PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on User-Centric Cell-Free Massive MIMO Systems

Shuaifei Chen, Jiayi Zhang, Jing Zhang, Emil Björnson, Bo Ai

arXiv:2104.13667v2cs.ITeess.SP

TL;DR

The paper addresses modest user-experienced data rates, large SNR or QoS variations, and the open challenge of efficiently supporting massive access with high spectral efficiency. It reviews CF mMIMO concepts, techniques, schemes, algorithms, and implementation approaches, concluding with lessons about performance tradeoffs and decentralized architectures.

  • Problem

    User-experienced data rates remain modest because of inter-cell interference and large SNR, while efficiently supporting massive access and high spectral efficiency remains an open issue.

  • Method

    The paper presents a comprehensive review of CF mMIMO concepts, techniques, state-of-the-art schemes, algorithms, and centralized combining implementation approaches.

  • Results

    The review highlights QoS variations within the coverage area and performance tradeoffs associated with preferred received signals and decentralized CF approaches.

  • Takeaways & Limitations

    The surveyed lessons indicate that decentralized CF approaches can use multiple antennas to dynamically achieve the best performance available under the approach.

Abstract

from arXiv · show

The mobile data traffic has been exponentially growing during the last decades, which has been enabled by the densification of the network infrastructure in terms of increased cell density (i.e., ultra-dense network (UDN)) and/or increased number of active antennas per access point (AP) (i.e., massive multiple-input multiple-output (mMIMO)). However, neither UDN nor mMIMO will meet the increasing data rate demands of the sixth generation (6G) wireless communications due to the inter-cell interference and large quality-of-service variations, respectively. Cell-free (CF) mMIMO, which combines the best aspects of UDN and mMIMO, is viewed as a key solution to this issue. In such systems, each user equipment (UE) is served by a preferred set of surrounding APs cooperatively. In this paper, we provide a survey of the state-of-the-art literature on CF mMIMO. As a starting point, the significance and the basic properties of CF mMIMO are highlighted. We then present the canonical framework, where the essential details (i.e., transmission procedure and mathematical system model) are discussed. Next, we provide a deep look at the resource allocation and signal processing problems related to CF mMIMO and survey the up-to-date schemes and algorithms. After that, we discuss the practical issues when implementing CF mMIMO. Potential future directions are then pointed out. Finally, we conclude this paper with a summary of the key lessons learned in this field. This paper aims to provide a starting point for anyone who wants to conduct research on CF mMIMO for future wireless networks.

I. INTRODUCTION

Cell densification and mMIMO increase capacity, but user-experienced rates remain constrained by inter-cell interference and large SNR variations. CF mMIMO combines distributed APs with cooperative mMIMO processing to address both limitations.

  • Motivation: User-experienced data rates, rather than average or peak rates, determine whether applications operate without interruption.The paper links perceived user fairness and consistent experience to improving rates for users in unfavorable propagation conditions.
  • Existing approaches: Cell densification increases simultaneous activity and SNR, while mMIMO adds directive transmission and spatial multiplexing through many AP antennas.mMIMO can achieve area-capacity gains with fewer APs, but requires more complicated hardware at each AP.
  • Limitations of densification: 10 APs per km2 is already a regime where increasing average SNR through densification can be outweighed by growing interference from additional cells.This observation assumes all APs transmit simultaneously.
  • Limitations of densification: Ultra-dense networks can use infrastructure inefficiently because most APs remain idle when there are many more APs than active UEs.Inactive APs avoid interference, but the resulting low utilization limits the efficiency of the deployed infrastructure.
  • Limitations of mMIMO: mMIMO is inefficient at overcoming the large SNR differences experienced by UEs across macro and micro cells.Thus, both densification and mMIMO may raise peak and average rates while leaving user-experienced rates modest.
  • Cell-free mMIMO: CF mMIMO uses geographically distributed APs without disjoint cells, serving each UE cooperatively through surrounding APs.Distributed antennas can transmit with coordinated power and phase shifts so their signals reach the intended UE synchronously.
  • Cell-free mMIMO: Cooperative mMIMO processing addresses interference and mitigates SNR variation by replacing a few large arrays with many distributed AP antennas.The paper presents this architecture as combining the beneficial aspects of ultra-dense networks and cellular mMIMO.

B. Related Technologies

Earlier distributed-cooperation approaches extended cellular networks through AP coordination, whereas CF mMIMO envisions a cell-free architecture with many cooperating APs. This survey organizes the related technologies, foundations, algorithms, implementation issues, and research directions.

