Distributed MIMO: Wireless Architectures

• Distributed MIMO is a wireless architecture that uses geographically separated antennas with cooperative processing to enhance spatial multiplexing and uniform service.
• It spans various configurations like cell-free, satellite, and industrial deployments, each tackling distinct challenges in fronthaul, synchronization, and calibration.
• Research in distributed MIMO highlights trade-offs in antenna deployment, quantization, and overhead management to balance performance and system complexity.

Searching arXiv for recent and foundational papers on distributed MIMO to ground the article in published work. arXiv search queries:

1. "Distributed MIMO"
2. "Distributed multiple input multiple output"
3. "Distributed Massive MIMO"
4. "cell-free massive MIMO distributed MIMO"
5. "Distributed MIMO 1-bit quantized reception"
6. "Distributed MIMO satellite"
7. "Distributed MIMO unlicensed"
8. "Distributed MIMO industrial channel model" Distributed Multiple Input Multiple Output (MIMO) denotes a class of wireless architectures in which antenna elements, access points, remote radio heads, satellites, or receive nodes are geographically separated yet cooperatively processed by a fusion center, central processing unit, central unit, or coordinated base station function. In contrast to co-located arrays, distributed MIMO uses spatial separation itself as a system resource: it can realize spatial multiplexing, macro-diversity, coverage-hole mitigation, and more uniform service, while also introducing stringent requirements on fronthaul, channel acquisition, synchronization, reciprocity calibration, and overhead management. The literature spans receive-side cloud MIMO with one-bit quantized nodes, cell-free and distributed massive MIMO, distributed phased arrays at mmWave, industrial and unlicensed deployments, and satellite and mobile distributed arrays (Choi et al., 2014, Tominaga et al., 2023, Abdelsadek et al., 2022, Said et al., 16 Apr 2025).

1. Architectural forms and system organization

A recurring architectural distinction is between co-located MIMO and distributed MIMO. In co-located systems, all antennas are mounted at a single site; in distributed systems, the same total number of antennas is partitioned across multiple geographically separated entities. Depending on the paper, these entities are single-antenna receive nodes forwarding quantized observations to a fusion center, access points jointly serving users, remote radio heads connected to a central unit, satellite access points in a LEO constellation, or distributed subarrays in a phased-array architecture (Choi et al., 2014, Tominaga et al., 2023, Abdelsadek et al., 2022, Zhang et al., 2019).

The receive-side formulation is exemplified by a transmitter with $N_t$ antennas that sends $N_t$ independent PSK streams to $K$ geographically separated single-antenna receive nodes. Each node quantizes its received signal using one bit for the real part and one bit for the imaginary part, and forwards the result to a receive fusion center that knows all channels and performs detection (Choi et al., 2014). This formulation is often described as cloud MIMO or distributed reception and is notable because each receive node processes only a single quantity regardless of the number of transmitted symbols.

The transmit-side formulation is more commonly associated with distributed massive MIMO, cell-free MIMO, or network MIMO. Here, $Q$ access points or remote radio heads, each with $S$ antennas or even a single antenna, jointly serve $K$ single-antenna users over a common resource block. The aggregate channel is formed by stacking the per-access-point channels, and a CPU applies linear combining or precoding over the distributed aperture (Tominaga et al., 2023, Björnson et al., 2015). In satellite variants, user terminals connect simultaneously to a cluster of $M$ satellites, with inter-satellite links carrying fronthaul data to a CPU, so that many satellites act as geographically distributed access points (Abdelsadek et al., 2022). In mobile distributed MIMO, the same cooperative principle is extended to wirelessly interconnected and potentially mobile remote units and even a mobile gNB, so that both the distributed array and its fronthaul can move (Said et al., 16 Apr 2025).

Several papers emphasize that distributed MIMO is not a single topology but a design space. Representative layouts include grid-distributed access points, linear radio stripes, fully distributed single-antenna access points, partially distributed multi-antenna access points, centralized benchmarks, user-centric satellite clusters, and distributed phased arrays composed of multiple subarrays (Tominaga et al., 2023, Elwekeil et al., 2023, Humadi et al., 2024, Zhang et al., 2019). This suggests that the term is best understood as a coordination paradigm rather than a fixed geometry.

