Ferkans — Interactive Telecom Tutor

Beyond MRC: Interference Suppression

MRC treats each user independently, ignoring multi-user interference. When $K$ is not negligible compared to $N_t$ , this interference dominates and the MRC rate saturates. Two alternatives exploit the excess spatial degrees of freedom to suppress interference:

Zero Forcing (ZF): Projects each user's signal onto the subspace orthogonal to all other users' estimated channels. This eliminates interference completely (given perfect estimates) but amplifies noise because the projection reduces the effective array dimension.
MMSE (Regularized ZF): Adds a regularization term that balances interference suppression against noise amplification. This is optimal among linear receivers in the MSE sense.

Both require inverting a $K \times K$ matrix, but the payoff is dramatic: the per-user rate with ZF/MMSE can be substantially higher than MRC, especially when $K$ approaches $N_t$ .

,

Definition:
Zero-Forcing Combining Vector

Let $\hat{\mathbf{H}} = [\hat{\mathbf{H}}_1, \ldots, \hat{\mathbf{H}}_{K}] \in \mathbb{C}^{N_t \times K}$ be the matrix of channel estimates. The ZF combining vector for user $k$ is the $k$ -th column of

$\mathbf{v}^{\text{ZF}} = \hat{\mathbf{H}} \left(\hat{\mathbf{H}}^H \hat{\mathbf{H}}\right)^{-1},$

i.e., $\mathbf{v}_{k}^{\text{ZF}} = \hat{\mathbf{H}} \left(\hat{\mathbf{H}}^H \hat{\mathbf{H}}\right)^{-1} \mathbf{e}_k$ , where $\mathbf{e}_k$ is the $k$ -th standard basis vector.

ZF requires $N_t \geq K$ for the pseudo-inverse to exist. In massive MIMO, $N_t \gg K$ is the typical operating regime, so this condition is easily satisfied.

Definition:
MMSE (Regularized ZF) Combining Vector

The MMSE combining vector for user $k$ is

$\mathbf{v}_{k}^{\text{MMSE}} = \left(\sum_{j=1}^{K} {P_t}_{j} \hat{\mathbf{H}}_j \hat{\mathbf{H}}_j^H + \mathbf{Z}\right)^{-1} \sqrt{{P_t}_{k}} \, \hat{\mathbf{H}}_k,$

where $\mathbf{Z} = \sum_{j=1}^{K} {P_t}_{j} (\beta_{j} - \gamma_j) \mathbf{I} + \sigma^2 \mathbf{I}$ accounts for the estimation error variance and noise.

When $\sigma^2 \to 0$ and the estimation is perfect ( $\gamma_k = \beta_{k}$ ), the MMSE combiner reduces to ZF. The regularization term $\mathbf{Z}$ prevents noise amplification by keeping the matrix inversion well-conditioned.

,

Theorem: Closed-Form UatF Rate with ZF Combining

Under i.i.d. Rayleigh fading with MMSE estimation and no pilot contamination, the UatF achievable rate for user $k$ with ZF combining is

$R_k^{\text{ZF}} = \log_2\!\left(1 + \frac{(N_t - K) \, {P_t}_{k} \, \gamma_k}{\sum_{j=1}^{K} {P_t}_{j} (\beta_{j} - \gamma_j) + \sigma^2}\right),$

valid for $N_t > K$ .

Comparing with the MRC formula, two changes stand out:

The array gain is $N_t - K$ instead of $N_t$ . ZF "uses up" $K - 1$ degrees of freedom to null interference, leaving $N_t - K$ for coherent combining. When $N_t \gg K$ , the loss is negligible.
The denominator contains $\beta_{j} - \gamma_j$ (the estimation error variance) instead of $\beta_{j}$ (the full channel power). ZF eliminates the component of interference along the estimated channels — only the part due to estimation error remains. With perfect estimation ( $\gamma_j = \beta_{j}$ ), the interference vanishes entirely and the denominator is just $\sigma^2$ .

