Ferkans — Interactive Telecom Tutor

From C-RAN to Cell-Free: Removing Cell Boundaries

Cloud-RAN (Section 25.2) centralizes processing but still assigns users to specific APs or clusters. The cell-free paradigm takes this one step further: there are no cells. Every AP in the network coherently serves every user, and the notion of a "cell edge" disappears entirely.

The appeal is both theoretical and practical. Theoretically, cell-free massive MIMO achieves uniformly good service across the network: no user suffers from being at the cell boundary. Practically, it can be implemented with simple conjugate beamforming at each AP and limited fronthaul, provided we accept some loss compared to fully centralized processing.

The information-theoretic question is: what rates are achievable with local processing at the APs and limited fronthaul? The answer reveals a rich tradeoff between decentralization and performance.

Cell-Free vs Cellular Network Topology

Side-by-side comparison: traditional cellular layout with cell edges where users suffer, versus cell-free massive MIMO where distributed APs jointly serve all users without boundaries.

Definition:
Cell-Free Massive MIMO System Model

Consider a network with $L$ single-antenna APs and $K$ single-antenna users ( $L \gg K$ ), all operating in the same time-frequency resource. The uplink received signal at AP $l$ is:

$Y_l = \sum_{k=1}^{K} g_{kl} X_k + Z_l$

where $g_{kl} = \sqrt{\beta_{kl}}\, h_{kl}$ is the channel from user $k$ to AP $l$ , $\beta_{kl}$ is the large-scale fading coefficient (path loss and shadowing), $h_{kl} \sim \mathcal{CN}(0, 1)$ is the small-scale fading, and $Z_l \sim \mathcal{CN}(0, \sigma^2)$ .

Each AP $l$ computes a local estimate of each user's signal using only its own observation and local CSI:

$\hat{s}_{kl} = a_{kl}^* Y_l$

where $a_{kl}$ is the local combining coefficient (e.g., matched filter: $a_{kl} = \hat{g}_{kl}$ , the channel estimate from uplink pilots).

The local estimates are sent to the CP, which forms the final estimate:

$\hat{s}_k = \sum_{l=1}^{L} \hat{s}_{kl} = \sum_{l=1}^{L} a_{kl}^* Y_l.$

Cell-free massive MIMO

A network architecture where a large number of distributed access points coherently serve all users without cell boundaries, with coordination through a central processor.

Related: Cloud-RAN, Fronthaul

Theorem: Achievable Uplink Rate with Local Processing

Under matched-filter combining at each AP ( $a_{kl} = \hat{g}_{kl}$ , where $\hat{g}_{kl}$ is the MMSE channel estimate), the uplink achievable rate for user $k$ is:

$R_k = \log\!\left(1 + \frac{p_k \left(\sum_{l=1}^{L} \gamma_{kl}\right)^2} {\sum_{k'=1}^{K} p_{k'} \sum_{l=1}^{L} \gamma_{kl} \beta_{k'l} + \sigma^2 \sum_{l=1}^{L} \gamma_{kl}}\right)$

where $p_k$ is user $k$ 's transmit power, $\gamma_{kl} = \mathbb{E}[|\hat{g}_{kl}|^2]$ is the mean-square channel estimate quality, and the expectation is over small-scale fading.

The numerator is the squared coherent combining gain: all $L$ APs' estimates add constructively for the desired user. The denominator has two terms: (1) interference from other users, which grows only as $\sum_l \gamma_{kl}\beta_{k'l}$ (not as the square), and (2) noise, which also grows linearly in $L$ . The key observation is that the signal grows as $L^2$ while interference and noise grow as $L$ , so the SINR grows linearly in $L$ . This is the channel hardening effect of massive MIMO, and it holds even with distributed APs and local processing.

Proof

Use-and-then-forget bound

The proof uses the "use-and-then-forget" (UatF) technique: treat the channel estimates as deterministic and the estimation errors as additional noise. The received signal after combining is:

$\hat{s}_k = \underbrace{\sqrt{p_k} \sum_{l=1}^L |\hat{g}_{kl}|^2}_{\text{desired signal}} X_k + \underbrace{\sqrt{p_k} \sum_{l=1}^L \hat{g}_{kl}^* \tilde{g}_{kl}}_{\text{estimation error}} X_k + \sum_{k' \ne k} \sqrt{p_{k'}} \sum_{l=1}^L \hat{g}_{kl}^* g_{k'l} X_{k'} + \text{noise}$

where $\tilde{g}_{kl} = g_{kl} - \hat{g}_{kl}$ is the estimation error.