2. Canonical signal models, channel models, and performance measures

The mathematical core of distributed MIMO is an aggregate channel representation. In quantized distributed reception, node $k$ observes

$y_k=\sqrt{\rho/N_t}\,\mathbf h_k^H\mathbf x+n_k,$

with $\mathbf h_k\in\mathbb C^{N_t}$ Rayleigh fading, $N_t$0, and $N_t$1 an $N_t$2-PSK transmit vector. The node forwards

$N_t$3

so the fusion center operates on a severely quantized but spatially distributed observation vector $N_t$4 (Choi et al., 2014).

In distributed massive MIMO for uplink machine-type communication, the CPU receives

$N_t$5

where $N_t$6 stacks all distributed channels from active devices, and each per-access-point channel obeys a correlated-Rayleigh model $N_t$7, with the industrial simplification $N_t$8 in the 3GPP-based setting considered there (Tominaga et al., 2023). In LEO distributed massive MIMO, the uplink or downlink link between user $N_t$9 and satellite access point $K$0 is modeled as Rician, with large-scale attenuation, Rician factor, phase, and Gaussian diffuse component explicitly separated (Abdelsadek et al., 2022). Under Rician fading for terrestrial D-MIMO, the per-user spectral efficiency depends critically on the factor $K$1, since increasing $K$2 increases inter-user channel correlation and can degrade the achievable spectral efficiency (Tominaga et al., 2024).

A standard downlink formulation writes

$K$3

where $K$4 is a linear precoder over the distributed access points or remote units (Abdelsadek et al., 2022). For uplink combining, papers commonly define

$K$5

with maximum-ratio combining, zero forcing, or MMSE as special cases (Tominaga et al., 2023, Tominaga et al., 2024).

The dominant performance metrics are spectral efficiency, sum spectral efficiency, outage probability, fairness, service time, and error-vector magnitude in testbed work. Representative expressions include

$K$6

for uplink data in block fading (Tominaga et al., 2023),

$K$7

for LEO downlink (Abdelsadek et al., 2022), and

$K$8

for experimental D-MIMO validation (Aabel et al., 2024). In unlicensed 6 GHz sub-networks, finite-blocklength rates with reliability target $K$9 and uplink fraction $Q$0 are used to reflect HRLLC constraints (Elwekeil et al., 2023).

3. Fronthaul, overhead, synchronization, and calibration

A central result across the literature is that distributed MIMO is frequently limited not by the access link alone but by the cost of coordination. One strand addresses protocol overhead directly. In orthogonally partitioned D-MIMO with joint multi-user beamforming, the effective sum-rate is

$Q$1

with overhead fraction $Q$2 and $Q$3, so that the raw scaling of zero-forcing beamforming can be negated by signaling overhead unless the system is partitioned into smaller orthogonal clusters (Lioumpas et al., 2012). The constrained partitioning problem is formulated as a bounded Knapsack problem with an explicit overhead budget.

Wireless fronthaul introduces a second class of constraints. In a coordinated RRH network, fronthaul and access use separate frequency resources, and the fronthaul rate to an RRH is modeled statistically under either multicast or zero-forcing beamforming. The fronthaul outage constraint

$Q$4

is then coupled to RRH placement and user access rate optimization. The resulting design conclusion is cautionary: fronthaul requires considerable bandwidth to enable joint service, and the requirement is relaxed by serving a low number of users on the same resource block (Ammar et al., 2021).

Physical synchronization is equally central. A TDD distributed MIMO testbed with 1-bit radio-over-fiber fronthaul demonstrated that downlink zero-forcing based on uplink channel estimates fails without reciprocity calibration: using uncalibrated UL-based ZF produced approximately $Q$5 and $Q$6 EVM at the two users, whereas over-the-air reciprocity calibration reduced these to $Q$7 and $Q$8, respectively (Aabel et al., 2024). In a later analytical treatment, over-the-air inter-access-point phase calibration was integrated into the TDD flow by shifting uplink/downlink switching points to create short calibration segments; the long-term spectral efficiency was then written as

$Q$9

making explicit the trade-off between calibration overhead and residual phase error (Ngo et al., 3 Sep 2025).

