Академический Документы
Профессиональный Документы
Культура Документы
in
As we progress towards the fifth generation (5G) of wireless networks, the bit-per-joule
energy efficiency (EE) metric becomes an important design criterion because it allows for
operation at practically affordable energy consumption levels. In this regard, one of the key
technology enablers for 5G is the recently proposed massive multiple-input multiple-output
(MIMO) technology, which is a special case of multiuser MIMO with an excess of base station
(BS) antennas. However, techniques for extracting large EE gains from massive MIMO (MM)
networks have not been actively investigated so far. We seek to address the above limitation
in this thesis by (i) reviewing MM technology from an EE perspective, (ii) critically analyzing
the state-of-the-art and proposing new research directions for EE-maximization in “hybrid
MM” networks, where massive MIMO operates alongside other emerging 5G technologies,
and (iii) proposing a novel resource allocation scheme for EE-maximization in MM networks.
The thesis consists of three main parts.
In the first part, we motivate the need for EE and explain why massive MIMO is promising
as an energy-efficient technology enabler for 5G networks. In the second part, we critically
analyze opportunities for EE-maximization in three types of hybrid MM networks, namely,
millimeter wave based MM networks, MM-based heterogeneous networks, and energy har-
vesting based MM networks. We analyze limitations in the state-of-the-art and propose
several promising research directions which, if pursued, will immensely help network opera-
tors in designing hybrid MM networks.
In the third part, we propose a novel EE-maximization scheme which optimizes resource
allocation in an MM network. Three communication resources, namely, the number of BS
ii
Abstract
antennas, pilot power, and data power are optimized for EE. Since the optimization problem
is difficult to solve in its original form, we propose a novel solution approach where each
iteration solves a sequence of difference of convex programming subproblems. Simulation
results render few interesting guidelines for network designers. For example, using higher
pilot power than data power can improve the system EE, particularly when SNR is high.
Also, the number of BS antennas should be optimized with the available power budget to
ensure operation at peak EE.
iii
Preface
The following publications have resulted from the research presented in this thesis:
Statement of Authorship
I am the primary author for the publication listed above. I have been responsible to develop
original ideas, derive mathematical solutions, and generate simulation results for these publi-
cations. Prof. Vijay K. Bhargava, who is my research supervisor, provided valuable guidance
and directions in identifying the research problems, developing solution methodologies, and
documenting the results. Some of the simulation results were obtained using the disciplined
convex optimization software CVX developed by Grant, Boyd & Ye [1].
One research contribution which is not presented in the thesis, but has been published
during my MASc program at UBC is listed below.
iv
Table of Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi
1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Expectations from 5G Cellular Networks . . . . . . . . . . . . . . . . . . . . 1
1.2 Need for Energy-Efficient Systems . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Background on MIMO and Massive MIMO . . . . . . . . . . . . . . . . . . 4
1.4 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
v
Table of Contents
vi
Table of Contents
vii
Table of Contents
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Appendix
viii
List of Tables
ix
List of Figures
x
List of Figures
4.1 Simulation setup under study: assuming that the same frequency band and
the same set of pilot sequences are reused in all the cells, we investigate the
benefits of RA in the center cell. . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Proof of convergence of the proposed solution methodology . . . . . . . . . 78
4.3 EE vs SNR budget Pmax for different RA schemes when M = 50, 100. . . . 79
4.4 Pilot and data SNR vs SNR budget Pmax for different RA schemes when
M = 50, 100. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 EE vs SNR budget Pmax for different RA schemes when (M, pp , pu ) are opti-
mized. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6 Pilot SNR, data SNR, and budget utilization vs Pmax for different RA schemes
when (M, pp , pu ) are optimized. . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.7 Optimal number of BS antennas M vs SNR budget Pmax when (M, pp , pu ) are
optimized. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.8 EE vs SNR budget Pmax under different pilot reuse scenarios for K = 5. . . 85
4.9 EE vs SNR budget Pmax under different pilot reuse scenarios for K = 10. . 85
xi
Mathematical Notations
We represent matrices using boldface capital letters (e.g. A), vectors using boldface
small letters (e.g. a), and scalars using small letters (e.g. a). The transpose of a matrix
A is represented as AT . The Hermitian transpose of a matrix A is represented as AH .
An M × M identity matrix is represented as IM and sometimes I when the dimensions are
clear from the context. If a is a circularly-symmetric complex Gaussian vector with mean
µ and covariance matrix Π, we represent its probability distribution as a ∼ CN (µ, Π). A
function f in variables (x, y, z) is represented as f (x, y, z). When the variables (y, z) in f
are assigned with values (y0 , z0 ), the resulting function is represented as f (x; y0 , z0 ). E{.}
denotes expectation with respect to the random variable under context. The covariance of
a vector a is represented by cov(.). |A| and ||A|| respectively denote the determinant and
the vector 2-norm of the square matrix A.
xii
List of Abbreviations
5G : Fifth Generation
ADC : Analog to Digital Converters
BS : Base Station
CCP : Convex Concave Procedure
Co-RTDD : Co-channel Reverse Time Division Duplexing
Co-TDD : Co-channel Time Division Duplexing
CMOS : Complementary Metal-oxide-semiconductor
CQI : Channel Quality Indicator
CSI : Channel State Information
DoF : Degrees of Freedom
DPC : Dirty Paper Coding
eICIC : Enhanced Inter-cell Interference Coordination
EE : Energy Efficiency
EH : Energy Harvesting
FDD : Frequency Division Duplexing
Gbps : Giga Bits Per Second
HetNet : Heterogenous Networks
ICT : Information and Communication Technology
i.i.d : Independent and Identically Distributed
IoT : Internet of Things
LTE : Long Term Evolution
xiii
List of Abbreviations
xiv
Acknowledgements
Firstly, I would like to thank my supervisor Professor Vijay K. Bhargava for his patience,
knowledge, and generous financial support. I thank him for providing me with an excellent
research atmosphere in his lab. This thesis would not have been possible without his support,
guidance, and encouragement.
Secondly, I would like to sincerely thank Professor Lutz Lampe and Professor David
Michelson for their valuable suggestions and critical comments on the thesis. Their excellent
advice has helped me in significantly improving the quality of the thesis.
Thirdly, I am very fortunate and grateful to have excellent colleagues at the Information
Theory and Systems laboratory, who offered me genuine and friendly support and assur-
ance to carry out my research. I should particularly thank Dr. Shankanaad Mallick for
his discussions, feedback and critical suggestions. I would also like to thank Buddhika Net-
tasinghe, Lina Elmorshedy, Abdelmalik Nasser Aljalai, Reza Ramezan, and Sudha Lohani
for their friendship and support during my MASc research. I am also thankful to them for
proofreading my thesis.
Lastly, I am profoundly indebted to my parents, sister, and best friend in India for their
unconditional love and blessings. Without their continuous support and encouragement, I
could not have completed this thesis.
xv
Dedication
To my father Gopala Krishna Koppisetti, mother Surya Kumari Koppisetti, sister Pavani
Sindhura Koppisetti, and best friend Gayatri Bhavani.
xvi
Chapter 1
Motivation
The information and communication technology (ICT) industry currently connects and man-
ages billions of devices across the globe. Currently, we are in the era of 4G and 4.5G networks,
which are referred to as Long Term Evolution (LTE) and LTE-Advanced networks respec-
tively by standardization bodies. Global trends suggest that future 5G networks should
handle up to a 1000-fold increase in the current traffic demands. In addition, a wide spec-
trum of services should be supported. See Fig. 1.1 for an overview of 5G services envisioned
by Huawei Technologies Co. Ltd. [2]. As we can observe from Fig. 1.1, the Internet of
Things (IoT), which promises to connect almost everything, is expected to be an integral
part of 5G networks. A host of emerging networks, such as, smart cities, vehicular networks,
and augmented reality hubs will co-exist and operate simultaneously within 5G. In terms of
technology demands, 5G networks should support latencies ranging from 1 millisecond (ms)
to a few seconds, peak data rates up to 20 Giga bits per second (Gbps), average data rates
up to 100 Mega bits per second (Mbps), seamless connectivity for millions of IoT devices
per square kilometre, and signaling loads ranging from 1% to almost 100% [2].
When 5G networks are designed to meet such huge service expectations, energy consumption
becomes a critical concern because mobile communication networks contribute towards a
1
1.2. Need for Energy-Efficient Systems
significant stake in the global carbon footprint. Trends [3] suggest that the mobile ICT sector
would emit more than 300 million tonnes of greenhouse gases per annum by 2020. Observe
from Figure 1.2 that a majority of these emissions come from mobile access and mobile
devices, i.e., from powering the wireless communications between the base stations (BSs) and
the user equipments (UEs). Therefore, for a sustainable evolution into future 5G networks, it
becomes critically important for future wireless technologies to not only address the multifold
increase in service expectations, but also to operate at reduced power consumption levels.
A key design parameter in this regard is the bit-per-joule energy efficiency, defined as
Throughput (bits/s)
Energy Efficiency (bits/Joule) = (1.1)
Power Consumption (Joule/s)
2
1.2. Need for Energy-Efficient Systems
Figure 1.2: Trends and forecast for greenhouse gas emissions by the mobile ICT sector [2]
As we can observe from (1.1), the energy efficiency (EE) of a wireless communication system
can be increased by using methods which maximize the system throughput or minimize
power consumption, or both. The focus of this thesis is on the massive multiple-input
multiple-output (MIMO) technology, which offers higher EE levels than the current LTE
and LTE-Advanced networks.
3
1.3. Background on MIMO and Massive MIMO
4
1.3. Background on MIMO and Massive MIMO
techniques are known to increase exponentially with the size of the system (for example,
with the number of antennas at the BS) [5]-[8]. As a result, even if large throughput gains
can be achieved in an MU-MIMO system by increasing the number of antennas at the BS,
we may not incur any EE gains (c.f. (1.1)). This limitation has prompted researchers to
investigate methods which can extract large throughput gains from MU-MIMO systems at
reduced levels of power expenditure.