Worst-case uncorrelated noise

Taking the expectation of the desired signal coefficient and treating the remainder as worst-case uncorrelated Gaussian noise (which is the standard technique from Hassibi and Hochwald, 2003), we obtain:

$\text{SINR}_k = \frac{p_k \left(\mathbb{E}\!\left[\sum_{l=1}^L |\hat{g}_{kl}|^2\right]\right)^2} {\text{Var}\!\left[\sum_{l=1}^L \hat{g}_{kl}^* g_{kl}\right] p_k + \sum_{k' \ne k} p_{k'} \mathbb{E}\!\left[|\sum_{l=1}^L \hat{g}_{kl}^* g_{k'l}|^2\right] + \sigma^2 \sum_{l=1}^L \gamma_{kl}}.$

Evaluating expectations

Using $\mathbb{E}[|\hat{g}_{kl}|^2] = \gamma_{kl}$ and the independence of channel estimates across different user-AP pairs:

Desired signal power: $p_k (\sum_l \gamma_{kl})^2$
Self-interference (estimation error): $p_k \sum_l \gamma_{kl}(\beta_{kl} - \gamma_{kl})$
Inter-user interference: $\sum_{k' \ne k} p_{k'} \sum_l \gamma_{kl} \beta_{k'l}$
Noise after combining: $\sigma^2 \sum_l \gamma_{kl}$

Combining all interference and noise terms yields the stated rate expression.

Example: Scaling Laws in Cell-Free Massive MIMO

Consider a cell-free system with $K = 10$ users and $L$ APs, where all large-scale fading coefficients are equal: $\beta_{kl} = \beta$ for all $k, l$ . All users transmit at equal power $p_k = p$ . The channel estimation quality is $\gamma_{kl} = \gamma = \beta \tau_p p / (\tau_p p \beta + \sigma^2)$ where $\tau_p$ is the pilot length.

(a) Show that the per-user rate grows as $\log(L)$ for large $L$ .

(b) How many APs are needed to achieve $R_k = 5$ bits/use per user when $\beta = 0.01$ , $p = 100$ mW, $\sigma^2 = -90$ dBm, and $\tau_p = 10$ ?

Solution

Simplify the rate expression

With equal coefficients, the SINR for any user becomes:

$\text{SINR}_k = \frac{p (L\gamma)^2}{Kp \cdot L\gamma\beta + \sigma^2 L\gamma} = \frac{p L\gamma}{Kp\beta + \sigma^2} = \frac{pL\gamma}{Kp\beta + \sigma^2}.$

For large $L$ , $\text{SINR}_k \propto L$ , so $R_k \approx \log(L) + \text{const}$ , confirming that the per-user rate grows logarithmically with the number of APs.

Compute the required number of APs

We need $\text{SINR}_k = 2^{R_k} - 1 = 2^5 - 1 = 31$ . With the given parameters:

$\sigma^2 = 10^{-12}$ W ( $-90$ dBm)
$p = 0.1$ W
$\gamma = \beta \cdot 10 \cdot 0.1 / (10 \cdot 0.1 \cdot 0.01 + 10^{-12}) \approx 0.01$ (since $\tau_p p\beta \gg \sigma^2$ )
$Kp\beta + \sigma^2 = 10 \times 0.1 \times 0.01 + 10^{-12} = 0.01$ W

Therefore $L = 31 \times 0.01 / (0.1 \times 0.01) = 31$ . We need approximately $L = 31$ APs for $K = 10$ users to achieve 5 bits/use per user, a ratio of about 3 APs per user.

Definition:
User-Centric Cell-Free Architecture

In the user-centric approach, each user $k$ is served by only a subset of APs $\mathcal{L}_k \subset [L]$ , typically those with the strongest large-scale fading coefficients $\beta_{kl}$ . Formally:

$\mathcal{L}_k = \{l \in [L] : \beta_{kl} \ge \beta_{\text{th}}\}$

or the $N_{\text{AP}}$ APs with largest $\beta_{kl}$ . The combining weights are:

$a_{kl} = \begin{cases} \hat{g}_{kl} & \text{if } l \in \mathcal{L}_k \\ 0 & \text{otherwise} \end{cases}$

The user-centric approach dramatically reduces the fronthaul load (only $|\mathcal{L}_k|$ APs forward estimates for user $k$ ) at the cost of reduced macro-diversity.

