11a1 PDF
11a1 PDF
11a1 PDF
Designs
Linköping 2015
Massive MIMO: Fundamentals and System Designs
⃝
c 2015 Hien Quoc Ngo, unless otherwise noted.
ISBN 978-91-7519-147-8
ISSN 0345-7524
Printed in Sweden by LiU-Tryck, Linköping 2015
Cảm ơn gia ñình tôi, cảm ơn Em,
In the rst part, we focus on fundamental limits of the system performance under
practical constraints such as low complexity processing, limited length of each coher-
ence interval, intercell interference, and nite-dimensional channels. We rst study
the potential for power savings of the Massive MIMO uplink with maximum-ratio
combining (MRC), zero-forcing, and minimum mean-square error receivers, under
perfect and imperfect channels. The energy and spectral eciency tradeo is inves-
tigated. Secondly, we consider a physical channel model where the angular domain
is divided into a nite number of distinct directions. A lower bound on the capacity
is derived, and the eect of pilot contamination in this nite-dimensional channel
model is analyzed. Finally, some aspects of favorable propagation in Massive MIMO
under Rayleigh fading and line-of-sight (LoS) channels are investigated. We show
that both Rayleigh fading and LoS environments oer favorable propagation.
v
In the second part, based on the fundamental analysis in the rst part, we pro-
pose some system designs for Massive MIMO. The acquisition of channel state
information (CSI) is very important in Massive MIMO. Typically, the channels
are estimated at the BS through uplink training. Owing to the limited length of
the coherence interval, the system performance is limited by pilot contamination.
To reduce the pilot contamination eect, we propose an eigenvalue-decomposition-
based scheme to estimate the channel directly from the received data. The pro-
posed scheme results in better performance compared with the conventional train-
ing schemes due to the reduced pilot contamination. Another important issue of
CSI acquisition in Massive MIMO is how to acquire CSI at the users. To address
this issue, we propose two channel estimation schemes at the users: i) a downlink
beamforming training scheme, and ii) a method for blind estimation of the ef-
fective downlink channel gains. In both schemes, the channel estimation overhead
is independent of the number of BS antennas. We also derive the optimal pilot
and data powers as well as the training duration allocation to maximize the sum
spectral eciency of the Massive MIMO uplink with MRC receivers, for a given
total energy budget spent in a coherence interval. Finally, applications of Massive
MIMO in relay channels are proposed and analyzed. Specically, we consider multi-
pair relaying systems where many sources simultaneously communicate with many
destinations in the same time-frequency resource with the help of a Massive MIMO
relay. A Massive MIMO relay is equipped with many collocated or distributed an-
tennas. We consider dierent duplexing modes (full-duplex and half-duplex) and
dierent relaying protocols (amplify-and-forward, decode-and-forward, two-way re-
laying, and one-way relaying) at the relay. The potential benets of massive MIMO
technology in these relaying systems are explored in terms of spectral eciency and
power eciency.
Populärvetenskaplig
Sammanfattning
Det har skett en massiv tillväxt av antalet trådlöst kommunicerande enheter de
senaste tio åren. Idag är miljarder av enheter anslutna och styrda över trådlösa
nätverk. Samtidigt kräver varje enhet en hög datatakt för att stödja sina app-
likationer, som röstkommunikation, realtidsvideo, lm och spel. Efterfrågan på
trådlös datatakt och antalet trådlösa enheter kommer alltid att tillta. Samtidigt
kan inte strömförbrukningen hos de trådlösa kommunikationssystemen tillåtas att
öka. Således måste framtida trådlösa kommunikationssystem uppfylla tre huvud-
krav: i) hög datatakt ii) kunna betjäna många användare samtidigt iii) lägre ström-
förbrukning.
vii
metoder för kanalskattning både för basstationen och för användarna, vilka ämnar
minimera eekten av pilotkontaminering och kanalovisshet. Den optimala pilot-
och dataeekten så väl som valet av längden av träningsperioden studeras. Till slut
föreslås och analyseras användandet av massiv MIMO i reläkanaler.
