Ferkans — Interactive Telecom Tutor

From Cellular to Cell-Free

Classical cellular networks divide the world into cells, each served by one base station. This architecture is a historical artifact: it simplified frequency reuse and handover in the 1980s. It is also the source of most cellular pathologies — inter-cell interference, asymmetric cell-edge vs cell-center performance, cumbersome handover under mobility. Cell-free massive MIMO eliminates cells: all APs jointly serve all UEs, coordinated by a central processing unit (CPU) over a fronthaul network. Under mobility, UEs experience smooth macro-diversity rather than discrete handovers. Combined with OTFS's high-mobility resilience, the cell-free OTFS architecture is the natural 6G vision for mobile wireless.

,

Definition:
Cell-Free Massive MIMO Architecture

A cell-free massive MIMO system consists of:

$L$ access points (APs) distributed over the service area, each with $N_a$ antennas (typical: $N_a = 1$ - $8$ ).
$K$ user equipments (UEs), mobile or static.
A central processing unit (CPU) connected to all APs via fronthaul links (eCPRI or O-RAN).
No cells, no handovers: all APs jointly transmit/receive for every UE.

Key advantages over cellular:

Macro-diversity: UE sees $L$ spatially-distributed paths — built-in redundancy.
No cell edge: nearest AP is always close; smooth coverage.
Interference becomes useful signal: what was inter-cell interference in cellular is now joint-beamforming signal.
Mobility handling: no handover — the CPU reassigns APs dynamically as UEs move.

Costs:

Heavy fronthaul bandwidth (APs forward all channel estimates and raw symbols to CPU).
Synchronization across $L$ APs (GNSS-PPS or PTP).
CPU compute scales as $L \cdot K$ .

,

Definition:
Cell-Free OTFS Channel Model

For UE $k$ at position $\mathbf{r}_k$ and AP $l$ at position $\mathbf{r}_l$ , the DD-domain channel vector is $\mathbf{H}^{(l, k)}[\ell, m] \;=\; \sum_{i=1}^{P_{l,k}} a_i^{(l,k)}\, \mathbf{a}_l(\theta_i)\, \delta[\ell - \ell_i^{(l,k)}, m - m_i^{(l,k)}]\, e^{-j 2\pi \nu_i^{(l,k)} \ell_i^{(l,k)}/(MN)},$ where:

$P_{l,k}$ is the number of paths between UE $k$ and AP $l$ .
$a_i^{(l,k)}$ is the gain of path $i$ (complex, distance-dependent).
$\ell_i^{(l,k)}, m_i^{(l,k)}$ are integer delay/Doppler indices.
$\mathbf{a}_l(\theta)$ is the AP array response (if $N_a > 1$ ).

The aggregate transmit signal from all APs to UE $k$ is $\mathbf{y}_k[\ell, m] \;=\; \sum_{l=1}^{L} \mathbf{H}^{(l, k)} \cdot \mathbf{v}^{(l, k)} s_k[\ell, m] \,+\, \text{interference} + \mathbf{w}_{k},$ where $\mathbf{v}^{(l, k)}$ is the precoder at AP $l$ for UE $k$ .

Macro-Diversity in the DD Domain

Each AP $l$ sees the UE's physical paths from its own geometric vantage point. The DD channel $\mathbf{H}^{(l, k)}$ has different $(\tau, \nu, \theta)$ triples for different APs — the same UE's motion creates different Doppler at different APs (angle-dependent radial velocity). The aggregate $\{\mathbf{H}^{(l, k)}\}_{l=1}^L$ is an $L$ -fold DD channel with rich spatial-temporal diversity that no single AP could capture. This is the geometric foundation of the 35% throughput gain.

Theorem: Cell-Free OTFS Capacity Scaling

The sum rate of cell-free OTFS with $L$ APs, $K$ UEs, and conjugate beamforming scales as $R_{\text{sum}} \;\geq\; K \log\!\left(1 + \frac{L \cdot P \cdot N_a \cdot \bar\alpha \cdot \text{SNR}}{K + (\text{interference})}\right),$ where $\bar\alpha$ is the average path gain, and the interference term depends on user separation (pilot contamination + spatial overlap).

Consequence: At fixed $K$ , increasing $L$ improves SINR linearly (macro-diversity gain). Compared to cellular (where only the serving BS contributes): $\sim L/1 = L\times$ gain in signal power, ameliorated by pilot contamination.

At $L = 64$ , $K = 16$ , $N_a = 4$ : effective SINR gain $\sim 256$ (24 dB) over single-BS. 30% of this translates to useful rate (rest to reduced interference): $\sim 7$ dB effective rate gain.

In cellular, the serving BS contributes one signal path; all other BSs contribute interference. In cell-free, every AP contributes signal (scaled by the effective channel between that AP and the UE). The aggregate is much larger than any single AP's contribution. Under mobility, the aggregate is also more stable — even if one AP's link degrades, others compensate. The 30% throughput gain at the 95%-likely point comes from this stability.

Proof

Per-AP SNR

For UE $k$ at distance $d_{l,k}$ from AP $l$ : $\mathrm{SNR}_{l,k} \propto d_{l,k}^{-\alpha}$ for path-loss exponent $\alpha = 3$ - $4$ .

Aggregate SINR

Sum across all APs: $\mathrm{SINR}_k = \sum_l \mathrm{SNR}_{l,k} / \sigma^2$ . For large $L$ : dominated by nearest APs ( $\sim \sqrt{L}$ of them).

Rate formula

Shannon: $R_k = \log(1 + \mathrm{SINR}_k)$ . Summed over UEs: $R_{\text{sum}} = K R_k$ (fair case).

