Ferkans — Interactive Telecom Tutor

Turning a Nonlinear Receiver into a Linear One

The quantizer $Q_b$ is nonlinear, so the observation $\mathbf{y}_q = Q_b(\mathbf{H}\mathbf{x} + \mathbf{w})$ has a non-Gaussian conditional distribution given $\mathbf{x}$ . Shannon capacity for a nonlinear channel has no closed form, and direct analysis of $\mathbf{y}_q$ forces us through a jungle of multinomial distributions. The Bussgang decomposition sidesteps the jungle by writing the quantizer output as a linear function of the input plus an uncorrelated distortion:

$\mathbf{y}_q = \mathbf{B}\mathbf{y} + \mathbf{d}, \qquad \mathbb{E}[\mathbf{d}\mathbf{y}^H] = \mathbf{0}.$

The matrix $\mathbf{B}$ absorbs the "correlated" part of the nonlinearity; $\mathbf{d}$ contains the rest. Downstream, we treat $\mathbf{d}$ like additive noise — it is not Gaussian and not independent of $\mathbf{y}$ in general, but its zero cross-correlation with $\mathbf{y}$ is enough to compute linear-receiver SINRs. The idea is due to Bussgang (1952) for scalar Gaussian inputs through memoryless nonlinearities and was introduced to MIMO quantization analysis by Mezghani and Nossek (2012). It is now the standard tool for 1-bit and low-resolution massive MIMO.

,

Theorem: Bussgang's Theorem (Scalar Gaussian Input)

Let $Y \sim \mathcal{N}(0, \sigma_Y^2)$ and let $g: \mathbb{R} \to \mathbb{R}$ be a memoryless (possibly nonlinear) function with $\mathbb{E}[g(Y)^2] < \infty$ . Write $Y_q = g(Y)$ . Then there exists a scalar $B_g \in \mathbb{R}$ such that $Y_q = B_g\, Y + d$ with $\mathbb{E}[d\, Y] = 0$ . Explicitly,

$B_g = \frac{\mathbb{E}[Y\, g(Y)]}{\sigma_Y^2} = \mathbb{E}\!\left[g'(Y)\right] \quad \text{(Stein's identity)}.$

For the 1-bit quantizer $g(y) = \operatorname{sign}(y)$ this gives $B_1 = \sqrt{\dfrac{2}{\pi\,\sigma_Y^2}}.$

We are projecting $Y_q$ onto the subspace of random variables linear in $Y$ . By the orthogonality principle, the residual is uncorrelated with $Y$ . Stein's identity expresses this projection as an expected derivative of the nonlinearity — for a hard limiter the derivative is a delta at zero and the resulting integral is $\sqrt{2/(\pi\sigma_Y^2)}$ .

Proof

Projection formula

The linear minimum-mean-square error (LMMSE) estimator of $Y_q$ given $Y$ is $\hat{Y}_q = \mathbb{E}[Y_q Y]/\mathbb{E}[Y^2] \cdot Y$ , so $B_g = \mathbb{E}[Y Y_q]/\sigma_Y^2$ . By construction the residual $d = Y_q - B_g Y$ satisfies $\mathbb{E}[dY]=0$ .

Stein's identity

For $Y$ Gaussian and any differentiable $g$ with $\mathbb{E}|g'(Y)| < \infty$ , $\mathbb{E}[Y g(Y)] = \sigma_Y^2\,\mathbb{E}[g'(Y)]$ (integration by parts against the Gaussian density). Thus $B_g = \mathbb{E}[g'(Y)]$ .