Acknowledgments
I would like to extend my sincere thanks to my supervisor, Prof. Erik G. Larsson,
for his valuable support and supervision. His advice, guidance, encouragement, and
inspiration have been invaluable over the years. Prof. Larsson always keeps an open
mind in every academic discussion. I admire his critical eye for important research
topics. I still remember when I began my doctoral studies, Prof. Larsson showed
me the rst paper on Massive MIMO and stimulated my interest for this topic. This
thesis would not have been completed without his guidance and support.
ix
Finally, I would like to thank my family and friends, for their constant love, encour-
agement, and limitless support throughout my life.
xi
PDF Probability Density Function
OFDM Orthogonal Frequency Division Multiplexing
QAM Quadrature Amplitude Modulation
RV Random Variable
SEP Symbol Error Probability
SIC Successive Interference Cancellation
SINR Signal-to-Interference-plus-Noise Ratio
SIR Signal-to-Interference Ratio
SISO Single-Input Single-Output
SNR Signal-to-Noise Ratio
TDD Time Division Duplexing
TWRC Two-Way Relay Channel
UL Uplink
ZF Zero-Forcing
Contents
Abstract v
Populärvetenskaplig Sammanfattning (in Swedish) vii
Acknowledgments ix
Abbreviations xi
I Introduction 1
1 Motivation 3
2 Mutiuser MIMO Cellular Systems 7
2.1 System Models and Assumptions . . . . . . . . . . . . . . . . . . . . 7
2.2 Uplink Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Downlink Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Linear Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.1 Linear Receivers (in the Uplink) . . . . . . . . . . . . . . . . 10
2.4.2 Linear Precoders (in the Downlink) . . . . . . . . . . . . . . . 13
2.5 Channel Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Channel Estimation in TDD Systems . . . . . . . . . . . . . 14
2.5.2 Channel Estimation in FDD Systems . . . . . . . . . . . . . . 16
3 Massive MIMO 19
3.1 What is Massive MIMO? . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 How Massive MIMO Works . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.1 Channel Estimation . . . . . . . . . . . . . . . . . . . . . . . 21
3.2.2 Uplink Data Transmission . . . . . . . . . . . . . . . . . . . . 21
3.2.3 Downlink Data Transmission . . . . . . . . . . . . . . . . . . 22
3.3 Why Massive MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Challenges in Massive MIMO . . . . . . . . . . . . . . . . . . . . . . 23
3.4.1 Pilot Contamination . . . . . . . . . . . . . . . . . . . . . . . 23
3.4.2 Unfavorable Propagation . . . . . . . . . . . . . . . . . . . . 24
3.4.3 New Standards and Designs are Required . . . . . . . . . . . 24
4 Mathematical Preliminaries 25
4.1 Random Matrix Theory . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Capacity Lower Bounds . . . . . . . . . . . . . . . . . . . . . . . . . 26
Introduction
1
Chapter 1
Motivation
During the last years, data trac (both mobile and xed) has grown exponentially
due to the dramatic growth of smartphones, tablets, laptops, and many other wire-
less data consuming devices. The demand for wireless data trac will be even
more in future [13]. Figures 1.1 shows the demand for mobile data trac and the
number of connected devices. Global mobile data trac is expected to increase to
15.9 exabytes per month by 2018, which is about an 6-fold increase over 2014. In
addition, the number of mobile devices and connections are expected to grow to
10.2 billion by 2018. New technologies are required to meet this demand. Related
to wireless data trac, the key parameter to consider is wireless throughput (bits/s)
which is dened as:
Clearly, to improve the throughput, some new technologies which can increase the
bandwidth or the spectral eciency or both should be exploited. In this thesis,
we focus on techniques which improve the spectral eciency. A well-known way to
increase the spectral eciency is using multiple antennas at the transceivers.
3
4 Chapter 1. Introduction
12.0
8.0 7.0 EB
4.0
4.4 EB
4.0 2.6 EB
1.5 EB
0.0 0.0
2012 2013 2014 2015 2016 2017 2018 2012 2013 2014 2015 2016 2017 2018
Year Year
(a) Global mobile data trac. (b) Global mobile devices and connections
growth.
Figure 1.1: Demand for mobile data trac and number of connected devices.
(Source: Cisco [3])
The eort to exploit the spatial multiplexing gain has been shifted from MIMO to
multiuser MIMO (MU-MIMO), where several users are simultaneously served by a
multiple-antenna base station (BS). With MU-MIMO setups, a spatial multiplexing
gain can be achieved even if each user has a single antenna [4]. This is important
since users cannot support many antennas due to the small physical size and low-
cost requirements of the terminals, whereas the BS can support many antennas.
MU-MIMO does not only reap all benets of MIMO systems, but also overcomes
most of propagation limitations in MIMO such as ill-behaved channels. Specically,
by using scheduling schemes, we can reduce the limitations of ill-behaved channels.
Line-of-sight propagation, which causes signicant reduction of the performance of
MIMO systems, is no longer a problem in MU-MIMO systems. Thus, MU-MIMO
has attracted substantial interest [49].
There always exists a tradeo between the system performance and the implemen-
tation complexity. The advantages of MU-MIMO come at a price:
• User scheduling: since several users are served on the same time-frequency
resource, scheduling schemes which optimally select the group of users de-
pending on the precoding/detection schemes, CSI knowledge etc., should be
considered. This increases the cost of the system implementation.