In this regard, a breakthrough contribution was provided by Marzetta in [9], where the
concept of massive MIMO was proposed. By definition, massive MIMO is an MU-MIMO
technology, where the BS is equipped with an excessively large number of antennas when
compared to the number of UEs in the system. Deploying a large number of antennas
at the BS results in an interesting propagation scenario, known as favourable propagation,
where the wireless channels become near-deterministic because the BS-UE channel vectors
become near-orthogonal to each other. This is in turn because the effects of small-scale
fading tend to disappear asymptotically when the number of antennas at the BS is increased
unboundedly [9]. Favorable propagation allows for some interesting design simplifications.
Firstly, large multiplexing gains can be extracted using simple linear processing techniques,
such as maximum-ratio combining (MRC) on the uplink and maximum-ratio transmission
(MRT) precoding on the downlink. Secondly, large array gains can also be extracted, thus
allowing for a substantial reduction in the radio frequency (RF) transmission power on both
the uplink and the downlink. Lastly, several operations at the BS, such as, user scheduling
and power control, can be performed over the large-scale fading time scale because the effects
of small-scale fading are averaged out under favourable propagation, thanks to the near-
deterministic channel vectors. Since massive MIMO offers large throughput gains, while also
allowing for reductions in the transmission power and the computational power, multiple
orders of EE gains can be achieved when compared to conventional MU-MIMO systems.
Based on this motivation, this thesis studies how massive MIMO technology can be used
to design energy-efficient systems for future 5G deployments.
5
1.4. Outline of the Thesis
6
1.4. Outline of the Thesis
observation leads us to identify few open research problems and to propose new research
directions for future work. Open research problems identified in the process are of critical
concern and would immensely benefit the network operators if addressed in an appropriate
manner. Also, given the number of opportunities observed, we believe that energy-efficient
hybrid massive MIMO systems are very much promising for deployment in 5G.
In Chapter 4, we solve a challenging resource allocation problem for maximizing EE of
communications in a massive MIMO system. Specifically, we investigate an EE-maximization
problem for uplink data transmissions in a massive MIMO system, where the optimization
variables are (i) the number of antennas per BS, (ii) the pilot signal power, and (iii) the
data signal power. Unlike most works on massive MIMO which assume equal pilot and
data signal powers, we treat pilot and data signal powers as separate optimization variables.
Also, unlike most studies which model circuit power as a fixed component in the total power
expenditure, we model circuit power as an increasing function in the number of BS antennas
and the number of UEs. The resulting optimization problem has a fractional objective
function and is very difficult to solve in general. To address this concern, we use principles
from fractional programming and propose an iterative algorithm based on Jagannathan’s
theorem [36], where each iteration uses an alternating optimization technique to decompose
the original problem into a sequence of solvable difference of convex (DC) programming
problems. Through simulation results, we observe that higher EE levels can be achieved
when the pilot and data signal powers are optimized separately. We also observe that higher
EE levels can be achieved if the number of antennas at the BS is optimized with respect to
the available power budget.
Finally, Chapter 5 provides few concluding remarks and topics for future work.
7
Chapter 2
2.1 Motivation
Firstly, we intend to develop necessary background on massive MIMO for appreciating var-
ious aspects of energy-efficient design discussed later in Chapters 3 and 4. Secondly, since
the EE metric relies on the power expenditure in the system, we intend to provide important
guidelines on how mathematical models can be developed for power consumption in massive
MIMO systems. This is of prime importance because inaccurate power consumption models
may invalidate theoretical designs from being implemented in practice. Lastly, we intend to
equip the reader with a critical understanding of the benefits as well as limitations associ-
8
2.3. Massive MIMO: A Multiuser MIMO Technology
ated with massive MIMO. To achieve the above mentioned objectives, we first provide an
overview of massive MIMO and develop a mathematical evaluation on how large EE gains
can be extracted from massive MIMO systems. Comparisons are drawn against conventional
multiuser MIMO systems to illustrate how signal processing requirements are significantly
simplified in the large M regime. Important design guidelines are provided on how accurate
power consumption models can be developed for massive MIMO systems. Subsequently, crit-
ical comments are provided on the practicality of massive MIMO and on some of the major
roadblocks in its acceptance as a future technology. The chapter is organized as follows.
Section 2.3 introduces massive MIMO as a multiuser MIMO technology and explains how
massive MIMO offers multiple orders of energy efficiency gains over current LTE networks.
Section 2.4 discusses few low-complexity linear detection methods which achieve near-optimal
throughput performance in massive MIMO systems. Section 2.5 provides few guidelines on
developing mathematical models for power consumption in massive MIMO systems. Section
2.6 briefs upon the practicality of massive MIMO. Section 2.7 discusses major unresolved
challenges in massive MIMO systems. Section 2.8 provides few concluding remarks.
9
2.3. Massive MIMO: A Multiuser MIMO Technology
UE 5
M
12 UE 4
UE K UE 3
UE 1 UE 2
Downlink
Uplink
Figure 2.1: Massive MIMO: a multi-user MIMO technology where K single-antenna UEs are
served by a BS with M >> K antennas.
the effects of small-scale fading tend to disappear asymptotically in the large M regime.
Significant EE gains can be achieved under favourable propagation because multiple orders
of multiplexing and array gains are realizable. For the purpose of illustration, let us consider
the uplink and downlink transmissions in a massive MIMO cell.
The asymptotic Shannon capacities on the uplink (CU L ) and the downlink (CDL ) for a
multiuser MIMO channel under favourable propagation are given by [42]
K
X
CU L = log2 (1 + pu,k M βk ),
k=1
K
(2.1)
X
CDL = max
P log2 (1 + pd,k M ak βk ),
(ak ≥0, ak ≤1)
k=1
where pu,k and pd,k are the uplink and downlink signal to noise ratios (SNRs) for the k th
UE, βk represents the large-scale fading coefficient for the k th UE, and {ak , k = 1, 2, . . . , K},
is an optimization vector to obtain CDL . For simplicity, if we neglect the effect of βk and
10
2.3. Massive MIMO: A Multiuser MIMO Technology
assume that each UE transmits with an average signal to noise ratio pu , the uplink Shannon
capacity simplifies to
CU L = K log2 (1 + M pu ). (2.2)
A similar argument can be made about downlink transmissions as well. The simplification
illustrated in (2.2) leads us to two important observations (i) the system throughput can
be improved by increasing K, i.e., by multiplexing parallel streams of data to more number
of UEs over the same time-frequency resource, and (ii) transmission power per UE can be
decreased by increasing M , i.e, the number of BS antennas, while still maintaining the same
throughput per UE. In other words, the simplification in (2.2) shows that we can achieve
O(K) multiplexing gains and O(M ) array gains under favourable propagation.
While the large array gains are a straightforward opportunity to reduce UE transmission
powers, massive MIMO also facilitates a drastic reduction in the circuit power consumed
in the system. As discussed next, this is because the BS can implement (i) linear signal
processing techniques and (ii) low-complexity user scheduling algorithms, and still achieve
near-optimal throughput performance.
In conventional multiuser MIMO systems, optimal capacities can be achieved if the BS im-
plements complex signal processing techniques, such as, maximum-likelihood (ML) multiuser
detection on the uplink and dirty paper coding (DPC) [6] on the downlink. Unfortunately,
such complex signal processing techniques incur large computational burdens which grow
exponentially with the size of the system, for example with the number of BS antennas M .
As a result, when M and K are large, such techniques consume large amounts of circuit
power, thus becoming highly unsuitable for massive MIMO operations.
Fortunately, in the large M regime, linear signal processing techniques, such as maximum-
11
2.3. Massive MIMO: A Multiuser MIMO Technology
ratio combining (MRC) on the uplink and maximum-ratio transmission (MRT) on the down-
link, can achieve near-optimal throughput performance. See Section 2.4 for a detailed dis-
cussion.
In conventional multiuser MIMO systems, simple linear precoding techniques, such as maximum-
ratio transmission (MRT), do not achieve optimal capacities on the downlink. To reduce the
performance gap, the BSs generally implement certain user scheduling methods which exploit
multi-user diversity in the system. Basically, the BS selects few UEs during each transmis-
sion interval and schedules them for simultaneous transmissions. Two seminal works on user
scheduling are the random beamforming (RBF) [15] and the semi-orthogonal user selection
(SUS) [8] methods. In the RBF method, the BS selects a group of UEs by matching them
to a pre-determined set of orthogonal beams transmitted on the downlink. The matching
is based on feedback provided by each UE, in terms of a channel quality indicator (CQI),
such as the signal to interference plus noise ratio (SINR), and the best beam index. In SUS
method, the BS acquires full channel state information (CSI) from all the candidate UEs
and selects a subset of UEs which have near-orthogonal channel vectors.
Conventional user scheduling methods, such as RBF or SUS, may not be appropriate for
massive MIMO systems due to a variety of reasons: (i) performance gains based on multi-
user diversity may not be significant in the large M regime because the effects of small-
scale fading are diminished (ii) such methods are computationally intensive and consume
significant amounts of circuit power when M is large − SUS incurs O(M 3 K) computational
complexity [8], and (iii) such methods often suffer from practical limitations − RBF schemes
do not perform well in systems with finite number of UEs [15] and SUS schemes are unscalable
because significant overhead is incurred when acquiring full CSI from all the candidate UEs.
Fortunately, in the large M regime, very simple user scheduling schemes, such as, selecting
a subset of UEs randomly [16], selecting a subset of UEs in the descending order of their
12
2.4. Linear Detection Methods for Massive MIMO
large-scale fading coefficients [16], or selecting UEs in a round-robin fashion [55], are known
to achieve near-optimal throughput performance. This is because the channel vectors become
near-orthogonal to each other and the effects of small-scale fading are diminished in the large
M regime.
Since low-complexity signal processing and user scheduling algorithms achieve near-
optimal throughput performance in massive MIMO systems, we observe that the circuit
power consumption is drastically reduced when compared to conventional multiuser MIMO
systems. Note that this reduction in circuit power consumption is in addition to the large
array gains, which allow for a significant reduction in the UE transmission powers. Conse-
quently, by achieving near-optimal throughput performance at reduced power consumption
levels, massive MIMO networks deliver multiple orders of EE gains over current LTE net-
works.