Evaluate for the sign function

$g(y) = \operatorname{sign}(y)$ , $g'(y) = 2\delta(y)$ . Then $\mathbb{E}[g'(Y)] = 2\,f_Y(0) = 2 \cdot \tfrac{1}{\sqrt{2\pi\sigma_Y^2}} = \sqrt{2/(\pi\sigma_Y^2)}$ . Substituting yields the stated $B_1$ . $\blacksquare$

,

Definition:
Bussgang Matrix and Distortion Covariance (MIMO)

Let $\mathbf{y} \in \mathbb{C}^{N_r}$ be zero-mean complex Gaussian with covariance $\boldsymbol{\Sigma}_{y} = \mathbb{E}[\mathbf{y} \mathbf{y}^H]$ and $\mathbf{y}_q = Q_b(\mathbf{y})$ the element-wise quantized vector. The Bussgang matrix is

$\mathbf{B} \triangleq \mathbb{E}[\mathbf{y}_q \mathbf{y}^H]\, \boldsymbol{\Sigma}_{y}^{-1}.$

By construction the residual $\mathbf{d} = \mathbf{y}_q - \mathbf{B} \mathbf{y}$ is uncorrelated with $\mathbf{y}$ , i.e. $\mathbb{E}[\mathbf{d}\mathbf{y}^H] = \mathbf{0}$ . Its covariance is $\boldsymbol{\Sigma}_{d} = \boldsymbol{\Sigma}_{y_q} - \mathbf{B}\,\boldsymbol{\Sigma}_{y} \mathbf{B}^H$ . For a 1-bit quantizer applied element-wise, the diagonal entries of $\mathbf{B}$ are

$B_{nn} = \sqrt{\tfrac{2}{\pi}}\, \frac{1}{\sqrt{[\boldsymbol{\Sigma}_{y}]_{nn}}},$

and the quantized covariance follows the arcsine law

$[\boldsymbol{\Sigma}_{y_q}]_{mn} = \tfrac{2}{\pi}\arcsin\!\left( \tfrac{[\boldsymbol{\Sigma}_{y}]_{mn}} {\sqrt{[\boldsymbol{\Sigma}_{y}]_{mm}\,[\boldsymbol{\Sigma}_{y}]_{nn}}}\right).$

The Bussgang matrix for the 1-bit quantizer depends only on the diagonal of $\boldsymbol{\Sigma}_{y}$ , so it is diagonal. The cross-antenna correlation is pushed into $\boldsymbol{\Sigma}_{d}$ , where the arcsine nonlinearity curbs the off-diagonals — a quantitative statement of "quantization de-correlates" used throughout the chapter.

,

Definition:
Bussgang Distortion Factor

For a uniform scalar quantizer with $b$ bits, optimized for a Gaussian input, define the Bussgang distortion factor

$\rho_b \triangleq 1 - \frac{\mathbb{E}[Y\,Q_b(Y)]^2} {\sigma_Y^2\,\mathbb{E}[Q_b(Y)^2]}.$

For small resolution $b$ a table (Max 1960, Lloyd-Max optimized quantizers, Gaussian source) gives the following values, which are ubiquitous in 1-bit-MIMO papers:

$b$	$\rho_b$	$\kappa_b = 1-\rho_b$ (effective SNR scaling)
1	$1 - 2/\pi \approx 0.3634$	$0.6366$
2	$0.1175$	$0.8825$
3	$0.03454$	$0.9655$
4	$0.009497$	$0.9905$
5	$0.002499$	$0.9975$

Every additional bit reduces $\rho_b$ by a factor of $\approx 4$ , matching the usual "6 dB/bit" SQNR heuristic.

The factor $\kappa_b = 1 - \rho_b$ is the effective SNR attenuation of the quantizer: a signal with pre-quantization SNR $\text{SNR}$ emerges with $\kappa_b\,\text{SNR}/(1 - \kappa_b\,\text{SNR}/(1+\text{SNR}))$ post-quantization. At low SNR this simplifies to $\kappa_b\,\text{SNR}$ , so quantization scales the SNR by a fixed factor.

