Ferkans — Interactive Telecom Tutor

From One Transmitter to Many

Chapters 5–6 treated a single $L$ -antenna transmitter. Chapter 8's C-RAN model had $N_\text{EN}$ cooperating ENs but was primarily about the fronthaul bottleneck between cloud and edge. This chapter zooms in on the multi-server problem: several independent transmitters (base stations), each with the full library, all cooperating to serve users.

The key question: how does the coded caching gain $t$ scale when there are $S$ cooperating servers? The answer, roughly speaking, is $\mathrm{DoF} = t + S L$ — the caching gain still adds linearly, and the spatial gain multiplies by the number of cooperating servers. This is effectively the Lampiris-Caire result with $L_\text{eff} = SL$ , under cooperative assumptions.

This is the theoretical foundation of cell-free massive MIMO + coded caching — the architecture where coded caching meets the most sophisticated wireless cooperation framework.

Definition:
Multi-Server Coded Caching Network

The multi-server coded caching network consists of:

$S$ servers (base stations), each with full access to the library $\mathcal{W} = \{W_1, \ldots, W_{N}\}$ and $L$ transmit antennas.
$K$ single-antenna users, each with a cache of size $M$ files.
A cooperative wireless channel: user $k$ receives a superposition of signals from all $S$ servers.

The received signal at user $k$ , channel use $m$ : $y_k[m] \;=\; \sum_{i=1}^S \mathbf{h}_{k,i}^H \mathbf{x}_i[m] + \mathbf{w}_{k}[m],$ where $\mathbf{x}_i \in \mathbb{C}^L$ is server $i$ 's transmit signal. Per-server power constraint: $\mathbb{E}[\|\mathbf{x}_i\|^2] \leq P$ . Full cooperation means the $S$ servers jointly design $\{\mathbf{x}_i\}$ to achieve the desired beamforming pattern.

The cooperating-servers assumption gives the strongest scheme. In practice, server cooperation requires low-latency inter-server communication (via fiber or dedicated backhaul). Limited-cooperation variants relax this and generally achieve less than the full cooperative DoF.

Theorem: Multi-Server DoF under Full Cooperation

For the multi-server coded caching network with $S$ cooperating $L$ -antenna servers, integer $t = K M/N$ : $\mathrm{DoF}(M) \;=\; \min(t + S L,\; K).$ The DoF is the Lampiris-Caire result with effective antenna count $L_\text{eff} = SL$ .

Under full cooperation, the $S$ servers act as a single transmitter with $SL$ antennas (joint precoding). Chapter 5's DoF formula applies directly with $L \to SL$ . The caching gain $t$ is unchanged — caches are per-user, independent of the transmitter count.

Proof

Achievability

Consolidate the $S$ servers into a single virtual transmitter with $SL$ antennas. Apply the Lampiris-Caire scheme (Ch 5) directly: DoF = $t + SL$ . The precoding is cooperative — each server's beam is a coordinate projection of the virtual beam.

Converse

Cut-set argument: the $S$ servers collectively output $SL$ DoF's worth of signal per channel use. Caches contribute $t$ DoF via XOR multicast. Total: $\min(t + SL, K)$ by the usual cut-set bound with $K$ as the saturation ceiling.

Tight match

Upper and lower bounds coincide. $\blacksquare$

Multi-Server DoF: Effect of Cooperation

DoF as a function of memory ratio, for varying number of cooperating servers $S$ . Blue solid: $S$ servers cooperating (DoF = $t + SL$ ). Red dashed: single server (DoF = $t + L$ ). The DoF gain from cooperation is $SL - L = (S-1)L$ — proportional to the cooperation scale.

Parameters

Users K20

Servers S2

Antennas per server L2

Example: Multi-Server DoF Computation

For $K = 30$ , $N = 60$ , $M = 10$ , $S = 3$ servers each with $L = 4$ antennas, compute the DoF under full cooperation.