In the next section, we present a discussion on three prominent linear multiuser detection
techniques for massive MIMO, namely maximum-ratio combining (MRC), zero-forcing (ZF)
detection, and minimum mean squared error (MMSE) detection. We explain how these
linear detection techniques are derived and show that these techniques achieve near-optimal
throughput performance in the large M regime.
When linear multiuser detection techniques are used, the BS multiplies the received signal
with a linear detection matrix so as to decode the data streams transmitted by the K UEs on
the uplink. For the purpose of illustration, let us consider uplink transmissions in a massive
MIMO cell, where pu is the average transmission SNR on the uplink, s = [s1 s2 . . . sk ]T ,
such that E{|sk |2 } = 1, is the vector of symbols transmitted by the K UEs, and H =
[h1 h2 . . . hK ] is the channel matrix. The received signal vector at the BS y is given by
13
2.4. Linear Detection Methods for Massive MIMO
√
y= pu Hs + n
K (2.3)
√ X
= pu hk sk + n,
k=1
where n ∈ CM ×1 is the additive noise vector. We assume that the elements of n are in-
dependent and identically distributed Gaussian random variables with zero mean and unit
variance, i.e, n ∼ CN (0, IM ). Now, if A ∈ CM ×K is the linear detection matrix, the signal
estimate ŷ = [ŷ1 ŷ2 . . . ŷk ]T , which can be used to decode the original transmission vector
s, is given by
√
ŷ = pu AH Hs + AH n (2.4)
Based on (2.4), the signal estimate which is used to decode the symbol sk transmitted by
the k th UE is given by
inter-user interference
desired signal
z }| { noise
z√ K
}| { √ X z}|{
ŷk = pu aH
k h k s k + p u aH
j h j s j + aH
k n, (2.5)
j=1,j6=k
where ak is the k th column of A. Thereby, the received signal to interference plus noise ratio
(SINR) is given by
pu |aH
k hk |
2
SINRk = PK (2.6)
j=1,j6=k |aH 2
k hj | + ||ak ||
2
As we can observe from (2.5), columns in the detection matrix A can be obtained by solving
an optimization problem which maximizes the “desired signal power” with respect to the
“inter-user interference plus noise power” in the system. Since this can be computationally
intensive, the BS can choose to tradeoff performance by neglecting the effects of inter-user
interference or noise in the system. Based on this analogy, different multiuser detection
techniques in the current literature choose to maximize different objective functions − MRC
14
2.4. Linear Detection Methods for Massive MIMO
detection maximizes the signal to noise ratio (SNR) in the system, ZF detection maximizes
the signal to interference ratio (SIR) in the system, and MMSE detection maximizes the
signal to interference plus noise (SINR) in the system. Due to this fundamental difference in
the objective functions, the detection matrices used by these techniques are different from
each other.
When the BS employs MRC detection, it neglects the effects of inter-user interference and
maximizes the received signal to noise ratio (SNR). Therefore, using (2.5), the k th column
in the detection matrix A can be obtained as follows
AM RC = H (2.8)
Observe from (2.4) and (2.8) that MRC detection requires minimal signal processing because
the BS should simply multiply the received signal with the conjugate-transpose of the channel
matrix H and proceed to decode each symbol independently. From an EE perspective, such
a low signal processing requirement is highly desirable in the large M regime because optimal
signal processing techniques, such as ML detection, consume prohibitively large amounts of
circuit power when the size of the system is large. Note that MRC detection experiences
15
2.4. Linear Detection Methods for Massive MIMO
When the BS employs ZF detection, it neglects the effects of noise and maximizes the
received signal to interference ratio (SIR). Therefore, using (2.5), the k th column in the
detection matrix A can be obtained as follows
To satisfy the requirement in (2.9), the columns {ak } in the ZF detection matrix can be
chosen such that
ak hk 6= 0,
(2.10)
ak hj = 0, ∀j 6= k
In other words, the columns {ak } can be constructed by projecting the signal estimate onto
the orthogonal component of the inter-user interference in the system. This condition is
satisfied if A is chosen as the pseudo-inverse of the channel matrix H, i.e.,
ŷ = (HH H)−1 HH y,
(2.12)
√
= pu s + (HH H)−1 HH n
Observe from (2.12) that the resulting signal estimate is free from inter-user interference in
the system. When compared to MRC, computational complexity is higher in ZF detection
16
2.4. Linear Detection Methods for Massive MIMO
When MMSE detection is employed, the BS attempts to maximize the received signal to
interference plus noise ratio (SINR). This can be achieved by minimizing the minimum mean
squared error between the signal estimate ŷ and the transmitted symbol vector s. In other
words, the MMSE detection matrix can be obtained as follows
17
2.4. Linear Detection Methods for Massive MIMO
Fig.2.2 plots the throughput performance of MRC, ZF, and MMSE detection methods when
the number of BS antennas is increased up to 100. We assume K = 10, pu = −10 dB,
and calculate the uplink sum-rates as K
P
k=1 E{log2 (1 + SINRk )}, where SINRk values are
calculated for the three detectors by substituting the detection matrices in (2.8), (2.11),
and (2.13) into the SINR expression in (2.6). When M is large, we observe that all the
linear detection methods achieve near-optimal throughput performance. The MRC detection
method does not perform as well as the other two linear detection methods because it neglects
the effects of inter-user interference in the system. Nevertheless, the performance of MRC
is comparable to the other two methods and is, in fact, the preferred choice for practical
deployments because (i) MRC requires minimal number of computations and hence incurs
low circuit power consumption and (ii) MRC does not suffer from noise amplification when
the channels are ill-conditioned, as often experienced in practice.
18
2.5. Modelling Power Consumption in Massive MIMO
On similar lines to our discussions for linear multiuser detection methods, linear pre-
coders can also be derived for downlink transmissions in a massive MIMO system. Three
prominent linear precoders in the current literature, namely maximum-ratio transmission
(MRT), zero-forcing (ZF), and minimum mean squared error (MMSE) precoders, have sim-
ilar operational meanings and exhibit similar properties as MRC, ZF, and MMSE detection
techniques respectively. Therefore, similar to the detection matrices given in (2.8), (2.11),
and (2.13), the precoding matrices WM RC , WZF , and WM M SE for the three linear precoders
mentioned above are given by
WM RC = H,
The sum power consumption P , aggregated over uplink and downlink transmissions in an
massive MIMO system, can be modelled as
P = PP A + PC + Psys , (2.15)
19
2.6. Practicality of Massive MIMO
where PP A represents the total uplink and downlink power consumed by the power amplifiers
(PAs) at the BS and the UEs, PC represents the total uplink and downlink circuit power
consumed by different analog and digital signal processing circuits at the BS and the UEs,
and Psys refers to the remaining system dependant component in P .
While PP A accounts for the sum power expenditure on RF transmissions, PC includes the
sum power consumption from RF chain components, such as, filters, mixers, and synthesizers,
as well as baseband operations, such as, digital up/down conversion, precoding/receiver
combining, channel coding/decoding, and channel estimation. Note that PC cannot be
modelled as per conventional practice as a constant term independent of (M, K) because
the hardware requirements and the number of circuit operations in the system grow with
M and K. For example, with the one RF chain per antenna design used in the current
LTE networks, the number of RF chains at the BS and the UEs grows affinely with M
and K, respectively. Additionally, the computational requirements for various baseband
operations are functions in M and K. For example [56], O(M, K 2 ) operations are required
for ZF precoding, O(M K) for MRC detection, O(M K) for minimum mean squared error
(MMSE) channel estimation, and O(K) for channel coding respectively. Therefore, realistic
models should treat PC as a function in (M, K) and the variability of PC with (M , K)
should be investigated during the design of energy-efficient massive MIMO networks. Lastly,
Psys accounts for the power consumed by site-specific and architecture-specific factors, such
as, BS and UE architectures, power supply, cooling system, backhaul, and other control
equipment. Psys will play an important role in characterizing EE for 5G networks because
several BS and UE types will co-exist in a multi-tier architecture with different cell sizes and
power consumption levels.
20
2.7. Major Challenges in Massive MIMO
and identically distributed (i.i.d) Rayleigh channels, achieved when M is increased unbound-
edly. There is a general notion that massive MIMO may not be practical because some of
the assumptions behind favourable propagation may not be valid in practice. For example, it
may not be feasible to increase M unboundedly. Also, due to rich scattering environments,
practical channels are known to exhibit fundamentally different propagation characteristics
when compared to theoretical i.i.d Rayleigh channels. Despite these concerns, recent field
studies [51] show that measured channels with large but finite M achieve a significant por-
tion of the multiplexing and array gains derived under theoretical assumptions. This shows
that massive MIMO is indeed a practical technology.
Nevertheless, several technological challenges continue to exist. For example, designing
compact massive MIMO antenna arrays is a challenge at the current sub-3GHz bands because
a minimum inter-antenna spacing of λ/2, where λ is the carrier wavelength, is required to
avoid spatial correlation. Other important challenges include mitigating the role of pilot
contamination and reducing channel state information (CSI) overhead in frequency division
duplexing (FDD) massive MIMO systems. Detailed discussions on these challenges are
presented in the next section.
Acquiring accurate CSI at the BS is very important in massive MIMO systems because the
achievable throughput and EE gains depend directly on the accuracy of the CSI at the BS.
