Massive MIMO

8 Massive multiple-input
multiple-output (MIMO) systems

Antti Tölli, Lars Thiele, Satoshi Suyama, Gabor Fodor, Nandana Rajatheva,
Elisabeth De Carvalho, Wolfgang Zirwas, and Jesper Hemming Sorensen
8.1 Introduction
As stated in Chapter 2, one of the main 5G requirements [1] is to support 1000 times
larger capacity per area compared with current Long Term Evolution (LTE) technology,
but with a similar cost and energy dissipation per area as in today’s cellular systems.
In addition, an increase in capacity will be possible if all three factors that jointly
contribute to system capacity are increased: More spectrum, a larger number of base
stations per area, and an increased spectral efficiency per cell.
Massive or large Multiple-Input Multiple-Output (MIMO) systems are considered
essential in contributing to the last stated factor, as they promise to provide
a substantially increased spectral efficiency per cell. A massive MIMO system is
typically defined as a system that utilizes a large number, i.e. 100 or more, of individu-
ally controllable antenna elements at least at one side of a wireless communications link,
typically at the Base Station (BS) side [2][3]. An example of such usage of massive
MIMO at the BS side is shown in Figure 8.1. A massive MIMO network exploits the
many spatial Degrees of Freedom (DoF) provided by the many antennas to multiplex
messages for several users on the same time-frequency resource (referred to as spatial
multiplexing), and/or to focus the radiated signal toward the intended receivers and
inherently minimize intra-cell and inter-cell interference [4]–[7]. Such focusing of
radiated signals in a particular direction is possible by transmitting the same signal
from multiple antenna points, but with a different phase shift applied to each of the
antennas (and possibly a different phase shift for different parts of the system band-
width), such that the signals overlap coherently at the intended target location. Note that
in the remainder of the chapter, the term beamforming is used when applying the same
phase shift at individual transmit antennas over the entire system bandwidth, while the
term precoding is used when applying different phase shifts for different parts of the
system bandwidth to tackle small-scale fading effects, for instance by applying phase
shifts in frequency domain. With this definition, beamforming can be seen as a subclass
of precoding algorithms. Regardless of whether precoding or beamforming is applied,
5G Mobile and Wireless Communications Technology, ed. A. Osseiran, J. F. Monserrat, and P. Marsch.
Published by Cambridge University Press. © Cambridge University Press 2016.
Downloaded from http:/www.cambridge.org/core. New York University Libraries, on 16 Dec 2016 at 15:21:33, subject to the Cambridge Core terms of use, available at
http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/CBO9781316417744.009
Massive multiple-input multiple-output (MIMO) systems 209
Figure 8.1 Multiuser and single user with massive MIMO base stations.
the gain of obtaining a coherent overlap of signals at the receive point is commonly
referred to as array gain.
Besides being used for access links, massive MIMO can also play a key role in
creating multi-Gbps backhaul links between infrastructure nodes that are deployed in
Frequency Division as well as Time Division Duplexing (FDD/TDD) systems.
Although large antenna arrays in radar systems have been in wide use since the late
1960s, the commercial deployment of massive MIMO systems, i.e. for the access and
backhaul of mobile communications systems, has been considered only recently.
In particular, recent research and experiments with practical implementations have
identified the key challenges that must be overcome in order to realize the potential
benefits of massive MIMO in cellular communications [7], and which are listed in the
sequel.
One of the most severe challenges is given by the need of accurate Channel State
Information (CSI) at the transmitter side. In principle, the CSI may be obtained through
the transmission of orthogonal pilot signals (also called reference signals) from each
transmit antenna element, and the subsequent feedback of the observed spatial channel
from the receiver to the transmitter side. This approach has the drawback that the pilot
signal overhead in terms of required CSI grows linearly with the number of transmit
antennas. Another option for obtaining the CSI at the transmitter side is to utilize channel
reciprocity, which is for instance possible in TDD systems. The cost in utilizing
reciprocity is that it requires array calibration in order to take the differences in the
transmit/receive Radio Frequency (RF) chains of the different antenna elements into
account. In time varying channels, the delay between pilot transmission, channel
estimation, channel feedback, beamformer calculation and the actual beamformed data
transmission will degrade the performance of a massive MIMO. Fortunately, channel
210 Antti Tölli et al.
prediction techniques could be used to mitigate this delay, as for instance discussed in
Section 9.2.1.
Another challenge is to understand the impact of massive MIMO on the design of
multi-cell multi-tier networks [8]. One of the problems is the impact of pilot contamina-
tion [4][9]. The inherent trade-off between the time, frequency, code, spatial and power
resources allocated to pilot signals and data transmission in multi-antenna systems is
well known, see for example the classical results in [10]–[13] as well as the more recent
contributions in [14]–[16]. In multi-cell multi-user massive MIMO systems, the pilot-
data resource allocation trade-off is intertwined with the management of inter-cell
interference (also known as contamination) both on the pilot and data signals and calls
for rethinking the pilot signal design of classical systems such as the 3rd Generation
Partnership Project (3GPP) LTE system. Recent works provide valuable insights into the
joint design of pilot and data channels in multi-cell massive MU-MIMO systems [17].
Furthermore, the exploitation of massive MIMO in a network with an ultra-dense
deployment of small cells may be difficult in practice due to the size of the antenna
arrays. At higher frequencies, such as millimeter Wave (mmW), this is not an issue, as
discussed in Section 6.3.2 of Chapter 6. One approach is to deploy massive MIMO at the
macro side, and to exploit the spatial DoF to lower the interference between the macro
and the small cells, in the case of a co-channel deployment between macro and small cell
layer. Another approach is to consider possible massive MIMO deployments at the small
cell side, in particular in high-frequency systems, where the small antenna dimensions
allow the deployment of large-scale antenna arrays with a realistic form factor.
Based on the general observations above, different technology solution components
are currently investigated, which aim to unleash the promising potential and enable an
efficient usage of massive MIMO in 5G.
The remainder of this chapter is structured as follows. First, Section 8.2 provides the
theoretical background on massive MIMO, investigating the principle capacity scaling
behavior that can be expected from massive MIMO systems. The design of pilot signals,
covering all associated challenges such as the mentioned pilot contamination, is
presented in Section 8.3. A discussion of methods of resource allocation as given in
Section 8.4 is essential to understand the utilization of spatial DoFs in a massive MIMO
environment. Then, Section 8.5 introduces digital, analog and hybrid beamforming as
the principle forms of precoding and beamforming implementation in hardware, provid-
ing different trade-offs between the extent of massive MIMO gains captured and the
cost-efficiency of the implementation. Finally, Section 8.6 gives some ideas on suitable
channel models for massive MIMO investigation.
8.1.1 MIMO in LTE

Historically, the 3GPP standard for LTE was designed with MIMO as a goal to increase
capacity. LTE has adopted various MIMO technologies. In LTE Release 8, the downlink
transmission supports up to four antennas at the BS. For the uplink transmissions, only
a single antenna is supported for transmission from the user. There is an option for
performing antenna switching with up to two transmit antennas. Multi-User MIMO
(MU-MIMO) is also supported in the uplink. Further, LTE Release 10 (also known as
LTE-Advanced) provided enhanced MIMO technologies. A new codebook and feed-
back design are implemented to support spatial multiplexing with up to eight indepen-
dent spatial streams and enhanced MU-MIMO transmissions. In the uplink, single user
MIMO is utilized with up to four transmit antennas at the user side. In LTE Release 12,
downlink MIMO performance was enhanced thanks to the definition of new CSI
reporting schemes, which permitted the BS to transmit with more accurate CSI.
The LTE Release 13 looked into higher-order MIMO systems with up to sixty four
antenna ports at the BS, to become more appropriate to the use of higher frequencies.
8.2 Theoretical background
This section provides insight on the fundamental bounds and behavior of wireless
networks where the BSs are equipped with a large number of antenna elements.
In particular, the massive MIMO foundations are reviewed including a brief introduction
to the fundamentals of multi-user/cell MIMO communications and its analysis via
random matrix theory.
After Marzetta’s pioneering work [4], massive MIMO gained significant attention in
both the academic and the industrial communities. Assuming the dimensions of a MIMO
system grow large, the results of Random Matrix Theory (RMT) can be applied to
provide simple approximations, for example, for the user-specific Signal to Interference
plus Noise Ratio (SINR) expressions [18][19].
The linear MIMO transmitter-receiver design in a multi-user (multi-cell) setting,
which involves resource allocation, has received considerable attention in the literature;
see for example [9][20][21]. In general, (near) optimal design to maximize a certain
optimization objective often leads to an iterative solution where each sub-problem is
presented as an optimization problem, and at each iteration some information needs to be
exchanged between adjacent cells. In a massive MIMO setting, the coordinated trans-
ceiver design methods can be potentially simplified, as compared to state-of-the-art
iterative optimization-based schemes.
In a special case, where the imbalance between the number of independently fading
transmit antennas nt or receiver antennas nr at the serving BS and the number of users K
in the cell becomes large, i.e. nt ≫ K, the processing can be simplified in a way that even
Matched Filter (MF) and Zero-Forcing (ZF) can be used in an ideal independent and
identically distributed (i.i.d.) channel for near optimal detection and precoding [4][5].
However, in a general setting and given practical non-ideal conditions such as non-zero
correlation between antennas and physical limitation of antenna array sizes, more
complicated precoder designs with some limited coordination between adjacent cells
can be still highly beneficial [22][23].
In order to understand the basics, more details of single user and multiple user MIMO
cases are considered next. Both scenarios are incorporated in the general single cell
MIMO system given in Figure 8.2, where nt is the number of transmit antennas, nr the
UE
Stream 1
Stream 2
BS UE
Stream
Stream UE
Figure 8.2 Generic single and multiple user MIMO system.
number of receive antennas and M the number of spatial data streams transmitted
simultaneously to all users.
8.2.1 Single user MIMO

Assuming a narrowband and time-invariant MIMO channel with nt transmit (Tx)
antennas and nr receive (Rx) antennas, the received signal vector at symbol time m
can be described by
y½m ¼ Hx½m þ n½m; ð8:1Þ
where x 2 ℂnt is the Tx signal, subject to TrðE½xxH Þ ≤ P, y 2 ℂnr is the Rx signal,

n ∼ CNð0; N0 Inr Þ denotes the complex white Gaussian noise with noise variance N0 ,
and H 2 ℂnr nt the channel matrix. The achievable rate of the channel H measured in
bits per channel use is upper bounded by the mutual information between the input x and
output y as
C¼ max Iðx; yÞ
TrðE½xxH Þ ≤ P

1
H
¼ max logInr þ HKx H ; ð8:2Þ
Kx : Tr½Kx ≤ P N0
where Kx ¼ QPQH is the covariance matrix of x ∼ CNð0; Kx Þ, Q is a unitary steering

matrix and P ¼ diag ðp1 ; . . . :; pnt Þ. When the channel is known at the transmitter side,
the optimal strategy is to assign Q to the right singular vectors V of H while the power
allocation P is found via waterfilling.
On the other hand, when the elements of H are i.i.d. CNð0; 1Þ and the channel is not
known at the transmitter, the optimal Kx is
P
Kx ¼ In : ð8:3Þ
nt t
In such a case, the capacity of the MIMO channel (8.2) is simplified to [24],