  • Related technologies: CoMP evolved existing cellular networks by coordinating neighboring APs to reduce inter-cell interference, rather than creating a cell-free network from scratch.This distinction separates the architectural premise of CoMP from the CF mMIMO vision.
  • Related technologies: CF mMIMO differs from earlier work through an operating regime with many more APs than UEs and mMIMO-inspired physical-layer processing.Earlier Network MIMO studies generally assumed perfect CSI, while practical CF mMIMO must handle imperfect CSI efficiently.
  • Related technologies: CoMP clusters preserve cellular structure, so cluster-edge UEs remain exposed to interference from neighboring clusters.CF mMIMO instead serves every UE through all surrounding APs.
  • Paper scope: The paper introduces CF mMIMO foundations through a canonical framework covering transmission procedures and the mathematical system model.This foundation supports the subsequent survey of resource allocation, signal processing, implementation, and future directions.
  • Paper scope: Its coverage spans state-of-the-art resource-allocation and signal-processing algorithms, practical implementation issues, future directions, and lessons learned.The roadmap assigns these topics to the paper’s later sections.
  • Contributions of This Survey: The survey addresses a literature gap because previous tutorials and surveys did not provide a comprehensive, up-to-date overview of applied CF mMIMO schemes and algorithms.The authors position the paper as a starting point for future research on CF mMIMO.

II. TECHNICAL FOUNDATIONS

The canonical CF mMIMO framework connects distributed APs to a CPU for cooperative transmission and reception over coherence blocks divided into pilot, uplink-data, and downlink-data phases. Channel estimation, combining, and achievable spectral-efficiency analysis support centralized or distributed processing, with imperfect channel knowledge requiring capacity bounds.

  • System model: CF mMIMO comprises K single-antenna UEs and L APs with N antennas, connected through fronthaul for coherent joint transmission and reception.The channels are modeled as block fading and may operate in TDD or FDD, although the basic model assumes TDD.
  • Transmission procedure: Each coherence block contains τp pilot symbols, τu uplink-data symbols, and τd downlink-data symbols, satisfying τc = τp + τu + τd.The block-fading abstraction can represent groups of OFDM subcarriers.
  • Channel estimation: Because τp < K in most practical scenarios, pilots are reused among UEs, causing pilot contamination and degraded channel estimation.Pilot-sharing UEs become harder to separate, reducing the effectiveness of coherent transmission and interference rejection.
  • Uplink processing: The CPU or APs estimate channels from received pilots, after which centralized or distributed combining estimates uplink data signals.Centralized combining uses all AP signals at the CPU, whereas distributed combining forms local estimates before CPU aggregation.
  • Performance analysis: With imperfect channel knowledge, exact ergodic capacity is unavailable, so achievable spectral-efficiency lower bounds are used for analysis.The MMSE-based bound is stated for MMSE channel estimation, while the distributed-combining bound can apply to any channel estimator.

3) Downlink Data Transmission:

Downlink transmission uses CPU- or AP-designed precoding, with power control and normalized precoding vectors determining the transmitted signals. Uplink-downlink duality can transfer uplink combining designs to downlink precoding, while downlink spectral efficiency depends jointly on all UEs’ precoders.

  • Downlink model: Downlink precoding vectors are normalized spatial-directivity vectors, while ρi ≥ 0 denotes the transmit power allocated to UE i.An achievable downlink spectral efficiency is computed using the hardening bound for a selected precoding scheme.
  • Joint design: Downlink spectral efficiency depends on the precoding vectors of all UEs, so precoding should be optimized jointly rather than independently per UE.This coupling distinguishes downlink design from uplink spectral-efficiency expressions based on an individual UE’s combining vector.
  • Uplink-downlink duality: Uplink-downlink duality permits downlink precoders to be selected from uplink combiners, followed by suitable downlink power control.The duality result preserves corresponding effective SINRs under the stated normalized-precoder and power-control construction.
  • Centralized precoding: Centralized precoding computes normalized vectors at the CPU using uplink channel estimates and channel reciprocity, then sends the formed downlink signal to APs.The CPU delegates downlink data encoding and forwards the resulting signal over the fronthaul.
  • Distributed precoding: Distributed precoding lets each AP select its local precoding vector from local channel estimates, requiring the CPU to send only downlink data signals.This design shifts precoder selection from the CPU to the APs.