This is because the performance of several BS operations, such as linear detection on the
uplink and linear precoding on the downlink, is subject to the availability of accurate CSI
at the BS. Due to moving UEs, the length of coherence intervals is generally limited and
consequently, only a finite number of orthogonal pilot sequences can be used within each
coherence interval for channel estimation. This necessitates the reuse of a finite set of pilot
21
2.7. Major Challenges in Massive MIMO
Cell I P1
P1
P2
P1
P2
P2
P4 P4
P3
P4
P3
Cell II P3
Cell III
Pilot Signal
Pilot Contamination
Figure 2.3: Pilot contamination in massive MIMO: since pilot sequences are reused across
the network, channel estimates at the BSs may be inaccurate.
sequences across the network during each coherence interval. When pilots are reused, the
channel estimates obtained in a given cell will be contaminated by pilot sequences transmitted
by UEs in other cells. This effect, called “pilot contamination”, reduces the throughput and
EE performance of the system. This is illustrated in Fig. 2.3, where the pilot sequences P1
and P2 are re-used in cells I and II and the sequences P3 and P4 are reused in cells II and
III respectively. Each interfering pilot signal represents itself as the desired pilot signal at a
given BS, thus causing imperfections in the CSI estimated at the BS. CSI imperfections are
known to introduce upper bounds on the achievable throughput and EE gains in the systems
[24].
Since pilot reuse is inevitable, pilot contamination continues to be a performance bottle-
neck in massive MIMO systems. Few researchers [10]-[13] have recently proposed methods
to reduce the effects of pilot contamination in massive MIMO but much of the research on
this topic is still ongoing.
22
2.7. Major Challenges in Massive MIMO
Worldwide, the number of licenses for the FDD mode of operation (> 400) is much more than
that for the time division duplexing (TDD) mode of operation (< 30) [14]. Therefore, FDD
systems facilitate better hardware re-use, easier software upgrades, and a smoother transition
into 5G when compared to TDD systems. Despite these advantages, most research works
on massive MIMO have focused on the TDD mode of operation because TDD systems incur
lower CSI overhead and offer better scalability than FDD systems. To make the argument
rigorous, let us compare the CSI acquisition overhead in TDD and FDD based massive MIMO
systems. The BS can implement very different channel estimation techniques, depending on
whether the system operates in TDD or in FDD.
For channel estimation in the TDD mode, the K UEs transmit orthogonal pilot sequences on
the uplink. Since the pilot sequences are already known to the BS, it uses this knowledge to
estimate uplink channels based on the received pilot signals. This process requires a minimum
of K channel uses per coherence interval. Since the uplink and downlink transmissions
occur over the same frequency resource, the BS can exploit channel reciprocity to precode
transmissions for the downlink. The UEs require information on the effective channel gain
to detect desired signals on the downlink. To facilitate this, the BS can beamform pilot
signals on the downlink and the UEs can estimate the effective channel gains based on
the received pilot signals. This process also requires a minimum of K channel uses per
coherence interval. Therefore, TDD systems incur a total CSI overhead of 2K channel uses
per coherence interval. In other words, TDD systems are subject to the design constraint
2K < T , where T is the channel coherence time in symbols.
23
2.7. Major Challenges in Massive MIMO
When operating in the FDD mode, the uplink and downlink channels are not reciprocal
because transmissions on the uplink and downlink occur over different frequency bands. For
uplink transmissions, the K UEs can transmit orthogonal pilot sequences and the BS can
estimate uplink channels based on the received pilot signals. This requires K channel uses per
coherence interval on the uplink. For precoding transmissions on the downlink, the BS needs
to acquire channel state information by transmitting M orthogonal pilot sequences to the
K UEs. The UEs estimate the M downlink channels based on the received pilot signals and
feedback the channel estimates to the BS on the uplink. This process requires a minimum of
M channel uses per coherence interval on the downlink and M channel uses per coherence
interval on the uplink. Combining the channel uses required to obtain CSI for uplink and
downlink transmissions, we observe that FDD systems are subject to the constraint M +K <
T on the uplink and the constraint M < T on the downlink. This makes M + K < T as
the universal design constraint for FDD systems. Since M is generally large in massive
MIMO systems, the CSI overheads are also generally large in FDD systems. Such large
CSI acquisition overheads can adversely affect the performance of the system, particularly
in high-mobility scenarios, because the CSI overhead consumes significant portions of the
relatively smaller coherence intervals.
Also, when compared to TDD systems, scalability can be severely curtailed in FDD
systems because the CSI overhead in FDD systems increases affinely with M . We illustrate
this in Fig. 2.4 by comparing the feasible (M, K) in FDD and TDD systems. We assume a
coherence interval of T = 300 symbols, corresponding to a coherence bandwidth of 200 kHz
and a coherence time of 1.5 ms. Thereby, based on our previous discussions, TDD systems
are subject to the constraint 2K < 300 and FDD systems are subject to the constraint
M + K < 300. We observe that the feasibility regions are much smaller in FDD systems
than in TDD systems. Better scalability is offered by TDD systems because more antennas
can be added at the BS without incurring any additional CSI overhead.
24
2.7. Major Challenges in Massive MIMO
1000
900
600
500
400
FDD
300
200
100
0
0 50 100 150 200 250 300
Number of User Equipments (K)
Figure 2.4: Possible range of (M, K) values in TDD and FDD for a coherence interval of 300
symbols
As evident from the above discussions, large CSI overheads are a major bottleneck for the
design of FDD massive MIMO systems. Although few recent studies [52] propose techniques
which exploit channel sparsity to reduce the CSI overhead in FDD massive MIMO systems,
the validity of underlying channel sparsity assumptions can be questionable, particularly
in the low frequency regime. Therefore, designing FDD systems continues to be a major
research challenge for massive MIMO.
To address issues such as non-contiguous spectrum and spectrum agility in 5G networks, mas-
sive MIMO technology should be complimented with the use of appropriate non-orthogonal
waveforms. The strict synchronization and orthogonality constraints imposed by the widely-
used OFDM waveforms are being viewed as a strong hindrance towards supporting services
such as low-latency tactile internet applications and highly sporadic machine-type commu-
25
2.8. Summary
nications (MTC). Other issues with OFDM such as the high peak to average power ratio
(PAPR), sensitivity to phase noise, poor spectrum localization, and large out-of-band (OOB)
emissions [17] are being viewed as potential drawbacks. Therefore, non-orthogonal waveforms
such as filter bank multi carrier (FBMC), single carrier modulation (SCM), bi-orthogonal
frequency division multiplexing (BFDM), and universal filtered multi carrier (UFMC) are
gaining prominence [89]. Going by the state-of-the-art, there is no clear consensus on a par-
ticular waveform combination with massive MIMO because each waveform comes with its
own share of advantages and disadvantages. For example, OFDM facilitates easier hardware
implementation and backward compatibility but, as mentioned earlier, suffers from syn-
chronization and orthogonality issues. SCM supports low-latency applications with reduced
PAPR, but offers lower throughputs and requires computationally intensive equalization
techniques. Similarly, FBMC offers flexible carrier aggregation, robustness against synchro-
nization errors, and low OOB emissions, but suffers from implementation issues such as the
use of long filter lengths and complex multiuser receivers. As a result, there in no clear con-
sensus on the choice of an appropriate non-orthogonal waveform for massive MIMO. This not
only offers huge scope for further research, but also reinforces the need for standardization
activities.
2.8 Summary
26
2.8. Summary
large multiplexing and array gains can be achieved at reduced power consumption levels in
the large M regime. To support this analysis, we compared massive MIMO against conven-
tional multiuser systems and explained how massive MIMO facilitates a drastic reduction in
the computational requirements at the BS. Simple computer simulations were presented to
illustrate that low-complexity linear detection methods, such as, maximum-ratio combining
(MRC), zero-forcing (ZF), and minimum mean squared error (MMSE) detection, achieve
near-optimal throughput performance in massive MIMO systems.
Section 2.5 focused on addressing the second objective and has provided important guide-
lines on modelling power consumption in the large M regime. A simple mathematical frame-
work was developed by attributing power consumption in massive MIMO systems to three
major operations, namely, power amplifier operations, circuit operations, and site-specific
operations. Special emphasis was laid on modelling circuit power as a function in the num-
ber of antennas at the BS and the number of UEs in the system. As will be evident later in
Chapter 4, such a realistic model helps us in developing practical solutions to challenging re-
source allocation problems in massive MIMO. The third and final objective was addressed in
Section 2.7, where a detailed discussion was presented on the practicality of massive MIMO
and on some of the major challenges to be overcome before accepting massive MIMO as a
future technology. Example scenarios were presented to develop a critical perspective on
why pilot contamination and FDD mode of operation continue to be major roadblocks for
massive MIMO. Since the intended objectives have been successfully achieved, we believe
that the reader is now better equipped with the necessary background to appreciate the
studies presented in the next two chapters.
27
Chapter 3
3.1 Motivation
Due to spectrum crunch and ever increasing traffic demands, spectrum efficiency (SE) has
always been the utmost design priority in our evolution over the last two decades from 1G
to 4G networks. While the data rates have significantly improved from kilobits per second
(Kbps) in 2G to gigabits per second (Gbps) in 4G, operators have continued to overlook the
power expenditure in achieving these data rates. As a result, when compared to SE gains,
energy efficiency (EE) gains have always lagged behind by several orders of magnitude in
our evolution so far. As highlighted in the previous chapter, energy consumption trends
suggest that it is no longer a sustainable strategy to prioritize SE gains over EE gains in
our evolution towards 5G. With this motivation, the current chapter critically analyzes how
massive MIMO technology can co-operate with other 5G technologies to realize multiple
orders of EE gains over current LTE networks.
28
3.2. Background and Contributions
Several emerging technologies, such as massive MIMO, millimetre wave (mmWave), dense
heterogenous networks (HetNets), energy harvesting, full duplex, and cloud based radio ac-
cess, are currently being investigated as potential enablers for future 5G networks. Given
the uniqueness of benefits offered, it is natural to expect that a subset of these technologies
will operate in unison under the 5G architecture. Despite growing evidence of such a coex-
istence [61]-[86], limited literature is currently available on how the coexistence of multiple
5G technology enablers can be exploited to design energy-efficient 5G networks.
