Moore's law)has resulted in tiny p time synchronized 177 computationt yMo procne i117)These ing a non-Iss B.Open Problem Th Since ELAAs provide a particular type of spatial channel the co sin a system with thousands of coop ratine correlation,the spectral and energy effi ciency can be com antennas [17],[54].One potential way to re olve this issue results on Mass MIMO with c s to in 28 sor for coding and decoding of the data pplied least in theory MR nre cessing is ver signals.This proc on nient for dis uted deployments sin there is no in the har h The way h ed to g ZF or MMSE 42 trees:there is an electrical connector at one side of the cable Hence.the first major open problem isto implemen the cable can be bent and shaped as you like,no permission is ining scheme uch a quired to deploy it,a d th s to 4(b) ing.in a di multiple radio strip for exan le on might only be we rthwhile to let a part of the ELAA serve per floor in the building.There is no need for fine-tuning the [39.It is not the computationl physica of the intenn when on an abundanc front- eve heir ba eband s S to stripes is that a larg number of antennas share a serial front ocated at a remote location such as an (edge)cloud 1741.181 hau ction to the baseband pro instead of having develop theoreti proce 741 Me pCe mes,circu nta Thi can greatly reduce the cost of the fronthaul infr When using a compact planar as in Fig.3(a).to serve with on ELAA of the type in Fig.4(b) array 28L.42 erenc of the in plan. Each antenna in an ElAa can he constructed using helpful when creating codebooks with beams that the users ins ead of industry-grade lec from,whi is common for system BSs i no It is like a divide and co ach to Bs der wavefronts might not be plane. Second the ant be arbitrarily deployed the array respons vectors ar W-per-ante ante nas in a pr ers n the r-hel he the er is Iw in both cases.Similarly. the uplink signal important but challengin problems.Measurement campaisn noise ratio improved by collecting and improved channel m leling based on physical pr with ray- are I the t channel gain If we to define standardized evaluation scenarios with deterministi an ELAA and a UDN with the same antenna locations. the ELAA achieves stronger receive gnals and less interferenc in compac the af divid has.for example bee en der strated nume rically for cell-free machine-type communications.it is hard to predict its impact networks in [4042 To achieve these gain ennas on ultra-reliable low atency communication services.On the in the hand, n e and ithm hased the air sio mall-scale fading including the sienal hlockases 、that car is described in 57]. indicating that this might be as occur at mmWave frequen s.On the other hand,if we chip scale atomi cks that ne c deplo t in Fig.4a)witth (RF component in Fig.4(b),the latter might be more
6 Moore’s law) has resulted in tiny processors that are incredibly computationally capable, essentially making the computational complexity of MIMO processing a non-issue [16], [17]. These facts can be exploited to achieve a cost-efficient deployment of ELAAs. The challenge lies in the interconnect of all the components in a system with thousands of cooperating antennas [17], [54]. One potential way to resolve this issue is to integrate the antennas and frontends into cables that can be attached to the facade of buildings and then connected to a baseband processor for coding and decoding of the data signals. This processor can either be physically close to the array or located in the basement or cloud. This type of cable is called radio stripes in [57]. The concept is in many ways analog to the string lights that are used to light up (Christmas) trees; there is an electrical connector at one side of the cable, the cable can be bent and shaped as you like, no permission is required to deploy it, and the system continues to operate even if one component breaks down. The deployment in Fig. 4(b) can be achieved using multiple radio stripes, for example, one per floor in the building. There is no need for fine-tuning the physical location of the antennas when one has an abundance of them; in fact, an irregular deployment might even enhance the spatial separability of the users. Another benefit of radio stripes is that a large number of antennas share a serial fronthaul connection to the baseband processor, instead of having a separated connection per antenna, as in the pCell technology [74] and initially considered for Cell-free Massive MIMO [40]. This can greatly reduce the cost of the fronthaul infrastructure. ELAAs will work efficiently even in a cellular deployment, with one ELAA of the type in Fig. 4(b) per cell, if MMSE processing methods are used to cancel inter-cell interference [28], [42]. Each antenna in an ELAA can be constructed using smartphone-grade hardware, instead of the industry-grade hardware used in contemporary BSs, and there are prospects of using even lower-grade hardware to cut cost [75], [76]. It is like a divide-and-conquer approach to BS deployment; if we compare 1 W transmission from a single antenna with 1 M W-per-antenna transmission from M antennas in an array, the latter can lead to M times stronger received signal at the user due to the signal focusing, even if the total output power is 1 W in both cases. Similarly, the uplink signalto-noise ratio (SNR) is improved by collecting more signal energy with an ELAA. This is also where ELAAs differ fundamentally from UDNs, where each user is served only by the antenna that gives the largest channel gain. If we compare an ELAA and a UDN with the same antenna locations, the ELAA achieves stronger received signals and less interference since it coherently combines signals from many antennas to achieve the aforementioned divide-and-conquer gains. This has, for example, been demonstrated numerically for cell-free networks in [40]–[42]. To achieve these gains, the antennas in the ELAA need to be phase-synchronized, which is more complicated to achieve in an ELAA than in a compact array. A synchronization algorithm based on over-the-air signaling is