B. Channel Model

The channel model represents propagation through line-of-sight and scattered components, yielding Rician fading when both are present and Rayleigh fading when only rich scattering is modeled. Unknown phase shifts can make perfect-phase analyses optimistic relative to practical operation.

  • Rician fading: A Rician channel coefficient is decomposed into a line-of-sight component with magnitude and phase and a non-line-of-sight scattered component.The scattered component is modeled as complex Gaussian, with its variance describing the aggregate scattered paths.
  • Phase uncertainty: With perfect receiver channel knowledge, the phase shift can be compensated and omitted from performance analysis.This simplification does not generally apply when the receiver must estimate the channel.
  • Phase uncertainty: Results obtained with a perfectly known phase shift provide an upper bound on practical performance when the phase cannot be perfectly tracked.The limitation is especially relevant because the phase affects the strong line-of-sight component.
  • Multi-antenna channels: For multi-antenna APs, the channel becomes an N-dimensional vector whose non-line-of-sight component has covariance Rkl.The line-of-sight vector, scattered component, and common phase shift extend the scalar model.
  • Rayleigh fading: Rayleigh fading models rich-scattering channels without a line-of-sight path, with hkl distributed as a zero-mean complex Gaussian variable.This model is widely used for basic wireless-propagation analysis.

C. Duplex Protocol

CF mMIMO commonly uses TDD because reciprocity supports downlink precoding from uplink CSI, whereas FDD requires additional CSI feedback. User-centric cooperation addresses scalability by limiting each AP’s served UEs, while channel hardening and favorable propagation support simpler resource allocation and signal processing.

  • Duplex protocol: In TDD, channel reciprocity lets APs use uplink-pilot CSI for downlink precoding, with overhead scaling with served UEs rather than AP antennas.FDD lacks uplink-downlink reciprocity because the two links occupy different frequency bands.
  • Duplex protocol: FDD-based CF mMIMO can exploit angle reciprocity in sufficiently sparse channels, making required overhead scale only with the number of served UEs.The relevant shared information includes angles of departure and dominant path gains rather than the full channel dimension.
  • Scalability issues: A network is scalable when channel estimation, data processing, fronthaul signaling, and power-control resource requirements remain finite as the UE count grows.Naive CF mMIMO is not scalable because processing and fronthaul load grow linearly or faster with the number of UEs.
  • Scalability issues: The user-centric approach assigns each AP a limited UE set Dl, so constant |Dl| yields constant per-AP complexity and fronthaul load as K →∞.APs compute channel estimates and combining or precoding vectors only for their assigned UEs and exchange only related data.
  • Propagation properties: Channel hardening makes effective channels nearly deterministic with many serving antennas, while favorable propagation makes distinct UE channels asymptotically orthogonal.Even limited hardening can support resource allocation based on long-term rather than small-scale channel statistics.
  • Resource allocation and signal processing: The survey covers resource-allocation and signal-processing schemes for channel estimation, combining, precoding, user association, and power control.These algorithms are presented as central tools for improving CF mMIMO system performance.

A. Channel Estimation

CF mMIMO channel estimation seeks accurate CSI under time variation and imperfect statistical knowledge. The survey compares pilot-based estimators, emphasizing MMSE accuracy against LS simplicity and EW-MMSE efficiency.

  • Motivation: Accurate, resource-efficient channel estimation is vital because coherent CF mMIMO processing requires CSI, while perfect CSI at APs and CPUs is unrealistic.The channels are time-varying, motivating practical estimation methods.
  • Estimation metric: NMSE measures relative channel-estimation error by normalizing summed MSEs with summed channel gains, so strong channels dominate the collective metric.This avoids comparing absolute errors without accounting for average channel gain.
  • Pilot-based estimators: MMSE minimizes MSE and NMSE and is optimal from that perspective, so the other estimators described provide larger MSEs.MMSE relies on channel statistics, unlike LS, which requires no prior statistical information.
  • Pilot-based estimators: EW-MMSE uses only diagonal spatial-correlation information, reducing statistical requirements and computational complexity but producing larger errors when correlations are ignored.Its performance loss disappears when all spatial correlation matrices are diagonal.
  • Implementation: CF mMIMO estimators can be computed separately at each AP because channel vectors are independent between APs, unlike centralized cellular-mMIMO processing.The estimator expressions are tailored to CF mMIMO because pilot contamination affects the systems differently.
  • Comparison: In spatially correlated Rayleigh fading, MMSE-type estimators significantly outperform LS, while EW-MMSE has larger average NMSE than MMSE.The comparison uses K = 50 UEs, L = 100 APs, and N = 4 antennas per AP across uplink transmit power p.