Particularly, there is no existing work which critically analyzes and develops a compre-
hensive understanding on how the coexistence of massive MIMO with other promising 5G
technologies can be exploited to design energy-efficient 5G networks. While few studies have
focused on designing spectrally efficient hybrid massive MIMO systems [20]-[23], a compre-
hensive study from an EE perspective is currently missing. This is an important shortcoming
from a system designer’s perspective because massive MIMO, which offers multiple orders
of EE gains by itself, can mutually benefit from one or more of the above mentioned 5G
technologies. The current chapter addresses this shortcoming by critically analyzing the
state-of-the-art on designing energy-efficient “hybrid massive MIMO” systems. We coin the
term “hybrid massive MIMO” systems to refer to wireless networks where massive MIMO
technology operates in conjunction with other emerging 5G technology enablers. The target
objective is to develop a comprehensive understanding of the opportunities and challenges
which arise when EE-maximization is attempted in hybrid massive MIMO systems. Our
contributions in this chapter are summarized next.
We take EE into perspective and critically analyze the potential of hybrid massive MIMO
systems in providing large EE gains over current LTE networks. The novelty of our work lies
in analyzing the potential of hybrid massive MIMO systems from an EE perspective. Firstly,
some of the most prominent EE-maximization techniques for massive MIMO systems, such
29
3.2. Background and Contributions
30
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
On similar lines to Section 3.4, we analyze the coexistence of heterogenous networks and
massive MIMO in Section 3.5 and the coexistence of energy harvesting and massive MIMO
in Section 3.6. Few concluding remarks are provided in Section 3.7.
MIMO Systems
Observe, from (1.1), that the energy efficiency of a massive MIMO network can be maximized
by achieving optimal throughput performance while operating at minimum levels of power
consumption. Based on this analogy, a number of research directions have been pursued for
the design of energy-efficient massive MIMO networks (see Fig 3.1 for a broad overview). Few
methods devise low-complexity algorithms for BS operations, such as, multi-user detection,
precoding, and user scheduling, so as to achieve near-optimal throughput performance at low
power expenditure. Few other methods, such as, transceiver re-design, antenna selection,
and power amplifier dimensioning, focus on improving resource utilization in the system so as
to relax hardware requirements and thereby, the power expenditure in the system. Available
literature also includes methods which minimize power losses in the system, such as, antenna
reservation and amplifier-aware design, and methods which relax hardware quality, i.e, by
introducing hardware imperfections, to reduce power expenditure in the system. In this
section, we study some of the techniques mentioned above and identify few open research
challenges thereof.
Due to favourable propagation in the large M regime, simple linear processing techniques,
such as Maximum-Ratio Combining (MRC) and Maximum-Ratio Transmission (MRT) pre-
coding, and simple user scheduling algorithms, such as random and round robin scheduling,
31
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
Figure 3.1: Overview of standard EE-maximization techniques for massive MIMO systems.
32
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
Significant energy efficiency gains can be achieved by minimizing power amplifier losses
because inefficient power amplifier operations in the current LTE networks discard as much
as 60% to 95% of the power input to the amplifiers [87]. Prominent research directions to
minimize power amplifier losses include (i) dimensioning the power amplifier, (ii) use of low
peak-to-average-power-ratio (PAPR) techniques, and (iii) PA-aware design
Losses at the power amplifiers can be minimized by operating the power amplifiers at points
close to the maximum allowed output. Unfortunately, most power amplifiers in current LTE
deployments are operated, on an average, at points much lower than the maximum allowed
output because of the high linearity requirements imposed by high peak to average power
(PAPR) waveforms such as OFDM. Power losses at linear power amplifiers can be reduced
significantly by adaptively dimensioning, i.e., adjusting, the power amplifier’s maximum
output power based on temporal variations in the traffic load in the system. Such load-
adaptive power amplifier dimensioning techniques are known to enhance the energy efficiency
of massive MIMO systems by up to 30 % [88]. A major drawback for power amplifier
dimensioning methods is that their performance is sensitive to the accuracy of available
information on temporal traffic variations in the system, which are generally very difficult to
predict.
As an alternative to linear power amplifiers, the BS can deploy non-linear power amplifiers,
which are known to be about 4-6 % more power efficient [91]. However, non-linear PAs require
low PAPR input signals to avoid signal distortions. To address this concern, researchers have
33
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
been exploring different PAPR reduction methods. For example, few antenna reservation
methods have been recently proposed [93], wherein the signals sent to one set of antennas
are deliberately clipped so as to achieve low PAPR, while correction signals are sent on the
remaining set of reserved antennas so as to compensate for the clipping. In [93], the authors
assume M = 100 and reserve 25 % of the antennas so as to achieve a PAPR reduction of
4 dB. A major drawback with antenna reservation methods is that they may not necessarily
increase the system energy efficiency because reserving antennas results in reduced throughput
and the reservation process increases signaling overhead in the system.
Few low-PAPR non-orthogonal waveforms, such as single carrier modulation (SCM), have
been proposed recently. However, designing appropriate non-orthogonal waveforms continues
to be a major research challenge because most of the recently proposed non-orthogonal
waveforms suffer from limitations, such as long filter lengths and complex receiver techniques
[89]. Linearity requirements of power amplifiers can also be relaxed using constant envelope
input signals. When appropriate precoding schemes are employed, constant envelope signals
can achieve similar throughputs as achieved by high-PAPR signals [92]. A major unresolved
challenge concerning constant envelope signal studies is the generation of perfectly constant
envelope continuous-time signals.
PA-Aware Design
A different approach to minimize power amplifier losses is the design of PA-aware mas-
sive MIMO systems. This can be beneficial because conventional massive MIMO systems,
designed with sum-power constraints per BS, do not impose constraints on the maximum
output power and the power loss per PA. Recent studies on simple MIMO systems [94] [95]
show that significant improvements in the throughput performance per UE can be achieved
when PA-aware design is implemented with realistic constraints on both the PA output power
and the power losses per PA. We observe that the design of PA-aware systems is a relatively
new research field with no literature currently available for massive MIMO systems.
34
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
Conventionally, MIMO precoding is performed digitally in the baseband. Since digital pro-
cessing requires dedicated baseband and RF chain components for each antenna element,
BS transceivers conventionally adopt a one RF chain per antenna design. Such a design
can result in significant circuit power consumption in the massive MIMO regime because
the number of RF chains at the BS increases affinely with M . Therefore, minimizing RF
chain requirements at the BS is an attractive strategy to improve energy efficiency in massive
MIMO networks. Prominent techniques which reduce RF chain requirements include hybrid
precoding, antenna selection, and transceiver redesign. Hybrid precoding techniques are
generally built on channel sparsity assumptions and are discussed in the context of mmWave
systems in Section 3.4.
Antenna Selection
35
3.3. Methods to Improve Energy Efficiency in Massive MIMO Systems
Data UE 1
RF Chain 1 1
Data UE 2
RF Chain 2
RF Switch
UE 1
Precoding
2
UE 2
RF Chain
Data UE K
N
M
UE K
Antenna
Selection
Selection
Criteria Channel
(N out of M)
Feedback
Figure 3.2: Antenna selection methods for massive MIMO: circuit power consumption is
reduced by operating with N < M RF chains at the BS.
Observe that, depending on the subset selection algorithms, such as, orthogonal match-
ing pursuit and gradient descent, antenna selection methods may add to the computational
burden at the BS. As a result, energy efficiency maximization using antenna selection be-
comes a non-trivial task. Current literature on antenna selection for massive MIMO is
mostly confined to simple single cell scenarios [96] [97]. Performance tradeoffs introduced by
design limitations, such as pilot contamination and antenna switching losses, have not been
evaluated so far.
36
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
chain transmitter based on a two-port matching network, but the transceiver performance is
subject to power losses in the matching network and mutual coupling in the antenna array.
Above mentioned examples show that, although transceiver redesign offers great promise to
improve energy efficiency in massive MIMO networks, existing works suffer from serious
practical limitations.
In the next few sections, we critically analyze state-of-the-art EE-maximization tech-
niques for hybrid massive MIMO systems where massive MIMO operates in conjunction
with other promising 5G technologies, namely, millimeter wave, heterogenous networks, and
energy harvesting networks.
MIMO Systems
The mmWave spectrum, which refers to spectrum in the 30-300 GHz band, is now being in-
vestigated for 5G operations because the current sub-3GHz bands have become overcrowded
and there is a need for additional spectrum to accommodate future traffic demands. By mov-
ing to the mmWave spectrum, significant throughput gains can be achieved because large
bandwidths of the order of multiple GHz are available (c.f. Fig. 3.3). Typically, mmWave
channels exhibit huge reflection and absorption losses and poor diffraction characteristics.
As a result, when compared to sub-3GHz bands, mmWave channels experience higher chan-
nel correlation, signal attenuation, and sensitivity to blockage. mmWave channel estimation
is still an active topic of research [62].
37
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
Figure 3.3: Large bandwidths of the order of GHz are available in the mmWave Spectrum
Massive MIMO implicitly offers the highly directional transmissions required to improve
signal strength and suppress interference in blockage-sensitive environments at mmWave
bands. On the other hand, mmWave makes massive MIMO realizable because (i) the small
wavelengths at mmWave frequencies allow a large number of antennas to be fit into very
small form factors - a 16 x 16 antenna array at 60 GHz can be fit into an 80 x 80 mm2
area [60], and (ii) the near-LOS channels in mmWave massive MIMO networks can be
estimated using direction of arrival (DoA) of the incident waves at the BS [63]. By replacing
the conventional pilot-based channel estimation techniques with such DoA-based channel
estimation techniques, the need for pilot reuse can potentially be eliminated. Thereby, the
pilot contamination effect, which is a major impediment for massive MIMO systems, may
potentially be overcome.
We observe that the current literature on the design of energy-efficient mmWave massive
MIMO systems has mostly been streamlined in few specific research directions. In this
section, we present an outline on the fundamental ideas behind prominent research directions
pursued so far. The observations we make in this section are summarized in Table 3.1 and
explained in detail here.
