Alexander von Humboldt Professorship Research Project:

Foundations and Architectures for the Next Generation of WirelessNetworks

Progress Report (April 2016 – March 2017)

Giuseppe Caire

Receiving institutions: Technische Universitat BerlinFraunhofer Institut fur Nachrichtentechnik (Heinrich-Hertz-Institut)

1 Research

1.1 Publications

This is a list of publications during the yhird year of the AvH project:

1. Book Chapter: G. Caire, “Massive-MIMO Scheduling Protocols,” in: V. Wong, R. Schober, D. W. K.Ng, L-C Wang (eds.), Key Technologies for 5G Wireless Systems, Cambridge Univ. Press, 2017.

2. Saeid Haghighatshoar; Giuseppe Caire “Massive MIMO Channel Subspace Estimation From Low-Dimensional Projections,” IEEE Transactions on Signal Processing, Year: 2017, Volume: 65, Issue:2 Pages: 303 - 318

3. K. Mahler; W. Keusgen; F. Tufvesson; T. Zemen; G. Caire “Tracking of Wideband Multipath Com-ponents in a Vehicular Communication Scenario,” IEEE Transactions on Vehicular Technology, Year:2016, Volume: PP, Issue: 99, Pages: 1 - 1.

4. Kim Mahler; Wilhelm Keusgen; Fredrik Tufvesson; Thomas Zemen; Giuseppe Caire “Measurement-Based Wideband Analysis of Dynamic Multipath Propagation in Vehicular Communication Scenar-ios,” IEEE Transactions on Vehicular Technology, Year: 2016, Volume: PP, Issue: 99 Pages: 1 -1

5. Junyoung Nam; Giuseppe Caire; Jeongseok Ha “On the Role of Transmit Correlation Diversity inMultiuser MIMO Systems,” IEEE Transactions on Information Theory, Year: 2016, Volume: PP,Issue: 99 Pages: 1 - 1

6. Bobak Nazer; Viveck R. Cadambe; Vasilis Ntranos; Giuseppe Caire “Expanding the Compute-and-Forward Framework: Unequal Powers, Signal Levels, and Multiple Linear Combinations,” IEEETransactions on Information Theory, Year: 2016, Volume: 62, Issue: 9 Pages: 4879 - 4909

7. Kittipong Kittichokechai; Giuseppe Caire “Secret Key-Based Identification and Authentication Witha Privacy Constraint,” IEEE Transactions on Information Theory, Year: 2016, Volume: 62, Issue: 11Pages: 6189 - 6203

1

8. Xinping Yi; Giuseppe Caire “Optimality of Treating Interference as Noise: A Combinatorial Perspec-tive,” IEEE Transactions on Information Theory, Year: 2016, Volume: 62, Issue: 8 Pages: 4654 -4673

9. Georgios Paschos; Ejder Bastug; Ingmar Land; Giuseppe Caire; Merouane Debbah “Wireless caching:technical misconceptions and business barriers,” IEEE Communications Magazine, Year: 2016, Vol-ume: 54, Issue: 8 Pages: 16 - 22

10. Vishnu V. Ratnam; Andreas F. Molisch; Giuseppe Caire, “Capacity Analysis of Interlaced Clusteringin a Distributed Transmission System With/Without CSIT,” IEEE Transactions on Wireless Commu-nications Year: 2016, Volume: 15, Issue: 4, Pages: 2629 - 2641.

11. M. Ji; G. Caire; A. F. Molisch “Wireless Device-to-Device Caching Networks: Basic Principles andSystem Performance,” IEEE Journal on Selected Areas in Communications Year: 2016, Volume: 34,Issue: 1, Pages: 176 - 189.

12. M. Ji; G. Caire; A. F. Molisch “Fundamental Limits of Caching in Wireless D2D Networks,” IEEETransactions on Information Theory Year: 2016, Volume: 62, Issue: 2, Pages: 849 - 869.

13. D. Bethanabhotla; G. Caire; M. Neely “WiFlix: Adaptive Video Streaming in Massive MU-MIMOWireless Networks,” IEEE Transactions on Wireless Communications Year: 2016, Volume: PP, Issue:99, Pages: 1 - 1.

14. M. Dai; B. Clerckx; D. Gesbert; G. Caire “A Rate Splitting Strategy for Massive MIMO with Im-perfect CSIT,” IEEE Transactions on Wireless Communications Year: 2016, Volume: PP, Issue: 99,Pages: 1 - 1.

15. Quality-Aware Streaming and Scheduling for Device-to-Device Video Delivery Joongheon Kim; GiuseppeCaire; Andreas F. Molisch IEEE/ACM Transactions on Networking Year: 2016, Volume: 24, Issue: 4Pages: 2319 - 2331

16. D. Bethanabhotla; O. Y. Bursalioglu; H. C. Papadopoulos; G. Caire “Optimal User-Cell Associa-tion for Massive MIMO Wireless Networks,” IEEE Transactions on Wireless Communications, Year:2016, Volume: 15, Issue: 3, Pages: 1835 - 1850.

17. K. Mahler; W. Keusgen; F. Tufvesson; T. Zemen; G. Caire “Tracking of Wideband Multipath Com-ponents in a Vehicular Communication Scenario,” IEEE Transactions on Vehicular Technology, Year:2016, Volume: PP, Issue: 99, Pages: 1 - 1.

18. Secure Massive MIMO Transmission With an Active Eavesdropper Yongpeng Wu; Robert Schober;Derrick Wing Kwan Ng; Chengshan Xiao; Giuseppe Caire IEEE Transactions on Information TheoryYear: 2016, Volume: 62, Issue: 7 Pages: 3880 - 3900

19. Kittipong Kittichokechai; Rafael F. Schaefer; Giuseppe Caire “Secret key generation through a relay,”2016 IEEE Information Theory Workshop (ITW) Year: 2016 Pages: 196 - 200,

20. Michal Kaliszan; Giuseppe Caire; Slawomir Stanczak, “On the throughput rate of wireless multipointmulticasting,” 2016 IEEE International Symposium on Information Theory (ISIT) Year: 2016 Pages:2998 - 3002

21. Ozgun Y. Bursalioglu; Chenwei Wang; Haralabos Papadopoulos; Giuseppe Caire “RRH based mas-sive MIMO with on the Fly pilot contamination control,” 2016 IEEE International Conference onCommunications (ICC) Year: 2016 Pages: 1 - 7

2

22. Kim Mahler; Wilhelm Keusgen; Fredrik Tufvesson; Thomas Zemen; Giuseppe Caire “PropagationChannel in a Rural Overtaking Scenario with Large Obstructing Vehicles,” 2016 IEEE 83rd VehicularTechnology Conference (VTC Spring) Year: 2016 Pages: 1 - 5

23. Xinping Yi; Giuseppe Caire “Topological interference management with decoded message passing,”2016 IEEE International Symposium on Information Theory (ISIT) Year: 2016 Pages: 550 - 55

24. Fabio D’Andreagiovanni; Giuseppe Caire “An unconventional clustering problem: User Service Pro-file Optimization,” 2016 IEEE International Symposium on Information Theory (ISIT) Year: 2016Pages: 855 - 859

25. Nikhil Karamchandani; Suhas Diggavi; Giuseppe Caire; Shlomo Shamai “Rate and delay for codedcaching with carrier aggregation,” 2016 IEEE International Symposium on Information Theory (ISIT)Year: 2016 Pages: 2724 - 2728

26. Xinping Yi; Giuseppe Caire “Topological coded caching,” 2016 IEEE International Symposium onInformation Theory (ISIT) Year: 2016 Pages: 2039 - 2043

27. Saeid Haghighatshoar; Peter Jung; Giuseppe Caire “Capacity and degree-of-freedom of OFDM chan-nels with amplitude constraint,” 2016 IEEE International Symposium on Information Theory (ISIT)Year: 2016 Pages: 900 - 904

28. Kittipong Kittichokechai; Giuseppe Caire “Privacy-constrained remote source coding,” 2016 IEEEInternational Symposium on Information Theory (ISIT) Year: 2016 Pages: 1078 - 1082

29. Antonio Forenza; Stephen Perlman; Fadi Saibi; Mario Di Dio; Roger van der Laan; Giuseppe Caire“Achieving large multiplexing gain in distributed antenna systems via cooperation with pCell tech-nology,” 2015 49th Asilomar Conference on Signals, Systems and Computers Year: 2015 Pages: 286- 293

1.2 Information theoretic study of caching wireless networks

Data traffic generated by wireless and mobile devices is predicted to increase by something between oneand two orders of magnitude [1] in the next five years, mainly due to wireless video streaming. Traditionalmethods for increasing the area spectral efficiency, such as using more spectrum and/or deploying morebase stations, are either insufficient to provide the necessary wireless throughput increase, or are too expen-sive. Thus, exploring alternative strategies that leverage different and cheaper network resources is of greatpractical and theoretical interest.