Proof

Step 1: ZF eliminates estimated-channel interference

By construction, $(\mathbf{v}_{k}^{\text{ZF}})^H \hat{\mathbf{H}}_j = 0$ for $j \neq k$ and $(\mathbf{v}_{k}^{\text{ZF}})^H \hat{\mathbf{H}}_k = 1$ . Therefore

$\mathbf{v}_{k}^{H} \mathbf{H}_{j} = \mathbf{v}_{k}^{H} (\hat{\mathbf{H}}_j + \tilde{\mathbf{H}}_j) = \mathbf{v}_{k}^{H} \tilde{\mathbf{H}}_j \quad (j \neq k).$

Step 2: Compute the signal mean

$\mathbb{E}[\mathbf{v}_{k}^{H} \mathbf{H}_{k}] = \mathbb{E}[(\mathbf{v}_{k}^{\text{ZF}})^H \hat{\mathbf{H}}_k] + \mathbb{E}[(\mathbf{v}_{k}^{\text{ZF}})^H \tilde{\mathbf{H}}_k] = 1 + 0 = 1.$ $The numerator of the UatF SINR is simply$ {P_t}_{k}$.

Step 3: Compute the interference power

For $j \neq k$ : $\mathbb{E}[|\mathbf{v}_{k}^{H} \mathbf{H}_{j}|^2] = \mathbb{E}[|\mathbf{v}_{k}^{H} \tilde{\mathbf{H}}_j|^2]$ . Since $\tilde{\mathbf{H}}_j \sim \mathcal{CN}(\mathbf{0}, (\beta_{j} - \gamma_j)\mathbf{I})$ and is independent of $\mathbf{v}_{k}$ :

$\mathbb{E}[|\mathbf{v}_{k}^{H} \tilde{\mathbf{H}}_j|^2] = (\beta_{j} - \gamma_j) \, \mathbb{E}[\|\mathbf{v}_{k}\|^2].$

Step 4: Use the trace identity for ZF norm

For i.i.d. channels, the expected ZF combining norm is

$\mathbb{E}[\|\mathbf{v}_{k}^{\text{ZF}}\|^2] = \mathbb{E}\!\left[[\left(\hat{\mathbf{H}}^H \hat{\mathbf{H}}\right)^{-1}]_{kk}\right] = \frac{1}{\gamma_k (N_t - K)},$

which follows from the distribution of the inverse Wishart matrix.

Step 5: Assemble the SINR

$\text{SINR}_k^{\text{ZF}} = \frac{{P_t}_{k}}{\frac{1}{\gamma_k(N_t - K)}\left(\sum_j {P_t}_{j} (\beta_{j} - \gamma_j) + \sigma^2\right)} = \frac{(N_t - K) \, {P_t}_{k} \, \gamma_k}{\sum_j {P_t}_{j}(\beta_{j} - \gamma_j) + \sigma^2}. \quad \blacksquare$ $

,

Theorem: UatF Rate with MMSE Combining (Large-System Approximation)

Under i.i.d. Rayleigh fading in the large-system limit ( $N_t, K \to \infty$ with $K/N_t \to \alpha \in (0,1)$ ), the UatF achievable rate with MMSE combining converges to

$R_k^{\text{MMSE}} \to \log_2\!\left(1 + {P_t}_{k} \gamma_k \, m_k(-\sigma^2)\right),$

where $m_k(z)$ is the Stieltjes transform of the asymptotic eigenvalue distribution of the interference-plus-noise covariance matrix, evaluated at $z = -\sigma^2$ . For equal power and equal path loss, this simplifies to

$R_k^{\text{MMSE}} \approx \log_2\!\left(1 + \frac{N_t \, P_t \, \gamma}{K \, P_t (\beta - \gamma) + \sigma^2} \cdot \frac{1}{1 + \alpha \cdot \delta}\right),$

where $\delta$ is the unique positive solution to $\delta = P_t \gamma / (\sigma^2 + K P_t(\beta - \gamma) / N_t + P_t \gamma \, \alpha \, \delta / (1 + \delta))$ .

The MMSE rate interpolates between MRC (when $\sigma^2$ dominates) and ZF (when interference dominates). The random matrix theory machinery is needed because the MMSE combiner couples all users through the matrix inversion, making the per-user SINR depend on the joint statistics of all channels.

For practical purposes, when $N_t / K \geq 5$ , the MMSE rate is very close to the ZF rate, and the simpler ZF formula suffices for system design.