Cell-Free vs Co-Located Massive MIMO

Compare the 95th-percentile (cell-edge) rate of cell-free massive MIMO with co-located massive MIMO having the same total number of antennas. Cell-free provides more uniform service quality by eliminating cell edges.

Parameters

Total antennas

L

64

Number of users

K

16

Area side length (m)500

Path-loss exponent

\alpha

3.5

Theorem: Cell-Free Massive MIMO Has Unlimited Capacity

Consider a cell-free massive MIMO system with $L$ APs and $K$ users distributed over a growing area, with spatial correlation modeled by a stationary ergodic random field. Under matched-filter processing with MMSE channel estimation, as $L \to \infty$ with $K/L \to \alpha \in (0, 1)$ and the AP density held fixed:

$R_{k} \xrightarrow{L \to \infty} \infty \quad \text{for each user } k,$

provided the spatial correlation function $\rho(d) = \mathbb{E}[\beta_{kl} \beta_{kl'}]$ satisfies $\sum_{l'} \rho(\|l - l'\|) < \infty$ (the correlation is summable).

More precisely, the per-user rate scales as:

$R_{k} = \Theta(\log L)$

which is the same scaling as co-located massive MIMO with $L$ antennas.

This result says that cell-free massive MIMO, even with simple matched-filter processing at each AP, achieves the same capacity scaling as a co-located array with the same total number of antennas. The "unlimited capacity" comes from the channel hardening: as we add more APs, the effective channel to each user becomes increasingly deterministic, and the interference averages out. The summable correlation condition ensures that distant APs contribute diminishing but non-zero gain, so the coherent combining gain grows without bound.

Proof

Signal power scaling

The desired signal power for user $k$ is $p_k (\sum_{l=1}^L \gamma_{kl})^2$ . Since $\gamma_{kl} \ge c\beta_{kl}$ for some constant $c > 0$ (depending on pilot power and noise), and $\sum_l \beta_{kl} \to \infty$ as $L \to \infty$ (because the AP density is fixed and the area grows), the signal power grows at least as $L^2$ .

Interference power scaling

The interference power is $\sum_{k'} p_{k'} \sum_l \gamma_{kl}\beta_{k'l}$ . For fixed $K/L$ ratio, this grows as $L$ (each AP contributes $O(1)$ interference, and there are $L$ APs). The noise term also grows as $L$ .

SINR scaling

Combining: $\text{SINR}_k = \Theta(L^2 / L) = \Theta(L)$ . Therefore $R_k = \log(1 + \Theta(L)) = \Theta(\log L) \to \infty$ .

The summable correlation condition is needed to ensure that the variance of the channel hardening estimate concentrates, so the use-and-then-forget bound remains tight.

🎓CommIT Contribution(2018)

Massive MIMO Has Unlimited Capacity

G. Caire — IEEE Transactions on Wireless Communications

Caire showed that under spatially correlated channels with summable correlation, cell-free massive MIMO with conjugate beamforming achieves per-user rates that grow as $\Theta(\log L)$ , the same scaling as co-located massive MIMO. This resolved a question about whether distributing antennas incurs a fundamental capacity penalty compared to co-locating them. The answer is no: the macro-diversity gain of distribution compensates for the loss of coherent array gain, provided the number of APs grows while maintaining spatial correlation.

massive-MIMOcell-freecapacity-scalingchannel-hardeningView Paper →

Historical Note: From Distributed Antennas to Cell-Free

2013-2019

The idea of distributing antennas across a coverage area dates to the 1980s distributed antenna systems (DAS). The modern cell-free massive MIMO concept was crystalized by Ngo, Ashikhmin, Yang, Larsson, and Marzetta in 2017, who combined Marzetta's massive MIMO vision (many antennas serving few users) with the distributed antenna concept. The term "cell-free" was deliberately chosen to emphasize the elimination of cell boundaries — the bane of cellular systems since their inception. Interestingly, the information-theoretic foundations were already present in the C-RAN literature (Simeone, Somekh, Poor, and Shamai), but it was the scalable matched-filter implementation that made the concept practical. The CommIT group's contribution was showing that this simple processing achieves the same capacity scaling as fully centralized MIMO processing, validating the practical appeal of the cell-free architecture.