The more antennas the BS is equipped with, the more degrees of freedom are oered
and hence, more users can simultaneously communicate in the same time-frequency
resource. As a result, a huge sum throughput can be obtained. With large antenna
arrays, conventional signal processing techniques (e.g. maximum likelihood detec-
tion) become prohibitively complex due to the high signal dimensions. The main
question is whether we can obtain the huge multiplexing gain with low-complexity
signal processing and low-cost hardware implementation.
In [13], Marzetta showed that the use of an excessive number of BS antennas com-
pared with the number of active users makes simple linear processing nearly optimal.
More precisely, even with simple maximum-ratio combining (MRC) in the uplink
or maximum-ratio transmission (MRT) in the downlink, the eects of fast fading,
intracell interference, and uncorrelated noise tend to disappear as the number of
BS station antennas grows large. MU-MIMO systems, where a BS with a hundred
or more antennas simultaneously serves tens (or more) of users in the same time-
frequency resource, are known as Massive MIMO systems (also called very large
MU-MIMO, hyper-MIMO, or full-dimension MIMO systems). In Massive MIMO,
it is expected that each antenna would be contained in an inexpensive module with
simple processing and a low-power amplier. The main benets of Massive MIMO
systems are:
(1) Huge spectral eciency and high communication reliability : Massive MIMO
inherits all gains from conventional MU-MIMO, i.e., with M -antenna BS and
K single-antenna users, we can achieve a diversity of order M and a multi-
plexing gain of min (M, K). By increasing both M and K , we can obtain a
huge spectral eciency and very high communication reliability.
(2) High energy eciency : In the uplink Massive MIMO, coherent combining can
achieve a very high array gain which allows for substantial reduction in the
transmit power of each user. In the downlink, the BS can focus the energy
into the spatial directions where the terminals are located. As a result, with
massive antenna arrays, the radiated power can be reduced by an order of
magnitude, or more, and hence, we can obtain high energy eciency. For a
xed number of users, by doubling the number of BS antennas, while reducing
the transmit power by two, we can maintain the original the spectral eciency,
and hence, the radiated energy eciency is doubled.
(3) Simple signal processing : For most propagation environments, the use of an
excessive number of BS antennas over the number of users yields favorable
propagation where the channel vectors between the users and the BS are
6 Chapter 1. Introduction
In the following, brief introductions to multiuser MIMO and Massive MIMO are
given in Chapter 2 and Chapter 3, respectively. In Chapter 4, we provide some
mathematical preliminaries which will be used throughout the thesis. In Chapter 5,
we list the specic contributions of the thesis together with a short description of
the included papers. Finally, future research directions are discussed in Chapter 6.
Chapter 2
Let H ∈ CM ×K be the channel matrix between the K users and the BS antenna
array, where the k th column of H, denoted by hk , represents the M ×1 channel
vector between the k th user and the BS. In general, the propagation channel is
modeled via large-scale fading and small-scale fading. But in this chapter, we ignore
large-scale fading, and further assume that the elements of H are i.i.d. Gaussian
distributed with zero mean and unit variance.
7
8 Chapter 2. Mutiuser MIMO Cellular Systems
Figure 2.1: Multiuser MIMO Systems. Here, K single-antenna users are served by
the M -antenna BS in the same time-frequency resource.
√ ∑
K
y ul = pu h k sk + n (2.1)
k=1
√
= puH s + n , (2.2)
M ×1
where is the average signal-to-noise ratio (SNR), n ∈ C
pu is the additive noise
T
vector, and s , [s1 ... sK ] . We assume that the elements of n are i.i.d. Gaussian
random variables (RVs) with zero mean and unit variance, and independent of H.
From the received signal vector y ul together with knowledge of the CSI, the BS will
coherently detect the signals transmitted from the K users. The channel model
(2.2) is the multiple-access channel which has the sum-capacity [27]
( )
Cul,sum = log2 det I K + puH H H . (2.3)
√
ydl,k = pdh Tk x + zk , (2.4)
where pd is the average SNR and zk is the additive noise at the k th user. We assume
that zk is Gaussian distributed with zero mean and unit variance. Collectively, the
received signal vector of the K users can be written as
√
y dl = pdH T x + z , (2.5)
( )
Csum = max log2 det I M + pdH ∗D q H T , (2.6)
{q }
∑k
qk ≥0, K
k=1 qk ≤1
√
ŝs = arg min ∥yy ul − puH s ∥2 (2.7)
s ∈S K
The BS can use linear processing schemes (linear receivers in the uplink and lin-
ear precoders in the downlink) to reduce the signal processing complexity. These
schemes are not optimal. However, when the number of BS antennas is large, it is
shown in [13, 14] that linear processing is nearly-optimal. Therefore, in this thesis,
we will consider linear processing. The details of linear processing techniques are
presented in the following sections.