F
Sparsity in mmWave channels has been exploited to design hybrid analog-digital beam-
38
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
Table 3.1: Existing EE-maximization Methods for mmWave Massive MIMO Systems
Method Key idea Impact
(a) Hybrid analog-digital • Add an analog precoding phase • Reduced RF chain requirements
precoding [62] • Exploit sparsity to reduce channel
dimension
(b) Multi-stage digital • Form UE groups with similar • FDD mode becomes realizable
precoding [52] covariance eigenspace
• Exploit channel sparsity to reduce
CSI overhead
(c) Overlaid massive • Network densification • Higher throughputs
MIMO macro tier [65] • Additional assistance to mmWave • Less frequent outage
cells • Improved QoS
(d) DoA-based channel • Use DoA to estimate channels • Potentially eliminate pilot reuse
estimation [63]
(e) One-bit quantization • Use low-resolution ADC in • Reduced circuit power expenditure
[64] RF chains • Simplified circuit complexity
forming techniques which relax RF chain requirements in the system [62]. For example, the
hybrid precoding technique shown in Fig. 3.4 reduces the number of RF chains from M
to NR , where S ≤ K, S ≤ NR ≤ M . The analog precoder applies phase-only control to
reduce the channel dimensionality from M x K to NR x S. The digital precoder applies
simple linear precoding techniques on the effective NR × S channel to extract multiplexing
gains. RF chain requirements are reduced because the digital precoders operate only on the
effective reduced dimensional channel.
Sparsity in mmWave channels has also been exploited to derive low overhead multistage
precoding techniques, such as shown in Fig. 3.5. The two-stage digital precoding techniques
shown in Fig. 3.5 exploits channel sparsity to partition UEs into different groups, where each
group comprises UEs with approximately the same channel covariance eigenspace, such that
the covariance eigenspaces of different UE groups are near-orthogonal to each other. When
such multistage digital precoding techniques are implemented, the FDD mode of operation,
which incurs large CSI overhead and is therefore impractical at the sub-3GHz bands, becomes
realizable at mmWave bands because the CSI training overhead is significantly reduced. To
understand how the overhead is reduced, let us first denote r, r ≤ M , as the rank of
channel covariance matrix and S, where (S ≤ K), as the number of independent streams
39
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
RF Chain 1
Analog Precoder
Digital Precoder
Data Streams RF Chain UE 1
2
N xS
Mx N
UE 2
S
N RF Chains
UE K
RF Chain
M
Figure 3.4: Hybrid analog-digital precoding with reduced RF chain requirements for
mmWave massive MIMO systems
40
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
1 UE
M x B, B << M
Data Streams RF Chain Group 1
Precoder II
Precoder I
BxS RF Chain 2 UE
S
Group 2
M RF Chains
RF Chain UE
M
Group G
Figure 3.5: Two-stage digital precoding with reduced training overhead for mmWave massive
MIMO systems
from an energy efficiency perspective because small cells offer considerable throughput gains
through extensive spatial frequency reuse.
The near-LOS propagation in mmWave systems facilitate channel estimation based on
the direction-of-arrival (DoA) of information signals at the BS. Few research works [63]
have exploited this new opportunity in mmWave systems to design energy-efficient systems.
Improvements in EE arise from replacing conventional pilot-based, channel estimation with
DoA-based estimation schemes. Since pilot contamination is known to adversely affect the
throughput rates in a system, improvements are observed in EE when pilot reuse is eliminated
through the use of DoA for channel estimation.
The conventional one RF chain per antenna design cannot be attempted for mmWave
massive MIMO BSs because the large number of antennas and large bandwidth operations
require proportionately higher requirements, such as, the chip area to ensure sufficient sep-
aration, local oscillator distribution, and power expenditure, particularly due to PAs and
wideband analog-to-digital converters (ADCs). Particularly, wideband ADCs consume un-
acceptably large amounts of power when operated at large bandwidths. Taking this into
perspective, few research works on conventional MIMO systems [64] demonstrate that ade-
41
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
Table 3.2: Proposed Research Directions for Designing Energy-efficient mmWave Massive
MIMO Systems
Method Limitations Proposed Research Directions
(Table 3.1)
(a) • Analog phase allows phase-only control Optimize EE subject to
• Number of data streams limited • Limited precision of phase control
by number of RF chains • Limited number of phase shifts
• Limited ADC resolution
(b) • Rely on wide sense stationarity of channel • Optimize UE grouping for EE
process • Develop low-complexity
• Channel covariance is susceptible to mobility covariance tracking methods
• Pilot contamination may affect performance • Optimize pilot sequence length and pilot
(not studied so far) symbol placement to maximize EE
(c) • Over densification may incur significant • Develop cell association, interference
increase in power expenditure coordination, and load balancing.
methods to maximize EE
(d) • Estimation is very sensitive to calibration • Develop DoA-based algorithms
• Relies on computationally intensive algorithms which optionally use pilots for accuracy.
(e) • Low-precision quantization introduces • Optimize ADC input distributions
severe non-linearity and thresholding to maximize EE
• Performance loss is inevitable
quate gains may still be achievable under the one RF chain per antenna design if appropriate
low-resolution ADCs are employed. This is because circuit power consumption at the RF
chains is drastically reduced when low-resolution ADCs are employed.
In this section, we present a critical analysis on the key ideas on which most existing works on
energy-efficient mmWave massive MIMO systems are built. We identify limitations with each
approach and propose new research directions which are not straightforward and obvious to
make. If addressed in an appropriate manner, we believe that the findings in this section
will immensely help network operators in planning for future mmWave massive MIMO de-
ployments. Table 3.2 presents a summary of the major limitations which we identify for
the existing research directions shortlisted in Table 3.1. It also summarizes new research
directions which we propose for future work. Details are presented next.
Despite evidence that multi-stage precoding techniques, such as shown in Fig. 3.5, can be
designed to reduce training overhead in mmWave massive MIMO systems, such techniques
42
3.4. Millimeter Wave (mmWave)-based massive MIMO Systems
have only been studied to a limited degree of extent (see [52] for a recent example). Tradeoffs
introduced by pilot contamination should be evaluated to assess the true EE gains offered
by such multi-stage precoding techniques. Missing in the existing literature are studies
which optimize the interference mitigation parameter B for energy efficiency. Other open
problems for future work include optimizing user grouping, covariance tracking, and inter
cell interference mitigation for energy efficiency. Similar is the situation with hybrid analog-
digital precoding techniques which relax RF chain requirements at the BS. These techniques
are invaluable for mmWave operations because mixed signal components in the RF chain,
particularly the high resolution ADCs, consume large amounts of power when operated at
large bandwidths. Notice that the analog precoding phase introduces new constraints in
the transceiver design, such as limited precision for phase control, limited number of phase
shifts, and limited ADC resolution. Existing literature does not discuss the energy efficiency
tradeoffs introduced by these new constraints, leaving scope for further research.
Due to severe blockage in mmWave environments - brick and concrete can cause upto 80
dB attenuation [61], small cell mmWave networks may experience frequent signal outages,
thus enforcing service imbalance in terms of coverage, connectivity, and other quality of ser-
vice (QoS) guarantees in the network. As a result, despite overlaying a massive MIMO macro
tier, the blockage-sensitive mmWave environments may require additional assistance in the
form of multi-hop relays and repeaters so as to achieve a network-wide service guarantee.
Very limited literature is currently available on EE-maximization in two-tier mmWave mas-
sive MIMO networks. Appropriate outage probability models ( see [66] for an example), cell
association techniques, interference coordination methods, and load-balancing mechanisms
should be developed in order to evaluate the EE gains offered by such a multi-tier network
architecture.
Although DoA-based channel estimation techniques are promising to eliminate pilot con-
tamination, these techniques are generally very sensitive to calibration issues. In addi-
tion, computational burden in DoA estimation may counterweigh the expected EE gains.
43
3.5. Massive MIMO-based Heterogenous Networks
Therefore, future works should investigate hybrid channel estimation techniques, where sub-
optimal and low-complexity DoA-based estimation algorithms receive optional assistance
from additional pilot-based training so as to improve estimation accuracy. There is cur-
rently no literature available on how one-bit quantization techniques can be used to improve
EE in mmWave massive MIMO systems, although few results are available for simple MIMO
systems [64]. Associated design challenges to be addressed include the development of opti-
mal input distributions and quantizer-thresholding techniques.
Another major bottleneck in the realization of energy-efficient mmWave massive MIMO
systems is the hardware design. Silicon-based CMOS technologies provide a simple and
cost-effective means to integrate several mmWave antennas with necessary analog and digital
circuitry onto a single package. However, the high frequency and large bandwidth operations
in the mmWave regime impose constraints on the design of transceiver components. For
example, high substrate absorption losses and high noise power levels become roadblocks
to the design and integration of highly directional antennas into CMOS packages. On-chip
ADCs and power amplifiers should operate at a low supply voltage so as to avoid damaging
other active on-chip components. In addition, improper isolation between active on-chip
components can result in mutual coupling, self-jamming, and signal distortion. Transceivers
which address these design complications have not been fabricated till date.
44
3.5. Massive MIMO-based Heterogenous Networks
depending on traffic demand, and (iii) high throughput gains can be achieved by offloading
traffic between outdoor and indoor SCs. Moreover, when M SCs are deployed per unit area
γ
and γ is the path loss exponent, O(M 2 ) array gains can be achieved because the average
1
BS-to-UE distance is reduced by M 2 . These array gains are larger than the O(M ) gains
offered by massive MIMO because γ > 2 for most propagation conditions.
Due to smaller coverage areas, SCs fail to ensure seamless connectivity and quality of service
(QoS) to UEs which are highly mobile. This limitation can be overcome by designing a
two-tier massive MIMO HetNet, wherein a macro cell tier formed by the massive MIMO
BSs is overlaid with an SC tier formed by small cells, such as pico cells and femto cells. The
macro cell tier ensures uniform service coverage and supports highly mobile UEs, while the
small cell tier caters to the local indoor and outdoor capacity requirements. Clearly, such an
γ
architecture can simultaneously extract the O(M 2 ) array gains offered by HetNets and the
O(K) multiplexing gains offered by massive MIMO. In addition, since the macro tier hosts a
large number of antennas, few antennas can be reserved for low power wireless backhaul to the
SC tier. Interference coordination in massive MIMO HetNets can be analyzed by using simple
tools from random matrix theory. This is highly beneficial because tools from stochastic
geometry, which are used to study interference coordination in single antenna HetNets,
cannot be easily applied to massive MIMO HetNets because beamforming introduces cross-
tier statistical dependencies.