The bulk of wireless video traffic is due to asynchronous video on demand, where users request videofiles from some library (e.g., iTunes, Netflix, Hulu or Amazon Prime) at arbitrary times. This type of trafficdiffers significantly from live streaming. The latter is essentially a lossy multicasting problem, for whichthe broadcast nature of the wireless channel can be naturally exploited (see for example [2, 3, 4, 5, 6]). Thetheoretical foundation of schemes for live streaming relies on well-known information theoretic settings forone-to-many transmission of a common message with possible refinement information, such as successiverefinement [7, 8, 9] or multiple description coding [10, 11, 12].

In contrast, the asynchronous nature of video on demand prevents from taking advantage of multicasting,despite the significant overlap of the requests (people wish to watch a few very popular files). Hence,even though users keep requesting the same few popular files, the asynchronism of their requests is largewith respect to the duration of the video itself, such that the probability that a single transmission fromthe base station is useful for more than one user is essentially zero. Due to this reason, current practicalimplementation of video on demand over wireless networks is handled at the application layer, requiring a

3

dedicated data connection (typically TCP/IP) between each client (user) and the server (base station), foreach streaming user, as if users were requesting independent information.

One of the most promising approaches to take advantage of the inherent asynchronous content reuseis caching, widely used in content distribution networks over the (wired) Internet [13]. In [14, 15], theidea of deploying dedicated “helper nodes” with large caches, that can be refreshed via wireless at thecellular network off-peak time, was proposed as a cost-effective alternative to providing large capacitywired backhaul to a network of densely deployed small cells. An even more radical view considers cachingdirectly at the wireless users, exploiting the fact that modern devices have tens and even hundreds of GBytesof largely under-utilized storage space, which represents an enormous, cheap and yet untapped networkresource.

Recently, a coded multicasting scheme exploiting caching at the user nodes was proposed in [16]. In thisscheme, the files in the library are divided in blocks (packets) and users cache carefully designed subsets ofsuch packets. Then, for given set of user demands, the base station sends to all users (multicasting) a com-mon codeword formed by a sequence of packets obtained as linear combinations of the original file packets.As noticed in [16], coded multicasting can handle any form of asynchronism by suitable sub-packetization.Hence, the scheme is able to create multicasting opportunities through coding, exploiting the overlap ofdemands while eliminating the asynchronism problem. For the case of arbitrary (adversarial) demands, thecoded multicasting scheme of [16] is shown to perform within a small gap, independent of the number ofusers, of the cache size and of the library size, from the cut-set bound of the underlying compound channel.1

However, the scheme has some significant drawbacks that makes it not easy to be implemented in practice:1) the construction of the caches is combinatorial and the sub-packetization explodes exponentially with thelibrary size and number of users; 2) changing even a single file in the library requires a significant recon-figuration of the user caches, making the cache update difficult. In [17], similar near-optimal performanceof coded caching is shown to be achieved also through a random caching scheme, where each user cachesa random selection of bits from each file in the library. In this case, though, the combinatorial complexityof the coded caching scheme is transferred from the caching phase to the (coded) delivery phase, where theconstruction of the multicast codeword requires solving multiple clique cover problems with fixed cliquesize (known to be NP-complete), for which a greedy algorithm is shown to be efficient.

Our contributions in 2016-2017: We have started an in-depth investigation of advanced PHY layertechniques (interference management, channel coding, MIMO beamforming, compress and forward relay-ing) in the presence of caching. In other words, cached information not only allows for the conventionalcaching gains (basically, making the content closer to the destination), but also enabled new interferencemanagement schemes that allow to improve also the underlying physical layer (PHY).

Main results:

• We have further extended our work on D2D network with caching, including new PHY techniquessuch as hierarchical cooperation and quantize-and-forward relaying. This work is preliminary, andhas been submitted to the Transactions on Information Theory and to ISIT 2017 (it will be presentedin ISIT 2017, Aachen, in June 2017).

• We studied new coding techniques to take advantage simultaneously of the beamforming and multi-plexing gain of MIMO, and the caching gain. This work is a collaboration with the group of Prof.Babak Khalaj, in Sharif University, and will be presented in the ISIT 2017 conference in Aachen, inJune 2017.

1The compound nature of this model is due to the fact that the scheme handles adversarial demands.

4

1.3 New architectures for super-dense cell-free massive MIMO

With the proliferation of mobile devices and services, industry predicts that the wireless data traffic is goingto increase by two to three orders of magnitude within a decade [18]. Although the definition of the nextgeneration of systems and standards is at its initial phase, it is widely agreed that the next generation ofwireless networks, generally referred to as “5G”, will involve a combination of multiuser MIMO technol-ogy, cell densification, and heterogeneous architectures based on nested tiers of smaller and smaller cellsoperating at higher and higher frequencies, in order to target traffic hot-spots [19]. These trends have mo-tivated the recent surge of research on massive and dense deployment of base station antennas, both in theform of Massive MIMO schemes, with hundreds of antennas at each cell site [20, 21, 22], and in the form ofmulti-tier networks of densely deployed small-cells [23, 24].

Massive MIMO promises dramatic increases in spectral efficiency by transmitting independent datastreams simultaneously to multiple users sharing the same transmission resource (time-frequency slot). Themassive MIMO regime [20, 21, 22] distinguishes itself from classical multiuser MIMO [25, 26] by the factthat the number of served users is significantly less than the (very large) number of base station antennas.Operating in Time-Division Duplexing (TDD) mode, massive MIMO can provide very large spectral effi-ciencies, simple per-cell processing, and very attractive power efficiency due to the large array gain [20].Thanks to the higher and higher carrier frequencies [27], it is possible to implement massive MIMO evenin relatively small base stations within a reasonable form factor. Hence, it is envisaged that massive MIMOwill not just be applied to large tower-mounted base stations, but also used in conjunction with small cells[28].

The heterogeneous wireless network framework mentioned above may include some of the followingfeatures: 1) base stations that may differ significantly by transmit power, number of antennas, and multiplex-ing gain (e.g., see [29] and references therein); 2) non-homogeneous user spatial distribution, characterizedby high-density hot-spots separated by less dense regions [30]; 3) Due to the large beamforming gain ofmassive MIMO, a user may be in good SINR conditions with respect to several base stations. As a conse-quence, the rationale that has driven for decades the conventional cellular system design and optimization,based on symmetric lattice-deployed cells (see for example [20, 21, 22]) and/or (roughly) uniform numberof users per cell (e.g., see [31] and references therein), must be abandoned in favor of more efficient schemesthat include user-cell association into the optimization problem.

Our contributions in 2016-2017: Dense large-scale antenna deployments are one of the most promisingtechnologies for delivering very large throughputs per unit area in the downlink (DL) of cellular networks.We consider such a dense deployment involving a distributed system formed by multi-antenna remote radiohead (RRH) units connected to the same fronthaul serving a geographical area. Knowledge of the DLchannel between each active user and its nearby RRH antennas is most efficiently obtained at the RRHs viareciprocity based training, that is, by estimating a user’s channel using uplink (UL) pilots transmitted by theuser, and exploiting the UL/DL channel reciprocity.

