Ferkans — Interactive Telecom Tutor

Bridging MRT and ZF

MRT maximises signal power but ignores interference. ZF eliminates interference but amplifies noise. Is there a middle ground? Regularized zero-forcing (RZF), also known as MMSE precoding, adds a regularization term $\alpha \mathbf{I}$ to the channel Gram matrix before inversion. By tuning $\alpha$ , we smoothly interpolate between MRT ( $\alpha \to \infty$ ) and ZF ( $\alpha \to 0$ ), achieving the best SINR tradeoff at any operating point.

Definition:
Regularized Zero-Forcing (RZF) Precoding

The RZF precoding matrix is

$\mathbf{W}^{\text{RZF}} = \mathbf{H}^{H} (\mathbf{H}\mathbf{H}^{H} + \alpha \mathbf{I})^{-1} \mathbf{D}_{\text{RZF}}$

where $\alpha > 0$ is the regularization parameter and $\mathbf{D}_{\text{RZF}}$ is a diagonal normalisation matrix ensuring unit-norm columns.

The per-user (unnormalised) precoding vector is

$\tilde{\mathbf{v}}_k^{\text{RZF}} = (\mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I})^{-1} \mathbf{h}_k$

using the matrix inversion lemma to write the equivalent form.

When $\alpha = 0$ , RZF reduces to ZF. When $\alpha \to \infty$ , the inverse approaches $(1/\alpha) \mathbf{I}$ and RZF reduces to MRT (up to scaling). The name "MMSE precoding" comes from the fact that the optimal $\alpha$ minimises the mean squared error between the transmitted and intended signals.

,

Theorem: Optimal Regularization Parameter

For i.i.d. Rayleigh fading with equal power allocation, the regularization parameter that maximises the asymptotic (large $N_t$ ) sum rate is

$\alpha^{\star} = \frac{K\, \sigma^2}{P_t}.$

This is the ratio of total noise power (across all users) to the transmit power.

The optimal $\alpha$ balances two costs: too small an $\alpha$ causes noise amplification (like ZF), while too large an $\alpha$ permits too much interference (like MRT). The sweet spot is where the regularization equals the "noise per degree of freedom," which is $K\sigma^2/P_t$ .

At high SNR ( $P_t/\sigma^2 \to \infty$ ), $\alpha^{\star} \to 0$ and RZF converges to ZF. At low SNR, $\alpha^{\star}$ is large and RZF behaves like MRT.

Proof

Formulate the SINR

The SINR under RZF with regularization $\alpha$ involves a tradeoff between:

Signal power: $|\mathbf{h}_k^H \tilde{\mathbf{v}}_k|^2$ , which decreases as $\alpha$ increases (the precoder moves away from the channel-inversion direction).
Interference: $\sum_{j \neq k} |\mathbf{h}_k^H \tilde{\mathbf{v}}_j|^2$ , which is zero at $\alpha = 0$ (ZF) and grows with $\alpha$ .
Noise amplification: $\|\tilde{\mathbf{v}}_k\|^2$ , which is large when $\alpha$ is small and decreases with $\alpha$ .

Large-system analysis

Using random matrix theory (Marchenko--Pastur law), as $N_t, K \to \infty$ with $K/N_t \to \beta \in (0,1)$ , the per-user SINR converges to a deterministic function of $\alpha$ . Differentiating with respect to $\alpha$ and setting to zero yields

$\alpha^{\star} = \frac{K\, \sigma^2}{P_t}. \quad \blacksquare$

,

Theorem: RZF SINR Expression

With RZF precoding, regularization $\alpha$ , and equal power allocation $p_k = P_t/K$ , the SINR at user $k$ is

$\text{SINR}_k^{\text{RZF}} = \frac{\frac{P_t}{K} |\mathbf{h}_k^H (\mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I})^{-1} \mathbf{h}_k|^2}{\frac{P_t}{K} \sum_{j \neq k} |\mathbf{h}_k^H (\mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I})^{-1} \mathbf{h}_j|^2 + \sigma^2 \sum_{j=1}^{K} \|(\mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I})^{-1} \mathbf{h}_j\|^2 \cdot c_k}$

where $c_k$ is a normalisation constant. In the large-system limit, this converges to a deterministic equivalent depending on $\alpha$ , $K/N_t$ , and $P_t/\sigma^2$ .

The expression is complex but the message is simple: RZF trades off residual interference (nonzero for $\alpha > 0$ ) against reduced noise amplification (better conditioned inverse). At $\alpha = \alpha^{\star}$ , the total "interference plus amplified noise" is minimised.

