Field Trials and the Practical-Theoretical Gap

Theory Meets Asphalt

Every section of this book so far — capacity scaling, channel hardening, linear precoding, codebook design, beam management — has been an argument about what should work in a massive MIMO system. Field trials measure what does work once the cellular network meets urban asphalt, indoor concrete, moving pedestrians, moving buses, and half-broken backhaul links. The gap between theoretical and measured performance is often 30-50% on sum rate and 1-3 dB on per-user SNR, and the reasons are not mysteries: CSI ageing at vehicular speeds, non-orthogonal pilots shared with neighbor cells, scheduler constraints that prefer legacy UEs, calibration drift on hot summer days.

This section collects published numbers from commercial 64T64R and 128T128R massive MIMO deployments, in Europe, China, and North America, from 2019 through 2023. The goal is to give a calibrated sense of the operating point — what sum rates, per-user throughputs, and spectral efficiencies are actually achieved — and to relate the measured gap to the specific theoretical assumptions that break in practice.

Definition:
Field Trial Metrics

Massive MIMO field trials are characterized by three operating-point metrics:

Cell sum throughput (bits/s): total data rate delivered by the cell, summed over all active UEs. A function of bandwidth, antenna count, UE count, and load.
Cell spectral efficiency (bits/s/Hz): cell sum throughput normalized by bandwidth. This is the metric most comparable to theoretical sum-rate predictions. Commercial 64T64R 3.5 GHz cells achieve 25-50 bits/s/Hz in uplink and 40-80 bits/s/Hz in downlink at typical load.
95%-ile or 5%-ile user throughput (bits/s): the rate exceeded by 95% (or 5%) of UEs. A fairness metric. 5%-ile throughput is the cell-edge experience and is what limits the commercial value proposition.

A useful shorthand is the massive MIMO gain: the ratio of massive MIMO SE to single-layer baseline SE under the same conditions. Typical measured gains are $4$ - $8\times$ in downlink and $3$ - $5\times$ in uplink — substantial but below the $K$ factor predicted by theory, which would be $10$ - $20$ for $K \in \{10, 20\}$ .

Theorem: UatF Lower Bound vs Measured Performance

Under the use-and-then-forget (UatF) lower bound (Chapter 4) with MMSE channel estimation and ZF precoding, the per-user downlink rate is $R_k^{\text{UatF}} = \log_2\!\left(1 + \frac{(N_t - K)\,\text{SNR}_{k}\,\gamma_k} {1 + K\,\text{SNR}_{k}\,(1 - \gamma_k)}\right),$ where $\gamma_k = \tau_p\,\text{SNR}_{k} / (1 + \tau_p\,\text{SNR}_{k})$ is the MMSE estimation quality. Measured per-user rates in commercial 64T64R deployments at 3.5 GHz are typically 50-70% of this lower bound — the 30-50% gap is attributable to: (i) CSI ageing from non-zero UE mobility, (ii) imperfect pilot orthogonality in multi-cell deployments (residual pilot contamination), (iii) scheduler grouping constraints that mismatch the ZF assumption of Gaussian random users, (iv) backhaul latency that widens the CSI feedback loop.

The UatF bound already incorporates imperfect CSI, but assumes a stationary channel over the coherence block. Real channels are non-stationary on the coherence-time scale at vehicular speeds, and the scheduler selects UEs in a correlated way based on QoS. Each of these is a 10-20% rate hit; stacked, they account for the observed gap.

Show Hint

Model CSI ageing as a phase rotation by $2\pi f_D T_{\text{CSI ageing}}$ on the per-user channel vector.

Pilot contamination adds a bias term that does not vanish with $N_t \to \infty$ (Marzetta 2010).

Non-Gaussian user selection breaks the channel-Gram concentration that ZF relies on.

Proof

CSI ageing rate penalty

If the BS uses a stale CSI estimate from $\Delta t$ earlier, the effective channel has an extra phase noise of variance $\sigma_\phi^2 \approx (2\pi f_D \Delta t)^2 / 2$ per spatial mode. This multiplies the ZF useful-signal term by $\exp(-\sigma_\phi^2 / 2)$ and adds back variance as residual interference. Typical $\Delta t = 5$ - $10$ ms at $\mu = 1$ gives $\sigma_\phi^2 \approx 0.01$ - $0.1$ rad $^2$ at pedestrian speeds — a 1-5% rate loss. At 60 km/h it rises to 10-20%.