10 Chapter 2. Mutiuser MIMO Cellular Systems
yul,1
ŝ2 yɶul,2
ŝK yɶul,K
yul,M
Each stream is then decoded independently. See Figure 2.2. The complexity is on
the order of K|S|. From (2.8), the k th stream (element) of ỹy ul , which is used to
decode sk , is given by
√ ∑ H
K
√
ỹul,k = pua H h k sk + pu a k h k′ sk′ + a H
k n, (2.9)
| {z
k
} |{z}
k′ ̸=k
desired signal | {z } noise
interuser interference
pu |a
a H h k |2
SINRk = ∑K . (2.10)
pu k′ ̸=k |a
aHk h k | + ∥a
′ 2 ak ∥2
With MRC, the BS aims to maximize the received signal-to-noise ratio (SNR)
of each stream, ignoring the eect of multiuser interference. From (2.9), the
2.4. Linear Processing 11
power(desired signal)
a mrc,k = argmax
a k ∈CM ×1 power(noise)
pu |a
aHk hk |
2
= argmax . (2.11)
a k ∈CM ×1 ∥aak ∥2
Since
pu |a
aHk hk |
2
pu ∥a
ak ∥2 ∥h
hk ∥2
≤ = pu ∥h
hk ∥2 ,
∥aak ∥2 ∥a
ak ∥2
and equality holds when ak = const · hk , the MRC receiver is: a mrc,k =
const · hk . Plugging a mrc,k into (2.10), the received SINR of the k th stream
for MRC is given by
pu ∥h
hk ∥4
SINRmrc,k = ∑K (2.12)
pu k′ ̸=k hH
|h k h k′ | + ∥h
2 hk ∥2
∥h
hk ∥4
→ ∑K , as pu → ∞. (2.13)
k′ ̸=k hH
|h k h k′ |
2
Advantage: the signal processing is very simple since the BS just mul-
tiplies the received vector with the conjugate-transpose of the channel
matrix H, and then detects each stream separately. More importantly,
MRC can be implemented in a distributed manner. Furthermore, at
low pu , SINRmrc,k ≈ pu ∥h
hk ∥2 . This implies that at low SNR, MRC can
achieve the same array gain as in the case of a single-user system.
b) Zero-Forcing Receiver:
The ZF receiver matrix, which satises (2.14) for all k, is the pseudo-inverse
of the channel matrix H. With ZF, we have
( )−1 √ ( )−1
ỹy ul = H H H H H y ul = pus + H H H H H n. (2.15)
12 Chapter 2. Mutiuser MIMO Cellular Systems
( )−1
where ñk denotes the k th element of H HH H H n. Thus, the received
pu
SINRzf,k = [( )−1 ] . (2.17)
H HH
kk
∑
K
{ H }
= arg min E |a
ak y ul − sk |2 . (2.19)
A ∈CM ×K
k=1
√ ( )−1
= pu puH H H + I M hk , (2.22)
{ }
where cov (vv 1 , v 2 ) , E v 1v H
2 , where v1 and v2 are two random column
vectors with zero-mean elements.
It is known that the MMSE receiver maximizes the received SINR. Therefore,
among the MMSE, ZF, and MRC receivers, MMSE is the best. We can see
2.4. Linear Processing 13
from (2.22) that, at high SNR (high pu ), ZF approaches MMSE, while at low
SNR, MRC performs as well as MMSE. Furthermore, substituting (2.22) into
(2.10), the received SINR for the MMSE receiver is given by
−1
∑
K
SINRmmse,k = puh H
k
pu h ih H
i + IM
hk . (2.23)
i̸=k
1
α= { }. (2.25)
E tr(W
W W H)
√ √ ∑
K
= αpdh Tk w k qk + αpd h Tk w k′ qk′ + zk . (2.27)
k′ ̸=k
x1
antenna 1
q1
Precoding x2
q2 Matrix antenna 2
(KxM)
qK
xM
antenna M
Figures 2.4 and 2.5 show the achievable sum rates for the uplink and the downlink
transmission, respectively, with dierent linear processing schemes, versus SNR , pu
for the uplink and SNR
∑K , p d for the downlink, with M = 6 and K = 4. The sum
rate is dened as k=1 E {log2 (1 + SINRk )}, where SINRk is the SINR of the k th
user which is given in the previous discussion. As expected, MMSE performs strictly
better than ZF and MRC over the entire range of SNRs. In the low SNR regime,
MRC is better than ZF, and vice versa in the high SNR regime.
