Ferkans — Interactive Telecom Tutor

How Good Is Cell-Free Massive MIMO?

We have described the cell-free architecture and its operational protocol. The natural question is: how does it actually perform? We now derive closed-form achievable rate expressions using the use-and-then-forget (UatF) bound (familiar from Chapter 4), and use these to compare cell-free massive MIMO against cellular networks and small cells. The punchline: the 95%-likely per-user rate — the metric that captures cell-edge performance — improves dramatically.

Theorem: Achievable Downlink SINR for Cell-Free Massive MIMO

Under conjugate beamforming with power control coefficients $\{\eta_{lk}\}$ and MMSE channel estimation, the achievable downlink SINR for user $k$ using the UatF bound is

$\text{SINR}_k^{\text{cf}} = \frac{\left(\sum_{l=1}^{L} \sqrt{\eta_{lk}} \, \gamma_{lk}\right)^2}{\sum_{j=1}^{K} \sum_{l=1}^{L} \eta_{lj} \, \gamma_{lj} \, \beta_{lk} + \sum_{j \neq k} \left(\sum_{l=1}^{L} \sqrt{\eta_{lj}} \, \gamma_{lj} \frac{c_{ljk}}{\gamma_{lj}}\right)^2 + \frac{\sigma^2}{P_t}}$

where $c_{ljk} = \frac{\tau_p p_p \, \beta_{lk} \, \beta_{lj}}{\tau_p p_p \sum_{k' \in \mathcal{P}_j} \beta_{lk'} + \sigma^2}$ captures the pilot contamination between users $j$ and $k$ when they share the same pilot.

When pilots are orthogonal ( $\mathcal{P}_k = \{k\}$ for all $k$ ), the pilot contamination term vanishes and the SINR simplifies to

$\text{SINR}_k^{\text{cf}} = \frac{\left(\sum_{l=1}^{L} \sqrt{\eta_{lk}} \, \gamma_{lk}\right)^2}{\sum_{j=1}^{K} \sum_{l=1}^{L} \eta_{lj} \, \gamma_{lj} \, \beta_{lk} + \frac{\sigma^2}{P_t}}$

The achievable rate is $R_k = \frac{\tau_d}{\tau_c} \log_2(1 + \text{SINR}_k^{\text{cf}})$ bits/s/Hz.

The numerator is a coherent macro-diversity gain: the effective channel gains $\gamma_{lk}$ from all $L$ APs add constructively (amplitudes, not powers). The denominator includes non-coherent interference from other users, pilot contamination (if pilots are shared), and noise. The distributed geometry ensures that the numerator is large for every user — not just users near a particular BS — because every user has several nearby APs with large $\gamma_{lk}$ .

Proof

Downlink signal model

User $k$ receives: $y_k = \sum_{l=1}^{L} \mathbf{H}_{lk}^{H} \mathbf{x}_l + w_k = \sum_{l=1}^{L} \sqrt{P_t} \sum_{j=1}^{K} \sqrt{\eta_{lj}} \, \mathbf{H}_{lk}^{H} \hat{\mathbf{H}}_{lj}^* \, s_j + w_k$

Desired signal term

The desired signal coefficient is $\text{DS}_k = \sum_{l=1}^{L} \sqrt{P_t \eta_{lk}} \, \mathbb{E}[\mathbf{H}_{lk}^{H} \hat{\mathbf{H}}_{lk}^*]$ . Since $\hat{\mathbf{H}}_{lk}$ and $\mathbf{H}_{lk}$ are jointly Gaussian with correlation $\mathbb{E}[\mathbf{H}_{lk}^{H} \hat{\mathbf{H}}_{lk}^*] = N \gamma_{lk}$ , we get $\text{DS}_k = N \sqrt{P_t} \sum_{l=1}^{L} \sqrt{\eta_{lk}} \, \gamma_{lk}$ .

UatF bound application

The UatF bound treats the channel estimation error as worst-case Gaussian noise. The effective noise variance includes (i) beamforming gain uncertainty, (ii) inter-user interference, (iii) pilot contamination from shared pilots, and (iv) thermal noise.

