d2 1 2 Proceedings of Second Special Session

HARP deliverable: D2.1.
2 Proceedings of second workshop/ special session at an international

conference
Project Title:
High capacity network Architecture with Remote radio heads & Parasitic antenna arrays
Grant Agreement Number: 318489
D2.1.2 Proceedings of second workshop/ special session at international conference

Authors:
Laura Cottatellucci, Eurecom
Constantinos Papadias, AIT
-1-
HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

conference
Table of Contents
1.
Introduction .......................................................................................................................... 3
2.
Second Workshop in an International Conference .............................................................. 3
3.
Special Session in an International Conference ................................................................... 6
3.1 General Information about the Special Session .................................................................... 7

3.2 Special Session Structure and Contributions ........................................................................ 8
4.
Special Stand-Alone Workshop ..................................................................................... 10
4.1 General Information about the GdR-ISIS workshop........................................................... 11

4.2 Workshop Structure and Contributions ............................................................................... 11
5.
Appendix ............................................................................................................................ 12
-2-

conference
1.
Introduction
This document describes the two major dissemination events which happened during Y3 of
HARP at major international scientific conferences. The first was a special workshop organized
at the 2015 IEEE 81st Vehicular Technology Conference (VTC2015-Spring), which took place in
Glasgow, UK, May 11-14, 2015, whereas the second was a special session organized at the 2015
European Signal Processing Conference (EUSIPCO 2015), which took place in Nice, France,
Aug. 31 Sept. 4, 2015. Both events were very well attended and received, stirring several
stimulating discussions and offering important technical insights. More details about them can be
found below:
2.
Second Workshop in an International Conference
An important aspect of the research work in HARP on the side of radio access is the enabling of
multi-antenna systems that can have a larger number of elements than what the available RF
hardware permits. The progress achieved in this direction within the project gives rise to
radically new MIMO technologies, such as the concept of cooperative transmission over remote
radio heads that employ a selected set of fixed beams shaped by parasitic arrays. Furthermore,
they are poised to play a catalytic role in 5G communication networks. This kind of work is
hence strongly related with the themes of Emerging MIMO Technologies and 5G Networks
that dominated the VTC2015-Spring workshop that was co-organized by HARP members.
Another two FP7 projects participated in the organization of this workshop: MiWaves
(www.miwaves.eu), MiWEBA (www.miweba.eu), which are both related with the other
important topic of the workshop, Millimeter-wave access and backhauling.
A brief description of the workshop concept can be found below:
Beyond 2020 Heterogeneous Wireless Networks with Millimeter-Wave Small Cell Access and
Backhauling Massive MIMO and millimetre-waves (mmW) are seen as key technology enablers
for future 5G wireless mobile networks. In fact, only taking advantage of multiple radio access
technologies will allow achieving higher capacities and bandwidths, reduced system power
consumption and lower electromagnetic field exposure. The lack of un-fragmented available
spectrum resources below 6 GHz and the considerable progress of mmW radio technologies over
the last few years have triggered a strong interest for the exploitation of mmW bands in future
wireless cellular networks for both backhauling and access. Massive MIMO is also considered
among the most promising technologies to achieve the challenging 5G system KPI and to take
full advantage of those in such new scenarios. This workshop will bring together academic
researchers and industrial professionals to identify and discuss technical challenges and recent
results related to mmW and massive MIMO in the context of future 5G mobile wireless networks.
2.1 General Information about the Special Session

Title: Emerging MIMO Technologies and Millimeter-waves for 5G Networks Workshop
Website Link: http://www.ieeevtc.org/vtc2015spring/workshops.php#wkshp_2/W3
-3-

conference
Hosting Conference: IEEE VTC2015-Spring
Date and Place: May 11, 2015 at Glasgow, UK
Workshop Chairs: Mathini Sellathurai (Heriot-Watt University, UK), Laurent Dussopt (CEALeti, France), Bruno Clerckx (Imperial College London, UK) and Tharmalingam Ratnarajah
(The University of Edinburgh, UK). Both Drs., Clerckx and Ratnarajah are members of HARP.
Organising Committee:
General Chairs:
Emilio Calvanese Strinati, CEA-LETI, France
Valerio Frascolla, Intel Mobile Communications, Germany
Thomas Haustein, Fraunhofer HHI, Germany
Zhinong Ying, Sony Mobile Communications, Sweden
TPC Chairs:
Laurent Dussopt, CEA-LETI, France
Mathini Sellathurai, Heriot Watt University, UK
Bruno Clerckx, Imperial College London, UK
Constantinos Papadias, Athens Information Technology, Greece
Tharmalingam Ratnarajah, The University of Edinburgh, UK
Kei Sakaguchi, Osaka University, Japan
Yue Frank Gao, Queen Mary University of London, UK
Publicity Chairs:
Thorsten Drger, National Instruments, Germany
Jessica Oueis, CEA-LETI, France
Number of Attendees: 40 on average.
Number of Invited Keynotes: 3, one of which was given by HARP member Dr. Constantinos
Papadias.
Number of TPC members/reviewers: 7, out of which 3 were HARP members.
2.2 Workshop Structure and Contributions

The workshop consisted of three sessions, each preceded by an invited Keynote presentation.
Each session consisted of a number of peer-reviewed papers in the corresponding areas. A
complete list of the workshop agenda can be found below:
Emerging MIMO Technologies and Millimeter-waves for 5G Networks Workshop
VTC2015-Spring, Glasgow, UK, May 11, 2015
Agenda:
-4-

conference
10:00-11:50 - Emerging Massive MIMO Technology
Chair: Mathini Sellathurai
Keynote:
Prof. C. Papadias (Athens Information Technology) HARP Member
Single-RF Transmission: An Emerging Technology
for both Link and Multi-User MIMO Systems
Papers:
1. M. Cui, B. Hu, X. Li (South China University of Technology), H. Chen (Guilin

University of Electronic Technology), A Novel Power Control Algorithm for Massive
MIMO Cognitive Radio Systems Based on Game Theory
2. M. Nasiri Khormuji (Huawei Technologies Sweden AB), Generalized Semi-Orthogonal
Multiple-Access for Massive MIMO
3. P. Harris, S. Zhang, A. Nix, M. Beach, S. Armour and A. Doufexi (Univ. of Bristol, UK),
A Distributed Massive MIMO Testbed to assess Real-World Performance & Feasibility
4. R. Hicheri, N. Hajri, N. Youssef (University of Carthage, Tunisia), M. Patzold
(University of Agder, Norway), T. Kawabata (University of Electro-Communications,
Japan), Statistical Analysis of the Channel Capacity Outage Intervals in Massive MIMO
Systems with OSTBC over Rayleigh Fading Channels
13:30-15:00 - MM-Wave Channel Modeling and Characterisation
Chair: Laurent Dussopt, CEA-LETI, France
Keynote: Dr. M. Nekovee (Samsung Electronics UK),

Unlocking spectrum above 6 GHz for 5G and IoT: Key Technologies and Uses
Papers:
1. S. Hfner, D. A. Dupleich, R. Mller, C. Schneider, R. S. Thom (Technische Universitt

Ilmenau, Germany), J. Luo, E. Schulz, X. Lu, T. Wang (Huawei Technologies),
Characterisation of Channel Measurements at 70GHz in Indoor Femtocells
2. Q. (Clara) Li, H. Shirani-mehr, T. Balercia, A. Papathanassiou, G. Wu (Intel Corp.), S.
Sun, M. Khalil Samimi, T. S. Rappaport (New-York University), Validation of a
Geometry-Based Statistical mmWave Channel Model Using Ray-Tracing Simulation
3. S. Montagner, N. Benvenuto (Univ. of Padova, Italy), P. Baracca (Bell Labs, AlcatelLucent, Germany), Channel Estimation Using a 2D DFT for Millimeter-Wave Systems
15:30-17:00 - MM-Wave Antennas and Channel
-5-

conference
Chair: Laurent Dussopt, CEA-LETI, France
Keynote: Dr. J. Putkonen (Nokia),

Opportunities of mmWave technology in 5G access and backhaul
Papers:
1. S. Le Hong Nguyen, K. Haneda, J. Jarvelainen, A. Karttunen (Aalto University, Finland),

J. Putkonen (Nokia, Finland), On the Mutual Orthogonality of Millimeter-wave Massive
MIMO Channels
2. K. Zhao (KTH Royal Institute of Technology, Sweden), J. Helander, Z. Ying (SONY
Mobile Communications AB, Sweden), D. Sjberg, M. Gustafsson (Lund University,
Sweden) and Sailing He (KTH Royal Institute of Technology, Sweden), mmWave
Phased Array in Mobile Terminal for 5G Mobile System with Consideration of Hand
Effect
3. K. Zhao (KTH Royal Institute of Technology, Sweden), Z. Ying (SONY Mobile
Communications AB, Sweden) and S. He (KTH Royal Institute of Technology, Sweden),
Human Exposure of mmWave Phased Array Antennas in Mobile Terminal for 5G
Mobile System
The workshop stirred many interesting discussions on the topics of future multi-antenna
technology, mmWave systems, channel models and transmission techniques and was cited as
highly successful.
The papers of the workshop are appended in a separate file.
3.
Special Session in an International Conference
The strategic goal of HARP is to bring distributed multi-antenna wireless access to reality by
combining two powerful emerging technologies. Namely, radio remote heads (RRHs), which
allow for widely geographically distributed access via radio-over-fiber connections to a central
base station, and load controlled antennas, which provide multi-antenna-like functionality with
a single active RF chain only. Recent developments in 5G identify Massive MIMO systems and
Cloud-based Virtual MIMO systems supported by cloud architectures (shortly Cloud-RAN) as
the most promising technologies. The concepts of RRH architectures and load controlled
antennas developed in HARP are expected to be key enablers for both massive MIMO networks
and Cloud-RAN. To further explore synergies, identify technical challenges and exchange results
and strategic visions, HARP promoted and organized a special session titled Massive and
Cloud-based Virtual MIMO: alternative, complementary or merging wireless network
technologies? at EUSIPCO 2015, the flagship conference of the European Association for
Signal Processing EURASIP.
As from the title, the special session aimed to explore complementarities of massive MIMO and
C-RAN in 5G with special attention to possibilities of developments for the new antenna array
architectures proposed in HARP.
-6-

conference
On the one hand, Massive MIMO systems employ a very large number of collocated antennas to
provide impressive advantages in combatting interference without the costly cooperation and/or
coordination, signalling and synchronization load of current cooperative multi-point (CoMP)
architectures. Crucial to this is that the capacity of networks with access points having a large
number of antennas, is not limited by data interference from adjacent cells, power per bit
vanishes as the number of antennas grows, and simplified linear detectors such as matched filters
can be used without substantial loss in performance.
On the other hand, Cloud-based Virtual MIMO systems, while also employing a large number of
remote radio heads (RRH), are envisioned to place these antennas in a random and distributed
manner throughout the coverage area, and to link these RRHs via optical fibres - to a cloud
architecture with considerable computational capabilities. This centralized processing is
envisioned to reduce the traditional cost relating to CoMPs, while at the same time is expected to
improve power efficiency by reducing the distances between transmitting and receiving nodes.
Spectral efficiency can also be greatly enhanced by the possibility to jointly processing signals
which enables interference avoidance in downlink and which transforms all received signals into
signals of interests in uplink. Additionally, the users can benefit from a better coverage and a
more uniform quality of service.
For some time, these two technologies have been widely considered as alternative technologies.
On the one hand, massive MIMO is thought to better capitalize on the existing cellular
infrastructures to achieve communications free of inter-cell interference. On the other hand,
Cloud-based virtual MIMO is envisioned to maintain some of the benefits of massive MIMO,
but also to transcend the cellular architecture and offer appealing features typically associated to
small cells.
The special session organized by HARP proposed a comparison between and conceivably even
a merging of - these two technologies and put emphasis on both theory and practice. From a
hardware point of view, the focus was in understanding the effects of novel hardware (HW)
antenna-array architectures such as load controlled antennas, which promise to enable merging
of the two technologies by enabling RRH with massive MIMO array-based on a single amplifier
chain, while at the same time offering gains in terms of cost, dissipated heat, and physical size.
The rest of this chapter is structured as follows. In Section 3.1 we provide general information
about the special session. Section 3.2 provides information about the invited paper authors and
gives an overview of the technical contributions. Finally, the whole contributions to the special
session by the invited authors are collected in Appendix.
3.1 General Information about the Special Session

Title: Massive and Cloud-based Virtual MIMO: alternative, complementary or merging wireless
network technologies?
Website Link: http://www.eusipco2015.org/content/special-sessions#SS12
Hosting Conference: EUSIPCO 2015
Date and Place: September 4, 2015 at Nice, France
-7-

conference
Workshop Organizer/Chair: Laura Cottatellucci
Petros Elia (EURECOM, France)
(HARP member, EURECOM, France),
Number of Attendees: About 25 in average.

Number of Invited papers: 5, two out of five were contributions from European FP7 projects,
included HARP.
Number of TPC members/reviewers: 13, two out of 13 were HARP members.
3.2 Special Session Structure and Contributions

The special session consisted in presentation of 5 invited papers by well-known scientists
working in the field of massive MIMO and C-RAN, both from Europe, USA and Canada.
The special session consisted of two papers focused in the analysis of massive MIMO and two
papers devoted to the design issues in C-RAN. A fifth paper considering practical design issues
inspired by load controlled antennas proposed in HARP, which provides a potential key
hardware enabler for both network architectures.
It worth noting that all the independent presentations in the special session fuelled many
discussions such that the one hour and 40 minutes special session extended to over two hours and
was ended just because of the beginning of the following session.
The following we list the invited authors and specify their contributions:
1) Prof. Erik G. Larsson from Linkping University, Sweden, member of FP7-MAMMOET.
He contributed with the paper Distributed massive MIMO in cellular networks: impact
of imperfect hardware and number of oscillators
2) Prof. Robert Heath from the University of Texas at Austin, USA which contributed with
the paper: Base station cluster patterns for semi-static multi-cell cooperation in irregular
network topologies.
3) Prof. Ralf Mller from Friedrich-Alexander-Universitt at Erlangen-Nrnberg, Germany,
member of the European project FP7-HARP. He contributed with the paper: Multidimensional Continuous Phase Modulation in uplink of MIMO systems.
4) Prof. Gerhard Kramer from the Technische Universitt Mnchen, Munich, Germany. He
contributed with the paper MIMO and Massive MIMO - Analysis for a Local Area
Scenario.
5) Prof. Wei Yu from University of Toronto, Canada which contributed with the paper:
Performance Comparison of Data-Sharing and Compression Strategies for Cloud Radio
Access Networks.
The contribution were presented in the order of the list above with the specific intention of
alternating works on massive MIMO systems and cloud-RAN while keeping in a central position
the gluing contribution on the load controlled antennas proposed in HARP. In the following we
provide an overview of those technical contributions and we follow the same order. Thus, the
story line could appear fragmented. We invite the reader to recall guidelines of the special
session to find a unitary perspective.
-8-

conference
The special session was open by the presentation of the contribution Distributed massive
MIMO in cellular networks: Impact of imperfect hardware & number of oscillators by
Bjornson, Matthaiou, Pitarokoilis, and Larsson which focuses on the impact of the hardware
non-idealities on distributed massive MIMO systems. This is a solution that goes in the direction
of merging massive MIMO systems with the distributed infrastructure of C-RAN. In fact,
distributed massive multiple-input multiple-output (MIMO) combines the array gain of coherent
MIMO processing with the proximity gains of distributed antenna setups. The authors analyze
how transceiver hardware impairments affect the downlink with maximum ratio transmission.
They derive closed-form spectral efficiencies expressions and study their asymptotic behavior as
the number of the antennas increases. This contribution proves a scaling law on the hardware
quality, which reveals that massive MIMO is resilient to additive distortions, while multiplicative
phase noise is a limiting factor. The authors can also conclude from their analysis that it is better
to have separate oscillators at each antenna than a single oscillator per BS.
The subsequent presentation was focused on issues that raise in clustering of cells in practical
C_RAN when the network topology is irregular. The work titled Base station cluster patterns
for semi-static multi-cell cooperation in irregular network topologies and authored by Park, Lee,
and Heath, proposes a clustering strategy for semi-static multicell cooperation. Semi-static
multicell cooperation exploits multiple predefined base station (BS) cluster patterns for
improving cell-edge user throughput. The proposed clustering guarantees that every user
communicates with their two closest BSs, so that users are protected from the dominant
interferer. The key idea of the proposed clustering is to use the 2nd-order Voronoi region to form
BS clusters. Each of the formed BS clusters is mapped into a particular cluster pattern by
exploiting the edge-coloring in graph theory. Through simulations, the performance is compared
to that of other conventional strategies. The major finding of this contribution is that the
proposed clustering provides performance gains for cell-edge users compared to that of the
conventional strategies.
The central presentation of the special session proposes some fundamental transmission scheme
inspired by the load controlled antennas proposed in HARP. The presentation titled Multidimensional Continuous Phase Modulation in uplink of MIMO systems by Sedaghat and Mller
focuses on a phase modulation scheme over an hypersphere (PMH) with constant instantaneous
power. This modulation scheme does not suffers from the problem Pick-to-Average Power Ratio
(PAPR) and thus presents very appealing features in terms of efficient energy utilization in
antenna transmission. The authors show that for an i.i.d. Gaussian channel, the capacity
achieving input distribution is approximately uniform on a hypersphere when the number of
receive antennas is much larger than the number of transmit antennas. Moreover, in the case that
channel state information is not available at the transmitter, it is proven that the capacity
achieving input distribution is exactly uniform on a hypersphere. The authors evaluate mutual
information between input and output of PMH with discrete constellation for an i.i.d. Gaussian
channel numerically. Furthermore, they propose a spherical spectral shaping method for PMH to
have Continuous Phase Modulation on the Hypersphere (CPMH). In CPMH, the continuous time
signal has a constant instantaneous sum power. It is shown that using a spherical low pass filter
in spherical domain followed by a Cartesian filter results in very good spectral properties.
-9-

conference
The forth work by Dierks, Zirwas, Jager, Panzner, and Kramer is centered on the comparison of
centralized and distributed MIMO systems and it is titled MIMO and Massive MIMO Analysis for a Local Area Scenario. The authors analyze the performance of centralized and
distributed deployments of massive MIMO in an office building both with and without
cooperation. They show that using twice as many base station antennas as data streams provides
most of the massive MIMO benefits. They propose a simple transmission scheme that achieves
user fairness and operates close to a capacity upper bound. Finally, they evaluate the tradeoff
between performance and cost for backhauling by comparing cooperation of distributed base
stations with a single central deployment.
The special session was concluded by the presentation of Prof. Wei Yu on Performance
Comparison of Data-Sharing and Compression Strategies for Cloud Radio Access Networks
which raised many discussions. Prof. Wei Yus work tackled the issue of comparing the systemlevel performance of two fundamentally different transmission strategies for the downlink of a
cloud radio access network. The two strategies, namely the data-sharing strategy and the
compression-based strategy, differ in the way the limited backhaul is utilized. While the datasharing strategy uses the backhaul to carry raw user data, the compression strategy uses the
backhaul to carry compressed beamformed signals. Although these strategies have been
individually studied in the literature, a fair comparison of the two schemes under practical
network settings was challenging because of the complexity in jointly optimizing user
scheduling, beamforming, and power control for system-level performance evaluation, along
with the need to optimize cooperation clusters for the data-sharing strategy and quantization
noise levels for the compression strategy. This contribution to the special session presents an
optimization framework for both the data-sharing and compression strategies, while taking into
account losses due to practical modulation in terms of gap to capacity and practical quantization
in terms of gap to rate-distortion limit. The main conclusion of this work is that the compressionbased strategy, even with a simple fixed-rate uniform quantizer, outperforms the data-sharing
strategy under medium to high capacity backhauls. However, the data-sharing strategy
outperforms the compression strategy under low capacity backhauls primarily because of the
large quantization loss at low backhaul capacity with compression.
4.
Special Stand-Alone Workshop
A special stand-alone workshop where to present the results of the HARP platform has been
organized. The opportunity was offered by the French Groupement de recherche (GdR) in the
area of Information Signal Image and viSion (ISIS), shortly GdR-ISIS, which sponsors meetings
and workshops under the patronage of the French CNRS (Centre national de la recherche
scientifique). HARP proposed and is promoting a workshop aligned with the central topics of the
project titled Cloud- and fog-based PHY communications in 5G: performance, feedback and
complexity.
As emphasized in the title, the workshop focuses on the 5G challenges, and will bring to the fore
new ideas and paradigms in wireless communications, with special preference on works relating
to massive MIMO, Cloud RAN, caching based techniques and edge-computing. In the setting of
massive MIMO, the idea is to equip base stations with a large number of antennas, in order to
allow faster rates, higher spectral and energy efficiency, as well as to allow for simpler
- 10 -

conference
algorithms that benefit from the fact that large systems often converge to their deterministic
limits. This relates somewhat to the concept of cloud RAN where now, a centralized entity
controls transmissions from a large number of remote radio heads that can be closer to the users.
In both these settings, which present their own opportunities and challenges, one can use the idea
of anticipatory distributed caching that allows for predictably popular content to be placed closer
to the users. All these ideas come together to give promise to a new way of communicating.
The proposed workshop aims to be an ideal breeding ground to both discuss technologies that
are at the front end of the current vision for wireless networks and to get inspiration and
understanding of the plethora of ways they could inter-operate.
The rest of this chapter is structured as follows. In Section 4.1 we provide general information
about the workshop. Section 4.2 details information about the invited speakers and gives an
overview of the technical contributions with special emphasis on the HARP contributions.
4.1 General Information about the GdR-ISIS workshop

Title: Cloud- and fog-based PHY communications in 5G: performance, feedback and complexity
Website Link: http://www.gdr-isis.fr/index.php?page=reunion&idreunion=291
Patronage and sponsor: CNRS (French National Center of Scientific Research) GdR-ISIS
Date and Place: November 20, 2015 at Paris, France
Workshop Organizer/Chair: Laura Cottatellucci
Petros Elia (EURECOM, France)
(HARP member, EURECOM, France),
Number of Invited Speeches: 6, two out of six were contributions from the European FP7
project HARP.
It is worth noticing that the information about the workshop are necessarily incomplete since the
workshop have not taken place yet at the time this deliverable was written.
4.2 Workshop Structure and Contributions

The workshop consists of 6 invited speeches and 3 or 4 contributed speeches.
The keynote speech by Prof. Giuseppe Caire will be followed by the invited speeches alternated
by shorter contributed speeches.
The following we list the invited speakers and the titles of their speeches:
1) Giuseppe Caire (TU-Berlin) presenting From Cloud to Fog: Massive MIMO, pCells,
Cellular Interference Alignment, and Device-to-Device.
2) Georgios Paschos (Huawei, Paris) presenting Load Balancing in Cloud (and Fog)
Computing.
- 11 -

conference
3) Dora Boviz (SUPELEC and Alcatel-Lucent, Paris) presenting C-RAN fronthaul
enhancements using Software Defined Networking.
4) Dirk Slock (EURECOM, Sophia Antipolis), presenting Sum Utility Optimization in
MIMO Multi-User Multi-Cell: Centralized and Distributed, Perfect and Partial CSIT,
Fast and Slow CSIT
5) Laurent Roullet (Alcatel Lucent, Paris), presenting CRAN challenges: the architecture
and solutions in HARP.
6) Stefan Valentin (Huawei, Paris), presenting Context-Aware and Anticipatory Resource
Management for the 5G RAN.
HARP contributed to this workshop with the presentations by the HARP members Dora Boviz
and Laurent Roullet which will illustrate and promote HARP achievements.
More specifically, Dora Boviz in his speech C-RAN fronthaul enhancements using Software
Defined Networking will illustrate the advantages offered by C-RAN both on the functional and
the hardware plane. The access network can be efficiently controlled using Software Defined
Networking and this is not the only benefit of SDN technology. She will illustrate how
optimization of various network features and elements can be done by plug-and-play applications
interfaced with the SDN controller. In fact, it orchestrates all the network elements from Remote
Radio Heads (RRHs) to the Baseband Unit (BBU) pool and all these network elements provide
real-time measurements that can be efficiently used by optimization algorithms. Boviz will
present SDN enabled C-RAN architecture and its advantages for network control, then she will
focus on various features that it facilitates. These can impact wireless transmissions (e.g. eICIC,
Network MIMO), fronthaul network (e.g. routing over Ethernet fronthaul) or computational
resources (e.g. load balancing). The presentation will highlight how SDN allows scalable and
reconfigurable realization of these features and how it can improve them with respect to
Distributed RAN or centralized architectures without collective control over network elements.
Laurent Roullet in his speech on CRAN challenges: the architecture and solutions in HARP
will address the fundamental issue of scalability in centralized C-RAN and the solution proposed
in HARP. More specifically, a key challenge in cellular networks is the ability to
face "exponential" demand that can be addressed by active antenna systems with larger MIMO
modes and denser cell distributions. To support this scalability, the HARP architecture focuses
on joint design of centralized cloud RAN architecture and load-controlled antenna arrays. The
remaining bottleneck is therefore the stringent fronthaul requirements in terms of bandwidth,
jitter and latency. Laurent Roullet will illustrate how these requirements have been relaxed in
the HARP project without (too much) impairments.
5.
Appendix
This appendix contains the following documents:

1) The paper Distributed massive MIMO in cellular networks: Impact of imperfect hardware
and number of oscillators by Emil Bjrnson, Michail Matthaiou, Antonios Pitarokoilis, and
Erik G. Larsson.
- 12 -

conference
2) The paper Base station cluster patterns for semi-static multi-cell cooperation in irregular
network topologies, by Jeonghun Park, Namyoon Lee, and Robert W. Heath Jr.
3) The paper Multidimensional continuous phase modulation in uplink of MIMO systems by

Mohammad A. Sedaghat and Ralf R.Mller.
4) The paper MIMO and massive MIMO Analysis for a local area scenario by Stefan
Dierks, Wolfgang Zirwas, Markus Jager, Berthold Panzner, and Gerhard Kramer.
5) The paper Performance comparison of data-sharing and compression strategies for cloud
radio access networks by Pratik Patil, Binbin Dai, and Wei Yu
- 13 -
A Novel Power Control Algorithm for Massive MIMO

Cognitive Radio Systems Based on Game Theory
Manman Cui, Bin-jie Hu, and Xiaohuan Li
Hongbin Chen
School of Electronic and Information Engineering

South China University of Technology
Guangzhou, China
eebjiehu@scut.edu.cn
emma2012_scut@163.com
Key Laboratory of Cognitive Radio and Information

Processing (Guilin University of Electronic Technology),
Ministry of Education
Guilin, China
AbstractIn this paper, a novel system model named as

massive multiple-input multiple-output (MIMO) cognitive radio
system (CRS) is built and an efficient uplink power control
algorithm based on noncooperative game theory with a selfadaptive power threshold scheme is proposed to improve the
power efficiency. And then, we give an analytical model for the
massive MIMO CRS and the detail of the self-adaptive power
threshold scheme. Moreover, we prove the existence of the Nash
Equilibrium and the convergence of the proposed algorithm. To
evaluate the performance of the proposed algorithm, we do
simulations and compare the system performance with two other
classical power control algorithms. Simulation results
demonstrate that the proposed algorithm can achieve a
preferable performance in signal-to-noise-plus interference ratio
(SINR) and a higher utility with lower transmission power and
faster convergence. It implies that the novel way can achieve
higher power efficiency and a better overall system performance.
Keywordscognitive radio; spectrum sharing; massive MIMO;
power control; game theory
I.
INTRODUCTION
In recent years, the explosive growth of wireless

applications directly leads to the crisis of limited available
spectrum resources. While the existing spectrum measurement
results show that the available spectrum resources are very
low utilized even with some spectrum holes unused in the time
domain and the spatial domain [1]. Cognitive Radio (CR) has
been proposed as a promising technology to deal with the low
spectrum utilization problem caused by the current inflexible
spectrum allocation policy [2].
In a CRS, cognitive users (CUs) share the spectrum
dynamically with primary users (PUs) without causing
unbearable interference to PUs. So, an effective power control
method applied in this situation turns out to be required and
essential. As one key technology in CR, power control has
been studied for many years, and there are some works in this
region. While among them, the power control algorithm based
on game theory has been a mainstream technology. It is more
flexible and fair than other methods, because it can set up a
specific utility function for each CU accordingly in a specific
power control scheme, such as to maximize the CRS capacity,
to reduce the energy consumption, or to reduce the algorithm
complexity, etc. Authors in [3] imported the power threshold

to iterative algorithm using game theory in the CRS to achieve
a better anti-noise performance and a lower system
complexity. Reference [4] proposed a new iterative algorithm
under flat-fading environment to meet the fairness among CUs.
Reference [5] engaged in the pricing issue for the power
control problem in code division multiple access (CDMA)
based CRS to maximize the revenue of the base station (BS).
However, these algorithms set a fixed power threshold for
each different CUs. It couldn't be the optimal selection for all
the CUs, thus it would inevitably limit the power efficiency of
the CRS. To break this limitation, we import a self-adaptive
power threshold scheme to power control algorithm. The
threshold for one CU will be adjusted according to its current
state of communication. It increases the flexibility and
reasonability of the algorithm and is also good to help the
CRS reach its Nash Equilibrium quickly.
Besides, in the CRS, CUs share the spectrum
opportunistically with PUs, so CUs would certainly suffer
from the interference from the other CUs and PUs. There
would be a part of transmission power consumed to overcome
the interference as well as the influence of the channel fading.
While the existing works ignored the problem to reduce this
consumption and save the energy. Recently, there has been a
great deal of interest in massive MIMO systems [6]-[9]. In a
massive MIMO system, the BS is equipped with a few
hundreds of antennas and simultaneously serves tens of
terminals[10]. The random channel vectors between the
terminals and the BS become pairwise orthogonal when the
number of the antennas grows larger [11]. Such systems can
substantially reduce the influence of intra-cell interference and
small-scale fading with simple signal processing, and achieve
a better communication reliability. Therefore, massive MIMO
systems can effectively reduce the transmission power of the
terminal, improve its SINR correspondingly, and thus get a
higher power efficiency . The concept of massive MIMO
systems was initially investigated for cellular networks, but
there is a positive potential in the area of CRS.
In this paper, we apply massive MIMO into the CRS to
exploit its advantages to improve the power efficiency of the
CRS. We name the new system as massive MIMO CRS, and
describe the new system structure in details. Then we propose
978-1-4799-8088-8/15/$31.00 2015 IEEE
an efficient uplink power control algorithm based on

noncooperative game theory with a self-adaptive power
threshold scheme in this system. This scheme increases the
flexibility and reasonability of the algorithm and is also good
to help the CRS reach its Nash Equilibrium quickly. At last,
we do simulations and compare the results with that of two
other classical power control algorithms. It is shown that the
proposed algorithm in this system can achieve higher power
efficiency. That reflects the superiority of the proposed
algorithm in the massive MIMO CRS.
The rest of this paper is organized as follows. Section II
introduces the system model of the massive MIMO CRS. The
uplink power control algorithm based on noncooperative game
theory with a self-adaptive power threshold scheme in this
system is described in Section III. Section IV presents the
simulation results. Finally, Section V concludes this paper.
II.
SYSTEM MODEL
In this section, we apply massive MIMO into the CRS to

exploit its advantages to improve the power efficiency of the
CRS. We name the new system as massive MIMO CRS, as
schematically shown in Fig.1. In this model, we consider a
single-cell with a primary network and a secondary network.
The primary network includes a primary base station (PBS)
and a single PU both equipped with one antenna. Within its
communication area, one cognitive base station (CBS)
equipped with M antennas and K single-antenna CUs compose
the secondary network, where M>>K. So, there exists the
mutual interference between PU and CUs.
noises (AWGNs) at the CBS with its each element

nk~CN(0, k2 ).
The channel matrix G incorporates the effects of small-scale
fading and large-scale fading, we can write G=HD1/2. Here H
is an MK small-scale fading coefficient matrix between the K
CUs and the CBS, and its elements are independent identically
distributed (i.i.d.) random variables (RVs) with zero mean and
unit variance. While D is an KK diagonal large-scale fading
coefficient matrix whose diagonal elements are [D]kk bk, and
depend on k but not on m [11]. We assume that the small-scale
fading coefficients for different users are independent, the
column channel vectors of H would be asymptotically
orthogonal as the number of antennas at the BS grows to
infinity. So, we have GHG=D1/2HHHD1/2MD1/2IKD1/2=MD.
Similarly, we define gp = hp p , where hp and bp are assumed
to be the small-scale fading coefficient and large-scale fading
coefficient between PU and the CBS, respectively.
We assume that the CBS has a known channel matrix G,
and processes the signals of each antenna with the maximumradio combining (MRC) linear detector. After the MRC linear
detector, the K1 signal vector r of CUs received at the CBS
is
r = GH y.
(2)
Let rk be the kth element of r . Then, we have

rk = g Hk g k
p k sk + g Hk g i
(3)
pi si + g kH g p s p + g Hk n.
ik
where gk is the kth column of G. From (3) we get the SINR of

the kth CU at the CBS as follows,
2
k =
g kH gk pk
K
g
ik
H
k
(4)
gi pi + g k g p p p + g k
H
2
k
Recalling that Ik = gH gi pi + gH gp pp + gH k2 and rearranging

i k
the terms, we rewrite (4) as

2
k =
Fig. 1. System model.
We consider the uplink transmission of the massive MIMO

CRS. All CUs and the PU are assumed to simultaneously send
data using the same spectrum. Therefore, the M1 received
vector of each CBS antenna is given by
y = GP1 2 S + g p p p s p + n.
(1)
where G is the MK channel matrix between the CBS and K

CUs, P=diag{p1, p2, , pk, , pK} is the transmission power
matrix used by each CUs, S=[s1, s2, , sK]T corresponds to the
transmitted symbols and E{|sk|2}=1, gp [gp, , gp] is the
channel vector between the PU and M CBS antennas with
same elements, pp is the transmission power of PU, sp is the
symbol transmitted from PU to the M CBS antennas and
E{|sp|2}=1 , and n is an M1 vector of additive white Gaussian
g kH gk pk
Ik
(5)
The expectation of k can be given by
g Hk gk pk
=
E[ k ] = E K
2
2
2
g Hk gi pi + g Hk g p p p + g kH k2
i k
III.
( M 1) k pk
p +
i
(6)
p p + k2
ik
GAME-THEORETIC FORMULATION
In this section, we formulate the power control problem as a

noncooperative game in the massive MIMO CRS and get the
corresponding power control algorithm with a self-adaptive
power threshold scheme. Then we prove the existence of the
Nash Equilibrium and the convergence of the power control
algorithm.
A. Utility Function and Derivation of the Nash Equilibrium

In the CRS, the CUs opportunistically use the authorized
spectrum of PUs. In order to let them share the spectrum
harmoniously, an appropriate utility function should be
investigated to be distributed among the CUs. So that the total
CRS utility is maximized, and the interference created to PUs
is as small as possible.
We suppose that the utility function of the kth CU is
where P-k is the power vector of all CUs excluding
the kth CU. The Nash Equilibrium in the CRS means that any
CU couldn't achieved any additional revenue through
unilaterally changing its transmission power from its Nash
*
Equilibrium point pk , i.e. Jk ( pk* , P*k ) J k ( pk , P*k ) . How to select a
utility function becomes an important problem.
J k ( pk , P k ) ,
In practical, a good CRS should satisfy the following two

basic goals: 1) to maximize the total CRS throughput; and 2)
to protect the primary system from harmful interference. So,
the transmission power and SINR of the CU are two critical
parameters for one utility function. After the comparison with
several utility functions, we define the following new utility
function,
1
J k ( pk , P k ) = ak ( pk ,th pk ) 2 + bk ( k min ) 2 .
(7)
where pk,th min{pk,max, pk,temp} is a self-adaptive threshold of

the kth CU, pk,max is the maximum allowable transmission
power of the kth CU, pk,temp is the transmission power constraint
determined by PU's interference temperature constraint, k is
in accordance with formula (4), min is the minimum SINR
requirement of CUs and assumed to be same for all CUs, ak
and bk are the correlation factors selected based on the
selection criteria which was mentioned in [4].
In the paper, we assume all CUs have the same maximum
allowable transmission power pk,max. Now we derive to get the
other power constraint pk,temp.
We assume Tth is the interference temperature constraint of
PU. According to the definition, we get
K
hip
Tth .
(8)
i =1
where hip is the channel coefficient between the ith CU and the
PBS. So, from (8) we can get pk,temp as follows
p k , tem p =
1
h kp
Tth
i k
p i hip
pk .
(9)
From (9), we see pk,temp can be self-adaptively adjusted based

on the current state of communication to constrain CU's
transmission power. Accordingly, we name pk,th min{pk,max,
pk,temp} as a self-adaptive power threshold scheme. It would be
more practical and reasonable to reduce the effects on PU and
help the CRS reach its Nash Equilibrium quickly.
Based on the necessary conditions for Nash Equilibrium, we
have
1
1
J k
b
a

= 0 = k ( k min ) 2 k k ( pk ,th pk ) 2 .
pk
pk 2
2
(10)
From (5), we can get the first derivation of k on pk as

g H gk
k
= k
pk
Ik
. So, we can get
g H gk
b
k = min + k k
ak
Ik
(p
k ,th
pk ) .
(11)
Thus we can further write the power as follows,

pk = k
Ik
= min
g Hk g k
Ik
2
g Hk g k
b
+ k
ak
(p
k , th
pk )
g kH g k
Ik
(12)
B. Power Control Algorithm

From (6) and (12), the power iterative algorithm
corresponding to the kth CU can be obtained. The pseudocodes
of the algorithm are listed as follows.
Algorithm: power control algorithm based on game theory
( 0)
Set n:=0, initialize powers pk , k = 1, 2,..., K .
Repeat{
For each CU, calculate the SINR

k( n) =
( M 1) k pk( n )
p
i
(n)
i
(13)
+ p p p + k2
ik
For each CU, update uplink transmission power
f k( n ) pk ( n) = pk
min
( n +1)
pk ( n )
(n)
b
+ k
ak
(n)
( p( ) p ( ) ) p ( )
n
k ,th
, if positive
,otherwise
(14)
Set n: = n+1 for each CU.

