Ferkans — Interactive Telecom Tutor

From Optimal to Simple: the Massive Regime Miracle

The optimal detector for the multi-user uplink is the successive interference cancellation (SIC) receiver, which achieves the MAC capacity region (ITA Ch. 16). SIC is exponentially complex in the number of users. In contrast, linear receivers — which simply multiply the received signal by a fixed matrix and then decode independently — are computationally trivial. In general, linear receivers sacrifice performance.

The central result of this section shows that in the massive MIMO regime, the sacrifice is negligible: MRC and ZF achieve the same sum rate scaling as the optimal nonlinear detector. This is one of the most important operational simplifications enabled by favorable propagation.

Definition:
Linear Uplink Receivers: MRC and ZF

Given the received signal $\mathbf{y} = \mathbf{H}\mathbf{s} + \mathbf{w}$ , a linear receiver computes the sufficient statistic $\hat{\mathbf{s}} = \mathbf{W}^{H}\mathbf{y} = \mathbf{W}^{H}\mathbf{H}\mathbf{s} + \mathbf{W}^{H}\mathbf{w},$ where $\mathbf{W} \in \mathbb{C}^{N_t \times K}$ is the combining matrix. Decoding is performed independently per user.

Two canonical choices:

MRC (Maximum Ratio Combining): $\mathbf{W}_{\text{MRC}} = \mathbf{H}$ . Each column is the matched filter to the corresponding user's channel. No matrix inversion required.
ZF (Zero Forcing): $\mathbf{W}_{\text{ZF}} = \mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}$ . Inverts the effective channel to force all inter-user interference to zero. Requires $N_t \geq K$ and a matrix pseudo-inverse.

The per-user post-combining SINR for user $k$ with combining vector $\mathbf{v}_{k}$ (the $k$ -th column of $\mathbf{W}$ ) is $\text{SINR}_k = \frac{P_k|\mathbf{v}_{k}^{H}\mathbf{h}_k|^2} {\sum_{j \neq k} P_j|\mathbf{v}_{k}^{H}\mathbf{h}_j|^2 + \sigma^2\|\mathbf{v}_{k}\|^2}.$

,

MRC (Maximum Ratio Combining)

A linear receiver that uses $\mathbf{v}_{k} = \mathbf{h}_k$ as the combining vector for user $k$ . Maximizes the SNR for each user in isolation by coherently combining all received signals weighted by the channel coefficients. Does not cancel inter-user interference.

ZF Combining (Zero Forcing)

A linear receiver using $\mathbf{W}_{\text{ZF}} = \mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}$ that projects each user's received signal onto the null space of all other users' channels. Eliminates inter-user interference but enhances noise.

Theorem: MRC Asymptotic Near-Optimality

For i.i.d. Rayleigh fading with equal power $P_k = P$ and channel $\mathbf{H} \sim \mathcal{CN}(\mathbf{0}, \mathbf{B})$ (where $\mathbf{B} = \text{diag}(\beta_{1},\ldots,\beta_{\ntn{nusers}})$ is the path-loss matrix), the MRC per-user SINR satisfies

$\text{SINR}_k^{\text{MRC}} = \frac{P\|\mathbf{h}_k\|^4} {\sum_{j \neq k} P|\mathbf{h}_k^H\mathbf{h}_j|^2 + \sigma^2\|\mathbf{h}_k\|^2} \xrightarrow{a.s.} \frac{P\,N_t\,\beta_{k}^{2}} {\sum_{j \neq k}P\,\beta_{k}\beta_{j} + \sigma^2\,\beta_{k}}$

as $N_t \to \infty$ . The asymptotic MRC sum rate satisfies

$C_{\text{sum}}^{\text{MRC}} \to \sum_{k=1}^{K} \log_2\!\left(1 + \frac{N_t\,P\,\beta_{k}}{\sum_{j\neq k}P\,\beta_{j} + \sigma^2}\right),$

which grows as $K\log_2(N_t) + \mathcal{O}(1)$ — the same asymptotic scaling as the optimal SIC receiver.

MRC works because favorable propagation simultaneously provides two things: (1) the desired signal power grows as $N_t^{2}\beta_{k}^{2}$ (from $\|\mathbf{h}_k\|^4$ ), while (2) the interference power grows as only $N_t\beta_{k}\beta_{j}$ (from $|\mathbf{h}_k^H\mathbf{h}_j|^2$ ). The signal grows faster than the interference, so SINR $\to \infty$ as $N_t\to\infty$ .

Show Hint

Substitute $\mathbf{W} = \mathbf{H}$ into the SINR expression. The numerator becomes $|\mathbf{h}_k^H\mathbf{h}_k|^2 = \|\mathbf{h}_k\|^4$ .

