Ferkans — Interactive Telecom Tutor

A Seamless Hot-Spot Architecture

The user-centric clustering framework developed in the previous section tells us which APs serve each user. But it leaves open a critical architectural question: who coordinates? In the original cell-free formulation, a single CPU handles everything. In a large-scale deployment, this central CPU becomes a bottleneck and a single point of failure. The Fog Massive MIMO architecture, proposed by Bursalioglu, Caire, and collaborators, resolves this by introducing a hierarchical structure with master APs and local coordination — a user-centric, seamless hot-spot network where intelligence is distributed to the network edge.

Definition:
Master AP

For each user $k$ , the master AP $m^*(k)$ is the AP in the serving cluster $\mathcal{M}_k$ with the strongest large-scale fading coefficient:

$m^*(k) = \arg\max_{m \in \mathcal{M}_k} \beta_{mk}.$

The master AP is responsible for:

Pilot assignment: deciding which pilot sequence user $k$ transmits
Power control: computing the transmit power coefficients for user $k$
Data routing: receiving the decoded data and forwarding it to the core network
Coordination: instructing other APs in $\mathcal{M}_k$ about the processing to perform

The master AP concept decentralizes the CPU's role. Instead of one CPU coordinating all $M K$ AP-user pairs, each master AP coordinates only its local cluster. If $|\mathcal{M}_k| = 10$ , the master AP manages 10 AP-user links — a bounded task.

Master AP

The designated access point in a user's serving cluster that is responsible for coordination tasks: pilot assignment, power control, data aggregation, and instructing other serving APs. Typically the AP with the strongest link to the user.

Definition:
Fog Massive MIMO Architecture

Fog Massive MIMO is a user-centric cell-free architecture with three key features:

User-centric AP clusters: each user $k$ is served by a dynamically selected cluster $\mathcal{M}_k$ of nearby APs, as in Section 12.2
Master AP coordination: a master AP $m^*(k)$ in each cluster handles pilot assignment, power control, and data routing for user $k$
Seamless hot-spot coverage: APs form a dense grid (inter-AP distance 50–100 m in urban areas), and every location is well-served by its local cluster — there are no coverage holes or hard cell boundaries

The term "fog" refers to fog computing: pushing intelligence from a distant cloud (the CPU) to the network edge (the APs), analogous to how fog hugs the ground while clouds float high above.

The Fog Massive MIMO architecture was proposed by Bursalioglu, Caire, Papadopoulos, and collaborators at Bell Labs and TU Berlin. It combines the macro-diversity benefit of cell-free with the scalability of user-centric clustering, and adds the practical element of distributed coordination via master APs.

🎓CommIT Contribution(2019)

Fog Massive MIMO: User-Centric Seamless Hot-Spot Architecture

O. T. Demir, E. Bjornson, O. Y. Bursalioglu, G. Caire — IEEE Access

Bursalioglu, Caire, and collaborators proposed the Fog Massive MIMO architecture, which formalizes the user-centric cell-free concept with three innovations:

(1) Dynamic AP clustering based on large-scale fading, with guaranteed bounded per-AP load — this is the scalability breakthrough that makes cell-free practical.

(2) The master AP concept for distributed coordination, eliminating the need for a single centralized processor. The master AP makes local decisions (pilot assignment, power control) using only cluster-level information.

(3) Seamless hot-spot coverage analysis, showing that the user-centric architecture provides uniform service quality across the deployment area with no coverage holes — a property that conventional cellular networks cannot achieve.

The key result is that the 95th-percentile per-user rate in the Fog Massive MIMO architecture closely tracks the full cell-free baseline while requiring orders of magnitude less computation and fronthaul capacity.

cell-freeuser-centricfog-computingscalabilityView Paper →

Theorem: SINR with Master-AP-Based Power Control

Under Fog Massive MIMO with master-AP-based max-min power control, the uplink SINR for user $k$ with serving cluster $\mathcal{M}_k$ and MRC processing is

$\text{SINR}_k^{\text{fog}} = \frac{p_k \left( \sum_{m \in \mathcal{M}_k} \gamma_{mk} \right)^2}{\sum_{k'=1}^{K} p_{k'} \sum_{m \in \mathcal{M}_k} \gamma_{mk} \beta_{mk'} + \sigma^2 \sum_{m \in \mathcal{M}_k} \gamma_{mk}}$

where the power coefficients $\{p_k\}$ are the solution of the max-min fairness problem

$\max_{\{p_k \geq 0\}} \min_k \text{SINR}_k^{\text{fog}} \quad \text{s.t.} \quad \sum_{k \in \mathcal{K}_m} p_k \leq P_t, \quad m = 1, \ldots, M.$

The per-AP power constraint ensures that each AP's total transmit power is bounded.