Pilot contamination residual

In the limit $N_t \to \infty$ , the UatF bound converges to a SINR of $\gamma_k / (\text{pilot reuse count})$ , which does not vanish. In practice the bound is tight for single isolated cells but loose by 10-30% for multi-cell deployments with dense pilot reuse. Caire 2018 showed that spatial correlation recovers some of this loss (Chapter 3).

Scheduler non-uniformity

The UatF bound assumes uniformly random user selection, but real schedulers pick users that are simultaneously high-SNR and bandwidth-needy. This violates the channel-Gram matrix concentration assumption and introduces a scheduler-dependent correction of order $1/K$ . In practice this is 5-15% of the rate. $\blacksquare$

FR1 vs FR2 Coverage at Matched EIRP

Downlink coverage (range at target SINR) for FR1 (3.5 GHz, 64T) and FR2 (28 GHz, 128T) at matched EIRP. The FR2 beam-forming gain partially compensates for the extra Friis loss, but the net coverage at cell edge is 3-5x smaller. This is the core reason FR2 deployments use small cells rather than macro sites.

Parameters

Target SINR (dB)5

EIRP (dBm)60

FR1

N_t

FR2

N_t

128

Example: Measured SE for a 64T64R 3.5 GHz Urban Macro

A Rel-15 commercial 64T64R AAU deployed at 3.5 GHz in a dense urban setting reports a mean cell downlink SE of 42 bits/s/Hz over a 100 MHz carrier. The cell carries 18 connected UEs with an average of 6 simultaneously scheduled per TTI. Compare this with the theoretical UatF prediction for the same configuration and identify the dominant source of gap.

Solution

Theoretical UatF prediction

With $N_t = 64$ , $K = 6$ simultaneously scheduled, $\text{SNR} \approx 10$ dB (mean SINR in urban macro), and perfect CSI, the UatF bound gives $R_k \approx \log_2(1 + (64-6) \cdot 10 / (1 + 6 \cdot 0.1)) \approx \log_2(363) \approx 8.5$ bits/s/Hz per user. Summed over 6 users: $\approx 51$ bits/s/Hz.

Observed-to-theory ratio

$42/51 \approx 0.82$ . The measured SE is 82% of the UatF prediction. This is at the good end of published ratios, reflecting a mature deployment and a moderate-mobility environment (urban pedestrian + some vehicular).

Dominant gap source

Decomposing the 18% gap: CSI ageing at 30 km/h average UE speed contributes $\approx 8\%$ ; pilot contamination from neighbor cells contributes $\approx 5\%$ ; scheduler imperfections (fairness, QoS) contribute $\approx 5\%$ . The three sources stack roughly additively at this load level.

Implication for 6G

If 6G is to close this 18% gap, the primary targets are CSI prediction (ML-based Kalman-style beam-forward prediction) and tighter pilot reuse through cell-free coordination. The latter is the Part III thread of this book returning to the 5G commercial problem. $\blacksquare$

FR1 vs FR2 Commercial Deployments (2023)

Property	FR1 (3.5 GHz, 64T)	FR2 (28 GHz, 128T)
Cell range (urban macro)	500-1500 m	150-400 m
Peak DL SE	40-80 bits/s/Hz	100-200 bits/s/Hz
Average DL SE (loaded)	25-50 bits/s/Hz	50-120 bits/s/Hz
Peak per-UE throughput	1-3 Gb/s	3-8 Gb/s
UE count per cell	50-200	10-50
Typical deployment	Macro sectorized	Dense small cell / fixed wireless access
Primary limiting factor	Interference and pilot contamination	Path loss and blockage
Measured mMIMO gain	4-8 $\times$	6-10 $\times$ (baseline is single-beam)

🔧Engineering Note

Deployment Realities Ignored by Theory

A partial catalogue of what commercial deployments face but analytical massive MIMO models do not:

Thermal drift of power amplifiers (0.1-0.5 dB per degree C) causes calibration errors that compound with array size.
Antenna-pattern shadowing from the BS tower and mounting hardware creates 2-5 dB beam-direction-dependent gain ripples.
UE capability heterogeneity forces the scheduler to group users by their CSI feedback mode (Type I vs Type II), creating scheduling bubbles.
Handover signaling during UE mobility interrupts CSI-RS processing for 10-50 ms per handover.
DRX (discontinuous reception) UEs are not measuring channels for 70-90% of the time — when they wake up, their CSI is stale and must be refreshed before scheduling.