Apply the almost-sure convergence results from Sections 1.2 and 1.3: $\|\mathbf{h}_k\|^2/N_t \to \beta_{k}$ and $|\mathbf{h}_k^H\mathbf{h}_j|^2/N_t^{2} \to 0$ .

Divide numerator and denominator by $N_t^{2}$ and take the limit.

Proof

MRC SINR expression

With $\mathbf{v}_{k} = \mathbf{h}_k$ : $\text{SINR}_k^{\text{MRC}} = \frac{P\|\mathbf{h}_k\|^4} {\sum_{j \neq k}P|\mathbf{h}_k^H\mathbf{h}_j|^2 + \sigma^2\|\mathbf{h}_k\|^2}.$

Normalize by $N_t^2$

Dividing numerator and denominator by $N_t^{2}$ : $\text{SINR}_k^{\text{MRC}} = \frac{P\left(\|\mathbf{h}_k\|^2/N_t\right)^2} {\sum_{j \neq k}P\left|\mathbf{h}_k^H\mathbf{h}_j/N_t\right|^2 + (\sigma^2/N_t)\left(\|\mathbf{h}_k\|^2/N_t\right)}.$

Take the almost-sure limit

By favorable propagation (Section 1.3):

$\|\mathbf{h}_k\|^2/N_t \xrightarrow{a.s.} \beta_{k}$
$|\mathbf{h}_k^H\mathbf{h}_j|^2/N_t^{2} \xrightarrow{a.s.} 0$ for $j \neq k$
$\sigma^2/N_t \to 0$

The denominator's interference terms vanish, leaving: $\text{SINR}_k^{\text{MRC}} \xrightarrow{a.s.} \frac{P\,\beta_{k}^{2}}{0 + 0} \to \infty.$ But at finite $N_t$ , keeping the $\sigma^2/N_t$ term gives the finite- $N_t$ approximation stated. $\blacksquare$

,

Theorem: ZF Combining Eliminates Interference Exactly

With $N_t \geq K$ and $\text{rank}(\mathbf{H}) = K$ almost surely, the ZF combining matrix $\mathbf{W}_{\text{ZF}} = \mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}$ satisfies: $\mathbf{W}_{\text{ZF}}^H\mathbf{H} = \mathbf{I}_{K},$ eliminating all inter-user interference exactly (not asymptotically). The resulting per-user SINR is: $\text{SINR}_k^{\text{ZF}} = \frac{P_k} {\sigma^2\left[(\mathbf{H}^{H}\mathbf{H})^{-1}\right]_{kk}}.$

ZF projects the received signal onto a direction orthogonal to all other users' channels. This eliminates interference exactly but the projection also removes some of the desired signal component and amplifies noise (the noise enhancement factor is $[(\mathbf{H}^{H}\mathbf{H})^{-1}]_{kk} \geq 1/\|\mathbf{h}_k\|^2$ ). In the massive regime, $[(\mathbf{H}^{H}\mathbf{H})^{-1}]_{kk} \to 1/(N_t\beta_{k})$ , so ZF approaches MRC performance.

Show Hint

Compute $\mathbf{W}_{\text{ZF}}^H\mathbf{H}$ directly using the definition.

The noise term in the output is $\mathbf{W}_{\text{ZF}}^H\mathbf{w}$ ; compute its covariance.

The $k$ -th diagonal of $(\mathbf{H}^{H}\mathbf{H})^{-1}$ bounds the noise power for user $k$ .

Proof

Interference cancellation

$\mathbf{W}_{\text{ZF}}^H\mathbf{H} = (\mathbf{H}^{H}\mathbf{H})^{-H}\mathbf{H}^{H}\mathbf{H} = (\mathbf{H}^{H}\mathbf{H})^{-1}(\mathbf{H}^{H}\mathbf{H}) = \mathbf{I}_{K}.$ So the post-combining output is $\hat{\mathbf{s}} = \mathbf{s} + (\mathbf{H}^{H}\mathbf{H})^{-1}\mathbf{H}^{H}\mathbf{w}$ — zero inter-user interference.