} until n=N.
There is an implicit assumption that whenever the
expression of power is negative, the assigned power will be
zero. The reason is that the power cannot be negative in
implementation. The algorithm will run in real time with
measurements potentially updated at every step of the
algorithm.
C. Existence of Nash Equilibrium
Part B gives the power control iterative equation (14). In
this part, we prove the existence of a solution for Nash
algorithm algebraic equation to ensure there exists a unique
solution to (14). We adduce the proof by using the implicit
function theorem in the following subsection.
From (6) and (12), we define that

Fk ( pk , P k , k , i , a 2k , b 2k , k2 ) = 0
= pk + min
p
i
ik
+ p p p + k2
( M 1) k
b
+ k
ak
(p
k ,th
pk )
Thus, the scalability has been proved.
( M 1) k
p
i
+ p p p + k2
ik
(15)
i, k = 1, 2,..., K , k i.
According to the implicit function theorem, there must be a

solution to (15) as long as the Jacobian matrix (the matrix of
partial derivatives Fk pk ) is nonsingular at the point of
existence. Note that the corresponding algebraic equations are
represented by the right-hand side of the above formula. So
that the corresponding Jacobian matrix has 1 ba ( M I 1 )
2
on the main diagonal and
min
i bk

k ak
(p
k , th
pk )
k i
Note that the solution exists if these terms such as

i , k , a 2k , b2k , k2 are all extremely small and will not have an
impact on the nonsingularity of the corresponding Jacobian
matrix. As demonstrated in [4], the suitable (bk/ak)2 is very
small. In practical, the value of min is not too large and the
attenuation coefficient bk can be very small (the order of 10-2),
which is relevant to the distance between the receiver and the
transmitter. So, when those values are reasonably set, the
Jacobian matrix will be nonsingular, that is to say, the solution
of the defined Nash algebraic equations exists.
D. Convergence of the Algorithm
If the power iterative algorithm satisfies the following three
properties (positivity, monotonicity, and scalability), and a
unique and fixed point of the algorithm exists, the algorithm
can converge to the unique and fixed point [12]. Part C has
proven the existence of the unique and fixed point, so we now
proceed to discuss the three properties.
1) positivity: fk(p)>0;
b
+ k
k ak
pk
(p
k ,th
pk )
k
pk
where the parameters are positive and pkpk,th. So fk(p)>0

always stand up. Therefore, it satisfies the positivity.
2) monotonicity: p1>p2, fk(p1)>fk(p2);
2
bk pk,th k
min
.
p
p
f
p
f
p
p
p
>
>
0,
(
)
(
)
(
)
Assuming 1 2

1
2
1
2
k
k
k ak p1 p2
2
To ensure fk(p1)>fk(p2), it requires
bk min p1 p2
< 2
k pk ,th k
ak
So the two cost factors should take appropriate values to

ensure the monotonicity of the algorithm.
3) scalability: "a>1, fk(ap)>afk(p).
From pkpk,th and "a>1,
2
IV. SIMULATION RESULTS

In this section, we show the data charts of power, SINR,
and utility through simulation tests. We compare the results
with that of the noncooperative power control game (NPCG)
algorithm [4] in a single-antenna system and the well-known
Koskie-Gajic (K-G) algorithm obtained through game theory ,
which is shown as follows,
outside the
I k2
main diagonal. So, the value of the Jacobian matrix is related

2
2
2
to min , pk ,th , pk , i , k , ak , b k , k .
According to (14), we get f k ( pk ) = min
From the certification process of convergence, we know

when the parameters are set reasonably, the algorithm would
converge.
p 1
b p
fk ( p) fk ( p) = k,th k (1) k < k k,th k k (1 ) < 0 .

p
a
p
k
k k
pk (
n +1)
p (n) a
= min k( n ) k
k 2bk
pk ( n )
( n) .
k
(16)
Firstly, we claim the system environment to which our

algorithm is applied. We consider a single hexagonal cell with
the radius (from center to vertex) of 1 km, where the CBS
with 200 transceiver antennas is centered at the origin and 20
single-antenna CUs are located uniformly in the cell. The PU
also works near the CBS, which may generate mutual
interference between the PU and CUs. We assume that no user
gets closer to the CBS than dh=100m. We ignore fast fading,
shadowing, and interference from adjacent cells.
The background noise power is assumed to be k2 =510W. The large-scale fading coefficient is determined
according to bk=zk/(dk/dh)a, where dk denotes the distance
between the kth CU to the CBS, zk is a log-normal random
variable with standard deviation sshadow, and a is the path loss
exponent. We define gp=8dB as the SINR of the PU, from
which we can get the value of Tth. We choose sshadow=8dB,
a=3.8, gmin=8dB, pk,max=14mW, p(0)=10-10W, (bk/ak)2=410-8,
pp=1W.
15
As shown in Fig.2, we find that both the proposed algorithm

and the NPCG algorithm can achieve convergence faster than
the K-G algorithm. That means the first two algorithms have a
superior performance of fast convergence to fit in the dynamic
network environment. Besides, the proposed algorithm can
reach a steady state a little faster than NPCG algorithm. That
owes to the self-adaptive power threshold scheme. In addition
to considering the maximum allowable transmission power of
CU, the scheme also considers the current interference to the
PU to constrain CU's transmission power. It helps the CU to
choose a power much closer to its Nash Equilibrium point thus
improving the convergence speed of the algorithm.
Furthermore, we can see that the first two algorithms both can
meet the SINR requirements. But the transmission power used
in proposed algorithm almost is half of that used in NPCG
algorithm, and the former can get a preferable overall SINR
performance than the latter. It is due to the advantages of
massive MIMO systems in reducing the influence of intra-cell
interference and small-scale fading. So the CUs in the massive
MIMO CRS can effectively reduce its transmission power and
get a better SINR.
Fig. 3 describes each CU's utility according to the utility
functions of the first two algorithms. It is obviously that in the
same environment each CU under the proposed algorithm can

achieve a higher utility than that under the NPCG algorithm.
All of these can reflect the superiority of the self-adaptive
threshold scheme and the massive MIMO CRS.
the proposed
NPCG
0.8
1.8
0.7
1.6
K-G
1.4
1.2
1.4
0.6
0.4
0.3
power(mW)
power(mW)
power(mW)
1.2
0.5
1
0.8
0.8
0.6
ACKNOWLEDGMENT
0.4
This work was supported by the National Natural Science

Foundation of China under Grant U1035002 and 61162008,
the Guangdong-Hong Kong Key Project of Science and
Technology
under
Grant
2011A011305001
and
2011A011302001, and the University-Industry Key Project of
Department of Education of Guangdong Province under Grant
CGZHZD1102.
0.6
0.2
0.4
0.1
0
0.2
0.2
10
20
Iteration
10
20
10
Iteration
20
Iteration
(a)
the proposed
NPCG
35
theory with a self-adaptive power threshold scheme in a

massive MIMO CRS is proposed. The massive MIMO is
applied to the CRS to reduce intra-cell interference and the
influence of small-scale fading. A self-adaptive power
threshold scheme is imported to the power control algorithm
to constrain each different CUs in time. Simulation results
show that the proposed algorithm in the massive MIMO CRS
can achieve a preferable SINR performance and a higher
utility with lower transmission power and faster convergence.
It implies that the novel way can achieve higher power
efficiency and a better overall system performance.
K-G
25
25
20
20
REFERENCES
30
[1]
25
8.2
8.1
15
8
4
15
SINR(dB)
8.3
20
SINR(dB)
SINR(dB)
8.02
8.01
8
10
10
15
10
10
SINR Threshold
SINR Threshold
SINR Threshold
10
10
Iteration
Iteration
10
Iteration
(b)
Fig. 2. Transmission power and SINR comparison of the proposed algorithm,
the NPCG algorithm and the K-G algorithm.
the proposed
NPCG
35
30
7
6
25
Utility
Utility
5
20
15
4
3
10
2
5
10
Iteration
15
10
15
Iteration
Fig. 3. Utility comparison of the proposed algorithm and the NPCG

algorithm.
V.
CONCLUSION
To achieve higher power efficiency of the CRS, an efficient

uplink power control algorithm based on noncooperative game
Federal Communications Commission, Spectrum Policy Task Force

Report. ET Docket, 2002.
[2] Q. S. Guan, F. R. Yu, S. M. Jiang, and G. Wei, Prediction-based
topology control and routing in cognitive radio mobile ad hoc
networks, IEEE Transaction on Vehicular Technology, vol. 59, no. 9,
pp. 4443-4452, Nov. 2010.
[3] F. Li, X. Tan, L. Wang, A new game algorithm for power control in
cognitive radio networks, IEEE Transaction on Vehicular Technology,
vol. 60, no. 9, pp. 4384-4391, Nov. 2011.
[4] J. H. Zhao, T. Yang, G. Yi, J. Wang, and L. FU, Power control
algorithm of cognitive radio based on non-cooperative game theory,
China Communications, vol. 10, no. 11, pp. 143-154, Nov. 2013.
[5] H. Yu, L. Gao, Z. Li, X. B. Wang, and E. Hossain, Pricing for uplink
power control in cognitive radio networks, IEEE Transactions on
Vehicular Technology, vol. 59, no. 4, pp. 1769-1778, May 2010.
[6] L. Lu, G. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, An
overview of massive MIMO: benefits and challenges, IEEE Journal of
Selected Topics in Signal Processing, vol. 8, no. 5, pp. 742-758, Oct.
2014.
[7] J. Hoydis, K. Hosseini, S. T. Brink, and M. Debbah, Making smart use
of excess antennas: massive MIMO, small cells, and TDD, Bell Labs
Technical Journal, vol. 18, no. 2, pp. 5-21, Sept. 2013.
[8] J. Choi, Massive MIMO with joint power control, IEEE Wireless
Communications Letters, vol. 3, no. 4, pp. 329-332, Aug. 2014.
[9] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive
MIMO systems with non-ideal hardware: energy efficiency, estimation,
and capacity limits, IEEE Transactions on Information Theory, vol. 60,
no. 11, pp. 7112-7139, Nov. 2014.
[10] J. Hoydis, S. T. Brink, and M. Debbah, Massive MIMO in the UL/DL
of cellular networks: how many antennas do we need? IEEE Journal on
Selected Areas in Communications, vol. 31, no. 2, pp. 160-171, Feb.
2013.
[11] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, Energy and spectral
efficiency of very large multiuser MIMO systems, IEEE Transactions
on Communications, vol. 61, no. 4, pp.1436-1449, Apr. 2013.
[12] R. D. Yates, A framework for uplink power control in cellular radio
systems, IEEE Journal on Selected Areas in Communications, vol. 13,
no. 7, pp. 1341-1347, Sep. 1995.
Generalized Semi-Orthogonal Multiple-Access for

Massive MIMO
Majid Nasiri Khormuji
Huawei Technologies Sweden AB, Stockholm
Emails: majid.nk@huawei.com
AbstractWe propose a novel framework to enable concurrent
transmission of multiple users to an access node equipped with a
massive multiple-input multiple-output (mMIMO), which encompasses and extends the conventional time-division duplex (TDD)
protocol and the recently proposed semi-orthogonal multipleaccess (SOMA). The new solution is referred to generalized semiorthogonal multiple-access (GSOMA). To enable GSOMA, the
users are grouped such that the users within each group are
assigned mutually orthogonal pilot sequences. However the pilot
sequences can be reused in different groups and the users in
different group are coordinated such that their pilot sequences
do not interfere by a semi-orthogonal mechanism enabled by a
resource-offset sequence. We describe the general framework and
analyze a case of GSOMA and show that a properly designed
GSOMA can outperform the conventional TDD and SOMA.
I. I NTRODUCTION TO M MIMO
Massive multiple-input multiple-output (mMIMO) communication is considered as one of promising physical-layer
solutions to enable multi-user communication for future communications systems [1][4] in response to ever-increased
demand for high data rate applications as well as to a more
homogenous quality-of-service across the service area. Timedivision duplex (TDD) protocol in which the uplink pilot
symbols are used to estimate both uplink and downlink is
proposed to cope with the pilot overhead for mMIMO arrays.
In TDD protocol [1][4], the transmission time over each
coherence time is divided into four non-overlapping phases:
Channel training to learn the channel between the users and the
base station, each user transmits some known pilot symbols;
Uplink Data: the uplink data of all users are transmitted
over the same timefrequency resources in a non-orthogonal
fashion such that the base station receives a superposition of
all transmitted symbols; Processing Time: the time which is
needed to perform the channel estimation and precode the
users data for the downlink transmission; and Downlink Data:
finally the downlink data of all users are precoded using
the estimated channels and transmitted over the same time
frequency resources in a non-orthogonal fashion. The duration
of uplink and downlink transmissions may vary and can be
adjusted based on the amount of the users data and the traffic
demands.
In this paper, we investigate the communication of K singleantenna users to a common receiver with a massive antenna
array over an uplink shared channel. To estimate the channel
between two antenna ports, the transmitting node sends pilot
symbols which are known at the receive node (the time
frequency location and the associated value are generally

preset). The pilot symbols from each user should have the
periodicity of N = Tc Bc symbols in order to track the channel
variation over the time and frequency where Tc and Bc denote
the coherence time and bandwidth, respectively. It is desirable
to coordinate as many users as possible for uplink transmission
such that the receiver can perform spatial-division multipleaccess (SDMA); i.e. the receiver is able to obtain interferencefree signals associated to each user via spatial filtering (e.g.
projection of the received signals in the space that is spanned
by the associated channels from the users to the antenna
array). In mMIMO, the receiver can asymptotically separate
the uplink data of different users without inter-user interference if the spatial channels are for example independent [1].
Hence ignoring the processing time, the maximum number
of the coordinated users, under successful decoding, can be
obtained by maximizing the total number of transmitted data
symbols that are separable at the receiver, which is given by
K(Bc Tc K), where the maximum is taken over the active
number of users, K, for a given coherence interval spanning N
symbols. Under assumption that all data symbols contain equal
information; i.e. all users employ the same modulation order,
this concludes that the coherence interval should be equally
divided between the channel training and data transmission
as considered in [1]. So the optimal number of active users
operating
over
the same timefrequency resources should be
set to 12 N and the total number

of datasymbols that can be
transmitted hence becomes 12 N 12 N symbols.
In [5], it is shown that the above number of users and the
associated aggregate rate can be improved by the so-called a
semi-orthogonal feature. In this paper, we further extend this
concept and present a generalized one which can bridge both
SOMA and the conventional TDD [1] and provide an enhanced
aggregate throughput.
The remainder of the paper is organized as follows.
Section II briefly reviews semi-orthogonal multiple-access
(SOMA) described in [5]. Section III generalizes SOMA
scheme to enhance its potential performance. Section IV
discusses two realizations of GSOMA and presents an iterative
receiver. This section also analyzes the sum-rate for two
modes of GSOMA using matched filtering (MF) and zeroforcing (ZF) receiver. Section V presents numerical results
and Section VI finally concludes the paper.
978-1-4799-8088-8/15/$31.00 2015 IEEE
Coherence Interval
User 1
...
...
User 2
...
...
Pilot symbol
Baseline UL data symbol

User K
...
...
NEW UL data symbol
User K+1
User K+2
...
...
...
...
NEW UL data symbol

of NEW users
Blank symbol
...
User N-1
Channel Training
UL Data
Fig. 1. Illustration of Semi-Orthogonal Multiple-Access (SOMA) transmission.
II. SOMA
We next briefly discuss the semi-orthogonal multiple-access
in [5] which is designed to address the uplink capacity shortage
problem in mMIMO. SOMA coordinates up to K N 1
users, where N = Tc Bc is the number of resources in time and
frequency over which the channel is approximately constant,
i.e. the coherence interval. Fig. 1 depicts an example of how
the uplink transmission is configured in which K = N 1
users are scheduled. Over each coherence block, the user
1 j N 1, transmits one pilot symbol and N j data
symbols such that the pilot symbols of users are transmitted over orthogonal timefrequency resources, and the data
symbols of user j, for all 1 j N 1, consume all the
timefrequency resources of users j + 1 k N 1, which
are used for both pilot and data transmission. The received
signals at the access node in a coherence window is illustrated
in Fig. 1. In a given time-slot, some of the users are silent
and appears orthogonal while some other users transmit nonorthogonally. For example, in time slot one, only user one
transmits its pilot and all other users appear orthogonal and
in the second time slot the first user transmits its data symbol
and second user transmits its pilot symbol and so forth; cf.
semi-orthogonal feature. SOMA when used with an access
node with a very high number of antennas can schedule K
users where user k can transmit N k asymptotically errorfree symbols. Therefore, for very large antenna arrays, SOMA
solution nearly doubles the throughput as compared to the
baseline TDD with optimal number of users. The receiver is
constructed using a sequential channel estimation and data
detection enabled by the designed semi-orthogonal feature
embedded in the transmitted signals.
III. G ENERALIZED SOMA
We next present the generalized SOMA (GSOMA). Fig. 2
depicts multi-user GSOMA transmission in which the users
are grouped into J groups where each group contains kj
PJ
users for j {1, 2, . . . , J} and K =
j=1 kj is the total
number of the users. The user i in group j uses the pilot

sequences si for all j {1, 2, . . . , J}. That is the pilot
sequences are reused. The pilot sequences within each group
are orthogonal such that it allows interference-free channel
estimation for the users in each group. The maximum number
of pilot sequences therefore is maxj kj . The conventional
reuse of pilot sequences where the pilot symbols interference
with one another results to pilot contaminations which
severely degrades the performance of the users. However with
this new solution, it is allowed to reuse the pilot sequences
in a controlled fashion. A pilot reuse is performed semiorthogonally to boost the spectral efficiency of the system.
However, the interference is controlled by a transmission of
a resource-offset (e.g. timing-offset or frequency-offset) such
that the received packets at the access node have the following
structure
The pilot signals of different groups are received nonorthogonally (for example non-overlapping time slots
with timing-offset)
The pilot sequences of the group j {1, 2, . . . , J}, are
only interfered by data symbols of users in the groups
1 to j 1. The pilot sequences of the first group are
received interference-free. That is the other users appear
silent at the receiver side.
With GSOMA, the user may use partial blanking which has
the same granularity as the length of pilot region to control
to the inter-group interference. GSOMA includes both SOMA
scheme described in [5] and the conventional TDD solution
proposed by Marzetta in [1] as special cases. When each group
contains one user and no blanking is used, GSOMA reduces to
SOMA. When there is only one group with maximum number
of users, then GSOMA reduces to the conventional TDD
wherein only orthogonal pilot sequences are used. Therefore,
a properly designed GSOMA can combine the advantages
of both SOMA and conventional TDD. The advantage of
GSOMA with respect to the conventional TDD is that it schedules more groups, which enhances the aggregate rate. Since
Resource-Offset Sequence
Access Node
Blanking Pattern
Channel Estimation
Users
Uplink data and pilot signals over

shared channels
DMUX
Detection
Pilot Sequence Index
Controller
Spatial Filtering
Group 1: users with orthogonal pilot sequences
Data
Data
Data
Pilot 11
Pilot
..
.
r1
Data
Data
Pilot 1
Group 2
Data
Data
Data
Pilot 11
Pilot
Blank Pilot
1
...
Pilot11
Pilot
Pilot
1
Pilot 11
Pilot
Pilot
1
Data
Data
Data
Pilot 11 Blank
Pilot
Data
Group J-1
...
Data
Data
Data
...
Users with pilot reuse

(i.e. nonorthogonal
pilot sequences)
Pilot 1
1
Pilot
Pilot
1
Group J
Data
Data
Blank Pilot 1
Data
3
2
L 1
L
L 1
Resource (Time/Frequency)
Resource-offset between group1 and group 2
Fig. 2. Generalized Semi-Orthogonal Multiple-Access (GSOMA) transmission.
within each group the pilot sequences are mutually orthogonal

as that in the conventional TDD, it allows performing joint
channel estimation without interference for all users in each
group and employing joint spatial filtering such as zero-forcing
per group to suppress the inter-user interference among the
users in each group. This also increases the aggregate rate.
IV. T WO -G ROUP GSOMA
In this section, we tailor the design of GSOMA to the cases
with two groups and discuss two modes of transmission.
A. Mode 1: without blanking
Fig. 3 shows the transmission protocol and the receiver for
the case with two groups where there is no blanking. The
users are grouped into two groups such that each group is
designed according the conventional TDD wherein the half of
the coherence time, i.e. 21 Tc , is used for pilot transmission and
the other half is consumed for the data transmission. To enable
the channel estimation without interference, a time-offset equal
to 12 Tc is used. The receiver first estimates the channel of the
users in the first group using the knowledge of the orthogonal
pilot sequences and then performs joint spatial filtering to
decode the data of the first group. The decoded data are fed
back to the channel estimator of the second group to perform
interference cancellation prior to the channel estimation. After
the interfere cancellation, the channels of the second group are
estimated using the known orthogonal pilot sequences used for
the users in the second group. The decoded data of the second
group is used to cancel the interference over the pilot symbols
of first group. This sequential channel estimation and decoding

is iteratively continued until all data are successfully decoded.
We next present the single-cell uplink throughput for the
protocol in Fig. 4 with L consecutive sub-frames when the
receiver has nt antennas and the channels are i.i.d. Rayleigh
fading with unit variance. We first consider matched-filtering
(MF) and then zero-forcing (ZF). The proofs, which are
omitted due the space limitation, can be obtained by similar
approaches as those in [5].
Proposition 1: The transmission protocol in Fig. 3 where
each group contains K users and the receiver employs MF
achieves the sum-rate
L
MF
RP
,1 =
K X
(R1,l + R2,l )
2L + 1
(1)
l=1
where

(nt 1)(1 Ne1,l )Pd1,l
R1,l = log 1 +
N0 + Ne1,l Pd1,l + Pp2,l + (K 1)Pd1,l

(nt 1)(1 Ne2,l )Pd2,l
R2,l = log 1 +
N0 + Ne2,l Pd2,l + Pp1,l+1 + (K 1)Pd2,l
N0 + KNe2,l1 Pd2,l1
Ne1,l =
N0 + Pp1,l + KNe2,l1 Pd2,l1
N0 + KNe1,l Pd1,l
Ne2,l =
N0 + Pp2,l + KNe1,l Pd2,l
and Ppj,l , Pdj,l denote the average power consumed for the
pilot and data symbols of the users in group j {1, 2} and
sub-frame l {1, 2, . . . , L}, and Nej,l and N0 respectively de-
Group 1 with BcTc/2 users with orthogonal pilot sequences

Pilot 1
Pilot11
Pilot
Data
Data
Data
Pilot 1
Pilot11
Pilot
Data
Data
Data
Pilot 1
Pilot11
Pilot
Data
Data
Data
Pilot 1
Pilot11
Pilot
Tc
Timing offset
Group 2
Data
Data
Data
Tc/2
a) Transmission Protocol
Spatial Filtering, Data

Detection or Decoding
Channel Estimation
Group 2
Spatial Filtering, Data

Detection or Decoding
DMUX
Channel Estimation
Group 1
b) Receiver
Fig. 3. Two-Group GSOMA without blanking.
note variance of AWGN at the receiver and channel estimation

error, and Ne1,0 = Pd1,0 = Pp1,L+1 = 0.
each group contains K users and the receiver employs ZF

L
(nt K)(1 Ne1,l )Pd1,l
K X
ZF
P
log 1 +
R ,1 =
2L + 1
N0 + KNe1,l Pd1,l + Pp2,l
l=1

(nt K)(1 Ne2,l )Pd2,l
+ log 1 +
(2)
N0 + KNe2,l Pd2,l + Pp1,l+1
where Nej,l is defined in Prop. 1.
B. Mode 2: with blanking
For the case the array contains not so many antenna
elements with respect to the scheduled number of users in
each group, it is beneficial to partially blank some part of
sub-frames to enhance the channel estimation and to improve
consequently the performance of spatial filtering which in turn
improves the spectral efficiency of the system. The blanking
pattern, should be chosen based on the inter-user interference
when for higher interference a blanking pattern with a higher
density is selected and similarly a sparser blanking pattern
should be selected for lower interference to improve the
spectral efficiency of the system. Fig. 4 shows the transmission
protocol and the receiver for the case with two groups where
the blanking is used. The users are grouped into two groups
and transmission is arranged such that the pilot symbols of first
group do not experience any interference. This is useful for the
case that one to enhance the channel estimation for the users in
the first group, this also enhances the interference cancelation
for channel estimation of the second group. Additionally, this
arrangement is also useful for the case where first and second
groups compromise far and near users, respectively.
each group contains K users and the receiver employs MF

(nt 1)(1 Ne1 )Pd1
1
MF
P
R ,2 = K log 1 +
3
N0 + Ne1 Pd1 + Pp2 + (K 1)Pd1

(nt 1)(1 Ne2 )Pd2
1
(3)
+ K log 1 +
3
N0 + Ne2 Pd2 + (K 1)Pd2
where Ppj , Pdj denote the average power consumed for the
pilot and data symbols of the users in group j {1, 2}, and
N0 denotes variance of AWGN at the receiver. The quantity
Nej = Nej,1 denotes the variance of channel estimation error
for users in group j {1, 2}.
each group contains K users and the receiver employs ZF

(nt K)(1 Ne1 )Pd1
1
ZF
K
log
1
+
RP
=
,2
3
N0 + KNe1 Pd1 + Pp2

1
(nt K)(1 Ne2 )Pd2
+ K log 1 +
(4)
3
N0 + KNe2 Pd2
where Nej = Nej,1 for j {1, 2} is given in Prop. 3.
From the sum-rates in Propositions 14, one can see that
the imperfections due to inter-user interference, the channel
estimation error and AWGN are linearly reduced by the
1
for MF and ZF, respectively. We
factor of nt11 and nt K
additionally see that ZF removes the inter-user interference
in each group due the data transmission which is one of the
advantages of GSOMA with respect to SOMA.
V. N UMERICAL E VALUATIONS
For comparison, we consider time-shared TDD as a baseline
when the resources are shared between two groups of users
when each group is designed according to the conventional
TDD. This, using ZF, gives the sum-rate

1
(nt K)(1 Ne1 )Pd1
ZF
RP
=
K
log
1
+
,TS
4
N0 + KNe1 Pd1

1
(nt K)(1 Ne2 )Pd2
+ K log 1 +
(5)
4
N0 + KNe2 Pd2
Group 1 with BcTc/2 users with orthogonal pilot sequences

Pilot 1
Pilot11
Pilot
Timing offset
Partially Blanked
Data
Data
Data
Pilot 1
Blank
Pilot 1
Pilot11
Pilot
Pilot 1
Pilot11
Pilot
Data
Data
Data
Pilot 1
Blank
Group 2
Pilot 1
Pilot11
Pilot
Tc/2
Tc
Fig. 4. Two-Group GSOMA with blanking.
180
160
140
Sumrate [bit/s/Hz]
0
where Nej = N0N
+Ppj .
We next discuss two numerical examples of the sum-rate.
Fig. 5 and 6 show the sum-rate of the schemes as a function of
number of antennas for N = Bc Tc = 100, where the number
of users in each group is K = 12 N = 50. We set Pd1,l = Pd1 ,
Pd2,l = Pd2 (i.e. uniform power allocation across the subframes) and N0 = 0 dB and the power of the associated
pilots are set 10 dB higher than the data for all users to
ensure a good channel estimation. In each figure, four schemes
are considered: time-shared TDD with ZF, SOMA with MF,
GSOMA with transmission protocol in Fig. 3 (i.e. Mode 1)
with ZF and L = 100, GSOMA with transmission protocol
in Fig. 4 with ZF (i.e. Mode 2). In Fig. 5 we set Pd1 = 0,
Pd2 = 20 dB and in Fig. 6 we set Pd1 = 10, Pd2 = 5
dB. In both cases, GSOMA provides an enhanced aggregate
rate as compared to the time-shared TDD. The gain is more
pronounced for the case that the groups are higher difference
in the received signal strength for which Mode 1 performs
better. Fig. 6, Mode 2 performs better than Mode 1 due to the
fact that the channel estimation for the high-power users in
group 1 is less degraded as compared to that in Mode 1.
120
100
80
60
Timeshared TDD
SOMA
GSOMA: Mode 1
GSOMA: Mode 2
40
20
0
200
400
600
Number of antennas: nt
800
1000
Fig. 5. The sum-rate of the schemes for Pd1 = 0, Pd2 = 20 dB.
VI. C ONCLUSIONS
We presented a new multiple-access solution and analyzed
its aggregate rate. The new solution is constructed using a
semi-orthogonal feature for a group of users wherein each
group employs the conventional TDD. The numerical evaluation showed that the new proposal can provide a higher
aggregate rate as compared to the conventional TDD solution.
[1] T. L. Marzetta, Noncooperative cellular wireless with unlimited numbers

of base station antennas, IEEE Transactions on Wireless Communications, vol. 9, no. 11, pp. 35903600, 2010.
[2] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, Massive
MIMO for next generation wireless systems, IEEE Communications
Magazine, pp. 186195, 2014.
[3] F. Rusek, D. Persson, B. K. 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.
4060, 2013.
[4] J. Hoydis, S. Ten Brink, and M. Debbah, Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need? IEEE Journal
on Selected Areas in Communications, vol. 31, no. 2, pp. 160171, 2013.
[5] M. N. Khormuji and B. M. Popovic, Semi-orthogonal multiple-access for
massive MIMO, Submitted to IEEE Trans on Wireless Communications,
2014.
Sumrate [bit/s/Hz]
R EFERENCES
200
150
100
Timeshared TDD
SOMA
GSOMA: Mode 1
GSOMA: Mode 2
50
200
400
600
Number of antennas: nt
800
1000
Fig. 6. The sum-rate of the schemes for Pd1 = 10, Pd2 = 5 dB.
A Distributed Massive MIMO Testbed to assess

Real-World Performance & Feasibility
Paul Harris, Siming Zhang, Andrew Nix, Mark Beach, Simon Armour and Angela Doufexi
Communication Systems & Networks Group, University of Bristol, Bristol, BS8 1UB, UK
Email: {paul.harris, sz1659, andy.nix, m.a.beach, simon.armour, a.doufexi}@bristol.ac.uk
Abstract This paper introduces a novel 128-element sub6 GHz Massive MIMO (Multiple-Input Multiple-Output)
testbed being developed by the University of Bristol in
collaboration with National Instruments and Lund University
as part of the BIO (Bristol is Open) city testbed. Distributed
Massive MIMO systems avoid large single-piece antenna facets
and their associated windage problems and also offer greater
spatial diversity, but node synchronization and phase
alignment becomes more challenging. Simulated network
performance of a co-located 128-element 3D Multi-User MIMO
system at a possible test site in Bristol are presented based on
our 3D channel propagation model.
Index Terms Massive MIMO, Distributed, Testbed, 5G.
I. INTRODUCTION
One possible constraint that may inhibit the realisation
of massive MIMO (multiple-input, multiple-output) BS
(Base Station) deployments operating in the sub-6 GHz
bands is the mounting of large antenna arrays, as the form
factor of these will need to adhere to strict planning
regulations and withstand strong wind conditions when
placed on a mast. A D-M-MIMO (distributed massive
MIMO) system would allow the BS antennas to be broken
down into more easily deployable segments, improving
spatial diversity and providing greater flexibility for finding
the optimal antenna configuration at a given site [1].
This paper outlines a state-of-the-art D-M-MIMO
testbed that is being developed at the UoB (University of
Bristol) through collaboration with NI (National
Instruments) and Lund University as part of the BIO (Bristol
is Open) city testbed. The system has been designed to allow
for both distributed and collocated deployments enabling
researchers to tackle both avenues. As a key part of Bristols
Programmable City Testbed [2], it is posed to provide a
highly flexible platform for tackling pragmatic massive
MIMO research that will ultimately help better establish the
feasibility and performance of the technology for use in
future 5G systems.
II. SYSTEM OVERVIEW
The Bristol D-M-MIMO system is an evolution of a
pioneering testbed recently developed at Lund University for
conducting real-world tests in the surrounding campus. It
will operate with 128 antenna elements, a bandwidth of 20
MHz, TDD (time division duplexing) and assume reciprocal
channels for precoding. The hardware used will permit
operating frequencies between 1.2 GHz and 6 GHz. As data
on the Lund system is already available in [3], this overview
will focus on the differences in the new D-M-MIMO

implementation at Bristol.
A. Hardware & Software Architecture
Leveraging the modular and hierarchical design used for
the Lund testbed, the Bristol D-M-MIMO system focuses on
allowing the NI PXI (PCI Extensions for Instrumentation)
chassis to be easily unshackled. The 4 slave PXI chassis
shown in [3] will be physically separable into nodes, each
with 32 antenna elements connected to 16 NI USRPs
(Universal Software-defined Radio Peripherals). These will
then communicate back to the master PXI chassis using
either short-range PCIe (Peripheral Component Interconnect
Express) links for collocated deployments or a pre-existing
fibre-optic network for distributed deployments. UE (User
Equipment) nodes will be provided by individual USRPs
connected to laptops, and as each USRP has two radios, each
node can either act as two single-antenna users or one user
with two spatial streams.
The NI graphical programming language, LabVIEW,
will be used to configure the operation of the FPGAs (Fieldprogrammable gate arrays) on-board both the USRPs and
PXI chassis, as well as the higher level processing on the
master PXI chassis controller. The LabVIEW FPGA
paradigm allows for rapid prototyping and abstracts the
underlying semantic complexities of FPGA development.
B. Antenna Array & Switchable Polarisation
Cross-polarised patch antenna elements are likely to be
used for the antenna array, but the form factor will be
segmented, allowing for 4 pieces to each break away with
their respective slave PXI chassis. In order to improve the
flexibility of the system for different tests and environments,
it was deemed beneficial to implement a simple way for
switching the polarisation of each array element and avoid
rewiring. By utilising the digital interface on the front of
each USRP and RF relays, reconfiguration of the entire array
will be possible through software at the click of a button.
III. SIMULATION RESULTS AND FUTURE WORK
In line with our testbed, theoretical network simulations
were performed as a baseline for the achievable cell
capacity. Our simulation evaluates the system performance
of massive 3D multi-user MIMO for realistic urban
deployments, i.e. Macro and Pico cells, and our current
configuration assumes a collocated 128-element BS antenna
array simultaneously serving 16 single-antenna UEs; D-MMIMO configurations will be implemented in the near
978-1-4799-8088-8/15/$31.00 2015 IEEE
future. A 3D laser-scanned environmental database (in this

study a database for the city of Bristol (UK)) was used as
input to UoBs 3D ray-tracing channel model along with
measured 3D antenna patterns for the individual BS and UE
antenna elements. The demonstrated BS sites (one Macrocell
and one Picocell) are chosen on the map to accurately match
our future test sites.
A. Antenna Characteristics and Configuration
Our BS is configured as a half-wavelength spaced
64(column)-by-2(row) antenna array, which could be easily
changed to any desired antenna design. For the Macrocell,
the BS antenna elements are cross-polarised patch antennas
with each polarisation counting as a logical antenna; for the
Picocell, cross-polarised dipole antennas are used. Each
element is measured in our anechoic chamber, and the 3D
complex voltage and polarimetric pattern is applied as a
spatial-polarisation-phase convolution process on the pure
3D channel model.
B. Channel Model and Ray-Tracing
The channel propagation study was performed using the
UoBs outdoor 3D ray-tracer, which was used to generate
many of the statistics now specified in the 3D extension of
the 3GPP/ITU channel model [4-6]. Table I summarises the
ray-tracing parameters for the chosen Macro and Pico cells.
The city database comprises buildings, foliage and terrain
layers with 10m resolution. Users were randomly dropped
within each cell.
C. Network Simulator and Parameters
In order to quantify the data throughput performance, a
MU-MIMO Downlink simulator was implemented in
MATLAB. Here we assumed the network operating
frequency was 2.6 GHz with a bandwidth of 20 MHz and
1200 data subcarriers. Closed-loop Eigen-Beamforming was
used as the channel precoding method. UE scheduling was
performed in a random manor from LOS (line-of-sight) and
non-LOS locations, and results were averaged over 100
iterations to maintain statistical relevance.
TABLE I.
SUMMARY OF RAY-TRACING PARAMETERS

Macro cells
Environment
Pico cells
To perform system level analysis in a computationally

efficient and scalable manner, a physical layer abstraction
technique, RBIR (Received Bit level mutual Information
Rate), was used to predict the packet error rate for 1000
channel realisations across the allocated OFDM (Orthogonal
Frequency Division Multiplexing) subcarriers. Due to the
intensive signal processing required in M-MIMO systems,
all simulations were executed on UoBs High Performance
Computing platform which comprises 5,500 Intel Xeon
cores; this has increased results turn-over more than 100
times.
D. Results and Future Work
The achievable throughput is strongly connected to the
channel richness, UE scheduling and the antenna
configurations at both the transmitter and receiver. Our
preliminary results show that the median data rate
approached 655Mbps in the Picocell and 378Mbps in the
Macrocell, notwithstanding the peak rate of just over
1.2Gbps. To better exploit large degrees of freedom and
provide enhanced array and diversity gain, future work will
use a more advanced UE scheduling algorithm, improved
channel precoding and BS antenna design, as well as the DM-MIMO BS configuration.
IV. CONCLUSIONS
This paper has presented a D-M-MIMO testbed that is
under development at the UoB with project partners. Coupled
with the UoBs capability for performing realistic city-wide
system evaluations with a state-of-the-art 3D channel model,
it will provide an unprecedented opportunity to tackle the
pragmatic research problems associated with both M-MIMO
and its distributed form factor.
ACKNOWLEDGMENT
The authors would like to acknowledge the technical
inputs of both Lund University (Sweden) and NI. This work
was supported by the Engineering and Physical Sciences
Research Council [EP/I028153/1] and the Centre for
Doctoral Training in Communications at the University of
Bristol.
17.6km2 area of central Bristol (UK)
REFERENCES
BS mounting
3m above rooftop level
On lamp-posts
[1]
Number of
BSs and UEs
User
locations
One 3-sector BS
300 UEs per sector
50-1000 m from BS
1.5m above ground level
One BS
with 150 users
5-150 m from BS
1.5m above ground level
[2]
2.6 GHz
Frequency
[3]
BS Tx power
43 dBm
30 dBm
BS height
42 m above ground level
5 m above ground level
Antennas
Isotropic at both ends of the link
Minimum
Rx
sensitivity
-120 dBm
(only links with two or more traced rays were
considered)
[4]
[5]
[6]
Madhow, U.; Brown, D.R.; Dasgupta, S.; Mudumbai, R., "Distributed

massive MIMO: Algorithms, architectures and concept systems,"
Information Theory and Applications Workshop (ITA), 2014 , vol.,
no., pp.1,7, 9-14 Feb. 2014
Bristol
is
Open
(n.d.)
[Online].
Available:
http://www.bristolisopen.com/
E. Luther, 5G Massive MIMO Testbed: From Theory to Reality,
National
Instruments,
Austin,
TX,
USA.
Available:
http://www.ni.com/white-paper/52382/en//. Accessed: Dec. 14, 2014.
T. Thomas, F. W. Vook, E. Visotsky et al., 3D extension of the
3GPP/ITU channel model, in Proceedings of the 77th IEEEVehicular
TechnologiesConference (VTC-Spring), Dresden, Germany, May
2013.
Text Proposal R1-130497, 3D ChannelModeling Issues and 3D
Channel Model Proposal, 3GPP TSG-RANWG1.
Text Proposal R1-130500, Detailed 3D Channel Model, 3GPP TSGRANWG1.
Statistical Analysis of the Channel Capacity Outage

