Ferkans — Interactive Telecom Tutor

Pilot Assignment Gets a New Structure

Pilot contamination is the Achilles' heel of massive MIMO (Chapter 3): users sharing the same pilot cause channel estimation errors that do not vanish with increasing antenna count. In conventional cell-free, pilot assignment is a global optimization over all $K$ users. In user-centric cell-free, the cluster structure provides a powerful new lever: two users can safely share a pilot if their serving clusters do not overlap, because the interference from a distant co-pilot user is attenuated by the large-scale fading at the serving APs. This section develops pilot assignment algorithms that exploit the user-centric cluster geometry.

Definition:
Pilot Contamination in User-Centric Systems

Let $\tau_p$ denote the number of orthogonal pilot sequences, and let ${\mathbf{S}_{i,k}}_{k} \in \{1, \ldots, \tau_p\}$ be the pilot index assigned to user $k$ . The set of co-pilot users (users sharing pilot $t$ ) is $\mathcal{P}_t = \{k : {\mathbf{S}_{i,k}}_{k} = t\}$ .

Under MMSE estimation, the channel estimate $\hat{g}_{mk}$ at AP $m$ for user $k$ is contaminated by users in $\mathcal{P}_{{\mathbf{S}_{i,k}}_{k}} \setminus \{k\}$ :

$\hat{g}_{mk} = \frac{\sqrt{p_k \tau_p} \, \beta_{mk}}{p_k \tau_p \beta_{mk} + \sum_{k' \in \mathcal{P}_{{\mathbf{S}_{i,k}}_{k}} \setminus \{k\}} p_{k'} \tau_p \beta_{mk'} + \sigma^2} \left( \sqrt{p_k \tau_p} \, g_{mk} + \sum_{k' \in \mathcal{P}_{{\mathbf{S}_{i,k}}_{k}} \setminus \{k\}} \sqrt{p_{k'} \tau_p} \, g_{mk'} + \tilde{w}_m \right)$

In user-centric systems, AP $m$ only processes users in $\mathcal{K}_m$ . The effective contamination at AP $m$ for user $k$ comes only from co-pilot users $k'$ that have significant $\beta_{mk'}$ — that is, users whose serving clusters overlap with $\mathcal{M}_k$ .

This is the structural advantage of user-centric clustering for pilot assignment: co-pilot users with non-overlapping clusters cause negligible contamination at the relevant APs.

Definition:
Pilot Contamination Metric

For two users $k$ and $k'$ with serving clusters $\mathcal{M}_k$ and $\mathcal{M}_{k'}$ , define the contamination metric

$\rho(k, k') = \sum_{m \in \mathcal{M}_k} \beta_{mk'} + \sum_{m \in \mathcal{M}_{k'}} \beta_{mk}.$

This measures the total interference that users $k$ and $k'$ would cause to each other's channel estimates if they shared a pilot. Small $\rho(k, k')$ means the users can safely share a pilot; large $\rho(k, k')$ means they should not.

The contamination metric is symmetric: $\rho(k, k') = \rho(k', k)$ . It captures the key observation that contamination depends on the large-scale fading at the serving APs of each user, not at all APs in the network.

Theorem: Safe Pilot Reuse Condition

Users $k$ and $k'$ can share a pilot with bounded contamination loss if

$\rho(k, k') \leq \epsilon \cdot \min\left\{ \sum_{m \in \mathcal{M}_k} \beta_{mk}, \; \sum_{m \in \mathcal{M}_{k'}} \beta_{mk'} \right\}$

for a tolerance parameter $\epsilon > 0$ . When this condition holds, the SINR degradation from pilot sharing is at most a factor of $(1 + \epsilon)^{-2}$ compared to contamination-free operation.

The condition says: the cross-contamination between users $k$ and $k'$ is small relative to the weakest user's own signal. If user $k$ is well-served by its cluster (large $\sum \beta_{mk}$ ), a small amount of contamination from $k'$ is tolerable. The parameter $\epsilon$ controls the tradeoff between pilot reuse (smaller $\tau_p$ ) and contamination.

