Ferkans — Interactive Telecom Tutor

Eliminating Interference Completely

MRC ignores interference and relies on favorable propagation to make it small. The zero-forcing (ZF) receiver takes the opposite approach: it designs the combining vectors to completely null out all inter-user interference. The price is noise enhancement — by forcing $\mathbf{g}_k^H \mathbf{h}_j = 0$ for $j \neq k$ , the combining vector may no longer point in the direction of maximum signal power.

The tradeoff between interference suppression and noise enhancement is governed by the condition number of the Gram matrix $\mathbf{H}^{H} \mathbf{H}$ . In the massive MIMO regime, the Gram matrix is well-conditioned, and the noise penalty of ZF vanishes — a satisfying convergence with the MRC result.

Definition:
Zero-Forcing (ZF) Receiver

The ZF receiver applies the combining matrix

$\mathbf{G}^{\text{ZF}} = \mathbf{H} (\mathbf{H}^{H} \mathbf{H})^{-1},$

so that $(\mathbf{G}^{\text{ZF}})^H \mathbf{H} = \mathbf{I}_{K}$ , i.e., the effective channel after combining is the identity. The soft estimate is

$\hat{\mathbf{x}}^{\text{ZF}} = (\mathbf{H}^{H} \mathbf{H})^{-1} \mathbf{H}^{H} \mathbf{y} = \mathbf{x} + (\mathbf{H}^{H} \mathbf{H})^{-1} \mathbf{H}^{H} \mathbf{w}.$

The interference is perfectly removed; the residual is shaped noise $\tilde{\mathbf{w}} = (\mathbf{H}^{H} \mathbf{H})^{-1} \mathbf{H}^{H} \mathbf{w}$ with covariance $\sigma^2 (\mathbf{H}^{H} \mathbf{H})^{-1}$ .

$\mathbf{G}^{\text{ZF}}$ is the Moore–Penrose pseudoinverse $\mathbf{H}^\dagger$ of the tall matrix $\mathbf{H}$ (since $N_t \geq K$ and we assume $\mathbf{H}$ has full column rank).

Theorem: ZF SINR Expression

With the ZF receiver and equal per-user power $P$ , the post-detection SINR for user $k$ is

$\text{SINR}_k^{\text{ZF}} = \frac{P}{\sigma^2 [(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk}}.$

Equivalently, defining $\text{SNR} = P / \sigma^2$ :

$\text{SINR}_k^{\text{ZF}} = \frac{\text{SNR}}{[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk}}.$

There is no interference term — ZF eliminates it by design. The denominator involves the $k$ -th diagonal entry of $(\mathbf{H}^{H} \mathbf{H})^{-1}$ , which quantifies the noise enhancement for user $k$ . A large value of $[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk}$ means the ZF combining vector for user $k$ is long (high norm), amplifying noise.

Proof

Write the ZF output

$\hat{x}_k^{\text{ZF}} = [\hat{\mathbf{x}}^{\text{ZF}}]_k = x_k + [(\mathbf{H}^{H} \mathbf{H})^{-1} \mathbf{H}^{H} \mathbf{w}]_k$ . The signal power is $P$ and there is zero interference.

Compute the noise variance

The noise term has variance $\mathbb{E}[|[(\mathbf{H}^{H} \mathbf{H})^{-1} \mathbf{H}^{H} \mathbf{w}]_k|^2] = \sigma^2 [(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk}$ , because $\mathbf{w} \sim \mathcal{CN}(\mathbf{0}, \sigma^2 \mathbf{I})$ and the noise covariance of the ZF output is $\sigma^2 (\mathbf{H}^{H} \mathbf{H})^{-1}$ .

Form the SINR

$\text{SINR}_k^{\text{ZF}} = P / (\sigma^2 [(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk})$ . $\blacksquare$

Definition:
Condition Number and Noise Enhancement

The condition number of the Gram matrix is

$\kappa(\mathbf{H}^{H} \mathbf{H}) = \frac{\lambda_{\max}(\mathbf{H}^{H} \mathbf{H})}{\lambda_{\min}(\mathbf{H}^{H} \mathbf{H})},$

where $\lambda_{\max}$ and $\lambda_{\min}$ are the largest and smallest eigenvalues. The noise enhancement factor satisfies

$[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk} \leq \frac{1}{\lambda_{\min}(\mathbf{H}^{H} \mathbf{H})},$

so a well-conditioned Gram matrix (small $\kappa$ ) ensures bounded noise enhancement for all users.

Theorem: ZF in the Massive MIMO Limit

Under i.i.d. Rayleigh fading with $N_t \to \infty$ and $K$ fixed, the ZF SINR for user $k$ satisfies

$\text{SINR}_k^{\text{ZF}} \xrightarrow{\text{a.s.}} P (N_t - K) \beta_k / \sigma^2.$

In the massive MIMO regime $N_t \gg K$ , this approaches the MRC asymptotic SINR $P N_t \beta_k / \sigma^2$ .

The factor $(N_t - K)$ rather than $N_t$ reflects the "degrees of freedom lost" by the ZF constraint: $K - 1$ dimensions are used to null interference, leaving $N_t - K + 1$ effective dimensions for signal collection. When $N_t \gg K$ , this cost is negligible.

Proof

Apply the Marchenko–Pastur result

For i.i.d. Rayleigh $\mathbf{H}$ with equal path loss $\beta$ , the Wishart matrix $\mathbf{H}^{H} \mathbf{H}$ has diagonal entries converging to $N_t \beta$ and the inverse Wishart distribution gives $\mathbb{E}[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk} = 1 / ((N_t - K) \beta_k)$ for $N_t > K$ .