P Xnmin P 2
C ¼ logInr þ HHH ¼ log 1 þ λ ; ð8:4Þ
nt N0 i¼1 nt N0 i
P
where nmin ¼ minðnt ; nr Þ, λi are the singular values of H and SNR ¼ .
N0
Let us now focus on the square channel n ¼ nt ¼ nr and define
Xn
λ2i
Cnn ðSNRÞ ¼ log 1 þ SNR : ð8:5Þ
i¼1 n
Assume now that the number of antennas becomes large, i.e. n→∞, then the distribution
pffiffiffi
of λi = n becomes a deterministic function
( pffiffiffiffiffiffiffiffiffiffiffiffiffi
1
f ðxÞ ¼ π 4 x 0 ≤ x ≤ 2;
2
ð8:6Þ
0 else:
For increasing n, the normalized capacity CðSNRÞ ¼ Cnn ðSNRÞ=n per spatial dimen-
sion becomes [24]
ð4
1 Xn λ2i n→∞
CðSNRÞ ¼ log 1 þ SNR → logð1 þ SNR xÞ f ðxÞdx ð8:7Þ
n i¼1 n 0
and the closed form solution to the integral is [24][25]

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 þ 4SNR þ 1 log e pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
CðSNRÞ ¼ 2log ð 4SNR þ 1 1Þ2 : ð8:8Þ
2 4SNR
Finally, when n→∞, the capacity of the n n point-to-point MIMO link can be
approximated as
lim Cnn ðSNRÞ

¼ CðSNRÞ→Cnn ðSNRÞ ≈ nCðSNRÞ ð8:9Þ
n→∞ n
The actual ergodic capacity EH ½CðHÞ is compared with the large-n approximation in
Figure 8.3 for n ¼ 2; 4; 8; 16 and 32. It can be seen that the approximation is very close
even for relatively small values of n. The result above can be extended to any fixed ratio
between transmit and receive antennas assuming the number of antenna elements grows
large at both ends of the transmission link [25]. Even though the capacity expression
becomes deterministic for large n, the rate optimal transmission strategy still requires in
general right singular vectors of H to be used as optimal transmit directions.
Assume now that nr ≫ nt , and the elements of H are i.i.d. CNð0; 1Þ. When nr becomes
very large, the columns of H ¼ ½h1 ; . . . ; hnt become close to orthogonal, i.e.
HH H
≈ Int : ð8:10Þ
nr
50
45 Approximate capacity
Ergodic capacity
40
n = 32
35
n = 16
Capacity [bps/Hz]
30
n=8
25
20
n=4
15
10
n=2
5
0
0 5 10 15 20
SNR [dB]
Figure 8.3 Capacity versus SNR.
By plugging in this approximation, the capacity of the MIMO link with or without CSIT
can be approximated as

P P
logInr þ HHH ¼ logInt þ HH H
nt N 0 nt N 0
Xnt
P‖hi ‖2
Pnr
≈ i¼1 log 1 þ ≈ nt log 1 þ : ð8:11Þ
nt N0 nt N0
Therefore, the MF receiver is the asymptotically optimal solution for nr ≫nt.

Similarly, when nt ≫nr and the elements of H are i.i.d. CNð0; 1Þ, the columns hi of
HT ¼ ½h1 ; . . . ; hnr become close to orthogonal, i.e.
HHH
≈ Inr : ð8:12Þ
nt
Then, the rate expression without CSIT is simplified to

P P
logInr þ HHH ≈ nr log 1 þ : ð8:13Þ
nt N0 N0
Thus, no array gain from having nt transmit antennas is attained. This is due to the lack of
channel knowledge at the transmitter, and hence, the power is evenly dissipated from all
nt antennas.
As the rows of H are asymptotically orthogonal, the nr dominant singular vectors

are asymptotically equivalent to the normalized rows of H, and hence the
transmit covariance matrix Kx can be approximated (at high SNR, assuming equal
power loading) as
P H
Kx ¼ VPVH ≈ H H; ð8:14Þ
nr nt
where the matrix V corresponds to the right singular vectors of H. Thus, using the MF
precoders at the transmitter is the asymptotically optimal solution. In such a case, the rate
expression with full CSIT can be simplified to

1 nt P
C ¼ logInr þ HKx HH ≈ nr log 1 þ ; ð8:15Þ
N0 nr N0
providing nt =nr -fold array gain.
8.2.2 Multi-user MIMO

8.2.2.1 Uplink channel
Assume a time-invariant uplink channel with K single-antenna users and a single BS
with nr receive antennas such that nr ≫K. The received signal vector at symbol time m is
described by
XK
y½m ¼ k¼1
hk xk ½m þ n½m ¼ Hx½m þ n½m; ð8:16Þ
where xk is the Tx symbol of user k, per user power constraints E½‖xk ‖2 ≤ Pk ; y 2 ℂnr is
the Rx signal, n ∼ CNð0; N0 Inr Þ denotes the complex Gaussian noise and
pffiffiffiffiffi
hk ¼ ak h k 2 ℂnr is the channel vector of user k, where ak is the large-scale fading
factor and h k is the normalized channel.
The sum capacity expression for the multiuser MIMO is equal to the Single User (SU)
MIMO without CSIT, i.e.

XK Pk 1

Csum ¼ logInr þ H
h h ¼ logInr þ H Kx HH j; ð8:17Þ
k¼1 N k k N
0 0
where H ¼ ½h1 ; . . . ::; hK and Kx ¼ diag ðP1 ; . . . :; PK Þ.

Assume nr ≫K, and that the elements of the normalized channel vectors h k are i.i.d.
CNð0; 1Þ, then
HH H
≈ AK ; ð8:18Þ
nr
where AK ¼ diag ða1 ; . . . :; aK Þ. Then, the sum rate can be approximated as

1 1
Csum ¼ logInr þ H Kx HH ¼ logIK þ Kx HH H
N0 N0
XK
Pk ak ‖h k ‖2 XK nr Pk ak
≈ log 1 þ ≈ log 1 þ ; ð8:19Þ
k¼1 N0 k¼1 N0
and again the matched filter receiver is the asymptotically optimal solution.
8.2.2.2 Downlink channel

The downlink problem is somewhat different due to the sum power constraint across
users. Assume a time-invariant downlink channel with K single-antenna users and
a single BS with nt transmit antennas such that nt ≫K. The received signal vector at
user k at symbol time m is
yk ½m ¼ hHk x½m þ nk ½m

pffiffiffiffiffi XK pffiffiffiffi
¼ hH
k uk pk dk ½m þ hH u pi di ½mwk ½m;
i¼1;i ≠ k i i
ð8:20Þ
where x 2 ℂnt is the Tx signal vector, subject to power constraint

P
EðTr½xxH Þ ¼ Kk¼1 pk ≤ P; uk 2 ℂnt is the normalized precoder with ‖uk ‖ ¼ 1; dk 2 ℂ
is the normalized data symbol with E½jdk j2 ¼ 1; yk 2 ℂ is the Rx signal,
pffiffiffiffiffi
nk ∼ CNð0; N0 Þ is the complex white Gaussian noise and hk ¼ ak h k 2 ℂnt is the
channel vector of user k, assumed to be ideally known at the transmitter.
Maximization of the downlink sum rate can be expressed via the dual uplink reformula-
tion, where the role of the transmitter and the receivers is reversed [24]. The sum rate
optimal solution is attained from the constrained optimization problem

1 XK H

max log2 Int þ qhh
k¼1 k k k
ð8:21Þ
qk N0
subject to
XK
k¼1
qk ≤ P; qk ≥ 0; k ¼ 1; . . . ; K;
where qk is the dual uplink power defined as the uplink power of the dual uplink
reformulation mentioned above such that the sum power between downlink and dual
P P
uplink powers holds Kk¼1 qk ¼ Kk¼1 pk ¼ P. When nt ≫K, the objective of (8.21) is
simplified to

1 XK qk nt ak
max logIK þ Kx HH H ≈ max log 1 þ ; ð8:22Þ
qk N0 qk k¼1 N0
HH H
where Kx ¼ diag ðq1 ; . . . :; qK Þ and ≈ diagða1 ; . . . ::; aK Þ. Since the inter-user
nt
interference vanishes as nt ≫K, the dual uplink power allocation is equal to the downlink
Table 8.1 Asymptotic capacity scaling behavior for massive MIMO assuming large imbalance between
transmit and receive antennas.
Mode Transmission CSIT Feedback Asymptotic capacity scaling behavior

P
SU-MIMO, nt ≫nr DL & UL No CSIT nr log 1 þ
N0

nt P
SU-MIMO, nt ≫nr DL & UL CSIT nr log 1 þ
nr N0

nr P
SU-MIMO, nr ≫nt DL & UL No CSIT nt log 1 þ
nt N0
XK
qk nt ak
MU-MIMO, nt ≫K DL CSIT max log 1 þ
qk N0
XK
k¼1

nr Pk ak
MU-MIMO, nr ≫K UL No CSIT log 1 þ
k¼1
N0
power allocation, i.e. pk ¼ qk 8k. Combining above relations, the optimal power alloca-
tion can be found via a simple waterfilling principle

N0
pk ¼ max 0; μ ; ð8:23Þ
nt ak
XK
where the optimal water level μ is found to satisfy the power constraint p ≤ P.
k¼1 k
8.2.3 Capacity of massive MIMO: a summary

The asymptotic capacity scaling behaviors of massive MIMO systems for various
MIMO modes and directions are summarized in Table 8.1, assuming a large imbalance
between transmit and receive antennas, and perfect CSI at the receiver. Note that for
finite imbalance the accuracy of the approximation depends on the correlation between
antennas.
8.3 Pilot design for massive MIMO
Spectral and energy-efficient operation relies heavily on the acquisition of accurate CSI
at the transmitters and receivers in wireless systems in general, and in Orthogonal
Frequency Division Multiplexing (OFDM) and massive multiple antenna systems in
particular. Consequently, channel estimation methods have been studied extensively and
a large number of schemes, including blind, data-aided, and decision-directed non-blind
techniques, have been evaluated and proposed in the literature. One reason for this is
that, for conventional coherent receivers, the effect of the channel on the transmitted
signal must be estimated in order to recover the transmitted information. As long as the
receiver accurately estimates how the channel modifies the transmitted signal, it can
recover the transmitted information. In practice, pilot signal-based data-aided techni-

ques are used due to their superior performance in fast fading environments and cost
efficiency and interoperability in commercial systems [26]–[28].
As the number of antennas at the BS grows large, it is desirable to have pilot-based
schemes that are scalable in terms of the required pilot symbols and provide high-quality
CSI for uplink data detection and downlink precoding. To this end, massive MIMO
systems rely on channel reciprocity and employ uplink pilots to acquire CSI at BSs.
Although solutions for non-reciprocal systems (such as systems operating in FDD
mode) are available [29], it is generally assumed that massive MIMO systems can
only feasibly operate in TDD mode [4][5].
Pilot reuse in general causes contamination of the channel estimates, which is
known as Pilot Contamination (PiC) or pilot pollution. As there are a large number of
channels to be estimated in massive MIMO systems, this becomes a significant
challenge due to the potentially limited number of pilots available. Consequently,
PiC limits the performance gains of non-cooperative MU-MIMO systems [4][30].
Moreover, PiC could cause a saturation effect in the SINR as the number of BS
antennas increases to a very large value. This is in contrast to the scenario where the
SINR increases almost linearly with the number of antennas without the presence of
PiC [30].
In the following subsections the impact of CSI error on the performance of pilots in
a massive MIMO setup is shown. Further, two methods for combating pilot contam-
ination in massive MIMO systems are presented. The first method, Pilot Power
Control (PPC) based on Open Loop Path loss Compensating (OLPC), provides
efficient measures against the detrimental effects of PiC. The second method, coded
random-access protocol for massive MIMO systems, brings a fundamentally new
approach to mitigate pilot contamination, that is, instead of avoiding intra-cell inter-
ference it embraces it.
8.3.1 The pilot-data trade-off and impact of CSI

