The road to 6G: Ten physical layer challenges for communicationsengineers

Matthaiou, M., Yurduseven, O., Ngo, H-Q., Morales-Jimenez, D., Cotton, S., & Fusco, V. (2021). The road to 6G:Ten physical layer challenges for communications engineers. IEEE Communications Magazine, 59(1).https://doi.org/10.1109/MCOM.001.2000208

Published in:IEEE Communications Magazine

Document Version:Peer reviewed version

Queen's University Belfast - Research Portal:Link to publication record in Queen's University Belfast Research Portal

Publisher rightsCopyright 2020 IEEE. This work is made available online in accordance with the publisher’s policies. Please refer to any applicable terms ofuse of the publisher.

General rightsCopyright for the publications made accessible via the Queen's University Belfast Research Portal is retained by the author(s) and / or othercopyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associatedwith these rights.

Take down policyThe Research Portal is Queen's institutional repository that provides access to Queen's research output. Every effort has been made toensure that content in the Research Portal does not infringe any person's rights, or applicable UK laws. If you discover content in theResearch Portal that you believe breaches copyright or violates any law, please contact openaccess@qub.ac.uk.

Download date:12. Jul. 2021

1

The Road to 6G: Ten Physical Layer Challenges forCommunications Engineers

Michail Matthaiou, Okan Yurduseven, Hien Quoc Ngo, David Morales-Jimenez,Simon L. Cotton, and Vincent F. Fusco

Abstract—While the deployment of 5G cellular systemswill continue well in to the next decade, much interestis already being generated towards technologies that willunderlie its successor, 6G. Undeniably, 5G will have trans-formative impact on the way we live and communicate,yet, it is still far away from supporting the Internet-of-Everything (IoE), where upwards of a million devices perkm3 (both terrestrial and aerial) will require ubiquitous,reliable, low-latency connectivity. This article looks at someof the fundamental problems that pertain to key physicallayer enablers for 6G. This includes highlighting challengesrelated to intelligent reflecting surfaces, cell-free massiveMIMO and THz communications. Our analysis coverstheoretical modeling challenges, hardware implementationissues and scalability among others. The article concludesby delineating the critical role of signal processing in thenew era for wireless communications.

I. INTRODUCTION

Each generation of cellular networks has an av-erage timescale of 10 years from original concep-tion to full commercialization. Now, in the eve anew decade, 5G is becoming a commercial realitywhile all mobile operators and vendors around theworld are competing for market primacy. 5G wasoriginally envisioned to be a technological shiftcompared to 4G in terms of aggregate spectral andenergy efficiency, as well as, latency and reliability[1]. After nearly eight years of intensive academicresearch and industrial testing, the lessons we havetaken are the following: a) 5G can indeed supportemerging data-hungry applications (e.g. ultra-fastbroadband, high-definition video streaming), mainlythrough advances in the massive MIMO (MaMi)space; b) 5G is still falling short of supportingthe so-called Internet-of-Everything (IoE), wheremyriads of devices in a geographic cube require ei-ther low-latency, ultra-reliable connectivity or wire-less Gpbs Internet access by availing of the mm-

The authors are with the Institute of Electronics, Communicationsand Information Technology (ECIT), Queen’s University, Belfast,Belfast BT3 9DT, U.K. (e-mail: m.matthaiou@qub.ac.uk).

wave/THz spectrum [2]. Having said the above, anumber of experts are asking the grand questions:“what comes after 5G?” and “are we approachingthe limits of wireless communications?”. The an-swers to these questions are not straightforward,especially taking into account the ever increasingnumber of connected devices for the years to come.In this context, the idea of 6G has tentatively startedto circulate within the wireless community, with thefirst overview articles appearing in the literature (seefor example [2], [3] and references therein). Thecommon ground of these narrative articles is that 6Gwill try to address the shortcomings of 5G by boldly(a) pushing the communication to higher frequencybands (mm-wave and THz), (b) creating smart radioenvironments through reconfigurable surfaces and(c) by removing the conventional cell structures(aka cell-free massive MIMO (CF MaMi)). Unfor-tunately, transforming these speculative academicconcepts into commercially viable solutions is avery challenging exercise. In this article, we identifyten of the most important challenges that need tobe addressed in at the physical layer to boost follow-up research in the 6G ecosystem. Our discussioncovers a broad spectrum of topics, from fundamentalelectromagnetic (EM) and signal processing (SP)topics to transceiver design and hardware implemen-tations.