We consider aggressive pilot reuse across an RRH system, whereby a single pilot dimension is simul-taneously assigned to multiple active users. We introduce a novel coded pilot approach, which allows eachRRH unit to detect pilot collisions, i.e., when more than a single user in its proximity uses the same pilotdimensions. Thanks to the proposed coded pilot approach, pilot contamination can be substantially avoided.As shown, such strategy can yield densification benefits in the form of increased multiplexing gain per ULpilot dimension with respect to conventional reuse schemes and some recent approaches assigning pseudo-random pilot vectors to the active users.

Main results:

• New coded pilot schemes allowing “on-the-fly” pilot contamination control. These schemes allow

5

users to connect instantaneously to all the RRHs that receive their UL pilots without contamination.

• We study the geometric problem of unique coverage in Boolean models, to determine the optimaldensity of active users for given density of RRHs.

• We have extended the approach to directional beamforming, suitable for random-access in mm-wavenetworks.

1.4 Optimization and implementation of massive MIMO schemes based on innovative hybrid digitalanalog signal processing

Massive MIMO (multiple-input multiple-output) systems are equipped with a large number (dozens or hun-dreds) of antenna elements at the base station (BS) [20, 32]. They are intended to be employed in a multi-userMIMO (MU-MIMO) setting, such that the number of BS antenna elements is much larger than the num-ber of users. Such an arrangement leads not only to very high spectral efficiency, but also to an importantsimplification of the signal processing: in the idealized regime of independent and isotropically distributedchannel vectors, in the limit of an infinite number of BS antennas, single-user beamforming, specificallyconjugate beamforming (i.e., maximum ratio combining in the receive mode, and maximum ratio trans-mission for the transmit mode) eliminates inter-user interference. Furthermore, the transmit power can bedrastically reduced, leading to less interference and a lower energy consumption of the BS. For all thesereasons, massive MIMO has received tremendous attention in the last years [33, 34, 35, 36, 37].

Massive MIMO is especially promising for systems operating at millimeter (mm-) Wave frequencies.Due to the short wavelength, very large arrays can be created with a reasonable form factor - a 100-elementlinear array is only about 50 cm long at a carrier frequency of 30 GHz. In light of the extremely largebandwidths that are available for commercial use (up to 7 GHz bandwidth in the 60 GHz band, and around1 GHz at 28 and 38 GHz carrier frequency), massive MIMO systems in the mm-Wave range are ideallysuited for high-capacity transmission and thus anticipated to form an important part of 5G systems. Whilethe first commercial mm-Wave products are intended for in-home, short-range communications (e.g., fortransmission of uncompressed video) [38], the potential of mm-Waves for cellular outdoor has recentlybeen investigated [39, 40, 41]. Experiments have shown a coverage range of more than 200 m even in nonline of sight (NLOS) situations [41]. Such long-range transmissions require high-gain adaptive antennas -something that massive MIMO implicitly provides.

For the downlink, massive MIMO systems at mm-Wave (or, for that matter, any other) frequencies re-quire channel state information at the transmitter (CSIT), for conjugate beamforming as well as for other,more advanced, forms of MU-MIMO precoding (see [42] and references therein). In most existing papers,it has been assumed that this CSIT can be obtained from the uplink sounding signals, based on the principleof channel reciprocity [20]. However, reciprocity only holds (approximately) in Time Division Duplex-ing (TDD) systems, where the duplexing time is much shorter than the coherence time of the channel. InFrequency Division Duplexing (FDD) systems, which are widely used in cellular communications, the spac-ing between uplink and downlink frequency is - for all practical systems - much larger than the coherencebandwidth of the channel [43]. Consequently, CSIT has to be provided through feedback - i.e., each usermeasures its channel vector in the downlink, and sends it to the BS in (quantized) form. Due to the largenumber of BS antenna elements, the overhead for this feedback can become overwhelming, and methodshave to be devised for reducing this load.2

Joint Spatial Division and Multiplexing (JSDM) is a recent technique proposed in [44] to achievemassive-MIMO like gains for FDD systems (or, more generally, for systems that do not make explicit use of

2TDD might also require feedback because accurate TDD calibration is difficult to achieve in practical hardware implemen-tations. This is the reason why the only existing commercial standard that considers MU-MIMO downlink, IEEE 802.11ac, alsoprescribes explicit downlink training and quantized CSIT feedback, even though it uses TDD.

6

channel reciprocity), with the added advantage of a reduced requirement for CSIT3. The idea is to partitionthe user space into groups of users with approximately similar covariances,4 and split the beamforming intotwo stages: a first stage consisting of a pre-beamformer that depends only on the second order statistics, i.e.,the covariances of the user channels, and a second stage comprising a standard MU-MIMO precoder forspatial multiplexing on the effective channel obtained after pre-beamforming. The instantaneous CSIT ofsuch an effective channel is easier to acquire thanks to the considerable dimensionality reduction producedby the pre-beamforming stage. Also, JSDM lends itself to a hybrid beamforming implementation, wherepre-beamforming (which changes slowly in time) may be implemented in the analog RF domain, while theMU-MIMO precoding stage is implemented by standard baseband processing. This approach allows theuse of a very large number of antennas with a limited number of baseband-to-RF chains; the latter dependson the number of independent data streams that we wish to send simultaneously to the users. A majorchallenge for massive MIMO in the mm-Wave region is the fact that the Doppler shift scales linearly withfrequency, and thus the coherence time is an order of magnitude lower than that of comparable microwavesystems. Thus, massive MIMO systems at mm-Wave frequencies need to be restricted to low-mobility sce-narios. For comparable speeds of motion, for example, at pedestrian speeds (1 m/s), coherence times are ofthe order of a few ms at mm-Wave frequencies. Since (outdoor) coherence bandwidths of mm-Wave chan-nels are similar to those of microwave channels [39, 46], the overall challenges of CSI feedback overheadare then comparable to those of higher-mobility (vehicular) microwave massive-MIMO systems. For exam-ple, a 30 GHz channel for a user moving at 1 m/s has the same coherence time and bandwidth of a 3 GHzchannel for a user moving at 10 m/s. In this work, we explicitly assume the availability of perfect channelstate information for simplicity (wherever required). In reality, devoting a certain amount of resource to thetraining phase would discount the achievable throughput by a certain factor [44].

The performance of JSDM depends on the type of channel statistics. Previous analysis was based on theone-cluster (local scattering) model, which means that the BS “sees” the incoming multi-path components(MPCs) under a very constrained angular range. This allows for an easy division of the users into sets, whoseassociated MPCs are disjoint in the angular domain, and can thus be separated by the pre-beamformers.However, this model does not represent many important scenarios. For example, in urban environments,high-rise buildings or street canyons can act as important “common clusters” that create spatially correlatedMPCs for many users [47], [48], [49]. Another important effect, which becomes particularly relevant atmm-Wave frequencies, is channel sparsity - in other words, the number of significant MPCs is much lowerthan that for a microwave system operating in a similar environment. The low number of MPCs enablesa further reduction of the CSIT that has to be fed back, and enables a new “degenerate” variant of JSDM,proposed in this paper and referred to as Covariance-based JSDM, that depends on the channel covarianceinformation only. In fact, it is well known that, as long as the scattering geometry relative to a given userremains unchanged, the fading channel statistics are wide-sense stationary (WSS). In particular, this meansthat the channel covariance matrix is time-invariant. In a typical scattering scenario, even if a user changesits position by several meters, the channel second order statistics remain unchanged [50, Chapter 4]. Hence,for a user moving at walking speed (1 m/s), the channel fading process is “locally” WSS over a time horizonof several seconds, spanning a very large number of symbol time slots (for example, a 20 MHz OFDMchannel has symbol duration of 4 µs, corresponding to 106 symbols over an interval of 4s, correspondingto a user position displacement of 4m). We conclude that it is effectively possible to learn very accuratelythe channel covariance matrix at the transmitter side, even without requiring very fast CSIT feedback. Thismakes our scheme particularly interesting for mm-Waves. The main goal of our work is to apply the JSDMapproach to realistic propagation channels inspired, inter alia, by the recent experimental observations ofmm-Wave channels in an urban outdoor environment [41].