Although pilot-based CSI acquisition is advantageous in fast fading environments, its
inherent trade-offs must be taken into account when designing channel estimation
techniques for various purposes. These purposes include demodulation, precoding or
beamforming, spatial multiplexing and other channel-dependent algorithms such as
frequency-selective scheduling or adaptive Modulation and Coding Scheme (MCS)
selection [11]–[13]. The inherent trade-offs between allocating resources to pilot and
data symbols include the following:
• Allocating more power, time or frequency resources to pilots improves the quality of
the channel estimates, but also leaves fewer resources for uplink or downlink data
transmission [11]–[13][16][31].
• Constructing longer pilot sequences (for example, employing orthogonal symbol
sequences such as those based on the well-known Zadoff-Chu sequences in LTE-
Advanced systems) helps to avoid tight pilot reuse in multi-cell systems, which helps
to reduce or avoid inter-cell pilot interference. On the other hand, spending a greater
number of symbols on pilots increases the pilot overhead and might violate the
coherence bandwidth [13][32][33].
• Specifically in multiuser MIMO systems, increasing the number of orthogonal pilot
sequences may increase the number of spatially multiplexed users at the expense of
spending more symbols when creating the orthogonal sequences [12].
In addition to these inherent trade-offs, the arrangement of the pilot symbols in the time,
frequency and spatial domains have been shown in practice to have a significant impact
on the performance of MIMO and massive MIMO systems in particular; see for example
[11][12][33].
8.1.1.1 Impact of channel state information errors on the throughput

of massive MIMO systems
When CSI is affected by channel estimation errors, the performance of channel adaptive
algorithms, such as symbol detection or MIMO transmission mode selection, degrades.
This leads to an overall degradation of the system performance in terms of spectral or
energy efficiency, as well as provided end-user data rates or Quality of Service (QoS) [34].
For example, a per-cell sum rate degradation due to erroneous CSIT in a system
employing 50 antennas at the BS from 17 bps/Hz down to 5 bps/Hz was reported in [34]
while [35] noted that the per-cell sum rate degradation depends heavily on whether the
BS uses Maximum Ratio Transmission (MRT) or ZF for the downlink transmission; see
Figure 8.4. The figure compares the per-cell sum rate performance of a 7-site TDD
30
MRT: LS
MRT: MMSE
MRT: Perfect CSI
25 ZF: LS
ZF: MMSE Superior ZF
Per−cell sum rate [bps/Hz]
ZF: Perfect CSI performance

with perfect CSI
20
15
Similar ZF and MRT performance

with non-perfect CSI
10
5
20 30 40 50 60 70 80 90 100
Number of antennas
Figure 8.4 Comparison of the impact of channel estimation error on the downlink sum rate
with MRT and ZF precoding.
system with an inter-site distance of 500 m, in which each site accommodates 3 cells and
serves 12 users per cell. Least Square (LS) and Minimum Mean Square Error (MMSE)
estimation methods are compared. It can be noticed that with imperfect CSI, the
performance degradation of ZF precoding is more severe than that of MRT precoding,
even with a moderate number of BS antennas.
8.3.2 Techniques to mitigate pilot contamination

In order to reduce the pilot overhead, pilot sequences in nearby cells can be reused at
the cost of intercell pilot interference; leading to PiC as introduced before. PiC has
been shown to limit the achievable performance of non-cooperative MU MIMO
systems [36].
Specifically, it has been found that PiC may cause the saturation of the SINR as
the number of BS antennas increases to infinity, while the SINR increases approxi-
mately linearly with the number of BS antennas nt in the absence of PiC. More
precisely, as pointed out by [36], when the number of users is comparable to the number
of antennas, the performance of a simple matched filter with contaminated estimate is
limited by the pilot interference. These insights led the research community to find
effective measures to mitigate the impact of PiC, both in non-asymptotic and asymptotic
regimes [37][38].
A large number of PiC mitigation schemes have been evaluated in the literature due to
the importance of accurate CSI and the strong negative impact on the quality of the
acquired CSI. These schemes differ in complexity and in the assumptions regarding
multi-cell cooperation, as well as in whether they operate in the time, frequency or power
domains.
A precoding scheme was proposed in [39], according to which each BS linearly
combines messages aimed to users of different cells that reuse the same pilot sequence.
This limited collaboration between BSs can resolve the PiC problem and allow for tight
pilot reuse.
Another approach to improve the channel estimation involves using a Bayesian
estimator that mitigates PiC for spatially well-separated users [34]. However, imple-
mentation of the Bayesian estimator relies on the knowledge of the second-order
statistics of the useful and crosstalk channels. Acquiring this knowledge entails some
overhead of estimating the covariance matrices and computation complexity.
An iterative filter may be employed that avoids explicit estimation of covariance
matrices, but its convergence remains an open problem [36].
A low-complexity Bayesian channel estimator known as the polynomial expansion
channel was proposed in [40]. This method is efficient in the presence of PiC.
Limited cooperation, which is based on the exchange of second-order channel
statistics and makes use of Bayesian channel estimation, is proposed in [34]. In this
method, PiC can be almost completely eliminated by allocating the same pilots to
spatially well-separated users. A different approach based on a less aggressive pilot
reuse is proposed in [37]. This approach can be effectively combined by spatially
separating users in different cells [38].
20
LS with max pilot power

15
LS: Maximum pilot power
LS: Pilot OLPC
Average NMSE [dB]
MMSE: Maximum pilot power

10
Gain of OLPC
Gain of
5 MMSE
estimation
0 MMSE
LS with pilot OLPC
–5
0 20 40 60 80 100
Number of antennas nt
Figure 8.5 The average normalized mean square error (NMSE) performance of LS and MMSE channel
estimation.
8.3.2.1 Pilot power control based on open loop path loss compensation
Within the framework of current LTE measurements and employing an LS estimator,
pilot power control based on the OLPC scheme of 3GPP LTE systems and pilot reuse
schemes can also provide feasible and efficient measures against the detrimental
effects of PiC [35]. The positive impact of employing the OLPC scheme on the
transmitted pilot symbols is illustrated in Figure 8.5. The figure shows the average
normalized mean square error of the estimated channel when LS or MMSE estimation
is used by the BS to acquire CSI. A multi-cell system is assumed where either full
power or OLPC for pilot transmissions is employed. As the figure shows, LS with
OLPC yields lower average NMSE than the much more complex MMSE estimation
using maximum pilot power.
8.3.2.2 Coded random access in massive MIMO systems

The rationale of using pilot sequences across the cellular network, in a way that
decreases the amount of inter-cell interference, can be challenged [41]. In fact, for
a very dense population of users with varying traffic patterns, a centralized pilot
allocation may become infeasible. Instead, a random access protocol may be a more
suitable choice. Consider a scheme where inside a cell a pool of pilot signals is available
for users to select randomly. The pools of different cells are orthogonal. Following this
random selection, users from the same cell will choose the same pilot sequence causing
a collision in the random access process. Hence, PiC is seen as a collision. However, PiC
is limited to within cells, where collisions can be processed and mitigated more easily.
In order to avoid intra-cell interference, conventional massive MIMO systems operate
with τ ≥ K, where τ is the length of the pilot sequences, and thereby the maximum
number of mutually orthogonal sequences, and K is the number of active users in a cell.
In [41] a fundamentally different approach to the mitigation of pilot contamination is
proposed where τ < K. The philosophy is to move the inevitable interference from
outside the cell to within the cell, such that it can be handled through appropriate
medium access procedures. The decrease of τ makes it possible to apply larger pilot
reuse factors, such that inter-cell interference is virtually non-existent. Hence, instead
of considering pilot contamination as an inter-cell interference problem that calls for
pilot planning, it is considered as an intra-cell interference problem that calls for
Medium Access Control (MAC) protocols. Especially in crowd scenarios (e.g. the large-
outdoor-event or stadium use cases presented in Chapter 2) with tens of thousands of
users served by a single cell, such an approach is interesting, since insisting on τ ≥ K
becomes prohibitively expensive. Furthermore, random MAC protocols, like the differ-
ent variants of ALOHA, are particularly suited for crowd scenarios with unpredictable
traffic patterns. An important motivation for coded random access is that the asymptotic
normalized throughput, i.e. decoded messages per resource block for increasing K,
approaches 1 [42], which is optimal.
Recently, random access protocols (see Chapter 7) have been further developed with
inspiration from the area of erasure coding [43][44]. Instead of considering collisions,
i.e. multiple users randomly selecting the same resource, as wasted resources, they are
processed using SIC. This potentially results in resolved collisions, which improves the
overall throughput. The combination of a random access protocol and SIC resembles an
erasure code, which makes the theory developed in this area applicable. In [41] the
framework of coded random access is adopted in the proposal of a joint pilot training and
data transmission scheme. Hence, the solution to the pilot contamination problem
becomes an integral part of the MAC protocol. An example of the coded random access
in the uplink transmissions for three time slots is depicted in Figure 8.6.
The joint channel and data acquisition is driven by low-complexity and low memory
requirements, which excludes collision resolution before spatial processing, demanding
storage of many received vector signals. This means that pilot contamination is not
solved: contaminated channels are used as matched filters transferring collisions to the
post-processing data domain. Two fundamental properties of massive MIMO are
User 1 Data Data
User 2 Data
User 3 Data Data
User 4 Data Data
User 5 Data Data
UL time slot 1 UL time slot 2 UL time slot 3

Figure 8.6 Example of uplink transmission using coded random access with 5 users ui and a set of 2 pilot
sequences s1 and s2 .
exploited to decode the data: (1) asymptotic orthogonality between user channels; and
(2) asymptotic invariance of the power received from a user over a short time interval.
In the following, the coded random access is analyzed for the uplink and downlink
transmissions.
Uplink
An uplink transmission is illustrated in Figure 8.6. Each user is active in a time slot with
probability pa. Active users transmit a pilot sequence sk ¼ ½sk ð1Þ sk ð2Þ . . . sk ðτÞ, chosen
randomly among a set of size τ, and send a data packet. One given user retransmits the
same data at each active time slot, which ensures each message is represented in several
encoded messages, similar to erasure codes. A½ν denotes all active users in time slot ν
and Aj ½ν denotes the subset applying sj . Following the example in Figure 8.6 with
A½2 ¼ fu2 ; u4 ; u5 g, A1 ½2 ¼ fu2 g, A2 ½2 ¼ fu4 ; u5 g, the received uplink pilot signal in
time slot ν can then be expressed as
Xτ X
S½ν ¼ j¼1
h ½νsj þ NS ½ν;
k 2 Aj ½ν k
ð8:24Þ
where hk ½ν is the channel vector from user k in time slot ν to the massive array with nt
antennas. Users are assumed to be equipped with a single antenna. S½ν is a nt τ matrix
grouping all the received vectorial signals in the pilot phase. NS ½ν is a matrix of i.i.d.
Gaussian noise components containing the noise vectors at the massive array.
All active users transmit a message of length T in the uplink data phase. The message
from the kth user is denoted xk . The received uplink data signal in time slot ν is then
expressed as
X
Y½ν ¼ h ½ν xk þ N½ν:
k 2 A½ν k
ð8:25Þ
If channel estimation is performed based on the received pilot signals, the pilot
contamination problem appears. The least squares estimate, ^h ½ν, based on the pilot
j
signal in time slot ν from users applying sj , is found as

