Ferkans — Interactive Telecom Tutor

Linear Processing Becomes Near-Optimal

One of the most profound consequences of operating in the massive MIMO regime is that linear processing — which is decidedly suboptimal for conventional $2 \times 2$ or $4 \times 4$ MIMO — becomes asymptotically optimal as $M \to \infty$ . The performance gap between the matched filter (MR) and the MMSE receiver, and between ZF precoding and dirty paper coding, vanishes. This means that the base station can serve $K$ users with $O(MK)$ computational complexity instead of the exponential cost of joint detection or DPC encoding. The theoretical foundation for this claim is the use-and-then-forget (UatF) bound, which provides a rigorous achievable rate that depends only on channel statistics.

Definition:
Linear Combining for Massive MIMO Uplink

In the uplink, the BS applies a combining vector $\mathbf{v}_{k} \in \mathbb{C}^M$ to the received signal $\mathbf{y}$ to estimate user $k$ 's symbol:

$\hat{x}_k = \mathbf{v}_{k}^{H} \mathbf{y} = \mathbf{v}_{k}^{H} \mathbf{h}_k x_k + \sum_{j \neq k}\mathbf{v}_{k}^{H}\mathbf{h}_j x_j + \mathbf{v}_{k}^{H}\mathbf{n}$

The three standard combining schemes in massive MIMO are:

Maximum Ratio (MR): $\mathbf{v}_{k} = \mathbf{h}_k$
Zero Forcing (ZF): $\mathbf{v}_{k} = \mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}\mathbf{e}_k$
MMSE: $\mathbf{v}_{k} = (\mathbf{H}\mathbf{H}^{H} + \sigma^2\mathbf{I}_M)^{-1}\mathbf{h}_k$

In the downlink, the corresponding precoding vectors $\mathbf{w}_k$ are:

MR (MRT): $\mathbf{w}_k = \mathbf{h}_k / \|\mathbf{h}_k\|$
ZF: $\mathbf{w}_k = \mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}\mathbf{e}_k / \|[\cdot]_k\|$
MMSE (RZF): $\mathbf{w}_k \propto (\mathbf{H}\mathbf{H}^{H} + \alpha\mathbf{I})^{-1}\mathbf{h}_k$

MR has complexity $O(MK)$ and requires no matrix inversion. ZF and MMSE require inverting a $K \times K$ matrix, costing $O(MK^2 + K^3)$ . Since $K \ll M$ , all three schemes are computationally tractable.

,

Theorem: Achievable Rate with MR Combining (Use-and-then-Forget Bound)

Under MR combining with perfect CSI, the uplink ergodic achievable rate for user $k$ in a massive MIMO system with i.i.d. Rayleigh fading is:

$R_k^{\text{MR}} = \log_2\!\left(1 + \frac{p_k M \beta_k^2} {\sum_{j \neq k} p_j \beta_j \beta_k + \sigma^2 \beta_k}\right) \xrightarrow{M \to \infty} \log_2\!\left(1 + \frac{p_k M \beta_k} {\sum_{j \neq k} p_j \beta_j + \sigma^2}\right)$

For equal power $p_k = p$ and equal path loss $\beta_k = \beta$ :

$R_k^{\text{MR}} \approx \log_2\!\left(1 + \frac{Mp\beta}{(K-1)p\beta + \sigma^2}\right)$

In the massive MIMO limit ( $M \to \infty$ with $K$ fixed):

$R_k^{\text{MR}} \to \infty \quad \text{(grows as } \log_2 M\text{)}$

With MR combining, the desired signal power grows as $M^2\beta_k^2$ (coherent combining of $M$ antennas), inter-user interference grows as $M\beta_j\beta_k$ (incoherent summation), and noise grows as $M\sigma^2\beta_k$ (noise enhancement from MR filter norm). The SINR therefore grows linearly with $M$ . In the large- $M$ limit, the noise becomes negligible and the rate is limited only by inter-user interference — but even this interference is suppressed at rate $M/(K-1)$ .

Proof

MR combining output

With $\mathbf{v}_{k} = \mathbf{h}_k$ , the combined signal is:

$\hat{x}_k = \mathbf{h}_k^H\mathbf{y} = \underbrace{\sqrt{p_k}\|\mathbf{h}_k\|^2 x_k}_{\text{desired}} + \sum_{j \neq k}\underbrace{\sqrt{p_j}\mathbf{h}_k^H\mathbf{h}_j x_j}_{\text{interference}} + \underbrace{\mathbf{h}_k^H\mathbf{n}}_{\text{noise}}$

Use-and-then-forget bound

The UatF bound treats the channel estimate as known and the estimation error as worst-case (Gaussian) noise. With perfect CSI, the effective SINR is:

$\text{SINR}_k = \frac{p_k \mathbb{E}[\|\mathbf{h}_k\|^2]^2} {\sum_{j \neq k} p_j \mathbb{E}[|\mathbf{h}_k^H\mathbf{h}_j|^2] + \sigma^2 \mathbb{E}[\|\mathbf{h}_k\|^2] + p_k\text{Var}[\|\mathbf{h}_k\|^2]}$