3 An approach that exploits the same directional structure of the channel covariance matrix used by JSDM, in order to eliminatepilot contamination in a multi-cell massive MIMO setting, was proposed concurrently and independently in [45].

4Usually caused by the fact that the multi-path components of such users have similar angles at the BS

7

Our contributions in 2016-2017: We have demonstrated for the first time our new SDR platform formassive MIMO. The demonstration took place in the ITG Workshop on Smart Antenna, in Berlin, March2016.

Furthermore, we have started the systematic investigation of beam acquisition techniques in mm-wavechannels with hybrid digital-analog (HDA) beamforming at both the base station and the user terminals side.

Main results:

• We have developed a software defined radio (SDR) massive MIMO system which follows a central-ized signal processing (CSP) approach, the CSP field programmable gate array (FPGA) device needsto handle huge amounts of digital I/O data sample streams. This problem can be naturally solved byusing the FPGAs integrated multi gigabit wireline transceiver (MGT) circuits, where each MGT offersdigital serial data bit streams on a single (differential) transmission line at data rates nowadays rangingup to several tens of Gbit/s. We have solved the MGT link design for such a system. Furthermorea custom MGT protocol is developed which is especially suited for this architecture. All the criticalparts in the system are experimentally evaluated with real world hardware and measurement resultsare provided. Finally it will be shown that in principle such a CSP architecture can be realized withtoday technology in terms of interconnect requirements for M-MIMO base stations with thousands ofantennas. A picture of our demonstrated testbed is shown below:

future base station deployment large-scale SDR platforms.The first M-MIMO hardware demonstrator was published

in [3]. This system is only capable of maximum ratio trans-mission (MRT) real-time precoding, since it is not capable ofdoing centralized signal processing. The advantage of MRT isthat the matrix inverse calculation degenerates to a conjugate-transpose operation and thus is very easy to implement ina distributed processing fashion. The advantage of such asystem is the easy scalability to more and more antennas [3].Nevertheless our paper will show that in principle with todayFPGA and MGT technology central signal processing (CSP)systems (featuring only a single signal processing FPGA) withthousands of antennas are possible to implement from aninterconnect technology point of view. In addition MGT chiparea and power consumption is decreasing faster [2] than thetypical digital scaling trend (which roughly follows Moore’slaw). This means that the number of MGTs which can berealized on a single chip will also increase considerably forthe coming years. So if the future requires M-MIMO systemswith even more than thousands of RF transceivers, this willalso be possible with the future FPGA and MGT technology.

The authors of [4] present a M-MIMO demonstrator whichis capable of real-time zero-forcing precoding. The digitalsignal processing is distributed across 50 FPGAs and ZFprecoding is distributed across the FPGAs for a certain numberof OFDM sub carriers. The system employs off the shelfhardware which uses PCI Express for the large amount ofFPGA to FPGA communications. This distributed processingmakes it a very complex system in terms of engineering effortto implement real time signal processing algorithms.

Our paper will show how to put control over the MGTs inthe designer’s hand and thus will enable an efficient reducedoverhead custom MGT protocol implementation which isespecially suited to the CSP style of SDR system. This is avital step towards a CSP M-MIMO SDR system. The CSPapproach using the custom MGT implementation will alsodecrease the number of required FPGAs considerably and thuswill decrease cost, size and power consumption of the system.

II. MGTS IN THE CONTEXT OF A CSP M-MIMO SDRSYSTEM

In an CSP M-MIMO SDR system, MGTs can be used totransport raw sample streams from the RF TRx’s ADCs andDACs to the CSP FPGA (see figure 1). Since most ADCs

Host PC

Bac

kpla

ne

PCI-E

N MGTlinks

MGT links1 . . . K

Serializer/Deserializer

& switch FPGA

RF TRx

RF TRx

...

RF TRx

...

MGT linksK + 1 . . . 2K

CSPFPGA

...

Serializer/Deserializer

& switch FPGA

RF TRx

RF TRx

...

RF TRx

Fig. 1. Central signal processing system architecture

Fig. 2. Experimental hardware setup of the interface technology evaluationplatform

and DACs are using conventional parallel CMOS or LVDSinterfaces, SERDES (serialization and de-serialization of theMGT streams) and switch FPGAs or ASICs are required insuch a system. These actually only function as sample streamexpanders, i.e. these high pin-count chips connect as many RFfront-ends via parallel interfaces as possible and they serializeADC sample streams and de-serialize DAC sample streams toone or more MGT link.

In the following we will investigate the implementationdetails and technical problems associated with the use ofMGTs in such an CSP M-MIMO SDR base station architecturefrom an embedded systems and hardware systems designpoint of view. Critical parts in the system are experimentallyevaluated with real-world hardware (see figure 2). The usedMGT setup will be described in detail, and implementationdetails about pre-emphasis and equalizer parametrization andcorresponding measured eye-diagrams will be provided. Theevaluated MGT communications channel, consisting of differ-ent forms of PCB transmission lines, special connectors andtwinaxial cables will be discussed in detail. Furthermore acustom MGT protocol especially suited for an CSP M-MIMOSDR will be presented which addresses issues like initial linksynchronization, skew compensation, word synchronizationand latency behaviour of the link. Just like MGT FPGA setups,frameworks and problems are discussed in the experimentalhigh energy physics community (see [7], [8], [6], [5]), thispaper will discuss the same in the context of large scale SDRsystems.

III. MGT CHANNEL

The main factors which limit maximum achievable datarates on impedance controlled wired MGT channels is the skineffect and dielectric loss. Since both result in increased losswith increasing frequency, the channel has generally a low-pass characteristic, with roll-off increasing with increasing tan� of the used transmission line medium and the total lengthof the link (see [10] for some example measurements).

• We have developed a model based on virtual beam decomposition, for which the beam acquisitionconsists of exploring the virtual beam directions at the user and base station side, until a pair of highlycoupled directions is found. The figure below shows a sketch of a channel with several multipathcomponents (scattering clusters) and its decomposition in virtual beams.

8

1

Beam Alignment in mm-Wave Massive MIMOSystems

Saeid Haghighatshoar, Member, IEEE, Giuseppe Caire, Fellow, IEEE

Abstract—Millimeter-wave (mm-Wave) cellular systems area promising option for a very high data rate communicationbecause of the large bandwidth available at mm-Wave frequen-cies. To compensate the large path-loss exponent in the mm-Wave range of the spectrum, directional beamforming witha large antenna gain at both the user side and base-stationside is necessary; this boost the signal-to-noise ratio to a levelsuitable for reliable communication. However, designing sucha directional beam-forming requires a reliable estimate of thechannel state, e.g., dominant direction-of-arrival and direction-of-departure at the base-station and user; this is especiallychallenging due to the very low SNR, and is pretty overlookedin the previous studies of mm-Wave channels.

In this paper, we propose an adaptive sampling schemeto speed up this estimation procedure. In our scheme, thebeamforming vector for taking every new sample, either atthe base-station or user side, is adaptively selected basedon all the previous beamforming vectors and the resultingoutput observations. We formulate the problem in the well-known setting of optimal design of experiments in statistics. Wealso propose simple suboptimal strategies and evaluate theirefficiency empirically via numerical simulations.

I. INTRODUCTION

Two key capacity-increasing techniques proposed for fu-ture massive MIMO cellular networks including 5G will benetwork densification and the use of higher frequency bandssuch as millimeter wave (mm-Wave) [1, 2]. The main chal-lenges to using mm-Wave frequencies are their high near-field path-loss (due to small effective antenna aperture) andvery poor penetration into buildings due to the large path-loss exponent in this range of the spectrum [3]. As a result,a very large array with many number of antennas, providinga very large beamforming gain, both at the user side and atthe base-station (BS) side is necessary in order to boost thesignal-to-noise ratio (SNR) to a level that is sufficient forreliable communication. A natural consequence of havinghigh path-loss in the channel and applying highly directionaltransmissions at both sides is that the inter-user interferencedue to the users inside the same cell is greatly reduced,such that, in most cases, mm-Wave channels turn out benoise- rather than interference-limited [4, 5]; this providesa big advantage compared with the sub-6 GHz networks.The drawback, as already noticed by many researchers, isthat the characteristics of the mm-Wave scattering channelmakes the channel estimation and acquisition much morechallenging that the sub-6 GHz counterpart.