X
^ j
h ½ν ¼ ðsj sH 1
þ NSj ½ν;
j Þ Y½ν sH
j ¼ h ½ν
k 2 Aj ½ν k
ð8:26Þ
where NSj ½ν is the post-processed noise term originating from NS ½ν. Instead of achiev-
ing the estimate of individual channel vectors, a sum of channel vectors is achieved, i.e.
intra-cell interference is experienced. The contaminated estimates are not discarded but
instead used as matched filters on the received signals in order to produce linear
combinations as follows:
X
f j ½ν ¼ ^
j j
h ½νH Y½ν ¼ k 2 Aj ½ν
‖hk ½ν‖2 xk þ N ½ν; and ð8:27Þ
X
g j ½ν ¼ ^
j j
h ½νH S½ν ¼ k 2 Aj ½ν
‖hk ½ν‖2 sj þ N S ½ν: ð8:28Þ
j j
N ½ν and N S ½ν contain post-processed noise terms as well as cross-user channel scalar
products which are null only asymptotically. Both space-time equation systems are
jointly exploited to recover the data. The coefficients of the linear combinations in
(8.27) and (8.28) are the two-norms, ‖hk ½ν‖2 , of the involved channels. In a massive
MIMO system, these can be assumed slowly fading, contrary to the fast fading of the
many individual channel coefficients. This enables a simplified application of SIC on the
filtered signals in order to solve the equation systems. Initially, signals with no contam-
ination are identified, which provide the corresponding channel powers and data directly.
These signals are then cancelled out, through subtraction from any other signals they
may appear in. The iterative process continues until all data has been recovered.
As a simple example, consider time slots 1 and 3 in Figure 8.6 and noise-free
reception. The channel energies ‖hk ½ν‖2 are stable in time, so that the time index is not
written. User 1 and user 4 collide in time slot 1: the channel estimate corresponding to
the transmission of pilot s1 is h1 þ h4 . Applying this contaminated estimate as
a matched filter in the training and data domain, two signals are obtained
ð‖h21 ‖ þ ‖h4 ‖2 Þs1 and ‖h1 ‖2 x1 þ ‖h4 ‖2 x4 . In time slot 3, user 1 has an uncontaminated
transmission: the channel estimate corresponding to the transmission of pilot s2 is h1 .
Applying it as a matched filter, the following signals are obtained ‖h1 ‖2 s2 and ‖h1 ‖2 x1 .
Knowing h1 and hence ‖h1 ‖2 ; x1 can be estimated in time slot 3. Removing the
contribution of user 1 in time slot 1 decontaminates user 4, enabling an estimation of
‖h4 ‖2 and thereby x4. This resembles belief propagation decoding in erasure codes. With
appropriate selection of pa , the degrees, i.e. the plurality of the collisions, can be shaped
to follow a distribution, which favors belief propagation decoding.
Existing literature does not consider random MAC protocols specifically designed for
massive MIMO systems. Hence, the only reference for performance comparisons is the
conventional slotted ALOHA protocol. In [41], it is shown that coded random access
approximately doubles the throughput (for nt ¼ 500) compared to slotted ALOHA. See
also Section 7.6.3 for more details on coded random access.
Downlink
Conventionally, downlink operation in a massive MIMO system relies on channel
estimates achieved in the uplink phase and the assumption of reciprocity. However,
such estimates are not guaranteed to be available when applying SIC in the uplink, since
only the channel norms are obtained after successful decoding. A solution is to perform
data assisted channel estimation. Provided the xk ’s are sufficiently long, it is a valid
assumption that they are mutually orthogonal for all k. Hence, after successfully decod-
ing xk , an estimate of hk ½ν can be found as
^
h k ½ν ¼ ðxk xH 1
k Þ Y½ν xH
k ¼ hk ½ν þ Nk ½ν: ð8:29Þ
Downlink transmission for user k is thus possible whenever an uplink message has been
successfully decoded.
8.4 Resource allocation and transceiver algorithms for massive MIMO
In order to exploit the advantages of massive MIMO, in particular, the spatial

multiplexing and array gain offered, the allocation of resources among the users
needs to be designed carefully. In particular, a number of issues shall be taken into
account: which users and how many shall be scheduled simultaneously, how the
antenna weights (at the transmitter and receiver) shall be chosen, shall the transcei-
ver (i.e. antenna weight vector) and scheduling be designed jointly, etc. Moreover,
this section investigates specifically two resource allocation schemes designed for
a massive MIMO system.
The first scheme suggests how to design the transceiver of a massive MIMO system in
order to avoid the full exchange of CSI centrally between BSs. In particular, the optimal
minimum power beamformers are obtained locally at each BS relying on limited back-
haul information exchange between BSs.
The second scheme is baptized interference clustering and user grouping. It allows
serving a large number of users in a massive MIMO system by properly combining user
clustering, grouping and precoding.
8.4.1 Decentralized coordinated transceiver design for massive MIMO

In general, coordinated resource allocation problems can be formulated as optimiza-
tion problems, maximizing a desired utility in the network subject to some constraints,
which can be solved iteratively along with exchange of some information between
nodes at each iteration [45]–[48]. In a Massive MIMO setting, however, the coordi-
nated transceiver design methods can be potentially simplified, as compared to state-
of-the-art iterative optimization based schemes. The coordinated multi-cell minimum
power beamforming approach, which has been studied widely in the past, is such
a problem that satisfies a given SINR for all users while minimizing the total
transmitted power [46][48][49]. This scheme is the basis for more general problems,
and various solutions have been suggested for solving this optimization problem,
based on uplink-downlink duality and optimization decomposition approaches.
Sharing instantaneous CSI between nodes under a delay constraint and limited back-
haul capacity becomes an important problem when the dimensions of the problem, i.e.
the number of antennas nt and the number of users K, grow large or when dealing with
a fast fading scenario.
A scheme to avoid the full exchange of CSI is based on the decoupling method
developed in [46] where the optimal minimum power beamformers can be obtained
locally at each BS relying on limited backhaul information exchange between BSs.
The original centralized problem is reformulated hereafter and named as decentralized
coordinated transceiver design, such that the BSs are coupled by real-valued inter-cell
interference terms. The centralized problem can be decoupled by a primal or dual

decomposition approach leading to a distributed algorithm. The optimal beamformers
are obtained based on convergence of an iterative solution that solves the local problems
and exchanges resulted Inter-Cell Interference (ICI) values at each iteration.
The decentralized coordinated transceiver scheme consists of a set of rules how to
design the transceiver of a massive MIMO system in order to avoid the full exchange of
CSI centrally between BSs. In particular, the optimal minimum power beamformers are
obtained locally at each BS relying on limited backhaul information exchange between
BSs. This scheme exploits the results of Random Matrix Theory (RMT) while making
use of the Stieltjes transform [22][23].
8.4.1.1 System model

The cellular system consists of NB BSs and K single antenna users; each BS has nt
transmitting antennas. Users allocated to BS b are in set Ub . The signal for user k
consists of the desired signal, intra-cell and inter-cell interference. Let hb;k be the
channel from the BS b to the user k. Assuming wk as the transmit beamformer for user
k, the following coordinated minimum power precoding problem subject to user-specific
SINR constraints γk 8k 2 Ub ; b 2 B is solved [46]:
X X
min b 2 B k 2 U ‖wk ‖2
b
2
jwH
k hbk ;k j
subject to X X ≥ γk 8k 2 Ub ; b 2 B and
σ2 þ l 2 U bk
jwH hbk ;k j2 þ
k l b ≠ bk
ε2b;k
X 2
l 2 Ub
jwH
l hbl ;k j ≤ εb;k 8k ≠ Ub ; b 2 B;
2
ð8:30Þ
where the optimization variables are wk and inter-cell interference variable ε2b;k .
A generic per-user channel correlation model is introduced such that
1
hb;k ¼ θb;k
2
gb;k ; ð8:31Þ
where θb;k is the correlation matrix of user k and gb;k is a vector with i.i.d. complex
entries with variance 1=nt . This per-user channel correlation model can be applied to
various propagation environments.
The optimization problem above can be solved iteratively using uplink-downlink
duality by first computing the power allocation and beamformers in the dual uplink [46]
[48]. It was shown that, by using large system analysis, the approximately optimal uplink
powers ƛk for the generic model are given by [22][23]
γk
ƛk ¼ ; ð8:32Þ
vbk ;θbk ;k ð1Þ
where vbk ;θbk ;k ðzÞ is the Stieltjes transform of a measure defined for z 2 C nRþ by
[23], and
!1
1 1X ƛl θbk ;l
vbk ;θbk ;k ðzÞ ¼ trθbk ;k zInt : ð8:33Þ
nt nt l 1 þ ƛl vbk ;θbk ;l ð1Þ
The approximation for downlink powers can be derived similarly, see [22][23] for
details. The above approximations result in an algorithm that gives the approximately
optimal uplink and downlink powers based on local CSI and statistics of other BS
channels. However, the error in approximations causes variations in the resulting SINRs
and rates. Thus, the SINR constraints cannot be guaranteed and those achieved SINRs
might be higher or lower than the target SINRs. In this case, the SINR constraints can be
met only asymptotically as the number of users and antennas grow large.
Next, an alternative approach for decoupling the sub-problems at the BSs is
described. Following the same logic as in [46], ICI is considered as the principal
coupling parameter among BSs, and the large dimension approximation for ICI
terms based on statistics of the channels is derived as [22]:
0
X 1 vbk ;θb ;l ð1Þ
ε2b;k ≈ l 2 Ub n
k
2 ; ð8:34Þ
t
1 þ ƛl vbk ;θbk ;l ð1Þ
0
where vbk ;θb ;l ð1Þ is the derivative of the Stieltjes transform vbk ;θbk ;l ðz ¼ 1Þ (see [23]),
k
δl is the downlink power weighting scalar that relates the optimal downlink and uplink
pffiffiffiffiffi
precoding/detection vectors of user k i.e. wk ¼ δk w e k is the MMSE uplink
e k , where w
detection vector of user k. This approximation allows the derivation of approximately
optimal ICI terms based on statistics of the user channels. Each BS needs the knowledge
about user-specific average statistics, i.e. user-specific correlation properties and path
loss values from other BSs (these statistics can be exchanged over the backhaul between
coordinating network nodes). In addition, each BS needs to know the local CSIT to each
user within the coordination cluster. Based on the statistics, each BS can locally and
independently calculate the approximately optimal ICI values. Plugging the approxi-
mate ICI into the primal problem decouples the sub-problems at the BSs and the
resulting SINRs satisfy the target constraints with slightly higher transmit power
compared to the optimal method. The fact that the coupling parameters depend only
on channel statistics results in a reduced backhaul exchange rate and processing load.
Moreover, the algorithm can be applied to fast fading scenarios as the channel statistics
(path loss, correlation properties) change slower than the instantaneous channel
realizations.
8.4.1.2 Performance results