• For the uplink transmission: the BS needs CSI to detect the signals transmit-
ted from the K users. This CSI is estimated at the BS. More precisely, the K
users send K orthogonal pilot sequences to the BS on the uplink. Then the
BS estimates the channels based on the received pilot signals. This process
requires a minimum of K channel uses.
1 In practice, the uplink and downlink channels are not perfectly reciprocal due to mismatches
of the hardware chains. This non-reciprocity can be removed by calibration [15, 29, 30]. In our
work, we assume that the hardware chain calibration is perfect.
2.5. Channel Estimation 15
35.0
M=6, K = 4
30.0
Sum Rate (bits/s/Hz)
25.0
20.0 MMSE
15.0
ZF
10.0
MRC
5.0
0.0
-20 -15 -10 -5 0 5 10 15 20
SNR (dB)
40.0
M=6, K = 4
35.0
30.0
Sum Rate (bits/s/Hz)
ZF
25.0
20.0 MMSE
15.0
10.0
MRT
5.0
0.0
-10 -5 0 5 10 15 20 25 30
SNR (dB)
≥ K (symbols) ≥ K (symbols)
T (symbols)
• For the downlink: the BS needs CSI to precode the transmitted signals, while
each user needs the eective channel gain to detect the desired signals. Due to
the channel reciprocity, the channel estimated at the BS in the uplink can be
used to precode the transmit symbols. To obtain knowledge of the eective
channel gain, the BS can beamform pilots, and each user can estimate the
eective channel gains based on the received pilot signals. This requires at
least K channel uses.
2
• For the downlink transmission: the BS needs CSI to precode the symbols
before transmitting to the K users. The M BS antennas transmit M orthog-
onal pilot sequences to K users. Each user will estimate the channel based
on the received pilots. Then it feeds back its channel estimates (M channel
estimates) to the BS through the uplink. This process requires at least M
channel uses for the downlink and M channel uses for the uplink.
• For the uplink transmission: the BS needs CSI to decode the signals trans-
mitted from the K users. One simple way is that the K users transmit K
orthogonal pilot sequences to the BS. Then, the BS will estimate the channels
based on the received pilot signals. This process requires at least K channel
uses for the uplink.
2 The eective channel gains at the users may be blindly estimated based on the received data,
and hence, no pilots are required [31]. But, we do not discuss in detail about this possibility in
this section.
2.5. Channel Estimation 17
T (symbols)
≥ M (symbols)
≥ K (symbols) ≥ M (symbols)
Massive MIMO
• TDD operation: as discussed in Section 2.5, with FDD, the channel estimation
overhead depends on the number of BS antennas, M . By contrast, with TDD,
the channel estimation overhead is independent of M . In Massive MIMO, M
is large, and hence, TDD operation is preferable. For example, assume that
the coherence interval is T = 200 symbols (corresponding to a coherence
bandwidth of 200 kHz and a coherence time of 1 ms). Then, in FDD systems,
the number of BS antennas and the number of users are constrained by M +
K < 200, while in TDD systems, the constraint on M and K is 2K < 200.
Figure 3.1 shows the regions of feasible (M, K) in FDD and TDD systems.
We can see that the FDD region is much smaller than the TDD region. With
TDD, adding more antennas does not aect the resources needed for the
channel estimation.
• Linear processing: since the number of BS antennas and the number of users
are large, the signal processing at the terminal ends must deal with large
dimensional matrices/vectors. Thus, simple signal processing is preferable.
In Massive MIMO, linear processing (linear combing schemes in the uplink
and linear precoding schemes in the downlink) is nearly optimal.
19
20 Chapter 3. Massive MIMO
1000
Number of BS Antennas (M)
800
600
TDD
400
200
FDD
0
0 20 40 60 80 100 120 140 160 180 200
Number of Users (K)
Figure 3.1: The regions of possible (M, K) in TDD and FDD systems, for a coher-
ence interval of 200 symbols.
• A massive BS antenna array does not have to be physically large. For example
consider a cylindrical array with 128 antennas, comprising four circles of 16
dual-polarized antenna elements. At 2.6 GHz, the distance between adjacent
antennas is about 6 cm, which is half a wavelength, and hence, this array
occupies only a physical size of 28cm×29cm [25].
• Massive MIMO is scalable: in Massive MIMO, the BS learns the channels via
uplink training, under TDD operation. The time required for channel estima-
tion is independent of the number of BS antennas. Therefore, the number of
BS antennas can be made as large as desired with no increase in the channel
estimation overhead. Furthermore, the signal processing at each user is very
simple and does not depend on other users' existence, i.e., no multiplexing
or de-multiplexing signal processing is performed at the users. Adding or
dropping some users from service does not aect other users' activities.