Intervals in Massive MIMO Systems with OSTBC
over Rayleigh Fading Channels
Rym Hicheri1 , Nazih Hajri1 , Neji Youssef1 , Matthias Patzold2 , and Tsutomu Kawabata3
1 Universit
e
de Carthage, Ecole Superieure des Communications de Tunis, 2083 EL Ghazala, Ariana, Tunisia
of Engineering and Science, University of Agder, NO-4898 Grimstad, Norway
3 Dept. of Commun. Eng. and Inf., University of Electro-Communications, 1-5-1 Chofu-shi, Tokyo 182, Japan
2 Faculty
E-mail:
rym.hicheri@supcom.tn,
{nazih.hajri, neji.youssef}@supcom.rnu.tn,
3 kawabata@ice.uec.ac.jp
AbstractThis paper studies approximate solutions for the

statistical properties of the outage intervals of the instantaneous
capacity in massive multiple-input multiple-output (MIMO) systems with orthogonal space-time block code (OSTBC) over
Rayleigh fading channels. We take advantage from the fact that
the probability density function (PDF) of the channel power gain
can be approximated by a left-truncated Gaussian distribution if
the number of transmit and receive antennas is large. Assuming
a symmetrical Doppler power spectral density (PSD), a closedform expression is presented for the Rice probability function
of the outage durations. This function, in general, approximates
the PDF of the outage intervals only over its initial part. By
applying the asymptotic behavior of level crossings of Gaussian
processes, an alternative expression, valid only for low capacity
thresholds, is obtained, which, in turn, is applied to determine an
approximate solution for the kth-order moments of the outage
intervals. Also presented is the PDF of the so-called instantaneous
capacity loss with respect to the average capacity. Simulation
results have been shown to be in good agreement with the
theoretical approximations, thereby confirming their validity.
Although, the focus of the study is on massive MIMO systems,
it has been observed that the approximations are accurate even
for a moderate number of antennas such as the case of 44
MIMO-OSTBC systems.
Index TermsMassive MIMO-OSTBC systems, capacity

processes, Rice probability function, outage durations, crossing
theory.
I.
I NTRODUCTION
It is well known that the second-order statistics, i.e., the

autocorrelation function (ACF), the level-crossing rate (LCR),
and the average duration of fades (ADF), play an important
role in the characterization and analysis of the capacity of
MIMO time-varying channels and provide additional insights
into its temporal behavior [1], [2]. For this reason, the investigation of the above quantities has been the subject of several
studies. For example, the ADF, the LCR, and the ADF of the
instantaneous capacity of MIMO Rayleigh channels have been
studied in [3] and [4] for the cases of symmetrical and asymmetrical Doppler PSDs, respectively. Moreover, the authors of
[5] investigated the impact of the spatial correlation on the
statistics of the capacity of MIMO Nakagami-m channels. In
the past, the study of the outage events has been mostly devoted
to the analysis of the LCR and ADF.
matthias.paetzold@uia.no,
Besides the above statistical quantities, the distribution

of the outage intervals is another important statistics, the
knowledge of which allows us to gain additional insight into
the dynamic behavior of the capacity of MIMO channels.
Owing to the difficulty of the level-crossing problem of random
processes [6], very few studies have been reported on the PDF
of the outage intervals of the instantaneous capacity of MIMO
systems. The derivation of the Rice probability function has
been reported in [7] for the case of single-input single-output
(SISO) Rayleigh fading channels based on the results presented
in [6]. The asymptotic analysis of the outage intervals in
MIMO-OSTBC Rayleigh fading channels has been conducted
in [8]. The drawback of the solution reported in [8] is its
limitation to very low outage probabilities.
In this paper, we address the statistics of the capacity
in massive MIMO-OSTBC systems and investigate approximate solutions for the PDF of the outage intervals for the
transmission over Rayleigh fading channels. More precisely,
we consider a large number of transmit antennas and receive
antennas which allow us to approximate the channel power
gain by a left-truncated Gaussian process [9]. The accuracy of
this approximation has been studied for a limited number of
antenna elements in [10, Ch. 2]. Based on the results regarding
the Rice probability function for Gaussian processes, we derive
the Rice probability function for the instantaneous channel
capacity. This function provides an accurate approximation for
the initial behavior of the PDF of the outage durations for any
outage threshold of capacity processes. Furthermore, drawing
upon known results on the asymptotic behavior of Gaussian
processes, we investigate approximate solutions for the PDF
and kth-order moments of the outage durations as well as the
distribution of the instantaneous capacity loss. The validity
of the derived expressions is verified by means of computer
simulations. The approximate solutions have been shown to be
accurate even for MIMO-OSTBC systems with a small number
of transmit/receive antennas.
This paper is organized as follows. Section II contains some
preliminaries and useful background material. Approximate
solutions for the PDF of the capacity outage intervals based
on the Rice probability function are investigated in Section III.
The asymptotic analysis of the outage intervals is the topic
of Section IV. Numerical examples and simulation results are
provided in Section V. Finally, Section VI concludes the paper.
978-1-4799-8088-8/15/$31.00 2015 IEEE
C(t) (bits/s/Hz)
cm
C
cl
Time, t (s)
Fig. 1: A simulated sample function of the instantaneous

capacity C(t) for a 4x4 MIMO-OSTBC Rayleigh channel.
II.
S YSTEM M ODEL AND P ROBLEM S TATEMENT
We consider a downlink MIMO-OSTBC system with NT

transmit antennas and NR receive antennas. If the channel
state information (CSI) is available at the receiver side, i.e.,
equal transmit power allocation, then the corresponding instantaneous capacity C(t) can be expressed as [11]

C(t) = log2 1 +
(t)
(1)
NT
where stands for the average signal-to-noise ratio (SNR)
and (t) is the instantaneous channel power gain. Since the
transmission is assumed to be over Rayleigh fading environments, (t) is a chi-square process with 2N
T NR degrees of
2N
R NT
freedom, which can be written as (t) = m=1
|m (t)|2 .
The processes m (t), m = 1, ..., 2NT NR , are given by
(1)
(2)
(k)
m (t) = m (t) + jm (t), where m (t), k = 1, 2, is a
zero-mean, real-valued, stationary Gaussian process with a
(k)
variance 02 . The normalized ACF of m (t) is denoted by
r(t). The mean and variance of the process (t) are given
by m = 2NT NR 02 and 2 = 4NT NR 04 , respectively. The
normalized ACF of (t) is r (t) = r2 (t).
Fig. 1 depicts a waveform example of the capacity process
C(t). Referring to this figure, our objective in this work is
to provide approximate solutions for the PDF of the duration
of the outage intervals , i.e., the time intervals during which
the capacity process C(t) is below a given threshold cl . We
consider the case of large antenna systems also termed as
massive MIMO systems. By invoking the central limit theorem,
the PDF of the process (t) can be approximated by a lefttruncated Gaussian distribution [10, Ch. 2] if 2NT NR .
Thus, the capacity process C(t) can be expressed as

C(t) = log2 1 +
( I(t) + m )
(2)
NT
where I(t) is a zero-mean unit variance Gaussian process.
Under the above conditions, we investigate two approximate
solutions to the distribution of the outage intervals of
the instantaneous channel capacity C(t). The first solution
is based on the derivation of the so-called Rice probability
function, while the second one is determined by undertaking
an asymptotic behavior analysis. The Rice function will be
denoted by QC,+ ( ; cl ), while p ( ; cl ) will stand for the
asymptotic PDF. Also, p ( ; cl ) will denote the true PDF.
III.
R ICE P ROBABILITY F UNCTION
In this section, we investigate the Rice probability function

of the instantaneous channel capacity C(t). This function, de-
noted by QC,+ ( ; cl ), represents the conditional probability

that the capacity process C(t) crosses upward a predetermined
threshold cl in the interval (t+, t+ +d ), given a downward
crossing of cl at the time instant t. According to [6, Sect. V],
the function QC,+ ( ; cl ) can be obtained as
+
0
c1 c2 pc1 c1 c2 c2 (cl , c1 , cl , c2 )
dc1 dc2 . (3)
QC,+ ( ; cl ) =
NC (cl )
0

dc denotes the
Here, the quantity NC (c) = 0 c pcc (c, c)
average LCR of the capacity process C(t) [12, Sect. 3.3.2], and
pc1 c1 c2 c2 (c1 , c1 , c2 , c2 ) refers to the joint PDF of c1 = C(t),
+ ),
c2 = C(t + ), and c2 = C(t
c1 = C(t),
where the dots stand for the first time derivative. In (3), the
quantities c1 and c2 are set to cl . Drawing upon the fact that
the channel capacity C(t) increases monotonically with the
Gaussian process I(t), it is shown in the Appendix that the
Rice probability function QC,+ ( ; cl ) of C(t) can then be
expressed as

(2cl 1)NT m
QC,+ ( ; cl ) = QI,+ ;
(4)

where QI,+ ( ; Il ) is the Rice probability function of the
Gaussian process I(t) [6, Sect. V] and Il = [(2cl 1)NT
m ]/[ ] denotes the threshold level. Finally, by using [6,
Eq. (47)] and (4) and letting r ( ) = r2 ( ), we obtain
3/2
M
QC,+ ( ; cl ) = 1 r( )4

2 2
cl
(2 1)NT m
r ( ) 1
J(s, m). (5)
exp

2(r2 ( ) + 1)
The quantities , M , J(s, m), s, and m are given by [6]

= 2 r 2 (0) + r(0)
r(0)
(6)
(7)
M = (1 r4 ( )) 4r 2 ( )r2 ( )
+ +
1
J(s, m) =
(x m)(y m)
2 1 s2 m
m

x2 + y 2 2sxy
exp
dxdy
(8)
2(1 s2 )

2
r(0) 1 r4 ( ) + 4r4 ( )r 2 ( )
r (0) + r(0)
(9)
s=
M
and

2r(
)r( ) ((2cl 1)NT m ) 1 r4 ( )
m =
(10)
(1 + r2 ( ))
M
respectively.
For the special case of large values of , i.e, r2 ( ) 0,
it can easily be shown that the Rice probability function
QC,+ ( ; cl ) tends to the LCR NC (cl ). In fact, for this
limiting case, the joint PDF pc1 c1 c2 c2 (cl , c1 , cl , c2 ) can be
expressed as the product of the joint PDFs pc1 c1 (cl , c1 ) and
pc2 c2 (cl , c2 ). This result, when substituted in (3), leads to the
expression QC,+ ( ; cl ) = NC (cl ).
The study of systems operating in low and high SNR
regimes, i.e., power limited and bandwidth limited systems,
is also of great importance. In the case of energy-limited
scenarios, for a given outage probability, the key performance
measure is the maximum number of bits per unit energy that
can be communicated. In fact, if 0, then long codes are

needed to provide sufficient redundancy for the recovery of
the desired signal. In this case, the channel capacity can be
approximated by C(t) (t)/[NT log(2)]. The corresponding Rice probability function can be obtained by considering
the mapping cl Il [NT cl log(2) m ]/[ ] in (4).
In the low SNR regime, the energy efficiency is the most
important measure, but at high SNR values, the main challenge
is to use the degrees of freedom efficiently. In the high SNR
regime ( , i.e., 2cl 1), the channel capacity can
be reduced to C(t) log2 ((t)/NT ). Consequently, the
function QC,+ ( ; cl ) can be computed by considering the
mapping cl Il [NT 2cl m ]/[ ] in (4).
Notice that the evaluation of (5) requires the numerical
calculation of two semi-infinite integrals. Also, due to its
complexity, (5) gives us no insight into the moments of the
outage intervals. For the important case of very low thresholds,
however, a simple approximate solution for the distribution of
the outage duration can be derived. This will be the topic of
the next section.
IV.
A SYMPTOTIC B EHAVIOR BASED S OLUTION
Here, we confine our attention to the asymptotic analysis of

the PDF of the outage intervals under low thresholds cl . This
regime is of particular interest in systems operating under a
low error probability. At low levels Il , the Gaussian process
I(t) behaves asymptotically as [13]

1 2
t 1 + t
(11)
I(t) Il
2
where is a Rayleigh random variable with a parameter
= 1. Using Il = [(2cl 1)NT m ]/[ ] and
substituting (11) in (2), the corresponding capacity process
C(t) can be approximated by

2m
(2cl 1)NT m 2
C(t) log2 2 2cl +
+
t
NT
2NT

+
t .
(12)
NT
Consequently, by setting C(t) to cl in (12), the duration of
the outage intervals is obtained to be
2
.
(13)
[(2cl 1)NT m ]
Then, by applying the concept of transformation of random
variables [14, pp. 130] to the relation in (13), we obtain the
following approximate PDF p ( ; cl ) of the outage durations
p ( ; cl ) =
[(2cl 1)NT m ]
4( )2

2
[(2cl 1)NT m ] 2
exp
. (14)
8( )2
The PDF in (14) can be identifiedas the Rayleigh distribution

with a parameter = 2 / [(2cl 1)NT m ] .
Moreover, considering
the result in (14), the kth-order

k mok
of
the
outage
intervals,
defined
as
E
=
ments
E
k
p ( ; cl ) d [14], can be approximated by
0

k
k
k/2 k
E 2 1+
(15)
2
where () is the gamma function [15, Eq. (8.310(1))]. Note

that for k = 1, E( ) in (15) provides an approximation for
at low values ofcl in
the ADF TC (cl ) of the channel
capacity

the form of TC (cl ) = 2/ [(2cl 1)NT m ] .
In addition to the above statistical quantities, the asymptotic
analysis allows us to also investigate the PDF of the instantaneous capacity loss, incurred during deep channel fades. We
define the instantaneous capacity loss C as the difference
between the average capacity and the minimum instantaneous
capacity occurring below cl . Referring again to Fig. 1, the
capacity loss C can be written as
C = cm cl + h
(16)
where cm denotes the average capacity and h is the capacity

loss below cl . Here, it should be emphasized that h changes
randomly over the outage events. It can be shown that the
minimum capacity occurs usually at t = /2. Taking this into
account and using (12) and (13), an approximation for C
can be obtained as

3( )2 2
cl 2m
Ccmlog2 22 +
+
.
NT
2NT [(2cl 1)NT m ]
(17)
For MIMO-OSTBC systems over Rayleigh fading channels, it has been shown that cm log2 (1 + NR )
2
2 log2 (e)NT NR 2 02 / [NT (1 + NR )] [16]. From (17), it
becomes clear that the PDF of C can be determined from
that of by applying the concept of transformation of random
variables [14, pp. 130]. This yields
NT log(2) [(2cl 1)NT m ] cm C
pC (C; cl )
2
3( )2
cl

(2 1)NT m
cm C
cl
2
2
N
exp
2
+2m
T
.
3( )2
(18)
For completeness, it should be mentioned that the theoretical
expressions described in (4)(10), (14), (15), and (18) remain
valid for single-input multiple-output (SIMO) and multipleinput single-output (MISO) systems. The corresponding quantities can be obtained by setting NT or NR to 1 as long as
the Gaussian approximation of the PDF of the channel power
gain remains accurate.
V.
N UMERICAL R ESULTS
In this section, we present some numerical results for the

validation of the theory presented in Sections III and IV. To this
end, we compare the analytical results with the corresponding
simulation data. The simulation system used is based on Rices
sum of sinusoids (SOS), the parameters of which have been
computed using the generalized method of exact Doppler
spread (GMEDS1 ) [17, Sect. III]. The average SNR is chosen
to be 17 dB. The maximum Doppler frequency fmax is set
to 91 Hz. In the case of isotropic scattering, the normalized
ACF r(t) is given by r(t) = J0 (2fmax t), where J0 ()
denotes the zeroth-order Bessel function [15, Eq. (8.411(1))].
To ensure the validity of the Gaussian approximation, the
number of antennas, i.e., NT and NR , have to be chosen such
as E(((t) m )4 ) 34 . It should be stressed that this
condition can be satisfied even for a small number of antennas.
In fact, in the following, we consider a 44 MIMO-OSTBC
VI.
C ONCLUSION
In this paper, we have investigated approximate solutions

for the PDF of the duration of the capacity outage intervals
in massive MIMO-OSTBC Rayleigh fading channels. The
assumption of a large number of transmit antennas and receive
antennas allows us to approximate the instantaneous channel
power gain by a left-truncated Gaussian process. The so-called
Rice probability function corresponding to the instantaneous
channel capacity has been investigated. This function is useful
to approximate the initial behavior of the distribution of the
capacity outage intervals. Moreover, the asymptotic analysis
at low capacity thresholds has also been conducted, and
approximate solutions for the PDF and the kth-order moments
180
Theory (QC,+ (; cl ))
PDF, QC,+( ; cl )
160
Simulation (p (; cl ))
cl = 7.8 bits/s/Hz
cl = 8 bits/s/Hz
140
120
100
80
60
40
20
0
0.005
0.01
0.015
0.02
0.025
Outage duration, (s)
0.03
Fig. 2: Comparison of the Rice probability function

QC,+ ( ; cl ) and the corresponding simulation results for a
44 MIMO-OSTBC system and two values of the threshold
cl .
1800
PDF, p ( ; cl )
system to produce Figs. 25. For comparison purposes, the

exact PDF, p ( ; cl ), will be the one obtained by simulation.
Fig. 2 presents the Rice probability function QC,+ ( ; cl )
and the simulated PDF p ( ; cl ) of the outage intervals for
cl = 8 and 7.7 bits/s/Hz. As can be seen, the Rice probability
function provides a good description of the initial behavior of
the PDF p ( ; cl ) of the outage durations, i.e., for small values
of . For large values of , the function QC,+ ( ; cl ) tends to
the average LCR NC (cl ) (approximately 85 and 78 crossings/s
for cl equal to 8 and 7.7 bits/s/Hz, respectively). Additionally,
Fig. 2 shows that when the threshold cl decreases, the part
of the PDF p ( ; cl ) that is approximated by the function
QC,+ ( ; cl ) increases. Now, we are interested in the behavior
of the PDF p ( ; cl ) of the capacity outage durations for low
thresholds cl or equivalently for low outage probabilities. The
capacity thresholds used in the following figures have been
selected by comparing the exact closed-form expression for
the ADF Tc (cl ) reported in [4, Eq. (20)] with that obtained
from the approximation in (15) for k = 1. The approximate
PDF p ( ; cl ) of the outage durations is presented in Fig. 3
together with the Rice probability function QC,+ ( ; cl ) and
the corresponding simulation results in the cases where cl is
6, 6.5,
c and 6.7 bits/s/Hz. The outage probability is defined
by 0 l pC (c)dc. The outage probabilities for the cases where
cl is 6, 6.5, and 6.9 bits/s/Hz are found to be 13%, 5.45%,
and 0.3%, respectively. As can be observed from Fig. 3,
there is an excellent agreement between the theoretical and
simulation results. It should also be noted that for deep
outage thresholds, the Rice probability function QC,+ ( ; cl )
represents a good approximation of the PDF p ( ; cl ). From
Figs. 2 and 3, it can be observed that the Rice probability
function QC,+ ( ; cl ) becomes more accurate in describing
the PDF p ( ; cl ) if the capacity threshold cl decreases.
Moreover, Fig. 3 shows that the average and spread of the
outage intervals decrease with decreasing values of cl . Since
our work is focusing on massive MIMO systems, we present
in Fig. 4 the approximate PDF p ( ; cl ) together with the Rice
probability function QC,+ ( ; cl ) for 1616, 3232, 6464,
and 128128 MIMO-OSTBC configurations by considering
in all cases the same threshold cl = 8 bits/s/Hz. Here, the
simulation results are omitted due to the large simulation time.
As can be observed from Fig. 4, increasing the number of
antennas results in a decrease of both the mean and variance of
the outage durations. Finally, Fig. 5 illustrates the approximate
PDF pC (C; cl ) with the corresponding simulation results
for three values of cl . Here again, a good fit can be seen
between the theoretical and simulation results.
1600
Theory (
p (; cl ))
1400
1200
Simulation (p (; cl ))
cl = 6.7 bits/s/Hz
1000
800
cl = 6.5 bits/s/Hz
600
cl = 6 bits/s/Hz
400
200
0
0
0.5
1.5
2
3
x 10
Fig. 3: Comparison of the approximate PDF p ( ; cl ), the

Rice probability function QC,+ ( ; cl ), and the corresponding
simulation results for a 44 MIMO-OSTBC system and three
values of the threshold cl .
of the outage durations as well as the distribution of the

instantaneous capacity loss have been derived. The correctness
and the validity of the derived analytical expressions have
been verified by means of computer simulations. Although,
the approximate results are intended to the case of massive
MIMO-OSTBC systems, their applicability has been shown to
be valid for a small number of antenna elements.
A PPENDIX
In this appendix, we present the derivation details of the
Rice probability function QC,+ ( ; cl ). It is clear from (3)
that the knowledge of the expressions of both the joint PDF
pc1 c1 c2 c2 (c1 , c1 , c2 , c2 ) and the LCR NC (cl ) is required for the
determination of the Rice probability function QC,+ ( ; cl ).
Moreover, (2) shows that the statistical properties of the
instantaneous channel capacity C(t) can be obtained from
x 10
for the Rice probability function QC,+ ( ; cl )
PDF, p ( ; cl )
Theory (
p (; cl ))
QC,+ ( ; cl ) =
1.5
128128 OSTBC-MIMO
1
cl
(2
pI1 I1 I2 I2
6464 OSTBC-MIMO
0.2
0.4
0.6
0.8
1
3
x 10
Fig. 4: Effect of the number of antennas on the approximate

PDF p ( ; cl ) for the same threshold cl = 8 bits/s/Hz.
3.5
Theory
PDF, pC (C; cl )
3
2.5
2
Simulation
cl = 6 bits/s/Hz
cl = 6.5 bits/s/Hz
1.5
1
cl = 6.7 bits/s/Hz
0.5
0
0
0
dI1
(2cl 1)NT m

1)NT m (2cl 1)NT

, I1 ,

NI
+

dI2 I1 I2
0

m
, I2 .
The left-hand side of (20) corresponds to the Rice probability

function QI,+ ( ; Il ) of the Gaussian process I(t) evaluated
at the level Il = [(2cl 1)NT m ]/[ ] [6, Eq. (47)].
This corresponds exactly to the result in (4).
1616 OSTBC-MIMO
0
0
(20)
3232 OSTBC-MIMO
0.5
Capacity loss, C (bits/s/Hz)
Fig. 5: Comparison of the approximate PDF pC (C; cl ) and

the corresponding simulation results for a 44 MIMO-OSTBC
system and three values of the threshold cl .
those of the Gaussian process I(t) by means of a transformation of random variables. Hence, by a transformation of random variables using I(t) = [(2C(t) 1)NT
= NT log(2)2C(t) C(t)/(
m ]/( ) and I(t)
), the joint
PDF pc1 c1 c2 c2 (cl , c1 , cl , c2 ) can be expressed as

pc1 c1 c2 c2 (cl , c1 , cl , c2 ) = |J| pI1 I1 I2 I2 Il , I1 , Il , I2
(19)
where pI1 I1 I2 I2 (I1 , I1 , I2 , I2 ) represents the joint PDF of
+ ) and Il
the processes I(t), I(t),
I(t + ), and I(t
denotes an arbitrary level which can be expressed as Il =
[(2cl 1)NT m ]/( ). In (19), the quantity |J| =
4
[2cl NT log(2)/ ( )] denotes the Jacobian of the underlying
transformation. Similarly, it can be shown that the LCR
NC (cl ) can be obtained as NC (cl ) = NI ([(2cl 1)NT
m ]/[ ]) = NI (Il ), where NI (Il ) is the LCR of the
process I(t) evaluated at the threshold Il . Therefore, after
substituting (19) in (3), we can find the following expression
R EFERENCES
[1] J. H. Winters, On the capacity of radio communication systems
with diversity in a radio fading environment, IEEE J. on Sel. Areas
Commun., vol. 5, no. 5, pp. 871878, Jun. 1987.
[2] I. E. Telatar, Capacity of multiantenna gaussian channels,, AT&T
Bell Labs Internal Tech. Memo., vol. 5, no. 5, Jun. 1995.
[3] A. Giorgetti, P. J. Smith, M. Shafi, and M. Chiani, MIMO capacity,
level crossing rates and fades: The impact of spatial/temporal channel
correlation, J. of Commun. and Networks (JCN), vol. 5, no. 2, pp.
104115, Jun. 2003.
[4] B. O. Hogstad, M. Patzold, N. Youssef, and V. Kontorovitch, Exact
closed-form expressions for the distribution, the level-crossing rate,
and the average duration of fades of the capacity of OSTBC-MIMO
channels, IEEE Trans. Veh. Technol., vol. 55, no. 2, pp. 10111016,
Feb. 2009.
[5] G. Rafiq, M. Patzold, and V. Kontorovich, The influence of spatial
correlation and severity of fading on the statistical properties of the
capacity of OSTBC Nakagami-m MIMO channels, in Proc. IEEE
Veh. Technol. Conf., IEEE VTC2009-Spring, Barcelona, Spain, April
2009.
[6] S. O. Rice, Distribution of the duration of fades in radio transmission:
Gaussian noise model, Bell Syst. Tech. J., vol. 37, pp. 581635, 1958.
[7] N. Youssef and T. Kawabata, On the probability density functions of
outage and interoutage durations of the capacity of Rayleigh fading
channels, IEEE Trans. Wireless. Commun., vol. 8, no. 2, pp. 529534,
Feb. 2009.
[8] N. Youssef, G. Rafiq, R. Hicheri, T. Kawabata, and M. Patzold, An
asymptotic approximate solution to the distribution of the capacity
outage intervals in OSTBC-MIMO Rayleigh fading channels, in 24th
Proc. Int. Symp. on Pers., Indoor and Mobile Radio Commun., IEEE
PIMRC13, London, UK, Sep. 2013, pp. 12961300.
[9] W. H. Greene, Econometric Analysis, Prentice Hall, 7th edition, 2003.
[10] G. E. P. Box, J. S. Hunter, and W. G. Hunter, Statistics for Experimenters: Design, Innovation, and Discovery, Wiley, 2nd edition, July
2005.
[11] A. J. Paulraj, N. Nabar, and D. Gore, Introduction to Space-Time
Wireless Communications, Cambridge, U.K: Cambridge Univ. Press,
2003.
[12] M. Patzold, Mobile Fading Channels, New York: Wiley, 2002.
[13] M. Kac and D. Slepian, Large excursions of Gaussian processes, Ann.
Math. Stat., vol. 30, pp. 12151228, Dec. 1959.
[14] A. Papoulis and U. S. Pillai, Probability, Random Variables and
Stochastic Processes, McGraw Hill, 4th edition, 2002.
[15] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, Academic Press, 7th edition, 2007.
[16] J. Perez, Jesus Iba~
nez, L. Vielva, and I. Santamara, Closed-form
approximation for the outage capacity of orthogonal STBC, IEEE
Commun. Letters, vol. 9, no. 11, pp. 261263, Nov. 2005.
[17] M. Patzold, C.-X. Wang, and B. O. Hogstad, Two new sum-ofsinusoids-based methods for the efficient generation of multiple uncorrelated Rayleigh fading waveforms, IEEE Trans. Wireless. Commun.,
vol. 8, no. 6, pp. 31223131, Jun. 2009.
Characterisation of Channel Measurements at

70GHz in Indoor Femtocells
Stephan Hfner, Diego A. Dupleich, Robert Mller,
Christian Schneider and Reiner S. Thom
Electronic Measurement Engineering Group
Technische Universitt Ilmenau
98684 Ilmenau, Germany
Email: stephan.haefner@tu-ilmenau.de
AbstractA wideband, fully polarimetric and directional radio

channel sounding campaign conducted at 70 GHz is presented.
Measurement site was chosen as an entrance hall, featuring
femtocell size. Transmitter and receiver were placed according
to an access-point scenario. Initial large-scale parameters like
excess-delay, delay-spread and angular spread are derived for
channel characterisation. It is not the papers goal to describe
a new channel model. The main focus is the identication
of propagation effects and characteristics, which have to be
considered in future modelling. It was found, that rst more
specular than diffuse components occur and second the channel
is spatially sparse.
Keywords 70 GHz, 5G, millimetre-Wave, UltraWideband, Access-Point Scenario, Indoor Channel Sounding,
Femtocell
I. I NTRODUCTION
Millimeter-wave frequency bands are seriously considered
for 5G cellular communication and Multi-Gigabit WLAN,
mainly because of the availability of huge bandwidth resources
[1]. There are, however, some challenges which makes physical layer design for frequencies at 20 GHz...80 GHz substantially different from the usual approaches used for the common
cellular and WLAN microwave frequency bands below approx.
6 GHz. The dominating difference seems to be the anticipated
usage of big ("massive") antenna arrays perhaps with more
than 100 elements. These massive arrays are necessary to
mitigate the increasing with f 2 free-space path loss. So at
the same time we will need agile adaptive beamforming with
split beams following the multiple propagation and user paths.
This inuences not only the implementation effort a lot, but
also the architecture of space-time signal processing.
Therefore, many channel sounding campaigns and modelling studies are currently being carried out to reveal the
characteristic propagation features for millimeter waves. Examples are [2] for outdoor environments and [3][5] for indoor
scenarios at 70 GHz. However, it is more or less common to
all published results that measurement bandwidth is too small,
directional resolution is too much limited, double directional
resolution (which is necessary for antenna independent channel characterization) is not carried out and, nally, polarization
Jian Luo , Egon Schulz ,

Xiaofeng Lu , Tianxiang Wang
Huawei
Technologies Duesseldorf GmbH

80992 Mnchen, Germany
Email: Jianluo@huawei.com
Communications Technology Lab
Huawei Technologies Co., Ltd
Email: Stan.Lu@huawei.com
is not considered.
A measurement equipment for 60 GHz and rst polarimetric
channel measurements by the authors were described in [6],
[7]. Only recently a new dual polarimetric setup for 70 GHz
was reported in [8] and new modelling issues for millimeter
wave beamforming assessment were elaborated in [9]. The
measurement results presented in [8] support more deterministic modelling for millimeter wave approaches as we are
used to apply in the lower bands. In essence, at millimeter
wave frequencies compared to micro waves, besides of ca 10
times shorter wavelength, we will have 10...100 times more
antennas in an array and about 10...100 times more bandwidth.
This separates the channel response more into single multipath
contributions with agile time varying deterministic behaviour
which manifests in less fading and contiguous phase/delay
evolution. At the same time we have shown that the temporaldirectional taps change from diffuse to more specular polarisation, which makes polarimetric ltering feasible.
Here we present further measurement results, which were
recorded with our recently developed dual-polarized ultra
wide-band multi-channel sounder (DP-UMCS) for 70 GHz
[10], [11]. The measurement campaign represents an indoor
access-point scenario, with the receiver (RX) located at several
positions of the users equipment (UE) and the transmitter (TX)
located at different heights playing the role of the access point
(AP). Directional channel resolution at both link sides (referred
as double directional resolution) was achieved by using rotating horn antennas at TX and RX side. Large scale statistical
parameters like excess delay, delay spread and angular spread
are derived. The main focus of this paper is to show how
those parameters change between an omnidirectional and a
high directional selective view to the channel.
This paper is organised as follows: Section II describes
the measurement campaign which was carried out University
building. The data pre-processing steps is described in section
III. In section IV, analysis results are presented in the delay,
angular, and joint angular-delay domain. Section V concludes
the paper.
978-1-4799-8088-8/15/$31.00 2015 IEEE
II. M EASUREMENT C AMPAIGN

A. Measurement Site and Scenario
RX1
25m
Measurement site was the northern entrance hall of the

Zuse-building at the campus of TU-Ilmenau. The site represents a typical modern entrance hall of an ofce building,
featuring a wide open area, a high ceiling height and three
oors. Furthermore, stairways to the next oor, concrete
pillars and doors to ofces and corridors are present. The
entrance wall and left side wall (see Figure 1) are glass-steel
constructions, whereas other walls are built of concrete. The
dimension of the entrance hall is approximately 25 m 13 m.
The measurement scenario was chosen as a static access
point scenario, with a RX (here the UE) located at different
points in the entrance hall and a TX (here the AP) located at
two different positions at a wall. Distances between RX and
TX were below 20 m, why the scenario features femtocell size.
For the rst TX position, TX and RX were located at the
ground oor (see Figure 2a) of the entrance hall. The TX
was elevated to 2.4 m, to have access point height. The RX
features a height of 1.4 m. The RX was positioned along a
straight line in 5 m steps starting at the entrance door, which
was assumed as the main path of a passing person. Here,
TX and RX feature line of sight (LOS) conditions. Figure
1 shows a schematic overview of the measurement site, with
the RX positions and the distances of RX and TX within the
scenario. During the measurement, the RX vehicle and the
table where the TX was located were covered with absorber
mats to exclude reections. It has to be noted, that for RX
position 3 only half of the measurements are available, because
of an system outage.
For the second TX position, the TX was located at the rst
oor, whereas the RX was located at the ground oor (see
Figure 2b). The height of the TX above ground was 6.5 m.
The RX positions RX1 up to RX4 were repeated (see Figure
1), to have equal RX positions for later on comparison. Here,
TX and RX feature LOS condition for the RX at rst position
and non-line of sight (NLOS) condition otherwise.
For directional channel resolution, horn antennas at TX and
RX in combination with positioning systems were used. At RX
side, rotation in the azimuth plane was conducted in a range
from 180 to 165 in 15 steps. At TX side, we limited
rotation in the azimuth plane to 90 to 90 , because of a
wall behind the TX. Elevation angles were set from 45 to
45 . Azimuth and elevation at TX were also varied in 15
steps. It has to be noted, that the azimuth rotation direction
at TX and RX was clock-wise. The 0 -azimuth direction of
TX and RX are depicted in Figure 1 as the direction of the
triangle-basis.
Fully polarimetric measurements were conducted,
which means that two orthogonal polarisations were
radiated/measured at TX/RX stage. At any angular position
of TX and RX and per TX polarisation, one impulse response
is recorded at each RX channel, which is called a snapshot.
RX2
RX3
RX4
13m
Figure 1. Schematic overview of the measurement site with the TX position

(red triangle) at ground oor and the RX positions (yellow triangles)
(a)
(b)
Figure 2. Picture from the measurements with the RX at position RX4 and
the TX located at: (a) ground oor and (b) rst oor
B. Measurement Setup
For the channel sounding campaign, the Dual-Polarised
Ultra-wideband Multi-Channel Sounder (DP-UMCS) developed at TU-Ilmenau [10], [11] was used. Figure 3 shows the
scheme of the channel sounder setup.
The Low-Noise-Ampliers (LNAs) at RX stage have 20 dB
gain and the Power Amplier (PA) at TX stage has a maximum
output power of 30 dBm. In order to avoid saturation of the
RX LNAs, the TX output power was attenuated to 0 dBm for
the ground oor measurements. For the measurements with the
TX at the rst oor, the TX output power was 10 dBm. Table
I summarises the main features of the measurement system.
Polarisation at the TX was switched manually. Therefore,
DP-UMCS TU-Ilmenau
0 dBm | 10 dBm (Ground oor|First oor)
74 GHz
6.75 GHz
606 ns
horn antennas (20 dB gain, 15 HPBW)
1x2 (#Tx x #Rx)
Table I
CH1
H Pol.
UWB Sounder
TX
LNA
PA
UWB Sounder
RX
V Pol.
TX azimuth []
Measurement System
Channel Sounder
Transmit power
Center frequency
Bandwidth
CIR length
Antennas
MIMO sub-channels
90
80
60
85
30
90
95
30
100
60
105
90
180
120
LNA
Multiplier
x12
CH2
Multiplier
x12
60
0
60
RX azimuth []
120
110
Figure 4. Azimuth-Power-Spectrum in dB for RX located at RX4 and TX at

ground oor
PA
6.75 GHZ
Oscillator
PA
PA 38dB
30 m
trade-off between purely rectangular window and purely Hann

window.
Figure 3. Scheme of the Dual-Polarised Ultra-wideband Multi-Channel

Sounder Setup
capturing of the full polarimetric channel information in one

rotation run of RX and TX was not possible.
A single clock signal was distributed to TX and RX, in order
to have same clock signals at both stages. Power ampliers
(PA) were necessary to have sufcient clock power at TX and
RX stage.
III. S YSTEM C ALIBRATION
In order to exclude the inuence of the measurement system
from the recorded channel impulse responses, calibration is
necessary. Here, calibration is referred as the deconvolution of
the measured data with the channel sounder impulse response,
which includes antennas, up/down-converters, LNAs, PAs and
the UWB Sounder TX/RX unit.
The impulse response of the sounder system is measured
within an in situ calibration measurement, where TX and RX
antenna are placed face-to-face in a known distance and an
impulse response is measured for each TX polarisation. To
exclude parasitic reections from the in situ environment, the
measured impulse responses are gated to get only the line-ofsight path. After extraction of the line-of-sight path, free-space
propagation delay is corrected and free-space attenuation is
corrected utilising the Friis-formula.
Because the spectrum is band limited, reduced signal magnitude occurs at the spectrum edges. Therefore the whole
bandwidth is not usable, because of the low signal-to-noise
ratio (SNR) in those regions. A window function is applied to
the data after de-convolution with the in situ measurements,
to cut the noise oor at the spectrum edges (approx. 0.5 GHz
per spectrum side) and to reduce the inuence of low SNR
values at the band pass lter slopes. We end up in a 3 dB
bandwidth of approx. 4 GHz. The applied window function
is a raised-cosine-window with a roll-off factor of 0.5 as a
IV. M EASUREMENT A NALYSIS

A. Angular Channel Characterisation
Figures 4 and 5 show as an example the azimuth power
spectrum at RX position 4 and the TX located at the ground
oor and at the rst oor, respectively. We averaged over the
elevation angles at TX to get the mean power contribution
from the scanned elevation range. Figure 5 shows, that most of
the signal arrives at azimuth of 0 and 180 , being reections
at the wall and entrance door, respectively. Therefore, multiple
reections also contribute. Furthermore, the channel is spatial
very sparse, because TX/RX transmit/receive sufcient power
in only a few distinct directions. Therefore, the calculated
azimuth spread of arrival (ASA), azimuth spread of departure
(ASD) and elevation spread of departure (ESD) [12] in Table
II for each RX position are very high.
Because of the low spatial resolution of our measurements,
spatial characteristics like direction of departure (DOD) and
direction of arrival (DOA) of each propagation path are
not available. High resolution methods become necessary, to
increase the resolution for a proper modelling of the spatial
channel characteristic to test e.g. beamforming algorithms
[13].
TX Pos.
RX Pos.
LOS/ NLOS
ASA [deg]
ASD [deg]
ESD [deg]
ground oor
RX1
RX2
RX3
RX4
LOS
LOS
LOS
LOS
74.27
60.86
76.57
53.10
53.26
51.44
29.07
26.77
27.82
rst oor
RX1
RX2
RX3
RX4
LOS
NLOS
NLOS
NLOS
92.15
79.87
81.85
92.33
51.18
52.31
57.45
62.20
31.30
30.11
32.24
33.07
Table II
L IST OF ESTIMATED ANGULAR PARAMETERS
90
110
60
115
30
120
125
30
130
60
135
140
PDP [dB]
TX azimuth []
150
160
170
90
180
120
60
0
60
RX azimuth []
140
120
Figure 5. Azimuth-Power-Spectrum in dB for RX located at RX4 and TX at

rst oor
180
100
200
300
Delay [ns]
400
500
600
Figure 7. Synthetic omni-directional PDP for the RX at position 4 and the

TX located at rst oor.
90
are increased in NLOS situations.
100
TX Pos.
PDP [dB]
110
120
RX Pos.
LOS/ NLOS
DS [ns]
MED [ns]
ground oor
RX1
RX2
RX3
RX4
LOS
LOS
LOS
LOS
28.57
38.60
32.29
161.62
161.03
153.48
rst oor
RX1
RX2
RX3
RX4
LOS
NLOS
NLOS
NLOS
40.74
47.96
56.72
44.32
228.44
230.96
284.44
222.81
130
140
150
160
100
200
300
Delay [ns]
400
500
600
Figure 6. Synthetic omni-directional PDP for the RX at position 4 and the