Each item is a 1-5% sum-rate hit. They stack, and they are the reason commercial deployments typically see 30-50% below the clean-theory UatF bound.

Practical Constraints

•
PA thermal drift must be compensated every minute at high load
•
Mounting hardware adds 3-5 dB ripple in the beam gain pattern
•
DRX UEs require CSI refresh on wake-up (5-20 ms delay)

📋 Ref: 3GPP TS 38.104 Section 9

🎓CommIT Contribution(2022)

CommIT Contributions to the Practical-Theoretical Gap

G. Caire, S. Shamai, H. Q. Ngo — TU Berlin CommIT group technical report and invited talks at IEEE GLOBECOM 2022 / ICC 2023

Caire's 2022-2023 invited retrospective on massive MIMO in commercial deployments makes three points that tie Part V back to Parts I-IV. First, the pilot contamination residual that Marzetta identified in 2010 is the dominant source of the theoretical gap at dense deployments — and the spatial-correlation exploitation of Caire 2018 (Chapter 3) is the correct remedy, not simply more antennas. Second, the scheduler cost of grouping heterogeneous- capability UEs is an underappreciated bottleneck that requires fundamentally new scheduler designs (Chapter 25 on AI/ML). Third, the gap between 5G NR multi-TRP and cell-free massive MIMO (Chapters 11-15) is the most promising 6G frontier: the cell-free architecture directly attacks both pilot contamination and scheduling heterogeneity, and is the logical successor to the multi-TRP Rel-16/17/18 evolution. These three points frame the narrative of Part V of this book.

massive-mimo5g-nrretrospectivefield-trials

Common Mistake: Reporting a Single-Snapshot SE Is Misleading

Mistake:

Published massive MIMO trial results often report a single peak SE number — "67 bits/s/Hz downlink" — as though this characterizes the deployment.

Correction:

A single-snapshot SE can be achieved under rare favorable conditions (quiet hour, few active UEs, all cell-center UEs) and does not reflect the operating-point average. A responsible report includes: (i) the mean SE at loaded conditions, (ii) the 5%-ile and 95%-ile user throughputs, (iii) the active-UE distribution during the measurement, and (iv) the mobility profile. Without these, comparing a "peak" number against a theoretical bound is meaningless. The field has gradually moved toward the four- statistic reporting template, but older papers still circulate peak-only claims.

Historical Note: 2019: The First Commercial Massive MIMO Deployments

2019-2023

The first large-scale commercial massive MIMO deployments happened in 2019, in China (China Mobile, Huawei 64T64R at 2.6 GHz), South Korea (KT, SK Telecom, Ericsson 64T at 3.5 GHz), and shortly after in Germany (Deutsche Telekom, Nokia AirScale at 3.6 GHz). Initial field reports were mixed: the 64T64R arrays worked as specified, but the commercial return on investment was debated because only a handful of UEs in each cell could actually use the MU-MIMO gain — most UEs were legacy LTE units with single-antenna capability. By 2021, UE support had improved and commercial massive MIMO became the standard mid-band configuration across the major vendors. The 2019 deployments are the dataset against which 6G cell-free proposals are benchmarked.

Why This Matters: Field Trials as the 6G Benchmark

Every 6G research proposal — cell-free massive MIMO, LEO NTN, ISAC, AI/ML-driven scheduling — will be evaluated against the measured performance of 5G NR massive MIMO deployments, not against the theoretical UatF bound. The bar is roughly "doubling the bits/s/Hz of a 64T64R 3.5 GHz cell in the same propagation environment," which is a concrete and hard target. The field trial numbers of this section are the baseline for Chapters 23-27.

64T64R Massive MIMO AAU

An active antenna unit with 64 transmit and 64 receive RF chains, typically arranged as an 8x8 or 4x16 dual-polarized panel. The dominant commercial massive MIMO configuration at FR1 mid-band (3.5 GHz) from 2019 to 2024.

5%-ile Spectral Efficiency

The user-perceived spectral efficiency exceeded by 95% of UEs in the cell — the cell-edge experience. The primary fairness metric for commercial massive MIMO deployments, more relevant than peak SE for customer-facing key performance indicators.

Quick Check

What is the typical ratio of measured commercial massive MIMO sum-rate to the UatF theoretical bound under realistic conditions?

$> 95\%$

$50$ - $70\%$

$20$ - $40\%$

$< 10\%$