Ferkans — Interactive Telecom Tutor

Letting the User Define the Cell

In conventional cellular networks, the network defines the cells, and users are assigned to them. In user-centric cell-free, we flip this: each user defines its own virtual cell — a dynamically selected cluster of nearby APs. No two users see the same "cell," and the clusters overlap extensively. This is not merely an optimization trick; it is an architectural paradigm shift that makes cell-free massive MIMO scalable.

Definition:
Cluster Indicator Matrix

Define the cluster indicator matrix $\mathbf{D} \in \{0, 1\}^{M \times K}$ with entries

$D_{mk} = \begin{cases} 1 & \text{if AP } m \text{ serves user } k, \\ 0 & \text{otherwise.} \end{cases}$

The serving cluster for user $k$ is $\mathcal{M}_k = \{m : D_{mk} = 1\}$ , and the user set for AP $m$ is $\mathcal{K}_m = \{k : D_{mk} = 1\}$ .

Full cell-free corresponds to $D_{mk} = 1$ for all $m, k$ (dense matrix). Conventional cellular corresponds to each column having exactly one nonzero entry. User-centric cell-free is the middle ground: each column has a moderate number of ones, determined by proximity.

The matrix $\mathbf{D}$ is the key design variable. Its sparsity pattern determines the computational complexity, fronthaul load, and achievable rate of the network.

Serving Cluster

For user $k$ , the set $\mathcal{M}_k$ of APs that actively participate in serving user $k$ — estimating its channel, computing combining/precoding weights, and forwarding processed signals. In user-centric cell-free, $|\mathcal{M}_k| \ll M$ .

Definition:
User-Centric Clustering Rule

A user-centric clustering rule assigns to each user $k$ a serving cluster $\mathcal{M}_k$ based on the large-scale fading coefficients. The simplest rule is largest-large-scale-fading (LLSF): for each user $k$ , select the $N_{\text{cl}}$ APs with the largest $\beta_{mk}$ values:

$\mathcal{M}_k = \underset{\mathcal{S} \subset \{1, \ldots, M\}, \, |\mathcal{S}| = N_{\text{cl}}}{\arg\max} \sum_{m \in \mathcal{S}} \beta_{mk}.$

Equivalently, sort the APs by decreasing $\beta_{mk}$ and take the top $N_{\text{cl}}$ .

LLSF is the simplest rule but not the only one. Threshold-based rules ( $\beta_{mk} \geq \beta_{\text{th}}$ ) allow variable cluster sizes. Optimization-based rules jointly design the clusters to maximize a network utility.

Largest-Large-Scale-Fading (LLSF)

A user-centric clustering rule that selects, for each user, the $N_{\text{cl}}$ APs with the strongest large-scale fading coefficients. Simple and effective, requiring only knowledge of path loss and shadow fading.

Definition:
Scalable Cell-Free Massive MIMO

A cell-free massive MIMO system is scalable if its per-AP computational complexity and fronthaul load remain bounded as $M$ and $K$ grow. Formally, we require:

$\max_m |\mathcal{K}_m| \leq L_{\max} \quad \text{and} \quad \max_k |\mathcal{M}_k| \leq N_{\max}$

for finite constants $L_{\max}$ and $N_{\max}$ that do not depend on $M$ or $K$ . The total network complexity is then $O(M L_{\max} + K N_{\max})$ , which scales linearly (not quadratically) with network size.

The scalability condition is a constraint on the sparsity of $\mathbf{D}$ : each row has at most $L_{\max}$ nonzeros, each column at most $N_{\max}$ . This is precisely the regime achieved by user-centric clustering with bounded cluster sizes.

Theorem: SINR Under User-Centric MRC Processing

Under user-centric clustering with cluster indicator $\mathbf{D}$ and conjugate beamforming, the UatF uplink SINR for user $k$ is

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

where the sums run only over the serving cluster $\mathcal{M}_k$ , not all $M$ APs.

This expression is identical in structure to the full cell-free SINR (TSINR Under Full Cell-Free MRC Processing) but with sums restricted to $\mathcal{M}_k$ . The numerator loses some coherent beamforming gain (fewer APs), but the interference terms are also evaluated only at the serving APs, which tend to be the ones that provide the strongest signal anyway.

Proof

Restrict combining to cluster APs

Under user-centric processing, the CPU forms $\hat{s}_k = \sum_{m \in \mathcal{M}_k} \hat{g}_{mk}^* y_m$ instead of summing over all $M$ APs. The derivation follows identically to TSINR Under Full Cell-Free MRC Processing, with the index set restricted to $\mathcal{M}_k$ .

Signal term

The desired signal component is $\sum_{m \in \mathcal{M}_k} \gamma_{mk} \sqrt{p_k} s_k$ , giving signal power $p_k \left(\sum_{m \in \mathcal{M}_k} \gamma_{mk}\right)^2$ .

