Cell-Free vs Small Cells vs Co-Located

A Fair Comparison Requires Equal Resources

Comparing cell-free massive MIMO to small cells or co-located massive MIMO is meaningful only when the total system resources are held constant: same total number of antennas, same total transmit power, same total fronthaul capacity. Without this constraint, any architecture can be made to "win" by giving it more resources. This section establishes a rigorous comparison framework that isolates the architectural gains from mere resource scaling.

Definition:
Three Baseline Architectures

Fix a total antenna budget of $M = LN$ antennas over a coverage area $\mathcal{A}$ :

1. Cell-free massive MIMO: $L$ APs with $N$ antennas each, distributed uniformly over $\mathcal{A}$ . All APs coherently serve all $K$ users via a central processing unit (CPU) connected by fronthaul links.

2. Small cells: $L$ small-cell BSs with $N$ antennas each, distributed uniformly over $\mathcal{A}$ . Each user is served by its nearest BS only (no cooperation). Inter-BS interference is treated as noise.

3. Co-located massive MIMO: A single BS with $M = LN$ antennas at the center of $\mathcal{A}$ , serving all $K$ users. Full CSI available at the BS; no fronthaul needed.

The total transmit power budget is $P_{\text{tot}} = L \cdot P_t$ for all three architectures.

Small cells use the same AP hardware as cell-free but without cooperation — the "zero-cost baseline." Co-located massive MIMO has perfect cooperation but suffers from the proximity problem: some users are far from the single BS.

Theorem: Cell-Free Outperforms Small Cells in Min-Rate

Under max-min fair power control, the 95%-likely per-user rate of cell-free massive MIMO with conjugate beamforming satisfies

$R_{5\%}^{\text{cf}} \geq R_{5\%}^{\text{sc}}$

where $R_{5\%}$ denotes the 5th percentile of the per-user rate CDF and the superscripts denote cell-free (cf) and small cells (sc). The inequality is strict when the coverage area contains users that are not in the near vicinity of any single AP.

More precisely, for any user $k$ in the small-cell architecture with SINR $\text{SINR}_k^{\text{sc}}$ , the cell-free SINR satisfies

$\text{SINR}_k^{\text{cf}} \geq \frac{\left(\sum_{l} \sqrt{\eta_{lk}} \gamma_{lk}\right)^2}{\text{SINR}_k^{\text{sc},-1} \cdot (\max_l \gamma_{lk})^2} \cdot \text{SINR}_k^{\text{sc}}$

so cell-free provides a gain factor that grows with the number of APs contributing significantly to user $k$ 's signal.

Small cells discard the signals from all but the nearest AP. Cell-free coherently combines contributions from all APs with reasonable channel quality. For a cell-center user (near one AP), both architectures perform similarly. For a cell-edge user (equidistant from multiple APs), cell-free turns the interferers into helpers — the SINR improvement can be 10 dB or more.

Proof

Small-cell SINR

In small cells, user $k$ is served only by its nearest AP $l^*(k) = \arg\max_l \beta_{lk}$ . The SINR is $\text{SINR}_k^{\text{sc}} = \frac{N \gamma_{l^*k} P_t}{\sum_{l \neq l^*} \beta_{lk} P_t + \sigma^2}$ , limited by inter-cell interference from all other small cells.

Cell-free SINR lower bound

In cell-free, the coherent combining gain includes contributions from all APs: $(\sum_l \sqrt{\eta_{lk}} \gamma_{lk})^2 \geq (\sqrt{\eta_{l^*k}} \gamma_{l^*k})^2$ . The interference is no worse than in small cells because each interfering user's power is now split across multiple APs (reducing per-AP interference).