The two algorithms developed in the previous section for a multi-cell system with large
dimensions provide good approximations even when the dimensions of the problem (i.e.
the number of users and antennas) are practically limited. In order to show the perfor-
mance of the approximate algorithms, some numerical examples are presented in this
section. The results for the algorithm based on ICI approximation are presented. This
algorithm satisfies the target SINRs for all users; however, the error in the approxima-
tions results in a somewhat higher transmit power at the BSs. A wrap-around network
with 7 cells is considered and users are equally distributed within cells. An exponential
path loss model is used for assigning the path loss to each user,
2:5
d0
ab;k ¼ ; ð8:35Þ
db;k
where db;k is the distance between BS b and user k. The path loss exponent is 2.5 and the
reference distance ðd0 Þ is 1m. The path loss from a BS to the boundary of the reference
distance of the neighboring BS is fixed to 60 dB. The correlation among channel entries
is introduced using a simple exponential model
½θb;k i;j ¼ ρjijj ; ð8:36Þ
where ρ represents the correlation coefficient which is 0.8 for the following simulations.
The users are dropped randomly for each trial and in total 1000 user drops are used for
calculating the average transmit power. The number of antennas at each BS varies
from 14 to 84 and the total number of users is equal to half the number of antennas at
each BS. Thus, the spatial loading is fixed as the number of antennas is increased.
Figure 8.7(a) illustrates the transmit SNR versus the number of antennas for a 0 dB
SINR target. It is clear that the gap between the approximated and optimal algorithm
(denoted as centralized) diminishes as the number of antennas and users increase.
The small gap for small dimensions indicates that the approximate algorithm can be
applied to the practical scenarios with a limited number of antennas and users. From
the results, it is clear that the centralized algorithm and the approximated ICI
algorithm outperform the ZF method. Note that the gap between ZF and optimal
and approximated method is fixed, which is due to the fixed ratio of the number of
antennas to the total number of users. The gap in performance is mainly because the
ZF algorithm wastes a degree of freedom for nulling the interference toward the
distant users, while the centralized algorithm finds the optimal balance between
interference suppression and maximizing the desired signal level. MF beamforming
must be dealt with more care since it completely ignores the interference (both intra-
and inter-cell), and hence the SINR target is below the target SINR shown in
Figure 8.7(b). Note that MF beamforming can satisfy the target SINR only asympto-
tically in a very special case, i.e. when the ratio of the number of antennas to the
number of users approaches infinity.
8.4.2 Interference clustering and user grouping

Although massive MIMO transmission reduces the inter-user interference significantly,
multi-cell interference is still an issue that limits the overall system performance.
A particular interference clustering and user-grouping scheme described in this section
allows tackling the interference by separating the multi-cell interference mitigation from
(a) Num. of users = /2, SINR = 0 dB

67
66
MF beamforming
ZF
65 Centralized
Aproximated ICI
Transmit SNR [dB]
64
63
62
61
60
59
14 28 42 84
Number of transmit antennas
(b)
0.5
MF Beamforming
0 ZF
Centralized
Aproximated ICI
−0.5
−1
MF SINR[dB]
−1.5
−2
−2.5
−3
−3.5
−4
14 28 42 84
Number of transmit antennas
Figure 8.7 Comparison of required transmit SNR for 0 dB SINR target.
the multi-user multiplexing at each BS. In particular, a joint approach combining user
clustering, grouping and regularized precoding [50] is considered.
The scheme is depicted in Figure 8.8. It can be divided into three steps applying a
two-stage beamforming precoding. It should be noted that the two-stage
Large Scale
Antenna Array
Inter-cell
3. Precoding interference
2. Grouping
Inter-group
interference
Group 1
1. Clustering
Selected
Users Group
Figure 8.8 Inter-group interference clustering, user grouping and inter-cell interference aware precoding with
massive MIMO antenna arrays.
beamforming-precoding concept can be either implemented in digital domain only or

in a hybrid form as shown in Section 8.5.3:
• First (beamforming) stage: Inter-group interference aware clustering of co-located

users into groups resulting in wide-band multi-user beamforming. The beamforming
is extracted from the channel’s second order statistics, more precisely channel covar-
iance information.
• Second (precoding) stage (per user group): Optimizes the multi-user SINR conditions,
e.g. based on regularized ZF precoding constraints. User selection for downlink trans-
mission on the same time-frequency resource, and inter-cell interference aware precoder
design PðkÞ are performed for selected users based on small-scale fading CSI feedback.
In the first step, the user clustering, the set of users K is divided into G groups of users
with similar second-order channel statistics. The Joint Spatial Division Multiplexing
(JSDM) [51] concept of first-stage beamforming is used to spatially separate the
channels of each user group. A density-based clustering algorithm called Density-
based Spatial Clustering of Applications with Noise (DBSCAN) is employed. This
algorithm clusters an adaptive number of groups with respect to a certain user density
that is adapted to the level of inter-group interference.
In the second step, which is done independently for each group, Semi-orthogonal User
Selection (SUS) [51] based grouping is used to find a subset of users for simultaneous
downlink transmission on the same time-frequency resource but on different spatial
layers. In combination with the selected groups for downlink transmission, this step
covers the resource allocation. Note that the groups can be selected in time and frequency,
whereas the SUS corresponds to spatial domain resource allocation. Here, the SUS
algorithm is adapted with the maximum sum-rate objective using projection-based rate
approximation according to [52]. Hence, it is ensured that the sum-throughput is increased
while the limited transmit power budget is divided among all active spatial data streams.
Finally, in the third step, the second stage precoders are designed at the BS. In order to
balance the desired signal and the suppression of multiuser interference with remaining
non-treated inter-cell interference, regularized ZF precoding is considered. This is an
important step as the BS can obtain the channel via reciprocity, but does not have any
knowledge about the interference or SINR situation at the user. Since the modulation and
coding scheme cannot be selected matching the SINR conditions at the user, regularized
zero-forcing precoding is used such that the inter-cell interference is considered for the
regularization weights. To obtain knowledge of the SINR condition at the user, a scalar
broadband power-value, obtained as the average of the diagonal elements from the
interference-covariance matrix measured at the user, is introduced [52]. This power-
value is fed back from the users to the BS independently of the time or frequency
division duplex system.
Combining all these three steps, large performance gains are shown in the order of 10
times for the sum-throughput compared to a baseline scenario with 8 transmit antennas.
8.4.2.1 Performance results

The following performance evaluation focuses on an FDD downlink. To evaluate the
first step of this approach, a single sector with a 1x256 antenna array and 6 physically
closely located clusters of users, each consisting of 6 users, is utilized. This clustering
into space-orthogonal groups has to be done with respect to inter-group interference.
The clustering with the well-known K-means and K-means++ algorithm [53] with
a joint spatial-division multiplexing (JSDM) scheme is not practical for this task since
the number of groups G is required as an input parameter, which is not known a priori.
An exhaustive search for the optimal G of each user constellation is hardly feasible.
Therefore, a density based clustering algorithm is introduced by adapting the DBSCAN
algorithm.
The K-means++ and DBSCAN algorithms are compared with the baseline scenario in
Figure 8.9. For the baseline, 8 transmit antennas and multi-user regularized zero-forcing
beamforming were considered. The gain for ZF precoding over JSDM with per-group
processing is relatively small and attributes to 11% for DBSCAN user clustering. Note
that in JSDM only multi-user interference within in the same user group is mitigated by
ZF beamforming (second stage) while inter-group interference is considered in the first
precoding stage only.
Assuming perfect channel knowledge of the downlink user channels at the BS,
Figure 8.10 shows the multiplexing gain by increasing the number of user candidates
K ¼ jKj from 40 to 120. Further, the following two user-grouping modes are
analyzed:
JSDM regZF
150 142
132
Spectral efficiency [bps/Hz]
113
100
127
105
50 87
8
0
Baseline k-means k-means++ DBSCAN
Figure 8.9 JSDM with per group precoding compared to joint regularized ZF for 8 (baseline) and 256
(massive MIMO) transmit antennas.
80
WUS ZF
SUS ZF
Median sum spectral efficiency [bps/Hz]
70 WUS RZF
SUS RZF
60
50
40
30
20
40 50 60 70 80 90 100 110 120
Number of user candidates in sector
Figure 8.10 Multi-cell performance over users with and without user selection comparing zero-forcing and
regularized zero-forcing with interference power feedback.
• WUS: Without User Selection (WUS), where the set of scheduled users is KWUS ¼ K
• SUS: With SUS, where the set of scheduled users is KSUS ⊂ K
The solid and dashed lines with upper triangles in Figure 8.10 consider interference
aware Regularized ZF (RZF) precoding. It can be observed that RZF precoding (com-
pared to ZF) yields 150% and 20% gain at K ¼ 120 for SUS and WUS, respectively.
As a main result, multi-user interference is reduced to the level of noise plus out-of-cell
interference resulting in higher signal power per user and smaller loss due to precoder
normalization. The gain from SUS over WUS for RZF vanishes for a small number of
users, i.e. K ¼ 40; and increases by up to 16% for K ¼ 120 users. Note that SUS requires
a larger set of users to provide feedback and thus significantly increases feedback
overhead. For the simple ZF precoding advanced user grouping (such as SUS) is
mandatory, as otherwise the system performance degrades significantly, refer to the
solid line with downwards pointing triangle.
8.5 Fundamentals of baseband and RF implementations in massive MIMO
8.5.1 Basic forms of massive MIMO implementation

Beside the theoretical analysis for SU and MU massive MIMO, one major limiting
aspect of massive MIMO is related to hardware constraints [54]. It is clear that RF
hardware complexity scales with the number of active antenna elements nt in the system.
This has to be taken for granted and cannot be avoided in the context of massive MIMO.
However, the complexity of massive MIMO implementations can differ strongly
depending on the extent to which precoding and beamforming are conducted in the
digital frequency/time or analog time domain. As noted at the beginning of this chapter,
the term precoding here relates to the usage of individual phase shifts for each antenna
and each sub-part of the system bandwidth, while the term beamforming relates to the
usage of common phase shifts over the entire system bandwidth. In this section, the
possible basic forms of massive MIMO hardware implementation are shown.
In order to utilize all degrees of freedom and extract all gains from the massive MIMO
channel, it is required to introduce a channel-dependent individual phase shift to each
antenna element and each sub-carrier. Consequently, frequency-selective precoding
should be conducted in the digital baseband, as illustrated in Figure 8.11. The amount
of baseband chains is denoted as L ≥ M, where M represents the number of MIMO
streams for all the users. The precoding is employed to cope with inter-beam interference
for the L beams. Note that the beamforming and the precoding can be performed
simultaneously by one matrix multiplication combining the precoding matrix with the
digital BF matrix. In this setup, the number of baseband chains is equivalent to the
number of RF chains, i.e. L ¼ nt . In the figure, the different signal processing blocks refer
to the application of an Inverse Discrete Fourier Transform (IDFT) and the introduction of
Stream
Up-
Digital precoder
IDFT +CP DAC

converter
Digital BF
Up-
IDFT +CP DAC
converter
Figure 8.11 Massive MIMO OFDM transmitter employing digital precoding.
Analog BF
Up-
IDFT +CP DAC
converter
Up-
IDFT +CP DAC
converter
Figure 8.12 Massive MIMO OFDM transmitter employing analog BF.
a Cyclic Prefix (CP) before Digital to Analog Conversion (DAC). This approach is clearly
very complex and costly, as the number of baseband signal processing chains have to equal
the number of transmit antennas, and one requires CSI for each single channel coefficient
on each sub-carrier, which would imply a large pilot and channel estimation and feedback
overhead, unless channel reciprocity can be utilized in TDD.
The implementation complexity and cost can be significantly reduced by applying
phase shifts over the entire system bandwidth, i.e. by applying beamforming in the
analog domain, as depicted in Figure 8.12. Here, only M different baseband signal
processing chains are employed, for the purpose of serving up to M different streams,
and for each stream configurable phase shifts are introduced among the transmit
antennas in the RF circuitry. As the searching time of the phase shifts for the analog
beamforming should be shortened, only a very limited set of different phase shift
configurations are available. This has the positive effect that the effort for CSI feedback
is substantially reduced, as it is sufficient to feed back the index of the preferred phase
shift configuration from the receiver to the transmitter. On the other hand, the analog
beamforming approach implies that the signals transmitted from the different antennas
cannot coherently align (constructively or destructively) as perfectly as in the case of
digital, frequency-selective precoding. This means that on one hand the array gains are
not as large as they could be in the digital precoding case, and also that a certain extent of
residual interference between streams or users is introduced. Further, the fact that all
phase shifts are applied equally over the entire system bandwidth is clearly suboptimal in
the case the propagation channel is highly frequency-selective.
For these reasons, a more suitable approach to precoding and beamforming in the
context of massive MIMO is to use a hybrid beamforming approach [55]–[58], i.e. where
some extent of frequency-selective precoding is performed in the digital baseband, and
further beamforming is applied in the analog RF circuitry. This approach is depicted in
Figure 8.13. Here, the M streams are precoded in digital baseband and on a per-sub-
carrier basis to L different signals per sub-carrier to cope with the residual interference
between the analog beams and extract additional beamforming gain, and then analog
beamforming is used to map these signals to the nt transmit antennas. In this case, only L
baseband signal processing chains are required. The performance is in this case clearly
a compromise between the fully digital precoding and fully analog beamforming case.
Analog BF
Stream
Up-
Digital precoder IDFT +CP DAC converter
Up-
IDFT +CP DAC converter
Figure 8.13 Massive MIMO OFDM transmitter employing a hybrid of digital precoding and analog
beamforming.
This approach is more suitable to be used for frequency-selective channels, but the
limited flexibility of the analog beamforming part still leaves some residual inter-stream
or inter-user interference. Also, the effort for CSI feedback is now a compromise
between the fully digital precoding and fully analog beamforming approach.
For the latter hybrid approach, there are different mechanisms to determine the precoding
weights in the digital baseband domain and phase shifts in the analog domain. In the sequel,
a particular scheme based on hybrid Fixed BF and CSI-based Precoding (FBCP) [59],
which performs a successive optimization of both beamforming domains, is presented.
8.5.2 Hybrid fixed BF with CSI-based precoding (FBCP)