II. INTELLIGENT REFLECTING SURFACES (IRS)

A reflective surface is a planar aperture synthe-sized using an array of sub-wavelength elements (orunit cells). Such a surface can be used to modulatethe incoming waves into a desired wavefront uponreflection. Due to their sub-wavelength unit cellsampling, reflective surfaces can be considered adistinct form of metasurfaces, synthesized using anarray of metamaterial unit cell elements. There hasbeen a substantial amount of research conductedin EM wave control using metasurface apertures

2

with applications ranging from imaging to EMinvisibility [4]. Despite the fact that these structuresare well understood within the applied EM com-munity, their adoption in wireless communicationnetworks has been under investigated to date. Withthe recent implementation of the 5G technology andthe upcoming 6G networks, reflective surfaces offerthe potential to substantially reduce the cost andcomplexity of the hardware layer while improvingthe total energy efficiency.

Wireless systems conventionally rely on a matureantenna technology to establish a communicationlink, which is known as phased arrays [5]. A phasedarray consists of individual antennas with dedicatedphased shifting circuits and power amplifiers tosynthesize the desired aperture wavefront. Phasedarrays can be power hungry and exhibit a rathercomplex hardware architecture. Different from thephased array technology, metasurfaces do not relyon phase shifting circuits to modulate the phaseof the incoming waves. Instead, metasurfaces relyon a holographic principle to achieve the desiredphase modulation. The incoming wave illuminatingthe aperture surface acts as a reference-wave, whichis converted to a desired wavefront upon reflectionfrom the reflective metasurface aperture. Hence, thereflective surface can be thought of as the product ofthe interaction between the incident reference-waveand the desired reflection wavefront. Their majoradvantage is that they can synthesize any arbitrarywaveform using this simple, yet strong, holographicprinciple without the need for expensive and powerhungry phase shifters. Designing a reflective surfacefor wireless systems exhibits two main challenges:

Challenge 1: Unit cell phase range and phasequantization levels

An important limitation in the design process ofa reflective surface is the achievable phase rangeof the unit cells synthesizing the aperture. Ideally,each unit cell across the reflective surface shouldprovide a full phase control across a phase range of0-2π. However, achieving this full-phase range mayrequire a rather complex unit cell topology, withmultiple layers complicating the hardware archi-tecture. Another important constraint in the designprocess of a reflective surface is the quantization ofthe phase range of the unit cells. Even if a full phaserange of 0-2π is achieved, the number of quan-tization levels used to discretize this phase rangehas a direct impact on the fidelity of the synthe-

sized wavefront upon reflection from the reflectivesurface. To investigate the effect of these designconstraints, we consider a reflective metasurface andstudy these cases individually. First, we assume thatthe unit cells can alter the phase response of the in-coming reference-wave across the full phase-range,0-2π. For this scenario, we investigate two differentquantization levels, 1-bit and 4-bit. We consider thatthe reflective metasurface is illuminated by an ar-bitrarily selected plane-wave incident along opticalaxis (z-axis) of the surface. The objective function isdefined to be a wavefront radiated in the broadsidedirection (θ=0◦, φ=0◦) upon reflection from thesurface. The reflective surface patterns shown inFigs. 1(a) and (b) are generated in MATLAB whilewe make use of a full-wave EM simulation software,CST Microwave Studio, to simulate the radiationpattern of the reflective surface. Fig. 1(c) shows theradiation pattern of the surface vs varying phasequantization levels.

Fig. 1: 3D Radiation pattern of the reflective surfaceas a function of phase quantization level (a) 1-bitunit cell phase quantization (b) 4-bit unit cell phasequantization (c) comparison of the radiation patternsalong the azimuth plane (H-plane) and elevationplane (E-plane). Colorbar in dB scale.