4) Other Methods:

Other channel-estimation methods address pilot contamination, missing statistics, fronthaul overhead, and FDD feedback constraints. The survey covers TOA-based, blind, and angle-reciprocity-based approaches with distinct practical trade-offs.

  • TOA-based estimation: A TOA-based pilot-contamination scheme estimates multipath arrival times and filters paths from distant APs, outperforming LS when TOA estimation is accurate.Its benefit is therefore conditioned on accurate TOA estimation.
  • TOA-based estimation: TOA-based estimation uses only the currently received signal and does not require channel statistics, making it attractive for practical implementation.This is identified as the scheme’s key advantage.
  • Blind estimation: Blind channel estimation is attractive in CF mMIMO because exchanging CSI between APs and the CPU creates additional fronthaul overhead.ICA has been used to separate and decode received signals while estimating channels.
  • Blind estimation: The ICA-based method suffers an error floor at high SINR because of inadequate ambiguity elimination.Reference bits are used to identify the desired signal among signals within a cell.
  • FDD estimation: In FDD CF mMIMO, a DFT-based scheme exploits uplink-downlink angle reciprocity, with CSI acquisition overhead scaling only with the number of served users.The method uses a likelihood function and tiny angle rotations to reduce training overhead.

B. Receive Combining

Receive combining uses estimated CSI to transform uplink channels into effective scalar channels, balancing desired-signal gain, interference suppression, complexity, and scalability. The survey spans MR, MMSE-like, ZF-like, and partial schemes.

  • Overview: Receive combining strengthens desired signals relative to interference and noise, but different schemes produce substantially different spectral efficiencies and computational complexities.It may operate centrally or in a distributed fashion.
  • MR combining: MR has the lowest complexity and coherently combines received desired-signal energy because its combining vector matches the desired UE’s channel.MR may be less preferable when interfering signals are present.
  • MMSE-like combining: C-MMSE maximizes each UE’s SINR, but its LN × LN matrix inversion creates relatively high computational complexity.MMSE-like schemes whiten the received signal before applying matched combining.
  • Scalable combining: L-MMSE and C-MMSE complexities grow with K because APs must compute all K MMSE channel estimates, making them unscalable.P-MMSE and LP-MMSE limit processing to UEs sharing partially the same serving APs.
  • ZF-like combining: FZF suppresses interference using pilot degrees of freedom, whereas P-FZF and PWP-FZF use fewer degrees of freedom and retain larger array gain.PWP-FZF additionally protects weak UEs by reducing intra-group interference.
  • ZF-like combining: Distributed FZF has much lower complexity than centralized ZF while supporting coordinated, scalable operation.Partial FZF further reduces complexity because the number of selected pilots is no larger than τp.

4) MMSE-SIC Combining:

The survey evaluates MMSE-SIC and scalable combining alongside uplink and downlink processing methods. The results emphasize interference suppression, near-equivalent scalable performance, and OTA precoding to reduce fronthaul signaling.

  • MMSE-SIC Combining: Non-linear MMSE-SIC provides only a minor average-uplink-SE gain over linear MMSE when favorable propagation exists.SIC is most effective with a few strongly interfering UEs, whereas CF mMIMO typically has many weakly interfering UEs.
  • MMSE-SIC Combining: MR trails ZF-based schemes, especially for UEs with large channel gains, because ZF-based combining suppresses inter-user interference.The comparison uses L = 25, K = 10, N = 8, τp = 7, and pk = 100 mW per UE with LSFD.
  • MMSE-SIC Combining: P-FZF and PWP-FZF outperform FZF by spending only τTl rather than τp degrees of freedom and gaining more array gain.PWP-FZF provides a higher 95%-likely SE than P-FZF by protecting weak UEs.
  • MMSE-SIC Combining: L-RZF and L-MMSE have a small performance gap and outperform the other evaluated combining schemes.This result is reported for the Fig. 6 uplink-SE CDF comparison.
  • Scalable combining: C-MMSE and P-MMSE outperform L-MMSE and LP-MMSE by exploiting more CSI to suppress interference, while scalable schemes have almost the same performance as their counterparts.The scalable schemes incur negligible performance loss from limiting the number of APs serving each UE.
  • Transmit precoding: Downlink precoding is more complicated than uplink combining because each UE’s downlink SE depends on all UEs’ precoding vectors.Uplink-downlink duality allows the surveyed linear combining schemes to design corresponding precoders.
  • Transmit precoding: OTA precoding iteratively exchanges CSI through an additional uplink signaling resource, allowing APs to acquire cross-term information without extensive fronthaul signaling.The procedure alternates AP transmission, UE combining, and AP precoding updates until a stopping criterion is met.