FBCP is referred sometimes as analog FBCP but was re-baptized herein as hybrid FBCP
in order to emphasize that it is implemented using a hybrid of precoding in the digital
baseband and analog beamforming. The FBCP approach is composed of two successive
steps. First, it selects the analog fixed BF weights whose number is larger than the
number of the streams and much less than the number of transmitter antennas. More
precisely, L fixed BF weights denoted by W are initially selected from some analog fixed
BF weight candidates based on the steering vector for arbitrary 2D angles and according
to a maximum total received power criterion [59].
Next, an Eigenmode (EM)-based precoding matrix PðkÞ is computed at the kth
subcarrier based on the SVD of an equivalent frequency-domain channel matrix
HðkÞW that multiplies the frequency-domain channel matrix HðkÞ by the selected
fixed BF weights W. Note that HðkÞW is estimated by exploiting a pilot signal with
the selected fixed BF.
8.5.2.1 Performance of FBCP

Link-level simulation results of single-user massive MIMO operating in TDD mode at
a carrier frequency of 20 GHz and employing the FBCP approach are shown in terms of
downlink throughput performances. The parameters used for the simulations are given in
Table 8.2. The number of transmitter antennas, nt , is set to 16 or 256, and the number of
receiver antennas, nr , is fixed to 16. The number of the streams, M, is fixed to 16. The UE
Table 8.2 Simulation parameters of Massive MIMO.
Concept Value
Transmission scheme Downlink Massive MIMO OFDM

Signal bandwidth 400 MHz
Active subcarriers Pilot: 32; data: 2000
No. of antennas nt : 16, 256; nr : 16
No. of data streams, M 16
Modulation scheme QPSK: 16QAM, 64QAM, 256QAM
Channel coding Turbo code: R = 1/2, 2/3, 3/4
Maximum bit rate 31.4 Gbps (256QAM, R = 3/4)
Antenna array structure UPA
Angular power spectrum Laplacian distribution in θ
Wrapped Gaussian distribution in ϕ
Average angle (θ, ϕ) Departure: (90 deg, 90 deg)
Arrival: (90 deg, 90 deg)
Angular spread (θ, ϕ) Departure: (5 deg, 5 deg)
Arrival: (20 deg, 20 deg)
Channel model Kronecker model
Fading channel Nakagami-Rice (K = 10 dB); 16-path
Figure 8.14 UPA for 3D BF in spherical coordinates.
receiver detects the spatially multiplexed streams by using a postcoding (receive weight)
matrix that is calculated from the SVD of the equivalent channel matrix HðkÞW. Both
the transmitter and the receiver employ a 2D Uniform Planar Array (UPA) as the antenna
array structure. As shown in Figure 8.14, θ and ϕ are zenith and azimuth angles.
The maximum bit rate reaches 31.4 Gbps according to the parameters set, where the
modulation and coding scheme is 256QAM with coding rate, R, of 3/4, using a turbo
code. Ideal adaptive modulation and coding (AMC) is assumed. The channel model was
based on the Kronecker model including Line of Sight (LOS) and Non-Line of Sight
(NLOS) components.
30
Average throughput [Gbps] Fully digital (EM)
25
FBCP ( = )
20
FBCP
( = )
15 FBCP
( = ) Fully digital (EM)
FBCP
( = ) with =
10 =
5 M == 16,
16,, AMC
AMC
LOS channel, K = 10 [dB]
w/o CSI error
0
0 5 10 15 20 25 30
Average SNR [dB]
Figure 8.15 Throughput of FBCP for different choices of L.
Figure 8.15 shows the throughput performance of the FBCP for different choices of L.
Ideal channel estimation is assumed and the pilot and feedback overhead is not taken into
account in the throughput calculation. Angular intervals for the possible analog BF steps
in zenith and azimuth are fixed to 5 degrees. For comparison, the throughput perfor-
mances of the EM precoding that employs the fully-digital Massive MIMO, i.e. with
L ¼ nt , are also plotted in Figure 8.15. It is shown (for the nt = 256 case) that as L
increases, the throughput of the FBCP approaches the fully digital EM-based precoding
and converges for L equal to 32. Moreover, in comparison with the conventional fully
digital MIMO with nt ¼ 32, the FBCP with nt ¼ 256 and L ¼ 32 can reduce the
required SNR for 20 Gbps throughput by more than 9 dB by exploiting higher BF and
diversity gains, while requiring the same number of baseband signal processing chains.
Figure 8.16 shows the throughput performance of the FBCP when this is subject to CSI
errors. Again, the throughput performance of the fully digital EM-based precoding is
added to the figure for comparison. For all schemes, the CSI error is generated according to
a complex Gaussian distribution with zero mean and a variance of σ 2e . σ 2n denotes the noise
power per antenna, and σ 2e is set to σ 2n – 10 dB or σ 2n – 20 dB. Figure 8.16 demonstrates
that as the CSI error increases, the throughput of the fully-digital EM-based precoding
approach drastically degrades, while the FBCP is robust to the CSI error. This can be
explained by the fact that the fully-digital approach requires accurate CSI to exploit the
full array gain. It is also found that when σ2e = σ 2n − 20 dB, the FBCP can achieve the same
throughput as the fully-digital approach, despite the significantly reduced complexity.
Finally, the throughput performance of the FBCP subject to phase errors is shown in
Figure 8.17. The phase errors reflect the hardware impairment in the RF phase shifters of
the FBCP and are generated by Gaussian distribution with zero mean and variance of σ 2p .
The standard deviation σ p is set to 3 degrees or 5 degrees. From Figure 8.17, it can be
seen that the FBCP achieves the same throughput irrespective of the phase error and is
30
w/o CSI error
Average throughput [Gbps] Fully digital (EM)
25
FBCP (
20
σ2 = σ 2 [dB]
15
FBCP
( σ2 = σ 2 [dB]
10
5 M== 16,
, AMC
Fully digital (EM)
LOS channel, K = 10[dB]
[dB]
nt = 256
0
0 5 10 15 20 25 30
Average SNR [dB]
Figure 8.16 Throughput performance of FBCP with CSI error.
30
25
Average throughput [Gbps]
20
FBCP
(
15 w/o phase error
σ [deg.]
Fully digital (EM) σ deg.]
10
5 AMC
LOS channel, K [dB]
σ2 = σ 2 [dB]
0
0 5 10 15 20 25 30
Average SNR [dB]
Figure 8.17 Throughput performance of FBCP with phase error.
fairly robust to this hardware impairment. Since the analog BF weights are chosen based
on the maximum received power criterion and inherently take the phase errors into
account, the FBCP is robust against the phase error.
8.5.3 Hybrid beamforming for interference clustering and user grouping

In this section, the advantage of using hybrid beamforming is shown for the interference
clustering and user-grouping scheme that was presented in Section 8.4.2. Figure 8.18
illustrates the interference clustering and user-grouping scheme implemented using
Data 1
RF Chain 1
Data 2
Digital RF Chain 2
precoding
(2nd stage) Data 3
RF Chain 3
Data
RF Chain
Spatial MUX at hot-spot
Analog beamforming (1st stage)
Figure 8.18 Beam steering control of antenna blocks with hybrid beamforming.
hybrid BF, where the number of radio frequency chains is L and the number of data
streams M. Furthermore, it is assumed a system with L ¼ M streams for K users and nt
antennas at the BS. In contrast to the hybrid BF shown in Figure 8.13, here it is assumed
that each RF chain feeds a single set of subarray antennas. Further, the antenna array is
partitioned into groups [17] of ϑ ¼ nt = L elements.
Mathematically, this means that the L nt matrix W is constrained to take on a block-
diagonal form with L diagonal blocks of dimension ϑ 1, each (column) block being the
steering vector for the corresponding beam-steering antenna. Consider for example the
case where the target is to serve M users belonging to the same group 1 and at the same
time create low interference to users in group 2. For an unconstrained first-stage
beamforming matrix W, this is easily achieved by block diagonalisation (see above
and details in [40]) in the case of the low-complexity hybrid BF implementation with
beam-steering antennas. However, when imposing a block-diagonal constraint on W,
the optimization of the beamforming matrix for the multi-user multi-cell downlink is far
from obvious. Driven by the LOS intuition this should be possible if the matrix W
defines multiple beams that point in the direction of the desired user-group 1 and put
a minimum in the direction of the undesired user-group 2. The multi-user multiplexing in
user-group 1 is enabled through the baseband precoding matrix PðkÞ and requires CSIT
incorporating the analog beamforming. However, whether this is possible and to what
extent one can actually achieve a multi-user multiplexing gain for each group of users,
has to be verified through an accurate statistical channel model and extensive system
simulation.
8.5.3.1 Performance of hybrid BF for interference mitigation

