Ferkans — Interactive Telecom Tutor

The Best of Both Worlds

MRC maximizes signal power but ignores interference. ZF eliminates interference but amplifies noise. The MMSE receiver — also called regularized ZF — strikes the optimal balance: it minimizes the total mean squared error $\mathbb{E}[\|\hat{\mathbf{x}} - \mathbf{x}\|^2]$ , jointly accounting for both interference and noise.

The MMSE filter reduces to MRC at low SNR (where noise dominates) and to ZF at high SNR (where interference dominates), smoothly interpolating between the two extremes. This makes it the uniformly best linear receiver across all operating regimes.

Definition:
MMSE (Regularized ZF) Receiver

The MMSE receiver for the uplink model $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{w}$ with $\mathbb{E}[\mathbf{x}\mathbf{x}^H] = P \mathbf{I}$ is

$\mathbf{G}^{\text{MMSE}} = \left(\mathbf{H} \mathbf{H}^{H} + \frac{\sigma^2}{P} \mathbf{I}_{N_t}\right)^{-1} \mathbf{H},$

or equivalently, using the matrix inversion lemma:

$\mathbf{G}^{\text{MMSE}} = \mathbf{H} \left(\mathbf{H}^{H} \mathbf{H} + \frac{\sigma^2}{P} \mathbf{I}_{K}\right)^{-1}.$

The soft estimate is $\hat{\mathbf{x}}^{\text{MMSE}} = (\mathbf{G}^{\text{MMSE}})^H \mathbf{y}$ .

The second form is computationally preferred when $K < N_t$ : it inverts a $K \times K$ matrix rather than an $N_t \times N_t$ one. The regularization term $\frac{\sigma^2}{P} \mathbf{I}$ is precisely what distinguishes MMSE from ZF — it prevents the noise enhancement that arises from inverting a near-singular Gram matrix.

Theorem: MMSE Is the Optimal Linear Receiver

Among all linear receivers $\hat{\mathbf{x}} = \mathbf{A}^H \mathbf{y}$ , the MMSE receiver minimizes the mean squared error:

$\mathbf{G}^{\text{MMSE}} = \arg\min_{\mathbf{A}} \mathbb{E}\left[\|\mathbf{A}^H \mathbf{y} - \mathbf{x}\|^2\right].$

The minimum MSE for user $k$ is

$\text{MMSE}_k = P - P^2 \mathbf{h}_k^H \left(\mathbf{H} \mathbf{P} \mathbf{H}^{H} + \sigma^2 \mathbf{I}\right)^{-1} \mathbf{h}_k,$

where $\mathbf{P} = \text{diag}(P_1, \ldots, P_{K})$ .

The MMSE receiver is the linear MMSE (LMMSE) estimator from estimation theory applied to the linear model $\mathbf{y} = \mathbf{H}\mathbf{x} + \mathbf{w}$ . The connection to FSI Ch. 12 is direct: $\mathbf{H}$ plays the role of the observation matrix, $\mathbf{x}$ is the unknown parameter vector, and the LMMSE solution uses the prior covariance $\mathbb{E}[\mathbf{x}\mathbf{x}^H] = P\mathbf{I}$ .

Proof

Set up the MSE minimization

The MSE is $J(\mathbf{A}) = \mathbb{E}[\|\mathbf{A}^H \mathbf{y} - \mathbf{x}\|^2] = \text{tr}(\mathbf{A}^H \mathbf{R}_y \mathbf{A} - \mathbf{A}^H \mathbf{R}_{yx} - \mathbf{R}_{xy} \mathbf{A} + P\mathbf{I})$ where $\mathbf{R}_y = \mathbb{E}[\mathbf{y}\mathbf{y}^H] = P \mathbf{H} \mathbf{H}^{H} + \sigma^2\mathbf{I}$ and $\mathbf{R}_{yx} = \mathbb{E}[\mathbf{y}\mathbf{x}^H] = P \mathbf{H}$ .