Computing each term for i.i.d. Rayleigh:

$\mathbb{E}[\|\mathbf{h}_k\|^2] = M\beta_k$
$\mathbb{E}[|\mathbf{h}_k^H\mathbf{h}_j|^2] = M\beta_k\beta_j$
$\text{Var}[\|\mathbf{h}_k\|^2] = M\beta_k^2$

Simplification

$\text{SINR}_k^{\text{MR}} = \frac{p_k M^2\beta_k^2} {\sum_{j \neq k} p_j M\beta_k\beta_j + \sigma^2 M\beta_k + p_k M\beta_k^2}KATEXPLACEHOLDER0END= \frac{p_k M \beta_k} {\sum_{j \neq k} p_j\beta_j + \sigma^2 + p_k\beta_k}$ $For$ M \gg 1 $, the term$ p_k\beta_k $in the denominator becomes negligible compared to$ p_k M\beta_k $in the numerator, yielding the stated result.$ \blacksquare$

,

Theorem: Achievable Rates with ZF and MMSE Processing

Under i.i.d. Rayleigh fading with perfect CSI, the uplink achievable rates with ZF and MMSE combining are:

ZF combining (requires $M > K$ ):

$R_k^{\text{ZF}} = \log_2\!\left(1 + p_k(M - K)\beta_k \cdot \frac{1}{\sigma^2}\right)$

The SINR grows as $(M - K)$ , reflecting the loss of $K$ degrees of freedom for interference nulling.

MMSE combining:

$R_k^{\text{MMSE}} = \log_2\!\left(1 + p_k\beta_k\, \mathbf{e}_k^T\!\left(\frac{1}{M}\mathbf{H}^{H}\mathbf{H} + \frac{\sigma^2}{M}\mathbf{I}_K\right)^{-1}\!\mathbf{e}_k\right)$

which converges for large $M$ to:

$R_k^{\text{MMSE}} \to \log_2\!\left(1 + \frac{p_k M\beta_k}{\sigma^2}\right)$

In the massive MIMO limit, all three schemes converge:

$R_k^{\text{MR}} \approx R_k^{\text{ZF}} \approx R_k^{\text{MMSE}} \approx \log_2\!\left(1 + \frac{p_k M\beta_k}{\sigma^2}\right)$

As $M \to \infty$ with $K$ fixed, favourable propagation makes the channels orthogonal, so there is no interference to null. ZF "wastes" $K$ degrees of freedom nulling already-negligible interference, but $M - K \approx M$ when $M \gg K$ . MMSE optimally balances noise enhancement and interference suppression, but when $M$ is large, all interference is already negligible and MMSE reduces to MR. The three schemes converge because the problem they solve — interference suppression — becomes negligible in the massive MIMO regime.

Proof

ZF SINR derivation

The ZF combining vector satisfies $\mathbf{v}_{k}^{H}\mathbf{h}_j = \delta_{kj}$ , eliminating all inter-user interference. The post-ZF noise variance involves $[(\mathbf{H}^{H}\mathbf{H})^{-1}]_{kk}$ .

For i.i.d. Rayleigh with $M > K$ , the diagonal elements of the inverse Wishart matrix satisfy:

$\mathbb{E}\!\left[\frac{1}{[(\mathbf{H}^{H}\mathbf{H})^{-1}]_{kk}}\right] = (M - K)\beta_k$

yielding $\text{SINR}_k^{\text{ZF}} = p_k(M-K)\beta_k/\sigma^2$ .

Asymptotic convergence

For $M \gg K$ :

MR: $\text{SINR} \sim M p_k\beta_k/\sigma^2$ (noise-limited)
ZF: $\text{SINR} \sim (M-K)p_k\beta_k/\sigma^2 \approx Mp_k\beta_k/\sigma^2$
MMSE: converges to ZF as interference vanishes

All three SINRs scale linearly with $M$ and converge. $\blacksquare$

, ,

Achievable Rate vs. Number of Antennas

Compare the achievable uplink rates with MR, ZF, and MMSE combining as a function of $M$ . Observe how all three curves converge as $M$ grows, confirming that linear processing becomes near-optimal in the massive MIMO regime.

Parameters

K

(number of users)10

SNR [dB]0

Channel Hardening Animation

Watch the effective channel gain $|\mathbf{v}_{k}^{H}\mathbf{h}_k|^2$ harden as the number of BS antennas $M$ increases from 4 to 256. Each frame shows a different random channel realisation, illustrating how the per-user rate fluctuations decrease dramatically with $M$ .

Parameters

K

(number of users)5

SNR [dB]0

Example: Rate Comparison MR vs. ZF for M=64, K=10

A massive MIMO base station has $M = 64$ antennas and serves $K = 10$ users with equal power $p = 1$ and equal path loss $\beta_k = 1$ . The noise variance is $\sigma^2 = 1$ (SNR = 0 dB per antenna). Compute the per-user achievable rate with MR and ZF combining.