Hardware impairments also shape coherent operation. In distributed massive MIMO downlink with maximum-ratio transmission, additive hardware distortions vanish asymptotically with large antenna counts, whereas multiplicative phase noise remains a limiting factor. Under the stated scaling law, separate oscillators at each antenna are preferable to a common local oscillator at the base station (Björnson et al., 2015). This is one reason why oscillator architecture, calibration periodicity, and fronthaul design cannot be treated as secondary implementation details in distributed MIMO.

4. Detection, combining, precoding, and open-loop transmission

The receive-side detection problem under coarse quantization is unusually stringent. For one-bit quantized distributed reception, the fusion center can implement an optimal maximum-likelihood detector that exhaustively searches over $S$0 PSK candidates, or a low-complexity zero-forcing-type detector

$S$1

followed by nearest-constellation slicing (Choi et al., 2014). The ML detector is asymptotically consistent as the number of receive nodes $S$2 grows, while the ZF-type receiver is comparable in the low-SNR regime but exhibits a high-SNR error-rate floor for fixed $S$3. That floor is reduced by increasing the number of receive nodes.

Channel estimation under the same quantized architecture admits parallel ML and ZF-type formulations. With block fading and two-bit quantization per node, the ML channel estimator solves a convex program in the real domain, while the ZF-type channel estimator

$S$4

achieves mean-squared error that decays as

$S$5

that is, on the order of $S$6 with training length $S$7 (Choi et al., 2014).

In more conventional distributed massive MIMO, linear combining and precoding dominate. The principal uplink combiners are MRC, ZF, and MMSE, with

$S$8

as standard forms (Tominaga et al., 2023). For LEO distributed massive MIMO with asynchronous satellite-to-user delays, precoding must additionally be phase-shift aware: $S$9 where $K$0 compensates the propagation-delay phase rotation (Humadi et al., 2024).

Not all distributed MIMO operation is CSI-rich. In the absence of instantaneous CSI at radio units, Alamouti-like orthogonal space-time-frequency block codes provide a robust open-loop mode. Orthogonality decouples symbols at the user and yields diversity order $K$1 for an $K$2-antenna code and $K$3 receive antennas, while inter-cluster interference is treated as noise (Kadan et al., 2023). At mmWave, the bottleneck is often beam management rather than linear precoding itself: Random Forest, MissForest, and conditional GAN models have been used to infer the best access point and beam from only a small subset of sounded beams, reducing sounding overhead by up to $K$4 at the cost of $K$5-$K$6 dB average SNR loss in the reported study (M et al., 2023). Wideband distributed phased-array MIMO adds another layer, since the common RF precoder must satisfy constant-modulus constraints across subarrays; the ADMM-AltMin method addresses this hybrid-precoding problem for OFDM with practical phase-shifter quantization (Zhang et al., 2019).

5. Performance trade-offs, design regimes, and common misconceptions

A persistent misconception is that distributing antennas is always superior to centralizing them. The available results are more conditional. In indoor industrial MTC simulations, grid-DmMIMO and linear-DmMIMO strictly outperform centralized massive MIMO under regular traffic, but centralized massive MIMO can outperform DmMIMO when channels are highly correlated or when active devices cluster around an alarm hotspot (Tominaga et al., 2023). Under Rician fading, the correlation among channel vectors increases with the Rician factor, degrading achievable spectral efficiency and altering the optimal number of access points and antennas per access point (Tominaga et al., 2024).

A second misconception is that “more access points” is always the correct direction. Multiple papers instead report a beamforming-versus-macro-diversity trade-off. With a fixed antenna budget $K$7, small coverage areas favor a few access points with many antennas each, while larger areas shift the optimum toward more access points with fewer antennas. In one study, a concave “sweet spot” appeared at roughly $K$8 APs/$K$9 for the considered parameter sets, and in another the optimal access-point count depended explicitly on the Rician factor and on whether MRC, ZF, or MMSE was used (Tominaga et al., 2023, Tominaga et al., 2024).