Differentiate and set to zero

Setting $\frac{\partial J}{\partial \mathbf{A}^*} = \mathbf{R}_y \mathbf{A} - \mathbf{R}_{yx} = \mathbf{0}$ , we obtain $\mathbf{A} = \mathbf{R}_y^{-1} \mathbf{R}_{yx} = (P\mathbf{H}\mathbf{H}^{H} + \sigma^2\mathbf{I})^{-1} P\mathbf{H}$ .

Simplify

Factoring out $P$ and absorbing it, we get $\mathbf{G}^{\text{MMSE}} = (P\mathbf{H}\mathbf{H}^{H} + \sigma^2\mathbf{I})^{-1} \mathbf{H} \cdot P / P = (\mathbf{H}\mathbf{H}^{H} + \frac{\sigma^2}{P}\mathbf{I})^{-1} \mathbf{H}$ .

The minimum MSE is obtained by substituting back. $\blacksquare$

Theorem: MMSE SINR Expression

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

$\text{SINR}_k^{\text{MMSE}} = P \mathbf{h}_k^H \left(\sum_{j \neq k} P \mathbf{h}_j \mathbf{h}_j^H + \sigma^2 \mathbf{I}\right)^{-1} \mathbf{h}_k.$

Equivalently,

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

The MMSE SINR has a beautiful structure: the interference-plus-noise covariance $\sum_{j \neq k} P \mathbf{h}_j \mathbf{h}_j^H + \sigma^2\mathbf{I}$ is inverted and then the signal is projected through it. This is precisely the Capon beamformer applied to the detection problem — it steers a "spatial null" toward the interference while collecting the signal.

Proof

Use the SINR identity for LMMSE estimators

From estimation theory (FSI Ch. 12), for the linear model $\mathbf{y} = \mathbf{h}_k x_k + \mathbf{n}_k^{\text{eff}}$ where $\mathbf{n}_k^{\text{eff}} = \sum_{j \neq k} \mathbf{h}_j x_j + \mathbf{w}$ , the LMMSE SINR equals the quadratic form $P \mathbf{h}_k^H \mathbf{R}_{n_k}^{-1} \mathbf{h}_k$ where $\mathbf{R}_{n_k} = \sum_{j \neq k} P \mathbf{h}_j \mathbf{h}_j^H + \sigma^2\mathbf{I}$ .

Derive the alternative form

By the matrix inversion lemma applied to the rank-one update $\mathbf{R}_y = \mathbf{R}_{n_k} + P \mathbf{h}_k \mathbf{h}_k^H$ , one can show that $\text{SINR}_k = 1/[(\mathbf{H}^{H} \mathbf{H} + \frac{\sigma^2}{P}\mathbf{I})^{-1}]_{kk} - 1$ , which is a convenient computational form. $\blacksquare$

Key Takeaway

MMSE interpolates between MRC and ZF. At low SNR ( $P/\sigma^2 \to 0$ ), the regularization term dominates and MMSE reduces to MRC (matched filter). At high SNR ( $P/\sigma^2 \to \infty$ ), the regularization vanishes and MMSE reduces to ZF. The MMSE receiver is never worse than either.

Example: MMSE vs. ZF for Two Correlated Users

Consider $N_t = 4$ , $K = 2$ with channel vectors $\mathbf{h}_1 = [1, 1, 0, 0]^T$ and $\mathbf{h}_2 = [1, 0.9, 0.1, 0]^T$ (highly correlated). Compare the SINR of ZF and MMSE at $\text{SNR} = P/\sigma^2 = 10$ dB.

Solution

Compute the Gram matrix and its regularized inverse

$\mathbf{H}^{H} \mathbf{H} = \begin{bmatrix} 2 & 1.9 \\ 1.9 & 1.82 \end{bmatrix}$ , $\det(\mathbf{H}^{H} \mathbf{H}) = 2 \times 1.82 - 1.9^2 = 0.03$ .

ZF: $[(\mathbf{H}^{H} \mathbf{H})^{-1}]_{11} = 1.82/0.03 = 60.67$ . $\text{SINR}_1^{\text{ZF}} = 10/60.67 = 0.165$ ( $-7.8$ dB).