Interference-plus-noise

The interference from user $k'$ is $\sum_{m \in \mathcal{M}_k} \hat{g}_{mk}^* g_{mk'} \sqrt{p_{k'}} s_{k'}$ . Using independence of the small-scale fading (conditioned on large-scale fading), the interference power from $k'$ is $p_{k'} \sum_{m \in \mathcal{M}_k} \gamma_{mk} \beta_{mk'}$ . Summing over all $k'$ and adding noise power yields the denominator. $\blacksquare$

Example: SINR Loss from Cluster Truncation

A user is located in a network with $M = 200$ APs. The large-scale fading coefficients (sorted in decreasing order) are $\beta_{(1)k} = -70$ dB, $\beta_{(2)k} = -75$ dB, ..., $\beta_{(10)k} = -95$ dB, with the remaining 190 APs having $\beta_{mk} < -100$ dB. Assuming $\gamma_{mk} \approx \beta_{mk}$ (high-SNR estimation regime), compare the signal power with $|\mathcal{M}_k| = 10$ vs. $|\mathcal{M}_k| = M = 200$ .

Solution

Signal power with full cell-free

The signal power is $p_k \left(\sum_{m=1}^{200} \gamma_{mk}\right)^2 \approx p_k \left(\sum_{m=1}^{200} \beta_{mk}\right)^2$ . In linear scale (dB to linear): the top 10 APs contribute approximately $10^{-7} + 3.16 \times 10^{-8} + \ldots + 3.16 \times 10^{-10} \approx 1.43 \times 10^{-7}$ . The remaining 190 APs contribute at most $190 \times 10^{-10} = 1.9 \times 10^{-8}$ , which is about 13% of the top-10 contribution.

Signal power with user-centric cluster of 10

Using only the top 10 APs: $\sum_{m \in \mathcal{M}_k} \gamma_{mk} \approx 1.43 \times 10^{-7}$ . The full sum is approximately $1.43 \times 10^{-7} + 1.9 \times 10^{-8} \approx 1.62 \times 10^{-7}$ .

Loss quantification

The signal power ratio is $\frac{(1.43)^2}{(1.62)^2} \approx 0.78$ , a loss of about $10 \log_{10}(0.78) \approx -1.1$ dB. Meanwhile, the computational cost is reduced by a factor of $200/10 = 20\times$ . Trading 1.1 dB of signal power for a 20 $\times$ reduction in complexity is an excellent bargain.

Definition:
Threshold-Based Clustering

An alternative to fixed-size clustering is threshold-based clustering, where

$\mathcal{M}_k = \{m : \beta_{mk} \geq \beta_{\text{th}}\}$

for a network-wide threshold $\beta_{\text{th}}$ . This allows variable cluster sizes: users at "hot spots" surrounded by many APs get large clusters, while isolated users get small clusters. The threshold is chosen to balance performance against a per-AP load constraint $\max_m |\mathcal{K}_m| \leq L_{\max}$ .

Threshold-based clustering naturally adapts to the spatial distribution of APs and users. In dense areas, users get many serving APs; in sparse areas, fewer. This is more efficient than forcing a uniform cluster size across the network.

LLSF User-Centric Clustering

Complexity:

O(K \, M \log M)

Input: Large-scale fading matrix

\{\beta_{mk}\}_{m=1,\ldots,M}^{k=1,\ldots,K}

, cluster size

N_{\text{cl}}

, per-AP load limit

L_{\max}

Output: Cluster indicator matrix

\mathbf{D} \in \{0,1\}^{M \times K}

1. Initialize

D_{mk} = 0

for all

m, k

2. for

k = 1, \ldots, K

do

3.

\quad

Sort APs by decreasing

\beta_{mk}

: let

(m_{(1)}, m_{(2)}, \ldots, m_{(M)})

be the sorted indices

4.

\quad

n \leftarrow 0

5.

\quad

for

i = 1, \ldots, M

do

6.

\quad\quad

if

|\mathcal{K}_{m_{(i)}}| < L_{\max}

then

7.

\quad\quad\quad

D_{m_{(i)}, k} \leftarrow 1

;

n \leftarrow n + 1

8.

\quad\quad\quad

if

n = N_{\text{cl}}

then break

9.

\quad\quad

end if

10.

\quad

end for

11. end for

The per-AP load constraint in line 6 prevents any single AP from being overloaded. Without this constraint, a centrally located AP could end up serving all users.

Dynamic User-Centric Clustering Visualization

Visualize a 2D network deployment with randomly placed APs and users. Each user's serving cluster is highlighted, showing the overlapping nature of user-centric clusters. Adjust the cluster size to see how clusters expand or shrink.

Parameters

M

(number of APs)60

K

(number of users)10

N_{\text{cl}}

(cluster size)5

Highlighted user1

Common Mistake: Fixed Cluster Size is Not Always Optimal

Mistake:

Setting the same cluster size $N_{\text{cl}}$ for all users, regardless of their location or the local AP density.

Correction:

Users near many APs benefit from larger clusters, while users in sparse areas may only have a few useful APs. Threshold-based clustering or optimization-based approaches that adapt $|\mathcal{M}_k|$ to the local geometry generally outperform fixed-size clustering, especially in networks with non-uniform AP placement.