Proof

Substitute RZF vectors into general SINR

Substituting $\tilde{\mathbf{v}}_j = (\mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I})^{-1} \mathbf{h}_j$ into the general SINR formula from Definition DSINR for Linear Precoding and accounting for the power normalisation yields the stated expression. The deterministic equivalent follows from applying the resolvent identity and trace lemma from random matrix theory. $\blacksquare$

Example: Effect of Regularization on Sum Rate

For $N_t = 16$ , $K = 8$ , and $P_t/\sigma^2 = 15$ dB, compute the sum rate $R_{\text{sum}} = \sum_{k=1}^{K} \log_2(1 + \text{SINR}_k)$ for $\alpha \in \{0, 0.1, 0.25, 0.5, 1, 10\}$ via Monte Carlo simulation. Verify that the optimum is near $\alpha^{\star} = K\sigma^2/P_t = 8/31.6 \approx 0.25$ .

Solution

Generate random channels

Draw $\mathbf{H} \in \mathbb{C}^{K \times N_t}$ with i.i.d. $\mathcal{CN}(0, 1)$ entries. Average the sum rate over 1000 realisations.

Compute precoders and SINRs

For each $\alpha$ , form $\tilde{\mathbf{W}} = \mathbf{H}^{H}(\mathbf{H}\mathbf{H}^{H} + \alpha \mathbf{I})^{-1}$ , normalise columns, allocate $p_k = P_t/K$ , and compute SINR per user.

Results

$\alpha$	0 (ZF)	0.1	0.25	0.5	1.0	10 (MRT-like)
$R_{\text{sum}}$	28.1	29.4	30.2	29.8	28.5	18.3

The optimal $\alpha \approx 0.25$ matches the theoretical prediction $\alpha^{\star} = K\sigma^2/P_t$ . RZF with optimal regularization gains $\sim 2$ bits/s/Hz over ZF and $\sim 12$ bits/s/Hz over MRT at this operating point.

RZF Sum Rate vs Regularization $\alpha$

Sweep the regularization parameter $\alpha$ and observe the sum rate. The vertical dashed line marks the optimal $\alpha^{\star} = K\sigma^2/P_t$ . Compare the sum rate at $\alpha = 0$ (ZF) and $\alpha \to \infty$ (MRT).

Parameters

N_t

32

K

8

P_t/\sigma^2

(dB)10

Sum Rate vs $N_t$ — MRT, ZF, RZF

Compare the sum rate of MRT, ZF, and RZF (with optimal $\alpha$ ) as the number of antennas grows. Observe that all three converge in the massive regime but differ significantly at moderate antenna counts.

Parameters

K

8

P_t/\sigma^2

(dB)10

MRT vs ZF vs RZF — Summary

Property	MRT	ZF	RZF (MMSE)
Precoding vector	$\mathbf{h}_k/\\|\mathbf{h}_k\\|$	$[\mathbf{H}^{H}(\mathbf{H}\mathbf{H}^{H})^{-1}]_{:,k}$ (normalised)	$[\mathbf{H}^{H}(\mathbf{H}\mathbf{H}^{H} + \alpha\mathbf{I})^{-1}]_{:,k}$ (normalised)
Interference	Nonzero (ignored)	Zero	Small (controlled)
Noise amplification	None	Severe when $K o N_t$	Moderate (regularized)
Complexity	$O(N_t K)$	$O(N_t K^2 + K^3)$	$O(N_t K^2 + K^3)$
Best regime	$N_t \gg K$ , low SNR	$N_t \gg K$ , high SNR	All regimes
Requires	Channel vectors	Full CSI + inversion	Full CSI + inversion + $\alpha$

Efficient RZF Precoder Computation

Complexity:

O(N_tK^{2} + K^{3})

, dominated by the matrix-matrix product in step 1 and the Cholesky factorisation in step 2. This is feasible for real-time operation with

K \leq 64

and

N_t \leq 256

on modern DSP hardware.

Input: Channel matrix

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

,

regularization

\alpha > 0

, power budget

P_t

1. Compute Gram matrix:

\mathbf{G} = \mathbf{H}\mathbf{H}^{H} + \alpha \mathbf{I}_{K}

\quad

//

O(N_tK^{2})

2. Cholesky factorisation:

\mathbf{G} = \mathbf{L}\mathbf{L}^H

\quad

//

O(K^{3})

3. Solve

\mathbf{L}\mathbf{L}^H \mathbf{B} = \mathbf{I}_{K}

for

\mathbf{B} = \mathbf{G}^{-1}

\quad

//

O(K^{3})

via back-substitution

4. Form unnormalised precoders:

\tilde{\mathbf{W}} = \mathbf{H}^{H} \mathbf{B}

\quad

//

O(N_tK^{2})

5. Normalise:

\mathbf{v}_{k} = \tilde{\mathbf{v}}_k / \|\tilde{\mathbf{v}}_k\|

for

k = 1, \ldots, K

6. Allocate power:

p_k = P_t/K

(equal allocation)

Output: Precoding vectors

\mathbf{v}_{1}, \ldots, \mathbf{v}_{\ntn{nusers}}

and powers

p_1, \ldots, p_{K}

Using the matrix inversion lemma, one can equivalently compute via the $N_t \times N_t$ matrix $\mathbf{H}^{H}\mathbf{H} + \alpha \mathbf{I}_{N_t}$ , which is preferred when $K > N_t$ (rare in practice).