Furthermore, each user may need partial knowledge of CSI to coherently detect
the signals transmitted from the BS. This information can be acquired through
downlink training or some blind channel estimation algorithm. Since the BS uses
linear precoding techniques to beamform the signals to the users, the user needs only
the eective channel gain (which is a scalar constant) to detect its desired signals.
Therefore, the BS can spend a short time to beamform pilots in the downlink for
CSI acquisition at the users.
techniques to detect signals transmitted from all users. The detailed uplink data
transmission was discussed in Section 2.2.
In (3.1), K is the multiplexing gain, and M represents the array gain. We can see
that, we can obtain a huge spectral eciency and energy eciency when M and K
are large. Without any increase in transmitted power per terminal, by increasing
K and M, we can simultaneously serve more users in the same frequency band. At
the same time the throughput per user also increases. Furthermore, by doubling
the number of BS antennas, we can reduce the transmit power by 3 dB, while
maintaining the original quality-of-service.
The above gains (multiplexing gain and array gain) are obtained under the condi-
tions of favorable propagation and the use of optimal processing at the BS. One
main question is: Will these gains still be obtained by using linear processing?
Another question is: Why not use the conventional low dimensional point-to-point
MIMO with complicated processing schemes instead of Massive MIMO with simple
linear processing schemes? In Massive MIMO, when the number of BS antennas
is large, due to the law of large numbers, the channels become favorable. As a
result, linear processing is nearly optimal. The multiplexing gain and array gain
can be obtained with simple linear processing. Also, by increasing the number of
BS antennas and the number of users, we can always increase the throughput.
3.4. Challenges in Massive MIMO 23
35.0
30.0
Shannon sum capacity
25.0
Sum Rate (bits/s/Hz)
MRC
20.0
15.0 ZF
10.0
MMSE
5.0
K = 10, pu= -10 dB
0.0
20 40 60 80 100
Number of BS Antennas (M)
Figure 3.3: Uplink sum rate for dierent linear receivers and for the optimal receiver.
Figure 3.3 shows the sum rate versus the number of BS antennas with optimal
receivers (the sum capacity is achieved) and linear receivers, at K = 10 and pu =
−10 dB. The sum capacity is computed from (2.3), while the sum rates for MRC,
ZF, and MMSE are computed by using (2.12), (2.17), and (2.23), respectively. We
can see that, when M is large, the sum rate with linear processing is very close
to the sum capacity obtained by using optimal receivers. When M = K = 10,
and with the optimal receiver, the maximum sum rate that we can obtain is 8.5
bits/s/Hz. By contrast, by using large M, say M = 50, with simple ZF receivers,
we can obtain a sum rate of 24 bits/s/HZ.
Mathematical Preliminaries
T T
• Let p , [p1 ... pn ] and q , [q1 ... qn ] be n×1 vectors whose elements
are independent identically distributed (i.i.d.) random variables (RVs) with
{ } { }
2 2
E {pi } = E {qi } = 0, E |pi | = σp2 , and E |qi | = σq2 , i = 1, 2, ..., n.
Assume that p and q are independent.
1 H a.s. 2
p p → σp , as n → ∞, (4.1)
n
1 H a.s.
p q → 0, as n → ∞, (4.2)
n
a.s.
where → denotes almost sure convergence.
1 d ( )
√ p H q → CN 0, σp2 σq2 , as n → ∞, (4.3)
n
d
where → denotes convergence in distribution.
25
26 Chapter 4. Mathematical Preliminaries
• Let X1 , X2 ,
... be a sequence of independent circularly symmetric complex
2
RVs, such that Xi has zero mean and variance
∑n σi 2. Further assume that the
i=1 σi → ∞, as n → ∞; and 2)
2
following conditions are satised: 1) sn =
σi /sn → 0, as n → ∞. Then by applying the Cramér's central limit theorem,
we have
∑n
Xi d
i=1 → CN (0, 1) , as n → ∞. (4.4)
sn
• Let X1 , X2 , ...Xn be independent RVs such that E {Xi } = µi and Var (Xi ) <
c < ∞, ∀i = 1, . . . , n. Then by applying the Tchebyshev's theorem, we have
1 1 P
(X1 + X2 + ... + Xn ) − (µ1 + µ2 + ...µn ) → 0, (4.5)
n n
P
where → denotes convergence in probability.
∑
K
yk = ak sk + ak′ sk′ + nk , (4.6)
|{z} |{z}
k′ ̸=k
desired signal | {z } noise
interference
where ak , k = 1, ..., K , are the eective channel gains, sk , is the transmitted signals
from the k th sources, and nk is the additive noise. Assume that sk , k = 1, ..., K ,
and nk are independent RVs with zero-mean and unit variance.