TX located at ground oor.
B. Delay Channel Characterisation

Delay characterisation is based on the power delay prole
(PDP) for omni-directional antennas at TX and RX, because
all transmitted/received paths have to be considered. The PDP
for omni-directional antennas is calculated as the average PDP
over all scanning angles at TX and RX.
In Figure 6 and 7, the omni-directional PDP for RX
at position 4 and TX at ground and rst oor is shown,
respectively. It is obvious, that more specular components
than diffuse components are present, as reported in [4], [9],
especially for the TX located at rst oor. However, for the
TX at ground oor, diffuse components are non-neglectable,
as seen in Figure 6.
Temporal parameters like delay spread (DS) and maximum
excess delay (MED) are calculated from the PDP, containing
only signicant paths. We dene all paths, which are 20 dB
below the maximum peak in the PDP as signicant, as shown
in Figure 6 and 7. For our analysis, a threshold of 30 dB was
to much, because noise samples were detected as paths too.
The delay difference between the last and rst detected path
denes the MED. The DS is calculated as the second central
moment over time [12]. The estimated temporal parameters
are summarised in Table III. It is obvious, that DS and MED
Table III
L IST OF ESTIMATED DELAY PARAMETERS
C. Angular-Delay Channel Characterisation

Highly directive antennas and beamforming are considered
for 5G communication systems [14], which acts as spatial
lters and therefore limits the number of received paths. We
investigate three different antenna scenarios: directive antennas
at TX and omni-directional antenna at RX, omni-directional
antenna at TX and directional antennas at RX, and directional
antennas at TX and RX. For the omni-directional case at
TX/RX-side, the PDPs are averaged over the scanning angles
at TX/RX-side. For each antenna scenario, we calculated the
DS.
Because spatial ltering reduces the received power, PDPs
are excluded from our calculations, to prevent that noise is
considered in the DS calculation. A PDP is discarded in
either of the three antenna scenarios, if the difference between
maximum and minimum magnitude in the PDP is lower than
20 dB. The probability of discarding a PDP is given in Table
IV. As visible, directional antennas increase the probability of
discarding a PDP, because of the channels spatial sparsity.
Therefore, beam adaption techniques as presented in [9] are
necessary to maintain the communication link, if the UE
moves. Such adaption techniques can be time, energy and
data rate exhaustive. Because the free-space path loss is
considerable for femtocells (96 dB at 75 GHz and 20 m
distance), low-gain and omni-directional antennas seem to
CDF
0.8
0.6
TX GF TX Omni
TX FF TX Omni
TX GF RX Omni
TX FF RX Omni
TX GF
TX FF
0.4
0.2
0
50
100
150
200
250
Delay Spread [ns]
300
350
400
Figure 8. CDF of the DS for the TX at ground oor (GF, solide lines) and rst
oor(FF, dashed lines). The CDF are calculated for: directional antennas at TX
and omni-directional antenna at RX (orange colour), omni-directional antenna
at TX and directional antennas at RX (blue colour), directional antennas at
TX and at RX (green colour).
be sufcient for communication. Further measurements are

necessary to investigate the performance gain of beamforming
in femtocell scenarios and the spatial channel characteristics
rate of change, to derive necessary adaption rates.
The cumulative distribution function (CDF) of the resulting
DS are displayed in Figure 8. It becomes obvious from
Figure 8, that using directional antennas at TX and RX side
narrows the delay spread, especially for the TX at rst oor
with mainly NLOS conditions. For the TX at ground oor,
directional antennas at TX are the best solution in terms of
the smallest DS. Directional antennas at both sides is the
optimal solution in terms of narrowing of the DS for the TX
at rst oor. In summary, the delay spread can be inuenced
by spatial ltering. This is quite important for the denition
of equalizer lengths or gap intervals.
Directional RX
Directional TX
Directional TX and RX
0%
TX at ground oor
1.92 %
31.91 %
1.04 %
TX at rst oor
34.89 %
87.00 %
Table IV
P ROBABILITY OF DISCARDING A PDP FOR DIFFERENT ANTENNA
SCENARIOS
V. C ONCLUSION
5G communication systems will feature higher bandwidth
in spatial domain (massive MIMO, pencil-beamforming) and
frequency domain (Gigabit transmission). The increase of
target system bandwidth necessitates channel measurements
with a higher bandwidth than the target system, to estimate
appropriate channel parameters for modelling. Because of
the limited spatial channel resolution with scanning measurements, high resolution parameter estimation is necessary, to
increase the spatial resolution.
As visible from our measurements, more specular components rather than diffuse scattering occurs, because of the
increased frequency bandwidth. Nevertheless, diffuse components are still present. Furthermore, because of the increased
spatial bandwidth, the channel occurs more directional and
scatterers become more spot like. Apart from the effects of
increased system bandwidth, propagation paths behave quasioptical like due to their small wavelength at millimetre wave
frequencies.
In summary, modelling of wave propagation at millimetre
wave frequencies for 5G systems differs from modelling at
micro wave frequencies for below 5G systems. New model
approaches are necessary, to account for the increased bandwidth. Therefore, known models like WINNER II have to
be extended, to support the high system bandwidth and must
include new modelling approaches.
R EFERENCES
[1] F. Boccardi, R. Heath, A. Lozano et al., Five disruptive technology
directions for 5G, Communications Magazine, IEEE, vol. 52, no. 2,
pp. 7480, February 2014.
[2] G. MacCartney and T. Rappaport, 73 GHz millimeter wave propagation
measurements for outdoor urban mobile and backhaul communications
in New York City, in Communications (ICC), 2014 IEEE International
Conference on, June 2014, pp. 48624867.
[3] N. Zhang, X. Yin, S. X. Lu et al., Measurement-based Angular Characterization for 72 GHz Propagation Channels in Indoor Environments,
in Mobile Communications in Higher Frequency Bands (MCHFB), 2014
Globecom Workshop on, December 2014, pp. 219223.
[4] K. Haneda, J. Jrvelinen, A. Karttunen et al., Indoor short-range
radio propagation measurements at 60 and 70 ghz, in Antennas and
Propagation (EuCAP), 2014 8th European Conference on, April 2014,
pp. 634638.
[5] S. Nie, G. R. MacCartney, S. Sun et al., 72 GHz millimeter wave indoor
measurements for wireless and backhaul communications, in Personal
Indoor and Mobile Radio Communications (PIMRC), 2013 IEEE 24th
International Symposium on, Sept 2013, pp. 24292433.
[6] A. Garcia, R. Mller, F. Wollenschlger et al., Dual-Polarized Architecture for Ultra-wideband Channel Sounding at 60 GHz with Digital/Analog Phase Control based on 0.25 mm SiGe BiCMOS and LTCC
Technology, in Antennas and Propagation (EUCAP), Proceedings of
the 5th European Conference on, 2011, pp. 1454 1458.
[7] A. Garcia Ariza, R. Mller, F. Wollenschlger et al., 60 GHz Ultrawideband Polarimetric MIMO Sensing for Wireless Multi-Gigabit and
Radar, Antennas and Propagation, IEEE Transactions on, vol. 61, no. 4,
pp. 16311641, April 2013.
[8] D. A. Dupleich, R. Mller, C. Schneider, E. M. Vittucci, F. Fuschini,
R. Thom, and V. Esposti, Directional characterization of the 60 GHz
indoor-ofce channel, in XXI General Assembly Beijing, August 2014.
[9] V. Degli-Esposti, F. Fuschini, E. Vitucci, M. Barbiroli, M. Zoli, L. Tian,
X. Yin, D. Dupleich, R. Mller, C. Schneider, and R. Thoma, RayTracing-Based mm-Wave Beamforming Assessment, IEEE Access,
vol. 2, pp. 13141325, October 2014.
[10] R. Mller, S. Hfner, D. A. Dupleich et al., Ultra-Wideband Channel
Sounder for Measurements at 70 GHz, 2014, accepted for VTC2015Spring.
[11] R. Mller, D. A. Dupleich, C. Schneider et al., Ultrawideband 3D
mmWave Channel Sounding for 5G, in XXI General Assembly Beijing,
August 2014.
[12] B. Fleury, First- and second-order characterization of direction dispersion and space selectivity in the radio channel, Information Theory,
IEEE Transactions on, vol. 46, no. 6, pp. 20272044, Sep 2000.
[13] S. Mohammad Razavizadeh, M. Ahn, and I. Lee, Three-Dimensional
Beamforming: A new enabling technology for 5G wireless networks,
Signal Processing Magazine, IEEE, vol. 31, no. 6, pp. 94101, Nov
2014.
[14] W. Roh, J.-Y. Seol, J. Park et al., Millimeter-wave beamforming as
an enabling technology for 5G cellular communications: theoretical
feasibility and prototype results, Communications Magazine, IEEE,
vol. 52, no. 2, pp. 106113, February 2014.
Validation of a Geometry-Based Statistical mmWave

Channel Model Using Ray-Tracing Simulation
Qian (Clara) Li, Hooman Shirani-Mehr, Tommaso Balercia, Apostolos Papathanassiou, Geng Wu
Intel Corporation
liqian@ieee.org, hooman.shirani-mehr@intel.com, tommaso.balercia@intel.com,
apostolos.papathanassiou@intel.com, geng.wu@intel.com
Shu Sun, Mathew Khalil Samimi, Theodore S. Rappaport

NYU WIRELESS, New York University
ss7152@nyu.edu, mks@nyu.edu, tsr@nyu.edu
AbstractNext-generation wireless communications systems
are expected to exploit frequency bands above 6 GHz. An
important transition towards such bands will be design of channel
models capable of supporting the design of efficient air-interface
and networks. In this paper, we describe a geometry-based
statistical channel model and apply ray-tracing simulation for its
validation. Focusing on path loss and root-mean-square (RMS)
delay spread as metrics, we show that the proposed modeling
approach is flexible and realistic.
I. I NTRODUCTION
Next-generation wireless communications systems are expected to explore the vast frequency resources in frequency
bands above 6 GHz (for brevity, we term the bands from
6 GHz to 300 GHz as millimeter-wave, mmWave, bands).
Unlike in lower bands where the wave propagation behavior
is dominant, for higher frequencies the particles behavior is
prominent. This leads to unique features of the wireless propagation channel at mmWave bands such as high diffraction loss,
high diffusion and sensitivity to attenuations in the propagation
environment, i.e., rain, foliage, etc. As a consequence, the existing channel models designed for bands below 6GHz are not
always directly applicable. New channel models considering
both accuracy and implementation efficiency would be needed
in air-interface design and system performance evaluation for
high frequency band communications [1], [2].
As a result of the extensive studies of the mmWave channel
[1]-[7], several models have been proposed. Examples are the
map-based ray-optical model proposed in [3], the geometrybased quasi-deterministic model proposed in [4], and the
statistical models based on power-delay-angular distributions
proposed in [5] and [6]. In the map-based ray-optical model,
a detailed map of the communication environment is used to
calculate how the electromagnetic energy interacts with the environment at hand. In the geometry-based quasi-deterministic
model, the channel model is modeled as one line-of-sight
(LOS) path, one non-line-of-sight (NLOS) path reflected from
the ground, and several random NLOS paths. The LOS path
and the NLOS path reflected from the ground are deterministically calculated from the geometry. The models based
on power-delay-angular distributions represent the channel by
its excess delay distribution, power-delay profile and powerangular spectra which are derived from channel measurements
and/or ray-tracing simulation.
All of such models represent the channel in different levels
of detail and with different levels of implementation complexity. In general, the more geometrical information is considered
in the model, the more accurate the channel is likely to
be, at the expense of complexity. On the other hand, when
abstractions on the geometry are applied, expected complexity
is low, but accuracy cannot be achieved without significant
channel measurement efforts. It goes without saying that an
effective channel model should be capable of portraying the
aspects of the channel with an impact on the air-interface
design, while remaining of low complexity and flexible enough
to represent a variety of scenarios.
In [7], we proposed a geometry-based channel model using
the environment geometrical statistics, the shadow fading
statistics and the time-of-arrival distribution as inputs for
generating the propagation channel. The proposed channel is
applicable both for bands below 6 GHz and the higher ones.
As only geometrical statistics are used, the complexity of the
model is in the same order as that of the models based on
power-delay-angular distributions. In [7], the channel model
was calibrated with NYU measurement results in Manhattan
[2] and showed good alignment. The calibration in [7], however, focuses on the path loss and is rough, as we do not
have the exact geometrical distributions of the environment in
which the measurements were taken. In this paper, we calibrate
the proposed channel model using ray-tracing simulation [8]
based on the Rosslyn city map. We first obtain the required
input geometry and channel statistics from ray-tracing. The
channel generated by the channel model using the obtained
geometry and channel statistics is then compared with the
channel obtained using ray-tracing on the same environment.
Using the omni path loss and root-mean-squared (RMS) delay
spread (DS) as the two metrics for comparison, we show that
the channel generated by the proposed model matches well
with the channel generated using ray tracing.
978-1-4799-8088-8/15/$31.00 2015 IEEE
II. G EOMETRY-BASED S TATISTICAL C HANNEL M ODEL

The double directional channel model [9] shown in (1)
provides a generally applicable framework for representing wireless channels. Each of the Mn paths within each
of the N clusters is defined by its gain Hn,m , random
phase n,m , excess delay n,m , azimuth-angle-of-departure
(AoD) AoD,n,m , azimuth-angle-of-arrival (AoA) AoA,n,m ,
zenith-angle-of-departure (ZoD) ZoD,n,m , and zenith-angleof-arrival (ZoA) ZoA,n,m .
loss by the kr and kd scattering event, which are randomly

drawn from the reflection loss and diffraction loss distributions. The path gain can finally be calculated as
4d 2
Hn (d) = A(d)SF (d)
,
(6)
c
where A(d) is the atmospheric attenuation which can be

obtained from measurements. With Hn (d), the remaining
paths in the cluster can be calculated by applying an intracluster power-delay-profile (PDP). Examples of such intraMn
N

cluster PDP at 60 GHz bands can be found in [4] and [10].
jn,m
H =
Hn,m e
( n,m )(R AoA,n,m )
It is worth noting that, since the shadow fading distribution
n=1 m=1
at
different
frequency bands have been intensively investigated,
(T AoD,n,m )(R ZoA,n,m )(T ZoD,n,m ).
the proposed modeling approach helps in relaxing channel
(1) measurement and calibration efforts. That is, in modeling the
In order to represent the behavior of mmWaves, we need channel at bands characterized by a mild frequency differto find ways of modeling the parameters appearing in (1), ence, the environment geometrical statistics can be reused
preferably at the cost of low measurement and calibration unmodified, and the shadow fading distributions can be taken
from literature. Moreover, unlike the geometry-based channel
efforts.
In [7], we proposed a geometry-based statistical channel models which specify geometrical locations, the proposed
model capable of generating the channel parameters by taking model only uses statistical geometrical distributions. This
the statistical geometrical distributions of the communication allows low implementation complexity.
In order to validate the proposed model, in the following
environment as inputs. Given the time of arrival (ToA) of a
given path, we can calculate the length of the path. Together sections, we first use ray-tracing to get the statistics of the
with the geometrical distribution of the scatterers within the needed inputs: ToA, geometrical distribution of the scatters
environment, and the statistics on the reflection/diffraction and reflection/diffraction losses. Then, the channel generated
losses, we can calculate the shadow fading for each path. using the proposed channel model is compared to the one
The path gain can in turn be calculated by adding the shadow obtained using ray-tracing.
fading with free-space path gain.
III. G EOMETRY AND C HANNEL S TATISTICS FROM
More specifically, given a path ToA tn (here we consider the
R AY-T RACING S IMULATION
central path of the nth cluster denoted by m = 0; subscript m
is removed for brevity), the length of the path can be calculated
We use the Wireless Insite [11] for ray-tracing simulation. A
as
street canyon scenario is considered using a map of Rosslyn,
dn = ctn ,
(2) Virginia, as shown in Fig. 1 and Fig. 2. By processing the
where c = 3 108 m/s is the speed of light in vacuum. The
probabilities of encountering kr reflections and kd diffractions
along the trajectory followed by the path at hand are instead
described by geometrical distribution of the environment. The
shadow fading can thus be calculated as

pr (kr |d)SFr (kr ) +
pd (kd |d)SFd (kd ), (3)
SF (d) =
kr
kd
where pr (kr |d) and pd (kd |d) are the probabilities of encountering kr reflections and kd diffractions at distance d,
respectively. As an example, if the environment geometry
follows the Poisson distribution, pr (kr |d) and pd (kd |d) are
given by
er d (r d)kr
(4)
pr (kr |d) =
kr !
and
pd (kd |d) =
ed d (d d)kd
,
kd !
(5)
with r and d are parameters used for calibration, SFr (kr )

and SFd (kd ) are respectively reflection loss and diffraction
propagation data generated from the ray-tracing simulations,

for each of the Tx-Rx link, we obtain the required inputs to the
channel model. The ray-tracing simulations were conducted
at 28 GHz and 73.5 GHz to facilitate cross-checking with
measurements done by NYU WIRELESS [2].
A. Geometry Distribution
We model geometrical distributions of the scatters using
the statistical distribution of the number of reflections and
diffractions experienced by paths of a given length as illustrated in Section II. We could observe a slight difference on
the distributions collected at 28 GHz and 73.5 GHz. As the
frequency increases, the environment observed becomes more
detailed, which affects the geometry distribution.
Given pr (d|kr ) and p(kr ), pr (kr |d) can be calculated as
pr (kr |d) =
pr (d|kr )p(kr )
pr (d|kr )p(kr )
=
.
p(d)
kr (p(d|kr )p(kr ))
(7)
Figs. 3, 4 and 5 show the obtained probability distributions at

28 GHz and 73.5 GHz. From Figs. 4 and 5, we can see that
0.25
Freq = 28 GHz
Freq = 73.5 GHz
0.2
PDF
0.15
0.1
0.05
0
0
Fig. 3.
Fig. 1.
Rosslyn map considered in ray-tracing simulation (3D)
x 10
3
4
Number of Reflections
PDF of the number of reflections
Freq = 28 GHz
Ray tracing, 0 reflection

Ray tracing, 2 reflections
Poisson, k=3
Poisson, k=4
Poisson, k=5
Poisson, k=6
Poisson, k=7
Poisson, k=8
Poisson, k=9
5
4
PDF
3
2
1
0
0
100
200
300
400
500
600
Travel Distance (m)
700
x 10

Poisson, k=3
Poisson, k=4
Poisson, k=5
Poisson, k=6
Poisson, k=7
Poisson, k=8
Poisson, k=9
4
PDF
Rosslyn map considered in ray-tracing simulation (2D)
ed (d)kr +3
,
pr (d|kr ) =
(kr + 3)!
1000
Freq = 73.5 GHz
the distribution of pr (d|kr ) can be approximated by a Poisson

distribution with probability density function
900
PDF of the traveling distance given the number of reflections at 28
Fig. 4.
GHz
Fig. 2.
800
3
2
1
(8)
where = 1/40.30 for 28 GHz and = 1/40.32 for 73.5

GHz.
The ray-tracing simulation considers up to 1 diffraction.
Fig. 6 shows the probability of diffraction with respect to the
path length, i.e., pd (kd = 0|d) and pd (kd = 1|d).
B. Shadow Fading Statistics
To obtain the shadow fading statistics, we differentiate the
paths by the number of reflections and diffractions encountered
and calculate the shadow fading of each path as the difference
of the actual path loss with the free-space path loss. The
actual path loss is obtained from ray-tracing, where reflection
loss is calculated from reflection coefficient of the surface
materials and diffraction loss is evaluated using uniform theory
of diffraction [12]. The free-space path loss is calculated using
0
0
Fig. 5.
GHz
100
200
300
400
500
600
Travel Distance (m)
700
800
900
1000
PDF of the traveling distance given the number of reflections at 73
the Friis equation from the path length. For the paths which
undergo the same number of reflections and diffractions, we
calculate the mean and the variance of the shadow fading. As
the ray-tracing tool simulates up to six reflections and one
diffraction, the paths can be categorized into 12 groups. As
it can be seen from Fig. 7, the shadow fading loss increases
with the number of reflections. Diffraction causes an additional
30 dB - 40 dB loss. Comparing 28 GHz and 73.5 GHz, the
difference in reflection loss is marginal while the difference
in diffraction loss is high.
x 10
Prob of 0 diffraction, 28 GHz

Prob of 1 diffraction, 28 GHz
Prob of 0 diffraction, 73.5 GHz
Prob of 1 diffraction, 73.5 GHz
0.8
Prob
10
PDF
0.4
0.2
0
0
200
400
Fig. 6.
600
800
Traveling Distance (m)
1000
Shadow Fading (dB)
6
4
Probability of diffraction
2
0
0
Mean, 0 Diffraction, Freq = 28 GHz

Std, 0 Diffraction, Freq = 28 GHz
Mean, 1 Diffraction, Freq = 28 GHz
Std, 1 Diffraction, Freq = 28 GHz
Mean, 0 Diffraction, Freq = 73.5 GHz
Std, 0 Diffraction, Freq = 73.5 GHz
Mean, 1 Diffraction, Freq = 73.5 GHz
Std, 1 Diffraction, Freq = 73.5 GHz
10
20
30
40
Fig. 9.
Carrier Frequency = 28 GHz
Fig. 7.
12 x 10
3
4
Number of Reflections
Shadow fading Statistics
Ray tracing, 28 GHz

Ray tracing, 73.5 GHz
Exp fitted, 28 GHz
Exp fitted, 73.5 GHz
10
NLOS PL Proposed Model

NLOS PL Closein Fit
LOS PL Closein Fit
NLOS PL Raytracing Sim
LOS PL Raytracing Sim
140
Path Gain (dB)
6
7
x 10
Inter-arrival time in the NLOS scenario for 28 GHz and 73.5 GHz
160
60
0
1
2
3
4
5
Interarrival Time, NLOS UE (sec)
180
50
120
100
80
8
PDF
1200
10
60
0
10
10
10
10
TxRx distance (m)
Fig. 8.
Ray tracing, 28 GHz

Ray tracing, 73.5 GHz
Exp fitted, 28 GHz
Exp fitted, 73.5 GHz
12
0.6
0
0
Fig. 10. Comparison of path loss between channel model and ray-tracing at
28 GHz
1
2
3
4
Excessive Delay w.r.t. LOS, LOS UE (sec)
6
7
x 10
Excess delay in the LOS scenario for 28 GHz and 73.5 GHz
C. ToA Distribution
The ToA distribution is used to generate the traveling time
of the paths. We calculate the ToA in the form of excess delay
with respect to the LOS path. From Figs. 8 and 9, we can
observe that the excess delay and the inter-arrival time match
well with the exponential distribution f ( ) = e u( ). For
28 GHz, the exponential distribution parameters are los =
91 ns and nlos = 78 ns; for 73.5 GHz, the parameters are
los = 86 ns and nlos = 89 ns.
IV. C HANNEL M ODEL VALIDATION WITH R AY-T RACING
We apply the obtained geometrical statistics to the proposed model so that the channel realizations are fit into the
communication environment used for conducting ray-tracing
simulations. We use the path loss and RMS DS in the NLOS
scenario as the metrics to validate the channel model as shown

in Figs. 10, 11 and 12.
The path gain of each Tx-Rx link is the omni path loss
calculated by summing up the path gains of all the paths, i.e.,
Nt,r
Ht,r =
Ht,r,n
(9)
n=1
where Nt,r is the number of paths of the Tx-Rx link indexed

by (t, r). Following the reference model defined in [1], we find
the MMSE fit of the omni path loss in the NLOS scenario as
P Lnlos [dB](d) = 61.38 + 32.2 log10 (d)
(10)
P Lnlos [dB](d) = 69.77 + 31.7 log10 (d)
(11)
and
for 28 GHz and 73.5 GHz, respectively. For the LOS scenario,
the path loss exponent of the omni path loss is 1.9 at 28 GHz
and 1.91 at 73.5 GHz. The omni path loss calculated from the
73.5 GHz, while the 90%-tile of the RMS DS is 230 ns at 28

GHz and 212 ns at 73.5 GHz.
A good match in the two metrics can be observed between
the channel generated by the proposed model and the channel
obtained from ray-tracing simulations. Note that due to the
limitations of the ray-tracing simulation in modeling diffuse
scattering and intra-cluster rays, the obtained channel statistics
may be subject to changes based on measurements or higher
resolution ray-tracing simulation.
Carrier Frequency = 73.5 GHz

200
NLOS PL Proposed Model
NLOS PL Closein Fit
LOS PL Closein Fit
NLOS PL Raytracing Sim
LOS PL Raytracing Sim
180
Path Gain (dB)
160
140
120
V. C ONCLUSION
100
80
60
0
10
10
10
10
TxRx distance (m)
Fig. 11. Comparison of path loss between channel model and ray-tracing at
73.5 GHz
Empirical CDF
1
0.9
0.8
R EFERENCES
0.7
[1] T. S. Rappaport, Robert W. Heath Jr., R. Daniels, J. Murdock, Millimeter Wave Wireless Communications, Pearson/Prentice-Hall, c. 2015
[2] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang,
G. N. Wong, J. K. Schulz, M. Shamimi, and F. Gutierrez, Millimeter
wave mobile communications for 5G cellular: It will work! IEEE
Access, vol. 1, pp. 335-345, May 2013
[3] METIS, Initial channel models based on measurements, ICT-317669METIS/D1.2, Apr. 2014
[4] MiWEBA, D5.1: channel modeling and characterization, FP7-ICT
368721/D5.1, June 2014
[5] M. K. Samimi and T. S. Rappaport, Ultra-wideband statistical channel
model for non line of sight millimeter-wave urban channels, in proc.
IEEE GLOBECOM 2014, pp. 1-5, Austin, Tx, Dec. 2014
[6] T. A. Thomas, H. C. Nguyen, G. R. MacCartney Jr., and T. S. Rappaport,
3D mmWave channel model proposal, in proc. IEEE VTC 2014 fall,
pp. 1-5, Sep. 2014
[7] Q. Li, G. Wu, and T. S. Rappaport, Channel model for millimeterwave communications based on geometry statistics, in Proc. IEEE
GLOBECOM 2014 Workshop on Mobile Communications in Higher
Frequency Bands, pp. 1-5, Austin, Tx, Dec. 2014
[8] G. Durgin, N. Patwari, and T.S.Rappaport, An advanced 3D raylaunching method for wireless propagation prediction, in Proc. IEEE
VTC Spring 1997, pp. 785-789, May 1997
[9] M. Steinbauer, A. F. Molisch, E. Bonek, The double-directional radio
channel, IEEE Antennas and Propagation Magazine, vol. 43, No. 4,
pp. 51-63, Aug. 2001
[10] A. Maltsev, et. al., Channel Models for 60 GHz WLAN Systems, IEEE document 802.1109/0334r8, May 2010, available at
https://mentor.ieee.org/802.11/
[11] Wireless Insite, Remcom, http://www.remcom.com/wireless-insite
[12] J. B. Keller, Geometrical theory of diffraction, Journal of the Optical
Society of America, vol. 52, pp. 116-130, Feb. 1962
0.6
F(x)
In this paper, we validate the proposed geometry-based

statistical mmWave channel model using ray-tracing. Specifically, we first get the geometry and channel statistics from
ray-tracing, which are then applied in the channel model
for generating the channel realizations. Using the generated
channel realizations, we calculate the NLOS path loss and the
RMS DS. Comparison with those obtained using ray-tracing
validated the proposed channel model. Further work includes a
refinement of the channel model from measurement and highresolution ray-tracing, and the derivation of power-angular
distributions from channel statistics.
0.5
0.4
0.3
28 GHz, raytracing
28 GHz, proposed model
73.5 GHz, raytracing
73.5 GHz, proposed model
0.2
0.1
0
0.2
0.4
0.6
RMS DS (s)
0.8
1
6
x 10
Fig. 12. Comparison of delay spread between channel model and ray-tracing
proposed channel model at each Tx-Rx distance is obtained

by averaging over multiple channel realizations.
For a Tx-Rx link with N paths of excess delay n (n =
1, , N ), the RMS DS is calculated as

P
2

(n ) Hn
RM S = n=1N
,
(12)
n=1 Hn
where
N
n Hn
= n=1
.
P
n=1 Hn
(13)
In calculating the DS, we apply a threshold to the path gain.

Only paths with gain greater than -160 dB are considered
useful and included in the RMS DS calculation. As it can
be seen from Fig. 12 for the considered Rosslyn city scenario,
the 50%-tile of the RMS DS is 81 ns at 28 GHz and 70 ns at
Channel Estimation Using a 2D DFT for

Millimeter-Wave Systems
S. Montagner , N. Benvenuto , and P. Baracca

Department of Information Engineering, University of Padova, Italy. Email: {montagne, nb}@dei.unipd.it
Bell Labs, Alcatel-Lucent, Germany. Email: paolo.baracca@alcatel-lucent.com
AbstractThe usage of the millimeter wave (MMW) band

in the 5th generation (5G) networks relies on beamforming to
compensate the strong path-loss suffered at higher frequencies.
To exploit the beamforming implemented by multiple antenna
devices, proper algorithms to estimate the channel need to be
designed. In this work we propose a novel channel estimation
method for MMW systems where both transmitter and receiver
are equipped with fewer radio frequency chains than antennas
and implement hybrid analog-digital beamforming. First, we
define a training sequence which includes a set of analog and
digital beamformers to probe the channel. Then, we develop
an algorithm which estimates the channel parameters directly
rather than the multiantenna channel matrix and it is based
on the two dimensional (2D) discrete Fourier transform (DFT)
of the received training samples. Numerical results show the
effectiveness of the proposed channel estimation method even
in low signal to noise ratio (SNR) conditions.
Index TermsMillimeter wave (MMW), multiple-input
multiple-output (MIMO), beamforming, channel estimation.
I. I NTRODUCTION
The large amount of bandwidth available at frequencies
between 30 and 300 GHz makes the millimeter wave (MMW)
technology an important component of the 5th generation (5G)
cellular networks to cope with the increasing demand for data
rates forecast in the near future [1]. For a given size of the
antenna array, several antennas can be packed and suitable
beamforming transmissions can be used to compensate the
huge path loss in MMW systems [2]. Channel estimation is
needed to fully benefit from the beamforming, and getting a
reliable channel estimate has been recognized as one of the
main challenges [3].
However, only few works in the literature have recently
targeted this issue. An hierarchical codebook is designed in
[4] to select the beamformers in a MMW system with only
analog beamforming. On the other hand, hybrid analog-digital
beamforming where the transmitter (receiver) is equipped with
fewer radio frequency (RF) chains than transmit (receive)
antennas has shown to be a good compromise in terms of
performance and complexity with respect to pure analog
or pure digital solutions [5]. Hence, hybrid analog-digital
beamforming is considered in [6] and a new multi-resolution
codebook is designed there by leveraging compressed sensing
tools to capture the limited scattering and the antenna correlation of the MMW channel [7]. Originally, this method
required a feedback channel for the iterative exchange of
information between transmitter and receiver. A new version of
the algorithm [8] does not require anymore a feedback channel,
however it is still rather complex by requiring an iterative

search within an hierarchical codebook. Also the classical
approach to estimate the channel matrix is not feasible due
to the constraint of analog beamformers and the numerous
transmit/receive antennas which would require a very long
training sequence.
In this work we also consider hybrid beamforming and we
design a novel channel estimation method. First, we define a
sequence of beamformers to probe the channel. Then, rather
than directly estimating the channel matrix, we exploit the
limited scattering of the MMW channel by designing an algorithm which a) leverages the two dimensional (2D) discrete
Fourier transform (DFT) of the received training samples,
and b) iteratively estimates angle of arrival/departure and gain
of the different scatters. We recall that a channel estimation
method working in the frequency domain can be found in
[9]. There, however, the transmitter does not make use of a
liner phase array. Hence, our channel model is more complex.
Moreover, in [9], after a DFT across the signals at receive
antennas, an algorithm making use of the discrete time Fourier
transform, with a continuous frequency, is outlined to improve
the estimate of angles of arrivals. In practical implementations,
for an increased angle resolution, we find it is simpler to
increase the DFT size (see also Fig. 4). Numerical results in
a typical 60 GHz scenario show that the proposed channel
estimation method provides good performance even in low
signal to noise ratio (SNR) conditions.
Notation: T stands for transpose while H stands for transpose and complex conjugate. [y]a:b denotes a vector obtained
by extracting elements of vector y from index a up to index b.
0N M denotes the matrix of size N M with all zero entries,
IN the identity matrix of size N , and diag (x) the diagonal
matrix with the entries of vector x on the diagonal.
II. S YSTEM M ODEL
We consider the single user MMW system with hybrid
analog-digital beamforming shown in Fig. 1 where a transmitter (Tx) with M antennas communicates symbol x to a
receiver (Rx) with N antennas, and it consists of a transmitter
equipped with MRF RF chains, with MRF M , and a
receiver with NRF RF chains, with NRF N . The transmitter
is assumed to apply a MRF 1 complex valued base-band
digital beamformer, fBB CMRF 1 , followed by a RF analog
beamformer FRF CM MRF . Similarly, the receiver is
constituted by a RF analog combiner URF CN NRF and a
978-1-4799-8088-8/15/$31.00 2015 IEEE
digital
xRF
xBB
[FRF ]1,1
fBB,1
analog
yRF
uBB,1
[URF ]1,NRF
[FRF ]1,MRF
DAC
[FRF ]M,MRF
ADC
[URF ]N,1
[FRF ]M,1
fBB,MRF
digital
[URF ]1,1
n1
DAC
yBB
ADC
uBB,NRF
[URF ]N,NRF
nN
Fig. 1. MMW base-band equivalent scheme with hybrid analog-digital beamforming both at transmitter and receiver.
base-band digital combiner uBB CNRF 1 . To simplify the

hardware implementation, each element of URF and FRF has
unitary magnitude, whoever it may have an arbitrary phase. If
H denotes the N M channel matrix, let xBB = fBB x and
after defining
yBB = UH
RF (HFRF xBB + n) ,
(1)
the received signal at the detection point can be written as

y = uH
(2)
BB yBB ,

where n CN 0N 1 , n2 IN is the thermal noise with
variance n2 . For later we also define the transmit and receive
antenna signals, i.e.,
xRF = FRF xBB ,
yRF = HxRF + n .
(3)
where

(r)
(r)
,
Ar = aN 1 , . . . , aN L

(t)
(t)
A t = a M 1 , . . . , a M L
,
(6)
Hg = diag ([g1 , ..., gL ]) .

We stress that in MMW systems while H can be a very
large matrix with, depending on the number of antennas, up
to hundreds of columns and rows, its parameters are only 2L
real phases and L complex gains, where L is usually much
smaller than the number of antennas. If x2 is the power of x,
as each term in (5) has unitary power, the average SNR for
each receive antenna is given by
AW GN =
x2
,
n2
(7)
2
under the power constraint at the Tx side FRF fBB = 1.

2
At the receiver (transmitter) we consider a uniform linear

array (ULA) with N (M ) elements having an inter-element
distance D. Due to the properties of the MMW propagation,
we assume a narrowband clustered channel representation,
based on the extended Saleh-Valenzuela model. For a N element ULA, the array response vector can be written as
[5]
T

aN () = 1 ejD sin ... ej(N 1)D sin , (4)
where = 2/, with the carrier wavelength, and represents the ray angle of arrival (departure) in the azimuth plane.
With the assumption that L rays are received with the same
delay, and by denoting for the -th ray with g CN (0, 1)
(r)
(t)
the complex random gain, the angle of arrival and the
angle of departure, we can write the N M channel matrix
as
1
(5)
H = Ar Hg AH
t ,
L
For a given channel matrix and by assuming URF uBB =

1, we can define the SNR at detection point as

2 x2
H

.
= uH
BB URF HFRF fBB
n2
(8)
The improvement of with respect to AW GN , defined in

(7), is given by

2
H
,
=
= u H
(9)
BB URF HFRF fBB
AW GN
and by averaging with respect to H, we also introduce
= EH [] .
(10)
III. C HANNEL E STIMATION BY USING A 2D DFT
By defining for = 1, 2, . . . , L,

(t)
(t)
(r)
(r)
= D sin , = D sin
,
(11)
the element (n, m) of the channel matrix H can be written as

L
1 jn(r) jm(t)

g e e
,
[H]n,m =
L =1
(12)
with n = 0, 1, ..., N 1, m = 0, 1, ..., M 1. Note that

(12) is the sum of L two dimensional complex modes, each
(t)
(r)
of phase and , and amplitude g . Hence, by using
the 2D DFT of the channel matrix we can estimate the
mode parameters. Actually, we perform the estimate of the
parameters by using the ND MD submatrix
= [H]
H
0:ND 1,0:MD 1 ,
(13)
where ND and MD , with ND N and MD M , are

two suitable integer parameters to be chosen depending on
can be
the length NT S of the training sequence. Matrix H
obtained by stacking the following NRF MRF submatrices
pN :(p+1)N 1,qM :(q+1)M 1 ,
(p,q) = [H]
H
RF
RF
RF
RF
(14)
with p
=
0, 1, ..., ND /NRF 1, and q
=
(p,q) we send
0, 1, ..., MD /MRF 1. In turn, to estimate H
(m)
a training sequence of vectors xBB , m = 0, 1, ..., MRF 1,
(m)
where xBB has 1/ MRF in position m and zero otherwise.
For each vector of this sequence we design the analog
(q)
transmit beamformer FRF in order to switch on only transmit
antennas qMRF , ..., (q + 1)MRF 1, i.e., we set
0(qMRF )MRF
(q)
RF
FRF =
(15)
,
F
0(M (q+1)MRF )MRF
RF is a MRF MRF square matrix, whose elwhere F
ements are chosen with unitary magnitude and such that
RF has full rank. At the same time we design the analog
F
receive beamformer by switching on only receive antennas
pNRF , ..., (p + 1)NRF 1, i.e., we set
0(pNRF )NRF
(p)
RF
URF =
(16)
,
U
0(N (p+1)NRF )NRF
RF is a NRF NRF square matrix, which, similarly
where U
to FRF , is full rank and has elements with unitary magnitude.