Gain for cell-edge users

For a user equidistant from $M_{\text{eff}}$ APs, the cell-free gain over the best single AP is $\sim M_{\text{eff}}^2$ in signal power, while interference from those same APs is eliminated. The 5th-percentile rate improvement follows because cell-edge users benefit most. $\blacksquare$

SINR CDF: Cell-Free vs Small Cells vs Co-Located

Compare the cumulative distribution function (CDF) of per-user SINR across the three architectures. Observe how cell-free dramatically improves the lower tail (cell-edge users) at the cost of a small reduction at the upper tail (cell-center users).

Parameters

L

(APs)64

K

(users)20

N

(antennas/AP)1

SNR (dB)10

Path-loss exponent

\alpha

3.8

Cell-Free vs Small Cells vs Co-Located Massive MIMO

Property	Cell-Free	Small Cells	Co-Located
Cooperation	All APs serve all users	No cooperation	Full cooperation (single BS)
Macro-diversity	High (geographically distributed)	None (single AP per user)	None (single site)
Cell-edge performance	No cell edges; uniform coverage	Poor (inter-cell interference)	Poor for far users (path loss)
Fronthaul requirement	High (all APs to CPU)	None (local processing)	None (co-located)
CSI acquisition	Distributed MMSE at each AP	Local at each BS	Centralized at single BS
95%-likely rate	Highest	Lowest	Medium
Peak rate	Medium (distributed overhead)	Low (interference-limited)	Highest (full array gain)
Scalability	User-centric clustering needed	Inherently scalable	Single-site bottleneck
Pilot contamination	Mitigated by AP diversity	Mitigated within cell	Severe across cells
Hardware complexity	Many simple APs	Many simple BSs	One complex BS

Example: When Does Cell-Free Win the Most?

Consider a 1 km $\times$ 1 km area with $M = 64$ total antennas and $K = 10$ users. Compare the 5th-percentile rate for: (a) cell-free with $L = 64$ single-antenna APs, (b) small cells with $L = 64$ single-antenna BSs, and (c) co-located with one 64-antenna BS at the center. Assume path-loss exponent $\alpha = 3.8$ and SNR = 10 dB.

Solution

Cell-free advantage at the edge

In cell-free, the worst user is surrounded by distributed APs on all sides. With 64 APs on a grid, the maximum distance to the nearest AP is about $1000/(2\sqrt{64}) = 62.5$ m. Multiple APs at distances 60-100 m contribute to the signal.

Small-cell disadvantage at the edge

In small cells, the same worst user is served by only the nearest BS. Other nearby BSs cause interference. The SIR at the cell boundary with one dominant interferer at the same distance is approximately 0 dB (no array gain from $N = 1$ ).

Co-located disadvantage for far users

With the BS at the center, users at the corners ( $d \approx 707$ m) have path loss $707^{-3.8} \approx 5.2 \times 10^{-11}$ , while users near the center ( $d \approx 50$ m) have $50^{-3.8} \approx 2.4 \times 10^{-7}$ — a 37 dB difference. The 5th-percentile rate is dominated by corner users.

Comparison

Numerical evaluation (see the interactive plot above) yields approximate 5th-percentile rates: cell-free $\approx 1.8$ bits/s/Hz, co-located $\approx 0.9$ bits/s/Hz, small cells $\approx 0.3$ bits/s/Hz. Cell-free provides a $\sim 6\times$ improvement over small cells and $\sim 2\times$ over co-located massive MIMO in the lower tail of the rate distribution.

When Co-Located Beats Cell-Free

Cell-free does not dominate in every metric. Co-located massive MIMO achieves higher peak rate for users near the BS, because the full $M$ -antenna array gain is concentrated at one site. If the goal is to maximize sum-rate or serve a small number of users near a known location (e.g., a stadium), co-located is preferable. Cell-free wins when the goal is uniform coverage and fairness — making the worst user's experience acceptable. The choice depends on the deployment objective.