,

Theorem: Bussgang SINR for Linear Combining

Consider the $K$ -user massive-MIMO uplink with $b$ -bit ADCs on every antenna, perfect CSI, and Gaussian inputs. Apply a linear combiner $\mathbf{v}_k$ to the quantized observation. Using the Bussgang decomposition, the achievable rate of user $k$ is

$R_k \,\geq\, \log_2\!\left(1 + \frac{\kappa_b\,|\mathbf{v}_k^H \mathbf{H}_{k}|^2\,P_k} {\kappa_b \sum_{j\neq k} |\mathbf{v}_k^H \mathbf{H}_{j}|^2 P_j + \mathbf{v}_k^H(\kappa_b \sigma^2\mathbf{I} + \boldsymbol{\Sigma}_{d}) \mathbf{v}_k}\right),$

where $\kappa_b = 1 - \rho_b$ is the effective-gain factor from Definition DBussgang Distortion Factor and $\boldsymbol{\Sigma}_{d}$ is the Bussgang distortion covariance.

The linear-combining analysis we already know from Chapter 9 carries over verbatim, with two modifications: (i) desired and interfering powers shrink by $\kappa_b$ , and (ii) an extra distortion term $\mathbf{v}_k^H \boldsymbol{\Sigma}_{d} \mathbf{v}_k$ joins the thermal noise. In the massive-MIMO limit $N_r \to \infty$ with i.i.d. users the distortion is diagonally dominant, so MRC and MMSE designs look almost identical to the infinite-precision case but with a constant penalty $\kappa_b$ in the SINR denominator.

Proof

Apply Bussgang

Write $\mathbf{y}_q = \mathbf{B}(\mathbf{H}\mathbf{x} + \mathbf{w}) + \mathbf{d}$ . After linear combining, $\mathbf{v}_k^H \mathbf{y}_q = \mathbf{v}_k^H \mathbf{B}\mathbf{H}\mathbf{x} + \mathbf{v}_k^H \mathbf{B}\mathbf{w} + \mathbf{v}_k^H \mathbf{d}$ . Absorb $\mathbf{B}$ into the effective channel $\tilde{\mathbf{H}} = \mathbf{B}\mathbf{H}$ .

Signal and interference powers

Because $\mathbf{B}$ is (approximately) diagonal with $B_{nn}\approx\sqrt{\kappa_b/[\boldsymbol{\Sigma}_{y}]_{nn}}$ , the effective signal power of user $k$ is $\kappa_b |\mathbf{v}_k^H\mathbf{H}_{k}|^2 P_k$ . Interferers scale the same way.

Noise and distortion

The effective noise is $\mathbf{v}_k^H \mathbf{B}\mathbf{w}$ , with covariance $\kappa_b\,\sigma^2\mathbf{I}$ . The distortion $\mathbf{v}_k^H\mathbf{d}$ is uncorrelated with signal and noise by construction, adding $\mathbf{v}_k^H\boldsymbol{\Sigma}_{d} \mathbf{v}_k$ .

Shannon lower bound

Treating the distortion as Gaussian worst-case noise gives the lower bound on the mutual information stated in the theorem, by the standard Gaussian-worst-case argument (cf. Book ITA Ch. 14). $\blacksquare$

,

Effective SNR After $b$ -Bit Quantization

Sweep the per-antenna SNR and plot the post-Bussgang effective SNR $\text{SNR}_{\text{eff}}(b) = \kappa_b\,\text{SNR}/(1 + \rho_b\,\text{SNR})$ for $b \in \{1, 2, 3, 4, \infty\}$ . Observe the asymptotic SNR ceiling for small $b$ at high SNR — 1 bit caps around $\sim 4$ dB, 4 bits essentially match infinite precision over any reasonable range.