A third misconception concerns quantization. One-bit reception or fronthaul does not imply unusable performance, but the operating regime matters. For distributed reception with one-bit quantization, the ZF-type detector exhibits a high-SNR floor that vanishes as the number of receive nodes increases, and the ML detector achieves vanishing symbol error rate as $M$0 (Choi et al., 2014). In the 1-bit radio-over-fiber testbed, point-to-point 16QAM at $M$1 MBd achieved $M$2 downlink EVM and $M$3 uplink EVM, while a $M$4-RRH, $M$5-UE over-the-air downlink with reciprocity-calibrated ZF achieved $M$6 and $M$7 EVM at the two users (Aabel et al., 2024).

A fourth misconception is that the main challenge is only the access link. In the unlicensed 6 GHz setting, listen-before-talk can create deferrals that directly affect latency, and adaptive power reduction is used as a genie-aided lower bound that guarantees zero deferments at the cost of reduced peak rates (Elwekeil et al., 2023). In LEO satellite DM-MIMO, cluster size, handover rate, and power allocation are co-optimized because service continuity is part of the system objective rather than a separate mobility layer (Abdelsadek et al., 2022). These examples show that distributed MIMO performance is frequently cross-layer.

6. Application domains and research directions

Industrial communication is a major target. A measurement campaign at $M$8 GHz with $M$9 MHz bandwidth and twelve fully coherent distributed dipole antennas in an industrial hall produced a spatially consistent D-MIMO channel model with explicit path gain laws for LoS and obstructed LoS, Rician small-scale fading, RMS delay-spread statistics, Markov obstruction-state transitions, and correlated shadowing across anchors (Nelson et al., 2024). The inclusion of large-scale fading correlations, spatial correlation, and tail distributions is particularly relevant for URLLC evaluation, because rare events rather than average rates dominate many industrial reliability questions.

Satellite networking is another active domain. Distributed massive MIMO for LEO networks treats multiple satellites as a cooperative serving cluster and has been coupled with a distributed joint power allocation and handover management framework, later approximated by a DNN for real-time operation (Abdelsadek et al., 2022). A related user-centric LEO formulation introduces dynamic cluster selection, cluster handover, and phase-shift-aware precoding to compensate for propagation-delay differences among satellites, achieving median spectral efficiency close to full cooperation while drastically reducing average cluster size (Humadi et al., 2024).

Deployment flexibility motivates mobile distributed MIMO. In MD-MIMO, the fiber-connected remote units of classical D-MIMO are replaced by mobile wireless-fronthaul nodes with local PHY processing, enabling use cases such as vehicular and UAV-based cooperation. In the reported vehicular case study, downlink zero-forcing MD-MIMO at $k$0 km yielded a relative capacity gain of $k$1 over the baseline co-located BS-to-UE link, and the uplink throughput rose from approximately $k$2 Mbps with one RU to approximately $k$3 Mbps with four RUs (Said et al., 16 Apr 2025).

Distributed MIMO also intersects with decentralized control and complexity reduction. A decentralized antenna-selection algorithm based on local neighborhood capacity comparisons and random mutation achieved sum-rates within $k$4-$k$5 of a centralized greedy solution for $k$6 and $k$7 users, while reducing per-iteration time by a factor of approximately $k$8 through the use of only $k$9 of subcarriers selected uniformly at random (Siljak et al., 2018). A plausible implication is that future distributed MIMO systems will increasingly combine cooperative physical-layer processing with distributed optimization, local decision rules, and learned surrogates for otherwise intractable control problems.

Distributed MIMO is therefore best viewed not as a single transmission technique but as a family of cooperative array architectures whose advantages depend on propagation regime, traffic pattern, fronthaul model, hardware quality, and the degree of achievable coherence. The literature consistently supports the same conclusion: when macro-diversity, user-centric clustering, calibration, and overhead are jointly engineered, distributed MIMO can approach or exceed the performance of centralized architectures in regimes where co-location is limited by form factor, path loss, blockage, or service continuity requirements.