Cell-Free: Fronthaul Load vs Number of Serving APs

In user-centric cell-free, each user is served by $N_{\text{AP}}$ nearest APs. Explore the tradeoff between per-user rate and total fronthaul load as $N_{\text{AP}}$ varies from 1 (no cooperation) to $L$ (full cooperation).

Parameters

Total APs

L

64

Number of users

K

16

Transmit SNR (dB)10

Common Mistake: Pilot Contamination in Cell-Free Systems

Mistake:

Assuming that with $L \gg K$ APs, pilot contamination is negligible because there are enough pilots for all users.

Correction:

The coherence interval is limited (typically $\tau_c \sim 200$ in 5G NR at 3.5 GHz with 30 kHz subcarrier spacing and 3 km/h mobility). With orthogonal pilots, at most $\tau_p \le \tau_c$ users can be assigned orthogonal pilots. When $K > \tau_p$ , some users must reuse pilots, causing pilot contamination that persists even as $L \to \infty$ . In the cell-free context, pilot contamination is actually worse than in co-located systems because every AP estimates every user's channel. The mitigation strategies include pilot assignment algorithms that exploit the user-centric structure: users sharing a pilot should have non-overlapping serving AP sets.

Key Takeaway

Cell-free massive MIMO eliminates cell boundaries and provides uniformly good service. With $L$ distributed APs serving $K$ users ( $L \gg K$ ), simple matched-filter processing achieves per-user rates scaling as $\Theta(\log L)$ -- the same as co-located massive MIMO. The user-centric approach, where each user is served by its nearest APs, makes the concept scalable by limiting fronthaul load. The price is pilot contamination management and the need for a central processor to aggregate local estimates.

Channel hardening

The phenomenon where the effective channel gain (after combining across many antennas or APs) concentrates around its mean, making the channel appear nearly deterministic. Formally, $\|\mathbf{H}\|^2 / \mathbb{E}[\|\mathbf{H}\|^2] \to 1$ as the number of antennas grows.

Related: Cell-free massive MIMO

Macro-diversity

Diversity gained by receiving signals from geographically separated access points, which experience independent large-scale fading (shadowing and path loss). Contrasts with micro-diversity (multiple antennas at the same location).

Quick Check

In cell-free massive MIMO with matched-filter processing, the per-user SINR scales as $\Theta(L)$ where $L$ is the number of APs. Why does the rate only grow as $\Theta(\log L)$ and not linearly in $L$ ?

Because of pilot contamination

Because the capacity formula $\log(1 + \text{SINR})$ is logarithmic in SINR

Because fronthaul capacity limits the rate

Correction:

Because the capacity formula

\log(1 + \text{SINR})

is logarithmic in SINR

Since SINR $= \\Theta(L)$ , we get $R = \\log(1 + \\Theta(L)) = \\Theta(\\log L)$ . This is a fundamental property of the additive Gaussian channel, not a limitation of the cell-free architecture.

Cell-Free Massive MIMO from an Information-Theoretic Perspective

From C-RAN to Cell-Free: Removing Cell Boundaries

Cell-Free vs Cellular Network Topology

Definition: Cell-Free Massive MIMO System Model

Cell-free massive MIMO

Theorem: Achievable Uplink Rate with Local Processing

Use-and-then-forget bound

Worst-case uncorrelated noise

Evaluating expectations

Example: Scaling Laws in Cell-Free Massive MIMO

Simplify the rate expression

Compute the required number of APs

Definition: User-Centric Cell-Free Architecture

Cell-Free vs Co-Located Massive MIMO

Parameters

Theorem: Cell-Free Massive MIMO Has Unlimited Capacity

Signal power scaling

Interference power scaling

SINR scaling

Massive MIMO Has Unlimited Capacity

Historical Note: From Distributed Antennas to Cell-Free

Cell-Free: Fronthaul Load vs Number of Serving APs

Parameters

Common Mistake: Pilot Contamination in Cell-Free Systems

Key Takeaway

Channel hardening

Macro-diversity

Quick Check

Definition:
Cell-Free Massive MIMO System Model

Definition:
User-Centric Cell-Free Architecture