Solution

Compute

$t = KM/N = 30 \cdot 10/60 = 5$ . $\mathrm{DoF} = \min(5 + 3 \cdot 4, 30) = \min(17, 30) = 17$ .

Decomposition

Caching gain $t = 5$ . Aggregate spatial gain $SL = 12$ . Sum: 17. Per-user: 17/30 ≈ 0.57 DoF.

Comparison

Single server with $L = 4$ : DoF = $t + L = 9$ . Three cooperating servers give DoF = 17 — almost doubling throughput. Critically, the caching contribution of $t = 5$ is unchanged; what multiplies by 3 is the spatial gain.

Multi-Server Cooperative Caching Topology

Three cooperating servers each with

L

antennas; users each with cache

\mathcal{Z}_k

. Full cooperation (dashed lines) provides effective

SL

-antenna joint precoding. DoF =

\min(t + SL, K)

generalizes the Lampiris-Caire formula to multi-transmitter networks.

Key Takeaway

Full server cooperation multiplies the spatial gain by $S$ ; the caching gain stays unchanged. DoF = $t + SL$ . This is the Lampiris-Caire result with effective $L = SL$ . Caching gain is a per-user property — multiplying transmitters does not multiply it. Practical deployments balance server cooperation (expensive — requires inter-server backhaul) against caching (cheap — local storage).

Cooperation Is Not Free

Full server cooperation has costs:

Low-latency backhaul. To jointly precode, servers must exchange user data in real time — microsecond-scale backhaul. Fiber between co-located servers is feasible; across a wide area it is not.
Shared CSI. All servers need accurate channel knowledge for every user; CSI is centrally aggregated and distributed. Chapter 7's pilot overhead analysis applies to the aggregate $SL$ - antenna virtual transmitter.
Clock synchronization. Phase-coherent transmission requires microsecond-scale clock sync across servers.

Realistic deployments often use partial cooperation: e.g., cluster servers into groups of 2-4 that cooperate fully, then time-division between clusters. This reduces backhaul and sync requirements while capturing most of the cooperation gain.

The CommIT group has studied this partial-cooperation regime extensively in the cell-free massive MIMO context (Lampiris- Bhattacharjee-Caire 2022+).

Multi-server coded caching topology — $S = 3$ cooperating servers, each with $L$ antennas. Users each have cache $\mathcal{Z}_k$ . Servers exchange data via inter-server backhaul (dashed lines). Wireless downlink is MIMO BC with $S L$ effective cooperating antennas.

⚠️Engineering Note

Cell-Free Massive MIMO Is the Deployment Target

The multi-server coded caching model maps directly to cell-free massive MIMO (CF-mMIMO), an emerging 6G architecture:

Many access points (APs). Instead of a few large base stations, CF-mMIMO uses dozens to hundreds of small APs spread through a coverage area.
Central processing unit (CPU). APs are connected to a CPU via fronthaul; CPU orchestrates cooperation.
Each AP has modest antennas ( $L \in [2, 16]$ ). Aggregate: $S \cdot L \in [100, 1600]$ .
Each AP can cache content. Pre-placement populates AP caches; cooperative delivery exploits combined storage.

In this architecture, Chapter 9's multi-server model is the information-theoretic baseline. The caching gain $t$ is practically a function of aggregate storage across APs; the spatial gain $SL$ is the number of cooperating APs times their local antennas. Both scale to the hundreds in aggressive CF-mMIMO designs.

Deployment path: 6G fog-based massive MIMO (2025+), which the CommIT group has prototyped extensively.

Practical Constraints

•
CF-mMIMO: 20-100 APs per coverage area, each with 2-8 antennas
•
Inter-AP backhaul: microsecond latency required for full cooperation
•
CSI aggregation at CPU: 10-100 ms periodicity
•
Per-AP cache: 10 GB - 1 TB typical for 6G prototypes

The Multi-Server Coded Caching Model