Parameters

Minimum SNR (dB)-15

Maximum SNR (dB)30

Show

b=1

Show

b=2

Show

b=3

Show

b=4

Example: Bussgang SINR in a Two-User Massive Uplink

A 1-bit massive-MIMO uplink has $N_r = 128$ antennas and $K = 2$ users. Both channels are i.i.d. Rayleigh with $\|\mathbf{H}_{k}\|^2 \approx N_r$ and both users transmit at per-antenna SNR $\text{SNR} = -5$ dB ( $\approx 0.316$ ). The base station uses MRC. Estimate the SINR and rate of user 1.

Solution

Array-combined SNR and parameters

$N_r\,\text{SNR} \approx 128\cdot 0.316 = 40.4$ . $\kappa_1 = 2/\pi \approx 0.637$ , $\rho_1 = 1 - 2/\pi \approx 0.363$ .

Bussgang SINR

Using Theorem TBussgang SINR for Linear Combining with MRC and the favorable propagation approximation $\mathbf{v}_1^H \mathbf{H}_{2} \approx 0$ : $\text{SINR}_1 \approx \dfrac{\kappa_1\,N_r\,\text{SNR}} {\kappa_1\,N_r\,\text{SNR}_{2}/N_r + 1 + \rho_1\,N_r\,\text{SNR}}.$ With $\text{SNR}_{2} = \text{SNR}$ and treating the residual inter-user term as $\kappa_1 \text{SNR}$ (small): $\text{SINR}_1 \approx \dfrac{0.637\cdot 40.4}{1 + 0.363\cdot 40.4 + 0.637\cdot 0.316} \approx \dfrac{25.7}{1 + 14.7 + 0.2} \approx 1.62$ .

Rate

$R_1 \approx \log_2(1 + 1.62) \approx 1.39$ bits/use. The ideal MRC rate would be $\log_2(1 + 40.4) \approx 5.37$ bits/use, so the 1-bit receiver retains roughly $26\%$ — worse than the $63.7\%$ of the low-SNR approximation because the array gain has pushed the effective SNR into the quantization-noise-dominated regime. $\blacksquare$

Bussgang-LMMSE Detector for a 1-Bit Uplink

Complexity:

\mathcal{O}(N_r^{3} + KN_r^{2})

per coherence block

Input: quantized observation y_q, channel H, powers P_k,

noise variance sigma^2, distortion factor rho_1 = 1 - 2/pi

Output: estimates x_hat for all users k = 1..K

// Step 1 — Covariance of the unquantized observation

Sigma_y <- H diag(P) H^H + sigma^2 I

// Step 2 — Bussgang diagonal matrix (1-bit)

for n = 1 .. N_r:

B[n,n] <- sqrt(2 / pi / Sigma_y[n,n])

// Step 3 — Quantized covariance via arcsine law

R_y <- diag(1/sqrt(diag(Sigma_y))) Sigma_y diag(1/sqrt(diag(Sigma_y)))

Sigma_yq <- (2/pi) * arcsin(R_y) // element-wise

// Step 4 — Distortion covariance

Sigma_d <- Sigma_yq - B * Sigma_y * B^H

// Step 5 — Bussgang-LMMSE combining

V <- (B H)^H ( B Sigma_y B^H + Sigma_d )^(-1)

// Step 6 — Decode

x_hat <- V * y_q

return x_hat

Steps 1–4 can be precomputed once per coherence block, amortizing their cost over hundreds of data symbols. The element-wise arcsine in Step 3 is the Van Vleck identity; for real-valued $R_y$ it is a scalar operation on each off-diagonal entry.

,

Common Mistake: Distortion Is Uncorrelated — Not Independent

Mistake:

Because the Bussgang residual $\mathbf{d}$ is orthogonal to the input by construction, it is tempting to treat it as statistically independent Gaussian noise and plug it straight into mutual-information formulas.