Since in this thesis, we consider linear processing, the end-to-end channel can be
considered as an interference SISO channel, and can be modeled as in (4.6). For
example, consider the downlink transmission discussed in Section 2.4. The received
signal at the k th user is given by (2.27) which precisely matches with the model
√
(4.6), where ak′ is αpdh Tk w k′ , k ′ = 1, ..., K . The capacity lower bound derived in
this section will be used throughout the thesis.
Let C be the channel state information (CSI) available at the receiver. We assume
that sk ∼ CN (0, 1). In general, Gaussian signaling is not optimal, and hence, this
assumption yields a lower bound on the capacity:
(a)
= log2 (πe) − h ( sk − αyk | yk , C) (4.8)
(b)
≥ log2 (πe) − h ( sk − αyk | C) , (4.9)
4.2. Capacity Lower Bounds 27
where (a) holds for any α, and (b) follows from the fact that conditioning reduces
the entropy. Note that, in (4.7), I(x; y) and h(x) denote the mutual information
between x and y, and the dierential entropy of x, respectively.
Since the dierential entropy of a RV with xed variance is maximized when the
RV is Gaussian, we obtain
{ ( { })}
2
Ck ≥ log2 (πe) − E log2 πe E |sk − αyk | C , (4.10)
which leads to
1
Ck ≥ E log2 { } . (4.11)
2
E |sk − αyk | C
{ }
2
To obtain the tightest bound, we choose α = α0 so that E |xk − α0 yk | C is
minimized:
{ }
2
α0 = arg min E |sk − αyk | C . (4.12)
α
E { yk∗ sk | C} E { a∗k | C}
α0 = { } = { }. (4.13)
2 2
E |yk | C E |yk | C
We have
{ } { } { }
2 2
E |sk − αyk | C = E |sk | −α∗ E { sk yk∗ | C}−α E { s∗k yk | C} + |α| E |yk | C
2 2
{ }
2
= 1 − α∗ E { a∗k | C} − α E { ak | C} + |α| E |yk | C .
2
(4.14)
{ } |E { a∗k | C}|
2
2
E |xk − αyk | C = 1 − ∑ { } . (4.15)
K 2
k=1 E |ak | C + 1
Plugging (4.15) into (11), we obtain the following lower bound on the capacity of
(4.6):
|E { ak | C}|
2
Ck ≥ E log2 1 + ∑ { } . (4.16)
K 2 2
k =1 E |ak | C − |E { ak | C}| + 1
′ ′
2
|E {a }|
Ck ≥ log2 1 + { } ∑ { } .
k
(4.17)
2 K 2
E |ak − E {ak }| + k′ ̸=k E |ak′ | + 1
This bound is often used in Massive MIMO research since it has a simple
closed-form solution. Furthermore, in most propagation environments, when
the number of BS antennas is large, the channel hardens (the eective channel
gains become deterministic), and hence, this bound is very tight.
{ ( )}
2
|ak |
Ck ≥ E log2 1 + ∑K 2
. (4.18)
k′ ̸=k |ak′ | + 1
Chapter 5
Summary of Specic
Contributions of the
Dissertation
29
30 Chapter 5. Contributions of the Dissertation
Paper B: The Multicell Multiuser MIMO Uplink with Very Large An-
tenna Arrays and a Finite-Dimensional Channel
We consider multicell multiuser MIMO systems with a very large number of an-
tennas at the base station (BS). We assume that the channel is estimated by using
uplink training. We further consider a physical channel model where the angular
domain is separated into a nite number of distinct directions. We analyze the
so-called pilot contamination eect discovered in previous work, and show that this
eect persists under the nite-dimensional channel model that we consider. In par-
ticular, we consider a uniform array at the BS. For this scenario, we show that when
the number of BS antennas goes to innity, the system performance under a nite-
dimensional channel model with P angular bins is the same as the performance
under an uncorrelated channel model with P antennas. We further derive a lower
bound on the achievable rate of uplink data transmission with a linear detector at
the BS. We then specialize this lower bound to the cases of maximum-ratio com-
bining (MRC) and zero-forcing (ZF) receivers, for a nite and an innite number
of BS antennas. Numerical results corroborate our analysis and show a comparison
between the performances of MRC and ZF in terms of sum-rate.
This paper consider a multicell multiuser MIMO with very large antenna arrays at
the base station. For this system, with channel state information estimated from
pilots, the system performance is limited by pilot contamination and noise limita-
tion as well as the spectral ineciency discovered in previous work. To reduce these
eects, we propose the eigenvalue-decomposition-based approach to estimate the
channel directly from the received data. This approach is based on the orthogonal-
ity of the channel vectors between the users and the base station when the number
of base station antennas grows large. We show that the channel can be estimated
from the eigenvalue of the received covariance matrix excepting the multiplicative
factor ambiguity. A short training sequence is required to solved this ambiguity.