We focus, here, on a massive MIMO scheme [6–9], inwhich uplink (UL) and downlink (DL) are organized in TimeDivision Duplexing (TDD), and the BS transmit/receivehardware is designed or calibrated in order to preserve UL-DL reciprocity [10, 11]. In such a massive MIMO system in

The authors are with the Communications and Information TheoryGroup, Technische Universitat Berlin ({saeid.haghighatshoar, caire}@tu-berlin.de).

sub-6 GHz, the BS estimates the channel vectors of the usersfrom UL training signals sent by the users on orthogonaldimensions. The resulting channel state is used to designbeamforming matrices for serving the users. Such a trainingscheme fails for mm-Wave channels due to the very smallSNR before suitable beamforming, which is impossible todo without knowing the channel state. This clearly indicatesthat estimating the channel state in mm-Wave has a quitedifferent nature than that in sub-6 GHz. This also shows thatdesigning fast and robust acquisition algorithms would playa crucial role for a reliable communication between the BSand the users [3, 12].

This problem has already been addressed by many re-searchers mainly by extending the sub-6 GHz channel esti-mation techniques such as sparse sampling and compressedsensing [13, 14]. However, none of these methods per-form quite well for the mm-Wave channels due to thevery low SNR of the channel. For example, in scenarioshown in Fig. 1 for a BS with M = 32 and a userwith N = 32 antennas, and assuming a SNR ⇠10 dB forreliable communication, the SNR at each antenna, thus, thewhole SNR before bemforming, can be as low as �20 dB,which is multiple orders of magnitude lower than the SNRrequired for traditional channel estimation algorithms towork robustly. However, the hope is that the SNR can beboosted by a factor MN after suitable beamforming to meetthe 10 dB working SNR.

✓i

•0

•d

•2d

•3d

•4d

•(M � 1)d

Base-station

�i

•0

•d

•2d

•3d

•4d

•(N � 1)d

User

...

Very sparse scattering channel

Fig. 1: mm-Wave scattering channel with few scatterers.

Recently, new algorithms based on adaptive search withmulti-stage codebooks have been proposed in which theuser and the BS jointly design their beamforming vectors(see [15–18] and the refs. therein). In short, the proposedmulti-stage (multi-resolution) beamforming algorithms canbe seen as different variants of binary search or bisectionalgorithm over all possible angle of arrivals (AoA) of thechannel. The search typically starts with two wide beamseach scanning half the range of possible AoAs. This pro-vides a rough estimate of the location of channel’s dominant

2

consideration.

maxB,[Wi],[Si]

E[ max[Fi],V

KX

i=1

(1� ⌧tr(b) + ⌧fb,i(Si)

T)Ri(Fi,V, Hi,Si)]

s.t. E[tr(BVV†B†)] Pt

(1)

where the expectation E(·) averages out the instantaneous channel realizations, Si = [Si1, ...,Sib] �i,Sij is the aUE ⇥ aUE receive beam selection matrix for transmit beam j which determines howmany and which elements of the e�ective channel should be fed back.. The selection matrix isdiagonal with entry of zeros and ones. Therefore, the feedback CSI can be expressed as non-zeroentries of Hi(:, j) = SijHi(:, j), where (:, j) denotes the j-th column of a matrix, T is the coher-ence block size, ⌧tr is the number of channel uses for downlink training depending on the numberof BS RF chains b, ⌧fb,i �i is the number of channel uses for uplink feedback of UEi dependingon the number of non-zero entries of Si.

The partially feedback mechanism can be viewed as a generalized “virtual sectorization”,where UE-specific beam selection is performed and only the channels of selected beams are fedback. For example, there are 2 UE RF chains and 16 BS RF chains.The e�ective channel H,includes 2 · 16 = 32 variables. However, there might be much fewer dominant components thatneed to inform BS since B includes orthogonal beamforming vectors serving other UEs. AsFig. 1 shows, either receive beam w1 or w2 can only acquire good array gain from 2 transmitbeams, i.e. b1,b2 and b3,b4 respectively, where W = [w1,w2], B = [b1, ...,b16]. Therefore,only 2 · 2 = 4 variables need to be fed back to the BS.

Figure 1: mm-wave channel model with multiple highly directional clusters

2.2 Development of Reduced Complexity Algorithms

As we described in sec.1, to solve problem (1) is very challenging in general. Meanwhile, it is pre-dictable that algorithms approximating the optimal solution will be too complex and impracticalfor implementation. Therefore, rather than investigating the full joint design over optimalityspace, we will investigate the development of reduced complexity algorithms. The essential ideais to decouple the interacting variables and separately design them.

3

Fig. 1. Coupling between the virtual beam space at the BS and the virtual beam space at the UE.

Chapter 5. Channel Characterization 59

(a) A LOS scenario - user location 2 (b) An NLOS scenario - user location 5

Figure 5.5: Fading and correlation patterns on the 128-element cylin-drical array, in (a) a LOS scenario, and (b) an NLOS scenario. Theantennas are re-indexed as follows: the first 64 are vertically-polarizedantennas and the last 64 are horizontally-polarized ones, antennas facingthe same direction are ordered successively from bottom to top.

been evaluated and confirmed based on measured channels, as reported in pa-pers I-III. Not only for users being located far apart, but closely-spaced userscan also be separated even with compact arrays, as shown in Paper III wherecrowded scenarios are investigated. Propagation e�ects among closely-spacedusers need to be investigated, as they have not been considered in conventionalMU-MIMO.

For users being far apart, interacting with di�erent scatterers in the prop-agation environment, it is relatively easy to spatially separate their signals.Closely-spaced users are most likely interacting with the same scatterers, butin di�erent ways [66]. More precisely, closely-spaced users may interact withdi�erent structural details of a physical scatterer, resulting in MPCs with dif-ferent angles, phases and amplitudes. Based on 3D ray tracing in an urbanenvironment, the study in [92] investigated the duration of MPCs along a path.Figure 5.6 illustrates the propagation e�ect, where a transmitter is moving anda receiver is fixed, after some distance, the interaction point on a scattererchanges and a di�erent MPC “arises”. We call this distance MPC lifetime.Simulation results in [92] show that most MPCs last for less than 1 m, onlya few last for above 10 m, and not much power is present in the long-lastingones. The observation supports that closely-spaced users have di�erent MPCsinteracting with the same scatterer. However, this propagation e�ect needs tobe further investigated using channel measurements.

BS virtual beam index

UE

virt

ual b

eam

inde

x

Fig. 2. Measured strength of the coupling between the Tx and Rx virtual beam spaces.

of RF chains (including amplification, modulation/demodulation, and A/D and D/A conversion) is limited by hardware

complexity and power consumption.

Research question 1: we wish to find efficient ways to estimate the physical channel channel second-order statistics,

described by the matrix RH, from m-dimensional observations at the BS side and n-dimensional observations

Two nodes (e.g., BS and UE) equipped with M and N antennas, and with m � M and n � N“antenna ports” (RF chains), respectively.

• Virtual beamspace channel model:

H =N∑

r=1

M∑

s=1

Hr,survHs

U and V are unitary matrices of dimensions N ×N and M ×M , respectively, corresponding to theleft and right generalized eigenvectors of the channel covariance matrix. The coefficients Hr,s for themoment are assumed to be uncorrelated ∼ CN (0, λr,s) and describe how the s-th right virtual beamdirection couples with the r-th left virtual beam direction.