Correction:

Orthogonality (zero cross-correlation) is not independence. $\mathbf{d}$ is a deterministic nonlinear function of $\mathbf{y}$ , so $\mathbf{d}$ and $\mathbf{y}$ share information through higher-order moments. What the Bussgang decomposition gives is (a) the linear part of the quantizer's input-output relationship exactly, and (b) a lower bound on the mutual information obtained by treating the distortion as the worst-case (Gaussian) noise with the correct covariance. The bound is tight at low SNR and becomes increasingly loose at high SNR — another reminder that 1-bit massive MIMO is a low-to-moderate SNR technology, not a high-SNR one.

Why This Matters: Bussgang Beyond Quantization

The Bussgang decomposition applies to any memoryless nonlinearity with finite second moment, not just ADC quantizers. It is the standard tool for analyzing nonlinear power amplifiers at the transmit side (Book CM Ch. 8), phase-noise distortion (Book FSP Ch. 9), and even nonlinear digital-predistortion residuals in modern base-station radios. The same linearization $\mathbf{y}_{\text{out}} = \mathbf{B} \mathbf{y}_{\text{in}} + \mathbf{d}$ will reappear in Chapter 20 when we study phase-shifter quantization in hybrid beamforming — mirror image of the 1-bit ADC on the transmit side.

Key Takeaway

Bussgang linearizes any memoryless nonlinearity at the cost of a distortion term that is only uncorrelated with the input. For a $b$ -bit quantizer the result is a rate expression that looks like the infinite-precision one with two edits: desired and interfering powers scale by $\kappa_b = 1 - \rho_b$ , and an extra diagonal-dominant distortion covariance $\boldsymbol{\Sigma}_{d}$ is added to the noise. For 1-bit, $\kappa_1 = 2/\pi \approx 0.637$ , recovering the $1.96$ dB low-SNR loss of the previous section.

Bussgang decomposition

For any zero-mean Gaussian input $\mathbf{y}$ passed through a memoryless nonlinearity $Q$ , the identity $Q(\mathbf{y}) = \mathbf{B}\mathbf{y} + \mathbf{d}$ with $\mathbb{E}[\mathbf{d}\mathbf{y}^H] = 0$ , where $\mathbf{B}$ is the Bussgang matrix and $\mathbf{d}$ the residual distortion. The workhorse for analyzing 1-bit and low-resolution MIMO receivers.

Quick Check

For a 1-bit uniform quantizer applied to a zero-mean Gaussian input, which value of the Bussgang distortion factor $\rho_1$ (i.e., the residual power fraction) is correct?

$\rho_1 = 0$

$\rho_1 = 1 - 2/\pi \approx 0.363$

$\rho_1 = 1/2$

$\rho_1 = 1 - 1/\sqrt{2\pi}$

Correction:

\rho_1 = 1 - 2/\pi \approx 0.363

The Bussgang factor for the sign quantizer of a unit-variance Gaussian is $\\kappa_1 = 2/\\pi$ , so the residual fraction is $\\rho_1 = 1 - 2/\\pi \\approx 0.363$ .

The Bussgang Decomposition

Turning a Nonlinear Receiver into a Linear One

Theorem: Bussgang's Theorem (Scalar Gaussian Input)

Projection formula

Stein's identity

Evaluate for the sign function

Definition: Bussgang Matrix and Distortion Covariance (MIMO)

Definition: Bussgang Distortion Factor

Theorem: Bussgang SINR for Linear Combining

Apply Bussgang

Signal and interference powers

Noise and distortion

Shannon lower bound

Effective SNR After bbb-Bit Quantization

Parameters

Example: Bussgang SINR in a Two-User Massive Uplink

Array-combined SNR and parameters

Bussgang SINR

Rate

Bussgang-LMMSE Detector for a 1-Bit Uplink

Common Mistake: Distortion Is Uncorrelated — Not Independent

Why This Matters: Bussgang Beyond Quantization

Key Takeaway

Bussgang decomposition

Quick Check

Definition:
Bussgang Matrix and Distortion Covariance (MIMO)

Definition:
Bussgang Distortion Factor

Effective SNR After $b$ -Bit Quantization