⚠️Engineering Note

Estimating $\alpha$ in Practice

The theoretical optimum $\alpha^{\star} = K\sigma^2/P_t$ assumes i.i.d. Rayleigh fading with perfect CSI. In practice:

Noise variance estimation: $\sigma^2$ is estimated from noise-only subcarriers or the off-diagonal elements of the received signal covariance. A 1--2 dB error in $\hat{\sigma^2}$ shifts $\alpha$ by the same factor.
Correlated channels: With spatial correlation, the optimal $\alpha$ depends on the eigenvalue spread of $\mathbf{H}\mathbf{H}^{H}$ . A practical rule is to use $\alpha = \text{tr}(\mathbf{H}\mathbf{H}^{H}) \cdot \sigma^2/(K\,P_t)$ .
Imperfect CSI: When the channel is estimated with error variance $\sigma_e^2$ , the effective regularization should be increased: $\alpha_{\text{eff}} = K(\sigma^2 + P_t\sigma_e^2)/P_t$ .

Historical Note: The MMSE Precoding Lineage

2003--2012

The idea of regularized channel inversion appeared independently in several groups around 2003--2005. Joham, Utschick, and Nossek (2005) derived it from the MMSE criterion for the transmit signal. Peel, Hochwald, and Swindlehurst (2005) approached it from the "vector perturbation" perspective, showing that linear regularized inversion is the first step toward nonlinear precoding. The large-system analysis by Wagner, Couillet, Debbah, and Slock (2012) provided the deterministic equivalent that made the optimal $\alpha$ analytically tractable in the massive MIMO regime.

,

Quick Check

What happens to the RZF precoding matrix as $\alpha \to \infty$ ?

It converges to the ZF precoder

It converges to scaled MRT (conjugate beamforming)

It converges to the identity matrix

It diverges

Correction:

It converges to scaled MRT (conjugate beamforming)

As $\alpha \to \infty$ , $(\mathbf{H}\mathbf{H}^{H} + \alpha\mathbf{I})^{-1} \to (1/\alpha)\mathbf{I}$ , so $\mathbf{W}^{\text{RZF}} \to (1/\alpha)\mathbf{H}^{H}$ , which is MRT up to a scalar.

Regularized Zero-Forcing (RZF)

Linear precoding with regularization: $\mathbf{W} = \mathbf{H}^{H}(\mathbf{H}\mathbf{H}^{H} + \alpha\mathbf{I})^{-1}$ . Bridges MRT ( $\alpha \to \infty$ ) and ZF ( $\alpha = 0$ ). Also called MMSE precoding. Optimal $\alpha = K\sigma^2/P_t$ .

Regularization Parameter

A positive scalar $\alpha$ added to the diagonal of a matrix before inversion to improve numerical conditioning and balance noise amplification against residual interference. In RZF precoding, $\alpha$ controls the MRT--ZF tradeoff.

Key Takeaway

RZF/MMSE precoding is the practical workhorse of MU-MIMO. With optimal regularization $\alpha^{\star} = K\sigma^2/P_t$ , it achieves the best linear precoding performance at any SNR and loading. It dominates MRT at high SNR, dominates ZF at high loading, and matches both in their respectively optimal regimes.

Regularized Zero-Forcing (MMSE Precoding)

Bridging MRT and ZF

Definition: Regularized Zero-Forcing (RZF) Precoding

Theorem: Optimal Regularization Parameter

Formulate the SINR

Large-system analysis

Theorem: RZF SINR Expression

Substitute RZF vectors into general SINR

Example: Effect of Regularization on Sum Rate

Generate random channels

Compute precoders and SINRs

Results

RZF Sum Rate vs Regularization α\alphaα

Parameters

Sum Rate vs NtN_tNt​ — MRT, ZF, RZF

Parameters

MRT vs ZF vs RZF — Summary

Efficient RZF Precoder Computation

Estimating α\alphaα in Practice

Historical Note: The MMSE Precoding Lineage

Quick Check

Regularized Zero-Forcing (RZF)

Regularization Parameter

Key Takeaway

Definition:
Regularized Zero-Forcing (RZF) Precoding

RZF Sum Rate vs Regularization $\alpha$

Sum Rate vs $N_t$ — MRT, ZF, RZF

Estimating $\alpha$ in Practice