3

In Fig. 1, the H-plane is the azimuth plane (orxz-plane) with φ=0◦. Similarly, the E-plane is theelevation plane (or yz-plane) with φ=90◦. Although,from a design perspective, using a 1-bit quantizationlevel would be rather simple, in comparison tothe 4-bit quantization case, the radiation patternexhibits substantially higher sidelobes (8 dB higher)and less directivity (21 dB lower in the broad-side direction). Increasing sidelobes is a significantconcern for wireless communication networks dueto the possibility to introduce strong interferenceacross other channels. Similarly, reduced directivityadversely affects the link budget. As a result, fromFig. 1, it is evident that there is a direct relationshipbetween the phase quantization level and the fidelityof the radiation pattern of the reflective surface.We now study the effect of the achievable unitcell phase range on the radiation pattern of thereflective surface. We select the quantization levelto be maximum (4-bit) and investigate two unitcell phase ranges, 0-π and 0-2π, respectively. Theradiation pattern of the reflectarray surface vs theachievable unit cell phase range is shown in Fig. 2.

Fig. 2: Radiation pattern of the reflective surface asa function of unit cell phase range.

From Fig. 2, limiting the unit cell phase range to0-π reduces the directivity of the surface by 7.3 dBwhile increasing the first sidelobe levels by 4 dBand secondary sidelobe levels by as much as 7 dB.From Figs. 1 and 2, to have a realistic estimate of thelink budget in a communication channel and a betterunderstanding of interference characteristics, it isimportant that the design limitations of reflectivesurfaces are taken into account. Yet, it is often as-sumed that the reflective surface is ideal, suggestingthat the link budget calculations do not consider thephase quantization errors and assume that unit cells

can modulate the phase response of the incidentreference-wave across the full-spectrum, 0-2π.

Challenge 2: Dynamic reconfigurability and IRSAlbeit producing highly desirable radiation char-

acteristics, conventional reflective surfaces arestatic, hence, beam characteristics are hard-codedinto the surface during the design process. Differentfrom static metasurfaces, an IRS has the capabilityto dynamically tune the reflection response of theaperture in an all-electronic manner. This is par-ticularly important as communication environmentshave dynamic characteristics, namely variations inthe number of connected users and non-static lo-cation distribution over time. Thus, the capabilityto intelligently change the characteristics of thereflected wavefront to meet their dynamic metricsplays a crucial role in future wireless communica-tion systems. The dynamic modulation of the IRScan be achieved using low-power semiconductorelements, e.g., PIN diodes [6]. This advantage ofIRS makes them a promising candidate to replacethe power-hungry phase shifting apertures.

Fig. 3: IRS aperture and electronically steered radi-ation patterns (a) (θ=-30◦, φ=-30◦) (b) (θ=30◦, φ=-30◦) (c) multibeam pointing in (θ=±30◦, φ=±30◦).

In Fig. 3, we study an IRS that can dynamicallycontrol the direction of the reflection. Similar to theunit cell phase studies, for the IRS shown in Fig. 3the reference-wave is selected to be incident alongthe optical axis of the reflective surface (z-axis) andit is fixed for all the beam-steering scenarios studiedin Fig. 3. Moreover, we consider a 4-bit phasequantization and assume that the unit cells exhibit aphase range of 0-2π. Once these design challengesare satisfied, IRS can produce reconfigurable radia-tion patterns to an excellent fidelity, rendering themperfectly suitable for 6G networks.

III. CELL-FREE MASSIVE MIMOCell-free massive MIMO (CF MaMi) has been

proposed in [7] to overcome the boundary effect

4

of cellular networks. In CF MaMi, many accesspoints (APs) distributed in a geographic coveragearea coherently serve many users in the same time-frequency resources. There are no cells, and hence,no boundary effects. Key points of CF MaMi:

• CF MaMi relies on MaMi technology. Moreprecisely, using many APs, CF MaMi offersmany degrees of freedom, high multiplexinggain, and high array gain. As a result, it canprovide huge energy efficiency and spectralefficiency. These gains can be obtained withsimple SP due to the favorable propagation andchannel hardening properties of MaMi tech-nology. Note that in CF MaMi, we still havechannel hardening, but potentially at a reducedlevel to that observed in colocated MaMi.