From (1), the base-band signal at the receiver becomes
H

(p,q,m)
(p)
(q) (m)
HFRF xBB + n(m)
= URF
yBB
(17)
(p,q)

H
RF x(m) + n
(m) ,
=U
F
RF H
BB
with m = 0, 1, ..., MRF 1, and where
(m)
= [n
(m)
(p,q)
(p,q,0)
]pNRF :(p+1)NRF 1 .
(p,q,1)
, yBB
(p,q,MRF 1)
, ..., yBB
(20)
(p,q) + n
=H
(0) (1)

RF 1 .
= MRF n
, ..., n
(MRF 1) F
,n
where n
RF we
RF and U
To guarantee the full rank of matrices F
use the Hadamard matrices [10, Ch. 6]. This guarantees also
that the noise in (20) is still white with variance n2 . If we
probe all values of p and q in (20) we obtain an estimate H
of H.
To estimate the parameters of the L complex modes we
on N
compute the 2D DFT of H
DF T NDF T samples, which
can be written as

M
D 1 N
D 1
1
j2 N im + N kn
DF
T
DF
T
W (k, i) =
[H]n,m e
MD ND m=0 n=0
(21)
for k, i = 0, 1, ..., NDF T 1.
Let us define

1
, 0 m < MD , 0 n < ND ,
(22)
d(n, m) = MD ND
0,
otherwise,
and its DFT

j N2k
D(k, i) = e
DF T
j N 2i
where sincN (x) =

that we can write
DF T

kND
NDF T

MD 1
iMD
2
sincMD
,
NDF T
ND 1
2
sincND
sin(x)
N sin(x/N ) .
(t)
=
2i
,
NDF T
(18)
],
(19)
(23)
If NDF T is large enough such

(r)
2k
,
NDF T
(24)
for two integers i and k , from (12) and (20) we can rewrite
(21) as
L
1
W (k, i) =
g D(k k , i i ) + N (k, i) , (25)
L =1

2
n
.
where N (k, i) CN 0, ND M
D
We estimate the channel parameters of the different modes
by using the iterative cancellation method outlined in Tab. I. In
detail, by indicating with W (k, i) the 2D DFT of the channel
matrix obtained by removing the first 1 estimated modes, at
iteration we first estimate the indexes (k , i ) that maximize
|W (k, i)|, i.e.,
(k , i ) = arg max |W (k, i)| .
By defining
YBB = [yBB
(p,q) from (17) by

we can estimate H

1
(p,q)
(p,q) = M
1
H
YBB (F
H
RF URF
RF )
k,i
(26)
Then, the angle of departure and arrival of the -th ray can be
computed from (11) and (24) as

2
i
2
k

(t)
(r)
, = sin1
,
= sin1
DNDF T
DNDF T
(27)
TABLE I
C ANCELLATION METHOD TO ESTIMATE THE CHANNEL PARAMETERS .
Input:
1.
2.
3.
4.
5.
6.
TABLE II
A LGORITHM FOR THE DESIGN OF FBB AND UBB .
Input:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
W (k, i), k, i = 0, 1, ..., NDF T 1

W1 (k, i) = W (k, i)
for = 1 to L
(k , i ) = arg max
k,i |W (k, i)|
g = W (k , i ) L
W+1 (k, i) = W (k, i) 1L g D(k k , i i )
end for
whereas the complex gain turns out to be g = W (k , i ) L

from (25).
In practice also the number of modes L should be estimated
by stopping the cancellation method when |
g | turns out much
smaller than |
g1 |. However, here we simply assume that L is
known.
Note that this method needs a training sequence of

ND
MD
NT S =
(28)
MRF
NRF
MRF
time slots. Moreover, the complexity of determining W (k, i)
2
in (25) is proportional to NDF
T log2 NDF T , due to the 2D
DFT on NDF T NDF T points.
Methods exist that estimate parameters of signal (20) with
better performance [11]. However, it is our feeling that they
are quite complex to be implemented in a dynamic radio
environment.
IV. B EAMFORMER D ESIGN
(t)
(r)
After estimating the channel parameters , , g , =
1, 2, . . . , L, from (5) we reconstruct the channel matrix, here
H / L. As the method proposed in

rH
gA
= A
denoted as H
t
[5] to design the analog and digital beamformers is quite
complex since it requires as target design the singular value
decomposition (SVD) of the estimated channel matrix H

which can be quite large, we propose here a much simpler
iterative algorithm, which is an example of application of the
method presented in [12]. First, by extending the approach in
t and
[5], the analog beamformers are obtained by stacking A
Ar , i.e.,
t, A
t , ..., A
t ]0:M 1 .
FRF = [A
RF
r, A
r , ..., A
r ]0:N 1 .
URF = [A
(29)
RF
Then, we design the digital beamformers fBB and uBB by

applying the method developed in [12] to the equivalent digital
channel matrix
(30)
G = UH
RF HFRF ,
whose dimension is only NRF MRF . In detail, the optimization problem can be written as
2

(fBB , uBB ) = arg max uH Gf
(31a)
f,u
s.t.
FRF , URF
H,
G = UH
RF HFRF
uBB = N1 [1, 1, ..., 1]T
RF
for i = 1 to NI
hM ISO = uH
BB G
H
fBB = hM ISO /||hM ISO ||
hSIM O = GfBB
uBB = hSIM O /||hSIM O ||
end for
fBB = fBB /||FRF fBB ||
uBB = uBB /||URF uBB ||
FRF f = 1 ,
URF u = 1 .
(31b)
We solve problem (31) by using the iterative solution exposed

in Table II, which requires NI iterations, with NI a design
parameter.
V. N UMERICAL R ESULTS
In this section we evaluate the performance of the proposed
channel estimate in a typical 60 GHz channel with L = 3 rays.
The Rx and Tx ULA arrays are made of antenna elements
separated by D = /2. The angle of arrival and departure
are random uniformly distributed in the interval between
min = 60 and max = 60 . Performance, in terms of
, are evaluated by averaging (9) for 5000 realizations of the
channel. The analog and digital beamformers are designed by
using the algorithm described in Section IV with NI = 7.
Fig. 2 shows the average SNR improvement versus
the number of antennas M = N with NDF T = 1024,
MD = MRF = ND = NRF = 16 (NT S = 16 from (28)),
and four values of AW GN . For a comparison we also report
the performance bound obtained by assuming that the channel
is perfectly known. We observe that performance using the
proposed channel estimate is close to the bound already with
AW GN = 0 dB.
In Fig. 3 we consider NDF T = 1024, AW GN = 5 dB, and
we plot the average SNR improvement versus the number of
antennas for four increasing values of MD = ND (which, from
(28), correspond to longer training sequences). We observe that
higher values of M (N ) require higher MD (ND ) to lower the
noise level in the channel estimate. Moreover, by plots similar
to those of Fig. 3 for different values of MRF = NRF , it
is seen that, when the training sequence is sufficiently long,
performance close to the bound can be achieved with few RF
chains.
In Fig. 4 we then investigate the effect of NDF T on the
2D estimate approach. Hence, we report versus the number
of antennas for four values of NDF T , with AW GN = 5 dB
and MD = MRF = ND = NRF = 16. We can see that
for a higher number of antennas also NDF T must increase,
40
35
35
30
30
[dB]
[dB]
40
25
known channel
AW GN = 5 dB
AW GN = 0 dB
AW GN = 5 dB
AW GN = 10 dB
20
15
20
40
60
80
100
120
M =N
known channel
MD = 16
MD = 12
MD = 8
MD = 4
[dB]
35
20
40
20
15
20
40
60
80
100
120
Fig. 4.
vs. M (and M = N ) for four values of NDF T : MD = MRF =
ND = NRF = 16, and AW GN = 5 dB.
R EFERENCES
25
20
25
proportional to the number of used RF chains, and b) taking

the 2D DFT of the received signals.
30
15
channel
= 1024
= 512
= 256
= 128
M =N
Fig. 2.
vs. M (and M = N ) for four values of AW GN : NDF T = 1024,
and MD = MRF = ND = NRF = 16 (NT S = 16).
40
known
NDF T
NDF T
NDF T
NDF T
60
80
100
120
M =N
Fig. 3.
vs. M (and M = N ) for four values of MD = ND : NDF T =
1024, AW GN = 5 dB, and MRF = NRF = 4.
corresponding to a better accuracy in the angle estimate,

otherwise performance deteriorates.
VI. C ONCLUSIONS
We considered the problem of channel estimate for MMW
systems with an hybrid implementation of the beamformer.
In this setup we developed an algorithm based on a training
sequence of beamformers which estimates the parameters of
the MMW channel by using a 2D DFT of the received training
samples. We evaluated the system performance in the presence
of channel estimate and it is seen that our method has performance close to that with the known channel. Furthermore,
our channel estimate has a low computational complexity
mainly because it just requires a) a training sequence of length
[1] J. G. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. C. K. Soong,

and J. C. Zhang, What will 5G be?, IEEE J. Sel. Areas Commun., vol.
32, no. 6, pp. 10651082, Jun. 2014.
[2] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and
F. Aryanfar, Millimeter-wave beamforming as an enabling technology for
5G cellular communications: theoretical feasibility and prototype results,
IEEE Commun. Mag., vol. 52, no. 2, pp. 106113, Feb. 2014.
[3] S. Rangan, T. S. Rappaport, and E. Erkip, Millimeter-wave cellular
wireless networks: potentials and challenges, Proc. of the IEEE, vol. 102,
no. 3, pp. 366385, Mar. 2014.
[4] S. Hur, T. Kim, D. J. Love, J. V. Krogmeier, T. A. Thomas, and A. Ghosh,
Millimeter wave beamforming for wireless backhaul and access in small
cell networks, IEEE Trans. Commun., vol. 61, no. 10, pp. 43914403,
Oct. 2013.
[5] O. El Ayach, R. W. Heath, S. Abu-Surra, S. Rajagopal, and Z. Pi, Low
complexity precoding for large millimeter wave MIMO systems, in Proc.
IEEE Int. Conf. Commun. (ICC), Ottawa (Canada), Jun. 2012.
[6] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Channel
estimation and hybrid precoding for millimeter wave cellular systems,
IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831846, Oct. 2014.
[7] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S. Rappaport,
and E. Erkip, Millimeter wave channel modeling and cellular capacity
evaluation, IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 11641179,
Jun. 2014.
[8] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Single-sided
adaptive estimation of multi-path millimeter wave channels, in Proc. IEEE
Int. Workshop on Signal Process. Advances in Wireless Commun. (SPAWC),
Toronto (Canada), Jun. 2014.
[9] D. Ramasamy, S. Venkateswaran, and U. Madhow, Compressive adaptation of large steerable arrays, in Proc. IEEE Information Theory and
Applications Workshop (ITA), San Diego (CA), Feb. 2012.
[10] N. Benvenuto and G. Cherubini, Algorithms for Communications Systems and their Applications. John Wiley & Sons, 2002.
[11] I. Maravic and M. Vetterli, Sampling and reconstruction of signals
with finite rate of innovation in the presence of noise, IEEE Trans. Signal
Process., vol. 53, no. 8, pp. 27882805, Aug. 2005.
[12] X. Zheng, Y. Xie, J. Li, and P. Stoica, MIMO transmit beamforming
under uniform elemental power constraint, IEEE Trans. Signal Process.,
vol. 55, no. 11, pp. 53955406, Nov. 2007.
On the Mutual Orthogonality of Millimeter-wave

Massive MIMO Channels
Sinh Le Hong Nguyen1 , Katsuyuki Haneda1 , Jan Jarvelainen1 , Aki Karttunen1 , and Jyri Putkonen2
1
Aalto University School of Electrical Engineering, Espoo, Finland, sinh.nguyen@aalto.fi

2
Nokia, Espoo, Finland, jyri.putkonen@nsn.com
AbstractMutual orthogonal user channels in multiuser (MU)multiple-input multiple-output (MIMO) systems are desirable
and can be approximately obtained under independent and
identically distributed (i.i.d.) Rayleigh fading assumption with
a very large number of base station antennas. However, it
has been shown that at millimeter-wave (mmW) frequencies,
this assumption is not valid due to the limited number of
multipath components and spatial channel correlation. In this
paper, we examine the mutual orthogonality of a realistic 60GHz outdoor propagation channel with practical large antenna
arrays, and determine the factors deciding it based on the channel
data generated by means of deterministic field prediction. The
results obtained reveal relationships between mutual orthogonality, inter-user distance, number of active users, transmit array
dimensions, and downlink system capacity at 60-GHz band,
which are useful for designing future mmW massive MU-MIMO
systems.
I. I NTRODUCTION
Millimeter-wave (mmW) massive multiple-input multipleoutput (MIMO) is a key technology for the upcoming 5G
communication networks, as this combination enjoys both
huge signal bandwidth at mmW frequencies and spatial multiplexing capability of large antenna arrays. One of the fundamental features that makes the massive MIMO attractive
is the asymptotic pairwise orthogonality between the channel
vectors under identically distributed (i.i.d.) Rayleigh fading
assumption. This feature allows multiple data streams spatially
multiplexed at a base station to be transmitted to multiple user
terminals simultaneously with just simple and linear precoding
techniques. The mutual orthogonal channels also provide the
optimal performance in terms of transmit power consumption
or sum-rate channel capacity [1].
When it comes to realistic design of multiuser (MU)-MIMO
systems operating at mmW frequencies, however, many practical considerations need to be accounted for to achieve
such desirable performances. One of the major issues is that
the number of independent multipath components (MPCs) at
mmW frequencies is typically limited, and that the vector
channels are not i.i.d. Rayleigh but rather correlated fading
[2]. As a result, making the number of antennas very large
does not necessarily guarantee orthogonal channels between
the users, especially when the users are in close proximity.
Another issue is that the number of Tx antennas in reality is
finite as it is bounded by the compact form of the antenna array
size. Increasing the number of Tx antennas without increasing
the array dimensions creates the mutual antenna coupling or
antenna correlation. Since neither the channels are i.i.d. nor
the number of Tx antennas is infinite or made very large, the

mutual orthogonal condition can not be perfectly satisfied in
reality, and hence the optimal system performance can not be
always achieved.
In this paper, we study the non-orthogonality of the mmW
massive MIMO channel under a realistic outdoor propagation
environment, and analyze the parameters that decide the spatial
correlation and the overall system performance. They include
the Tx antenna array size and antenna number, the inter-user
distance between two users, and the number of users that could
be simultaneously active/served. The realistic channel data for
this study is based on directional channel measurements at 60
GHz in an open square in downtown Helsinki, Finland (Fig. 1)
and a deterministic field prediction tool. The results obtained
from this work are useful for designing future mmW MUMIMO systems especially in urban outdoor propagation environments, in which multiple antenna design, transmit/receive
beamforming, and multiuser scheduling are major techniques
that need to be considered.
II. C HANNEL G ENERATION BY M EANS OF
D ETERMINISTIC F IELD P REDICTION
A. Point cloud-based channel propagation prediction
Fig. 1. Photo of the measurement location in Narinkkatori, Helsinki, Finland.
In order to obtain the channel data for the environment of

interest, we use a point cloud-based propagation prediction
method calibrated with measured channel data. The propagation prediction method utilizes laser scanning to obtain
accurate descriptions of the propagation environments in the
form of a point cloud [3]. The prediction method relies on
a single-lobe directive scattering mode [4] to calculate the
backscattering from each point in the point cloud, and the
contributions coming from distinct points are combined to give
978-1-4799-8088-8/15/$31.00 2015 IEEE
Fig. 2.
Illustration of the laser scanned point cloud and the simulated 2 Tx locations and 648 Rx locations.
the total field. In particular, we assume that the field consists

of a LOS path along with single-bounce scattering from the
point cloud. The effect of shadowing caused by blocking
objects in the environment is taken into account by assuming
that points found inside the Fresnel zone induces additional
attenuation to a path [5]. The point cloud obtained from the
laser scans for the open square is illustrated in Fig. 2. The
square is approximately 100100m2 , surrounded by buildings
on four sides. In addition to the buildings and the ground, the
propagation environment is affected by lamp posts and people.
The people cause the line-of-sight (LOS) to be obstructed, i.e.,
obstructed-line-of-sight (OLOS) in some places.
The scattering model contains two parameters, a scattering
coefficient S and a scattering lobe width R , which relate
to the material properties of the local surface. Since it is
not feasible to obtain material parameter values for all local
surfaces, we calibrate S and R such that the predicted
and measured root-mean-square (rms) delay spread agree as
well as possible using the same parameter values for all
points in the prediction of a single channel. In the open
square, the scattering model parameters were found to be
S = 1.0, R = 50. The propagation paths are identified as the
peaks in the power delay profile (PDP). Amplitude, delay, and
azimuth and elevation angles are obtained for each multipath
for both the Tx as well as for the Rx.
B. Massive MU-MIMO channel generation
To generate the MU-MIMO channel, an N N uniform
planar array (UPA) is created at each Tx location. Each UPA
is a horizontal planar of M = N 2 half-wave vertical dipoles
with D antenna spacing in both dimensions. The planar array
is of size (N 1)D (N 1)D at each Tx. For routes 1-4,
the broadside of the Tx planar array points to 135 (see Fig.
1). For routes 5-8, the broadside of the Tx planar array points
to 315 . Each route has 81 Rx locations. For the simplicity,
each Rx is assumed to have only one omni-directional antenna.
Using the classical multipath model, the M 1 channel vector
between a Tx and Rx k on a route is given by
hk =
Lk

l=1
ejkl gkl aH (kl , kl ) aV (kl ),
(1)
where denotes the Kronecker product, and aH and aV are

the phase response vectors for horizontal and vertical linear
arrays to path l in the form
1
D
j2 sin cos
1
e
aH (, ) =
,
..
Lk
.
ej2
(N 1)D
sin cos
T
1
j2 D cos
(N 1)D
aV () =
ej2 cos ,
1e
Lk
respectively. In (1), k denotes the large-scale fading coefficient accounting for path loss and shadowing of channel hk ,
gkl CN (0, 1) denotes the complex amplitude of path l of
user k, (kl , kl ) are the elevation and azimuth angles of arrival at Tx of path l originating from user k, and Lk is the total
number of MPCs from user k to the Tx. kl [/2, /2]
is the uniformly distributed random phase for the the smallscale fading of different MPCs, and independent to the user
and MPC indices.
TABLE I
C ORRELATION DISTANCE ( IN METERS ) OF LARGE - SCALE PARAMETERS
LS parameter
DS
ASD
ASA
ESD
ESA
SF
K
LOS
1.3
23
1
15
17
7.8
7.9
OLOS
1.4
3.2
8.1
1.8
0.8
3.6
N/A
By (1) we assume the plane wavefront and the wide-sense

stationary (WSS) of the channel over the Tx antenna array. By
assuming that, all of Lk propagation paths for a particular user
k to the Tx are subject to the same large-scale fading across
all Tx antennas. All cluster/scatterer parameters including path
loss (PL) and shadow fading (SF), azimuth and elevation
spreads of departure and arrival (ASD, ASA, ESD, and ESA),
and K-factor do not vary along the antenna locations on the
MPCs of Tx1Rx50 and Tx1Rx57 links on route 3
MPCs of Tx1Rx10 and Tx1Rx30 links on route 3
Common scatterers
Tx1
Tx1
Rx57
Rx30
Rx10
10
5
Rx50
0
5
10
15
16
14
12
10
0
5
10
10
15
15
20
Fig. 3. Visual comparison of MPCs in the open square scenario with (left) no common scatter when two users are well separated, and (right) three common
scatterers when two users are 0.8 m spaced.
array axes. These assumptions are reasonable as long as the

array dimensions are smaller than the correlation distance of
the large-scale parameters obtained from point cloud channel
data as in Table I, and when the Tx-Rx link distance is in the
far-field range. The following results and analysis are applied
for the OLOS cases, which have the shortest correlation
distance of 0.8 m considering all the large-scale parameters.
III. M UTUAL O RTHOGONALITY OF MU-MIMO
P ROPAGATION C HANNEL
Let H = [h1 h2 . . . hK ] denote the M K composite
MU-MIMO channel matrix. Mutual orthogonal channel requires every pair of column vectors of H satisfies
hH
i hj = 0 for all i = j.
(2)
Under the i.i.d. Rayleigh channels assumption (i.e., the NLOS

rich scattering environment), hk CN (0, k IM ), condition
(2) is asymptotically achieved with very large Tx antenna
array, by the law of large numbers
1 H
h hj 0 as M for i = j.
(3)
M i
Denote (H) be the condition number of H, which is defined
as the ratio between the largest singular value max and
smallest singular value min of H, i.e.,
max
.
(H) =
min
If all the user channel vectors are normalized to have the same
norm, e.g. hk 2 = M , the power imbalance due to largescale fading between the user channels is removed, and the
condition number of the i.i.d. MU-MIMO channels is equal to
1. This ideal propagation condition provides optimal performance for MU-MIMO channels: downlink spatial streaming
can achieve maximum sum-rate, by using only simple linear
precoding methods at Tx such as MF to decompose the
channel to multiple independent spatial interference-free subchannels.
The mutual orthogonal condition, however, can not be
satisfied in reality, especially in our interested mmW outdoor
propagation channel, for practical reasons previously discussed
including spatial correlation and compact form of the Tx
arrays. The spatial correlation between user channel vectors
occurs in our experiments where the users are close to each
other and hence their MPCs often belong to the same cluster
of scatters. In other cases, spatial correlation is also introduced
by common scatterers (walls in street canyon, high buildings
in city center, etc.) seen by different users even when they are
not well co-located [6], [7]. Fig. 3 shows the MPCs obtained
from point cloud channel data for two scenarios of no common
scatterer and three common scatterers when the inter-user
distance are 8 m and 2.8 m, respectively.
Beside the the richness of the scattering and the distribution of the scatterers in the propagation environment, other
factors that affect the mutual orthogonality of H include the
number of Tx antennas and the antenna spacing, and the interuser distance. One way to measure the amount of lack of
orthogonality or spatial correlation is examining the condition
number (H) that quantifies the spread in magnitude of the
singular values of H. The smaller and closer to 1 the condition
number is, the better suited the channel is to support spatial
multiplexing and the better performance of the precoder that
results in higher capacity. When considering only two users,
the pairwise orthogonality between two channel vectors hi and
hj can be measured by the normalized scalar product
ij =
1 H
|hH
i hj |
=
|h hj |.
hi hj
M i
(4)
The smaller and closer to zero this product is, the smaller
the downlink inter-user interference between user i and j
is created using the MF precoder. In the next section, we
analyze the impact of mutual orthogonality on the downlink
capacity of the MU-MIMO channels associated with MF and
ZF precoders, respectively.
IV. I MPACTS OF MUTUAL ORTHOGONALITY ON DOWNLINK
CHANNEL CAPACITY
Consider the above described MU-MIMO system operating

under the narrow-band channel. With the assumptions of timedivision duplexing (TDD) and channel reciprocity, the received
signal at user k is given by

hTk wj sj + nk , (5)
yk = hTk x + nk = hTk wk sk +
j=k
where nk is the additive white Gaussian noise (AWGN) at

user k, nk CN (0, 2 ). In (5), s CK is the normalized
information vector of length K satisfying E{ssH } = IK . x
K
E{log2 (1 + k )},
k=1
where k is the signal-to-interference-plus-noise ratio (SINR)

at user k, computed as
|hTk wk |2
.
T
2
2
j=k |hk wj | +
k =
(6)
The precoding matrix W is designed to maximize certain performance metric of MU-MIMO system. As we are interested
in system capacity performance, two sub-optimal but more
practical linear precoding methods are considered: matched
filtering (MF) and zero-forcing (ZF).
A. Matched filtering (MF)
MF is the simplest and least complex precoding technique,
in which each precoding vector wk is designed to maximize
the intended users receive energy. That is,
wk = pk hk / hTk ,
K
where pk is the power allocated to user k, k=1 pk = P , the
total power available at the Tx. With this notation, the received
SINR at user k in (6) becomes
pk
k =
2 p + 2 /M .
j=k kj j
In this method, since we aim to maximize the received signal
power solely, the inter-user interference is not accounted for
and hence there is some SNR loss in the case of nonorthogonal channels (kj = 0) as compared to asymptotically
orthogonal ones (i.d.d. channels with very large M ). The
asymptotic sum-rate for i.i.d. channels is
MP
) as M ,
K 2
if all the users are allocated the same power pk = P/K.
Riid, ZF K log2 (1 +
B. Zero-forcing (ZF)
In ZF precoding, the precoding matrix at the Tx is the
pseudo-inverse of the downlink channel matrix

1

,
W = HT = H HT H
where is the normalization parameter that guarantees
the average total transmit power of P , i.e., E{ x 2 } =
E{tr(WH W)} = P. For each random channel realization,
the normalization is computed as

1
.
(7)
= P/E tr HT H
In the ideal scenario of i.i.d channels, all the user channel vectors are mutually orthogonal at the limit of infinite number of
Tx antennas, and hence the inter-user interference approaches
The SINR has the same value of k = / 2 = M P/(K 2 )

across all the users, and the average achievable sum-rate in
this ideal scenario using ZF precoding is

Riid, ZF = KE log2 (1 + 2 )
MP
) as M ,
K log2 (1 +
K 2
which is equal to the asymptotic sum-rate Riid, ZF (The two
precoders are identical when all columns of H are mutually
orthogonal).
In our practical scenario, however, the vector channels are
not always perfectly mutually orthogonal, neither uncorrelated.
The presence of strong spatial correlation induces situations
when the channel is ill-conditioned, reducing or the SNR at
the users. In those case, a loss in the receive SNR and hence
the sum-rate with respect to i.i.d. channels.
V. N UMERICAL R ESULTS AND D ISCUSSIONS
In the simulation, for each experiment 500 small-scale
channel realizations for each Tx-Rx pair are generated for
a total of 648 pairs at different Tx and Rx locations. Our
first interest is how the amount of correlation of the two user
channel vectors varies along the inter-user distance and the Tx
antenna array size under the realistic propagation environment.
Fig. 4 shows the average number of common scatters (left)
and the average correlation coefficient (right) over all TxRx pairs of all 8 routes of the experiment versus the interuser distance. The antenna spacing of UPA at each Tx is
D = 2, so the largest array dimensions are 38 38
or approximately 0.20.2 m2 . It can be observed the clear
dependency of the number of common cluster and spatial
channel correlation with respect to the inter-user distance.
When the number of Tx antennas exceeds 1616, increasing it
does not necessarily decrease the spatial correlation by much.
We can also observe that on the average, with 1010 UPA
and D = 2 the correlation becomes lower than 0.1 when the
inter-user distance is greater than 16 m (approximately twice
of OLOS shadowing correlation distance).
8
Correlation coefficient
R=
zero. In this case, is also the receive energy at each user since
the MU-MIMO channel can be decomposed as K independent
channels as (5) becomes
yk = sk + nk .
Average no. of
common scatterers
is the transmit (precoded) vector of length M , related to s by

the precoding matrix W CM K as x = Ws.
Assuming that the CSI is perfectly known at the Tx, then
the average achievable sum-rate of all K user is given by
6x6 array
10x10 array
16x16 array
20x20 array
0.5
4
2
0
0
10
20
30
Interuser distance (m)
0
0
10
20
30
Fig. 4.
Average number of common scatterers (left) and correlation
coefficient with D = 2 (right) over all routes versus inter-user distance.
1
0.8
CDF
Fig. 5 shows the scatter plot of average achievable sum-rate

of 2 users when Tx is equipped with 10 10 UPA, D = 2,
versus the inter-user distance. The total power P = 10 dB
and noise power 2 = 1/10 dB (high SNR regime). It can
be seen that the larger distance between the two users, due to
the decreasing of the correlation, the closer of sum-rate of the
MF precoder to that of the ZF precoder. When this distance
is larger than 16 m, the average sum-rate (not shown here) of
ZF achieve 80% of the asymptotic i.i.d. sum-capacity.
0.6
D = 0.5
4 users
8 users
12 users
16 users
0.4
0.2
Sumrate (bits/channel use)
20
0
0
10
20
30
40
Condition number
50
60
15
Fig. 6. C.d.f. of condition number of H for different number of users, 1010
UPA with D = 0.5 at Tx.
10
10
20
0.8
30
Fig. 5. Sum-rate versus inter-user distance, 1010 UPA with D = 2,

P = 10 dB, 2 = 1/10 dB.
When K 2, each Tx with a 10 10 UPA with D = 0.5

chooses to serve 4Kr Rxs from 4 routes (Kr Rx locations with
a maximum separation on each route). Fig. 6 shows the cumulative distribution function (CDF) of the condition number of
H for K = {4, 8, 12, 16} users. It can be seen that the wellconditioned 4-user channel is capable of supporting 4 parallel
paths for spatial multiplexing. The channel becomes more illconditioned as K increases, suggesting that the transmission
of higher numbers of spatial streams is less feasible. Indeed, it
can be observed from Fig. 7 that at P = 5 dB and 2 = 1, the
sum-rate with 4 users using MF is close to that using ZF and
of i.i.d. case, since the channel matrix is nearly orthogonal.
At K = 16, both ZF and MF precoders have far smaller sumcapacity as compared to the i.i.d. case, since the channel now
is ill-conditioned indicated by large condition number (H).
In this low SNR regime, MF performs poorly as the signal
energy in (7) at each
1 user is low, due to the small singular
.
values of HT H
VI. C ONCLUSIONS
We have studied the mutual orthogonality of mmW massive
MU-MIMO channels and its impact to the spatial multiplexing
capability. The results obtained have invalidated the i.i.d. channels assumption as well as the asymptotic mutual orthogonal
channels of mmW massive MU-MIMO. The dependency of
the channel mutual orthogonality and system capacity on
the inter-user distance, number of active users, and antenna
array dimensions at 60-GHz has also been analyzed. This
study has revealed that the number of active users in mmW
CDF
0
0
i.i.d
ZF
MF
0.6
0.4
0.2
0
0
ZF, K = 4
MF, K = 4
i.i.d., K = 4
ZF, K = 16
MF, K = 16
i.i.d., K = 16
20
40
60
Sumrate (bits/channel use)
80
Fig. 7. C.d.f. of sum-rate for different number of users, Tx has 1010 UPA
with D = 2, P = 5 dB, 2 = 1.
should be small as compared to microwave systems. The users

should also be well-separated, in a range of at least two times
of correlation distance of shadowing, for an efficient spatial
multiplexing with linear processing precoders.
R EFERENCES
[1] T. L. Marzetta, Noncooperative cellular wireless with unlimited numbers
of base station antennas, IEEE Trans. Wireless Commun., vol. 9, no. 11,
pp. 35903600, Nov. 2010.
[2] A. L. Swindlehurst et al., Millimeter-wave massive MIMO: the next
wireless revolution? IEEE Commun. Mag., vol. 52, no. 9, pp. 5662,
Sep. 2014.
[3] J. Jarvelainen and K. Haneda, Sixty gigahertz indoor radio wave propagation prediction method based on full scattering model, Radio Sci.,
vol. 49, no. 4, pp. 293305, Apr. 2014.
[4] V. Degli-Esposti et al., Measurement and modelling of scattering from
buildings, IEEE Trans. Antennas Propag., vol. 55, no. 1, pp. 143153,
Jan. 2007.
[5] J. Jarvelainen et al., 70 GHz radio wave propagation prediction in a
large office, in Proc. Loughborough Ant. Prop. Conf., Loughborough,
UK, Nov. 2014.
[6] A. Adhikary et al., Joint spatial division and multiplexing for wave
channels, IEEE J. Select. Areas in Commun., vol. 32, no. 6, pp. 1239
1255, Jun. 2014.
[7] J. Poutanen et al., Multi-link MIMO channel modeling using geometrybased approach, IEEE Trans. Antennas Propag., vol. 60, no. 2, pp. 587
596, Feb 2012.
mmWave Phased Array in Mobile Terminal for 5G

Mobile System with Consideration of Hand Effect
Kun Zhao1,2, Jakob Helander2,3, Zhinong Ying2, Daniel Sjberg3, Mats Gustafsson3 and Sailing He1
1. Department of Electromagnetic Engineering, KTH Royal Institute of Technology, Stockholm, Sweden
2. Network Technology, SONY Mobile Communications AB, Lund, Sweden
3. Department of Electrical and Information Technology, Lund University, Lund, Sweden
Abstract In this paper, the latest study of the millimeter
wave (mmWave) phased array in mobile terminals for the 5G
communication is presented. New parameters for evaluating the
mmWave phased array in mobile terminals are introduced.
Furthermore, the hand effect on the phased array in mobile
terminal at 15 GHz is also investigated.
KeywordsmmWave, 5th Generation Mobile System, mobile
terminal, phased array, hand effect
I. INTRODUCTION
Due to the shortage of the frequency spectrum below 6
GHz, the mmWave (10 GHz to 300 GHz) band has been seen
the potential carrier frequency for the 5th generation mobile
communication (5G) [1]-[2]. However, from the Friis formula
(1), we can see the free space path loss will be much higher in
mmWave band than current mobile networks due to the shorter
wavelength. Consequently, antenna gains of both in base
stations and mobile terminals needs to be increased to
compensate the higher path loss without consuming any more
power.
20 log
(a)
60
(1)
For mobile terminals, thanks to the small electrical length

of mmWave antennas, antenna arrays can be implemented in
terminals to achieve a high gain. However, the high gain
directly translates into a relatively narrow beamwidth, which
will reduce the space coverage of terminal antennas, and
increase the possibility of poor links [3]. Therefore, the beam
steering function for arrays becomes necessary to maintain a
good link when incoming signals come from different angles
[4]-[5], and this function can be realized by phased arrays.
As the beamsteering of phased arrays have a limited
scanned angular range, it is important to characterize the
achievable space coverage of such a system. In [5] the total
scanned pattern and the coverage efficiency are introduced.
The total scan pattern is regained from all array patterns
corresponding to different phase shifts, by drawing out the best
achievable gain at every angular distribution point ( , ) (Fig.
1(a)). For example, in Fig. 1(b), a total scanned pattern of a
four elements antenna array in a mobile terminal with
continuously phase shifts from -180 to 180 is plotted.
Comparing to the radiation pattern of fixed phased arrays, the
total scan pattern can show all the possible space coverage.
(b)
Fig. 1. (a) Gain threshold level G
and concept of total scan pattern of
phased array gain (pictures from [5]) (b) the total scanned pattern of a phased
array comparing to
Furthermore, we bring up the coverage efficiency to

quantify the space coverage of a phased array, which is defined
as the ratio between the covered area and the total area.

(2)
The coverage area can be defined as the total scan pattern

coverage with respect to a threshold gain level. The total area
includes the whole surrounding sphere and the threshold gain
will depend on the path loss, transmitting/receiving power level
and transmitting antenna gain according to (1).
The coverage efficiency of some typical phased arrays in
mobile terminals has been studied in [3]-[4]. In this paper,
978-1-4799-8088-8/15/$31.00 2015 IEEE
Anntenna arrays att 15GHz are studied since th

his frequency has
h
beeen selected as a demo candid
date for the 5G
G communicatiion
by NTT Docom
mo. We furtheer develop the antenna arrray
connfiguration an
nd bring abou
ut the diverssity and MIM
MO
funnctions by usin
ng two subarraays in this pap
per. Moreover, as
moobile terminals mainly operatte when they are
a hold by useers,
thee hand effect on
n the antenna array
a
has also been
b
investigateed.
Alll simulations are done by CST 2014. Since
S
the mob
bile
term
minal antennass will face to arrbitrary directio
ons in reality, the
t
inccoming wave is assumed to
t uniformly distribute in an
isottropic environ
nment. Since base stationss can track the
t
pollarization of terrminal antennaas and optimizee the polarizatiion
of transmitted signals, we assume that no polarizatiion
missmatch exists in
n the system.
(a)
Phasse
Shifteer
Phi 1
2*Phi1
33*Phi1 4*Phi 1
Phi 2
Power Input
2*Phi 2
3*Phi2 4*P
Phi 2
Power Input
(b)
Fig. 2. (aa) the array configuuration and its elem
ments radiation pattterns in a 6inch
mobile teerminal; (b) the scchematic of our phaase array system
II. ANTENN
NA ARRAY
A. Antenna Arrayy Setup
In this study, the phased arrray is composeed by eight nottch
anttennas and sepaarated slightly less than 0/2 at 15 GHz in a 6
incch terminal is studied,
s
which is shown in Fiig. 2(a). All arrray
elem
ments are locaated in a row and on the to
op of the mob
bile
term
minal. Each fo
our elements co
ompose a sub array
a
and a phaase
shift progression that is indepen
ndent of the ad
djacent subarraay,
whhich is shown in
i Fig. 2(b). Th
hus, their beam
ms can be steerred
sepparately, makin
ng the pattern diversity
d
feasib
ble. Furthermo
ore,
it aalso increases the
t degree of freedom
f
for steeering beams and
a
thee coverage efficciency can be increased
i
comp
paring to a sing
gle
arraay with eight elements, wh
hile a trade-offf is made with
w
resppect to maxim
mum achievablee gain (Fig. 3) [4]. Furthermo
ore,
we can switch between
b
the antenna
a
diverssity scheme and
a
MIIMO scheme by two subaarrays based on the chann
nel
sceenarios.
In Fig.4, the beamsteering
b
of
o subarray 1 with
w 0 and 15
50
phaase shift in freee space are presented. Its totaal scanned patteern
from
m +180 to -18
80 phase shiftt can also be fo
ound in Fig. 1(b).
Wee will also present the measurement results in the
t
preesentation.
Fig. 3. Illlustration of patterrn diversity receiving scheme with tw

wo subarrays
SSubarray 2
Sub
ubarray 1
Subarrray 2
((a)
Subarrayy 1
(b)
Fig. 4. A
Array pattern for ssubarray 1 at (a) 00 phase shift andd (b) 150 phase
shift
Subarrray 2 Subarray 1
4
3
2
1
(a)
(b)
(cc)
User cases are studdied in this paper: (a) single hand grripping, (b) dual
Fig. 5. U
hand grippping and (c) the ssingle finger touchhing
E-plane
H-plane
B. Hannd Effect
Thee Hand effectt is always a critical issuee for mobile
terminaal. In order too simulate the hand effect, tthe dielectric
propertty of human haand at 15 GHz must be selectted carefully.
The skiin depth of eleectrical field att 15 GHz on hhuman skin is
around 2 mm, and tthe skin on paalm is the thicckest part on
human body which caan be as thick as 2mm [6]. T
Therefore, the
skin layer plays the most important role in the hand effect at 15
GHz. In our study, for simplicity, a homogenous hand phantom
with skin material is used for simulations. The permittivity of
the skin at 15 GHz is set to 26.40 and the loss tangent to 0.63
[7].
Three user cases are studied: the single hand gripping, the
dual hand hold gripping and the single finger touching, which
are shown in Fig. 4
In the single hand gripping case, coverage efficiencies and
total scanned patterns of subarray 1 and subarray 2 are almost
the same as in the free space. The hand effect at 15GHz is
lower comparing to the current cellular frequency, the reason is
that the effective distance between the hand and the antennas
are larger due to the shorter wavelength.
(a)
In the dual hand case, coverage efficiencies of subarray 1

and subarray 2 are presented in Fig. 7. Unlike the single hand
case, both subarrays show a dramatic degrading of the
coverage efficiency, which most likely is an effect of the close
proximity of the index finger. Furthermore, the total scan
pattern is also changed by the hand phantom; as the main beam
of the subarray is reflected towards the space aligned with the
right hand (
120).