D. User Access and Association

User access and association in CF mMIMO assigns UEs serving APs and pilot resources while addressing pilot contamination, AP capacity, scalability, and crowded-network access.

  • User access: CF mMIMO user access covers AP selection, pilot assignment, user activity detection, and AP switch on/off strategies.These procedures assign resources when a UE accesses the network.
  • AP selection: Serving every UE with all APs is impractical because pilot shortage limits how many UEs each AP can serve.The practical design therefore selects a subset of APs for each UE.
  • AP selection: Large-scale AP selection chooses up to L dominant APs using the largest large-scale fading coefficients and associates UEs and APs in a user-centric manner.A threshold can indicate the settled percentage of total received power contributed by selected APs.
  • AP selection: Competition-based AP selection lets UEs compete for APs, prioritizing better channel conditions while limiting an AP to at most τp UEs.The procedure retains an AP for a UE that loses all other competitions, helping prevent weak UEs from being abandoned.
  • Pilot assignment: Pilot contamination degrades channel estimation and interference rejection, making pilot assignment critical for CF mMIMO performance.Random, greedy, clustering, graph-coloring, tabu, Hungarian, joint, K-means, and user-group approaches are surveyed; some are unscalable because complexity grows polynomially with APs and UEs.
  • Pilot assignment: Scalable pilot schemes include joint AP selection and assignment, K-means clustering, and user-group methods, while the survey identifies efficient access in crowded CF mMIMO systems as an open issue.The joint method assigns a low-interference pilot through a Master AP and informs selected neighboring APs.

4) AP Switch On/Off:

AP switch on/off (ASO) treats AP activity as an optimization variable to serve dynamic traffic efficiently, improve energy efficiency, and reduce unnecessary active infrastructure.

  • AP Switch On/Off: ASO dynamically turns APs on or off according to served-user locations and generated data traffic.The objective is to exploit only competent APs needed for current traffic rather than keeping all APs active.
  • AP Switch On/Off: ASO aims to improve system energy efficiency by using some, not all, competent APs for dynamic traffic requests.Neighboring APs may fill users’ spectral-efficiency requirements without universal AP activation.
  • ASO algorithms: A globally optimal ASO solution can be obtained by solving a mixed-integer second-order cone program, but its non-convexity motivates low-complexity heuristics.Other methods use location and propagation-loss information to construct heuristic ASO algorithms.
  • ASO algorithms: ChiS-ASO, KS-ASO, and LSE-ASO significantly outperform RS-ASO by matching active-AP spatial distributions to UE distributions.OG-ASO provides the best SE, whereas MPL-ASO offers a complexity–SE tradeoff with a minor performance penalty.
  • Resource allocation: CF mMIMO resource allocation surveys max-min fairness, max sum SE, and max EE utilities alongside corresponding power-control and power-allocation methods.UE uplink and AP downlink transmit powers are optimized through system-wide utility functions.
  • Resource allocation: System-wide max-min methods become unscalable as the UE count grows, motivating distributed and heuristic schemes in large CF networks.Local decisions can limit each device’s involvement with other devices.

2) Max Sum SE:

Max sum SE optimization targets overall spectral efficiency but is generally non-convex, so CF mMIMO studies use iterative, machine-learning, and energy-efficiency-oriented approaches under practical constraints.

  • Max Sum SE: Max sum SE represents overall network spectral efficiency, allowing most UEs to gain larger SE while minimally affecting poorly served UEs.This objective differs from maximizing the SE of a specific UE.
  • Max Sum SE: The max sum SE problem is usually non-convex, making the optimal solution difficult to obtain and often requiring local-optimum methods.Alternating optimization is one approach for addressing the non-convexity.
  • Optimization methods: Successive convex approximation replaces non-convex terms with convex approximations to optimize sum SE iteratively.The approach has been applied under hardware impairment, short-term power constraints, and low-resolution ADCs.
  • Machine learning: Machine-learning methods use inputs such as UE positions or large-scale fading information to predict max sum SE power-control policies.Surveyed approaches include artificial neural networks, deep convolutional neural networks, and deep neural networks.
  • Energy efficiency: Energy-efficiency optimization accounts for information transmission and total power consumption, including transmit, analog-processing, digital-processing, and transceiver-chain terms.The resulting max EE problem is non-convex and is addressed with alternating optimization, SCA, geometric programming, or related methods.
  • Energy efficiency: A first-order non-convex-programming method achieves the same performance with faster runtime than second-order optimization methods.Second-order methods do not scale favorably with network size, motivating first-order alternatives.
  • Practical implementation: Practical CF mMIMO must address limited fronthaul capacity and hardware impairment because many surveyed algorithms assume error-free fronthaul and perfect hardware.Fronthaul remedies include signal quantization and structured lattice codes.
  • Practical implementation: Estimate-compress-forward reduces fronthaul load by compressing channel estimates and data signals at each AP before centralized CPU detection.Low-resolution ADCs are also presented as a promising option for reducing load, power consumption, and hardware cost.

