t^2∣hkH​hj​∣2/Nt2​ vs NtN_tNt​

Parameters

Common Mistake: Favorable Propagation Is Not the Same as Orthogonal Channels

Channel Hardening vs. Favorable Propagation

Quick Check

Definition:
Favorable Propagation

Favorable Propagation: $|\mathbf{h}_k^H\mathbf{h}_j|^2 / N_t^2$ vs $N_t$