(b)
Fig. 7. (a) Coverage efficiency in dual hand gripping positiona and (b) the
total scanned pattern of subarray 1
(a)
(b)
Fig. 6. (a) Coverage efficiency in the single hand gripping positiona and (b)
the total scanned pattern of subarray 1
Subarray 1
Subarray 2
Fig. 8. Coverage efficiency and total scanned pattern in the single finger
touching position
Since the wavelength is relatively small at 15GHz, which

makes the hand can only effect the array propagation when it is
proximate to the array. Therefore, we design an extreme case
that one of the sub arrays is totally blocked by the index finger,
which is shown in Fig. 5(c). With single finger touching case,
we can see from Fig. 6, when the minimum gain requirement is
larger than 1, the coverage efficiency of subarray 1 is
decreased rapidly, and its total scanned pattern is totally

changed by the finger. However, the subarray 2 is almost not
affected by the hand phantom. In this case, the diversity
technology can be used to select the best subarray in terms of
the signal to noise ratio (SNR).
REFERENCES
[1]
[2]
CONCLUSION
In the paper, the latest study of the mmWave phased array
in mobile terminals for 5G communication is presented. The
recently introduced parameters; the total scan pattern and the
coverage efficiency, are presented with the aim to emphasize
their importance for evaluating phased array antennas in the
mobile terminal. Different antenna array schemes are
introduced and the advantages of using two subarrays are
presented. Furthermore, we have also investigated the hand
effect on the antenna array at 15 GHz. Where we can see that
the hand effect can change the array pattern and total scanned
pattern dramatically when it get close to the array. However,
with two sub arrays configuration, the switch diversity can be
used to against the hand effect.
[3]
[4]
[5]
[6]
[7]
Rappaport, T. S., Sun, S., et al., Millimeter Wave Mobile

Communications for 5G Cellular: It Will Work!, IEEE Access, Vol. 1,
May 2013
Roh, W., Seol, J-Y., et al., Millimeter-Wave Beamforming as an
Enabling Technology for 5G Cellular Communications: Theoretical
Feasibility and Prototype Results, IEEE Communications Magazine,
Vol. 52, No. 2, Feb. 2014
T. Bai and R. Heath, Jr., Coverage and Rate Analysis for MillimeterWave Cellular Networks, IEEE Trans. Commun., vol. 14, no. 2, pp.
11001114, Feb. 2015.
Wonbin Hong; Kwang-Hyun Baek; Youngju Lee; Yoongeon Kim;
Seung-Tae Ko, "Study and prototyping of practically large-scale
mmWave antenna systems for 5G cellular devices," Communications
Magazine, IEEE , vol.52, no.9, pp.63-69, September 2014
Helander, J., Millimeter Wave Antenna Technology in Mobile
Terminal and Station, Masters Thesis, Lund University, July 2014
Gillian S. Ashcroft, Teresa Greenwell-Wild, and Mark W. J. Ferguson
(1999). "Topical Estrogen Accelerates Cutaneous Wound Healing in
Aged Humans Associated with an Altered Inflammatory Response". The
American Journal of Pathology155 (4): 11371146.
C Gabriel, S Gabriel and E Corthout The dielectric properties of
biological tissues: I. Literature survey Phys. Med. Biol. 41 (1996)
22312249.
Human Exposure to mmWave Phased Array

Antennas in Mobile Terminal for 5G Mobile System
1.
Kun Zhao1,2, Zhinong Ying2 and Sailing He1

Department of Electromagnetic Engineering, KTH Royal Institute of Technology, Stockholm, Sweden
2. Network Technology, SONY Mobile Communications AB, Lund, Sweden
AbstractIn this paper, results of the latest study regarding

human exposure to the mmWave phase array in mobile terminal
for 5G mobile communication are presented. FCC guideline has
been reviewed and power density is used to evaluate the human
exposure in the mmWave band. The purpose of the paper is to
present the basic power density property of the mmWave phased
array for terminal applications and the defect of the current
standards of human exposure in the mmWave band. An IFA
antenna phased array at 15GHz is simulated and the result shows
the challenges of power density evaluation for this application.
KeywordsmmWave, 5th Generation Mobile System, mobile
terminal, Human Exposure, Power density
I. INTRODUCTION
The human exposure level of mobile terminals is always a
critical issue for wireless communication systems since the
maximum output power of mobile terminals is limited by the
strength of the human exposure. Today, the human exposure
level of mobile terminals in cellular bands is evaluated by
Specific Absorption Rate (SAR). However, for the 5th
generation mobile network (5G), due to the shortage of
spectrum in current cellular bands (<6GHz), people has begun
to focus on the mmWave band (10GHz to 200GHz) as its rich
spectrums resources [1]-[3]. In the mmWave band, the human
exposure is currently evaluated by the power density [4], but as
the mmWave communication has not implemented in mobile
phones yet, the current guideline is well defined for mobile
terminal applications.
The current power density limitation for public exposure is
10 W/m2 peak value with at least 5 cm away from the devices
in The Federal Communications Commission (FCC) guideline,
which is only valid in the far field region. However, for the
devices like mobile phones which will contact directly on
human body, the near field evaluation of the power density is
also necessary, but the evaluation method and the limitation are
still unknown. The measurement of power densities in the near
field will be more complicated, and in the current FCC
guideline, no measurement method is suggested for near field
power density but computer simulations are recommended.
Another innovation on antennas for 5G is that phased
arrays with high gain and the beam steering function become
necessary due to the high path loss in the mmWave band [5][6]. Comparing to conventional mobile antennas, the power
density evaluation on the phased array is more complicated as
the value is impacted by the angle of phase shift.
exposure in the mmWave band. In this study, we simulate an

IFA type phased array at 15GHz by CST 2014, its power
density is calculated and the peak value of the power density is
presented against the distance to the array and the phase shift
value.
II. RESULT AND DISSCUTION
A. Simulation Setup
A terminal mockup configures with eight identical IFA
elements is shown in Fig. 1(a). All eight elements are placed on
the top of the mockup with the half wavelength distance. Each
four of them compose a sub array [5], and the two sub array
can operate either in diversity or MIMO mode. In this study,
we will only exam the case that sub array 1 transmits and sub
array 2 is terminated. The radiation pattern of sub array 1 with
0 phase shifts at 15GHz is shown in Fig. 1(b), and the peak
gain is 9.6dB in this case.
20mW power is fed into one sub array and equally splitted
into four elements (5mW for one element). The power density
of each cutting plane is shown in Fig. 2. We can see that the
attenuation rate is different with increased distance to the
antenna array.
The field close to the antenna can be divided into three
zones: the reactive near field, the radiative near field and the
transition zone. In current mobile terminals, due to the
thickness of the display and the phone case, there are 2-3mm
gap between the array and human body. At 15GHz, this is a
critical distance as it is close to the maximum reactive near
field region distance, where the power density fluctuates
rapidly.
(a)
(b)
Fig. 1. (a) diagram of the sub array configuration and (b) the raidaiton pattern
of sub array 1 with 0 phase shift at 15GHz
The purpose of the paper is to present the basic power

density property of the mmWave phased array for terminal
applications and the defect of the current standards of human
978-1-4799-8088-8/15/$31.00 2015 IEEE
to that the summation of complex field at each observing point

could be either constructive or deconstructive. Therefore, the
power density value of a phased array needs to be evaluated
under all possible phase shift angles.
(a)
(b)
(c)
Fig. 2. Power density distribution on (a) XZ-plane (b) XY-plane and (c) YZplane
0:
57.8 W/m2
60:
68.5 W/m2
120:
91.9 W/m2
Fig. 4. Peak value of power density against phase shift value of sub array 1 in
Z plane
CONCLUSION
Fig. 3. Peak value of power density against distance of sub array 1 in X, Y

and Z direction
The peak value of power density against the distance in X,

Y and Z directions are plotted in Fig. 3. First, it can be seen
that power density values are much higher than the 10W/m2
limitation within the near field region (<10mm) in all
directions even though our input power is relatively low, which
shows that the current power density limitation is not
applicable for mobile terminal applications. Second, the decay
of power density values on different directions also varies a lot,
and this is related to the direction of the main radiation beam:
we can see the power density is higher on the Y-direction as it
is alone with the main beam; but meanwhile, the power density
on the X-direction drops fast as it is orthogonal to the main
beam. Furthermore, the values of power density changes
extremely fast against the distance, which will make the
evaluation to become difficult in practical.
Another issue is that the peak values of power density also
changes dramatically with the phase shift. The peak values of
power density of sub array 1 against different phase shifts are
shown in Fig. 4. The power density is calculated on the XY
plane and 3mm away from the antenna array. We can see that
the peak value almost double at 120 phase shift comparing to
0 phase shift. The variation of the power density mainly due
In this paper, we reviewed the current FCC guideline of

human exposure in the mmWave band, and implement it to
the future 5G communication. In the current guideline, the
power density evaluation in the farfield region is well defined.
However, the power density in the near field is also critical for
mobile terminal applications since they will contact directly
on human body, and the corresponding measurement method
has not been clear stated yet. We also presented the basic
power density property of a phase array in a mobile terminal
at 15GHz. From the study, we can see that the peak value of
the power density decays fast in the near field region and also
fluctuates rapidly with different phase shift. These properties
make the evaluation of power density for the phased array to
be challenge and complicated, new standards need to be raised
for this special application.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Rappaport, T. S., Sun, S., et al., Millimeter Wave Mobile

Communications for 5G Cellular: It Will Work!, IEEE Access, Vol. 1,
May 2013
Rappaport, T. S., Gutierrez, F., et al., Broadband Millimeter-Wave
Propagation Measurements and Models Using Adaptive-Beam Antennas
for Outdoor Urban Cellular Communication, IEEE Transactions, Vol.
61, No. 4, Apr. 2013
Roh, W., Seol, J-Y., et al., Millimeter-Wave Beamforming as an
Enabling Technology for 5G Cellular Communications: Theoretical
Feasibility and Prototype Results, IEEE Communications Magazine,
Vol. 52, No. 2, Feb. 2014
Federal Communications commission office of engineering &
Technology Evaluation Complianc with FCC Guideline for Human
Exposure to Radiofrequency Electromagentic Fields June 2001
Helander, J., Millimeter Wave Antenna Technology in Mobile
Terminal and Station, Masters Thesis, Lund University, July 2014
Wonbin Hong; Kwang-Hyun Baek; Youngju Lee; Yoongeon Kim;
Seung-Tae Ko, "Study and prototyping of practically large-scale
mmWave antenna systems for 5G cellular devices," Communications
Magazine, IEEE , vol.52, no.9, pp.63-69, September 2014
23rd European Signal Processing Conference (EUSIPCO)
DISTRIBUTED MASSIVE MIMO IN CELLULAR NETWORKS:

IMPACT OF IMPERFECT HARDWARE AND NUMBER OF OSCILLATORS
Emil Bjornson , Michail Matthaiou , Antonios Pitarokoilis , and Erik G. Larsson
Dept. of Electrical Engineering (ISY), Linkoping University, Linkoping, Sweden
ECIT Institute, Queens University Belfast, Belfast, Northern Ireland, U.K.
Department of Signals and Systems, Chalmers University of Technology, Sweden

{emil.bjornson, antonios.pitarokoilis, erik.g.larsson}@liu.se, m.matthaiou@qub.ac.uk
ABSTRACT
Distributed massive multiple-input multiple-output (MIMO)
combines the array gain of coherent MIMO processing with
the proximity gains of distributed antenna setups. In this paper, we analyze how transceiver hardware impairments affect
the downlink with maximum ratio transmission. We derive
closed-form spectral efficiencies expressions and study their
asymptotic behavior as the number of the antennas increases.
We prove a scaling law on the hardware quality, which reveals
that massive MIMO is resilient to additive distortions, while
multiplicative phase noise is a limiting factor. It is also better
to have separate oscillators at each antenna than one per BS.
1. INTRODUCTION
Cellular radio access networks (RANs) have conventionally
consisted of one single-antenna base station (BS) per cell that
served one user equipment (UE) per time-frequency resource.
Since the increasing data traffic calls for higher spectral efficiencies [bit/symbol/cell], the RAN structure is now evolving to enable coherent downlink (DL) transmission to multiple UEs per resource symbol. The LTE-A standard has basic
support for multi-user MIMO using a handful of co-located
antennas. The massive MIMO concept from [1] takes multiuser MIMO to the 5G era by using hundreds of BS antennas
to serve tens of UEs in parallel on each resource block.
As the cellular concept is evolving, we can also question
whether future BSs should be in the cell centers as in the past
or distributed over the cells. The cloud RAN concept from [2]
provides an efficient way to operate distributed antenna arrays and perform the coherent processing required by massive
MIMO. While the majority of works on massive MIMO considers co-located arrays, the recent works [35] show that distributed massive MIMO can provide even higher spectral efficiencies than co-located deployments, due to proximity gains.
A potential showstopper for distributed massive MIMO
would be if the technology is too sensitive to transceiver hardware impairments; for example, phase noise in local oscillators (LOs), amplifier non-linearities, non-ideal analog filters,
This research has received funding from ELLIIT, CENIIT, and the EU
7th Framework Programme under GA no ICT-619086 (MAMMOET).
978-0-9928626-3-3/15/$31.00 2015 IEEE
2481
and finite-precision analog/digital converters. The impact of

hardware impairments on massive MIMO has received considerable attention in recent years [511], but only [5] considered distributed arrays. The paper [6] showed that it is of fundamental importance to include hardware impairments in the
performance analysis, since this can be a main limiting factor
in systems with many antennas. Nevertheless, [5, 6] showed
that massive MIMO is resilient to additive distortions originating from the BS. Multiplicative distortions such as phase
noise can, however, hinder the system performance. These
works use simplified stochastic impairment models, but the
validity of the results has been confirmed in [11] by simulations based on sophisticated and realistic models.
For distributed arrays, an important question is whether
the antennas should share a common LO (CLO) or if each
antenna should be equipped with a separate LO (SLO). A
number of recent works have looked into how this design
choice impacts the severeness of the phase noise [5, 710].
The papers [5, 7, 8, 10] seem to establish the consensus that
a setup with SLOs is preferable in the uplink (UL), since the
independent phase rotations average out over the BS antennas. However, the answer is still open when it comes to the
DL; [9] showed that a CLO is preferable for non-fading channels, while [10] considered fading single-cell systems and
claimed that CLO prevails for few BS antennas (per user) or
high SNR, and SLOs are desirable in the opposite cases.
In this paper, we extend our previous UL work in [5] to
the DL. We consider a multi-cell massive MIMO system with
distributed arrays and three kinds of hardware impairments:
phase noise, distortion noise, and noise amplification. We derive new spectral efficiency expressions for maximum ratio
transmission (MRT), which establish a performance baseline
in hardware-impaired multi-cell scenarios. These expressions
are used to prove how the hardware quality may scale with
the number of antennas. The analysis shows that SLOs is systematically a better choice than CLO also in the DL.
2. SYSTEM MODEL
We consider a cellular network with L cells that operate in
a synchronized time-division duplex (TDD) mode. Each cell
Uplink data
Uplink pilots
Downlink data
t {UL +1, . . . , 0} t {1, . . . ,B} t {B+1, . . . , B+ DL }
Fig. 1. Illustration of the TDD protocol where each coherence
block consists of T = UL + DL + B symbols.

serves K single-antenna UEs using a BS equipped with N antennas, which can be arbitrarily distributed over the coverage
area. The TDD protocol divides the time-frequency resources
into coherence blocks, as illustrated in Fig. 1. Each block consists of T symbols with time indices t = UL + 1, . . . , B +
DL , whereof UL are UL data symbols, B are UL pilots, and
DL are DL data symbols. Note that T = UL +DL +B.
Let ()T and ()H denote the transpose and conjugate transpose, respectively. The channel response between UE k in
(1)
(N )
cell l and BS j is a constant vector hjlk , [hjlk . . . hjlk ]T
(n)
CN within each block, where hjlk is the channel response for

the nth BS antenna. The channels are assumed to be Rayleigh
fading as
hjlk CN (0, jlk ),
(1)
(1)
(N )
where the covariance matrix is jlk , diag(jlk , . . . , jlk ).

(n)
The average channel attenuation jlk 0 is different for each

combination of cell indices, UE index, and BS antenna index
n. It depends, for example, on how the BS antennas are distributed in the cell and on the UE positions.
A main goal of this paper is to investigate how transceiver

hardware impairments impact the DL spectral efficiency. We
mainly consider impairments at the BSs, since the massive
MIMO deployment constraints (e.g., cost, power, synchronization, and size restrictions) are likely to lead to BS hardware of lower quality than in contemporary networks.
To this end, we adopt the UL system model from [5] and
generalize it to also cover the DL. Since the BSs in massive
MIMO use channel estimates from the UL to perform transmit precoding in the DL, we need to model both directions of
the links. As in [5], the received UL signal yj (t) CN in
cell j at symbol time t {UL + 1, . . . , B} is modeled as
L
X
Hjl xl (t) + j (t) + j (t)
2. Separate LOs (SLOs): All jn (t) are independent.

The above represent having one LO that feeds all antennas at
BS j or one separate LO connected to each of the N antennas.
Moreover, j (t) CN (0, j (t)) is additive distortion
noise (e.g., from finite-precision quantization, non-linearities,
and interference leakage in the frequency domain). It is proportional to the received signal power at the antenna and uncorrelated between antennas [5, 13]:

L X
K
X
(1) 2
(N ) 2
2
UL
j (t) , UL
plk diag |hjlk | , . . . , |hjlk |
(3)
l=1 k=1
where UL 0 is the proportionality coefficient.

2
Finally, j (t) CN (0, BS
IN ) is the receiver noise with
2
variance BS (including noise amplification in circuits).
2.2. Downlink Model with Hardware Impairments
Similar to the UL, we model the received DL signal zjk (t)
C at UE k in cell j at time t {B + 1, . . . , B + DL } as
!
K
L
X
X
H
wlm (t)slm (t) + l (t) +jk (t)
hljk Dl (t)
zjk (t) =
m=1
l=1
(4)
where slm (t) is the DL data symbol (with power pDL
jk =
(1)
2.1. Uplink Model with Hardware Impairments
yj (t) = Dj (t)
jn (t) N (jn (t 1), ) where 0 is the variance of the

phase-noise increments. We consider two implementations:
1. Common LO (CLO): j1 (t) = . . . = jN (t) within a cell.
(2)
l=1
where xl (t) = [xl1 (t) . . . xlK (t)]T CK contains pilot/data

symbols from UEs in cell l and the channel matrix from these
UEs to BS j is Hjl , [hjl1 . . . hjlK ] CN K . The sym2
bols from UE k in cell j have power pUL
jk = E{|xjk (t)| },
where E{} denotes the expected value of a random variable.
The matrix Dj (t) , diag ej1 (t) , . . . , ejN (t) models
the multiplicative effect of phase noise (with = 1). The

variable jn (t) is the phase rotation at the nth BS antenna in
cell j at time t, and it is modeled as a Wiener process [12]:
2482
(N )
E{|slm (t)|2 }) and wlm (t) , [wlm (t) . . . wlm (t)]T CN is

the corresponding linear precoding vector. The receiver noise
2
2
is the variance (includ), where UE
is jk (t) CN (0, UE
ing noise amplification). The phase-noise matrix Dj (t) was
defined earlier, while j (t) CN (0, j ) is the additive distortion in the DL (e.g., due to non-linearities and leakage in
the frequency domain). Similar to (3), the distortion at a certain antenna is proportional to the transmit power at this antenna and uncorrelated with the distortions at other antennas:
K

X
(1)
(N )
2
2
j , 2DL
pDL
jk diag |wjk (t)| , . . . , |wjk (t)|
k=1
where DL 0 is the proportionality coefficient.

This system model is used in the next section to compute achievable DL spectral efficiencies. These depend on
the level of hardware impairments, as characterized by the
variance of the phase-noise increments , the distortion noise
proportionality coefficients UL , DL , and the receiver noise
2
2
variances BS
, UE
. The results are applicable for any pDL
jk
and pUL
jk , for each j and k, thus under arbitrary power control.
3. DOWNLINK PERFORMANCE ANALYSIS
In this section, we derive the DL spectral efficiency per UE
and study its asymptotic behavior (when N is large) to understand the impact of hardware impairments.
SINRjk (t) =
pDL
jk
L P
K
P
pDL
lm
l=1m=1
2
|E{hH
jjk (t) jk (t)}|
E{k jk (t)k2 }

PN
(n)
(n) 2
2
2
2
E{|hH
n=1 E{|hljk | | lm (t)| }
ljk (t) lm (t)| }+DL
E{k lm (t)k2 }
3.1. Uplink Channel Estimation

In order to perform coherent transmit precoding in the DL,
each BS acquires the channels to its UEs by using the UL
pilots. The pilot sequence of UE k in cell j is defined as
jk , [xjk (1) . . . xjk (B)]T CB1 . The analysis in this
x
paper holds for arbitrary pilot sequences (with |xjk (b)|2 =
pUL
jk for b = 1, . . . , B), while we consider columns from a
Fourier matrix in Sec. 4 (to achieve mutual orthogonality and
constant energy per symbol). Since the effective channels
hjlk (t) , Dj (t) hjlk
(5)
depend on the phase-noise and are different at every symbol

time t, we need a channel estimator that provides new estimates at each t. Such a linear minimum mean-squared error
(MMSE) estimator was derived in [5] and is described below,
since the expressions are used in the forthcoming analysis.

T
Lemma 1 (Th. 1 in [5]). Let j , yjT (1) . . . yjT (B)
CBN denote the combined received signal in cell j from the
pilot transmission. The LMMSE estimate of hjlk (t) at any
symbol time t {UL + 1, . . . , B + DL } for any l and k is

jlk (t) = x
Hlk D(t) jlk 1
h
(6)
j j
where denotes the Kronecker product,

D(t) , diag e 2 |t1| , . . . , e 2 |tB| ,

j ,
L X
K
X
`=1 m=1
2
X`m j`m + BS
IBN ,
(13)
2
+ UE
Lemma 2. Suppose that UE k in cell j knows the channel and

interference statistics, but not the channel realizations. An
achievable lower bound on the ergodic capacity of this UE is
Rjk =
B+

1 XDL
log2 1 + SINRjk (t)
T
[bit/symbol] (11)
t=B+1
where SINRjk (t) is given in (13) at the top of this page.

Proof. As in [7], we compute one spectral efficiency for each
t {B+1, B+DL } since the effective channels vary with t.
The expression is obtained by using the signal received over
the average channel E{hHjjk (t) jk (t)} for decoding, while
treating the signal received over the uncorrelated deviation
from this average value, the inter-user interference and distortion noise as worst-case Gaussian noise in the decoder.
(7)
(8)
Proof. Follows from straightforward computation of the expectations, whereof some are the same as in [5, Th. 2].
T
lk jlk ).
Cjlk (t) = jlk (
xHlk D(t) jlk )1
j (D(t) x
Proof. This lemma follows from adapting results in [5] to the

notation and power constraints considered herein.
3.2. Downlink Spectral Efficiency
Next, we derive achievable DL spectral efficiencies, using
normalized linear precoding vectors of the general form
jk (t)
.
E{k jk (t)k2 }
2
|E{hH
jjk (t) jk (t)}|
E{k jk (t)k2 }
The expression in Lemma 2 is a reasonable bound on the

practical performance that can be achieved using simple signal processing at the UE (i.e., detect the useful signal and
treat everything unknown as Gaussian noise). The SINR expression in (13) contains a number of expectations that can be
computed numerically for any choice of precoding vectors.
Next, we provide closed-form expressions for MRT.
Theorem 1. If MRT is used, then the expectations in SINRjk (t)
of Lemma 2 are computed as in (14)(15) at the top of the
next page (where en denotes the nth column of IN ).
and the element at position (b1 , b2 ) in X`m CBB is

(
2
pUL
b 1 = b2 ,
`m (1 + UL ),
[X`m ]b1 ,b2 =
x`m (b1 )x`m (b2 )e 2 |b1 b2 | , b1 6= b2 .

(9)
The corresponding error covariance matrix is
wjk (t) = p
pDL
jk
3.3. Asymptotic Behavior and Scaling Laws

Next, we investigate the behavior at large N . For tractability,
we consider A < spatially separated subarrays each with
N
A antennas. Recall that these antennas are either controled
by a common LO that sends clock signals or separate LOs at
each antenna. The channel covariance matrices then factorize
as
(A) I N
(16)
jlk =
jlk
A
(1) , . . . ,
(A) ) CAA and
(a) is the
(A) = diag(
where
jlk
jlk
jlk
jlk
average channel attenuation between subarray a in cell j and
UE k in cell l. By letting the number of antennas in each
subarray grow large, we obtain the following property.
Corollary 1. If MRT is used and the channel covariance matrices can be factorized as in (16), then
(10)
SINRjk (t) =
L P
K
P
pDL
jk Sjk
pDL
lm Ilmjk
l=1m=1
jjk (t).
With this notation, MRT is given by jk (t) = h
2483
1
pDL
jk Sjk +O( N )
(17)
E{k jk (t)k2 } = E{hHjjk (t) jk (t)} = tr

E{|hHljk (t) lm (t)|2 } + 2DL
N
X
n=1
(n)

T
Hjk D(t) jjk 1
x
x
D
jk
jjk
j
(t)
(14)
(n)
E{|hljk |2 | lm (t)|2 }
(15)

T
Hlm D(t) llm 1
= (1 + 2DL )tr ljk x

x
D
llm
(t) lm
l
N N

P P (n1 ) (n1 ) (n2 ) (n2 ) H
1
lm D(t) eHn1 1
lm en2
DT(t) x
llm ljk llm ljk x
(Xjk 2UL pUL
jk IB ) en1 en2 l
l
n1 =1n2 =1
+

2
tr x
Hlm D(t) llm 1
jk ljk
DT(t) x
l
N
2

P
(n) (n)
1
2
UL
2
H
T
Hlm D(t) eHn 1
llm ljk
x
(
p
I
+
X
)
e
e
D
x
B
jk
n
lm
n
n
UL jk
DL
l
l
(t)
+ n=1

N
2

P
(n) (n)
1
2
T
H
H
T
Hlm D(t) eHn 1
llm ljk
x
((1
+
)X
D
x
x
D
)
e
e
D
x
e
jk
n n
n
DL
l
l
(t) jk jk (t)
(t) lm
if CLO
if SLOs
if CLO
if SLOs
n=1
where the signal part is

(A) )
e 1 (D(t) x
(A) )
jk
Sjk = tr (
xHjk D(t)
j
jjk
jjk
e j , PL PK X`m
(A) + 2 IAB , where the
with
BS
`=1
m=1
j`m
interference terms Ilmjk with a CLO are
A P
A

P
(a1 )
(a1 ) (a2 ) (a2 ) x
Hlm D(t) eHa1
llm ljk llm ljk

a =1a =1
CLO

Ilmjk
= 1 2
1 T
e (D x
Hlm D(t) llm
)
tr x
jk
ljk
l
(t)

1
1
H
T
e
e
lm ea2 )
(Xjk 2UL pUL

l
jk IB ) ea1 ea2 l (D(t) x
and the interference terms with SLOs are

1 T
SLOs
e (D x
Hlm D(t) llm
Ilmjk
= tr x
l
(t) jk ljk ) .
The notation O( N1 ) is used for terms that go to zero as
faster when N , while ea is the ath column of IA .
1
N
or
Proof. Follows from dividing all terms in SINRjk (t) by

and then analyze the expressions for MRT in (14)(15).
N
A
2
2
This corollary does not contain UL , DL , BS
, or UE
,
thus it shows that the impact of distortion noise and receiver
noise vanishes as N . The asymptotic SINRs are only
limited by the channel distributions, pilot-contaminated interference, and phase noise. This means that distributed massive
MIMO systems can handle larger additive distortions than
conventional systems, as manifested by the next corollary.
Corollary 2. Suppose that 2UL = 2UL,0 N z1 , 2DL = 2DL,0 N z1 ,

2
2
2
2
BS
= BS,0
N z2 , UE
= UE,0
N z2 , and = 0 (1 + ln(N z3 )),
for some scaling exponents z1 , z2 , z3 0 and constants
2
2
UL,0 , DL,0 , BS,0
, UE,0
, 0 0. The SINRs, SINRjk (t),
with MRT converge to non-zero limits as N if
(
max(z1 , z2 ) 12 and z3 = 0
for a CLO
(18)
B|
1
max(z1 , z2 ) + z3 0 |DL
for SLOs.
2
2
Proof. Based on (14)(15); see [5] for a similar proof.
2484
This corollary shows that the DLcan handle additive distortions with variances that scale as N (i.e., z1 = z2 = 12 ),
while achieving decent performance. The scaling law also
shows that the phase noise variance with SLOs can increase
logarithmically with N , while this is not allowed with a CLO.
This proves that massive MIMO with SLOs are preferable in
the DL, at least when the number of antennas is large. The
scaling law holds also for precoders that are better than MRT.
4. NUMERICAL RESULTS
The analytic results are corroborated for the distributed massive MIMO setup in Fig. 2. This is a wrap-around topology
with 16 cells of 400 400 meters, each consisting of A = 4
subarrays with N
A antennas located 100 meters from the cell
center. The K = 15 UEs per cell are uniformly distributed,
with a minimum distance of 25 meters from the subarrays.
UL
The transmit powers are pDL
jk = pjk = 50 dBm/Hz for all j
and k (e.g., 100 mW over 10 MHz). The channel attenuations
(n)
(n)
(n)
are modeled as in [5]: jlk = 10sjlk 1.53 /(djlk )3.76 , where

(n)
djlk is the distance in meters between BS antenna n in cell j

(n)
and UE k in cell l and sjlk N (0, 3.16) is shadow-fading.

The hardware impairments are characterized by the distortion proportionality coefficients UL = DL = 0.03, the
variance of phase noise increments = 1 105 , and the re2
2
ceiver noise powers BS
= UE
= 169 dBm/Hz (with 5 dB
noise amplification). These are also the initial constants when
we scale the hardware quality based on Corollary 2.
Fig. 3 shows the average spectral efficiency per UE. The
coherence block contains T = 300 symbols, whereof B = 15
symbols are used for pilot sequences and DL = 285 for DL
payload data. Hardware impairments incur a performance
loss as compared to ideal hardware. The gap is small with
SLOs, but larger with a CLO. This validates the analytic observation that SLOs is the better choice in massive MIMO.
The figure also illustrates the scaling law established by
Corollary 2. The middle curves show the behavior when sat-
Cell 11
Cell 12
Cell 9
Cell 10
Cell 11
Cell 12
Cell 9
Cell 10
Cell 15
Cell 16
Cell 13
Cell 14
Cell 15
Cell 16
Cell 13
Cell 14
Cell 3
Cell 4
Cell 1
Cell 2
Cell 3
Cell 4
Cell 1
Cell 2
Cell 7
Cell 8
Cell 5
Cell 6
Cell 7
Cell 8
Cell 5
Cell 6
Cell 11
Cell 12
Cell 9
Cell 10
Cell 11
Cell 12
Cell 9
Cell 10
Cell 15
Cell 16
Cell 13
Cell 14
Cell 15
Cell 16
Cell 13
Cell 14
Cell 3
Cell 4
Cell 1
Cell 2
Cell 3
Cell 4
Cell 1
Cell 2
Cell 7
Cell 8
Cell 5
Cell 6
Cell 7
Cell 8
Cell 5
400 meter
Subarray
Cell 6
Baseband UEs
unit
Fig. 2. Illustration of the multi-cell distributed massive MIMO
scenario considered in the numerical evaluation.
Average Spectral Efficiency per UE [bit/symbol]
3.5
Ideal Hardware
With Impairments: SLOs
With Impairments: CLO
3
2.5
z1 = z2 = z3 = 0
(Fixed imperfections)
2
1.5
z1 = z2 (= z3)= 0.48
(Satisfies scaling laws)
z1 = z2 (= z3)= 0.96
0.5
0
(Faster than scaling laws)

0
50
100
150
200
Number of BS Antennas (N)
250
300
Fig. 3. Average DL spectral efficiency for distributed massive
isfying the scaling law (z1 = z2 = 0.48 with a CLO and

adding also z3 = 0.48 with SLOs). By gradually degrading
the hardware with N , there is a performance loss at every N ,
but the curves are still increasing with N . The performance
loss is small for SLOs, but very large for a CLO. The curves at
the bottom are for a case when the scaling law is not satisfied,
which gives a performance that goes to zero as N .
5. CONCLUDING REMARKS
We have analyzed the DL performance of distributed massive
MIMO systems, with focus on the impact of hardware impairments. We have proved that additive distortions have smaller
impact on massive MIMO than
conventional networks, since
the variance may increase as N with little performance loss.
Multiplicative phase noise can be more severe, but the performance is better if each BS antenna has a separate oscillator.
The DL analytic results in this paper are in line with previous UL results in [5, 7, 8, 10]. This is natural since the ULDL duality for systems with linear processing implies that
the same performance is achievable in both directions (if the
power allocation is optimized). However, our results stand
in contrast to the recent works [9, 10] where the DL behave
differently than the UL when it comes to phase noise. This
is due to different system models: [9] considers high SNRs
in non-fading single-user cases, while [10] considers a single
cell with relatively good CSI. In comparison, we consider a
generalized multi-cell setup with more inter-user interference
and thus lower SINRs.
REFERENCES
[1] T. L. Marzetta, Noncooperative cellular wireless with
unlimited numbers of base station antennas, IEEE
Trans. Wireless Commun., vol. 9, no. 11, pp. 3590
3600, 2010.
MIMO with fixed or increasing hardware impairments.

[4] H. Yin, D. Gesbert, and L. Cottatellucci, A coordinated
approach to channel estimation in large-scale multipleantenna systems, IEEE J. Sel. Topics Signal Process.,
vol. 8, no. 5, pp. 942953, 2014.
[5] E. Bjornson, M. Matthaiou, and M. Debbah, Massive
MIMO with non-ideal arbitrary arrays: Hardware scaling laws and circuit-aware design, IEEE Trans. Wireless Commun., To appear.
[6] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive MIMO systems with non-ideal hardware: Energy efficiency, estimation, and capacity limits, IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112
7139, 2014.
[7] A. Pitarokoilis, S. K. Mohammed, and E. G. Larsson,
Uplink performance of time-reversal MRC in massive
MIMO systems subject to phase noise, IEEE Trans.
Wireless Commun., vol. 14, no. 2, pp. 711723, 2015.
[8] A. Pitarokoilis, E. Bjornson, and E. G. Larsson, Optimal detection in training assisted SIMO systems with
phase noise impairments, in Proc. IEEE ICC, 2015.
[9] M. R. Khanzadi, G. Durisi, and T. Eriksson, Capacity
of SIMO and MISO phase-noise channels with common/separate oscillators, IEEE Trans. Commun., To
appear.
[10] R. Krishnan et al., Linear massive MIMO precoders
in the presence of phase noisea large-scale analysis,
IEEE Trans. Veh. Technol., To appear.
[11] U. Gustavsson et al., On the impact of hardware impairments on massive MIMO, in Proc. IEEE GLOBECOM, 2014.
[2] China Mobile Research Institute, C-RAN: The road

towards green RAN, White Paper, Oct. 2011.
[12] D. Petrovic, W. Rave, and G. Fettweis, Effects

of phase noise on OFDM systems with and without
PLL: Characterization and compensation, IEEE Trans.
Commun., vol. 55, no. 8, pp. 16071616, 2007.
[3] K. T. Truong and R.W. Heath Jr., The viability of

distributed antennas for massive MIMO systems, in
Proc. Asilomar CSSC, 2013, pp. 13181323.
[13] W. Zhang, A general framework for transmission with

transceiver distortion and some applications, IEEE
Trans. Commun., vol. 60, no. 2, pp. 384399, 2012.
2485
BASE STATION CLUSTER PATTERNS FOR SEMI-STATIC MULTI-CELL COOPERATION

IN IRREGULAR NETWORK TOPOLOGIES
Jeonghun Park, Namyoon Lee, and Robert W. Heath Jr.
Department of Electrical and Computer Eng.
The University of Texas at Austin
Austin, TX 78712 USA
Email: {jeonghun, namyoon.lee, rheath}@utexas.edu
ABSTRACT
This paper proposes a clustering strategy for semi-static multicell cooperation. Semi-static multicell cooperation exploits
multiple predefined base station (BS) cluster patterns for improving cell-edge user throughput. The proposed clustering
guarantees that every user communicates with their two closest BSs, so that users are protected from the dominant interferer. The key idea of the proposed clustering is to use the
2nd-order Voronoi region to form BS clusters. Each of the
formed BS clusters is mapped into a particular cluster pattern
by exploiting the edge-coloring in graph theory. Through simulations, the performance is compared to that of other conventional strategies. Our major finding is that the proposed clustering provides performance gains for cell-edge users compared to that of the conventional strategies.
1. INTRODUCTION
One main benefit of cloud-based radio access networks
(Cloud-RANs) [1], featured by distributed base stations (BSs)
connected to centralized processing units, is that it inherently
has an efficient structure for BS cooperation. BS cooperation is a strategy for mitigating inter-cell interference where
the transmission strategy are cooperated, e.g., beamforming,
scheduling, and power control. Since exchanging transmitted
data or sharing estimated channel coefficients between cooperating BSs are often needed in BS cooperation strategy, in
conventional cellular networks, an additional core processor
is required to help the exchanging and sharing procedures [2].
In cloud-RANs, however, BS cooperation is rather a default
option due to the centralized structure of processing units [3].
Considering BS cooperation in an entire network, a BS
cluster, defined as a subset of BSs, is indispensable for practically implementing BS cooperation because cooperating all
the BSs in the network is infeasible due to the unreasonable
amount of overhead [4]. Therefore, it is natural to cooperate
within a cluster. When cooperating with a limited number of
BSs, however, the performance of a user is mainly limited by
unmanageable out-of-cluster interference, especially when a
978-0-9928626-3-3/15/$31.00 2015 IEEE
user is in an edge-region [4, 5]. For this reason, the way of

forming a BS cluster has a substantial influence to the performance of BS cooperation approaches.
Dynamic BS clustering [6, 7], where each BS cluster is
formed by each user based on the users conditions, e.g., a
distance to each BS and channel quality, is an effective BS
clustering strategy. There are limitations though with dynamic BS clustering: the overheads in dynamically forming
the clusters and the high complexity of the user scheduling
[8]. To overcome these shortcomings, in [9, 10], semi-static
BS clustering was proposed. The main concept is to use multiple predefined cluster patterns so that edge-users can have
a chance to be protected from dominant interference. Since
the BS clusters are predefined, problems caused in dynamic
clustering are relieved. The existing semi-static BS clustering
approaches [9, 10], however, mainly focused on grid network
topologies where the BSs locations are regularly placed on a
grid. This network model is far from a practical cellular environment, and also it is trivial to design BS cluster patterns in a
regular network topology. In a more practical network model
where BSs are irregularly placed, it is not clear how to form
a BS cluster and how to construct the cluster patterns. This
is mainly because when the network size is large, considering all the BS clusters jointly is infeasible, which requires a
general rule for making BS clusters and their corresponding
patterns.
In this paper, we propose a BS clustering strategy applied
in irregular network topologies. The main benefit of the proposed clustering is to guarantee that every user in the network is able to communicate with their two closest BSs, so
that users are protected from the dominant interference, i.e.,
the interference coming form the closest interfering BS. The
key feature of the proposed clustering is to use the 2nd-order
Voronoi region, defined as
V2 (di , dj )

= d R2 {kd di k kd dk k}

{kd dj k kd dk k} , k Z+ , k 6= i, j .
(1)
The geometric meaning of the 2nd-order Voronoi region is a
2486
10
2.2. Cooperation Model
Base station
8
6
Y coordinate
4
2
0
2
4
6
8
10
10
2
0
2
X coordinate
10
Fig. 1. An example of the network model. This example is

generated by a repulsive point process where a distance between any two BSs is larger than 1.4km.
set of points that is closer to di and dj than any other points.

