Ferkans — Interactive Telecom Tutor

The Fronthaul Bottleneck

Every cooperation level requires some information exchange between the APs and the CPU. Even Level 1 must forward $|\mathcal{D}_m|$ complex scalars per channel use. In practice, these complex numbers must be quantized to a finite number of bits for digital transmission over the fronthaul link. The quantization introduces distortion that degrades the effective SINR. The system designer faces a tradeoff: more quantization bits reduce distortion but increase the fronthaul rate requirement. This section develops the theory of fronthaul-aware distributed processing: how many bits are needed, how to allocate them across APs and users, and what happens when the fronthaul capacity is the binding constraint.

Definition:
Scalar Quantization of Local Estimates

In distributed processing (Levels 1–3), AP $m$ computes local estimates $\hat{s}_{mk} \in \mathbb{C}$ for each $k \in \mathcal{D}_m$ and quantizes them before fronthaul transmission. Using a $b$ -bit uniform scalar quantizer on the real and imaginary parts independently, the quantized estimate is

$\hat{s}_{mk}^{(q)} = \hat{s}_{mk} + e_{mk}$

where $e_{mk}$ is the quantization error with variance

$\sigma^2_{q} = \frac{\Delta^2}{12} = \frac{(\hat{s}_{\max} - \hat{s}_{\min})^2}{12 \cdot 2^{2b}}$

per real dimension, where $\Delta = (\hat{s}_{\max} - \hat{s}_{\min}) / 2^b$ is the quantization step size. The fronthaul rate per AP is

$R_{\text{fh},m} = 2b |\mathcal{D}_m| f_s \quad [\text{bit/s}]$

where $f_s$ is the sampling rate (equal to the system bandwidth for Nyquist sampling).

The factor of 2 accounts for quantizing both real and imaginary parts. With $b = 8$ bits, $|\mathcal{D}_m| = 10$ users, and $f_s = 20$ MHz: $R_{\text{fh},m} = 2 \times 8 \times 10 \times 20 \times 10^6 = 3.2$ Gbit/s per AP. This is within the capacity of a single 10G Ethernet fronthaul link.

Definition:
Fronthaul Capacity Constraint

The fronthaul link between AP $m$ and the CPU has a finite capacity $C_{\text{fh}}$ (in bit/s). The quantization resolution is constrained by

$2 b_m |\mathcal{D}_m| f_s \leq C_{\text{fh}}$

which gives the maximum number of quantization bits per AP:

$b_m \leq \frac{C_{\text{fh}}}{2 |\mathcal{D}_m| f_s}$

For Level 4 (centralized processing), the fronthaul must carry the full received signal $\mathbf{y}_m \in \mathbb{C}^N$ :

$2 b_m N f_s \leq C_{\text{fh}} \implies b_m \leq \frac{C_{\text{fh}}}{2 N f_s}$

Since typically $N > |\mathcal{D}_m|$ for multi-antenna APs, Level 4 requires higher fronthaul capacity — or fewer quantization bits — than Levels 1–3.

Theorem: SINR Degradation Under Fronthaul Quantization

Under $b$ -bit uniform scalar quantization of the local estimates, the effective SINR of user $k$ with LSFD (Level 3) is

$\text{SINR}_k^{(3,q)} = \frac{p_k |\boldsymbol{\alpha}_k^H \mathbf{b}_k|^2}{\boldsymbol{\alpha}_k^H \left( \mathbf{E}_k + \sigma^2_{q} \, \text{diag}(\boldsymbol{\alpha}_k \circ \boldsymbol{\alpha}_k^*) \right) \boldsymbol{\alpha}_k}$

where $\sigma^2_{q}$ is the per-component quantization noise variance. In the high-resolution regime ( $b \geq 5$ ), the SINR loss relative to unquantized LSFD is approximately

$\Delta \text{SINR}_k \approx \frac{\sigma^2_{q} \|\boldsymbol{\alpha}_k\|^2}{\boldsymbol{\alpha}_k^H \mathbf{E}_k \boldsymbol{\alpha}_k} \approx \frac{c}{2^{2b}}$

for a constant $c$ that depends on the signal dynamic range.