SINR expression

Dividing the desired signal power $|\text{DS}_k|^2$ by the total effective noise power (with $N = 1$ for single-antenna APs) yields the stated SINR expression. The key structure is: coherent sum in numerator, non-coherent sums in denominator. $\blacksquare$

Example: Achievable Rate in a Simple Cell-Free Setup

Consider $L = 40$ single-antenna APs and $K = 5$ users with orthogonal pilots. Suppose all APs have equal power control $\eta_{lk} = \eta$ for all $l, k$ , and the estimation qualities are $\gamma_{lk} = \beta_{lk}$ (high pilot SNR). A particular user $k$ has 5 nearby APs with $\beta_{lk} = 10^{-8}$ and 35 distant APs with $\beta_{lk} = 10^{-11}$ . Compute the effective desired signal strength and compare to a cellular system with a single BS at the same distance as the nearest AP.

Solution

Cell-free desired signal (numerator)

The coherent sum: $\sum_{l=1}^{L} \sqrt{\eta} \, \gamma_{lk} \approx 5 \sqrt{\eta} \cdot 10^{-8} + 35 \sqrt{\eta} \cdot 10^{-11}$ . The 5 nearby APs contribute $5 \times 10^{-8} = 5 \times 10^{-8}$ , while 35 distant APs contribute $35 \times 10^{-11} = 3.5 \times 10^{-10}$ . The nearby APs dominate by a factor of $\sim 140$ . The squared sum gives $\sim 25 \eta \cdot 10^{-16}$ .

Cellular desired signal

A single BS with $N_t = 40$ antennas at the same distance gives $N_t \cdot \beta = 40 \times 10^{-8} = 4 \times 10^{-7}$ (linear, not squared). For fair comparison, the cellular desired power is $N_t^{2} \cdot \eta \cdot \beta^{2} = 1600 \eta \cdot 10^{-16}$ .

Comparison

The cell-free system achieves a desired signal power that is $25/1600 \approx 1.6\%$ of co-located massive MIMO for this particular user. But the cell-free advantage is not in peak power — it is in uniformity. Every user has nearby APs, while in cellular, cell-edge users have much weaker signals. The 95%-likely rate favors cell-free.

Theorem: Max-Min Fair Power Control

The max-min fair power control problem for cell-free massive MIMO is:

$\max_{\{\eta_{lk} \geq 0\}} \; \min_{k=1,\ldots,K} \; \text{SINR}_k^{\text{cf}}(\boldsymbol{\eta})$

subject to $\sum_{k=1}^{K} \eta_{lk} \, \gamma_{lk} \leq 1$ for all $l = 1, \ldots, L$ .

This is a quasi-concave optimization that can be solved via bisection: for a target SINR $t$ , we check feasibility of $\text{SINR}_k^{\text{cf}} \geq t$ for all $k$ subject to the per-AP power constraints, which reduces to a second-order cone program (SOCP).

Max-min fairness ensures that the worst user in the network gets the highest possible rate. This is precisely the metric that addresses the cell-edge problem: instead of maximizing average throughput (which favors cell-center users), we maximize the minimum rate. Cell-free massive MIMO is particularly well-suited to max-min fairness because the distributed geometry ensures every user has nearby APs.

Proof

Quasi-concavity

The minimum SINR $\min_k \text{SINR}_k^{\text{cf}}(\boldsymbol{\eta})$ is quasi-concave in $\boldsymbol{\eta}$ because each $\text{SINR}_k^{\text{cf}}$ is a ratio of concave over convex functions in the power coefficients.

Bisection

Binary search on the target $t$ : for each $t$ , check whether there exist $\{\eta_{lk}\}$ such that $\text{SINR}_k^{\text{cf}} \geq t$ for all $k$ and $\sum_k \eta_{lk} \gamma_{lk} \leq 1$ for all $l$ . This feasibility check is an SOCP (linear constraints + SOC constraint from the SINR definition).

Convergence

Bisection converges to the optimal $t^*$ in $O(\log(1/\epsilon))$ iterations, each requiring one SOCP solve. The total complexity is polynomial in $L$ and $K$ . $\blacksquare$

Sum Rate vs AP Density

Explore how the sum rate and 95%-likely per-user rate scale with the number of distributed APs. As AP density increases, both metrics improve, but the 95%-likely rate improves faster because the distributed geometry eliminates coverage holes.