Proof

SINR with contamination

When users $k$ and $k'$ share a pilot, the estimation quality degrades as $\gamma_{mk} \to \gamma_{mk}' = \frac{p_k \tau_p \beta_{mk}^{2}}{p_k \tau_p \beta_{mk} + p_{k'} \tau_p \beta_{mk'} + \sigma^2}$ . The SINR numerator becomes $p_k (\sum_{m \in \mathcal{M}_k} \gamma_{mk}')^2$ .

Bounding the degradation

Using the condition $\sum_{m \in \mathcal{M}_k} \beta_{mk'} \leq \epsilon \sum_{m \in \mathcal{M}_k} \beta_{mk}$ , we bound $\gamma_{mk}' \geq \gamma_{mk} / (1 + \epsilon)$ for each AP $m$ . Therefore $\sum_{m \in \mathcal{M}_k} \gamma_{mk}' \geq (1 + \epsilon)^{-1} \sum_{m \in \mathcal{M}_k} \gamma_{mk}$ , and the SINR degrades by at most $(1 + \epsilon)^{-2}$ . $\blacksquare$

Graph-Coloring-Based Pilot Assignment

Complexity:

O(K^{2} + K \tau_p)

Input: Users

\{1, \ldots, K\}

, clusters

\{\mathcal{M}_k\}

, large-scale fading

\{\beta_{mk}\}

,

\tau_p

pilots, tolerance

\epsilon

Output: Pilot assignment

\{{\mathbf{S}_{i,k}}_{k}\}_{k=1}^{K}

1. Build interference graph

G = (V, E)

:

2.

\quad

Vertices:

V = \{1, \ldots, K\}

(one per user)

3.

\quad

Edges:

(k, k') \in E

if

\rho(k, k') > \epsilon \cdot \min\{\sum_{m \in \mathcal{M}_k} \beta_{mk}, \sum_{m \in \mathcal{M}_{k'}} \beta_{mk'}\}

4. Graph color

G

with

\tau_p

colors using a greedy algorithm:

5.

\quad

Order users by decreasing degree in

G

6.

\quad

for

k

in ordered user list do

7.

\quad\quad

Assign

{\mathbf{S}_{i,k}}_{k}

= smallest color not used by

k

's neighbors in

G

8.

\quad

end for

9. if chromatic number

> \tau_p

:

10.

\quad

Relax

\epsilon

or increase

\tau_p

The greedy graph coloring produces a valid assignment when $\tau_p$ exceeds the chromatic number of $G$ . The interference graph is typically sparse in user-centric systems because only users with overlapping clusters create edges — distant users with disjoint clusters can always share a pilot.

Example: Pilot Reuse in a User-Centric Network

Consider a network with $M = 100$ APs, $K = 40$ users, and cluster size $N_{\text{cl}} = 10$ . Suppose the average cluster overlap between any two users is 2 APs (i.e., on average, $|\mathcal{M}_k \cap \mathcal{M}_{k'}| = 2$ for nearby users and 0 for distant users). Estimate the number of orthogonal pilots needed with (a) full cell-free and (b) user-centric pilot assignment.

Solution

Full cell-free baseline

In full cell-free, every AP processes every user, so all users can potentially contaminate each other. With $K = 40$ users, we ideally need $\tau_p = 40$ orthogonal pilots to avoid any contamination. With a coherence block of $\tau_c = 200$ , this leaves $200 - 40 = 160$ symbols for data — a pilot overhead of 20%.

User-centric pilot assignment

With cluster size 10 and average overlap of 2 APs, each user has significant interference with approximately $10 \times 2 / 10 \approx 2$ neighbors (those whose clusters overlap significantly). The interference graph has average degree $\approx 4$ , so the chromatic number is approximately 5. With $\tau_p = 5$ orthogonal pilots, the pilot overhead drops to $5/200 = 2.5\%$ .

Improvement

User-centric pilot assignment reduces the pilot overhead from 20% to 2.5% — a factor of 8 improvement. This translates to $(200 - 5)/(200 - 40) = 1.22$ , or a 22% increase in spectral efficiency from the data phase alone, on top of the reduced contamination.

Pilot Contamination Under User-Centric Assignment

Visualize how the pilot contamination level varies with cluster size, number of users, and number of pilots. Observe that user-centric clustering reduces the effective number of co-pilot interferers seen at each user's serving APs.