Due to this property, if the BSs located at di and dj forms
a BS cluster and serve a user in V2 (di , dj ), the served user
is able to communicate with their two closest BSs. Applying
this strategy in whole network, the network plane is tessellated into the 2nd-order Voronoi regions and users in each
region are served by the corresponding BS pair, i.e., a user in
V2 (di , dj ) are served by BS pair {di , dj }. The formed BS
clusters are mapped into multiple BS cluster patterns so that
they can be used without conflict. To do this, edge-coloring
in graph theory is exploited. For demonstration, the Signalto-Interference ratio (SIR) coverage probability is compared
to other conventional methods via Monte-Carlo simulations.
Our main novelty is to propose a clustering strategy that
can be applied in any BS deployment scenario, and in the proposed clustering every user is ensured to communicate without the dominant interference, so that the performance improvement of cell-edge users is achieved.
It is assumed that a BS pair located at {di , dj } for di , dj N

can cooperate. The reason behind that only pair-wise BS cooperation is assumed is not only its simplicity, but also that
the pair-wise BS cooperation is considered as the optimal cooperation size in particular setups. In [11], it was shown that
forming a BS cluster including more than two BSs actually
degrades the spectral efficiency performance when considering the signaling overheads for estimating channel coefficients within the cluster. Obviously, this is only true in a frequency division duplexing (FDD) system, while the most of
current cellular systems are operated based on the FDD system. When the BS i and BS j form a BS cluster, they exchange the data and transmit it to one user, i.e., one user is
served per one BS cluster.
2.3. Signal-to-Interference Ratio
We focus on a user located at u, denoted as the tagged user.
By shifting BS is location to d0i = di u for i Z+ , we
are able to assume that the tagged user is located on the origin
without loss of generality. For simplicity, we write d0i as di
and u = 0. It is assumed that a BS pair located at {d0 , dc }
jointly serves the tagged user by transmitting the same data
through the designed precoder. Denoting that hi is the channel fading coefficient from BS i to the tagged user, the precoder of the BS i is designed as vi = hi / |hi | if i {0, c}.
If i Z+ \{0, c}, vi is independent to the interfering channel
hi . The perfect channel state information at the transmitter
(CSIT) is assumed in this paper. Then, as in [3], the SIR is
given by
kd0 k
SIR = P
Hc
Hi
H0 + kdc k
di N \{0,c}
kdi k
(2)
2. SYSTEM MODEL
In this section, we first explain the network model and the
cooperation model. Then we define the SIR metric.
where Hi = |hi vi | . The SIR coverage probability is defined

as
#
"
kd0 k H0 + kdc k Hc
> , (3)
P [SIR > ] = P P
Hi
di N \{0,c} kdi k
where is a SIR target.
2.1. Network Model

3. MAIN RESULTS
A fixed downlink cellular network where each BS has a single
antenna is considered. The location of BS i is denoted as di
2
for i Z+ and di
R , and
a set of BSs location is denoted
2
as N = di R i Z+ . As in the conventional cellular
system, each network plane is tessellated into |N | numbers
of the 1st-order Voronoi regions V1 (di ) for di N . An
example of the considered network is illustrated in Fig. 1.
In this section, we first explain the concept of forming an efficient BS cluster, which guarantees that a user is served by
the two nearest BSs. Next, we propose how the proposed BS
cluster is implemented with the multiple cluster patterns in a
network. A toy example is exploited for an intuitive explanation, which is subsequently extended to a general network.
2487
d3
V2 (d1 , d3 )
d3
V2 (d0 , d3 )
V2 (d0 , d1 )
d0
V2 (d2 , d3 )
V2 (d0 , d2 )
d2
d1
V2 (d1 , d2 )
P1 = {V2 (d0 , d1 ), V2 (d2 , d3 )}
P2
P3
P1
d0
P2 = {V2 (d0 , d2 ), V2 (d1 , d3 )}

P3 = {V2 (d0 , d3 ), V2 (d1 , d2 )}
A=
3
[
i=1
Pi
d1
P3
P1
P2
d2
Fig. 3. The graph where vertices in the example Fig. 2 and

Fig. 2. The considered toy example where the network plane is
a finite triangular plane A, and each BS is located at di for i

{0, 1, 2, 3}. By using the edge-coloring for the constructed
graph as in Fig. 3, each 2nd-order Voronoi region is mapped
into particular cluster pattern, resulting in that the 2nd-order
Voronoi region whose the shade pattern is same is included in
the same cluster pattern.
3.1. Toy Example
In a toy example, it is assumed that there are four BSs, denoted as BS i for i {0, ..., 3}. BS i is located at di as
illustrated in Fig. 2, thereby N = { di A| i {0, 1, 2, 3}}.
In this network, we tessellate the network plane into the
2nd-order Voronoi regions defined in (1). The tessellated
network plane is also described in Fig. 2. Then, from the
geometric meaning of the 2nd-order Voronoi region, a user
in V2 (di , dj ) has the two nearest BSs pair {BS i, BS j}.
When BS i and BS j form a BS cluster, therefore, a user in
V2 (di , dj ) is served from the two nearest BSs, which guarantees that the served user is protected from the dominant
interference. When applying this clustering strategy in the
whole network, however, it is possible that a BS conflict occurs [12], i.e., the situation where more than two users want
to be served from one BS simultaneously. For instance, assume that there are two users, where user 1 is in V2 (d0 , d1 )
and user 2 is in V2 (d0 , d2 ). Since BS 0 cannot transmit
two data symbols simultaneously, user 1 or user 2 cannot
be served from the two nearest BSs. To avoid this, each of
the 2nd-order Voronoi regions needs be covered by a different time-frequency resource. When the 2nd-order Voronoi
regions covered by the same time-frequency resource are
mapped into the same cluster pattern, it is guaranteed that every user in the network communicates with their two nearest
BSs without the dominant interference by using the multiple
cluster patterns.
Now we propose how to map each of the 2nd-order
Voronoi region into a particular cluster pattern. The key idea
is to use the edge-coloring in graph theory. Assuming an
arbitrary graph, edge-coloring assigns a color to each edge in
the graph so that any edge sharing the same vertex does not
have the same color with the minimum number of colors. To
edges are made by the Delaunay triangulation corresponding

to the toy example. By the edge-coloring, each edge is assigned by the cluster patterns Pi for i = {1, 2, 3}.
this end, we construct the graph G (N ) corresponding to the
considered network. In G (N ), each vertex is the location of
each BS, and each edge is the Delaunay triangulation defined
in N . The Delaunay triangulation in N is defined by the
three following conditions: 1) It is a triangulation of a plane,
2) Each vertex in N is a vertex of a triangle, and 3) The
circumcircle of each triangle should not include any vertex in
N . The following Lemma 1 provides the core feature of the
graph G (N ) related to the 2nd-order Voronoi region.
Lemma 1. The 2nd-order Voronoi region V2 (di , dj ) for
i, j Z+ and i 6= j is not empty if and only if di and dj is
connected by the Delaunay triangulation.
Proof. See [13] and the references therein.
By the Lemma 1, when considering cluster patterns

as colors, edge-coloring is equivalent to map each nonempty 2nd-order Voronoi region into the cluster patterns.
Since there is no vertex (BS) sharing the edges assigned
by the same color (the same cluster pattern), a BS conflict
is avoided. Fig. 3 describes the example of edge-coloring
of the graph corresponding to the toy example. It is observed that no vertex has the same colored edges, i.e., the
edges mapped into the same cluster pattern. Exploiting this
edge-coloring, the cluster patterns are constructed as P1 =
{V2 (d0 , d1 ), V2 (d2 , d3 )}, P2 = {V2 (d0 , d2 ), V2 (d1 , d3 )},
and P3 = {V2 (d0 , d3 ), V2 (d1 , d2 )}. By allocating different
time-frequency resources to each cluster pattern, no BS conflict occurs and therefore every user in A is guaranteed to be
protected from the dominant interference.
The required number of colors for edge-coloring is equivalent to the required number of time-frequency resources to
avoid a BS conflict. In this example, since three colors (cluster patterns) are used for edge-coloring, three time-frequency
resources are needed to avoid a BS conflict.
2488
10
cooperation, where the exploited BS cluster and the cluster

patterns are predefined. Since the proposed clustering only
depends on the BS geometry, unless the BSs locations are
changed, the designed BS cluster and cluster patterns are
preserved irrespective of the instantaneous channel condition.
Conventionally, the BS geometry would not be changed on
the order of months or years.
Base station
8
6
Y coordinate
4
2
0
2
4
6
4. SIMULATION RESULTS
8
10
10
2
0
2
X coordinate
10
Fig. 4. The graph corresponding to N in Fig. 1. Each vertex

of G (N ) is di N , and each edge is the Delaunay triangulation.
3.2. General Network

Now we extend the toy example to a general network. The
procedure is equivalent to the toy example case. We first tessellate the network plane R2 into multiple 2nd-order Voronoi
regions V2 (di , dj ) for di , dj N and i 6= j, and a BS
pair located at {di , dj } serves a user in V2 (di , dj ) by transmitting the same data symbol for the user. To design cluster
patterns, we draw the graph G (N ) according to the Delaunay
triangulation of N . When N is corresponding to the example
illustrated in Fig. 1, the graph G (N ) is drawn as in Fig. 4.
Once G (N ) is made, by using the edge-coloring for G (N ),
each edge in G (N ) is assigned by particular color (cluster
pattern). Since edge-coloring is not restricted to a specific
topology of vertices, the edge-coloring used in a general network is similar to that applied in the toy example, as illustrated in Fig. 3. Due to the page limitation, we provide more
rigorous explanations in our extended work [14]. Exploiting
the result of the edge-coloring, different time-frequency resources are allocated to each cluster pattern, and thanks to
this a BS conflict is prevented. With the proposed clustering
strategy, every user in the network is guaranteed to receive the
desired data symbol from two closest BSs without the dominant interference.
Denoting the constructed cluster patterns as Pi for i
{1, ..., L} and i{1,...,L} Pi = R2 , L time-frequency resources are required to cover the whole network without a
BS conflict. In contrast to the toy example, however, it is not
trivial to find L in a general case. It turns out that the value
of L is closely related to the chromatic index of the graph
G (N ), which is determined by the maximum degree, i.e., the
maximum number of edges connected to a certain vertex in
G (N ). We explore this point further in our other work [14].
One point worth noticing is that although drawing G (N )
and solving the general edge-coloring for G (N ) might cause
high computational complexity, the complexity is not an
important issue in the proposed clustering strategy. This is
mainly because the proposed clustering is a semi-static BS
For the simulation, we use the same network model with the
example described in Fig. 1. To generate this network model,
we first drop each BS by a homogeneous Poisson point process with density = 3 107 per m2 , and perform dependent thinning for each point if a distance between two points
is less than 1.4km. This mimics the actual BS deployment
scenario where each BS is implemented with a guard distance
to prevent that BSs are located very close together. Although
the particular network model is considered for the simulation,
one should note that the proposed strategy is not limited by
specific network topologies. The proposed clustering can be
applied in any network topology regardless of the distribution
of BSs. We assume that the tagged user is uniformly dropped
in the network and we shift di u for di N , where u is a
location where the user is initially dropped. By this construction, the tagged user is located on the origin. The pathloss
exponent is set to be 4, which is a typical value for a terrestrial wireless environment [11]. The channel coefficients
are assumed to follow Rayleigh fading, so that Hi follows
the exponential distribution with unit mean. For performance
comparison, two baseline methods are considered, i.e., single
cell operation and fractional frequency reuse. In single cell
operation, each BS serves its own user without interference
management technique. In fractional frequency reuse, one
of two orthogonal sub-bands is randomly allocated to each
cell [15], so that approximately half of the total interference
affects to the performance of the tagged user. Monte-Carlo
simulations are performed by 10 103 times.
The simulation result for the SIR coverage probability of
each case is illustrated in Fig. 5. As observed in Fig. 5, the
proposed clustering provides better SIR coverage probability
in the low SIR regime. Specifically, at = 0dB, the proposed
clustering has the performance gain of 12.5% compared to
fractional frequency reuse, and 23% compared to single cell
operation. The performance gain comes from the fact that
with the proposed clustering, the tagged user is ensured to be
protected from the dominant interference, while in fractional
frequency reuse, there is a non-zero possibility that the tagged
user is exposed to the dominant interference. Combined with
this fact, in the low SIR regime, one dominant interference
has significant effect to the SIR performance. From these two
reasons, although the number of the total interfering sources
is clearly smaller in fractional frequency reuse, the proposed
clustering is able to provide the better SIR coverage in the
2489
[4] A. Lozano, R. W. Heath, and J. G. Andrews, Fundamental limits of cooperation, IEEE Trans. Info. Th.,
vol. 59, no. 9, pp. 52135226, Sep. 2013.
Proposed Clustering
Fractional Frequency Reuse
Single Cell Operation
0.9
SIR coverage probability
0.8
0.7
[5] K. Huang and J. Andrews, An analytical framework

for multicell cooperation via stochastic geometry and
large deviations, IEEE Trans. Info. Th., vol. 59, no. 4,
pp. 25012516, Apr. 2013.
0.6
0.5
0.4
0.3
[6] A. Papadogiannis, D. Gesbert, and E. Hardouin, A dynamic clustering approach in wireless networks with
multi-cell cooperative processing, in Proc. IEEE Int.
Conf. on Comm., May 2008, pp. 40334037.
0.2
0.1
0
10
5
10
SIR threshold, (dB)
15
20
[7] A. Papadogiannis and G. Alexandropoulos, The value

of dynamic clustering of base stations for future wireless networks, in Proc. IEEE Int. Conf. on Fuzzy Sys.,
Jul. 2010, pp. 16.
Fig. 5. The SIR coverage probability depending on the SIR
threshold when = 4 and each channel coefficient follows

Rayleigh fading.
[8] S. Ramprashad and G. Caire, Cellular vs. Network

MIMO: A comparison including the channel state information overhead, in Proc. IEEE Int. Symp. on Pers.,
Indoor and Mobile Radio Comm., Sep. 2009, pp. 878
884.
low SIR regime. This can be concluded into the following

statement: By using the proposed clustering, cell-edge users
whose the SIR threshold is conventionally low, have benefit
of the SIR coverage performance.
[9] J.-Y. Hwang, J. Kim, T. Kim, and Y. Han, A periodic

frequency band rotation scheme for multi-cell coordination clustering, IEEE Comm. Lett., vol. 15, no. 9, pp.
956958, Sep. 2011.
5. CONCLUSIONS
In this paper, we propose a semi-static clustering strategy to
guarantee every user to communicate their two nearest BSs
without the dominant interference. The key idea is to use the
2nd-order Voronoi region for forming an efficient BS cluster,
and to exploit the edge-coloring for mapping each BS cluster
into cluster patterns for implementing the BS clusters while
avoiding BS conflicts. Through the simulation, we demonstrate that the proposed clustering provides the SIR coverage
gain compared to other methods for cell-edge users.
Acknowledgement
This research was supported by a gift from Huawei Technologies Co. Ltd.
REFERENCES
[1] China Mobile Research Institute, C-RAN: The road toward Green RAN, White Paper, Oct. 2011.
[2] B. Clerckx, H. Lee, Y.-J. Hong, and G. Kim, A practical cooperative multicell mimo-ofdma network based
on rank coordination, IEEE Trans. Wireless Comm.,
vol. 12, no. 4, pp. 14811491, Apr. 2013.
[3] N. Lee, R. Heath, D. Morales-Jimenez, and A. Lozano,
Base station cooperation with dynamic clustering in
super-dense cloud-ran, in Proc. IEEE Glob. Comm.
Conf., Dec. 2013, pp. 784788.
[10] H. Purmehdi, R. Elliott, W. Krzymien, and J. Melzer,

Rotating clustering with simulated annealing user
scheduling for coordinated heterogeneous mimo cellular networks, in Proc. IEEE Int. Conf. on Comm., Jun.
2014, pp. 52935298.
[11] N. Lee, D. Morales-Jimenez, A. Lozano, and R. Heath,
Spectral efficiency of dynamic coordinated beamforming: A stochastic geometry approach, IEEE Trans.
Wireless Comm., vol. 14, no. 1, pp. 230241, Jan. 2015.
[12] X. Hou, E. Bjrnson, C. Yang, and M. Bengtsson, Cellgrouping based distributed beamforming and scheduling for multi-cell cooperative transmission, in Proc.
IEEE Int. Symp. on Pers., Indoor and Mobile Radio
Comm., 2011, pp. 19291933.
[13] I. Fischer and C. Gotsman, Fast approximation of
high-order Voronoi diagrams and distance transforms
on the GPU, J. Graphics Tools, vol. 11, no. 4, pp. 39
60, 2006.
[14] J. Park, N. Lee, and R. Heath, Cooperative base
station coloring for pair-wise multi-cell coordination,
Submitted to IEEE Trans. Comm., 2015. [Online].
Available: http://arxiv.org/abs/1503.01102
[15] J. Andrews, F. Baccelli, and R. Ganti, A tractable approach to coverage and rate in cellular networks, IEEE
Trans. Comm., vol. 59, no. 11, pp. 31223134, Nov.
2011.
2490
MULTI-DIMENSIONAL CONTINUOUS PHASE MODULATION IN UPLINK OF MIMO

SYSTEMS
Mohammad A. Sedaghat , Ralf R. Muller
Norwegian University of Science and Technology, Trondheim, Norway

Friedrich-Alexander Universitat Erlangen-Nurnberg, Erlangen, Germany
ABSTRACT
Phase Modulation on the Hypersphere (PMH) is considered

in which the instantaneous sum power is constant. It is shown
that for an i.i.d. Gaussian channel, the capacity achieving input distribution is approximately uniform on a hypersphere
when the number of receive antennas is much larger than the
number of transmit antennas. Moreover, in the case that channel state information is not available at the transmitter, it is
proven that the capacity achieving input distribution is exactly uniform on a hypersphere. Mutual information between
input and output of PMH with discrete constellation for an
i.i.d. Gaussian channel is evaluated numerically. Furthermore, a spherical spectral shaping method for PMH is proposed to have Continuous Phase Modulation on the Hypersphere (CPMH). In CPMH, the continuous time signal has a
constant instantaneous sum power. It is shown that using a
spherical low pass filter in spherical domain followed by a
Cartesian filter results in very good spectral properties.
Index Terms Phase modulation, multiple-input multipleoutput (MIMO) systems, peak-to-average power ratio (PAPR),
single-RF transmitters, continuous phase modulation (CPM),
spherical filtering.
1. INTRODUCTION
Load modulated single-RF Multiple-Input Multiple-Output
(MIMO) transmitters have been proposed recently in [1] as
an efficient implementation method of MIMO transmitters.
In [1] and [2], load modulated massive MIMO transmitters
have been shown to have high power efficiency and allow
for a compact implementation in massive MIMO base stations. In load modulated MIMO transmitters, a central power
amplifier is used for all antennas in contrast to the standard
implementation method which uses one amplifier per antenna. The power efficiency of the central power amplifier in
load modulated MIMO transmitters is affected by the Peak
to Average Sum Power Ratio (PASPR). Due to a large number of antennas at massive MIMO base stations, by using
This work was supported in part by the European FP7 project HARP
(High capacity network architecture with remote radio heads & parasitic antenna arrays) under grant agreement no. 318489.
978-0-9928626-3-3/15/$31.00 2015 IEEE
2491
load modulated MIMO transmitters, high Peak to Average

Power Ratio (PAPR) signals such as Orthogonal FrequencyDivision Multiplexing (OFDM) signals can be transmitted
using an efficient power amplifier with low back-off [1]. This
is, in fact, because of the PASPR decaying with the number
of antennas [1].
Using load modulated MIMO transmitters in user terminals does not reduce PASPR as much as in base stations. This
is due to the low number of antenna elements in user terminals.
In [3], a novel modulation technique called Phase Modulation on the Hypersphere (PMH) has been proposed to use
in load modulated single-RF MIMO transmitters with low to
moderate number of antennas in order to have better power
efficiency1. In PMH, the sum power is fixed, i.e., x x =
constant; therefore, the central power amplifier in load modulated single-RF MIMO transmitters requires no back-off. In
fact, the signal at the central power amplifier has a PASPR of
0dB. PMH can be considered as a generalized form of classical phase modulation for multi-antenna applications. Note
that PMH is a different approach than the method proposed
in [4] for downlink of massive MIMO systems.
In [3], the sum capacity of PMH in an identity channel2
has been derived and it has been shown that the capacity is
achieved by a signal distributed uniformly on the surface of
a hyperball. Discrete constellation PMH has been also proposed in [3] as a set of points distributed uniformly on a
hypersphere. Moreover, in [3], the performance of discrete
constellation PMH has been evaluated by using some known
bounds on spherical codes.
In this paper, we investigate PMH in Gaussian i.i.d. channels. It is shown that in two scenarios, the capacity is achieved
by the uniformly distributed signal on a hypersphere: 1) when
the number of receive antennas is much larger than the number of transmit antennas, 2) when Channel State Information
(CSI) is only available at the receiver. Note that both of these
assumptions are valid in a massive MIMO uplink channel.
We also evaluate the mutual information between the input
1 Note that the new modulation is called Multi-dimensional Phase Modulation in [3]; however, in the current paper we change the name to Phase
Modulation on the Hypersphere for better intuition.
2 By identity channel we refer to the channel y = x + n.
and the output of a discrete constellation PMH in Gaussian

i.i.d. channels, numerically.
Furthermore, in this paper, we introduce Continuous
Phase Modulation on the Hypersphere (CPMH) which has
constant instantaneous sum power in continuous time domain, i.e., x(t) x(t) = constant. CPMH is constructed by
applying a pulse shaping filter to PMH signals. In this paper, a novel spherical pulse shaping filter is proposed using
the spherical filtering introduced in [5]. The spherical pulse
shaping filter does not affect the peak to average ratio of the
sum power. It is shown that to get an appropriate spectrum,
we need a Cartesian pulse shaping filter after the spherical
pulse shaping filter. Note that the spectral shaping method
proposed in this paper can be also used in CPM per antenna
proposed in [4].
The rest of this paper is organized as follows: in Section
2, capacity analysis of PMH in a point to point MIMO system
with large number of receive antennas is presented. Section 3
discusses PMH capacity with no CSI at the transmitter. Spectral shaping is presented in Section 4. Section 5 shows the
numerical results and finally Section 6 concludes the paper.
2. CHANNEL CAPACITY OF PMH IN A POINT TO
POINT MIMO CHANNEL WITH LARGE NUMBER
OF RECEIVE ANTENNAS
In a massive MIMO system, the aggregate number of all active user antennas is much smaller than the number of base
station antennas. In this section, a massive MIMO uplink
channel with only one active user is considered, i.e., the interference by the other users is neglected. Assume that the
base station and the user are equipped with N and M antennas, respectively, with N M (e.g., M = 2 and N = 100).
Consider the discrete channel model
y = Hx + n,
(1)
where x is the input, y is the output, H is the channel matrix

and n N (0, 2 I) is the noise vector. The transmitter sends
a PMH signal with fixed sum power. Let
x x = M,
(2)
without loss of generality. The elements of the channel matrix

are assumed to be i.i.d. Gaussian and perfectly known at both
transmitter and receiver. The capacity in such a system is
described as
C=
max
f (x), x x=M
I(x; y),
(3)
where f (x) is the probability density function (pdf) of input signal. By using the singular value decomposition of the
channel matrix, we have
y = Hx + n = U V x + n,
(4)
2492
where U and V are unitary matrices and is an N M matrix with at most M nonzero diagonal elements. To simplify
the problem, we consider an equivalent channel
y p = xp + np ,
(5)
where xp = V x, np = U n and y p = U y. Note that np

has the same statistics as n and xp xp = M ; therefore, the
channel described in (5) is equivalent to the channel in (4).
Let u be the matrix containing the first M rows of . It
can be shown that the capacity of channel (5) is equal to the
capacity of
y p,u = u xp + np,u ,
(6)
where y p,u and np,u consist of the first M elements of y p and

np , respectively.
The nonzero singular values of H are the square roots
of the eigenvalues of H H. Furthermore, for N/M growing large, all the eigenvalues of H H converge to the same
values. Thus, the matrix u converges to a diagonal matrix
with identical diagonal elements. Using this asymptotics and
considering the findings in [3], one can conclude that the optimum input distribution is approximately the uniform distribution on the hypersphere x x = M .
Note that, the exact capacity achieving input distribution
is not known in this case. Furthermore, it is not even clear
whether the exact capacity achieving input distribution in a
known channel is continuous or discrete. However, the analN
ysis in this section shows that if M
1, uniform distribution
on a hypersphere is a good approximation for the capacity
achieving distribution.
3. CHANNEL CAPACITY IN A POINT TO POINT
MIMO SYSTEM WITH NO CSI AT TRANSMITTER
In this section, channel capacity of PMH in a MIMO system
is discussed when there is no CSI at the transmitter and perfect CSI at the receiver. This is, in fact, the common scenario
considered in massive MIMO uplinks. In massive MIMO systems, the users send pilots and the base stations estimate the
channel, hence there is no CSI at transmitter.
Lets consider the channel model in (1). Furthermore, let
the elements of H be i.i.d. Gaussian. The mutual information
between the transmitter and the receiver is calculated as [6]
I(x; (y, H)) =
=
I(x; H) + I(x; y|H) = I(x; y|H)

E {I(x; y|H = H 0 )}
H0
(7)
To calculate the capacity achieving input distribution, we consider

y = HA Ax + n,
(8)
with A being a unitary matrix. It can be shown that HA and

H have the same distribution when the channel coefficients
are i.i.d. Gaussian [6]. Therefore, in the case of no CSI at the

transmitter, an equivalent channel is
y = HAx + n.
(9)
This actually shows that xopt and Axopt have the same distribution, where xopt denotes the capacity achieving input.
Furthermore, A can be any arbitrary unitary matrix. Therefore, xopt has to be distributed uniformly on the hypersphere
x x = M .
In [3], PMH with discrete constellation has been proposed. It is a more feasible modulation to use in practice. The
constellation is made of uniformly distributed points on a hypersphere. To have an estimate of the mutual information of
PMH with uniformly distributed inputs on a hypersphere, we
consider discrete constellation PMH. The mutual information
between the input and the output in this case is presented in
the numerical results section.
4. SPECTRAL SHAPING IN CPMH
In this section, we present a novel spectral shaping method to
construct the continuous time modulation CPMH. In most of
wireless communication systems, pulse shaping is done using a low-pass filter in digital domain. For instance, RootRaised-Cosine (RRC) filters are used widely in communication systems. Nevertheless, this type of pulse shaping increases PAPR. In PMH, constellation points are on the surface
of a hyperball and therefore, it meets PASPR = 0dB in discrete domain. However, if we use an RRC filter, the PASPR
increases.
To design a pulse shaping method for CPMH, we need a
method to find a continuous signal on a hypersphere which
passes through all the constellation points while is has good
spectral properties. Note that a geodesic path does not do a
good job due to corners.
In the classical modulation techniques in MIMO systems,
pulse shaping is done for every antenna individually. However, for CPMH we need to do a joint pulse shaping. In the
single-antenna case, CPMH is identical to the classical CPM
modulation. Therefore, one can employ any low-pass filter
to the phase component of the complex plane, e.g., a Gaussian filter in GMSK. However, in the case of several antennas,
pulse shaping is not straightforward and applying a low-pass
filter to the various phase components in a spherical coordinate system results in a wide spectrum. This is in fact due
to the multiplications of sine and cosine functions when converting from spherical to Cartesian coordinates. One may use
interpolation methods on the hypersphere, e.g., [7], as pulse
shaping. However, we look for a general solution of filtering
on hypersphere.
In [5], spherical filtering has been introduced by Buss
et al. which in summary is as following: consider a filter
with impulse response fn . In a Cartesian coordinate system, filtering a data stream xn can be done by the convolu-
2493
Fig. 1. Averaging in spherical coordinate system.

tion i fni xi which is a weighted average with coefficients
fni . Now, to apply the filter in the spherical domain, the key
point is the relation between taking average in spherical and
Cartesian coordinate systems. Note that in [5], it is shown
that for spherical filtering, the filter coefficients should meet
fi 0 and
fi = 1.
(10)
Let si be some constellation points on a hypersphere and also

sav be the average of the points on the hypersphere. Fig. 1
shows an example of a three dimensional case in which sav
i
is at the south pole of the sphere. Moreover, let the point s
be obtained by projecting the points si on the tangent plane
at the point sav , where the projection is done such that the
distance between si and sav on the hypersphere is equal to
i and sav on the tangent plane. Then,
the distance between s
the point sav is the average of points si in the spherical coi in the
ordinate system, if sav is the average of the points s
Cartesian coordinate system.
Filtering on a hypersphere cannot be done in a single step
but requires iterations [5]. In [5], some iterative methods are
proposed based on the similarity mentioned above for filtering in spherical and Cartesian coordinate systems. Here, we
employ the algorithm called Algorithm A1 in [5]. For a hypersphere with radius 1, the algorithm is summarized as follows [5]:
1. Select
an initial point sav .

i fni si .

fni si
A good initial guess is
2. Project all the points on hypersphere to the tangent plane

i .
at the point sav . We call them s

i .
av = i fni s
3. Calculate the weighted average s
av is close enough to sav , project s
av back to the
4. Unless s
surface of hypersphere, call it sav , and go back to step 2.
12
12
11
Mutual information (bits/sym)
EH log2 det 1 + H H/

2
10
Mutual information (bits/sym)
10
8
6
32 points
16 points
8 points
4 points

EH log2 det 1 + H H/2
9
8
128 points
64 points
32 points
16 points
8 points
4 points
2
1
0
0
10
SNR (dB)
15
0
0
20
Fig. 2. The mutual information of discrete PMH with 2 antennas vs. SNR in i.i.d. Gaussian channel.
g2
In this paper, we use the filter coefficients kg2 for spheri i

ical filtering, where gi are the RRC filter coefficients. We call
this filter RRC2 . Our investigations show that filtering in the
spherical domain results in insufficient sidelobe suppression;
therefore, we propose to use a low-pass filter in the spherical
domain followed by a low-pass filter in the Cartesian domain.
The good point is that the second filter hardly increases the
PASPR since it only affects sidelobes with very low energy.
5. NUMERICAL RESULTS
5.1. Numerical results on mutual information of discrete
constellation PMH in Gaussian i.i.d. channels
To show the performance of discrete constellation PMH, we
estimate the mutual information of it using the toolbox proposed in [8]. Furthermore, we use the spherical codes as the
constellation points as explained in [3].
The channel coefficients are assumed to be i.i.d. Gaussian. For each realization of the channel, we first estimate the
mutual information using 105 symbols, and then, averaging is
done over 104 channel realizations. The number of the antennas at receiver and transmitter are assumed to be equal. The
results for 2 and 3 antennas are shown in Fig. 2 and Fig. 3, respectively. The mutual information for Gaussian input is also
plotted for sake of comparison. The figures show that PMH
with fixed constellation has an acceptable mutual information
compared to the mutual information of Gaussian input.
2494
SNR (dB)
10
15
Fig. 3. The mutual information of discrete PMH with 3 anten-
nas vs. SNR in i.i.d. Gaussian channel.

5.2. Numerical results on spectral shaping
Here, the numerical results of the spherical filtering are presented. An oversampling factor of 32 was used to simulate
continuous time signals. We consider a standard RRC filter
with length 1024 and roll-off factor 0.3 .
Fig. 4 shows the result for 4 antennas. The figure shows
that the RRC2 filter in the spherical domain results in only
17dB sidelobe suppression. However, by accepting a small
PASPR, we can use a post filtering by adding an RRC filter in
the Cartesian coordinate system, i.e., per antenna filtering. In
fact, the spherical filter shapes the spectrum approximately,
and the second filter kills the side lobes. The main point is
that the first filter keeps PASPR at 0dB and the second filter
operates on side lobes and changes PASPR only slightly. Note
that for the spherical RRC filter, reducing the roll-off factor
does not improve the spectrum and a roll-off around 0.3 has
the best performance.
Due to lack of space, further investigations are referred to
the extended version of this paper.
Next, the resulted PASPR due to using the second filter is
plotted in Fig. 5 versus the number of antennas. It is observed
that the spherical filtering followed by a Cartesian post filtering reduces PASPR by about 0.5dB 0.8dB compared to the
Cartesian RRC filtering. Note that in this paper we use spherical codes as constellation points which have phase shifts of
up to . In the longer version of this paper, we propose some
other mapping with better PASPR properties.
Note that the best Inter Symbol Interference (ISI) free low
pass filter for CPMH is unknown and designing appropriate
filters can still be done in the future.
ing input is uniformly distributed on a hypersphere when

channel coefficients are i.i.d. Gaussian. Numerical results on
mutual information of discrete constellation PMH in an i.i.d.
Gaussian channel were also presented.
Furthermore, a new spectral shaping method for CPMH
was proposed using spherical filtering. It was shown that using two filters in spherical and Cartesian domains results in
excellent spectral properties while the PASPR of the systems
is hardly affected.
Sphe. RRC2
Sphe. RRC2 + Cart. RRC
10
0
PSD (dB)
10
20
30
40
Acknowledgment
50
The authors would like to thank Ali Bereyhi for helpful comments.
60
70
REFERENCES
80
2.5 2 1.5 1 0.5 0 0.5 1 1.5 2 2.5
Normalized frequency (f T )
Fig. 4. PSD versus normalized frequency for different filters
for 4 antennas.
4.5
[2] R. R. Muller, M. A. Sedaghat, and G. Fischer, Load

Modulated Massive MIMO, in IEEE GlobalSIP, Massive MIMO Symposiumum, Atlanta, USA, 2014, pp. 14.
[3] M. A. Sedaghat and R. R. Muller, A novel modulation scheme for user devices equipped with a single-RF
MIMO tansmitter, in Proc. ITG Workshop on Smart Antennas (WSA), Ilmenau, Germany, Mar. 2015.
Cart. RRC (uniform mapping)

Sphe. RRC2 + Cart. RRC (uniform mapping)
PASPR (dB)
3.5
[4] S. K. Mohammed and E. G. Larsson, Per-antenna constant envelope precoding for large multi-user MIMO systems, Communications, IEEE Transactions on, vol. 61,
no. 3, pp. 10591071, 2013.
3
2.5
[5] S. R. Buss and J. P. Fillmore, Spherical averages and

applications to spherical splines and interpolation, ACM
Transactions on Graphics (TOG), vol. 20, no. 2, pp. 95
126, 2001.
2
1.5
1
1
[1] M. A. Sedaghat, R. R. Muller, and G. Fischer, A Novel

Single-RF Transmitter for Massive MIMO, in Smart Antennas (WSA), 2014 18th International ITG Workshop
on. Erlangen, Germany: VDE, 2014, pp. 18.
Number of antennas
[6] E. Telatar, Capacity of multi-antenna gaussian channels, European transactions on telecommunications,

vol. 10, no. 6, pp. 585595, 1999.
10
Fig. 5. PASPR versus the number of antennas for spherical RRC2 filtering followed by Cartesian post filtering and
a Cartesian RRC filtering.
6. CONCLUSION
In this paper, some new results on the capacity of PMH in
Gaussian channels were presented. It was proven that in the
uplink of massive MIMO with one user, the capacity achiev-
2495
[7] A. Gfrerrer, Rational interpolation on a hypersphere,

Computer Aided Geometric Design, vol. 16, no. 1, pp.
2137, 1999.
[8] Z. Szabo, Information theoretical estimators toolbox,
The Journal of Machine Learning Research, vol. 15,
no. 1, pp. 283287, 2014.
MIMO AND MASSIVE MIMO - ANALYSIS FOR A LOCAL AREA SCENARIO

Stefan Dierks , Wolfgang Zirwas , Markus Jager , Berthold Panzner , and Gerhard Kramer
Institute for Communications Engineering

Technische Universitat Munchen, Munich, Germany
{stefan.dierks, markus.jaeger, gerhard.kramer}
@tum.de
ABSTRACT
The performance of centralized and distributed deployments
of massive MIMO in an office building is analyzed both with
and without cooperation. It is shown that using twice as
many base station antennas as data streams provides most of
the massive MIMO benefits. A simple transmission scheme
achieves user fairness and operates close to a capacity upper
bound. An example scheduling algorithm improves efficiency
only for less than twice the number of base station antennas
as data streams. The tradeoff between performance and cost
for backhauling is evaluated by comparing cooperation of
distributed base stations with a single central deployment.
Index Terms MIMO; massive MIMO; network MIMO;
5G; two stripe building
1. INTRODUCTION
One goal of new mobile communication standards (e.g. 5G)
is to increase the spectral efficiency per unit area or volume. One way to increase spectral efficiency is by using
multiple-input multiple-output (MIMO) methods. MIMO allows one node to transmit several streams to one or more user
equipments (UE) using spatial degrees-of-freedom. Massive MIMO, a vast over-provisioning of base station (BS)
antennas, lets simple transmission schemes achieve large
performance gains over todays systems [1].
The terminology massive MIMO is not clearly defined.
Massive MIMO may refer to any MIMO configuration beyond the highest MIMO mode in current LTE (at present 8x8),
or it may refer to simply a large number of antennas at the
BSs. A somewhat more precise way to define massive MIMO
is to relate it to the ratio of serving antennas to active UEs.
Most massive MIMO studies consider wide area outdoor
deployments [2]. However most mobile traffic is generated by
indoor users [3]. We analyze the performance of centralized
and distributed deployments with and without cooperation for
the 3GPP two stripe office building [4]. We fix the number of
S. Dierks, M. J
ager and G. Kramer were supported by the German
Ministry of Education and Research in the framework of an Alexander von
Humboldt Professorship.
978-0-9928626-3-3/15/$31.00 2015 IEEE
2496
Nokia Networks
Munich, Germany
{wolfgang.zirwas, berthold.panzner}
@nokia.com
active, single antenna UEs and sweep the ratio of total number of BS antennas to the number of active UEs from one to
10-times more serving antennas. We find that a ratio of twice
as many BS antennas provides most of the massive MIMO
benefits. We further find that this ratio is a good tradeoff
between number of antennas versus spectral efficiency. We
present suboptimal transmission schemes that approach a capacity upper bound. An example scheduler improves performance if there are fewer than twice as many BS antennas.
We analyze fairness using Jains index. Placing a single massive MIMO BS at the center of a building is intuitively not
an optimal choice as the UEs suffer from large transmitter-toreceiver distances and high wall penetration loss. We compare this deployment to distributed BSs with cooperation. We
find that distributed indoor BSs with cooperation achieve a
substantial performance gain at the cost of a backhaul connection, while the gain achieved with cooperation between
outdoor small cells and a single indoor BS is smaller.
The following aspects of this study are novel as compared
to [5] and [6]: the per-BS power constraint, indoor-outdoor
cooperation, the example scheduling algorithm and the analysis of a capacity upper bound and fairness.
2. SYSTEM MODEL
Consider the downlink in the two stripe building in Fig. 1,
defined by 3GPP as the A1 indoor office scenario in the WINNER II deliverable [7]. The UEs are served by BSs located
inside and outside the building.
We consider single antenna UEs and orthogonal frequencydivision multiplexing (OFDM). For each subcarrier we obtain
a MISO broadcast channel. The received signal of the k-th
UE for one subcarrier is
yk = hH
k x + zk
k = 1, . . . , K
(1)
T
T
where hk = [hT
k,1 , . . . , hk,NBS ] is the collection of the channel coefficients from all NBS BSs to the k-th UE. The i-th
BS has Mi BS antennas with the channel coefficients hk,i .
The length
PNBS of hk is equal the total number of BS antennas
M = i=1
Mi . The transmitted signals vectors are collected
y = Hx + z
(2)
where H = [h1 , . . . , hK ]H and z = [z1 , . . . , zK ]T .