Macro-diversity

95th percentile of SINR scales much more favorably than 10th percentile. Cell-edge UEs get most benefit: $\sim L\times$ vs cellular. $\blacksquare$

,

Key Takeaway

Cell-free OTFS fixes the cell-edge problem. The 95%-likely per- user throughput — the rate guaranteed to 95% of users — is where cellular systems suffer most (cell-edge penalty). Cell-free OTFS achieves $\sim 35\%$ improvement here: the CommIT contribution of Mohammadi-Ngo-Matthaiou-Caire.

Definition:
User-Centric AP Clustering

User-centric clustering: each UE is served by a subset of APs — those with significant channel quality — rather than all $L$ APs. Typical cluster size: $L_k = 5$ - $10$ APs per UE.

Scalability: per-UE fronthaul bandwidth is $\mathcal{O}(L_k)$ , not $\mathcal{O}(L)$ . For $L = 1000$ , $L_k = 8$ : $125\times$ fronthaul reduction.

Selection criterion: APs with $\|\mathbf{H}^{(l, k)}\|^2 > \gamma \max_l \|\mathbf{H}^{(l, k)}\|^2$ for threshold $\gamma \in [0.1, 0.5]$ .

Dynamic reconfiguration: cluster is re-evaluated every 100-1000 frames as UE moves. In OTFS, this is natural: the DD-channel's path strengths stabilize over time, giving the CPU clear criteria for AP selection.

Example: Urban Cell-Free OTFS: Architecture Scale

A dense urban deployment: square kilometer with $L = 100$ APs (each $N_a = 4$ antennas), $K = 200$ UEs. 28 GHz mmWave.

(a) Compute total antenna count. (b) Estimate per-UE SINR advantage over cellular (~5 BS cells). (c) State fronthaul bandwidth with user-centric clustering ( $L_k = 6$ ).

Solution

Total antennas

$L \cdot N_a = 100 \cdot 4 = 400$ antennas distributed across 1 km². Compare to 4-sector macro BS: $\sim 64$ antennas total. $6\times$ more spatial resources.

SINR advantage

Per-UE average signal power: sum over nearest-10 APs vs single BS. Signal gain: $\sim 10\times$ . Cellular interference: reduced by coordinated transmission. Net SINR gain: $\sim 10$ dB for cell-edge UEs.

Fronthaul bandwidth

Per-UE: $L_k = 6$ APs × per-AP DD-channel = 6 × (MN × 16 bytes) = ~50 kB per UE per frame. At 100 Hz frame rate, 200 UEs: 1 GB/s aggregate fronthaul. Comfortable for eCPRI (10 GbE).

Summary

Dense cell-free OTFS deployable with existing 5G fronthaul infrastructure. Main cost: AP hardware + CPU compute. Gain: 6-10 dB at cell edge, 30% throughput improvement under mobility.

Cell-Free OTFS Coverage Map

2D visualization: spatial SNR distribution across a service area for cellular (1 BS) vs cell-free (L APs). Shows the smoothed, cell-edge-free coverage pattern of cell-free.

Parameters

L

APs25

Grid size50

Ref SNR (dB)20

🔧Engineering Note

Cell-Free Deployment (2024-2028)

Cell-free massive MIMO deployment status:

2019-2021: Academic testbeds (Linköping University, KU Leuven, Princeton). Proof-of-concept at small scale ( $L = 10$ - $30$ APs).
2022-2024: Industry trials (Ericsson, Nokia, Mavenir). O-RAN community standardizes fronthaul for cell-free architectures. Still OFDM-based; OTFS prototype at CommIT/ Mohammadi 2023.
2024-2028: Gradual commercial rollout. Initial deployments in dense stadium, airport, conference hall scenarios. 5G NR Rel. 18+ supports cell-free mode.
2028+: Cell-free OTFS for mobile scenarios (6G). Combined with LEO integration (Ch. 18) for global coverage.

Deployment barriers:

Fronthaul cost: eCPRI requires fiber + 10-25 GbE per AP. ~ $10k-$ 50k per AP deployment. Subsidy from operator.
Synchronization: PTP-1588v2 or GNSS-PPS with sub- $\mu$ s accuracy. Feasible but additional complexity.
CPU compute: $L \cdot K$ scaling. For $L = 100$ , $K = 200$ : $\sim 10^5$ operations per frame. Modern server CPUs handle it.

Practical Constraints

•
Fronthaul: eCPRI + fiber, $10-50k per AP
•
Synchronization: PTP / GNSS-PPS
•
CPU compute: O(LK) per frame
•
Initial deployments 2024-2028

,

Common Mistake: Too Dense Is Worse Than Too Sparse

Mistake:

Assuming that more APs always improves performance. Beyond a certain density, pilot contamination (same pilot used by different UEs, received by same AP) limits rate. Very dense deployments can actually degrade performance.

Correction:

Design for AP density / UE density ratio $L/K \approx 5$ - $10$ . Below this, insufficient diversity. Above, pilot contamination bottlenecks. Use longer pilot sequences or pilot reuse schemes to mitigate. At the CommIT 35% gain operating point: $L/K \approx 5$ - $6$ for urban scenarios.

Why This Matters: Global Coverage: From Terrestrial to Satellite

Cell-free OTFS extends nicely across terrestrial networks. Chapter 18 takes the next step: LEO satellite constellations providing global coverage. A LEO constellation is, in effect, a "cell-free network in the sky" — hundreds of satellites, each seen by many UEs, coordinated from ground stations. The CommIT contribution of Buzzi-Caire-Colavolpe extends the DD-domain treatment to this orbital scale, where Doppler reaches extremes and OFDM cannot compete.

Cell-Free Architecture and the DD Channel