Substitute into the SINR

$\text{SINR}_k^{\text{ZF}} \to \frac{P}{\sigma^2 / ((N_t - K) \beta_k)} = \frac{P (N_t - K) \beta_k}{\sigma^2}.$ $As$ N_t \to \infty $with fixed$ K $,$ N_t - K \approx N_t $.$ \blacksquare$

Common Mistake: ZF Fails When the Channel is Rank-Deficient

Mistake:

Applying ZF detection when $\mathbf{H}$ does not have full column rank (e.g., more users than antennas, or highly correlated channels) causes $(\mathbf{H}^{H} \mathbf{H})^{-1}$ to blow up or become undefined.

Correction:

ZF requires $\text{rank}(\mathbf{H}) = K$ , which holds almost surely for $N_t \geq K$ with i.i.d. fading. For correlated channels with colinear users, the Gram matrix can be near-singular and the MMSE (regularized ZF) receiver should be used instead.

Example: Noise Enhancement in a 2-User ZF System

Consider $N_t = 4$ , $K = 2$ with channel vectors $\mathbf{h}_1 = [1, 1, 0, 0]^T$ and $\mathbf{h}_2 = [1, 0, 1, 0]^T$ . Compute the noise enhancement factor $[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{11}$ and compare with MRC.

Solution

Compute the Gram matrix

$\mathbf{H}^{H} \mathbf{H} = \begin{bmatrix} \|\mathbf{h}_1\|^2 & \mathbf{h}_1^H \mathbf{h}_2 \\ \mathbf{h}_2^H \mathbf{h}_1 & \|\mathbf{h}_2\|^2 \end{bmatrix} = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix}.$

Invert the Gram matrix

$(\mathbf{H}^{H} \mathbf{H})^{-1} = \frac{1}{4 - 1}\begin{bmatrix} 2 & -1 \\ -1 & 2 \end{bmatrix} = \begin{bmatrix} 2/3 & -1/3 \\ -1/3 & 2/3 \end{bmatrix}.$

Read off the noise enhancement

$[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{11} = 2/3$ . The ZF noise variance for user 1 is $\sigma^2 \times 2/3$ , which is lower than the MRC noise variance $\sigma^2 \times \|\mathbf{h}_1\|^2 = 2\sigma^2$ (though MRC also has the interference term). The condition number $\kappa = 3/1 = 3$ , indicating mild ill-conditioning.

ZF SINR vs. $N_t$ : Noise Enhancement Effect

Compare the ZF and MRC SINR as a function of $N_t$ . Notice that ZF starts below MRC at small $N_t/K$ ratios (due to noise enhancement) but converges to the same asymptote for large $N_t$ . The gap $K - 1$ "lost degrees of freedom" is visible in the finite-antenna regime.

Parameters

K

(users)8

P/\sigma^2

(dB)10

Zero-Forcing (ZF) Receiver

A linear detector that uses the pseudoinverse $\mathbf{H}^\dagger = \mathbf{H}(\mathbf{H}^{H} \mathbf{H})^{-1}$ as the combining matrix, completely eliminating multi-user interference at the cost of noise enhancement.

Noise Enhancement

The increase in effective noise power that results from ZF combining. Quantified by $[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{kk}$ , which can be large when the channel vectors are nearly colinear.

⚠️Engineering Note

Computational Cost of ZF Detection

ZF detection requires inverting the $K \times K$ Gram matrix $\mathbf{H}^{H} \mathbf{H}$ , costing $\mathcal{O}(K^{3})$ via Cholesky decomposition. For $K = 16$ this is manageable, but for large user counts the cubic cost becomes a bottleneck.

In practice, the Gram matrix is computed once per coherence interval and reused for all OFDM symbols within that interval. The per-symbol cost is $\mathcal{O}(N_t K)$ (matrix-vector multiply), same as MRC. The dominant cost is the one-time matrix inversion at the start of each coherence block.

Practical Constraints

•
Cholesky decomposition: $K^{3}/3$ FLOPs per coherence block
•
Must be recomputed when the channel changes (every $T_c$ seconds)
•
For $K > 32$ , iterative approximations (Section 9.6) become attractive

Quick Check

In the massive MIMO limit, the ZF SINR for user $k$ is $P(N_t - K)\beta_k / \sigma^2$ . What does the factor $(N_t - K)$ represent?

The number of users being served

The effective degrees of freedom after nulling interference

The rank of the channel matrix

The number of pilot symbols needed

Correction:

The effective degrees of freedom after nulling interference

ZF uses $K - 1$ spatial dimensions to null the $K - 1$ interferers for each user, leaving $N_t - K + 1 \approx N_t - K$ dimensions for signal collection. This is the array gain after subtracting the interference-nulling cost.

Zero-Forcing (ZF) Receiver

Eliminating Interference Completely

Definition: Zero-Forcing (ZF) Receiver

Theorem: ZF SINR Expression

Write the ZF output

Compute the noise variance

Form the SINR

Definition: Condition Number and Noise Enhancement

Theorem: ZF in the Massive MIMO Limit

Apply the Marchenko–Pastur result

Substitute into the SINR

Common Mistake: ZF Fails When the Channel is Rank-Deficient

Example: Noise Enhancement in a 2-User ZF System

Compute the Gram matrix

Invert the Gram matrix

Read off the noise enhancement

ZF SINR vs. NtN_tNt​: Noise Enhancement Effect

Parameters

Zero-Forcing (ZF) Receiver

Noise Enhancement

Computational Cost of ZF Detection

Quick Check

Definition:
Zero-Forcing (ZF) Receiver

Definition:
Condition Number and Noise Enhancement

ZF SINR vs. $N_t$ : Noise Enhancement Effect