Solution

MR combining rate

$\text{SINR}^{\text{MR}} = \frac{M p\beta}{(K-1)p\beta + \sigma^2} = \frac{64 \times 1}{9 \times 1 + 1} = \frac{64}{10} = 6.4KATEXPLACEHOLDER0ENDR^{\text{MR}} = \log_2(1 + 6.4) = \log_2(7.4) \approx 2.89\;\text{bits/s/Hz}$ $

ZF combining rate

$\text{SINR}^{\text{ZF}} = \frac{(M-K)p\beta}{\sigma^2} = \frac{(64-10) \times 1}{1} = 54KATEXPLACEHOLDER0ENDR^{\text{ZF}} = \log_2(1 + 54) = \log_2(55) \approx 5.78\;\text{bits/s/Hz}$ $

Comparison

ZF provides about 2.9 bits/s/Hz more than MR per user, nearly double the rate. This gap is significant for $M/K = 6.4$ , which is not yet deep in the massive MIMO regime. For $M = 256$ ( $M/K = 25.6$ ), the gap narrows substantially. $\blacksquare$

Quick Check

In the massive MIMO limit ( $M \to \infty$ , $K$ fixed), which statement about linear processing is correct?

ZF always outperforms MR by a factor of K

MR combining achieves the same rate as the optimal joint detector

MMSE combining requires matrix inversion that grows as O(M^3)

Linear precoding cannot approach the DPC capacity region

Correction:

MR combining achieves the same rate as the optimal joint detector

As $M \to \infty$ , favourable propagation eliminates inter-user interference, so MR is asymptotically optimal. The SINR of all three schemes converges to $Mp_k\beta_k/\sigma^2$ .

Why This Matters: Massive MIMO in 5G NR

5G NR (New Radio) has adopted massive MIMO as a core technology for both sub-6 GHz and mmWave deployments. Key specifications:

Antenna configurations: Up to 256 antenna elements at the gNB (base station), typically 64T64R (64 transmit, 64 receive chains) for sub-6 GHz and 256--512 elements for mmWave.
CSI framework: Type I and Type II codebook-based feedback for FDD; channel reciprocity (SRS-based) for TDD.
Processing: ZF or RZF (regularised ZF) precoding with up to 16 spatial layers.
Beam management: SSB (Synchronisation Signal Block) beams for initial access; CSI-RS for refined beam tracking.

Field deployments by operators worldwide have demonstrated 3--5 $\times$ spectral efficiency gains over 4G LTE MIMO (typically $2 \times 2$ or $4 \times 4$ ), with per-cell throughputs exceeding 1 Gbps in the downlink using 64 antenna ports.

See full treatment in Chapter 19

Key Takeaway

In the massive MIMO regime ( $M \gg K$ ), the simplest linear processing scheme — matched filter / MR combining with complexity $O(MK)$ — achieves the same asymptotic rate as the optimal joint decoder (exponential complexity). This means that massive MIMO converts a hard signal processing problem into a straightforward one through sheer antenna numbers.

Deeper Treatment in the MIMO Book

The signal processing foundations of massive MIMO — linear combining, MMSE estimation, the capacity of MIMO channels — build directly on the MIMO theory developed in Chapters 15--17 of this book. The MIMO specialised book extends the treatment to spatially correlated channels (beyond the i.i.d. Rayleigh model of this chapter), random matrix theory for deterministic equivalents, and advanced precoding techniques including successive encoding and dirty paper coding. Readers interested in the information- theoretic foundations of multi-antenna systems should consult that resource.

Use-and-then-Forget Bound

A lower bound on the ergodic achievable rate that treats the known part of the channel as deterministic and the unknown part (estimation error, interference) as worst-case Gaussian noise. Formally: $R_k = \log_2(1 + \text{SINR}_k)$ where the SINR is computed from channel statistics rather than instantaneous values.

Maximum Ratio (MR) Combining

A linear combining scheme that sets $\mathbf{v}_{k} = \mathbf{h}_k$ , maximising the received SNR for user $k$ but not suppressing inter-user interference. Also called matched filtering (MF). In massive MIMO, MR becomes near-optimal due to favourable propagation.

Uplink and Downlink Processing

Linear Processing Becomes Near-Optimal

Definition: Linear Combining for Massive MIMO Uplink

Theorem: Achievable Rate with MR Combining (Use-and-then-Forget Bound)

MR combining output

Use-and-then-forget bound

Simplification

Theorem: Achievable Rates with ZF and MMSE Processing

ZF SINR derivation

Asymptotic convergence

Achievable Rate vs. Number of Antennas

Parameters

Channel Hardening Animation

Parameters

Example: Rate Comparison MR vs. ZF for M=64, K=10

MR combining rate

ZF combining rate

Comparison

Quick Check

Why This Matters: Massive MIMO in 5G NR

Key Takeaway

Deeper Treatment in the MIMO Book

Use-and-then-Forget Bound

Maximum Ratio (MR) Combining

Definition:
Linear Combining for Massive MIMO Uplink