Researchers have begun attempting standard EE-maximization techniques for HetNets, such
as, BS sleeping and cell zooming in the context of massive MIMO-based HetNets [53] [54]
[71]. Few other energy-efficiency maximization techniques have been designed by jointly
45
3.5. Massive MIMO-based Heterogenous Networks
Table 3.4: Proposed Research Directions for Designing Energy-efficient Massive MIMO-based
HetNets
Method Limitations Proposed Research Directions
(Table 3.3)
(a) • Rely on sharing control between • Attempt EE-maximization
network tiers using sub-optimal schemes
• Optimal beamforming can be • Investigate traffic-adaptive
computationally intensive sleep modes
(b) • Covariance-based blanking relies on • Optimize sacrificed degrees of
quasi-static channels freedom (DoFs) in each tier for EE
• Very sensitive to pilot • Optimize pilot symbol placement
contamination for blanking
• Develop low-complexity
covariance tracking algorithms
(c) • Quality of interference estimation and • Dynamically switch between
rejection varies significantly co-TDD and co-RTDD
between co-TDD and co-RTDD with variations in traffic.
• Requires tight timing synchronization • Implement interference-temperature
of all devices power control
• Channel estimation suffers from • Design location-dependent scheduling
interference and pilot contamination
(d) • Network management becomes • Develop QoS-aware antenna
complicated due to partitioning and allocation to aid decentralization.
distributed control • Jointly optimize cell association
and traffic offloading
46
3.5. Massive MIMO-based Heterogenous Networks
Macrocell
UE
MM MM
Macro BS Macro BS
SC UE SC UE
Macrocell
UE SC BS SC BS
Figure 3.6: Co-channel TDD and co-channel reverse TDD deployment modes for massive
MIMO HetNets.
exploiting the properties of massive MIMO and HetNet technologies. For example, soft-cell
coordination techniques have been proposed for massive MIMO HetNets wherein BSs in
the macro and small cell tiers can jointly serve the UEs through low-complexity multiflow
beamforming. Such techniques offer high EE gains because they are known to drastically
reduce hardware requirements at the massive MIMO BSs − [72] shows that the number of
massive MIMO BS antennas can be reduced by more than 50% if a few single antenna SCs
are overlaid on the massive MIMO cell.
In addition, co-channel deployment modes, where the available spectrum is fully utilized
in both macro and SC tiers, have been recently attempted. To explain the underlying ideas,
example scenarios are illustrated in Fig. 3.6 as the co-channel TDD (co-TDD) and the co-
channel reverse TDD (co-RTDD) modes. In the co-TDD mode, the macro and SC tiers
are time-synchronized to simultaneously transmit in the uplink or the downlink. In the
co-RTDD mode, the order of uplink and downlink transmissions are reversed in one of the
tiers, i.e., macro tier operates in the downlink when SC tier operates in the uplink and vice
versa. Since the entire spectrum is utilized in both the tiers, simultaneous and uncoordinated
47
3.5. Massive MIMO-based Heterogenous Networks
transmissions can introduce significant inter-tier and intra-tier interference when co-channel
deployment modes are practiced.
Fortunately, the BSs in macro and SC tiers can not only estimate the channels to their
intended UEs, but also the covariance of interfering signals. As a result, few research works
have exploited channel reciprocity to design precoding vectors which sacrifice certain degrees
of freedom (DoFs) on the downlink so as to blank out the strongest interference subspace.
When such spatial blanking techniques are used and the number of sacrificed DoFs are opti-
mized, throughput gains can be achieved in the SC tier at the cost of a negligible throughput
loss in the macro tier [67]. Observe that spatial blanking techniques do not introduce ad-
ditional overheads into the system because no explicit cross-tier information exchange is
required. In addition, spatial blanking techniques are more resource-efficient than interfer-
ence mitigation techniques, such as almost blank subframes and fractional frequency reuse,
which are proposed under the enhanced Inter Cell Interference Coordination (eICIC) in the
current LTE standards [69]. This is because spatial blanking techniques avoid the need for or-
thogonalizing time-frequency resources by spatially blanking out the dominating interference
subspace during each time-frequency slot.
Co-channel deployment modes have also been applied within each network tier by parti-
tioning the network tier into groups of BSs and fully utilizing the available spectrum in each
BS group. For example, [70] studies a co-RTDD massive MIMO HetNet system where the
massive MIMO BSs serve the UEs and also provide wireless backhaul to few BSs in the SC
tier. co-RTDD mode is applied not only between the macro and SC tiers but also within
the SC tier by partitioning the SC tier into BS groups. Power savings are observed when
compared to single-tier massive MIMO networks as well as massive MIMO HetNets with
wired backhaul.
EE-maximization has also been attempted when co-channel deployment modes are used
in propagation environments with highly directional channel vectors, such as observed at
mmWave frequencies or when the UEs are concentrated at hotspots in certain areas (see [71]
48
3.5. Massive MIMO-based Heterogenous Networks
for an example). The key idea has been to exploit directionality of the channel vectors so as
to attempt spatial blanking based on low-overhead multi-stage precoding techniques, such as
discussed earlier in Fig 3.5. Further enhancement in EE gains have been demonstrated by
implementing these precoding techniques alongside other EE-maximization techniques for
HetNets, such as, dynamically turning on/off a few SCs, scheduling hotspots, and offloading
traffic across network tiers [71].
Lastly, EE-maximization in massive MIMO HetNets has also been attempted through
the use of appropriate user association techniques. Conventional user association techniques
based on reference signal received power (RSRP) or reference signal received quality (RSRQ)
may not be suitable for massive MIMO HetNets because these techniques do not perform well
for cells with asymmetric transmission powers, number of antennas, and load distributions.
Additionally, biasing techniques, which artificially scale the RSRP by a bias term to offload
traffic from macro-cells to SCs, may not be effective because these methods do not balance
traffic within each network tier and are generally based on average performance metrics.
Generally, the joint optimization problem of user association, precoding design, and power
allocation is known to be non-deterministic polynomial-time hard (NP-hard). Fortunately,
in the massive MIMO regime, this problem is decoupled because the asymptotic UE rates are
independent of each other’s cell association. Thanks to this simplification, studies such as
[73]−[75] propose optimal user association algorithms which achieve efficient load balancing,
both within and across network tiers.
In this section, we explain some of the limitations associated with the key ideas which are
exploited by most existing works on EE-maximization in massive MIMO HetNets. This
has allowed us to propose interesting research directions for future work. Details on the
limitations and the proposed research directions are summarized in Table 3.4 and explained
next.
49
3.5. Massive MIMO-based Heterogenous Networks
50
3.5. Massive MIMO-based Heterogenous Networks
uplink, co-RTDD can attempt spatial blanking to achieve large throughput gains in the SC
tier (see [67] for an example), while this cannot be done in co-TDD. On the other hand, when
in the macro-tier downlink, co-RTDD offers lower throughput gains than co-TDD because
co-RTDD renders lower interference rejection. This is in turn because the co-RTDD mode
can only reject interference between massive MIMO and SC BSs.
As a result, there is no clear winner among co-TDD and co-RTDD. This calls for the
design of innovative co-channel deployment modes, which can simultaneously reap the ben-
efits and overcome the limitations of co-TDD and co-RTDD. One example would be design
co-channel deployment modes which dynamically switch between co-TDD and co-RTDD
modes with varying traffic conditions. Appropriate pilot assignment methods should be
developed to contain pilot contamination, which can be particularly severe in co-channel
deployments. Since co-channel deployment modes can incur significant increase in intra-tier
and inter-tier interference, appropriate interference-temperature power control and location-
dependent scheduling algorithms should be designed. To aid decentralization in cell associa-
tion techniques for massive MIMO HetNets, future works should develop QoS-aware antenna
allocation methods. Another unexplored subject in cell association for massive MIMO Het-
Nets is to optimize cell association jointly with traffic offloading. Resource efficient inter-tier
offloading techniques based on load-adaptive cell zooming, dynamic antenna activation in
the macro tier, and mobility-aware handover policies, should be designed under practical
constraints such as limited backhaul and load asymmetries.
51
3.6. Energy Harvesting (EH)-based massive MIMO Networks
Networks
Incorporating energy harvesting capabilities at the BS and the UEs in a massive MIMO
system introduces several new design constraints into the system. For example, unlike in grid-
powered networks, the energy harvested from most renewable resources fluctuates randomly,
although over a smaller range of magnitude and a larger timescale than the communication
channel amplitudes. Also, the energy conversion efficiency is generally low - about 15%
for solar [76]. Signal interference may not be desirable from a throughput perspective,
but is desirable from an RF energy harvesting perspective. Also, power sensitivity levels
are radically different for RF energy harvesters (about −10 dBm) and information receivers
(about −60 dBm) [81]. In addition, energy harvesting networks are subject to a new causality
constraint: energy consumed until a given time cannot exceed the energy harvested until
then. These new constraints in energy harvesting networks enforce the transmission policies
to depend not only on channel fading but also on the energy arrival and data backlog.
Powering the BSs in massive MIMO networks with energy harvested from renewable re-
sources, such as solar, wind, and thermal, can result in reduced carbon footprint and in-
creased network lifetime. The UEs in a massive MIMO network can also benefit from energy
harvesting capabilities because they are usually powered by limited-capacity batteries. Un-
like at the BSs, energy harvesting rates at the UEs should be controllable because battery
drains can potentially lead to loss of network connectivity. To achieve this, the BSs can
transmit dedicated RF signals on the downlink and perform energy beamforming so as to
provide uninterrupted wireless energy transfer (WET) to the UEs. Alternatively, the BSs can
52
3.6. Energy Harvesting (EH)-based massive MIMO Networks
attempt simultaneous wireless information and power transfer (SWIPT), where the downlink
RF signals are used to simultaneously transport both energy and information to the UEs.