• In CF MaMi, the service antennas (APs) aredistributed over the whole network, and hence,we can obtain macro-diversity gains. As aresult, CF MaMi can offer a very good networkconnectivity, i.e., it provides good services forall users in the networks. There is no deadzone. Figures 4a–4b show the downlink achiev-able rates displayed with scaled colors forCF MaMi and colocated MaMi, respectively.Clearly, CF MaMi can offer much more uni-form connectivity for all users.

• Different with colocated MaMi where the basestation is equipped with very large antennas, inCF MaMi each AP has a few antennas. Thus,CF MaMi is expected to be built by low-cost,low-power components and simple SP APs.

The above benefits (in particular the high networkconnectivity) fulfil the main requirements of futurewireless networks. Therefore, CF MaMi has becomeone of the promising technologies of beyond 5G andtowards 6G wireless networks, and attracted a lot ofresearch interest, see [8] and references therein.

Challenge 3: Practical user-centric approachesIn canonical CF MaMi [7], all APs participate in

serving all users through the backhaul connectionswith one or several central processing units (CPUs).This is not scalable in the sense that it is notimplementable when the network size (number ofAPs and/or number of users) grows large. Designinga scalable structure is one of the main challengesof CF MaMi. It is shown in [7] that, owing tothe path loss, only 10-20% of the total numberof APs really participate in serving a given user.

-1 -0.5 0 0.5 1

Position (km)

-1

-0.5

0

0.5

1

Positio

n (

km

)

1.5

2

2.5

3

3.5

4

4.5

Downlink Rate

(a) CF MaMi

-1 -0.5 0 0.5 1

Position (km)

-1

-0.5

0

0.5

1

Positio

n (

km

)

1

2

3

4

5

6Downlink Rate

(b) Colocated MaMi

Fig. 4: The downlink achievable rates for differentuser locations displayed with scaled colors (obtainedusing simulation approach described in [7]); (a): 100APs are uniformly located at random in a 2km×2kmarea; (b): all 100 service antennas are located at theoriginal point (i.e. the center of the square area).

Thus, each user should be served by a subset (notall) of APs. There are two ways to implement this:network-centric and user-centric approaches. In thenetwork-centric approach, the APs are divided intodisjoint clusters. The APs in a cluster coherentlyserve the users in their joint coverage area. Network-centric-based systems still have boundaries, andhence, are not suitable for CF MaMi. By contrast,in user-centric approach, each user is served by itsselected subset of APs. There are no boundaries,and hence, user-centric approach is a suitable wayto implement CF MaMi. There are several simplemethods to implement user-centric approach, suchas each user chooses some of its closest APs orchooses a subset of APs which contribute most

5

of the total received power of the desired signal[9]. Yet, existing methods are not optimal, andstill require huge connections from all APs to theCPUs. In addition, the cluster formed by each userchanges quickly depending on the user locations.This requires more control signaling. Hence, design-ing a practical user-centric approach is a challengingexercise and requires urgently intensive research.

Challenge 4: Scalable power controlPower control is central in CF MaMi since it

controls the near-far effects and the interuser inter-ference to optimize the objectives (e.g. the max-min fairness or the total energy efficiency) we wantto achieve. Ideally, power control is done at theCPU under the assumption that the CPU perfectlyknows all large-scale fading coefficients. Then, theoptimal power control coefficients will be sent tothe APs (for the downlink transmission) and to theusers (for the uplink transmission). This requireshuge front/back-hauling overhead. Yet, it is verydifficult for the CPU to have perfect knowledgeof large-scale fading coefficients associated witha potentially unprecedented number of APs andusers. Thus, the above power control method is notscalable with the network size. To make the systemscalable, power control should be done distributedat the APs with local knowledge of the channelconditions. This is again problematic because it ishard to control the near-far effects and interuserinterference without full channel knowledge of alllinks from all APs and users. Some heuristic powercontrol schemes have been proposed, which, how-ever, are developed based on a specific assumptionof the propagation environment, and hence, it ishard to evaluate how well these schemes work inpractice. Promising approaches based on machinelearning (ML) and deep learning (DL) have recentlybeen proposed in the IoT context [10]. An importantquestion is whether these approaches are also scal-able, in order to meet the foreseeable dimensionsand decentralization of CF MaMi networks.

