Connection to Cell-Free Massive MIMO

From Coded Caching to 6G Architecture

Multi-server coded caching is the information-theoretic baseline for cell-free massive MIMO (CF-mMIMO) with edge caching — a leading candidate architecture for 6G. In this vision:

Many "access points" (APs) are deployed, each with modest antenna count and local storage.
Users are not tied to a single AP; all nearby APs jointly serve.
Pre-placement populates AP caches off-peak.
Delivery uses cooperative precoding + coded XOR multicast.

This combines all the machinery of Chapters 5 (multi-antenna), 8 (cloud-RAN), and 9 (multi-server). The CommIT group has positioned fog massive MIMO — their specific version of CF-mMIMO with caching — as the flagship research program for 6G content delivery.

Definition:
Cell-Free Massive MIMO with Caching

The cell-free massive MIMO (CF-mMIMO) caching architecture consists of:

$S$ access points (APs), each with $L$ antennas and cache $M$ files.
A central processing unit (CPU) coordinating APs.
$K$ users scattered across the coverage area.
Per-user channel $\mathbf{h}_k = [\mathbf{h}_{k,1}; \ldots; \mathbf{h}_{k,S}] \in \mathbb{C}^{SL}$ (stacked across APs).

The central parameters: $t = S M/N$ (aggregate caching gain, one "cache copy" per AP). $L_\text{eff} = S L$ (cooperative antennas). Total users supported at full DoF: $\min(t + L_\text{eff}, K)$ .

The model differs from Chapter 8's C-RAN in that: (a) No fronthaul constraint — APs are assumed to share all information via the CPU. (b) Spatial gain scales with $S \cdot L$ , not just a single transmitter.

Theorem: CF-mMIMO Caching DoF

For the CF-mMIMO caching architecture with $S$ APs, $L$ antennas per AP, per-AP cache $M$ , and $K$ users: $\mathrm{DoF}(M) \;=\; \min(S M/N + S L,\; K).$ Here the aggregate caching gain is $t = S M/N$ (not the MAN $KM/N$ , because caches are per-AP, not per-user).

AP caches pool with user demands via cooperative precoding. The aggregate cache budget is $S M$ files; the MAN structure applies across this aggregate cache. Combined with cooperative MIMO (DoF $= SL$ ), DoF sums to $SM/N + SL$ .

Proof

Achievability

Run MAN placement on $S$ caches (one per AP). Each AP caches a distinct subset of $\binom{S}{t}$ subfiles per file (with $t = SM/N$ ). Deliver via cooperative Lampiris-Caire with $L_\text{eff} = SL$ .

DoF

Standard Lampiris-Caire argument: sum-DoF $= t + SL$ where $t$ is the relevant aggregate caching gain (here $SM/N$ ).

Converse

Cut-set with $S$ APs × $L$ antennas = $SL$ independent data streams per channel use, plus caching gain $SM/N$ . $\blacksquare$

Example: A 6G Fog-mMIMO Scenario

A 6G fog-mMIMO deployment: $S = 20$ APs, $L = 4$ antennas each, per-AP cache $M = 1000$ files, library $N = 10^4$ , serving $K = 100$ users.

Solution

Aggregate caching gain

$t = SM/N = 20 \cdot 1000/10^4 = 2$ .

Spatial gain

$SL = 20 \cdot 4 = 80$ .

DoF

$\mathrm{DoF} = \min(t + SL, K) = \min(82, 100) = 82$ . Per-user DoF = 82/100 = 0.82.

Comparison to single-cell

Single-cell: 1 AP with L=4 antennas; no caching-user-per-AP benefit. DoF = 4. Per-user = 4/100 = 0.04. CF-mMIMO with caching: 20× higher per-user DoF. The aggregate cache + cooperation is transformative at scale.

Design levers

DoF scales with $SL + SM/N$ . To double DoF: double $S$ (more APs) or double $M$ (more cache per AP). Both are deployment options in the fog-mMIMO vision.

Per-Server Load vs User Population

Per-server delivery load as the user population scales. With balanced cooperation, load is distributed evenly across $S$ servers; with no cooperation, individual server load scales with $K$ directly. The CF-mMIMO sweet spot is in the balanced regime.

Parameters

Servers S4

Memory ratio M/N0.25

Key Takeaway

Fog massive MIMO = CF-mMIMO + caching generalizes all previous chapters. The DoF formula $t + SL$ combines per-server caching contribution ( $SM/N$ ) with cooperative spatial multiplexing ( $SL$ ). Scaling to 6G dense networks (20+ APs): both terms can contribute substantially to per-user DoF.

Fog mMIMO vs. Cloud-RAN: What's Different?

Chapter 8's C-RAN model and Chapter 9's CF-mMIMO look similar at first but differ in essence:

Property	C-RAN (Ch 8)	CF-mMIMO (Ch 9)
Processing	Centralized at cloud	Distributed across APs
Fronthaul	Bottleneck (NDT metric)	Not explicitly modeled
APs	Thin radio-only	Active participants
Primary metric	NDT (latency)	DoF (throughput)
Caching gain	Aggregate $N_\text{EN} M/N$	Aggregate $SM/N$

Both are legitimate cache-aided wireless architectures. The theoretical essence is the same — cache + antenna cooperation. The deployment differences matter operationally: C-RAN optimizes for low-fronthaul regimes; CF-mMIMO optimizes for dense cooperation without fronthaul constraints.

Historical Note: The Path to Fog Massive MIMO

2014–2025

The multi-server coded caching idea evolved through several stages:

~2014: Multi-cell caching studied with non-cooperative schemes. Rate gains over single-cell noted.
~2016: Cooperation + caching joint analysis (Bölcskei et al.). Linear DoF gain from cooperation established.
~2017: Lampiris-Caire et al. establish $\mathrm{DoF} = t + L$ for single-transmitter; cooperative extension immediate via effective antenna count.
~2019: Parrinello-Unsal-Elia (2020 Trans. IT) comprehensive multi-cache analysis; shared vs dedicated comparisons.
~2022+: Cell-free mMIMO + caching: CommIT "fog mMIMO" program (Lampiris-Bhattacharjee-Caire 2023).
~2025: 6G pre-standardization: cell-free architecture is a leading candidate; coded caching extensions under active research.

The trajectory shows the field becoming increasingly practical: from abstract information theory to deployable architectures. The challenge now is reducing subpacketization, fronthaul, and CSIT requirements to make the theory realizable at scale. Chapter 14 (subpacketization) and Chapter 8 (NDT) address these in detail.

Common Mistake: Caching Gain in CF-mMIMO Is $SM/N$ , Not $KM/N$

Mistake:

Applying the MAN formula $t = KM/N$ to the CF-mMIMO + caching architecture where caches are at APs.

Correction:

When caches are at $S$ APs (not $K$ users), the relevant aggregate cache is $SM$ , not $KM$ . Caching gain $t = SM/N$ . This is typically smaller than the user-cache version for realistic $S \ll K$ . The compensation is the spatial multiplexing gain $SL$ , which can be much larger.

The correct DoF for CF-mMIMO + AP caching: $SM/N + SL$ . For user caching at $K$ mobiles: $KM/N + SL$ (caches at users provide per-user XOR decoding, caches at APs don't).

The Multi-Server MAN Scheme Chapter Summary