Parameters

K

(users)10

Total Tx power (dBm)40

Path-loss exponent

\alpha

3.8

Area side length (m)1000

Metric

SINR Uniformity: Cellular vs Cell-Free

Compare the 95%-likely per-user rate between cellular, small-cell, and cell-free deployments. The cell-free architecture provides dramatically more uniform service: the ratio of 95th-percentile to 5th-percentile rate is much smaller than in cellular networks.

Parameters

Number of APs/BSs40

K

(users)10

Total antennas (

M

)40

SNR (dB)10

,

Example: Cell-Free vs Small Cells: The Ngo et al. Comparison

Ngo et al. (2017) compare cell-free massive MIMO ( $L = 100$ single-antenna APs) against small cells (100 single-antenna BSs, each serving users in its Voronoi cell) with $K = 40$ users in a $1 \times 1$ km area. Both systems have the same total number of antennas and total transmit power. Explain why cell-free achieves 5--10 times higher 95%-likely per-user throughput.

Solution

Small-cell architecture

In the small-cell system, each BS serves only users in its cell. A cell-edge user is served by a single AP and interfered with by all others. The SINR is limited by inter-cell interference, just as in conventional cellular (albeit with smaller cells and hence shorter distances).

Cell-free architecture

In the cell-free system, every AP serves every user. A "cell-edge user" does not exist because there are no cells. Every user benefits from coherent transmission from all 100 APs. The nearby APs provide macro-diversity gain, while distant APs contribute marginally (their interference is also marginal).

95%-likely rate

The 5th percentile user in the small-cell system is at a cell edge, receiving signal from one AP at large distance and interference from several nearby APs. In the cell-free system, the 5th percentile user still has several nearby APs serving it. The key metric — $\min_k R_k$ under max-min power control — is 5--10 times higher for cell-free because the power control can focus the network's resources on weak users through the $\eta_{lk}$ coefficients.

Theorem: Channel Hardening in Cell-Free Massive MIMO

In a cell-free massive MIMO system with $L$ single-antenna APs serving user $k$ , define the aggregate channel gain as $G_k = \sum_{l=1}^{L} |\mathbf{H}_{lk}|^2 = \sum_{l=1}^{L} \beta_{lk} |g_{lk}|^2$ where $g_{lk} \sim \mathcal{CN}(0,1)$ are independent. Then

$\frac{G_k}{\mathbb{E}[G_k]} \xrightarrow{L \to \infty} 1 \quad \text{in probability}$

provided that $\max_l \beta_{lk} / \sum_{l=1}^{L} \beta_{lk} \to 0$ , i.e., no single AP dominates the aggregate channel. The variance satisfies

$\frac{\text{Var}(G_k)}{(\mathbb{E}[G_k])^2} = \frac{\sum_{l=1}^{L} \beta_{lk}^{2}}{\left(\sum_{l=1}^{L} \beta_{lk}\right)^2} \leq \frac{\max_l \beta_{lk}}{\sum_{l=1}^{L} \beta_{lk}} \to 0$

Channel hardening in cell-free systems works differently from co-located massive MIMO. In co-located systems, many antennas at one location average out the fading fluctuations. In cell-free systems, many APs at different locations average out the fading — but they also average out the shadowing, which co-located arrays cannot do. This is why cell-free channel hardening can be stronger than co-located channel hardening.

Proof

Mean of aggregate channel

$\mathbb{E}[G_k] = \sum_{l=1}^{L} \beta_{lk} \, \mathbb{E}[|g_{lk}|^2] = \sum_{l=1}^{L} \beta_{lk}$ .

Variance of aggregate channel

Since $|g_{lk}|^2 \sim \text{Exp}(1)$ independently, $\text{Var}(G_k) = \sum_{l=1}^{L} \beta_{lk}^{2}$ .