Parameters

M

(number of APs)200

K

(number of users)40

N_{\text{cl}}

(cluster size)10

\tau_p

(number of pilots)10

Common Mistake: Global Pilot Assignment is Overkill in User-Centric Systems

Mistake:

Solving a global optimization over all $K$ users for pilot assignment, treating the problem as if every user interferes with every other user.

Correction:

In user-centric systems, the interference graph is sparse — only users with overlapping clusters create significant contamination. A local graph-coloring approach that considers only neighboring clusters is both computationally cheaper ( $O(K L_{\max})$ instead of $O(K^{2})$ ) and produces nearly identical pilot assignments.

Definition:
Cluster-Aware Pilot Assignment

A pilot assignment is cluster-aware if it exploits the user-centric cluster structure to minimize contamination. Formally, the assignment $\{{\mathbf{S}_{i,k}}_{k}\}$ minimizes a network contamination cost:

$\min_{\{{\mathbf{S}_{i,k}}_{k}\}} \sum_{k=1}^{K} \sum_{\substack{k' \neq k \\ {\mathbf{S}_{i,k}}_{k'} = {\mathbf{S}_{i,k}}_{k}}} \rho(k, k')$

subject to ${\mathbf{S}_{i,k}}_{k} \in \{1, \ldots, \tau_p\}$ for all $k$ . This is equivalent to a weighted graph coloring problem on the interference graph.

The master AP $m^*(k)$ can locally solve a simplified version of this problem: assign pilot ${\mathbf{S}_{i,k}}_{k}$ to minimize the contamination at the APs in $\mathcal{M}_k$ , considering only the pilots already assigned to users whose clusters overlap with $\mathcal{M}_k$ . This distributed approach requires only local information exchange.

Key Takeaway

User-centric clustering creates natural pilot reuse opportunities. Users with non-overlapping serving clusters can safely share pilots with negligible contamination. This structural advantage reduces pilot overhead from $O(K)$ (full cell-free) to $O(\chi(G))$ where $\chi(G)$ is the chromatic number of the (sparse) interference graph. In typical deployments, $\chi(G) \ll K$ .

⚠️Engineering Note

Pilot Overhead in 5G NR Cell-Free Deployments

In 5G NR, the SRS (Sounding Reference Signal) is used for uplink channel estimation in TDD mode. The SRS resource configuration supports up to 4 antenna ports and various comb factors (2 or 4) for frequency-domain multiplexing. For cell-free deployments, the SRS configuration must be carefully designed to accommodate the user-centric pilot assignment:

With $\tau_p = 10$ orthogonal pilots and a comb-4 SRS, the system can support 40 users with 4 SRS symbols per slot (14 OFDM symbols at 30 kHz SCS).
The SRS overhead is $4/14 \approx 28.6\%$ per slot — significant, but most slots do not carry SRS (typical periodicity: every 5–20 slots).
The effective pilot overhead per coherence block depends on the SRS periodicity and the coherence time of the channel.

Practical Constraints

•
3GPP TS 38.211 Section 6.4.1.4: SRS sequence generation and resource mapping
•
SRS bandwidth configuration: 4–272 resource blocks
•
Maximum 4 SRS ports per UE in Release 16

📋 Ref: 3GPP TS 38.211, Section 6.4.1.4

Pilot Contamination

The phenomenon where users sharing the same pilot sequence cause mutual interference in channel estimation, degrading the quality of channel estimates. In massive MIMO, pilot contamination is the dominant source of residual interference and does not vanish as the number of antennas grows.

Interference Graph

A graph $G = (V, E)$ where vertices represent users and edges connect users whose serving clusters overlap significantly. Used for graph-coloring-based pilot assignment: users connected by an edge should not share a pilot.

Quick Check

Two users $k$ and $k'$ have non-overlapping serving clusters: $\mathcal{M}_k \cap \mathcal{M}_{k'} = \emptyset$ . The users are far apart with $\beta_{mk'} \approx 0$ for all $m \in \mathcal{M}_k$ . Can they safely share a pilot?

No — pilot sharing always causes contamination

Yes — because the contamination metric $\rho(k, k') \approx 0$

Only if they are in different cells