Proof

Sketch: Random matrix theory approach

The MMSE SINR for user $k$ involves $\mathbf{v}_{k}^{H} \mathbf{H}_{k}$ where $\mathbf{v}_{k} = (\hat{\mathbf{H}} \hat{\mathbf{H}}^H + \alpha \mathbf{I})^{-1} \hat{\mathbf{H}}_k$ . In the large-system limit, the resolvent $(\hat{\mathbf{H}} \hat{\mathbf{H}}^H + \alpha \mathbf{I})^{-1}$ concentrates around a deterministic equivalent characterized by the Stieltjes transform of the Marchenko-Pastur distribution. The rigorous proof uses the rank-1 perturbation formula and the matrix inversion lemma — see Hoydis, ten Brink, and Debbah (2013) for the full derivation. $\blacksquare$

,

Historical Note: Random Matrix Theory Enters Wireless Communications

1996-2013

The application of random matrix theory (RMT) to wireless communications began with the landmark papers of Telatar (1999) and Foschini (1996), who showed that MIMO capacity scales linearly with the minimum of $N_t$ and $N_r$ . The Marchenko-Pastur law, originally developed in the context of nuclear physics (1967), became an essential tool for analyzing large MIMO systems.

Hoydis, ten Brink, and Debbah (2013) brought RMT-based analysis to massive MIMO, deriving deterministic equivalents for the SINR under MMSE processing. Their results showed that RMT predictions are accurate even for modest system sizes ( $N_t = 64$ , $K = 16$ ), vindicating the large-system approach for practical 5G design.

,

Historical Note: From Interference Cancellation to Zero Forcing

1990s-2010s

Zero-forcing detectors have a long history in communications, dating back to the equalization of inter-symbol interference in single-antenna channels. The extension to multi-user MIMO was developed in the early 2000s as part of the BLAST architecture at Bell Labs. The key realization for massive MIMO was that ZF — previously considered impractical due to the matrix inversion — becomes nearly optimal and computationally feasible when $N_t \gg K$ , because the Gram matrix $\hat{\mathbf{H}}^H \hat{\mathbf{H}} / N_t$ converges to a well-conditioned diagonal matrix.

MRC vs. ZF vs. MMSE Combining

Property	MRC	ZF	MMSE
Combining vector	$\hat{\mathbf{H}}_k$	$\hat{\mathbf{H}}(\hat{\mathbf{H}}^H \hat{\mathbf{H}})^{-1} \mathbf{e}_k$	$(\hat{\mathbf{H}}\hat{\mathbf{H}}^H + \alpha \mathbf{I})^{-1} \hat{\mathbf{H}}_k$
Array gain	$N_t$	$N_t - K$	$\approx N_t - K$ (large system)
Interference handling	None (interference-limited)	Fully eliminated (estimated part)	Optimally balanced
Denominator noise	$\sum_j {P_t}_{j} \beta_{j} + \sigma^2$	$\sum_j {P_t}_{j}(\beta_{j} - \gamma_j) + \sigma^2$	Implicitly via Stieltjes transform
Complexity per symbol	$O(N_t K)$	$O(N_tK + K^{3})$	$O(N_tK + K^{3})$
Best regime	$N_t \gg K$	$N_t > 2K$	All regimes

ZF and MMSE SINR vs. Number of Antennas

Compare ZF and MMSE achievable rates as a function of $N_t$ . Notice how ZF loses degrees of freedom (the curve starts at $N_t = K + 1$ ) while MMSE gracefully handles the transition.

Parameters

K

(users)10

\text{SNR}

[dB]10

\beta

[dB]-10

Example: When Does ZF Outperform MRC?

For equal power, equal path loss, and perfect estimation ( $\gamma_k = \beta$ ), find the condition on $N_t$ , $K$ , and $\text{SNR} = P_t \beta / \sigma^2$ under which $R_k^{\text{ZF}} > R_k^{\text{MRC}}$ .

Solution

Simplify with perfect estimation

With $\gamma_k = \beta$ for all users:

$R_k^{\text{MRC}} = \log_2\!\left(1 + \frac{N_t \text{SNR}}{K \text{SNR} + 1}\right), \quad R_k^{\text{ZF}} = \log_2\!\left(1 + (N_t - K) \text{SNR}\right).$

Find the crossover

$R_k^{\text{ZF}} > R_k^{\text{MRC}}$ iff

$(N_t - K) \text{SNR} > \frac{N_t \text{SNR}}{K \text{SNR} + 1}.$

Simplifying:

$(N_t - K)(K \text{SNR} + 1) > N_t,$

$N_t K \text{SNR} - K^{2} \text{SNR} + N_t - K > N_t,$

$K \text{SNR}(N_t - K) > K,$

$\text{SNR} > \frac{1}{N_t - K}.$

So ZF outperforms MRC whenever $\text{SNR} > 1/(N_t - K)$ , which is almost always satisfied in practice (e.g., $N_t = 64$ , $K = 10$ gives the threshold at $-17.3$ dB).