Quantization adds independent noise at each AP's fronthaul output. This noise is equivalent to increasing the thermal noise variance by $\sigma^2_{q}$ per AP. With $b = 5$ bits, $2^{2b} = 1024$ , so the quantization noise is roughly 30 dB below the signal — negligible in most scenarios. With $b = 3$ bits, the noise is only 18 dB below, which can degrade the SINR by 1–2 dB.

Proof

Additive quantization noise model

The quantized local estimate is $\hat{s}_{mk}^{(q)} = \hat{s}_{mk} + e_{mk}$ where $e_{mk}$ is independent of $\hat{s}_{mk}$ (additive quantization noise model, valid for high-resolution quantization). The variance of $e_{mk}$ is $\sigma^2_{q}$ per real dimension.

Effect on the CPU estimate

The CPU forms $\hat{s}_k = \boldsymbol{\alpha}_k^H (\hat{\mathbf{s}}_k + \mathbf{e}_k)$ where $\mathbf{e}_k = [e_{1k}, \ldots, e_{Mk}]^T$ . The quantization noise power is $\mathbb{E}[|\boldsymbol{\alpha}_k^H \mathbf{e}_k|^2] = \sigma^2_{q} \sum_m |\alpha_{mk}|^2 = \sigma^2_{q} \|\boldsymbol{\alpha}_k\|^2$ .

Modified SINR

Adding the quantization noise to the denominator of the LSFD SINR gives the stated result. The SINR loss $\Delta \text{SINR}_k$ follows from a first-order expansion assuming $\sigma^2_{q} \ll$ noise floor. $\blacksquare$

Fronthaul Rate vs. Quantization Distortion

Explore the tradeoff between fronthaul rate (determined by quantization bits $b$ ) and the resulting SINR degradation. Observe that 4–6 bits per dimension are sufficient for near-lossless performance in most scenarios.

Parameters

M

(APs)100

N

(antennas/AP)4

K

(users)20

\text{SNR}

(dB)10

Example: Required Quantization Bits for 1 dB SINR Loss

A cell-free network operates at an average unquantized SINR of 15 dB. The dynamic range of the local estimates is $\hat{s}_{\max} - \hat{s}_{\min} = 4\sigma_s$ where $\sigma_s^2 = 0.1$ is the variance of the local estimates. How many quantization bits $b$ are needed to limit the SINR degradation to at most 1 dB?

Solution

Quantization noise variance

With $b$ bits and dynamic range $4\sigma_s$ : $\sigma^2_{q} = \frac{(4\sigma_s)^2}{12 \cdot 2^{2b}} = \frac{16 \sigma_s^2}{12 \cdot 4^b} = \frac{4\sigma_s^2}{3 \cdot 4^b}$

SINR degradation condition

The SINR loss is approximately $\Delta \text{SINR} \approx \sigma^2_{q} / \sigma^2_{\text{eff}}$ where $\sigma^2_{\text{eff}}$ is the effective noise floor. For 1 dB loss: $10 \log_{10}(1 + \Delta) \leq 1$ dB, so $\Delta \leq 10^{0.1} - 1 \approx 0.259$ .

At 15 dB SINR, the effective noise is $\sigma^2_{\text{eff}} = \sigma_s^2 / 10^{1.5} \approx 0.00316$ . Condition: $\sigma^2_{q} \leq 0.259 \times 0.00316 = 8.18 \times 10^{-4}$ .

Solve for $b$

$\frac{4 \times 0.1}{3 \times 4^b} \leq 8.18 \times 10^{-4}$

$4^b \geq \frac{0.4}{3 \times 8.18 \times 10^{-4}} = \frac{0.4}{0.00245} \approx 163$

$b \geq \frac{\log(163)}{2 \log(2)} \approx \frac{5.09}{1.386} \approx 3.67$

So $b = 4$ bits suffice. The fronthaul rate per AP with $|\mathcal{D}_m| = 10$ users and $f_s = 20$ MHz is $2 \times 4 \times 10 \times 20 = 1600$ Mbit/s = 1.6 Gbit/s.