Historical Note: From Distributed Antenna Systems to Cell-Free

1987-2017

The idea of distributing antennas across a coverage area predates cell-free massive MIMO by decades. Distributed Antenna Systems (DAS) were studied in the 1990s and 2000s, primarily for indoor coverage. Saleh, Rustako, and Roman (1987) analyzed distributed antennas in buildings. Choi and Andrews (2007) provided one of the first rigorous multi-cell analyses. However, DAS treated each distributed antenna as an independent entity with limited cooperation. The breakthrough of cell-free massive MIMO (Ngo et al., 2017) was to combine distribution with massive MIMO principles: TDD reciprocity for scalable CSI acquisition, coherent joint processing, and max-min fair power control. This marriage of distributed deployment with coherent massive processing is what enables the dramatic fairness gains.

Common Mistake: Comparing Architectures with Unequal Total Antennas

Mistake:

Comparing cell-free with $L = 100$ APs to co-located with $M = 64$ antennas, then claiming cell-free is "better" because it has more total antennas.

Correction:

A fair comparison requires $M_{\text{total}} = L \cdot N$ to be the same across all architectures. If cell-free uses $L = 100$ single-antenna APs, the co-located system should have $M = 100$ antennas, and the small-cell system should have 100 single-antenna BSs. Only then do differences reflect the architectural advantage rather than resource scaling.

Quick Check

Which users benefit most from cell-free massive MIMO compared to small cells?

Users near the center of a cell (cell-center users)

Users at cell boundaries (cell-edge users)

Users with line-of-sight to many APs

All users benefit equally

Correction:

Users at cell boundaries (cell-edge users)

Cell-edge users suffer from weak desired signal and strong interference in small cells. Cell-free converts nearby interfering APs into signal sources via coherent combining, providing the largest improvement where it is needed most.

Definition:
Macro-Diversity Gain

Macro-diversity refers to the signal strength improvement achieved by receiving (or transmitting) from multiple geographically separated nodes. In cell-free massive MIMO, the macro-diversity gain for user $k$ is

$G_{\text{macro},k} = \frac{\left(\sum_{l \in \mathcal{M}_k} \gamma_{lk}\right)^2}{\max_{l \in \mathcal{M}_k} \gamma_{lk}^2}$

which measures the squared coherent combining gain relative to the best single AP. For a user equidistant from $|\mathcal{M}_k|$ APs with equal channel quality, $G_{\text{macro},k} = |\mathcal{M}_k|^2$ — a quadratic gain. For a user very close to one AP, $G_{\text{macro},k} \approx 1$ (the nearest AP dominates).

Macro-diversity is fundamentally different from micro-diversity (multiple antennas at one site). Micro-diversity averages out small-scale fading; macro-diversity averages out large-scale fading (path loss and shadowing). Cell-free systems exploit both.

Macro-Diversity

Signal strength improvement from combining signals across geographically separated access points. Averages out path loss and shadowing variations, providing more uniform coverage than co-located arrays.

95%-Likely Per-User Rate

The 5th percentile of the CDF of per-user achievable rates across random user locations. Equivalently, the rate that 95% of users achieve or exceed. Used as the primary fairness metric in cell-free massive MIMO literature.

🔧Engineering Note

AP Density in Practice

Current 5G small-cell deployments use inter-site distances (ISD) of 100-500 m in urban areas. Cell-free massive MIMO with single-antenna APs requires higher density (ISD 20-50 m) to achieve the macro-diversity gains shown in the analysis. This means 400-2500 APs per km $^2$ . While this seems extreme, it aligns with the vision for ultra-dense 6G networks. The key enabler is the simplicity of each AP: a single antenna, a single RF chain, and a fronthaul connection. At scale, the per-AP cost is projected to be comparable to a Wi-Fi access point.

Practical Constraints

•
Urban lamppost density: ~40-80 per km^2 (sufficient for ISD 100-200 m)
•
Indoor ceiling density: ~100-400 per km^2 (sufficient for ISD 50-100 m)
•
Power-over-Ethernet (PoE) can supply single-antenna APs with ~15 W

Achievable Rate Expressions Fairness and Coverage