2) Compute-and-Forward:

Compute-and-forward schemes reduce fronthaul requirements by forwarding integer-linear combinations, while expanded variants support unequal UE powers and different noise levels. The broader discussion also covers quantization, hardware impairments, cooperation, and deployment trade-offs in CF mMIMO.

  • Compute-and-Forward: C&F forwards integer-linear combinations of all UEs’ transmitted signals, potentially achieving the minimum fronthaul requirement with lossless transmission.The served UEs are selected through coefficient vectors constrained by the computation rate.
  • Compute-and-Forward: C&F is limited to symmetric scenarios with equal UE transmit powers, whereas E-C&F accommodates unequal power allocation and different noise levels.Unequal power can compensate for pathloss, and E-C&F distributes UE power unequally to minimize computation-rate loss.
  • Compute-and-Forward: E-C&F noticeably outperforms conventional C&F in achievable sum-rate because its power optimization facilitates exploitation of channel conditions.The comparison uses K = 8, L = 20, and p = 200 mW while varying the number of APs.
  • Quantization and Fronthaul: Quantization reduces fronthaul load but uniform quantization introduces error and can cause considerable performance loss.Different transmission schemes also require different CSI/data fronthaul allocations and AP processing capabilities.
  • Quantization and Fronthaul: EMCFW provides larger SE than EMCF because its optimized receiver-filter coefficients maximize the SNR, while EMCF can outperform other schemes in numerical results.The schemes are compared as average SE versus quantization bits, with EMF and EMFW using perfect ADCs and no AP compression.
  • Practical Implementation: CF mMIMO implementation requires balancing performance gain against deployment cost, energy consumption, hardware impairments, and fronthaul topology.Dummy APs can reduce deployment cost by compressing and forwarding received signals, but at the cost of performance loss; practical networks may use multiple CPUs or radio-stripes topologies.
  • Hardware Impairments: Hardware-impairment effects at APs vanish as the number of APs increases, but secure transmission generally does not follow the same hardware-quality scaling law.Other results report that LSFD provides the largest SE under hardware impairments and that an arbitrary UE’s achievable rate converges to a finite ADC-resolution-independent limit as APs increase.
  • Future Directions: Future CF mMIMO algorithms should exploit detailed fronthaul topology, including multiple CPUs and parallel or sequential connections, rather than assuming a single-CPU star topology.The canonical star topology is considered unlikely for geographically large networks, motivating adaptations to sequential processing and alternative architectures.

B. Synchronization

CF mMIMO requires accurate synchronization and reciprocity calibration because distributed APs must coordinate coherent transmission. Practical implementations use wired or over-the-air synchronization, while perfect timing synchronization remains physically unattainable over large networks.

  • Synchronization: Coherent signal processing requires sufficiently accurate relative timing among APs and regular synchronization to limit reciprocity and synchronization errors.The system must maintain coherence during transmission, with corrections applied at regular intervals.
  • Synchronization: AP synchronization and TDD reciprocity calibration are critical problems for enabling CF mMIMO.Unsynchronized APs and local oscillators complicate coordinated operation, especially when fronthaul cannot provide a sufficiently accurate common reference.
  • Synchronization: Wired synchronization can use IEEE 1588 PTPv2, while GPS and over-the-air signaling provide alternatives when precise common references are unavailable.AirSync detects OFDM slot boundaries within the cyclic prefix and predicts instantaneous carrier phase.
  • Synchronization: Theoretical CF mMIMO studies often assume perfect timing synchronization, but this assumption is physically impossible over large networks.This gap leaves room for further theoretical advances addressing nonideal synchronization.
Loading 2104.13667v2…