Challenge 5: Advanced distributed SPOne of the ultimate aims of CF MaMi research

is designing a SP scheme which offers good per-formance and can be implemented in a distributedmanner. Otherwise, the system is not scalable. Thisis very challenging. In canonical CF MaMi [7],conjugate beamforming is proposed to use sinceit can be implemented in distributed manner andperforms well. However, compared to other linear

processing schemes such as zero-forcing (ZF) andminimum mean-square error (MMSE), the perfor-mance of conjugate beamforming is far below. Tocover the gap between conjugate beamforming andZF/MMSE, we need to have additionally very largenumber of service antennas. CF MaMi with localzero-forcing scheme was proposed in [11]. How-ever, this scheme requires that each AP has a largenumber of antennas. It is more challenging for theuplink designs. Currently, there are no distributedSP schemes available for the uplink in CF MaMi.Even with the simple matched filtering, we need tosend the (processed) signals from each AP to theCPUs for signal detection.

IV. MOVING TO HIGHER FREQUENCY BANDS

Future wireless systems for 6G will potentiallyrely on: i) Millimetre-wave technologies (30 to 300GHz); ii) THz technologies (300GHz to 3 THz); andiii) Free space optics (FSO). For example, utilizingspectrum available from 100 GHz to 1 THz, willhave the potential to deliver data rates in excess of10 Gbps [12].

Challenge 6: Packaging/interconnect techniquesIn addition to providing for a physical enclosure,packaging must provide a reliable interconnectionbetween interior and exterior operating environ-ments. Some of the key driving factors includeintegration of high-speed semiconductor integratedcircuits with advanced antenna systems and inte-gration with optoelectronics. Materials used mustbe characterized at THz frequencies as this data isnot readily available. At higher frequencies, bondwires cause considerable signal degradation. Theeffects of bond wires are difficult to characterize forlarge signal applications, such as power amplifiersand also for phase critical applications, such asbeamformers for phased arrays, where these canintroduce undesirable side lobe levels. Traditionalmetallic split-block packages provide excellent per-formance but are bulky and heavy. Recent activ-ities indicate noticeable progress in interconnectand packaging technologies for THz applications.Cutting-edge techniques in micro-machining andLTCC technology yield compact and low-cost so-lutions. Additive manufacturing techniques, such asmetal coated 3D printing of plastic devices canrealize low-cost, light weight and compact devices.

Challenge 7: Transceiver designThe compact physical size and power efficiency

6

requirements become more challenging at mm-waveand THz frequencies. Hybrid beam forming willbe best suited to implement large number of an-tenna elements along with high efficiency ampli-fiers. Performance parameters, such as the noisefigure, output power and power efficiency degradesignificantly for high operating frequencies. De-modulation of higher order modulated signals alsobecomes more challenging as phase noise increasesat higher frequencies. Advanced array SP techniquesmust complement the transceiver design to addressthese challenges. Novel techniques, such as spatiallyoversampled antennas and new phased array archi-tectures can be leveraged to provide a solution tosize, weight and power consumption of large mm-wave/THz antenna arrays.

Challenge 8: Measurements & standardizationTwo popular approaches for phase-sensitive mea-surements at THz use either vector network ana-lyzers (VNAs) or time domain spectrometers [13].In both cases, calibration, verification, and measure-ment traceability at THz frequency bands remainsa major challenge. For time domain systems, amajor challenge is the establishment of standardizedmeasurement and calibration, whereas for VNAsystems, solutions are being sought for high pre-cision waveguides and interconnects. Electro-opticsampling (EOS) shows much promise as a comple-mentary approach to THz measurements, althoughit is yet to extend the bandwidth to 1.5 THz andto improve resolution. Beyond this, there is also apressing need for standardization. More precisely,the establishment of a robust framework which en-ables meteorological traceability to the InternationalSystem of Units for THz measurements.