For linear precoding the transmitted signals vector x is
x = Ws
(3)
where W = [w1 , . . . , wK ] is the matrix of the precoding vectors and s = [s1 , . . . , sK ] is the vector of transmit symbols.
We consider a per-BS sum-power constraint
NSC

X
(f ) 2
xi Pi
f =1
i {1, NBS }
(4)
where NSC is the number of subcarriers. We omit the subcarrier index (f ) for clarity in the remaining paper.
We assume that ideal hardware and perfect channel state
information of the complete network is available at all nodes.
40
20
2
distance [m]
T
T
in x = [xT
1 , . . . , xNBS ] and zk is independent proper com2
plex thermal Gaussian noise with variance N
. HH denotes
the complex conjugate transpose of matrix H. We place the
BS antennas with an antenna spacing of /2. Mutual coupling between array elements is ignored. The number of UEs
is K, which is equal to the number of receive antennas as we
consider single antenna UEs. The received signals of all UEs
y = [y1 , . . . , yK ]T for one subcarrier are
1
2
-20
-40
-40
-20
0
distance [m]
20
40
Fig. 1. Deployments in the two stipe building.
3.2. Network MIMO

For Network MIMO we assume a perfect backhaul between
the BSs and treat distributed BSs as a single BS. Then we can
apply ZFBF either in a central unit or by exchanging cooperation messages. We initialize the power allocation as for
massive MIMO and fulfill the per-BS power constraint (4) by
scaling all (f ) . A better approach would be to design the precoders and allocate the power (considering a maximal modulation scheme) to maximize the spectral efficiency under a
fairness constraint, but this adds complexity.
3. TRANSMISSION SCHEMES
3.3. Local Precoding
We use zero forcing beamforming (ZFBF) where the linear
precoders are determined according to an interference zero
forcing objective. The optimal solution given a sum power
constraint is the pseudo-inverse combined with a power allocation [8]
W = HH (HHH )1 diag()
(5)
where is the power allocation vector of non-negative reals.
With this choice of precoding matrix the received signals are
y = HWs + z = HHH (HHH )1 diag() s + z
= diag() s + z.
If the backhaul cannot enable Network MIMO or any other

kind of cooperation or coordination, a BS determines its precoders locally while treating interference from the other BSs
as noise. Each UE connects to the BS with the maximum average SNR. If the number of UEs connected to a BS is larger
than the number of BS antennas, scheduling with channel decomposition using a capacity upper bound [9] is used. We call
this scheme Local Precoding.
(6)
(7)
3.1. Central Massive MIMO

If there is only one BS and all antennas are located at one
physical site, ZFBF eliminates all interference. For a sum
rate maximization with a total power constraint, the power
allocation can be solved using water filling [8]. We choose
to distribute the power equally to the transmission to each
UE to ensure fairness between the UEs. The distribution of
the per-UE power on the physical resource blocks (PRB) is
determined by water filling assuming Gaussian signals. We
consider scenarios where the total number of BS antennas M
is larger than or equal to the number of UEs K. Hence we
are able to schedule all UEs in every time frame and on every
PRB with ZFBF. We call this Central Massive MIMO.
2497
4. DEPLOYMENTS
We define six different BS deployments which are shown in
Fig. 1. The indoor BSs are rectangular arrays mounted underneath the ceiling, while the outdoor BSs are uniform linear arrays (ULA). Indoor Central Massive MIMO is a single
central BS that uses Central Massive MIMO (1 in Fig. 1).
Indoor Local Precoding uses Local Precoding for the four BSs
on the corridors (2), while Indoor Network MIMO requires
a backhaul to employ Network MIMO for the corridor BSs
(2+A). Outdoor Only uses only the outdoor ULAs with
Local Precoding (3). If we add a central BS to the outdoor
BSs we obtain Indoor-Outdoor Local Precoding (1+3)
and Indoor-Outdoor Network MIMO (1+3+B), which
employ Local Precoding, respectively Network MIMO. Note
that the BSs are not necessarily optimally placed.
5. SIMULATION RESULTS
24 CQAM bits 1200 14

,
(8)
S=
1ms 20 MHz
where 24 is the number of UEs, CQAM is the instantaneous capacity of a memoryless channel with QAM input and continuous output [11], 1200 is the number of subcarriers, 14 is the
number of OFDM blocks per subframe, 1ms is the duration of
one subframe and 20 MHz is the bandwidth. For 256-QAM
the maximal spectral efficiency is S = 161.28 bit/s/Hz.
In Fig. 2 the average spectral efficiencies of the deployments are shown for 24 to 240 total BS antennas. The 5%-tile
and the 95%-tile spectral efficiencies follow the same trend.
The Indoor Central Massive MIMO, Indoor Network MIMO
and Indoor-Outdoor Network MIMO deployments perform
poorly for fully loaded MIMO systems with 24 BS antennas.
The performance improves significantly when few antennas
are added (Scheduling improves the performance for fully or
close to fully loaded MIMO systems as discussed in Section
7). Adding more antennas increases the performance, but the
gain from each additional antenna decreases. A ratio of twice
as many BS antennas seems to be a good tradeoff between
achieved performance and number of antenna elements. The
performance loss of the Indoor Central Massive MIMO versus the Indoor Network MIMO deployment (mainly due to
2498
Number of drops
Number of channel realizations per drop
300
10
Table 1. Simulation parameters.

average spectral efficiency [bit/s/Hz]
We fix the number of UEs and compare the spectral efficiencies of the deployments for different numbers of total BS antennas. We define one drop as a random placement of 24 UEs
within the two stripe building. For each drop we generate 10
channel realizations.
The channel coefficients for indoor BSs are generated according to the WINNER II A1 indoor channel model [7]. The
A1 channel model provides parameter sets for line-of-sight
(LOS) and non line-of-sight (NLOS) conditions. For each
BS-UE pair, the number of walls between their positions are
determined and the appropriate condition is selected. A wall
penetration loss of 12 dB (as we assume heavy walls) for every wall beyond the first is applied. When determining the
number of walls, paths along the corridors are considered as
alternatives to the direct path, which might penetrate more
walls. The channel coefficients for outdoor BSs are generated
according to the WINNER II B4 outdoor (Urban Micro-Cell)
to indoor channel model [7]. Here we assume a LOS path
from the BS to the outside wall of the building. The number
of walls is determined as mentioned above. We use the QUAsi
Deterministic RadIo channel GenerAtor (QuaDRiGa) [10] to
generate the channels, which we enhanced to count the number of walls and apply the wall penetration loss.
We use a bandwidth of 20 MHz around a carrier frequency
of 2.1 GHz. According to LTE we obtain 100 PRBs, where
the precoders and the power allocations are determined per
PRB. The simulation parameters are summarized in the Table
1. With these parameters the spectral efficiency in the building without considering control signaling overhead is
2.1 GHz
20 MHz
18 MHz
15 kHz
1200
100
/2
12 dB
26 dBm
124.6 dBm
256-QAM
24
Carrier frequency
Bandwidth
Used bandwidth
Subcarrier spacing
Number of subcarriers
Number of PRBs
Antenna Spacing
Wall penetration loss
Sum power constraint
Noise level
Largest modulation scheme
Number of UEs
160
100
40
24
48
96
192
number of total BS antennas

Indoor Central Massive MIMO
Indoor Local Precoding
Indoor Network MIMO
Outdoor Only
Indoor-Outdoor Local Precoding
Indoor-Outdoor Network MIMO
Fig. 2. Average spectral efficiencies.
wall penetration loss) is analyzed in more detail in [5]. The

performance loss of the Indoor-Outdoor Network MIMO versus the Indoor Network MIMO is due to wall penetration
loss and building penetration loss. Note that Indoor Network
MIMO approaches the maximal spectral efficiency for only
48 BS antennas, while Indoor Central Massive MIMO requires more than 240 antennas to approach the maximal spectral efficiency. The non-cooperative schemes perform worse
due to interference. Their performances improve little per additional antenna.
6. COMPARISON WITH CAPACITY UPPER BOUND
Each deployment can be upper bounded by a broadcast channel (BC) with a sum power constraint. We allow all BSs of a
300
dashed: Gaussian signals
solid: 256 QAM
dotted: 16 QAM
210
120
30
24
48
96
192
180
All UEs scheduled

Example Scheduling
Capacity upper bound
100
dashed: Gaussian signals
solid: 256 QAM
dotted: 16 QAM
20
24
48
96
192

Capacity upper bound Indoor Central Massive MIMO
Indoor Network MIMO
Capacity upper bound Indoor Network MIMO
Fig. 4. Average spectral efficiencies for the Indoor Central
Massive MIMO deployment with and without scheduling.
Fig. 3.
Average spectral efficiencies and capacity upper

bounds for the Indoor Central Massive MIMO and Indoor
Network MIMO deployment.
deployment to cooperate and to act as one BS with distributed
antennas, and relax the per-BS power constraint to a sum
power constraint. The capacity of a BC is achieved by dirtypaper coding. We find the optimal transmission policy with
the algorithms in [12]. As the spectral efficiency is limited
by the modulation scheme we examine Gaussian signaling
and 16-QAM in addition to 256-QAM. In Fig. 3 the average
spectral efficiencies and the capacity upper bounds of the Indoor Central Massive MIMO and Indoor Network MIMO deployment are shown. While the 16-QAM and the 256-QAM
schemes limit the average spectral efficiency, Gaussian signaling allows the spectral efficiency to grow and to approach
the capacity upper bound. Notice that for a ratio of two BS
antennas per UE the gap between the capacity upper bound
and our schemes is already small and a good tradeoff between
performance and number of BS antennas is achieved also for
Gaussian signaling and 16-QAM.
channel gains. Algorithm 1 utilizes this idea. Note that fairness is not considered explicitly, but implicitly as the per UE
transmit powers are equally distributed.
Algorithm 1 Example scheduling algorithm
Initialize all PRBs as scheduled for all UEs
repeat
for all PRBs do
ZFBF with all UEs scheduled on current PRB
end for
for all UEs do
water filling for current UE
determine new schedule for current UE as all PRBs,
where a power larger than zero is allocated
end for
until Convergence of all schedules
In Fig. 4 the average spectral efficiencies for the Indoor
Central Massive MIMO deployment are shown for all UEs
scheduled and for the example scheduling. As expected a performance gain with scheduling can be observed only for less
than twice as many BS antennas as UEs. Similar results are
obtained for the other deployments.
7. WHEN IS SCHEDULING NECESSARY

8. FAIRNESS
One claim of massive MIMO is that for sufficient randomness
and many antennas the channels hardens and scheduling all
UEs is optimal, which means that more advanced scheduling
does not provide gains [2]. We use an example scheduler to
analyze how many BS antennas are required for this claim to
be valid in our scenario.
When we use the water filling power allocation as described in Section 3 there is usually no power allocated to
some UEs on some PRBs. When no power is allocated to
an UE the interference zero-forcing constraint for that UE is
unnecessary on that PRB and may be removed. This creates
more degrees-of-freedom when determining the precoders
for the other UEs, which leads to potentially higher effective
2499
We analyze fairness qualitatively with Jains index [13]

J (r1 , r2 , . . . , rK ) =
PK
2
i=1 ri )
PK
i=1 ri
(9)
where rk is the rate achieved by the k-th UE. Jains index

approaches 1 when all UEs achieve similar rates and is 1/K
when only one UE achieves a non-zero rate. The fairness indices are plotted in Fig. 5. The deployments with cooperation
approach a fairness index of 1 for an increasing number of
antennas as more and more UEs are served with the maximal
spectral efficiency. Scheduling helps to increase fairness for a
Jains fairness index
Trans. Wireless Commun., vol. 9, no. 11, pp. 3590

3600, Nov. 2010.
[2] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta,
Massive MIMO for next generation wireless systems,
IEEE Commun. Mag., vol. 52, no. 2, pp. 186195, Feb.
2014.
0.7
[3] J. Zhang and G. de la Roche, Eds., Femtocells: Technologies and Deployment, Wiley, 2013.
0.4
24
48
96
[4] 3GPP, TR36.814 - Further advancements for E-UTRA

physical layer aspects, Tech. Rep. v9.0.0, Mar. 2010.
192

Indoor Local Precoding
Indoor Network MIMO
Outdoor Only
Indoor-Outdoor Local Precoding
Indoor-Outdoor Network MIMO
Indoor Central Massive MIMO with Example Scheduling
Fig. 5. Jains fairness index.
[5] S. Dierks, M.B. Amin, W. Zirwas, M. Haardt, and

B. Panzner, The Benefit of Cooperation in the Context of Massive MIMO, in Proc. of 18th Int. OFDM
Workshop 2014, Aug. 2014, pp. 18.
[6] B. Panzner, W. Zirwas, S. Dierks, M. Lauridsen, P. Mogensen, K. Pajukoski, and D. Miao, Deployment and
Implementation Strategies for Massive MIMO in 5G,
in IEEE Global Telecommun. Conf. (Globecom) 2014
Workshop on Massive MIMO: From Theory to Practice, Dec. 2014.
small number of antennas. The deployments without cooperation saturate at a fairness index of less than 0.7.
[7] IST, D1.1.2 - WINNER II Channel Models, Tech.

Rep. v1.2, 2008.
9. CONCLUSIONS
[8] A. Wiesel, Y.C. Eldar, and S. Shamai, Zero-Forcing

Precoding and Generalized Inverses, IEEE Trans. Signal Proc., vol. 56, no. 9, pp. 44094418, Sep. 2008.
We compared the performance of different deployments in

the 3GPP two stripe office scenario. The same performance
is achieved by a single massive MIMO BS or by distributed
BSs with cooperation and less antennas. The costs of antenna
elements can be traded off with the costs for a backhaul to
achieve the same performance. A ratio of twice as many BS
antennas as served UEs offers many of the massive MIMO
benefits. User fairness and spectral efficiency close to capacity are achieved with a simple transmission scheme.
A spectral efficiency of 113 bit/s/Hz without considering overhead is achievable with the Indoor Central Massive
MIMO deployment with 96 antennas and with Indoor Network MIMO with less than 28 antennas. Considering an overhead of 50% the required bandwidth to achieve the goals of
the METIS (Mobile and wireless communications Enablers
for the Twenty-twenty Information Society) project [14] is
for the TC1 virtual reality office:
0.1 Gbit/s/m2 x 5000 m2 / 56.5 bit/s/Hz 8.85 GHz
(More UE antennas, more base stations or larger QAM
constellations could reduce the required bandwidth),
for the TC2 dense urban information society:
0.7 Mbit/s/m2 x 5000 m2 / 56.5 bit/s/Hz 62 MHz.
REFERENCES
[1] T.L. Marzetta, Noncooperative Cellular Wireless with
Unlimited Numbers of Base Station Antennas, IEEE
2500
[9] X. Zhang and J. Lee, Low complexity MIMO scheduling with channel decomposition using capacity upperbound, IEEE Trans. Commun., vol. 56, no. 6, pp. 871
876, Jun. 2008.
[10] S. Jaeckel, L. Raschkowski, K. Brner, L. Thiele,
F. Burkhardt, and E. Eberlein, QuaDRiGa-Quasi Determinsitic Radio Channel Generator, User Manual and
Documentation, Tech. Rep. v1.2.3-307, Fraunhofer
Heinrich Hertz Institute, 2014.
[11] G. Ungerboeck, Channel coding with multilevel/phase
signals, IEEE Trans. Inf. Theory, vol. 28, no. 1, pp.
5567, Jan. 1982.
[12] N. Jindal, W. Rhee, S. Vishwanath, S.A. Jafar, and
A. Goldsmith, Sum power iterative water-filling for
multi-antenna Gaussian broadcast channels, IEEE
Trans. Inf. Theory, vol. 51, no. 4, pp. 15701580, Apr.
2005.
[13] R. Jain, D.M. Chiu, and W. Hawe, A Quantitative
Measure of Fairness and Discrimination for Resource
Allocation in Shared Computer Systems, Tech. Rep.
301, DEC Research Report, 1984.
[14] Mobile and wireless communications Enablers for the
Twenty-twenty Information Society (METIS), Deliverable D1.1 - Scenarios, requirements and KPIs for 5G
mobile and wireless system, Tech. Rep., Apr. 2013.
PERFORMANCE COMPARISON OF DATA-SHARING AND COMPRESSION STRATEGIES

FOR CLOUD RADIO ACCESS NETWORKS
Pratik Patil, Binbin Dai, and Wei Yu
Department of Electrical and Computer Engineering
University of Toronto, Toronto, Ontario M5S 3G4, Canada
Emails: {ppatil, bdai, weiyu}@comm.utoronto.ca
ABSTRACT
This paper provides a system-level performance comparison
of two fundamentally different transmission strategies for the
downlink of a cloud radio access network. The two strategies,
namely the data-sharing strategy and the compression-based
strategy, differ in the way the limited backhaul is utilized.
While the data-sharing strategy uses the backhaul to carry raw
user data, the compression strategy uses the backhaul to carry
compressed beamformed signals. Although these strategies
have been individually studied in the literature, a fair comparison of the two schemes under practical network settings is
challenging because of the complexity in jointly optimizing
user scheduling, beamforming, and power control for systemlevel performance evaluation, along with the need to optimize
cooperation clusters for the data-sharing strategy and quantization noise levels for the compression strategy. This paper
presents an optimization framework for both the data-sharing
and compression strategies, while taking into account losses
due to practical modulation in terms of gap to capacity and
practical quantization in terms of gap to rate-distortion limit.
The main conclusion of this paper is that the compressionbased strategy, even with a simple fixed-rate uniform quantizer, outperforms the data-sharing strategy under medium to
high capacity backhauls. However, the data-sharing strategy outperforms the compression strategy under low capacity
backhauls primarily because of the large quantization loss at
low backhaul capacity with compression.
1. INTRODUCTION
The ultra-dense cell deployment in the next generation (5G)
wireless networks calls for efficient management of intercell interference. Cloud radio access network (C-RAN) has
emerged as a promising cellular architecture that allows joint
signal processing across base-stations (BSs) for interference
mitigation purposes whereby the BSs are connected to a
centralized cloud-computing based processor. This paper
compares the performance of two fundamentally different
transmission strategies for the downlink C-RAN, where the
BSs essentially act as relays in transmitting data from the
central processor to the remote users.
978-0-9928626-3-3/15/$31.00 2015 IEEE
2501
In the data-sharing strategy, the central processor shares

the data of each user to a cluster of BSs which then compute the beamformed signals to be transmitted. In the compression strategy, the central processor itself computes the
beamformed signals to be transmitted by each BS, which
are then quantized and sent to the BSs through capacitylimited backhaul links. Individually, both the data-sharing
and compression strategies have been studied in the context
of C-RAN. However, a fair system-level comparison between
the two strategies under practical network settings is still not
available in the literature due to the challenges in solving
the corresponding network optimization problems involving
user scheduling, beamforming, power control, along with the
optimization of clusters for the data-sharing strategy and the
optimization of quantization noise levels for the compression
strategy. This paper tackles such a system-level performance
evaluation and tries to answer the question of under what
condition one strategy performs better than the other.
One contribution of this paper is that we model and take
into account loss due to practical modulation schemes in
terms of gap to capacity for both strategies. In addition, for
the compression strategy, we introduce a similar notion of gap
to rate-distortion limit to account for quantization losses due
to non-ideal quantizers used in practice. Further, we propose
a novel algorithm for the joint optimization of the beamformers and quantization noise levels for the compression
strategy based on an equivalence between weighted sum rate
(WSR) maximization and weighted minimum mean square
error (WMMSE) optimization.
We show through simulations on a heterogeneous cellular topology that whether one strategy is superior to the other
largely depends on the backhaul capacity constraint in the system. If the available backhaul capacity is medium to high,
the compression strategy outperforms the data-sharing strategy, even with a simple fixed-rate uniform scalar quantizer.
However, if the available backhaul capacity is low, the datasharing strategy outperforms the compression strategy. Intuitively, under low backhaul capacity the quantization noises
introduced in the compression strategy dominate the interference, in which case it is better to just share the data directly
with a limited set of BSs rather than to compress.
We note that in our previous work [1], a comparison between the data-sharing strategy and the compression strategy
is made. But the system considered in [1] is limited to only a
sum backhaul constraint, instead of the per-BS backhaul constraints considered here. Moreover, in [1], the data-sharing
strategy does not select an optimized cluster of BSs for each
user; the compression strategy does not consider the joint optimization of the beamformers and the quantization noise levels; further only a fixed user scheduling is assumed.
This paper restricts attention to linear precoding strategies
and does not consider nonlinear precoding based on dirtypaper coding [2]. A hybrid between the data-sharing and
compression strategies is also possible and is discussed in [1].
For more references on the data-sharing strategy, we refer the
readers to [3] and for the compression strategy to [4].
2. SYSTEM MODEL
Consider a downlink C-RAN consisting of L single-antenna
BSs serving K single-antenna remote users. All L BSs are
connected to a central processor with capacity-limited backhaul links. (We use the term backhaul, because the links
carry digital data. These links are sometimes referred to as
fronthaul links in the C-RAN literature, especially when they
carry compressed analog signals.) The capacity of the backhaul link connecting lth BS to the central processor is denoted
by Cl , l = 1, . . . , L. We assume one data stream per user, and
that the central processor has access to the data and perfect
CSI for all K users in the network.
Let xl denote the complex signal transmitted by BS l and
x CL1 = [x1 , . . . , xL ]T be the aggregate signal from all
the BSs. The received signal at user k can be written as
y k = hH
k x + zk ,
k = 1, 2, . . . , K
(1)
where hk CL1 = [h1,k , . . . , hL,k ]T is the channel to the

user k from all the BSs, and zk is the additive complex Gaussian noise with zero-mean and variance 2 . Each BS l has a
transmit power budget denoted by Pl . Let sk denote the data
of kth user distributed as complex Gaussian with zero-mean
and unit variance, which is available at the central processor.
3. DATA-SHARING STRATEGY
In the data-sharing strategy, a cluster of BSs locally form
beamformers to cooperatively serve each user. The data for
that user is replicated at all the participating BSs in the cluster via the backhaul links. A crucial decision is to select an
appropriate cluster of BSs for each user for interference mitigation, while staying under the limited backhaul capacity.
Let the beamforming vector for user k from all the BSs be
wk CL1 = [w1,k , w2,k , . . . , wL,k ]T , where wl,k denotes
the component of the beamformer from BS l. If BS l does
not participate in cooperatively serving user k, then wl,k =
2502
0. The beamformed signals transmitted from all the BSs can

then be written as
K
X
x=
wk s k .
(2)
k=1
At user k, the signal-to-interference-plus-noise ratio (SINR)

can be expressed as
SINRk = P
2
|hH
k wk |
.
H
2
2
j6=k |hk wj | +
(3)
The information theoretical achievable rate for user k is related to SINR as Rk = log(1+SINRk ). However, this rate expression assumes Gaussian signaling, while in practice QAM
constellations are typically used for the Gaussian channel in
the moderate and high SINR regime. With moderate coding,
to achieve a given data rate we still need an SINR higher than
what is suggested above. This extra amount of power is usually captured by a so-called SNR gap. Denoting the gap by
m , we can rewrite the achievable rate for user k as
!
SINRk
.
(4)
Rk = log 1 +
m
The optimization problem of finding the optimal set of
BS clusters and beamformers for the data-sharing scheme can
now be formulated as a WSR maximization problem under
per-BS power constraints and per-BS backhaul constraints:
maximize
wl,k
subject to
K
X
k Rk
(5a)
k=1
K
X
|wl,k |2 Pl ,
(5b)
k=1
K
X

1 |wl,k |2 Rk Cl , l
(5c)
k=1
where k denotes the priority

weight associated with user k
and the indicator function 1 |wl,k |2 denotes if BS l participates in beamforming to user k, and if so, the user rate Rk
is included in the backhaul constraint Cl . The beamforming
coefficients are computed at the central processor, and are assumed to be transmitted to the BSs without any error. We
neglect the backhaul consumption for transmitting the beamformers. This formulation considers joint design of BS clustering and beamforming. It also implicitly does power control and user scheduling. This optimization problem is solved
repeatedly and the BS clusters are dynamically optimized in
each time slot as the priority weights are updated.
The presence of the backhaul constraint (5c) makes the
optimization problem challenging. In this paper, we follow
the approximation suggested in [3] to first write the indicator function as a l0 norm which is then approximated as a
weighted l1 norm as

1 |wl,k |2 = |wl,k |2 0 l,k |wl,k |2 ,
(6)
where l,k is a constant weight associated with BS l and user

k and is updated iteratively according to
1
l,k =
, k, l
(7)
|wl,k |2 +
for some regularization constant > 0 and |wl,k |2 from the
previous iteration. This simplifies the constraint (5c) to
K
X
l,k |wl,k |2 Rk Cl ,
(8)
The optimization of transmit beamformers {wk } under fixed

{uk }, {k } and fixed {Rk } is the following quadratically
constrained quadratic programming (QCQP) problem:
minimize
wl,k
subject to
Algorithm 1 WSR maximization for data-sharing strategy

Initialization: {l,k }, {wk }, {Rk };
Repeat:
1. For fixed {wk }, compute the MMSE receivers {uk } and
the corresponding MSE {ek } according to (11) and (9);
2. Update the MSE weights {k } according to (10);
3. For fixed {uk }, {k }, and {Rk } in (12c), find the optimal
transmit beamformer {wl,k } by solving (12);
4. Update {l,k } as in (7). Compute the achievable rate
{Rk } according to (4);
Until convergence
The quantities used in the WMMSE approach in the above
algorithm are as follows. The mean square error (MSE) for
user k is defined as
!
X
2
2
2
+ |hH
|hH
ek = |uk |2 m
k wk |
k wj | +
(9)
j6=k
2
H
Re{uH
k hk w k }
+ 1.
The optimal MSE weight k under fixed {wk } and {uk } is

given by
k = ek1 .
(10)
The optimal receive beamformer uk under fixed {wk } and
{k } is given by
1
!
X
2
2
2
+ |hH
uk = m
hH
|hH
k wk |
k wk .
k wj | +
j6=k
(11)
2503
wkH Ak wk Re{bH
k wk }
(12a)
|wl,k |2 Pl ,
(12b)
k=1
K
X
k=1
k=1
which is equivalent to a generalized power constraint, if Rk is

assumed fixed and heuristically chosen from the previous iteration. The resulting optimization problem can then be solved
using an equivalence between the WSR maximization and the
WMMSE problem.
The only difference between the formulation (5) and that
in [3] is the gap factor m . We can easily verify that the
equivalence between WSR optimization and WMMSE extends even with the gap m . Below we summarize the overall algorithm for the optimization of the data-sharing strategy.
Though we do not have theoretical guarantee of its convergence in general, it is observed to converge in simulations.
K
X
K
X
Rk l,k |wl,k |2 Cl ,
(12c)
k=1
where Ak CLL and bk CL1 are defined to be

X
2
H
Ak =
j j |uj |2 m hj hH
j + k k |uk | hk hk
(13)
j6=k
bk = 2k k uk hk
(14)
4. COMPRESSION STRATEGY
In the compression strategy, the central processor computes
the beamformed analog signals to be transmitted by the BSs.
These signals have to be compressed before they can be forwarded to the corresponding BSs through the finite-capacity
backhaul links. The process of compression introduces quantization noises; the quantization noise levels depend on backhaul capacities.
In the data-sharing strategy, the beamformed signal x as
given by (2) is computed at the BSs. In the compression strategy, x is computed at the central processor, then compressed,
sent over the backhaul links, and reproduced by the BSs. We
model the quantization process for x as
= x + e,
x
(15)
where e is the quantization noise with covariance matrix Q

CLL modelled as complex Gaussian and assumed to be in . The received SINR for user k can then be
dependent of x
written as
SINRk = P
2
|hH
k wk |
.
H
H
2
2
j6=k |hk wj | + + |hk Qhk |
(16)
This paper considers independent quantization at each BS, in

which case Q is a diagonal matrix with diagonal entries ql .
(Multivariate compression is also possible and has been studied in [4].) Assuming an ideal vector quantizer, the quantization noise level ql and the backhaul capacity Cl (from ratedistortion theory) are related as
!
PK
2
k=1 |wl,k |
log 1 +
Cl .
(17)
ql
However, the quantizers used in practice for compression can
be far from ideal. In order to capture these losses, we introduce a notion of gap to rate-distortion limit. Following [5],
we note that operational distortion achieved by virtually all

practical quantizers at high resolution follow the relation
(R) = q var(X)2R .
(18)
where var(X) is the variance of the signal being quantized, R

is the rate of quantizer, and q is a constant that depends on
the particular choice of quantizer. For example, for a fixedrate (uncoded) uniform scalar quantizer, q = 23 , which
is around 2.72. For a uniform scalar quantizer followed by
variable-rate entropy coding we get q = e
6 which is around
1.42. Note that q = 1 corresponds to the distortion achievable by the best possible vector quantization scheme. Accounting for this, we can rewrite the relation above as
log 1 +
PK
k=1
|wl,k |2
ql
!
Cl .
(19)
Algorithm 2 WSR maximization for compression strategy

Initialization: {wk }, {ql };
Repeat:
1. For fixed {wk }, {ql }, compute the MMSE receivers {uk }
and the corresponding MSE {ek } according to (11) and
(9) with 2 replaced by 2 + |hH
k Qhk | in both equations;
2. Update the MSE weights {k } according to (10);
3. For fixed {uk } and {k }, find the optimal transmit beamformers {wk } and quantization noise levels {ql } by solving the convex optimization problem (21);
Until convergence
fixed {uk } and {k } is the following convex program:
min
wl,k ,ql
maximize
wl,k ,ql
subject to
K
X
k Rk
(20a)
k=1
K
X
k=1
K
X
|wl,k |2
2Cl 1
ql 0,
q
|wl,k |2 + ql Pl ,
(20b)
(20c)
k=1
The constraint (20b) is just a reformulation of (19) while the

constraint (20c) is the power constraint on the compressed
signal xl . Finding the globally optimal solution to the above
problem is challenging. An iterative approach based on the
majorize-minimization (MM) algorithm has been suggested
in [4]. The algorithm in [4] transforms wk wkH into a nonnegative definite matrix variable Rk and ignores the rank constraint on Rk in the optimization. In this paper, we propose
a novel way to solve (20) by reformulating it as an equivalent WMMSE problem and then using the block coordinate
descent method between the transmit beamformers {wk } and
the quantization noise levels {ql }, the receive beamformers
{uk }, and the MSE weights {k }. The algorithm can be
shown to reach a stationary point of (20). The explicit equivalence is not stated here for brevity. The numerical procedure
is presented as Algorithm 2.
In Algorithm 2, the optimization of the transmit beamformers {wk } and the quantization noise levels {ql } under
2504
2
H
wkH Ak wk Re{bH
k wk } + |uk | m |hk Qhk |
k=1
(21a)
PK
Note that k=1 |wl,k |2 is the power of the signal that is quantized for BS l. The achievable rate for user k, Rk , is again as
given by (4).
The design of the compression strategy can now be stated
as a WSR maximization problem over the transmit beamformers and the quantization noise levels:
K
X
s.t.
K
X
k=1
K
X
|wl,k |2
2Cl 1
ql 0,
q
|wl,k |2 + ql Pl ,
(21b)
(21c)
k=1
where Ak and bk are as defined in (13) and (14).

We further observe that the convex optimization problem (21) has a particular structure that can be exploited.
Observe that the two constraints (21b) and (21c) provide a
lower and an upper bound on {ql }, respectively. Since the
objective (21a) is monotonically decreasing in {ql }, we can
replace the inequality with equality in the constraint (21b)
and substitute {ql } from (21b) into the objective (21a) and
the constraint (21c). This results in a QCQP problem in only
a single set of variables {wk }, which can be efficiently solved
by standard solvers.
5. PERFORMANCE EVALUATION
We consider a 7-cell wrapped-around two-tier heterogeneous
network with simulation parameters as listed in Table 1. All
the macro-BSs and pico-BSs are connected to a centralized
processor by capacity-limited backhaul links. We compare
the performance of the two strategies under varying backhaul
capacities. The combined background noise and interference
caused by two tiers of cells outside the 7-cells is estimated to
be -150 dBm/Hz. We assume an SNR gap of m = 9 dB
(corresponding to uncoded QAM transmission) and a gap to
rate-distortion limit of q = 4.3 dB (corresponding to uncoded fixed-rate uniform scalar quantizer). At each time slot,
we solve the respective network optimization problems and
update the weights in WSR maximization according to the
proportional fair criterion.
10 MHz
0.8 km
30
1
3
43 dBm
30 dBm
15 dBi
169 dBm/Hz
128.1 + 37.6 log 10 (d)
140.7 + 36.7 log 10 (d)
8 dB
0 dB
9 dB
4.3 dB
1
0.9
Cumulative Distribution Function
Channel bandwidth
Distance between cells
Number of users/cell
Number of macro-BSs/cell
Number of pico-BSs/cell
Max. Tx power at macro-BS
Max. Tx Power at pico-BS
Antenna gain
Background noise
Path loss from macro-BS to user
Path loss from pico-BS to user
Log-normal shadowing
Rayleigh small scale fading
SNR gap (m )
Rate-distortion gap (q )
0.8
0.7
0.6
0.5
Strongest 1 BS (98, 52) Mbps
Data Sharing (40, 20) Mbps
Compression (40, 20) Mbps
Full Cooperation
0.4
0.3
0.2
0.1
0
LongTerm Average User Rates (Mbps)
Table 1. Simulation Parameters

Fig. 1. Comparison of cumulative distribution of user rates for
Fig. 1 shows the cumulative distribution of user rates under varying backhaul capacities for both strategies. For reference, we also include the full cooperation case with infinite
backhaul capacity and the baseline scheme of no cooperation
with each user connected to the strongest BS. When the backhaul capacity is low at 40 Mbps/macro-BS and 20 Mbps/picoBS, the data-sharing strategy outperforms the compression
strategy. The 50-percentile rate for the data-sharing strategy is about 3 times that of the compression strategy. If we
double the backhaul capacity to 80 Mbps/macro-BS and 40
Mbps/pico-BS, the compression strategy becomes comparable to the data-sharing strategy and both have about the same
50-percentile user rates. At this operating point, the sum
backhaul capacity is about 6 times that of the average sum
rate per cell. We also observe that the compression strategy
favours low rate users while the data-sharing strategy favours
high rate users. A reason for this is that the compression strategy under low backhaul capacity is limited by the quantization noises which are about the same for all the BS signals
resulting in more uniform user rates.
We observe that with moderate-to-high backhaul capacity
of 160 Mbps/macro-BS and 80 Mbps/pico-BS, the compression strategy outperforms the data-sharing strategy with the
50-percentile rate for the compression strategy more than 2.5
times than that of data-sharing. Increasing the backhaul in this
regime improves the compression strategy drastically, while
the data-sharing strategy sees only a moderate increase. This
is because, at low backhaul capacity, the performance of the
compression strategy is limited by the quantization noises.
An increase in backhaul capacity reduces the quantization
noise levels exponentially, while a similar increase in the
backhaul capacity does not buy as much for the data-sharing
strategy. Finally with a backhaul of 240 Mbps/macro-BS
and 120 Mbps/pico-BS, the compression strategy performs
close to the full cooperation limit, while for the data-sharing
strategy, backhaul capacities of 1200 Mbps/macro-BS and
600 Mbps/pico-BS are needed to get as close.
2505
the data-sharing and compression strategies.

6. CONCLUSIONS
This paper compares two fundamentally different strategies,
the data-sharing and the compression strategy, for the downlink C-RAN under realistic network settings considering various practical aspects. Our main conclusion is that the backhaul capacity constraint is crucial in deciding which strategy
to adopt. The compression strategy offers better user rates for
moderate-to-high backhaul capacity, due to its ability to have
full cooperation before quantization. But it suffers from high
quantization loss at low backhaul capacity in which case it is
better to do data-sharing with limited cooperation cluster.
REFERENCES
[1] P. Patil and W. Yu, Hybrid compression and messagesharing strategy for the downlink cloud radio-access network, in Proc. Inf. Theory Applicat. Workshop, Feb.
2014.
[2] O. Simeone, O. Somekh, H. Poor, and S. Shamai,
Downlink multicell processing with limited-backhaul
capacity, EURASIP J. Advances Singal Process., Feb.
2009.
[3] B. Dai and W. Yu, Sparse beamforming and usercentric clustering for downlink cloud radio access network, IEEE Access, vol. 2, pp. 13261339, 2014.
[4] S.-H. Park, O. Simeone, O. Sahin, and S. Shamai, Joint
precoding and multivariate backhaul compression for the
downlink of cloud radio access networks, IEEE Trans.
Signal Process., vol. 61, no. 22, pp. 56465658, Nov.
2013.
[5] R. M. Gray and D. L. Neuhoff, Quantization, IEEE
Trans. Inf. Theory, vol. 44, no. 6, pp. 23252383, Oct.
1998.

d2 1 2 Proceedings of Second Special Session

Încărcat de

Informații document

Descriere originală:

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

d2 1 2 Proceedings of Second Special Session

Încărcat de

Drepturi de autor:

Formate disponibile

HARP deliverable: D2.1.

2 Proceedings of second workshop/ special session at an international

D2.1.2 Proceedings of second workshop/ special session at international conference

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

Second Workshop in an International Conference .............................................................. 3

Special Session in an International Conference ................................................................... 6

3.1 General Information about the Special Session .................................................................... 7

Special Stand-Alone Workshop ..................................................................................... 10

4.1 General Information about the GdR-ISIS workshop........................................................... 11

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

Second Workshop in an International Conference

2.1 General Information about the Special Session

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

2.2 Workshop Structure and Contributions

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

1. M. Cui, B. Hu, X. Li (South China University of Technology), H. Chen (Guilin

Keynote: Dr. M. Nekovee (Samsung Electronics UK),

1. S. Hfner, D. A. Dupleich, R. Mller, C. Schneider, R. S. Thom (Technische Universitt

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

Keynote: Dr. J. Putkonen (Nokia),

1. S. Le Hong Nguyen, K. Haneda, J. Jarvelainen, A. Karttunen (Aalto University, Finland),

Special Session in an International Conference

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

3.1 General Information about the Special Session

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

(HARP member, EURECOM, France),

Number of Attendees: About 25 in average.

3.2 Special Session Structure and Contributions

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

Special Stand-Alone Workshop

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

4.1 General Information about the GdR-ISIS workshop

(HARP member, EURECOM, France),

4.2 Workshop Structure and Contributions

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

This appendix contains the following documents:

HARP deliverable: D2.1.2 Proceedings of second workshop/ special session at an international

3) The paper Multidimensional continuous phase modulation in uplink of MIMO systems by

A Novel Power Control Algorithm for Massive MIMO

School of Electronic and Information Engineering

Key Laboratory of Cognitive Radio and Information

AbstractIn this paper, a novel system model named as

In recent years, the explosive growth of wireless

complexity, etc. Authors in [3] imported the power threshold

978-1-4799-8088-8/15/$31.00 2015 IEEE

an efficient uplink power control algorithm based on

In this section, we apply massive MIMO into the CRS to

noises (AWGNs) at the CBS with its each element

Let rk be the kth element of r . Then, we have

where gk is the kth column of G. From (3) we get the SINR of

Recalling that Ik = gH gi pi + gH gp pp + gH k2 and rearranging

the terms, we rewrite (4) as

Fig. 1. System model.

We consider the uplink transmission of the massive MIMO

where G is the MK channel matrix between the CBS and K

The expectation of k can be given by

In this section, we formulate the power control problem as a

A. Utility Function and Derivation of the Nash Equilibrium

In practical, a good CRS should satisfy the following two

where pk,th  min{pk,max, pk,temp} is a self-adaptive threshold of

From (9), we see pk,temp can be self-adaptively adjusted based

where pk,th min{pk,max, pk,temp} is a self-adaptive threshold of