MMSE with regularization

$\mathbf{H}^{H} \mathbf{H} + 0.1 \mathbf{I} = \begin{bmatrix} 2.1 & 1.9 \\ 1.9 & 1.92 \end{bmatrix}$ , $\det = 2.1 \times 1.92 - 1.9^2 = 0.422$ .

$[(\mathbf{H}^{H} \mathbf{H} + 0.1\mathbf{I})^{-1}]_{11} = 1.92/0.422 = 4.55$ . $\text{SINR}_1^{\text{MMSE}} = 1/4.55 - 1 = -0.78$ ... Let us recompute using the direct formula:

$\text{SINR}_1^{\text{MMSE}} = P \mathbf{h}_1^H (P \mathbf{h}_2 \mathbf{h}_2^H + \sigma^2\mathbf{I})^{-1} \mathbf{h}_1$ .

With $P = 10\sigma^2$ : the interference covariance is $10\sigma^2 \mathbf{h}_2 \mathbf{h}_2^H + \sigma^2\mathbf{I}$ and the MMSE SINR evaluates to approximately $2.1$ dB — a 10 dB improvement over ZF for these correlated channels.

Interpretation

The correlated channels ( $\mathbf{h}_1^H \mathbf{h}_2 / (\|\mathbf{h}_1\|\|\mathbf{h}_2\|) = 0.998$ ) cause catastrophic noise enhancement in ZF. MMSE accepts some residual interference to avoid this noise explosion.

SINR Comparison: MRC vs. ZF vs. MMSE

Compare the average per-user SINR of MRC, ZF, and MMSE receivers as a function of $N_t$ . Observe that MMSE always dominates, ZF suffers at low $N_t/K$ ratios, and all three converge in the massive MIMO regime.

Parameters

K

(users)8

P/\sigma^2

(dB)10

Comparison of Linear Receivers

Property	MRC	ZF	MMSE
Combining matrix $\mathbf{G}$	$\mathbf{H}$	$\mathbf{H}(\mathbf{H}^{H}\mathbf{H})^{-1}$	$\mathbf{H}(\mathbf{H}^{H}\mathbf{H} + \frac{\sigma^2}{P}\mathbf{I})^{-1}$
Interference handling	Ignores	Nulls completely	Balances suppression and noise
Noise enhancement	None	Can be severe	Bounded (regularized)
Per-symbol complexity	$\mathcal{O}(N_tK)$	$\mathcal{O}(N_tK)$	$\mathcal{O}(N_tK)$
One-time complexity	None	$\mathcal{O}(K^{3})$	$\mathcal{O}(K^{3})$
Optimal at	Low SNR, $N_t \gg K$	High SNR, well-conditioned $\mathbf{H}$	All regimes
Massive MIMO SINR	$N_t P \beta_k / \sigma^2$	$(N_t - K) P \beta_k / \sigma^2$	$N_t P \beta_k / \sigma^2$

Common Mistake: MMSE Is Not Just 'ZF with Diagonal Loading'

Mistake:

Students sometimes view MMSE as merely adding a small constant to the Gram matrix diagonal to fix numerical issues. This trivializes the fundamental statistical optimality of MMSE.

Correction:

The regularization $\sigma^2/P \cdot \mathbf{I}$ is not arbitrary — it is the exact ratio of noise power to signal power dictated by the Bayesian LMMSE estimator. Changing this ratio degrades performance. MMSE is the unique linear receiver that minimizes the MSE, and its regularization strength adapts to the SNR.

MMSE Receiver

The linear minimum mean squared error receiver, which minimizes $\mathbb{E}[\|\hat{\mathbf{x}} - \mathbf{x}\|^2]$ . Equivalent to the LMMSE estimator from Bayesian estimation theory and to the regularized ZF (or Wiener) filter.

ZF vs MMSE: Noise Enhancement Tradeoff

Constellation diagrams for ZF and MMSE receivers under well-conditioned and ill-conditioned channels. When the channel is well-conditioned, both receivers produce tight constellations. When the condition number is large, ZF noise enhancement scatters the constellation widely (variance

\propto (\mathbf{H}^{H}\mathbf{H})^{-1}

), while MMSE trades a small bias toward the origin for dramatically reduced noise variance.

MMSE (Regularized ZF) Receiver