Any beam pattern at the Tx (right) and Rx (left) side, can be written as

A = VA, B = UB

where A ∈ CM×m and B ∈ CN×n are beam coefficients with the normalizations

tr(AHA) = m, tr(BHB) = n

A channel probing round (from right to left) consists of

←−Y = BH

(√PbsHA+

←−Z)

obtained by sending m orthogonal signature sequences over the m Tx antenna ports. In the reversedirection (from left to right) we have

−→Y = AH

(√PueH

HB+−→Z)

obtained by sending n orthogonal signature sequences over the n Tx antenna ports.

We have devised strategies for bidirectional interactive probing consisting of a sequence of right-leftand left-right probing rounds, where the beamforming matrices A and B can be modified as a functionof the incoming signal. The figures below show qualitatively a “bisection” procedure, where the pairof highly coupled virtual beam directions is “trapped” by successive bisection.

9

3

1

2

3

4

22

21

33

34

43

44

11

34

41

� �a

� �b

� �c

MSBS

MS

BS

……

31

32

42

Fig. 1. Illustration of bipartite iterative beam alignment method. (a)Searching block decomposition between BS and UE. (b) Block labeling

scheme. (d) Block labeling example.

III. TRAINING METHOD DESCRIPTION

We proposal a bipartite iterative beam alignment method to search the strongly coupled BS and UE beams

corresponding to some strong scatters, as illustrated in Fig. 5. Before the reliable data transmission, the location

of the scatter and their strengths are unknown to the BS and UE. During a training stage, they try to identify the

location of a scatter in S according to the received base band signal power Qu at UE side and Qb at BS side. The

training procedures are as follows. The BS divides the current beam directions (which need to be trained) into two sets

with respect to two beam forming vectors a(t1) and a(t2). Also the UE divides the current beam directions (which

need to be trained) into two sets with respect to two beam combining vectors b(t1) and b(t2). Then BS firstly transmits

pilot symbols along the beam directions decided by a(t1), UE receives the signal successively with combining vector

b(t1) and b(t2) while in listening mode. BS continues to transmit pilot symbols using beamforming vector a(t2)

and UE again receives signal successively with combining vector b(t1) and b(t2). Now UE has received 4 different

signals with power {Qu,1, Qu,2, Qu,3, Qu,4} corresponding to the 4 blocks shown in Fig. 5(b). The UE sorts the power

Qu,(1) � Qu,(2) � Qu,(3) � Qu,(4) and selects the maximum power Qu,(1) block with respect to vector b(tl), l 2 {1, 2}as the detected block index (BlockUE) at UE side. Then UE transmits pilot symbols with beamforming vector b(t+l )

along the beam directions decided by b(tl). BS receives signals successively with combining vector a(t1) and a(t2).

Now BS has received 2 different signals with power {Qb,1, Qb,2} corresponding to 2 blocks. The BS sorts the power

Qb,(1) � Qb,(2) and selects the maximum power Qb,(1) block with respect to vector a(tl0), l0 2 {1, 2} as the detected

10

6

10 20 30

5

10

15

20

25

30

10 20 30

5

10

15

20

25

30

10 20 30

5

10

15

20

25

30

10 20 30

5

10

15

20

25

30

10 20 30

5

10

15

20

25

30

Fig. 2. SNR = 20 dB, M = N = 32, scatter number = 4, keeping training until the last iteration

1.5 Applications of compressed sensing to massive MIMO

Consider a multiuser MIMO channel formed by a base-station (BS) with M antennas and K single-antennamobile users in a cellular network. Following the current massive MIMO approach [51, 22, 52, 53], uplink(UL) and downlink (DL) are organized in Time Division Duplexing (TDD), and the BS transmit/receivehardware is designed or calibrated in order to preserve UL-DL reciprocity [54, 55] such that the BS can es-timate the channel vectors of the users from UL training signals sent by the users on orthogonal dimensions.Since there is no multiuser interference on the UL training phase, in this paper we shall focus on the basicchannel estimation problem for a single user.

In massive MIMO systems, the number of antennasM is typically much larger than the number of usersK scheduled to communicate over a given transmission time slot (i.e., the number of spatially multiplexeddata streams). Letting D denote the duration of a time slot (expressed in channel uses), τD channel usesfor some τ ∈ (0, 1), are dedicated to training and the remaining (1 − τ)D channel uses are devoted todata transmission, where it is assumed that D is not larger than the channel coherence block length, i.e., thenumber of channel uses over which the channel is nearly constant [51]. It turns out that for isotropicallydistributed channel vectors with min{M,K} ≥ D/2, it is optimal to devote a fraction τ = 1/2 of the slot

11

to channel estimation while serving only D/2 out of K users in the remaining half [51].5

In many relevant scenarios, the channel vectors are highly correlated since the propagation occursthrough a small set of Angle of Arrivals (AoAs). This correlation can be exploited to improve the systemmultiplexing gain and decrease the training overhead. A particularly effective scheme is the Joint Space Di-vision and Multiplexing (JSDM) approach proposed and analyzed in [28, 56, 57, 58, 59]. JSDM starts fromthe consideration that for a user with a channel vector h ∈ CM the signal covariance matrix S = E[hhH]is typically low-rank6. Moreover, according to the well-known and widely accepted Wide-Sense StationaryUncorrelated Scattering (WSSUS) channel model, S is invariant over time and frequency. In particular,while the small-scale fading has a coherence time between 0.1s and 10 ms for motion speed between 1 m/sto 10 m/s at the carrier frequency of 3 GHz, the time over which the channel vector can be considered WSSis of the order of tens of seconds, i.e., from 2 to 4 orders of magnitude larger. Hence, estimating the signalsubspace of a user is a much easier task than estimating the instantaneous channel vector h on each coher-ence time slot. This is especially important in mm-wave channels (e.g., carrier frequency of the order of30 GHz) since, due to the higher carrier frequency, the Doppler bandwidth of these channels is large andtherefore D is small, i.e., the multiplexing gain of D/2 achieved by estimating the channels by TDD oneach given slot as in [51] is significantly impaired.

When the subspace information for the users can be accurately estimated over a long sequence of timeslots, JSDM partitions the users into G > 1 groups such that users in each group have approximately thesame dominant channel subspace [28, 56, 57]. The overall multiplexing gain is obtained in two stages, asthe concatenation of two linear projections. Namely, groups are separated by zero-forcing beamforming thatuses only the group subspace information. Then, additional multiuser multiplexing gain can be obtained byconventional linear precoding applied independently in each group. In this way, the system multiplexinggain can be boosted by G such that a decrease in D can be compensated by a larger G [60].

Furthermore, JSDM lends itself naturally to a Hybrid Digital Analog (HDA) implementation, where thegroup-separating beamformer can be implemented in the analog (RF) domain, and the multiuser precodinginside each group is implemented in the digital (baseband) domain. The analog beamforming projectionreduces the dimensionality from M to some intermediate dimension m�M . Then, the resulting m inputs(UL) are converted into digital baseband signals, and are further processed in the digital domain. This hasthe additional non-trivial advantage that only m � M RF chains (A/D converters and modulators) areneeded, thus reducing significantly the massive MIMO BS receiver/transmitter front-end complexity andpower consumption.

From what said, it is apparent that a central task at the BS side consists in estimating, for each user, asubspace containing a significant amount of its received signal power. Since in an HDA implementation wedo not have direct access to all the M antennas, but only to m � M analog output observations, we needto estimate this subspace from snapshots of a low-dim projection of the signal.

Our contributions in 2016-2017: We have solved the problem of reconstructing the dominant channelsubspace from low-dimensional projections of its sample covariance matrix. In general, we assume that wecan observe only low-dim sketches of the received signal via m�M linear projections. In the case wherethe projection matrix contains a single non-zero element equal to 1 in each row, we recover the case ofarray subsampling as a special case. In particular, we shall consider a coprime sampling scheme requiringm = O(2

√M). Coprime subsampling was first developed by Vaidyanathan and Pal in [61, 62], where

they showed that for a given spatial span for the array, one obtains approximately the same resolution as auniform linear array by nonuniformly sampling only a few array elements at coprime locations. We proposeseveral algorithms for estimating the signal subspace and cast them as convex optimization problems that can

5When K > D/2, then groups of D/2 users are scheduled over different time slots such that all users achieve a positivethroughput (i.e., rate averaged over a long sequence of scheduling slots).