To evaluate the two-stage beamforming-precoding approach with hybrid BF, the follow-
ing deployment is considered, as depicted in Figure 8.19. Two transmitting BSs facing
toward each other and 2 receiver groups at the cell border, each with 8 users, are
considered. The focus of this section is to demonstrate the potential of the hybrid BF,
so the deployment is kept intentionally simple. User-groups 1 and 2 are forced to be
served by BS 1 and 2, respectively.
BS 1
4
Height [m]
2
user-
group 2 BS 2
0
0 user-
group 1
−50
Y− 100
Po −100 50
sit
io −150 0 ]
n
[m n [m
] −200
−50
os itio
−100 X−P
Figure 8.19 Deployment of users and BSs.
user-group 1 0
0
Normalized received power [dB]

user-group 2 −5
BS 1 −10
−50 BS 2
−15
Y−direction [m]
−20
−100 −25
−30
−150 −35
−40
−45
−200
−50
−100 −50 0 50 100
X−direction [m]
Figure 8.20 Received power of the analog beamformed data stream 1 transmitted from the first antenna block
in Figure 8.18.
The BS transmitter antenna is a uniform linear array consisting of ϑ ¼ 4 blocks

separated by 2λ, where each block is formed by 4 antennas with λ=2 spacing as shown in
Figure 8.18. For more details on the assumptions, see [17].
In order to utilize the multiplexing gain for a user group, each of the antenna blocks
generates the same “broad beam” with the same beamforming vector wL selected from
a DFT matrix. In a more general form, the beamforming matrix W ¼ w1 . . . wL1 wL is
a concatenation of multiple row vectors chosen from any codebook matrix, where each
beamforming vector wL could have different weight entries. The effective channel
HðkÞWgðkÞ of such a beam, with gðkÞ being the channel gain on frequency block k,
for a single antenna-block is illustrated in Figure 8.20, where multi-path components are
neglected to simplify the visualization. It is seen that the power is steered to user-group 1
and at the same time interference toward group 2 is mitigated.
0.9
0.8
0.7
P(Sum SE<x)
0.6
without HBF with HBF
0.5
0.4
0.3
HBF 1BS
HBF 2BS
0.2 2λ, 1BS
2λ, 2BS
0.1 λ/2, 1BS
λ/2, 2BS
0
0 5 10 15 20 25 30
Sum spectral effiency [bps/Hz]
Figure 8.21 Performance comparison of one and two BSs with and without hybrid beamforming.
Performance results from system simulations are shown in Figure 8.21. For compar-
ison, the performance results of a ULA first with λ=2 and then with 2λ antenna
spacing are shown. The optimal performance in this scenario is considered as the
interference free single BS case with one user group labeled as 1BS. The case of two
BSs and two user groups as in Figure 8.19 is labeled with 2BS. Looking at the bold lines
with one transmitter and user group, a beamforming gain from λ=2 to 2λ spacing can be
observed.
Without beamforming and interference mitigation, the sum spectral efficiency (SE)
of two transmitters is less than the single transmitter. In contrast, the sum spectral
efficiency with the beamforming in Figure 8.21 increases from 13 to 20 bps/Hz.
Dividing the sum performance by the number of transmitters results in 10 bps/Hz
per BS, which is only a loss of 3 bps/Hz compared to the optimal scenario of two
transmitters serving simultaneously the two user groups without any mutual interference.
This loss is due to the residual interference from multi-path components in the channel,
which cannot be fully eliminated by the analog beamforming projection due to its
constrained block-diagonal form.
8.6 Channel models
Massive MIMO and joint transmission have enormous potential to improve the spectral
efficiency in mobile communication systems. The realistic performance assessment of
such techniques requires having channel models that reflect the true behavior of the radio
channel [60]. Existing geometry-based stochastic channel models such as the Spatial
Channel Model (SCM), WINNER, and QuaDRiGa model allow the separation of
antenna and propagation effects and are thus ideal candidates for the evaluation of
new transmission schemes. However, many existing models lack features such as
time-evolution and full-3D propagation as well as parameter tables for many impor-
tant scenarios. In addition, the validation of such models against measured data is
still an open issue.
The physical characteristics of the radio channel are fundamentally defining the
potential and limitations for a new radio system like 5G. For theoretical analysis, the
assumptions in this respect are often quite ideal, like a fully scalable number of Tx
antennas, easily tractable Rayleigh fading channels or perfectly calibrated Tx-arrays
generating perfect beam patterns with eventually infinitely small half power beam width,
negligible multi-user or inter-cell interference, etc.
The definition of new channel models as such, ranging from RF frequencies below
6 GHz to cmW as well as mmW, can be found in Chapter 13. Furthermore, the 3GPP
extension of the spatial channel model enhanced to the so-called full dimension MIMO
channel model might be of interest [61]. Time-evolutions and full-3D propagation
modeling were also included in the QuaDRiGa channel model [62].
For mmW, extensive measurement campaigns have been and are being conducted.
The interested reader is referred to [63][64]. For the cmW, interesting results can be
found in [65] demonstrating the feasibility of strong spatial multiplexing even in the
28 GHz frequency band over few hundreds of meters.
Massive MIMO arrays in local area scenarios inside of buildings are another very
specific application, which is affected by outdoor-to-indoor as well as wall penetration
losses (see some results in [66] or [67]).
8.7 Conclusions
This chapter has covered massive MIMO as one of the clear key technology components
for 5G. While the analyzed capacity scaling behavior for large numbers of antennas and
users appears highly promising, it has become clear that massive MIMO inherits various
challenges that have to be overcome.
One key challenge is the aspect of so-called pilot contamination, i.e. the fact that
for large massive MIMO constellations with many channel components to be
estimated it is unavoidable that channel estimation becomes subject to interference,
if pilot overhead is to be kept reasonable. Multiple options to overcome pilot
contamination have been discussed, for instance based on pilot power control,
coded pilots exploiting sparse channel properties or the random usage of pilot
sequences within cells.
Further challenges are related to resource allocation, i.e. the right grouping of users to
be served by massive MIMO, and the actual transceiver design, for instance in the
context of hybrid beamforming, where part of the precoding is performed in the digital
baseband, and part in the analog domain.
References
[1] ICT-317669 METIS project, “Scenarios, requirements and KPIs for 5G mobile
and wireless system,” Deliverable D1.1, May 2013, www.metis2020.com/docu
ments/deliverables/
[2] E. G. Larsson, O. Edfors, Fr. Tufvesson, and T. L. Marzetta, “Massive MIMO
for next generation wireless systems,” IEEE Communications Magazine, vol. 52,
no. 2, pp. 186–195, 2014.
[3] 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, 2013.
[4] T. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base
station antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3590–3600,
November 2010.
[5] 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,” IEEE Signal Processing Mag., vol. 30, no. 1, pp. 40–60, January 2013.
[6] J. Hoydis, S. T. Brink, and M. Debbah, “Massive MIMO: How many antennas do
we need?,” in Annual Allerton Conference on Communication, Control and
Computing, Monticello, IL, September 2011, pp. 545–550.
[7] L. Lu, G. Li, A. 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, October 2014.
[8] M. Kountouris and N. Pappas, “HetNets and Massive MIMO: Modeling, Potential
Gains, and Performance Analysis,” in IEEE-APS Topical Conference on Antennas
and Propagation in Wireless Communications, Torino, September, 2013.
[9] J. Jose, A. Ashikhmin, T. Marzetta, and S. Vishwanath, “Pilot contamination and
precoding in multi-cell TDD systems,” IEEE Transactions on Wireless
Communications, vol. 10, no. 8, pp. 2640–2651, August 2011.
[10] B. Hassibi and B. M. Hochwald, “How much training is needed in multiple
antenna wireless links?,” IEEE Transactions on Information Theory, vol. 49,
no. 4, pp. 951–963, 2003.
[11] T. Kim and J. G. Andrews, “Optimal pilot-to-data power ratio for MIMO-OFDM,”
in IEEE Global Telecommunications Conference, St. Louis, December 2005,
pp. 1481–1485.
[12] T. Kim and J. G. Andrews, “Balancing pilot and data power for adaptive MIMO-
OFDM Systems,” in IEEE Global Telecommunications Conference, San Francisco,
December 2006.
[13] T. Marzetta, “How much training is needed for multiuser MIMO?,” in Asilomar
Conference, Pacific Grove, October 2006, pp. 359–363.
[14] N. Jindal and A. Lozano, “A unified treatment of optimum pilot overhear in
multipath fading channels,” IEEE Transactions on Communications, vol. 58,
no. 10, pp. 2939–2948, October 2010.
[15] K. T. Truong, A. Lozano, and R. W. Heath Jr., “Optimal training in continuous
block-fading massive mimo systems,” in European Wireless Conference,
Barcelona, May 2014.
[16] G. Fodor and M. Telek, “On the pilot-data power trade off in single input multiple
output systems,” in European Wireless Conference, Barcelona, May 2014,
pp. 485–492.
[17] M. Kurras, L. Thiele, and G. Caire, “Interference mitigation and multiuser multi-
plexing with beam-Steering antennas,” in International ITG Workshop on Smart
Antennas, Ilmenau, March 2015, pp. 1–5.
[18] R. Couillet and M. Debbah, Random Matrix Methods for Wireless
Communications. Cambridge, UK: Cambridge University Press, 2011.
[19] S. Wagner, R. Couillet, M. Debbah, and D. T. Slock, “Large system analysis of
linear precoding in correlated MISO broadcast channels under limited feedback,”
IEEE Trans. Inform. Theory, vol. 58, no. 7, pp. 4509–4537, 2012.
[20] D. Gesbert, S. Hanly, H. Huang, S. Shamai Shitz, O. Simeone, and W. Yu, “Multi-
cell MIMO cooperative networks: A new look at interference,” IEEE Journal on
Selected Areas in Communications, vol. 28, no. 9, pp. 1380–1408, December
2010.
[21] A. Tölli, M. Codreanu, and M. Juntti, “Cooperative MIMO-OFDM cellular system
with soft handover between distributed base station antennas,” IEEE Trans.
Wireless Commun., vol. 7, no. 4, pp. 1428–1440, April 2008.
[22] H. Asgharimoghaddam, A. Tölli, and N. Rajatheva, “Decentralizing the optimal
multi-cell beamforming via large system analysis,” in IEEE International
Conference on Communications, Sydney, June 2014, pp. 5125–5130.
[23] H. Asgharimoghaddam, A. Tölli, and N. Rajatheva, “Decentralized multi-cell
beamforming via large system analysis in correlated channels,” in European
Sign. Proc. Conf., Lisbon, September 2014, pp. 341–345.
[24] D. Tse and P. Viswanath, Fundamentals of Wireless Communication, Cambridge,
UK: Cambridge University Press, 2005.
[25] A. M. Tulino and Sergio Verdú, “Random matrix theory and wireless
communications,” Foundations and Trends in Communications and Information
Theory, vol. 1, no. 1, pp. 1–182, 2004.
[26] M. K. Ozdemir and H. Arslan, “Channel estimation for wireless OFDM systems,”
IEEE Communications Surveys and Tutorials, vol. 9, no. 2, pp. 18–48, 2007.
[27] S. Coleri, M. Ergen, A. Puri, and A. Bahai, “Channel estimation techniques based
on pilot arrangement in OFDM systems,” IEEE Trans. Broadcasting, vol. 48,
no. 3, pp. 223–229, September 2002.
[28] Y-H. Nam, Y. Akimoto, Y. Kim, M-i. Lee, K. Bhattad, and A. Ekpenyong,
“Evolution of reference signals for LTE-advanced systems,” IEEE
Communications Magazine, vol. 5, no. 2, pp. 132–138, February 2012.
[29] J. Choi, T. Kim, D. J. Love, and J-Y. Seol, “Exploiting the preferred domain of
FDD massive MIMO systems with uniform planar arrays,” in IEEE International
Conference on Communications, pp. 3068–3073, London, June 2015.
[30] B. Gopalakrishnan and N. Jindal, “An analysis of pilot contamination on multi-
user MIMO cellular systems with many antennas,” in International Workshop on
Signal Processing Advances in Wireless Communications, San Francisco, June
2011, pp. 381–385.
[31] K. Guo, Y. Guo, G. Fodor, and G. Ascheid, “Uplink power control with MMSE
Receiver in multi-cell MU-massive-MIMO systems,” in IEEE International
Conference on Communications, Sydney, June 2014, pp. 5184–5190.
[32] N. Jindal and A. Lozano, “A unified treatment of optimum pilot overhead in