described in [57], indicating that this might be a solvable problem. There are also chip-scale atomic clocks that can be used to keep distributed radio-frequency (RF) components time synchronized [77]. B. Open Problems Since ELAAs provide a particular type of spatial channel correlation, the spectral and energy efficiency can be computed using known results on Massive MIMO with correlated fading [3], [20], [28], [78]. The standard channel estimation, transmit precoding, and receive combining schemes can be readily applied—at least in theory. MR processing is very convenient for distributed deployments since there is no need to share channel knowledge between the antennas in the array [40] but this benefit comes at a huge performance loss compared to using ZF or MMSE processing [28], [42], [79]. Hence, the first major open problem is to implement interference-rejecting precoding/combining schemes, such as ZF and MMSE processing, in a distributed or hierarchical way. When the channel gain variations are large over the array, it might only be worthwhile to let a part of the ELAA serve each user [39], [80]. It is not the computational complexity that is the issue but the front-haul capacity requirements that would be extreme if thousands of antennas need to send their baseband samples to a common processing unit, possibly located at a remote location such as an (edge) cloud [74], [81]. It is important to develop theoretical distributed processing architectures, optimized resource allocation schemes, circuit implementations, and prototypes. Channel modeling is another open problem for ELAAs. When using a compact planar array, as in Fig. 3(a), to serve LoS users located in the far-field, the array response vector can be computed as a function of the azimuth and elevation angles of the incoming plane wave [3, Section 7]. This is helpful when creating codebooks with beams that the users can select from, which is common for systems operating in FDD mode. It is challenging, if not impossible, to do the same for ELAAs. First, the users are in the near-field thus the wavefronts might not be plane. Second, the antennas can be arbitrarily deployed so the array response vectors are unknown, a priori, even for users located in the far-field. Hence, the modeling of channels, as well as the acquisition of the channel features in a particular deployment scenario, are important but challenging problems. Measurement campaigns and improved channel modeling based on physical properties and ray-tracing are needed. If the near-field behaviors turn out to be too complicated to model statistically, it will be necessary to define standardized evaluation scenarios with deterministic channels. Prototyping is also an important challenge and will naturally be more complicated and costly than in compact MIMO systems. While an ELAA is very suitable for broadband and machine-type communications, it is hard to predict its impact on ultra-reliable low-latency communication services. On the one hand, a large number of distributed antennas and an extreme spatial resolution can bring macro diversity against small-scale fading, including the signal blockages that can occur at mmWave frequencies. On the other hand, if we compare the conventional deployment in Fig. 4(a) with the ELAA in Fig. 4(b), the latter might be more susceptible to
large-scale shadowing.Clearly.the reason for deploying BSs in(2)from the latter two.Hence.a()is reconstructed a elevated arge the .In et h .opuica ent but case performance.unless the deployment locations,antenna ase,we can instead illuminate it by density.and operation are well optimized for ultra-reliability. the conjugate e III.DIRECTION 2:HOLOGRAPHIC MASSIVE MIMO minating wave is a pilot waveform emitted by a user device The capacity of Massive MIMO grows monotonically with antennas 3 so wave might be generate suaCe the num t n in a practically viable way?The previous section described sing Mr precoding(also known as conjugate beamforming). e option:physically very large arrays of classicante that are sm Two approaches are currently being taken to approximately Iand well sep realize a continuous microwave aperture: opti 1)The first nach use in the form of a spatially continuous transmitting/receiving aperture. This requires a radically new way of designing using a dense array of conventional fraction-of-spaced mght not ante the g e result will be aphic RF System[59例 are gen mixin differenee the desired RFcarrier frequ ,84 2)The second approach mixes an RF reference signal with num of antennas is the asymptotic limit of Massive MIMO. How can we model communication when using a contin implementation described]uses a single RFpor the receptio on the of the which is connecte the defining characteristic of optical holography where the received field is recorded and later reconstructed.The Fig 6(a).Using the holographic terminology,the RF recording medium a photog phic emulsion)only re port generates the refer ence wave and the ution he tial din h。 network is the hol ographic display.If multip the phase is lost.The holographic rece position of"eams"can be tra ived to process trated in Fig.5.circumvents this issue by mixing enable spatial multiplexing.This is conceptually similar wn n ference wave.Suppose to the hybrid ar a(.)+e+)reaches the [861 and medium.which will record design instead of the complicated beam-training proce 1b(红，2=la(红，)+ea+P dures commonly considered for hybrid beamforming =a(红，P+1+2(a(红，)e-iar+w）.Q) Another embodiment of rec which are :su The last t d he faces that are not actively emitting RF signals.but consis many discrete meta-atom with electronically steerab e是shin d by a ignal sent from another location.the metasurface can control In pr ei(o+)b(,y)2 =a(，+a'(,)e2ax+8 in the ven fre ncy band.Metasurfaces are also knowr +(la(红，gP+1)ear+ (2) as intelligent reflecting surfaces [65].[87]and there are eral variations on these names [88191].N Note that the that the RF ed in