Sum Rate Comparison: MRC vs. ZF vs. MMSE

Compare the sum spectral efficiency of all three combining schemes as a function of the number of users $K$ for a fixed $N_t$ . MRC saturates at high $K$ while ZF and MMSE continue to grow.

Parameters

N_t

(antennas)128

\text{SNR}

[dB]10

\beta

[dB]-10

Common Mistake: Forgetting the Degrees-of-Freedom Loss in ZF

Mistake:

Writing the ZF SINR with array gain $N_t$ instead of $N_t - K$ . This overestimates the ZF rate, especially when $K$ is not negligible compared to $N_t$ .

Correction:

ZF projects onto the $(N_t - K + 1)$ -dimensional orthogonal complement of the other users' channels. The effective array gain is $N_t - K$ , not $N_t$ . Always use the correct formula:

$\text{SINR}_k^{\text{ZF}} = \frac{(N_t - K) \, {P_t}_{k} \, \gamma_k}{\sum_j {P_t}_{j}(\beta_{j} - \gamma_j) + \sigma^2}.$

Regularized Zero Forcing (RZF)

A linear combining/precoding scheme that adds a regularization (Tikhonov) term to the ZF pseudo-inverse: $\mathbf{v}_{k} = (\hat{\mathbf{H}}^H \hat{\mathbf{H}} + \alpha \mathbf{I})^{-1} \hat{\mathbf{H}}_k^H$ . The optimal regularization $\alpha$ equals $K \sigma^2 / P_t$ , recovering the MMSE combiner. Also called MMSE combining.

Related: Effective SINR

Inverse Wishart Distribution

If $\mathbf{X} \in \mathbb{C}^{n \times p}$ has i.i.d. $\mathcal{CN}(0,1)$ entries with $n > p$ , then $(\mathbf{X}^H \mathbf{X})^{-1}$ follows an inverse Wishart distribution. The diagonal entries have mean $1/(n - p)$ , which determines the ZF combining norm and hence the degrees-of-freedom loss.

Related: Regularized Zero Forcing (RZF)

Quick Check

With perfect channel estimation ( $\gamma_k = \beta_{k}$ for all $k$ ), what does the ZF denominator reduce to?

$\sum_j {P_t}_{j} \beta_{j} + \sigma^2$

$\sigma^2$ only

$\sum_j {P_t}_{j} \gamma_j + \sigma^2$

Zero — ZF achieves infinite rate with perfect CSI

Correction:

\sigma^2

only

With $\gamma_j = \beta_{j}$ , the estimation error variance $\beta_{j} - \gamma_j = 0$ for all users, so the denominator is just $\sigma^2$ . ZF with perfect CSI is noise-limited, not interference-limited.

Key Takeaway

ZF combining eliminates inter-user interference at the cost of $K$ degrees of freedom, yielding an effective array gain of $N_t - K$ . MMSE combining optimally balances interference suppression and noise enhancement. Both achieve strictly higher rates than MRC whenever $\text{SNR} > 1/(N_t - K)$ . The gap between ZF and MMSE is small when $N_t / K \geq 5$ .

Achievable Rates with ZF and MMSE

Beyond MRC: Interference Suppression

Definition: Zero-Forcing Combining Vector

Definition: MMSE (Regularized ZF) Combining Vector

Theorem: Closed-Form UatF Rate with ZF Combining

Step 1: ZF eliminates estimated-channel interference

Step 2: Compute the signal mean

Step 3: Compute the interference power

Step 4: Use the trace identity for ZF norm

Step 5: Assemble the SINR

Theorem: UatF Rate with MMSE Combining (Large-System Approximation)

Sketch: Random matrix theory approach

Historical Note: Random Matrix Theory Enters Wireless Communications

Historical Note: From Interference Cancellation to Zero Forcing

MRC vs. ZF vs. MMSE Combining

ZF and MMSE SINR vs. Number of Antennas

Parameters

Example: When Does ZF Outperform MRC?

Simplify with perfect estimation

Find the crossover

Sum Rate Comparison: MRC vs. ZF vs. MMSE

Parameters

Common Mistake: Forgetting the Degrees-of-Freedom Loss in ZF

Regularized Zero Forcing (RZF)

Inverse Wishart Distribution

Quick Check

Key Takeaway

Definition:
Zero-Forcing Combining Vector

Definition:
MMSE (Regularized ZF) Combining Vector