Convergence

The normalized variance is $\frac{\text{Var}(G_k)}{(\mathbb{E}[G_k])^2} = \frac{\sum_l \beta_{lk}^{2}}{(\sum_l \beta_{lk})^2}$ . By the condition $\max_l \beta_{lk} / \sum_l \beta_{lk} \to 0$ , $\sum_l \beta_{lk}^{2} \leq \max_l \beta_{lk} \cdot \sum_l \beta_{lk}$ , so the normalized variance vanishes. By Chebyshev's inequality, $G_k / \mathbb{E}[G_k] \to 1$ in probability. $\blacksquare$

,

🎓CommIT Contribution(2017)

Cell-Free Massive MIMO: The Founding Paper

H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, T. L. Marzetta — IEEE Trans. Wireless Communications, vol. 16, no. 3, pp. 1834–1850

This foundational paper established the cell-free massive MIMO concept and provided the first rigorous performance analysis. The key results include: (i) closed-form achievable rate expressions under conjugate beamforming with imperfect CSI; (ii) max-min fair power control via bisection over SOCPs; and (iii) a comparison against small cells showing 5--10 times improvement in 95%-likely per-user throughput. The paper demonstrated that the simplest possible processing — conjugate beamforming with local MMSE channel estimation — is sufficient to achieve near-uniform service when the number of distributed APs is large. This work launched a major research direction that continues to shape 6G network architecture.

cell-freemassive-mimofounding-paperView Paper →

Common Mistake: Cell-Free Does Not Beat Co-Located at Peak Rate

Mistake:

Assuming that cell-free massive MIMO provides higher peak throughput than co-located massive MIMO with the same total number of antennas.

Correction:

Cell-free trades peak rate for uniformity. A user near a co-located massive MIMO BS with $M$ antennas enjoys array gain $M$ and coherent beamforming. In a cell-free system with the same $M$ antennas distributed as $L$ single-antenna APs, the coherent gain from nearby APs is smaller because only a subset of APs are close. The average rate may be similar, but the 95%-likely (cell-edge) rate is far superior in the cell-free case. Cell-free optimizes the tail of the distribution, not the head.

,

Historical Note: Precursors: Distributed Antenna Systems and Network MIMO

1987–2017

The idea of distributing antennas across a coverage area did not begin with cell-free massive MIMO. Distributed Antenna Systems (DAS) were studied in the 1980s for indoor coverage enhancement. Network MIMO (Venkatesan, Lozano, Valenzuela, 2007) and cooperative multi-cell processing (Gesbert et al., 2010) formalized the idea of BSs jointly processing signals. What cell-free massive MIMO added was the insight that (i) a massive number of simple APs with (ii) simple local processing (conjugate beamforming, no centralized precoding) suffices to provide excellent performance — a crucial simplification that makes the concept scalable. The "cell-free" framing also sharpened the conceptual message: the cell boundary is not a physical constraint but an architectural choice that can be abandoned.

,

SINR Map: Cellular vs Cell-Free — Left: SINR heatmap for a cellular network with 7 BSs ( $N_t = 64$ per BS). The cell boundaries are clearly visible as low-SINR valleys. Right: SINR heatmap for a cell-free deployment with $L = 100$ single-antenna APs and the same total antenna count. The SINR is much more uniform — no valleys, no boundaries.

Max-Min Fair Power Control via Bisection

Complexity:

O(\log(1/\epsilon))

SOCP solves, each with

O(L \cdot K)

variables.

Input: Large-scale fading

\{\beta_{lk}\}

, estimation qualities

\{\gamma_{lk}\}

, noise

\sigma^2

Output: Power control coefficients

\{\eta_{lk}^*\}

, max-min SINR

t^*

1. Set

t_{\min} \leftarrow 0

,

t_{\max} \leftarrow \bar{t}

(upper bound from single-user case)

2. while

t_{\max} - t_{\min} > \epsilon

do

3.

\quad t \leftarrow (t_{\min} + t_{\max}) / 2

4.

\quad

Solve feasibility: find

\{\eta_{lk}\}

s.t.

5.

\quad\quad \text{SINR}_k^{\text{cf}}(\boldsymbol{\eta}) \geq t

for all

k = 1, \ldots, K

6.

\quad\quad \sum_{k=1}^{K} \eta_{lk} \gamma_{lk} \leq 1

for all

l = 1, \ldots, L

7.

\quad\quad \eta_{lk} \geq 0

for all

l, k

8.

\quad

if feasible then

t_{\min} \leftarrow t

; store

\{\eta_{lk}\}

9.