7 large-scale shadowing. Clearly, the reason for deploying BSs at elevated locations is to avoid that the signals are blocked by large objects. There is a risk that ELAAs will greatly improve the average user-performance, but degrade the worstcase performance, unless the deployment locations, antenna density, and operation are well optimized for ultra-reliability. III. DIRECTION 2: HOLOGRAPHIC MASSIVE MIMO The capacity of Massive MIMO grows monotonically with the number of antennas [3] so it would be desirable to have nearly infinitely many antennas, but how can that be deployed in a practically viable way? The previous section described one option: physically very large arrays of classic antennas that are small and well separated to enable essentially invisible deployment. Another option is to integrate an uncountably infinite number of antennas into a limited surface area, in the form of a spatially continuous transmitting/receiving aperture. This requires a radically new way of designing and analyzing antenna arrays; in fact, “array” might not be the right terminology anymore. Research in this direction is taking place under the names of Holographic RF System [59], Holographic Beamforming [60], and Large Intelligent Surface [61], [82]. When a spatially continuous aperture is being used to transmit and receive communication signals, we refer to it as Holographic Massive MIMO since having an infinite number of antennas is the asymptotic limit of Massive MIMO. How can we model communication when using a continuous aperture? Interestingly, the reception and transmission of an electromagnetic field over a continuous aperture is the defining characteristic of optical holography [83], where the received field is recorded and later reconstructed. The recording medium (e.g., a photographic emulsion) only responds to the intensity |a(x, y)| 2 of the received field a(x, y), where x, y are the spatial coordinates on the detector, thus the phase is lost. The holographic recording/reconstruction process, illustrated in Fig. 5, circumvents this issue by mixing the desired wavefront with a known reference wave. Suppose the plane wave e i(αx+βy) is used as reference wave, then the combined wave b(x, y) = a(x, y) + e i(αx+βy) reaches the medium, which will record |b(x, y)| 2 = |a(x, y) + e i(αx+βy) | 2 = |a(x, y)| 2 + 1 + 2< a(x, y)e −i(αx+βy) . (1) The last term in (1) depends on the phase of a(x, y) and the known phase of the reference wave, which means that the phase information has been implicitly recorded. To reconstruct a(x, y), the transparent holographic display is illuminated by a replica of the reference wave, as shown in Fig. 5(b), which yields the reconstructed electromagnetic field e i(αx+βy) |b(x, y)| 2 = a(x, y) + a ∗ (x, y)e i2(αx+βy) + |a(x, y)| 2 + 1 e i(αx+βy) . (2) If the received wave is of finite angular extent, then a judicious choice of the reference wave separates the first component in (2) from the latter two. Hence, a(x, y) is reconstructed and indistinguishable from the original field. In effect, optical holographic recording constitutes a distributed, homodyne receiver. If the display is not transparent but acts as a mirror in the reconstruction phase, we can instead illuminate it by the conjugate e −i(αx+βy) of the reference wave, which leads to emitting the conjugate wavefront a ∗ (x, y) in the opposite direction. Analogously, in wireless communications, the illuminating wave is a pilot waveform emitted by a user device and the reflecting object is the propagation environment, and the reference wave might be generated inside the surface. By emitting the conjugate wavefront from the surface, we can effectively transmit back to the user device using MR precoding (also known as conjugate beamforming). Two approaches are currently being taken to approximately realize a continuous microwave aperture: 1) The first approach uses a tightly coupled array of discrete, active antennas [59]. This can be implemented using a dense array of conventional fraction-of-λ-spaced antennas, connected to RF chains, but the result will be costly and bulky. Alternatively, the RF signals are generated by mixing two optical signals whose frequency difference equals the desired RF carrier frequency [84]. 2) The second approach mixes an RF reference signal with a large number of nearly passive reflecting elements having electronically steerable reflection parameters [62], [85], known as a reconfigurable reflectarray. The specific implementation described in [60] uses a single RF port on the backside of the surface, which is connected to an electronically steerable RF distribution network with radiating elements that emit the wavefront; see Fig. 6(a). Using the holographic terminology, the RF port generates the reference wave and the distribution network is the holographic display. If multiple RF ports are connected to the same distribution network, a superposition of “beams” can be transmitted and received, to enable spatial multiplexing. This is conceptually similar to the hybrid architectures considered in the mmWave literature (see [86] and references therein), but differs in the use of holographic recording for beamforming design instead of the complicated beam-training procedures commonly considered for hybrid beamforming. Another embodiment of reconfigurable reflectarrays is software-controlled metasurfaces, which are contiguous surfaces that are not actively emitting RF signals, but consist of many discrete “meta-atoms” with electronically steerable reflection properties (e.g., shift in phase, polarization, and amplitude) [63], [64]. Hence, when illuminated by an RF signal sent from another location, the metasurface can control the reflections and, for example, beamform the RF signal in a preferable way; see Fig. 6(b). In principle, the metasurface can be reconfigured to behave as an arbitrarily shaped mirror in the given frequency band. Metasurfaces are also known as intelligent reflecting surfaces [65], [87] and there are several variations on these names [88]–[91]. Note that the key difference from the second approach, described in the previous paragraph, is that the RF signal is not generated inside or close