V. THE ROLE OF SP IN THE 6G ERA

The journey towards 6G will inevitably be hur-dled by significant challenges in the SP arena. Mas-sively populated and decentralized (cell-free) net-works, supporting unprecedented Internet of Every-thing connectivity, will produce high-dimensionaland increasingly complex signals, which are subjectto increased interference and other impairments(e.g., related to synchronization, temporal corre-lation etc.) that have so far been largely over-looked. Current SP methods, generally based onlow-dimensional signals and classic stationarity as-sumptions, will indeed need to be rethought. Wenow discuss two of these methods, namely channel

estimation and adaptive filtering, and their associ-ated challenges.

Challenge 9: Channel estimationExtensive research in the context of 5G has

been concerned with the need to reduce trainingoverheads in pilot-based channel estimation. Whilesuch need will be most critical in 6G networks,viable solutions become extremely challenging dueto the massive scale-up and connectivity demands,in particular: (i) supporting high data rates (Gbps)in high-mobility scenarios—a prime concern ofoperators—will require dealing with much shorterchannel coherence times; (ii) ultra-low latency re-quirements will see transmission intervals substan-tially shortened, and (iii) the number of parametersto estimate will be massively large as a consequenceof the scaling (not only of antennas/access points,but also in the numbers of users/devices). Thehigh-dimensional channels will therefore need tobe estimated in severe under-sampling constraints,which might render pilot-based (coherent) estima-tion approaches unfeasible, particularly under high-mobility or low-latency requirements. Blind (non-coherent) estimation approaches—which do not re-quire dedicated pilot signals—stand as promisingalternatives and might play a key role. However,these approaches typically require knowledge ofthe (high-dimensional) received signal covariancematrix [14], which will again need to be acquiredfrom a limited number of samples. To that end, re-search in the fields of random matrix theory (RMT)and high-dimensional statistics will be relevant; inparticular, to develop accurate estimators of largecovariance matrices (and their eigen-spectrum) un-der limited sampling. Other, more recent approachesbased on machine learning (ML) might also play asignificant role (see [10] and references therein).

Challenge 10: Adaptive filteringIn beamforming, transmitted signals are dynam-

ically adapted (via digital precoding) to the propa-gation conditions, effectively mitigating interferenceand noise. Adaptive beamforming can be seen asa linear filter with a particular design objective,e.g., to maximize the signal-to-noise ratio (SNR).Optimal solutions for the beamformer (and associ-ated receiver filters) require the covariance matrixof the aggregated interference and noise; unknownin practice, this needs to be estimated from ob-served samples. Current solutions rely on classicalestimators, such as the sample covariance matrix

7

(SCM), which will return a poor estimate in high-dimensional 6G scenarios, due to:

• Finite sampling (scarcity of samples): Whilethe numbers of antennas and user devices willscale up massively, strict low-latency and high-mobility requirements will impose a rather lim-ited number of training (observed) samples.

• Temporal correlation: The ultra-dense andhighly decentralized deployments, e.g., withthousands of distributed access points in CFnetworks, will be subject to non-perfect syn-chronization (e.g., between interference anddesired signals) which might in turn producenon-stationarity effects.

• Outlying samples: With millions of intercon-nected devices (from electrical/smart appli-ances to connected vehicles), it is also rea-sonable to expect multiple sources of impul-sive noise and, in security-sensitive applica-tions/environments, eventual sources of inten-tional interference (jamming).

Traditional estimators (e.g., SCM) rely on the suffi-cient availability of samples (the number of samplesshould be far greater than the number of signals)and on the assumption that these samples are i.i.d.Under the conditions above, however, the mismatchbetween true and estimated covariance leads tohighly inaccurate filters with severe performancelosses, in terms of connectivity, reliability, and datarates. A major challenge is then to develop filter-ing solutions which are robust to the effects offinite-sampling, temporal correlation, and corrupted(outlying) samples. Relevant tools and promisingdirections can be leveraged from the fields of robuststatistics, RMT, and high-dimensional covarianceestimation (see, e.g. [15] and references therein).