\quad

else

t_{\max} \leftarrow t

10. end while

11. return

\{\eta_{lk}^*\} \leftarrow

stored solution,

t^* \leftarrow t_{\min}

The feasibility check (lines 4--7) is a second-order cone program that can be solved efficiently using interior-point methods. For moderate $L$ and $K$ , the total runtime is a few seconds.

⚠️Engineering Note

Practical Deployment Considerations

Deploying cell-free massive MIMO in practice raises several engineering challenges:

AP placement: APs should be distributed to ensure coverage uniformity. Regular grids, random (PPP) deployments, and optimized placements all give different trade-offs.
Synchronization: Coherent JT requires tight time and phase synchronization across distributed APs. GPS-disciplined oscillators or over-the-air synchronization protocols are needed.
Power supply: Each AP needs power. Options include Power-over-Ethernet (PoE, up to 90W for PoE++), solar, or conventional mains. PoE limits the per-AP transmit power to roughly 200 mW (23 dBm).
Cost: While each AP is simpler than a macro BS, the total infrastructure cost (fronthaul, power, installation) for $L = 100$ APs may exceed that of 7 macro BSs.

Practical Constraints

•
PoE limit: 200 mW per AP (23 dBm)
•
GPS synchronization accuracy: ~10 ns (sufficient for sub-6 GHz)
•
Fronthaul: Ethernet-based, 1–10 Gbps per AP for Level 1–2 processing

,

Use-and-then-Forget (UatF) Bound

A technique for deriving achievable rate lower bounds in massive MIMO systems with imperfect CSI. The estimated channel is used for beamforming ("use"), and then the estimation error is treated as worst-case uncorrelated noise ("forget"). This yields closed-form rate expressions that depend only on channel statistics, not on instantaneous channel realizations.

Macro-Diversity

Spatial diversity obtained from geographically separated transmitters/receivers. In cell-free massive MIMO, macro-diversity arises because each user receives signals from many distributed APs at different locations, providing resilience against shadowing and path loss variations.

Quick Check

Under max-min fair power control in cell-free massive MIMO, which user receives the most total transmit power from the network?

The user closest to the most APs (strong user)

The user with the weakest aggregate channel (weak user)

All users receive equal power

It depends on the number of APs

Correction:

The user with the weakest aggregate channel (weak user)

Max-min fairness maximizes the minimum rate. To equalize SINRs, the power control coefficients $\eta_{lk}$ are largest for the weakest user, directing more network power toward users that need it most.

Key Takeaway

The initial performance analysis of cell-free massive MIMO reveals its fundamental advantage: the 95%-likely per-user rate is 5 to 10 times higher than small cells with the same total antenna count. This gain comes from (i) the absence of cell boundaries eliminating the worst-case geometry, (ii) macro-diversity from distributed APs, and (iii) max-min fair power control that focuses network resources on the weakest users. The simplest processing — conjugate beamforming with local CSI — is already highly effective.

Initial Performance Analysis

How Good Is Cell-Free Massive MIMO?

Theorem: Achievable Downlink SINR for Cell-Free Massive MIMO

Downlink signal model

Desired signal term

UatF bound application

SINR expression

Example: Achievable Rate in a Simple Cell-Free Setup

Cell-free desired signal (numerator)

Cellular desired signal

Comparison

Theorem: Max-Min Fair Power Control

Quasi-concavity

Bisection

Convergence

Sum Rate vs AP Density

Parameters

SINR Uniformity: Cellular vs Cell-Free

Parameters

Example: Cell-Free vs Small Cells: The Ngo et al. Comparison

Small-cell architecture

Cell-free architecture

95%-likely rate

Theorem: Channel Hardening in Cell-Free Massive MIMO

Mean of aggregate channel

Variance of aggregate channel

Convergence

Cell-Free Massive MIMO: The Founding Paper

Common Mistake: Cell-Free Does Not Beat Co-Located at Peak Rate

Historical Note: Precursors: Distributed Antenna Systems and Network MIMO

SINR Map: Cellular vs Cell-Free

Max-Min Fair Power Control via Bisection

Practical Deployment Considerations

Use-and-then-Forget (UatF) Bound

Macro-Diversity

Quick Check

Key Takeaway