Full Cell-Free vs. User-Centric vs. Cellular

Property	Full Cell-Free	User-Centric Cell-Free	Cellular
Serving set per user	All $M$ APs	$\|\mathcal{M}_k\| \ll M$ nearby APs	1 BS (or 2–3 in CoMP)
Computational complexity	$O(M K)$	$O(M L_{\max})$ , $L_{\max}$ bounded	$O(K)$ per cell
Fronthaul load	Very high (grows with $MK$ )	Moderate (bounded per AP)	Low (local processing)
Cell-edge performance	Excellent (no cell edges)	Very good (soft boundaries)	Poor (hard cell edges)
Fairness (95%-likely rate)	Best	Near-optimal	Poor
Practical scalability	No	Yes	Yes
Pilot assignment flexibility	Network-wide optimization	Cluster-aware assignment	Cell-level reuse

SINR CDF: Full Cell-Free vs. User-Centric vs. Cellular

Compare the cumulative distribution function of the downlink SINR under three architectures: full cell-free (all APs serve all users), user-centric cell-free (each user served by a cluster of nearby APs), and conventional cellular (each user served by one BS). The 5th percentile SINR measures fairness.

Parameters

M

(number of APs)200

K

(number of users)40

N_{\text{cl}}

(cluster size)10

Number of cells (cellular)9

Key Takeaway

User-centric clustering resolves the scalability–performance tradeoff. By restricting each user's serving set to $N_{\text{cl}} \ll M$ nearby APs, the complexity drops from $O(MK)$ to $O(M L_{\max})$ with bounded per-AP load. The SINR loss from cluster truncation is typically 1–3 dB — negligible compared to the 20–100 $\times$ reduction in computational and fronthaul cost.

🔧Engineering Note

Cluster Update Rate in Practice

The user-centric clusters $\{\mathcal{M}_k\}$ depend on large-scale fading coefficients $\{\beta_{mk}\}$ , which change on a timescale of seconds (shadow fading) to minutes (user mobility). Clusters need not be updated every coherence block — they can be recomputed periodically (e.g., every 100–1000 ms) with negligible performance loss. This amortizes the $O(K M \log M)$ sorting cost of the LLSF algorithm over many coherence blocks.

Practical Constraints

•
Cluster updates at 1–10 Hz are sufficient for pedestrian mobility (3 km/h)
•
Vehicular mobility (60–120 km/h) may require updates at 10–50 Hz
•
Large-scale fading measurements available via uplink SRS or RSRP reports

Why This Matters: User-Centric Clustering in 5G and Beyond

The user-centric philosophy is already partially realized in 5G NR through multi-TRP (Transmission/Reception Point) operation and CoMP. In 5G NR Release 16/17, a UE can receive downlink PDSCH from up to 2 TRPs simultaneously (multi-TRP), and the network selects the serving TRPs based on CSI feedback. Cell-free massive MIMO with user-centric clustering extends this to many more APs per user, with TDD-based channel acquisition replacing explicit CSI feedback. The O-RAN architecture with disaggregated RU/DU/CU provides the infrastructure for user-centric cell-free deployments.

See full treatment in Downlink Fronthaul Strategies

Quick Check

In user-centric cell-free massive MIMO, can two users share the same AP in their serving clusters?

No — each AP serves exactly one user

Yes — AP clusters overlap, and a single AP may serve multiple users

Only if the users are assigned the same pilot

Correction:

Yes — AP clusters overlap, and a single AP may serve multiple users

Correct. The serving clusters $\mathcal{M}_k$ overlap extensively. An AP near the boundary between two users will likely serve both. This is captured by $|\mathcal{K}_m|$ being the number of users served by AP $m$ .

User-Centric Clustering

Letting the User Define the Cell

Definition: Cluster Indicator Matrix

Serving Cluster

Definition: User-Centric Clustering Rule

Largest-Large-Scale-Fading (LLSF)

Definition: Scalable Cell-Free Massive MIMO

Theorem: SINR Under User-Centric MRC Processing

Restrict combining to cluster APs

Signal term

Interference-plus-noise

Example: SINR Loss from Cluster Truncation

Signal power with full cell-free

Signal power with user-centric cluster of 10

Loss quantification

Definition: Threshold-Based Clustering

LLSF User-Centric Clustering

Dynamic User-Centric Clustering Visualization

Parameters

Common Mistake: Fixed Cluster Size is Not Always Optimal

Full Cell-Free vs. User-Centric vs. Cellular

SINR CDF: Full Cell-Free vs. User-Centric vs. Cellular

Parameters

Key Takeaway

Cluster Update Rate in Practice

Why This Matters: User-Centric Clustering in 5G and Beyond

Quick Check

Definition:
Cluster Indicator Matrix

Definition:
User-Centric Clustering Rule

Definition:
Scalable Cell-Free Massive MIMO

Definition:
Threshold-Based Clustering