The SINR expression is identical to the user-centric case (TSINR Under User-Centric MRC Processing), but the power control is now designed to maximize the minimum SINR across all users. The key insight is that the master AP $m^*(k)$ can compute the relevant terms in the SINR using only information from the cluster $\mathcal{M}_k$ and the large-scale fading coefficients $\{\beta_{mk'}\}$ — it does not need global CSI.

Proof

SINR structure

The SINR expression follows from TSINR Under User-Centric MRC Processing. The UatF bound with MRC combining over the cluster $\mathcal{M}_k$ yields the same form.

Max-min formulation

The max-min fairness problem can be solved by bisection: fix a target SINR $t$ and check whether there exist feasible powers $\{p_k\}$ such that $\text{SINR}_k^{\text{fog}} \geq t$ for all $k$ and $\sum_{k \in \mathcal{K}_m} p_k \leq P_t$ for all $m$ . This is a linear feasibility problem in $\{p_k\}$ .

Distributed computation

The master AP $m^*(k)$ can evaluate $\sum_{m \in \mathcal{M}_k} \gamma_{mk}$ and $\sum_{m \in \mathcal{M}_k} \gamma_{mk} \beta_{mk'}$ using local channel estimates and large-scale fading information shared within the cluster. The power control iterations require inter-cluster message passing (each cluster shares its power allocation with overlapping clusters), but the messages are scalar-valued and the convergence is fast. $\blacksquare$

Example: Master AP Selection Strategies

Consider a network with 3 APs and a user $k$ with large-scale fading coefficients $\beta_{1k} = -80$ dB, $\beta_{2k} = -75$ dB, $\beta_{3k} = -90$ dB, and estimation qualities $\gamma_{1k} = 0.8 \beta_{1k}$ , $\gamma_{2k} = 0.9 \beta_{2k}$ , $\gamma_{3k} = 0.5 \beta_{3k}$ . Compare the master AP chosen by (a) largest $\beta_{mk}$ and (b) largest $\gamma_{mk}$ .

Solution

Strategy (a): Largest path gain

In linear scale: $\beta_{1k} = 10^{-8}$ , $\beta_{2k} = 3.16 \times 10^{-8}$ , $\beta_{3k} = 10^{-9}$ . The largest is $\beta_{2k}$ , so $m^*(k) = 2$ .

Strategy (b): Largest estimation quality

$\gamma_{1k} = 0.8 \times 10^{-8} = 8 \times 10^{-9}$ , $\gamma_{2k} = 0.9 \times 3.16 \times 10^{-8} = 2.84 \times 10^{-8}$ , $\gamma_{3k} = 0.5 \times 10^{-9} = 5 \times 10^{-10}$ . The largest is $\gamma_{2k}$ , so $m^*(k) = 2$ again.

When do the strategies differ?

In this example, both strategies agree because the estimation quality $\gamma_{mk}$ is approximately proportional to $\beta_{mk}$ . They can differ when pilot contamination disproportionately degrades the estimate at the closest AP — for instance, if AP 2's pilot is heavily contaminated while AP 1's is clean, we could have $\gamma_{1k} > \gamma_{2k}$ despite $\beta_{1k} < \beta_{2k}$ . In such cases, the $\gamma$ -based strategy is preferable.

Master AP Selection Strategies Comparison

Compare different master AP selection strategies (largest path gain, largest estimation quality, lowest load) and their impact on the per-user rate distribution. A good master AP selection improves coordination quality and overall network fairness.

Parameters

M

(number of APs)100

K

(number of users)30

N_{\text{cl}}

(cluster size)8

Common Mistake: Master AP is Not a Base Station

Mistake:

Treating the master AP as a "small base station" that exclusively serves user $k$ , reverting to a cellular mindset with the master AP as the cell center.

Correction:

The master AP $m^*(k)$ is a coordinator, not an exclusive server. It serves as master for user $k$ while simultaneously being a regular serving AP for other users. An AP can be the master for one user, a regular serving AP for five others, and not involved with the remaining users. The master role is a coordination function, not a service exclusivity.

Historical Note: From Cloud to Fog to Cell-Free

2012–2019

The term "fog computing" was coined by Cisco in 2012 to describe an architecture where computation, storage, and networking are distributed to the network edge — closer to the "ground" where data is generated, like fog compared to clouds. The application to massive MIMO by Bursalioglu, Caire, and collaborators recognized that the same principle applies to wireless signal processing: instead of centralizing all combining and precoding at a distant cloud (the CPU), push the intelligence to the APs themselves. Each AP processes only its local users, and the master AP coordinates only its small cluster. This fog architecture bridges the gap between fully centralized cell-free (theoretically optimal but unscalable) and fully distributed small cells (scalable but losing the macro-diversity benefit).