Definition:
Vector Quantization for Fronthaul Compression

Instead of quantizing each local estimate independently (scalar quantization), vector quantization jointly quantizes the vector $\hat{\mathbf{s}}_m = [\hat{s}_{m1}, \ldots, \hat{s}_{m|\mathcal{D}_m|}]^T$ of all local estimates at AP $m$ . By exploiting the correlation structure (nearby users have correlated channels), vector quantization achieves the same distortion as scalar quantization with fewer bits.

The rate-distortion bound for Gaussian sources gives the minimum fronthaul rate:

$R_{\text{fh},m} \geq \sum_{i=1}^{|\mathcal{D}_m|} \max\left(0, \frac{1}{2} \log_2 \frac{\lambda_i}{D}\right) \quad [\text{bit/sample}]$

where $\lambda_1 \geq \cdots \geq \lambda_{|\mathcal{D}_m|}$ are the eigenvalues of the local estimate covariance matrix and $D$ is the target distortion level.

Vector quantization provides diminishing returns when the local estimates are nearly uncorrelated (well-separated users). The largest gain occurs when $|\mathcal{D}_m|$ is large and users share similar channel subspaces.

Scalar vs. Vector Quantization for Fronthaul

Aspect	Scalar Quantization	Vector Quantization
Rate (bit/sample)	$2b \|\mathcal{D}_m\|$	$\sum_i \frac{1}{2} \log_2(\lambda_i / D)$ (rate-distortion bound)
Distortion	$\Delta^2 / 12$ per component	Optimally distributed across eigenvalues (water-filling)
Complexity	$O(\|\mathcal{D}_m\|)$	$O(\|\mathcal{D}_m\|^2)$ (covariance estimation + KLT)
Rate saving	Baseline	20–40% for correlated users
Implementation	Simple: standard ADC	Requires eigendecomposition of local estimate covariance

Optimal Bit Allocation Across APs

Complexity:

O(M b_{\max})

where

b_{\max}

is the maximum allowed bits

Input: Fronthaul budget

C_{\text{fh}}

per AP, user set

\mathcal{D}_m

,

large-scale coefficients

\{\beta_{mk}\}

, LSFD weights

\{\alpha_{mk}\}

Goal: Allocate quantization bits

\{b_m\}_{m=1}^M

to maximize the sum SINR

subject to per-AP fronthaul constraint.

1. Initialize:

b_m = \lfloor C_{\text{fh}} / (2 |\mathcal{D}_m| f_s) \rfloor

for all

m

2. Compute the marginal SINR gain from adding one bit at each AP:

\Delta_m = \text{SINR}_k(b_m + 1) - \text{SINR}_k(b_m)

3. Greedy allocation: While fronthaul budget allows:

a. Find

m^{\star} = \arg\max_m \Delta_m

(AP with largest marginal gain)

b. If

2(b_{m^{\star}} + 1) |\mathcal{D}_{m^{\star}}| f_s \leq C_{\text{fh}}

:

set

b_{m^{\star}} \leftarrow b_{m^{\star}} + 1

, update

\Delta_{m^{\star}}

c. Else: remove

m^{\star}

from candidates

4. Return

\{b_m\}

The greedy algorithm gives a near-optimal solution because the SINR gain from each additional bit is a concave function of $b_m$ (diminishing returns). APs with strong LSFD weights $|\alpha_{mk}|$ receive more bits because their quantization noise has a larger impact on the final estimate.

🚨Critical Engineering Note

Fronthaul Technologies and Their Capacities

The choice of fronthaul technology determines the achievable cooperation level:

Dark fiber: 10–100 Gbps per link. Supports Level 4 with high-resolution quantization. Cost: high (fiber deployment), power: 0.5–2 W per link.
Ethernet (eCPRI): 10–25 Gbps per link. Supports Level 4 with moderate quantization or Level 3 with high resolution. The O-RAN standard mandates eCPRI for the fronthaul interface.
Millimeter-wave wireless: 1–10 Gbps per link. Supports Level 2–3. Used in urban small cells where fiber is not available.
Sub-6 GHz wireless: 100 Mbps–1 Gbps. Supports only Level 1–2 with aggressive compression.

In practice, heterogeneous fronthaul is common: some APs have fiber (enabling Level 4) while others use wireless backhaul (limited to Level 2–3). The cooperation level can be selected per-AP based on the available fronthaul capacity.