Noise covariance

The noise component is $\tilde{\mathbf{n}} = (\mathbf{H}^{H}\mathbf{H})^{-1}\mathbf{H}^{H}\mathbf{w}$ . Its covariance is $\mathbb{E}[\tilde{\mathbf{n}}\tilde{\mathbf{n}}^H | \mathbf{H}] = \sigma^2(\mathbf{H}^{H}\mathbf{H})^{-1}$ . User $k$ sees noise power $\sigma^2[(\mathbf{H}^{H}\mathbf{H})^{-1}]_{kk}$ , yielding the stated SINR. $\blacksquare$

Key Takeaway

MRC uses $N_t^{2}$ signal growth; ZF kills interference exactly. MRC has no interference cancellation but its signal grows as $N_t^{2} \beta_{k}^{2}$ while interference grows as only $N_t\beta_{k}\beta_{j}$ — the SINR diverges. ZF kills interference at finite $N_t$ but pays a noise enhancement penalty. For $N_t \gg K$ with i.i.d. channels, MRC and ZF achieve the same asymptotic sum rate; ZF outperforms MRC at small $N_t$ or under high inter-user interference (correlated channels).

Uplink Linear Combining: MRC and ZF

Complexity: MRC:

\mathcal{O}(N_t K)

. ZF:

\mathcal{O}(N_t K + K^3)

(Gram matrix inversion dominates at large

N_t

).

Input: Received signal

\mathbf{y} \in \mathbb{C}^{N_t}

, estimated channel

\hat{\mathbf{H}} \in \mathbb{C}^{N_t \times K}

Output: Per-user symbol estimates

\hat{\mathbf{s}} \in \mathbb{C}^K

MRC Combining:

1.

\mathbf{V}_{\text{MRC}} \leftarrow \hat{\mathbf{H}}

2.

\hat{\mathbf{s}} \leftarrow \mathbf{V}_{\text{MRC}}^H \mathbf{y}

3. Normalize:

\hat{s}_k \leftarrow \hat{s}_k / \|\hat{\mathbf{h}}_k\|^2

for each

k

4. return

\hat{\mathbf{s}}

ZF Combining:

1. Compute Gram matrix:

\mathbf{G} \leftarrow \hat{\mathbf{H}}^H \hat{\mathbf{H}}

(

K \times K

)

2.

\mathbf{V}_{\text{ZF}} \leftarrow \hat{\mathbf{H}} \mathbf{G}^{-1}

3.

\hat{\mathbf{s}} \leftarrow \mathbf{V}_{\text{ZF}}^H \mathbf{y}

4. return

\hat{\mathbf{s}}

The ZF inversion is a $K \times K$ (not $N_t \times N_t$ ) matrix inversion — far cheaper than inverting the full channel matrix. For $K \ll N_t$ , both combiners have similar computational cost.

MRC vs. ZF vs. MMSE Uplink Combining

Property	MRC	ZF	MMSE
Interference cancellation	None (relies on favorable propagation)	Exact (requires $N_t \geq K$ )	Partial (optimal linear tradeoff)
Noise behavior	Low noise enhancement ( $\\|\mathbf{h}_k\\|^2$ )	Noise enhancement: $[(\mathbf{H}^H\mathbf{H})^{-1}]_{kk}$	Minimum MSE: $[(\mathbf{H}^H\mathbf{H} + \sigma^2\mathbf{I})^{-1}]_{kk}$
Complexity (per coherence interval)	$\mathcal{O}(N_t K)$	$\mathcal{O}(N_t K + K^3)$	$\mathcal{O}(N_t K + K^3)$
Optimal for	Large $N_t$ , i.i.d. channels	Moderate $N_t$ , correlated channels, low SNR	All regimes (approaches ZF at high SNR)
Achieves MAC capacity?	Asymptotically ( $N_t \to \infty$ )	Asymptotically ( $N_t \to \infty$ )	Asymptotically ( $N_t \to \infty$ )

MRC vs ZF Sum Rate vs. $N_t$

Compare per-user achievable rate for MRC and ZF combining as a function of $N_t$ at fixed $K$ and SNR. Observe the crossover where ZF overtakes MRC and how both converge to the same asymptote.

Parameters

Users

K

4

SNR (dB)10

Show optimal (MAC capacity)

Common Mistake: ZF Can Fail When Channels Are Nearly Parallel

Mistake:

Using ZF combining in scenarios where two users have highly correlated channel vectors (e.g., same angular direction), and assuming ZF interference cancellation always works.

Correction:

ZF requires inverting the Gram matrix $\mathbf{H}^{H}\mathbf{H}$ . When two users have nearly parallel channels, the Gram matrix is nearly rank-deficient, its inverse has very large entries, and noise is massively amplified. The ZF SINR can be far worse than MRC in this case. Regularized ZF (MMSE combining, also called MMSE precoding) avoids this by adding $\sigma^2\mathbf{I}$ to the Gram matrix before inversion: $\mathbf{W}_{\text{MMSE}} = \mathbf{H}(\mathbf{H}^{H}\mathbf{H} + \sigma^2\mathbf{I})^{-1}$ . This is the optimal linear receiver and is treated in detail in Chapter 4.