6This is especially true in the case of a tower-mounted BS and/or in the case of mm-wave channels, as experimentally confirmedby channel measurements (see [57] and references therein).

12

be solved efficiently. We also compare via simulation the performance of our algorithms with other state-of-the-art algorithms in the literature. The relevance of the proposed approach for JSDM is demonstratedvia a representative example, where the dominant subspace of users with different channel correlations areestimated and grouped according to the Grassmanian quantization scheme introduced in [56]. Then, JSDMis applied to the estimated user groups. We compare the achieved sum-rate of our scheme with the idealcase, where the users’ channel covariances are perfectly known, as in [56], and we find that the performancepenalty incurred by our proposed method is negligible, even for very short training lengths.

Main results:

• We have new algorithms based on an Approximated Maximum-Likelihood approach, to reconstructthe channel subspace.

• We have new low-complexity implementation of the Approximated Maximum Likelihood, whichreduces to a version of the modified Lasso with `2,1 norm for block sparsity, that is shown to coincideasymptotically with the solution of the semi-definte program resulting from approximated maximumlikelihood.

• We have extended this approach to wideband channel estimation, including interpolation over a combof subcarrier as in OFDM systems, and introduced a method for pilot decontamination based onseparating useful signal and interference in the expected domain of angle of arrivals and delay.

13

References

[1] “White Paper: Cisco Visual Networking Index: Global Mobile DataTraffic Fore-cast Update, 2010 – 2015,” http://www.scribd.com/doc/63529506/Cisco-White-Paper-c11-520862, Feb. 2010.

[2] F-L. Luo, Mobile Multimedia Broadcasting Standards: Technology and Practice, Springer Verlag,2008.

[3] U. Reimers, “Digital video broadcasting,” Communications Magazine, IEEE, vol. 36, no. 6, pp.104–110, 1998.

[4] U. Ladebusch and C.A. Liss, “Terrestrial dvb (dvb-t): A broadcast technology for stationary portableand mobile use,” Proceedings of the IEEE, vol. 94, no. 1, pp. 183–193, 2006.

[5] O.Y. Bursalioglu, M. Fresia, G. Caire, and H.V. Poor, “Lossy multicasting over binary symmetricbroadcast channels,” Signal Processing, IEEE Transactions on, vol. 59, no. 8, pp. 3915–3929, 2011.

[6] Y. Li, E. Soljanin, and P. Spasojevic, “Three schemes for wireless coded broadcast to heterogeneoususers,” Physical Communication, 2012.

[7] W.H.R. Equitz and T.M. Cover, “Successive refinement of information,” Information Theory, IEEETransactions on, vol. 37, no. 2, pp. 269–275, 1991.

[8] O.Y. Bursalioglu, M. Fresia, G. Caire, and H.V. Poor, “Lossy joint source-channel coding using raptorcodes,” International Journal of Digital Multimedia Broadcasting, vol. 2008, 2008.

[9] M.R. Chari, F. Ling, A. Mantravadi, R. Krishnamoorthi, R. Vijayan, G.K. Walker, and R. C, “Flophysical layer: An overview,” Broadcasting, IEEE Transactions on, vol. 53, no. 1, pp. 145–160, 2007.

[10] V.K. Goyal, “Multiple description coding: Compression meets the network,” Signal Processing Mag-azine, IEEE, vol. 18, no. 5, pp. 74–93, 2001.

[11] Y. Wang, A.R. Reibman, and S. Lin, “Multiple description coding for video delivery,” Proceedings ofthe IEEE, vol. 93, no. 1, pp. 57–70, 2005.

[12] R. Ahlswede, “On multiple descriptions and team guessing,” Information Theory, IEEE Transactionson, vol. 32, no. 4, pp. 543–549, 1986.

[13] Erik Nygren, Ramesh K Sitaraman, and Jennifer Sun, “The akamai network: a platform for high-performance internet applications,” ACM SIGOPS Operating Systems Review, vol. 44, no. 3, pp. 2–19,2010.

[14] N. Golrezaei, A.F. Molisch, and A.G. Dimakis, “Base station assisted device-to-device communica-tions for high-throughput wireless video networks,” IEEE Communications Magazine, in press., 2012.

[15] N. Golrezaei, K. Shanmugam, A. G Dimakis, A. F Molisch, and G. Caire, “Femtocaching: Wirelessvideo content delivery through distributed caching helpers,” CoRR, vol. abs/1109.4179, 2011.

[16] M.A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” arXiv preprint arXiv:1209.5807,2012.

[17] M.A. Maddah-Ali and U. Niesen, “Decentralized caching attains order-optimal memory-rate tradeoff,”arXiv preprint arXiv:1301.5848, 2013.

D-1

[18] “Cisco visual networking index: Global mobile data traffic forecast update, 2013-2018.,” .

[19] Hiroyuki Ishii, Yoshihisa Kishiyama, and Hideaki Takahashi, “A novel architecture for LTE-B: C-plane/U-plane split and phantom cell concept,” in IEEE Globecom Workshops, 2012, pp. 624–630.

[20] T.L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,”Wireless Communications, IEEE Transactions on, vol. 9, no. 11, pp. 3590–3600, 2010.

[21] Jakob Hoydis, Stephan Ten Brink, and Merouane Debbah, “Massive MIMO: How many antennas dowe need?,” in IEEE 49th Annual Allerton Conference on Communication, Control, and Computing,2011, pp. 545–550.

[22] H. Huh, G. Caire, H.C. Papadopoulos, and S.A. Ramprashad, “Achieving massive MIMO spectralefficiency with a not-so-large number of antennas,” IEEE Trans. on Wireless Commun., vol. 11, no. 9,pp. 3226–3239, 2012.

[23] Jakob Hoydis, Mari Kobayashi, and Merouane Debbah, “Green small-cell networks,” IEEE VehicularTechnology Magazine,, vol. 6, no. 1, pp. 37–43, 2011.

[24] V. Chandrasekhar, J. Andrews, and A. Gatherer, “Femtocell networks: a survey,” IEEE Commun.Magazine,, vol. 46, no. 9, pp. 59–67, 2008.

[25] “Quantenna 4× 4 MIMO technology,” .

[26] “Broadcom 6× 6 MIMO press release,” .

[27] T.S. Rappaport, Shu Sun, R. Mayzus, Hang Zhao, Y. Azar, K. Wang, G.N. Wong, J.K. Schulz,M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular: It will work!,”IEEE Access, vol. 1, pp. 335–349, 2013.

[28] Ansuman Adhikary and Giuseppe Caire, “Joint spatial division and multiplexing: Opportunistic beam-forming and user grouping,” arXiv preprint arXiv:1305.7252, 2013.

[29] Amitabha Ghosh, Nitin Mangalvedhe, Rapeepat Ratasuk, Bishwarup Mondal, Mark Cudak, EugeneVisotsky, Timothy A Thomas, Jeffrey G Andrews, Ping Xia, Han Shin Jo, et al., “Heterogeneouscellular networks: From theory to practice,” IEEE Communications Magazine, vol. 50, no. 6, pp.54–64, 2012.

[30] 3GPP technical specification group radio access network, “Further advancements for E-UTRA: LTE-Advanced feasibility studies in RAN WG4,” Tech. Rep., 3GPP TR 36.815, March 2010.

[31] Harpreet S Dhillon, Radha Krishna Ganti, Francois Baccelli, and Jeffrey G Andrews, “Modeling andanalysis of k-tier downlink heterogeneous cellular networks,” IEEE J. on Sel. Areas in Commun., vol.30, no. 3, pp. 550–560, 2012.