VI. CONCLUSION

Although the 6G era is a decade away, it isextremely timely to understand what are the mainchallenges for communications engineers that re-quire immediate research attention. This article hasidentified ten of the most pressing challenges whoseinvestigation will cross-leverage expertise in sig-nal processing, information theory, electromagneticsand physical implementation.

REFERENCES

[1] J. G. Andrews, et al., “What will 5G be?,” IEEE J. Sel. AreasCommun., vol. 32, no. 6, pp. 1065–82, June 2014.

[2] W. Saad, M. Bennis, and M. Chen, “A vision of 6G wireless sys-tems: Applications, trends, technologies, and open research prob-lems,” [Online], arxiv, 2019. https://arxiv.org/abs/1902.10265.

[3] P. Yang et al., “6G wireless communications: Vision and po-tential techniques,” IEEE Netw., vol. 33, no. 4, pp. 70–75, Jul.2019.

[4] H. T. Chen, A. J. Taylor, and N. Yu, “A review of metasur-faces: Physics and applications,” Reports on Progress in Physics,Vol. 79, No. 7, 076401, 2016.

[5] R. J Mailloux, Phased Array Antenna Handbook, Artech house,2017.

[6] O. Yurduseven et al., “Dynamically reconfigurable holographicmetasurface aperture for a Mills-Cross monochromatic mi-crowave camera,” Optics Express, vol. 26, no. 5, 2018.

[7] H. Q. Ngo, et al., “Cell-free massive MIMO versus small cells,”IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1834–1850,Mar. 2017.

[8] G. Interdonato et al., “Ubiquitous cell-free massive MIMOcommunications,” EURASIP J. Wireless Commun. Netw., vol.2019, no. 1, p. 197, 2019.

[9] H. Q. Ngo et al., “On the total energy efficiency of cell-freemassive MIMO,” IEEE Trans. Green Commun. Netw., vol. 2,no. 1, pp. 25–39, 2017.

[10] M. Chen et al., “Artificial neural networks-based machine learn-ing for wireless networks: A tutorial,” IEEE Commun. SurveysTuts., vol. 21, no. 4, pp. 3039–3071, 2019.

[11] G. Interdonato et al., “Local partial zero-forcing precod-ing for cell-free massive MIMO,” [Online], arxiv, 2019,https://arxiv.org/abs/1909.01034

[12] C. Han et al., “Terahertz communications (TeraCom): Chal-lenges and impact on 6G wireless systems,” [Online], arxiv,2019, http://arxiv.org/abs/1912.06040.

[13] M. Naftaly et al., “Metrology state-of-the-art and challenges inbroadband phase-sensitive terahertz measurements,” Proc. IEEE,vol. 105, no. 6, pp. 1151–1165, Jun. 2017.

[14] H. Q. Ngo and E. G. Larsson, “EVD-based channel estimationin multicell multiuser MIMO systems with very large antennaarrays,” in Proc. IEEE ICASSP, pp. 3249–3252, 2012.

[15] N. Auguin et al., “Large-dimensional behavior of regularizedMaronna’s M-estimators of covariance matrices,” IEEE Trans.Signal Process., vol. 66, no. 13, pp. 3529–3542, Jul. 2018.

AUTHORSMichail Matthaiou is a Reader at Queen’s University

Belfast (QUB), UK. His research spans signal processing,massive MIMO, and mm-wave systems.

Okan Yurduseven is currently a Senior Lecturer at QUB.His research interests include microwave and mm-wave imag-ing, MIMO radar, antennas and metamaterials.

Hien Quoc Ngo is a Lecturer at QUB. His main researchinterests include massive MIMO, cell-free massive MIMO,physical layer security, and cooperative communications.

David Morales-Jimenez is a Lecturer at QUB. His researchinterests span statistical signal processing, random matrixtheory, and high-dimensional statistics.

Simon L. Cotton is a Professor of Wireless Communica-tions at QUB. Among his research interests are propagationmeasurements and statistical channel modeling.

Vincent F. Fusco is the Chair of High Frequency ElectronicEngineering at QUB. His research interests include activeantenna front-end techniques and self-tracking antenna arrays.