🔧Engineering Note

Computational Complexity at Scale: $N_t = 256$ , $K = 16$

At $N_t = 256$ antennas and $K = 16$ users:

MRC requires a $256 \times 16$ multiply: $\approx 8000$ multiply-accumulates per sample. At 30.72 MHz symbol rate, this is $2.5 \times 10^{11}$ FLOPS/s. Achievable with current DSP/FPGA technology.
ZF additionally inverts a $16 \times 16$ matrix ( $\approx 4000$ FLOPS) — negligible compared to the initial multiply. Total ZF complexity is $\approx 1\%$ more than MRC.
SIC (optimal, non-linear) would require $2^{16}$ candidate evaluations per symbol vector — computationally intractable in real time.

This confirms the practical advantage: linear receivers require hardware commensurate with the antenna count, while near-optimal performance requires only the same scaling.

Practical Constraints

•
5G NR BS processors must achieve <0.5 ms processing latency (3GPP TS 38.133)
•
Commercial 64-antenna AAUs (e.g., Ericsson AIR 6488) use custom ASICs for 256-QAM processing
•
Real-time MMSE inversion for K=32 users is the practical limit with 2024-era hardware

📋 Ref: 3GPP TS 38.133, Section 8.1

Example: MRC vs ZF SINR: $3 \times 2$ System

Consider $N_t = 3$ BS antennas and $K = 2$ users with channel matrix $\mathbf{H} = \begin{pmatrix} 1+j & 1-j \\ 2 & -j \\ j & 1 \end{pmatrix}.$ Both users transmit with $P = 1$ and noise variance $\sigma^2 = 1$ .

(a) Compute the MRC SINR for user 1. (b) Compute the ZF SINR for user 1. (c) Which performs better and why?

Solution

(a) MRC SINR for user 1

$\mathbf{h}_1 = (1+j, 2, j)^T$ . $\|\mathbf{h}_1\|^2 = |1+j|^2 + |2|^2 + |j|^2 = 2 + 4 + 1 = 7$ . Signal power: $|\mathbf{h}_1^H\mathbf{h}_1|^2 = \|\mathbf{h}_1\|^4 = 49$ . $\mathbf{h}_2 = (1-j, -j, 1)^T$ . Cross-term: $\mathbf{h}_1^H\mathbf{h}_2 = (1-j)(1-j) + 2(-j) + (-j)(1) = (1 - 2j - 1) + (-2j) + (-j) = -5j$ . So $|\mathbf{h}_1^H\mathbf{h}_2|^2 = 25$ . $\text{SINR}_1^{\text{MRC}} = \frac{P \cdot 49}{P \cdot 25 + \sigma^2 \cdot 7} = \frac{49}{25 + 7} = \frac{49}{32} \approx 1.53$ .

(b) ZF SINR for user 1

Gram matrix: $\mathbf{G} = \mathbf{H}^H\mathbf{H}$ . $G_{11} = \|\mathbf{h}_1\|^2 = 7$ , $G_{22} = \|\mathbf{h}_2\|^2 = |1-j|^2 + 1 + 1 = 4$ . $G_{12} = \mathbf{h}_1^H\mathbf{h}_2 = -5j$ . $\det(\mathbf{G}) = 7 \cdot 4 - |{-5j}|^2 = 28 - 25 = 3$ . $[\mathbf{G}^{-1}]_{11} = G_{22}/\det = 4/3$ . $\text{SINR}_1^{\text{ZF}} = P_1 / (\sigma^2 [\mathbf{G}^{-1}]_{11}) = 1/(1 \cdot 4/3) = 3/4 = 0.75$ .

(c) Comparison

MRC SINR $= 1.53 >$ ZF SINR $= 0.75$ at this small $N_t = 3$ . ZF eliminates interference but the noise enhancement factor $[\mathbf{G}^{-1}]_{11} = 4/3$ exceeds $1/\|\mathbf{h}_1\|^2 = 1/7$ : ZF increases the effective noise by $4/3 \cdot 7 \approx 9.3\times$ . In this low- $N_t$ regime, MRC outperforms ZF because the interference is modest ( $|\mathbf{h}_1^H\mathbf{h}_2|^2/\|\mathbf{h}_1\|^4 = 25/49 \approx 51\%$ ) but the noise enhancement is severe. For $N_t \gg K$ , both SINRs diverge and converge to the same rate. $\blacksquare$

Quick Check

In the asymptotic regime $N_t \to \infty$ with MRC combining, what happens to the inter-user interference term in the post-combining signal for user $k$ ?

It grows linearly with $N_t$

It remains constant

It vanishes (grows sublinearly compared to the desired signal)

It grows as $N_t^2$ (same as the desired signal)