[32] F. Rusek, D. Persson, Buon Kiong Lau, E.G. Larsson, T.L. Marzetta, O. Edfors, and F. Tufvesson,“Scaling up mimo: Opportunities and challenges with very large arrays,” Signal Processing Magazine,IEEE, vol. 30, no. 1, pp. 40–60, 2013.

[33] H. Q. Ngo, E. Larsson, and T. Marzetta, “Energy and Spectral Efficiency of Very Large MultiuserMIMO Systems,” IEEE Trans. on Commun., vol. 61, no. 4, pp. 1436–1449, 2013.

[34] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of Cellular Networks: HowMany Antennas Do We Need?,” IEEE Journal on Selected Areas in Communications, vol. 31, no. 2,pp. 160–171, 2013.

D-2

[35] T. L. Marzetta, G. Caire, M. Debbah, I. Chih-Lin, and S. K. Mohammed, ,” .

[36] H. Huh, G. Caire, H. Papadopoulos, and S. Ramprashad, “Achieving “Massive MIMO” Spectral Effi-ciency with a Not-so-Large Number of Antennas,” IEEE Transactions on Wireless Communications, ,no. 9, pp. 3226–3239, September 2012.

[37] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO for next generation wirelesssystems,” IEEE Communications Magazine, vol. 52, no. 2, pp. 186–195, 2014.

[38] Eldad Perahia and Michelle X Gong, “Gigabit wireless lans: an overview of ieee 802.11 ac and 802.11ad,” ACM SIGMOBILE Mobile Computing and Communications Review, vol. 15, no. 3, pp. 23–33,2011.

[39] T.S. Rappaport, Shu Sun, R. Mayzus, Hang Zhao, Y. Azar, K. Wang, G.N. Wong, J.K. Schulz,M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5g cellular: It will work!,”Access, IEEE, vol. 1, no. 1, pp. 335–349, May 2013.

[40] Mathew Samimi, Kevin Wang, Yaniv Azar, George N. Wong, Rimma Mayzus, Hang Zhao, Jocelyn K.Schulz, Shu Sun, Felix Gutierrez Jr., and Theodore S. Rappaport, “28 ghz angle of arrival and angleof departure analysis for outdoor cellular communications using steerable beam antennas in new yorkcity,” in Vehicular Technology Conference Fall (VTC 2013-Fall), 2013 IEEE 74th, 2013.

[41] Y. Azar, G.N. Wong, K. Wang, R. Mayzus, J.K. Schulz, Hang Zhao, F. Gutierrez, D. Hwang, and T.S.Rappaport, “28 ghz propagation measurements for outdoor cellular communications using steerablebeam antennas in new york city,” in Communications (ICC), 2013 IEEE International Conference on,2013, pp. 5143–5147.

[42] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser mimo achievable rates with downlinktraining and channel state feedback,” Information Theory, IEEE Transactions on, vol. 56, no. 6, pp.2845–2866, 2010.

[43] Andreas F Molisch, Wireless communications, Wiley. com, 2010.

[44] A. Adhikary, Junyoung Nam, Jae-Young Ahn, and G. Caire, “Joint spatial division and multiplexing:The large-scale array regime,” Information Theory, IEEE Transactions on, vol. 59, no. 10, pp. 6441–6463, 2013.

[45] Haifan Yin, David Gesbert, Miltiades Filippou, and Yingzhuang Liu, “A coordinated approach tochannel estimation in large-scale multiple-antenna systems,” IEEE JOURNAL ON SELECTED AREASIN COMMUNICATIONS, vol. 31, no. 2, 2013.

[46] T.S. Rappaport, F. Gutierrez, E. Ben-Dor, J.N. Murdock, Y. Qiao, and J.I. Tamir, “BroadbandMillimeter-Wave Propagation Measurements and Models Using Adaptive-Beam Antennas for Out-door Urban Cellular Communications,” IEEE Transactions on Antennas and Propagation, vol. 61, no.4, pp. 1850 – 1859, 2013.

[47] J. Fuhl, A.F. Molisch, and E. Bonek, “Unified channel model for mobile radio systems with smartantennas,” Radar, Sonar and Navigation, IEE Proceedings -, vol. 145, no. 1, pp. 32–41, 1998.

[48] H. Asplund, A.A. Glazunov, A.F. Molisch, K.I. Pedersen, and M. Steinbauer, “The cost 259 directionalchannel model-part ii: Macrocells,” Wireless Communications, IEEE Transactions on, vol. 5, no. 12,pp. 3434–3450, 2006.

D-3

[49] M. Toeltsch, J. Laurila, K. Kalliola, A.F. Molisch, P. Vainikainen, and E. Bonek, “Statistical charac-terization of urban spatial radio channels,” Selected Areas in Communications, IEEE Journal on, vol.20, no. 3, pp. 539–549, 2002.

[50] T. S. Rappaport, Wireless communications: principles and practice, Prentice Hall, 1996.

[51] Thomas L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station anten-nas,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010.

[52] Jakob Hoydis, Stephan Ten Brink, and Merouane Debbah, “Massive mimo in the ul/dl of cellularnetworks: How many antennas do we need?,” IEEE J. on Sel. Areas on Commun. (JSAC), vol. 31, no.2, pp. 160–171, 2013.

[53] Erik Larsson, Ove Edfors, Fredrik Tufvesson, and Thomas Marzetta, “Massive mimo for next genera-tion wireless systems,” Communications Magazine, IEEE, vol. 52, no. 2, pp. 186–195, 2014.

[54] Clayton Shepard, Hang Yu, Narendra Anand, Erran Li, Thomas Marzetta, Richard Yang, and LinZhong, “Argos: Practical many-antenna base stations,” in Proceedings of the 18th Annual InternationalConference on Mobile Computing and Networking. ACM, 2012, pp. 53–64.

[55] Ryan Rogalin, Ozgun Y Bursalioglu, Helene Papadopoulos, Giuseppe Caire, Andreas F Molisch, An-tonios Michaloliakos, Viorel Balan, and Konstantinos Psounis, “Scalable synchronization and reci-procity calibration for distributed multiuser mimo,” IEEE Trans. on Wireless Commun., vol. 13, no. 4,pp. 1815–1831, 2014.

[56] Junyoung Nam, Ansuman Adhikary, Jae-Young Ahn, and Giuseppe Caire, “Joint spatial division andmultiplexing: Opportunistic beamforming, user grouping and simplified downlink scheduling,” IEEEJ. of Sel. Topics in Sig. Proc. (JSTSP), vol. 8, no. 5, pp. 876–890, 2014.

[57] Ansuman Adhikary, Ebrahim Al Safadi, Mathew K Samimi, Rui Wang, Giuseppe Caire, Theodore SRappaport, and Andreas F Molisch, “Joint spatial division and multiplexing for mm-wave channels,”IEEE J. on Sel. Areas on Commun. (JSAC), vol. 32, no. 6, pp. 1239–1255, 2014.

[58] Ansuman Adhikary, Harpreet S Dhillon, and Giuseppe Caire, “Massive-MIMO meets HetNet: Inter-ference coordination through spatial blanking,” IEEE J. on Sel. Areas on Commun. (JSAC), 2014.

[59] Ansuman Adhikary, Harpreet S Dhillon, and Giuseppe Caire, “Spatial blanking and inter-tier coordi-nation in massive-mimo heterogeneous cellular networks,” in Globecom Workshops (GC Workshop).IEEE, 2014, pp. 1229–1234.

[60] Junyoung Nam, Giuseppe Caire, Young-Jo Ko, and Jeongseok Ha, “On the role of transmit correlationdiversity in multiuser MIMO systems,” CoRR, vol. abs/1505.02896, 2015.

[61] Palghat P Vaidyanathan and Piya Pal, “Sparse sensing with co-prime samplers and arrays,” IEEETransactions on Signal Processing, vol. 59, no. 2, pp. 573–586, 2011.

[62] PP Vaidyanathan and Piya Pal, “Theory of sparse coprime sensing in multiple dimensions,” SignalProcessing, IEEE Transactions on, vol. 59, no. 8, pp. 3592–3608, 2011.

D-4