multipath fading channels,” IEEE Trans. Communications, vol. 58, no. 10,
pp. 2939–2948, October 2010.
[33] G. Fodor, P. Di Marco, and M. Telek, “Performance analysis of block and comb
type channel estimation for massive MIMO systems,” in International Conference
on 5G for Ubiquitous Connectivity, Levi, Finland, November 2014, pp. 1–8.
[34] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach to channel
estimation in large-scale multiple-antenna systems,” IEEE Journal of Selected
Areas in Communications, vol. 31, no. 2, pp. 264–273, February 2013.
[35] V. Saxena, G. Fodor, and E. Karipidis, “Mitigating pilot contamination by pilot
reuse and power control schemes for massive MIMO systems,” in IEEE Vehicular
Technology Conference, Glasgow, May 2015, pp. 1–6.
[36] N. Krishnan, R. D. Yates, and N. B. Mandayam, “Uplink linear receivers for
multi-cell multiuser MIMO with pilot contamination: Large system analysis,”
IEEE Trans. Wireless Communications, vol. 13, no. 8, pp. 4360–4373, August 2014.
[37] Y. Li, Y.-H. Nam, B. L. Ng, and J. C. Zhang, “A non-asymptotic throughput for
massive MIMO cellular uplink with pilot reuse,” in IEEE Global
Telecommunications Conference, Anaheim, December 2012, pp. 4500–4504.
[38] M. Li, S. Jin, and X. Gao, “Spatial orthogonality-based pilot reuse for multi-cell
massive MIMO transmissions,” in International Conference on Wireless
Communications and Signal Processing, Hangzhou, October 2013, pp. 1–6.
[39] A. Ashikhmin and T. Marzetta, “Pilot contamination precoding in multicell large
scale antenna systems,” in IEEE International Symposium on Information Theory,
Cambridge, July 2012, pp. 1137–1141.
[40] N. Shariati, E. Björnsson, M. Bengtsson, and M. Debbah, “Low complexity
channel estimation in large scale MIMO using polynomial expansion,” in IEEE
International Symposium on Personal, Indoor and Mobile Radio Communications,
London, September 2013, pp. 1–5.
[41] J.H. Sørensen, E. de Carvalho, and P. Popovski, “Massive MIMO for crowd scenar-
ios: A solution based on random access,” in IEEE Global Telecommunications
Conference Worshops, Austin, December 2014, pp. 352–357.
[42] E. Paolini, G. Liva, and M. Chiani, “Graph-based random access for the collision
channel without feedback: Capacity bound,” in IEEE Global Telecommunications
Conference, Houston, December 2011, pp. 1–5.
[43] G. Liva, “Graph-based analysis and optimization of contention resolution diver-
sity slotted ALOHA,” IEEE Transactions on Communications, vol. 59, no. 2,
pp. 477–487, February 2011.
[44] C. Stefanovic, P. Popovski, and D. Vukobratovic, “Frameless ALOHA protocol for
wireless networks.” IEEE Communications Letters, vol. 16, no. 12, pp. 2087–2090,
December 2012.
[45] Q. Shi, M. Razaviyayan, Z.-Q. Luo, and C. He, “An iteratively weighted MMSE
approach to distributed sum utility maximization for a MIMO interfering broad-
cast channel,” IEEE Trans. Signal Processing, vol. 59, no. 9, pp. 4331–4340,
September 2011.
[46] A. Tölli, H. Pennanen, and P. Komulainen, “Decentralized minimum power
multi-cell beamforming with limited backhaul signaling,” IEEE Trans. Wireless
Commun., vol. 10, no. 2, pp. 570–580, February 2011.
[47] P. Komulainen, A. Tölli, and M. Juntti, “Effective CSI signaling and decentralized
beam coordination in TDD multi-cell MIMO systems,” IEEE Trans. Signal
Processing, vol. 61, no. 9, pp. 2204–2218, May 2013.
[48] H. Dahrouj and W. Yu, “Coordinated beamforming for the multicell multiantenna
wireless system,” IEEE Transactions on Wireless Communications, vol. 9, no. 5,
pp. 1748–1759, 2010.
[49] S. Lakshminarayana, J. Hoydis, M. Debbah, and M. Assaad, “Asymptotic analysis
of distributed multi-cell beamforming,” in IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications, Istanbul, 2010,
pp. 2105–2110.
[50] M. Kurras, L. Raschkowski, M. Talaat, and L. Thiele, “Massive SDMA with large
scale antenna systems in a multi-cell environment,” in AFRICON, Pointe-Aux-
Piments, September 2013, pp. 1–5.
[51] A. Adhikary, J. Nam, J.-Y. Ahn, and G. Caire, “Joint spatial division and multi-
plexing: The large-scale array regime,” IEEE Transactions on Information Theory,
vol. 59, no. 10, pp. 6441–6463, 2013.
[52] M. Kurras, L. Thiele, and T. Haustein, “Interference aware massive SDMA with
a large uniform rectangular antenna array,” in European Conference on Networks
and Communications (EUCNC) Bologna, 2014, pp. 1–5.
[53] D. Arthur and S. Vassilvitskii, “K-means++: The advantages of careful seeding,”
in ACM-SIAM Symposium on Discrete Algorithms, New Orleans, January 2007,
pp. 1027–1035.
[54] U. Gustavsson, C. Sanchez-Perez, T. Eriksson, F. Athley, G. Durisi, P. Landin,
K. Hausmair, C. Fager, and L. Svensson, “On the impact of hardware impairments
on massive MIMO,” in IEEE Global Telecommunications Conference Workshops,
San Diego, December 2014, pp. 294–300.
[55] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr. “Hybrid precoding for
millimeter wave cellular systems with partial channel lnowledge,” in Information
Theory and Applications, San Diego, February 2013.
[56] O. El Ayach, R. W. Heath, Jr., S. Abu-Surra, S. Rajagopal, and Z. Pi, “Low
complexity precoding for large millimeter wave MIMO systems,” in IEEE
International Conference on Communications, pp. 3724–3729, Ottawa, June
2012.
[57] A. Alkhateeb, J. Mo, N. G. Prelcic, and R. W. Heath, Jr., “MIMO precoding and
combining solutions for millimeter-wave systems,” IEEE Communications
Magazine, vol. 52, no. 12, pp. 122–131, December 2014.
[58] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Channel estimation
and hybrid precoding for millimeter wave cellular systems,” IEEE Journal
of Selected Topics in Signal Processing, vol. 8, no. 5, pp. 831, 846, October
2014.
[59] T. Obara, S. Suyama, J. Shen, and Y. Okumura, “Joint fixed beamforming and
eigenmode precoding for super high bit rate massive MIMO systems using higher
frequency bands,” in IEEE International Symposium on Personal Indoor and
Mobile Radio Communications, Washington, September 2014, pp. 1–5.
[60] International Telecommunications Union Radio (ITU-R), “Requirements related
to technical performance for IMT-Advanced radio interface(s),” Report ITU-
R M.2134, November 2008, www.itu.int/pub/R-REP-M.2134-2008.
[61] T. A. Thomas, F. W. Vook, E. Mellios, G. S. Hilton, and A. R. Nix, “3D extension

of the 3GPP/ITU channel model,” in IEEE Vehicular Technology Conference,
Dresden, June 2013, pp. 1–5.
[62] S. Jaeckel, L. Raschkowski, K. Borner, and L. Thiele, “QuaDRiGa: A 3-D
multi-cell channel model with time evolution for enabling virtual field trials,”
IEEE Transactions on Antennas and Propagation, vol. 62, pp. 3242–3256, June
2014.
[63] G. R. MacCartney and T. S. Rappaport, “73 GHz millimeter wave propagation
measurements for outdoor urban mobile and backhaul communications in
New York City,” in IEEE International Conference on Communications, Sydney,
June 2014, pp. 4862–4867.
[64] T. A. Thomas, H. C. Nguyen, G. R. MacCartney, Jr., and T. S. Rappaport, “3D
mmWave channel model proposal,” in IEEE Vehicular Technology Conference,
Vancouver, September 2014, pp. 1–6.
[65] S. Suyama, J. Shen, H. Suzuki, K. Fukawa, and Y. Okumura, “Evaluation of
30 Gbps super high bit rate mobile communications using channel data in
11 GHz band 24x24 MIMO experiment,” in IEEE International Conference on
Communications, Sydney, June 2014, pp. 5203–5208.
[66] A. O. Martınez, E. De Carvalho, and J. Ø. Nielsen, “Towards very large aperture
massive MIMO: A measurement based study,” in IEEE Global Telecommunications
Conference Workshops, Austin, December 2014, pp. 281–286.
[67] S. Dierks, W. Zirwas, B. Amin, M. Haardt, and B. Panzner, “The benefit of
cooperation in the context of massive MIMO,” in International OFDM Workshop,
Essen, August 2014.

Massive MIMO

Uploaded by

Document Informationclick to expand document informationMassive MIMO

Document Informationclick to expand document information

Copyright:

Available Formats

Massive MIMO

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Massive MIMO

Uploaded by

Copyright:

Available Formats

8 Massive multiple-input

multiple-output (MIMO) systems

8.1.1 MIMO in LTE

8.2 Theoretical background

Figure 8.2 Generic single and multiple user MIMO system.

8.2.1 Single user MIMO

y½m ¼ Hx½m þ n½m; ð8:1Þ

where x 2 ℂnt is the Tx signal, subject to TrðE½xxH Þ ≤ P, y 2 ℂnr is the Rx signal,

where Kx ¼ QPQH is the covariance matrix of x ∼ CNð0; Kx Þ, Q is a unitary steering

and the closed form solution to the integral is [24][25]

lim Cnn ðSNRÞ

Therefore, the MF receiver is the asymptotically optimal solution for nr ≫nt.

Then, the rate expression without CSIT is simpliﬁed to

As the rows of H are asymptotically orthogonal, the nr dominant singular vectors

providing nt =nr -fold array gain.

8.2.2 Multi-user MIMO

where H ¼ ½h1 ; . . . ::; hK and Kx ¼ diag ðP1 ; . . . :; PK Þ.

where AK ¼ diag ða1 ; . . . :; aK Þ. Then, the sum rate can be approximated as

8.2.2.2 Downlink channel

yk ½m ¼ hHk x½m þ nk ½m

where x 2 ℂnt is the Tx signal vector, subject to power constraint

Mode Transmission CSIT Feedback Asymptotic capacity scaling behavior

8.2.3 Capacity of massive MIMO: a summary

8.3 Pilot design for massive MIMO

recover the transmitted information. In practice, pilot signal-based data-aided techni-

8.3.1 The pilot-data trade-off and impact of CSI

8.1.1.1 Impact of channel state information errors on the throughput

ZF: Perfect CSI performance

Similar ZF and MRT performance

8.3.2 Techniques to mitigate pilot contamination

LS with max pilot power

MMSE: Maximum pilot power

LS with pilot OLPC

8.3.2.2 Coded random access in massive MIMO systems

User 1 Data Data

User 3 Data Data

User 4 Data Data

User 5 Data Data

UL time slot 1 UL time slot 2 UL time slot 3

signal in time slot ν from users applying sj , is found as

8.4 Resource allocation and transceiver algorithms for massive MIMO

In order to exploit the advantages of massive MIMO, in particular, the spatial

8.4.1 Decentralized coordinated transceiver design for massive MIMO

interference terms. The centralized problem can be decoupled by a primal or dual

8.4.1.1 System model

8.4.1.2 Performance results

½θb;k i;j ¼ ρjijj ; ð8:36Þ

8.4.2 Interference clustering and user grouping

(a) Num. of users = /2, SINR = 0 dB

beamforming-precoding concept can be either implemented in digital domain only or

• First (beamforming) stage: Inter-group interference aware clustering of co-located

8.4.2.1 Performance results

8.5 Fundamentals of baseband and RF implementations in massive MIMO

8.5.1 Basic forms of massive MIMO implementation

IDFT +CP DAC

Figure 8.12 Massive MIMO OFDM transmitter employing analog BF.

8.5.2 Hybrid ﬁxed BF with CSI-based precoding (FBCP)

8.5.2.1 Performance of FBCP