Furthermore, to improve the performance of our approach, we investigate the join
eigenvalue-decomposition-based approach and the Iterative Least-Square with Pro-
jection algorithm. The numerical results verify the eectiveness of our channel
estimate approach.
We consider the massive MIMO downlink with time-division duplex (TDD) op-
eration and conjugate beamforming transmission. To reliably decode the desired
signals, the users need to know the eective channel gain. In this paper, we propose
a blind channel estimation method which can be applied at the users and which does
not require any downlink pilots. We show that our proposed scheme can substan-
tially outperform the case where each user has only statistical channel knowledge,
and that the dierence in performance is particularly large in certain types of chan-
nel, most notably keyhole channels. Compared to schemes that rely on downlink
pilots (e.g., [41]), our proposed scheme yields more accurate channel estimates for
a wide range of signal-to-noise ratios and avoid spending time-frequency resources
on pilots.
Authored by Hien Quoc Ngo, Himal A. Suraweera, Michail Matthaiou, and Erik G.
Larsson.
power of each source and of the relay station can be reduced by 1.5 dB if the pilot
power is equal to the signal power, and by 3 dB if the pilot power is kept xed,
while maintaining a given quality of service.
Can the channel be blindly estimated? Can payload data help improve
channel estimation accuracy? How much can we gain from such schemes?
37
38 Chapter 6. Contributions of the Dissertation
Should each user estimate the eective channel in the downlink? And
how much gain we can obtain?
• In our work, Paper F, we considered the case where the transmit powers during
the training and payload data transmissions are not equal, and are optimally
chosen. The performance gain obtained by this optimal power allocation was
studied. However, the cost of performing this optimal power allocation may
be an increase in the peak-to-average ratio of the emitted waveform. This
should be investigated in future work.
• In current works, we assume that each user has a single antenna. It would
be interesting to consider the case where each user is employed with several
antennas. Note that in the current wireless systems (e.g. LTE), each users can
have two antennas [48]. The transceiver designs (transmission schemes at the
BS and detection schemes at the users) and performance analysis (achievable
rate, outage probability, etc) of Massive MIMO systems with multiple-antenna
users should be studied.
[1] Qualcom Inc., The 100× Data Challenge, Oct. 2013. [On-
line]. Available: http://www.qualcomm.com/solutions/wireless-
networks/technologies/1000x-data.
[2] Ericson, 5G Radio AccessResearch and Vision, June 2013. [Online]. Avail-
able: http://www.ericsson.com/res/docs/whitepapers/wp-5g.pdf.
[3] Cisco, Cisco Visual Networking Index: Global Mobile Data Traf-
c Forecast Update, 20132018, Feb. 2014. [Online]. Available:
http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-
networking-index-vni/white-paper-c11-520862.html.
[8] P. Viswanath and D. N. C. Tse, Sum capacity of the vector Gaussian broad-
cast channel and uplink-downlink duality IEEE Trans. Inf. Theory, vol. 49,
no. 8, pp. 19121921, Aug. 2003.
41
42 References
[11] N. Jindal and A. Goldsmith, Dirty-paper coding vs. TDMA for MIMO broad-
cast channels, IEEE Trans. Inf. Theory, vol. 51, no. 5, pp. 17831794, May
2005.
[16] J. Hoydis, S. ten Brink, and M. Debbah, Massive MIMO in the UL/DL of
cellular networks: How many antennas do we need?, IEEE J. Sel. Areas
Commun., vol. 31, no. 2, pp. 160171, Feb. 2013.
[20] , Eect of oscillator phase noise on uplink performance of large MU-
MIMO systems, in Proc. of the 50-th Annual Allerton Conference on Com-
munication, Control, and Computing, 2012.
[21] W. Yang, G. Durisi, and E. Riegler, On the capacity of large-MIMO block-
fading channel, IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 117132,
Feb. 2013.
[26] S. Payami and F. Tufvesson, Channel measurements and analysis for very
large array systems at 2.6 GHz, in Proc. 6th European Conference on Anten-
nas and Propagation (EuCAP), Prague, Czech Republic, Mar. 2012.
[45] Hien Quoc Ngo, Himal A. Suraweera, Michail Matthaiou, and Erik G. Larsson,
Multipair full-duplex relaying with massive arrays and linear processing,
IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 17211737, Sep. 2014.
[48] 3GPP TR 36.211 V 11.1.0 (Release 11), Evovled universal terrestrial radio
access (E-UTRA), Dec. 2012.
Fundamentals of Massive
MIMO
47
Papers
The articles associated with this thesis have been removed for copyright
reasons. For more details about these see:
http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-112780