⚠️Engineering Note

Fronthaul Requirements in Fog Massive MIMO

In the Fog Massive MIMO architecture, fronthaul links connect each AP to its master AP (or directly to the CPU in a two-tier deployment). The fronthaul load per AP $m$ is

$F_m = |\mathcal{K}_m| \times B_{\text{quant}} \times (\tau_c - \tau_p) \times \frac{1}{T_c}$

where $B_{\text{quant}}$ is the quantization bits per complex sample and $T_c$ is the coherence time. With $|\mathcal{K}_m| \leq L_{\max} = 20$ , $B_{\text{quant}} = 16$ bits (8 I + 8 Q), $\tau_c = 200$ , $\tau_p = 10$ , and $T_c = 1$ ms, the per-AP fronthaul rate is $20 \times 16 \times 190 / 10^{-3} = 60.8$ Mbps. This is feasible over standard Ethernet or fiber-to-the-AP links.

Practical Constraints

•
Per-AP fronthaul rates of 50–100 Mbps achievable with standard Ethernet
•
Fronthaul quantization: 8–12 bits per I/Q component, per eCPRI standard
•
O-RAN 7.2x split places the bulk of PHY processing at the AP (DU collocated with RU)

📋 Ref: O-RAN Alliance WG4: Fronthaul Interface (v7.0)

Definition:
Levels of AP Cooperation

Fog Massive MIMO supports multiple levels of AP cooperation within each user-centric cluster:

Level 1 (Non-coherent): Each AP in $\mathcal{M}_k$ processes its signal independently. The CPU (or master AP) combines the local estimates by simple averaging.

Level 2 (Coherent combining): APs share their channel estimates with the master AP, which computes optimal combining weights $\{a_{mk}\}_{m \in \mathcal{M}_k}$ and instructs each AP. The combined estimate is $\hat{s}_k = \sum_{m \in \mathcal{M}_k} a_{mk} \hat{g}_{mk}^* y_m$ .

Level 3 (Local MMSE): Each AP computes a local MMSE combining vector using only its own channel estimates. Results are forwarded to the master AP for final combining.

Level 4 (Centralized MMSE): The master AP has access to all channel estimates within the cluster and computes a cluster-wide MMSE combiner. This achieves the best performance at the cost of highest fronthaul and computation.

Moving from Level 1 to Level 4 increases both performance and fronthaul requirements. The right level depends on the fronthaul capacity and the desired quality of service. Chapter 13 develops the distributed processing theory for each level in detail.

Quick Check

What is the primary role of the master AP in the Fog Massive MIMO architecture?

It transmits with the highest power to user $k$

It performs final data detection for user $k$

It coordinates the serving cluster: pilot assignment, power control, and data routing

It stores user $k$ 's data for caching purposes

Correction:

It coordinates the serving cluster: pilot assignment, power control, and data routing

Correct. The master AP is a coordinator — it handles pilot assignment, power control decisions, and routes the decoded data to the core network. This distributes the CPU's functions across the network.

Fog Massive MIMO

A user-centric cell-free massive MIMO architecture that distributes coordination intelligence to the network edge via master APs. Each user is served by a dynamically selected cluster of nearby APs, with one designated master AP handling coordination. Proposed by Bursalioglu, Caire, and collaborators.

Related: Master AP, User Centric, Cell Free

Fog Massive MIMO and O-RAN

The Fog Massive MIMO architecture maps naturally onto the O-RAN (Open Radio Access Network) framework. The O-RAN 7.2x functional split places lower-PHY processing (FFT, beamforming, channel estimation) at the RU (Radio Unit), which corresponds to the AP in our framework. The DU (Distributed Unit) handles higher-layer processing and can act as the master AP for a cluster of RUs. The CU (Centralized Unit) provides network-level coordination — analogous to the CPU in cell-free. This hierarchical mapping suggests that Fog Massive MIMO is not merely a theoretical concept but an architecture that can be implemented within the existing O-RAN ecosystem.

Fog Massive MIMO

A Seamless Hot-Spot Architecture

Definition: Master AP

Master AP

Definition: Fog Massive MIMO Architecture

Fog Massive MIMO: User-Centric Seamless Hot-Spot Architecture

Theorem: SINR with Master-AP-Based Power Control

SINR structure

Max-min formulation

Distributed computation

Example: Master AP Selection Strategies

Strategy (a): Largest path gain

Strategy (b): Largest estimation quality

When do the strategies differ?

Master AP Selection Strategies Comparison

Parameters

Common Mistake: Master AP is Not a Base Station

Historical Note: From Cloud to Fog to Cell-Free

Fronthaul Requirements in Fog Massive MIMO

Definition: Levels of AP Cooperation

Quick Check

Fog Massive MIMO

Fog Massive MIMO and O-RAN

Definition:
Master AP

Definition:
Fog Massive MIMO Architecture

Definition:
Levels of AP Cooperation