Practical Constraints

•
eCPRI fronthaul adds 1–2 microseconds latency per hop
•
Wireless fronthaul adds 10–50 microseconds, limiting real-time cooperation
•
Fronthaul power scales with the data rate: approximately 0.1 W per Gbps
•
Multi-hop fronthaul (daisy-chain topology) accumulates latency and limits cooperation

Common Mistake: Using the Additive Noise Model for Low-Resolution Quantization

Mistake:

Applying the additive quantization noise model ( $\hat{s}^{(q)} = \hat{s} + e$ with $e$ independent of $\hat{s}$ ) when $b \leq 3$ bits.

Correction:

The additive noise model is accurate only for $b \geq 4$ – $5$ bits. For very low resolution ( $b = 1$ – $3$ ), the quantization error is correlated with the input and the Bussgang decomposition should be used instead: $\hat{s}^{(q)} = \alpha_{\text{B}} \hat{s} + e_{\text{B}}$ where $\alpha_{\text{B}} = 1 - 2^{-2b}/3$ is the Bussgang gain and $e_{\text{B}}$ is uncorrelated with $\hat{s}$ . The Bussgang model accounts for the signal attenuation caused by coarse quantization.

Quick Check

An AP with $N = 4$ antennas has a 10 Gbps eCPRI fronthaul link. The system bandwidth is $f_s = 100$ MHz. What is the maximum number of quantization bits per real dimension for Level 4 operation?

$b = 6$ bits

$b = 2$ bits

$b = 16$ bits

$b = 1$ bit

Correction:

b = 6

bits

Level 4 requires forwarding $N = 4$ complex samples per channel use: $2 b \times 4 \times 100 \times 10^6 \leq 10 \times 10^9$ , so $b \leq 10^{10} / (8 \times 10^8) = 12.5$ . Actually, let us recompute: $b \leq 10 \times 10^9 / (2 \times 4 \times 100 \times 10^6) = 10^{10}/(8 \times 10^8) = 12.5$ . So up to $b = 12$ bits. However, with overhead (control, timing), the effective capacity is about $50\%$ , giving $b \approx 6$ .

The Fronthaul Perspective on Cooperation Levels

The fronthaul analysis provides a unified view of the cooperation levels:

Level 1–2 (local combining): Fronthaul carries $|\mathcal{D}_m|$ complex scalars per channel use. With 4–6 bit quantization and $|\mathcal{D}_m| = 10$ : 0.8–1.2 Gbps per AP.
Level 3 (LSFD): Same fronthaul as Level 2, plus LSFD weights updated every 100–1000 ms (negligible additional overhead).
Level 4 (centralized MMSE): Fronthaul carries $N$ complex vectors per channel use. With 6–8 bit quantization and $N = 4$ : 4.8–6.4 Gbps per AP.

This quantifies the fundamental tradeoff: Level 3 provides 70–90% of Level 4's SINR performance at 15–25% of the fronthaul cost. When fronthaul capacity is the bottleneck — which it is in most deployments — Level 3 is the rational choice.

Key Takeaway

Fronthaul quantization with 4–6 bits per dimension is sufficient for near-lossless distributed processing. The SINR loss from quantization is less than 0.5 dB in most scenarios, which is far smaller than the gap between cooperation levels. The system designer should choose the cooperation level first (based on fronthaul capacity and AP antennas), then allocate quantization bits within the chosen level. Vector quantization provides 20–40% rate savings over scalar quantization when users are spatially correlated, but the added complexity is justified only in fronthaul-constrained deployments.

Fronthaul Compression

The process of quantizing and encoding the signals transmitted between access points and the central processing unit over capacity-limited fronthaul links. Includes scalar quantization, vector quantization, and information-theoretic compression based on Wyner-Ziv coding.

Bussgang Decomposition

A technique for analyzing nonlinear systems (such as quantizers) applied to Gaussian inputs. The output is decomposed as $Q(x) = \alpha x + e$ where $\alpha$ is a deterministic gain and $e$ is uncorrelated with $x$ . Used to analyze low-resolution ADCs and fronthaul quantization in massive MIMO systems.

Fronthaul Requirements and Compression