Massive MIMO may be suitable more suitable than omnidirectional antenna systems for such
RF energy harvesting applications because the directional transmissions in massive MIMO
systems can increase the energy transfer efficiency of RF signals. This can be beneficial
because the energy transfer efficiency of RF signals can be very low when the atmospheric
attenuation is severe [82].
Energy
Energy Supply
Energy Supply
Supply
UE
UE
UE
Energy Supply
UE
Battery
Information Transfer
Energy Transfer Renewable Power
Energy Sharing Energy Grid
Energy harvesting (EH) is a natural option to improve energy efficiency in massive MIMO
networks because it opens up several new opportunities for minimizing power consumption
from non-renewable resources, such as the power grid. Despite the well-known benefits of
incorporating EH capabilities into cellular wireless networks, limited literature is currently
53
3.6. Energy Harvesting (EH)-based massive MIMO Networks
54
3.6. Energy Harvesting (EH)-based massive MIMO Networks
Despite the well-known benefits of using a large number of antennas at the BS, limited
literature is currently available on energy harvesting massive MIMO networks. One possible
reason is that the energy transfer and conversion efficiency values can be very low [84] [85].
Even otherwise, most transmission policies available for conventional MIMO networks cannot
be generalized to the massive MIMO regime because these policies either assume zero circuit
power consumption or assume that the circuit power consumption (PC ) is a constant term
which is independent of (M, K). As discussed earlier in Section 2.5, such assumptions on
PC are not valid in the massive MIMO regime. Very few studies, such as [78], use a realistic
model for PC but make other unrealistic assumptions, such as, perfect CSI and full knowledge
of future energy arrival. This leaves scope for future work, particularly on the achievable
energy efficiency values under constraints, such as, battery imperfections, delay-sensitive
traffic, and lossy energy sharing architectures.
RF energy harvesting capabilities have also not been thoroughly investigated in massive
MIMO networks. Few studies, such as [83], study energy beamforming for WET in massive
MIMO networks under practical constraints such as imperfect CSI and delay-sensitive traffic.
However, tradeoffs introduced by other practical constraints, such as, finite battery capacity,
energy leakages, and transceiver imperfections, are yet to be fully understood. Few other
studies [86] discuss SWIPT-enabled massive MIMO networks, but several important concerns
are yet to be addressed. For example, information leakage concerns, which arise when
downlink signals are amplified to improve RF energy harvesting rates at the UEs, have not
been addressed so far. In addition, energy-aware medium access control (MAC) protocols,
which optimize time resource allocation between channel access and energy harvesting, are
yet to be designed.
Several interesting research directions can be pursued in the future to realize the en-
ergy efficiency gains offered by energy harvesting massive MIMO networks. Traffic-adaptive
transceiver activation methods should be designed to efficiently utilize the energy harvested
55
3.7. Summary
at the BS. Reliability concerns in energy harvesting massive MIMO networks should be ad-
dressed through the use of hybrid power supplies, cooperative relays, and energy cooperation
techniques. Another major research direction is to study optimal energy management poli-
cies in the presence of uncertainties in the battery state information − circuit components
are known to introduce uncertainty errors as high as 30 % [77]. In this context, partially-
observed MDP frameworks may be utilized to arrive at trade-offs between the accuracy of
battery state information and the amount of energy spent in acquiring it. Lastly, energy
harvesting devices and communication protocols should be standardized so as to simplify
network planning and management.
3.7 Summary
This chapter analyzed the state-of-the-art on methods to improve energy efficiency in mas-
sive MIMO systems. A brief discussion was initially presented on few EE-maximization
techniques which are applicable to any massive MIMO system. The limitations identified
in the process comprise both theoretical and practical issues, have been mostly longstand-
ing, and are not straightforward to address. Therefore, we believe that significant efforts
are required from both academia and industry in order to overcome the said limitations.
We also analyzed the state-of-the-art on EE-maximization techniques for “hybrid massive
MIMO systems”, where massive MIMO operates alongside other 5G technologies, namely,
millimeter wave, heterogenous networks, and energy harvesting networks.
A detailed study on how massive MIMO benefits mutually from the abovementioned
5G technologies allowed us to identify new opportunities and challenges for the design of
energy-efficient systems. Our observations from the state-of-the-art reveal that the available
literature on hybrid massive MIMO systems is streamlined in few specific directions. By
analyzing the key ideas behind these specific research directions, we could identify few open
research problems. We have categorically shortlisted a few ideas for future research, which,
56
3.7. Summary
if addressed, will immensely help network operators in realistically extracting large EE gains
from hybrid massive MIMO systems. Given the number of opportunities identified in this
chapter, we also believe that energy-efficient hybrid massive MIMO systems are very much
promising for deployment in future 5G networks.
57
Chapter 4
As with any general wireless communication system, the performance of a massive MIMO
system is subject to constraints imposed by the wireless channel, such as, limited capacity
and limited coherence interval. Therefore, the wireless channel should be taken into active
consideration when designing massive MIMO systems. In this regard, one natural design
strategy is to intelligently allocate communication resources, such as, the number of antennas,
and transmission power, based on any available channel information at the transmitter. This
chapter studies a challenging resource allocation problem in a massive MIMO system where
communication resources, namely, the pilot signal power, data signal power, and the number
of BS antennas, are optimized so as to maximize the energy efficiency (EE) of uplink data
transmissions.
58
4.1. Research Contributions
that the BSs employ maximum-ratio combining (MRC) technique for multiuser detection,
we study the problem of maximizing EE of uplink data transmissions when the optimization
variables are the number of antennas per BS, the pilot signal power, and the data signal power
and the constraints are in terms of the antenna budget and the power budget available in the
system. Unlike most studies which discuss simplistic single-cell massive MIMO systems, we
study a practical multi-cell system where pilot reuse causes pilot contamination and thereby,
results in reduced throughput rates.
To develop the EE metric (c.f. (1.1)), we propose a realistic power consumption model
which incorporates the role of circuit power as an increasing function in the number of BS
antennas and the number of UEs. Additionally, unlike most studies which assume equal
transmit power for pilot and data signals, we treat pilot and data signal powers as separate
variables and exploit this additional degree of freedom to achieve higher EE values. Incor-
porating these features makes our optimization problem unique and different from existing
works. The problem under investigation has not been studied so far.
The resulting optimization problem has a non-convex fractional objective function which
is difficult to solve in its original form, with no existing literature on possible solution method-
ologies. To address this concern, we propose a novel solution approach which uses Jagan-
nathan’s theorem [36] to first transform the fractional objective into an equivalent parametric
form and then to derive an iterative RA algorithm. In each iteration of the proposed al-
gorithm, an alternating optimization technique is used to solve the objective function by
decomposing it into a sequence of solvable difference of convex (D.C) programming sub-
problems. Through simulation results, we observe that higher EE levels can be achieved
by optimizing the pilot and data powers independently, particularly when operating in the
high SNR regime. We also observe that the number of antennas per BS should be optimized
with respect to the available power budget in order to operate at high EE levels. Pilot
contamination was observed to have an adverse effect on the achievable EE levels. Lastly,
increasing the number of UEs in the system diminished the EE improvements achieved by
59
4.2. Related Works
Several recent studies have actively investigated resource allocation for the design of energy-
efficient massive MIMO systems but most works have either ignored the dependence of circuit
power expenditure on the number of BS antennas in the system or have ignored the idea
of treating pilot and data signal powers separately. For example, the authors in [24] -[25]
study resource allocation for energy efficiency in massive MIMO systems but use unrealistic
power consumption models, where the circuit power is either completely ignored or it is
modelled as a constant term independent of (M, K). Realistic power consumption models
have been proposed in [26]−[28]. [26] studies how the system energy efficiency changes when
parameters in the power consumption model are varied. A similar analysis is performed
in [27], in addition to optimizing the system energy efficiency with respect to (w.r.t) to
the system throughput. [28] proposes energy-efficient resource allocation algorithms for
downlink massive MIMO systems. However, the studies in [26] -[28] do not consider pilot
reuse and therefore, neglect the impact of pilot contamination on the system design. Pilot
contamination in multi-cell massive MIMO systems has only been briefly investigated in [55],
where a simplified special case of a multi-cell scenario with symmetric user distribution and
propagation conditions is studied.
Lastly, although using different pilot and data powers lends an additional degree of free-
dom in maximizing the system energy efficiency, this idea has not been fully explored for the
design of energy-efficient massive MIMO systems. This is a bit surprising because current
LTE systems already employ pilot signal powers which are higher than data signal powers
by upto 6 dB so as to achieve better channel estimation accuracy [29]. The authors in [30],
[31] assume different pilot and data powers to study resource allocation in multi-cell massive
MIMO systems. Specifically, [30] investigates spectral efficiency maximization under power
60
4.3. System Model
budget constraints and [31] investigates energy efficiency maximization under per-user SINR
and per-user power constraints. However, both [30] and [31] neglect the role of circuit power
consumption in their analysis.
This chapter is organized as follows. In Section 4.3, we present the system model and
provide a mathematical derivation for the achievable uplink rates in a pilot-contaminated
multicell massive MIMO system. A realistic power consumption model for massive MIMO is
proposed in Section 4.4. Thereby, an energy efficiency maximization problem is formulated
in Section 4.5 and a solution methodology is proposed in Section 4.6. Simulation results are
presented in Section 4.7, followed by some concluding remarks in Section 4.8.
We study uplink transmissions in a massive MIMO system comprising L cells, all operating
over the same frequency band of bandwidth B Hz. Each cell has K single-antenna UEs
which are serviced by a massive MIMO BS with M >> K antennas over the same time-
frequency resource. To facilitate channel estimation at the BS, the K UEs in each cell
transmit orthogonal pilot sequences of length τ symbols per coherence interval. Therefore,
we have τ ≤ T , where T is the coherence interval in symbols. Since T is generally limited due
to moving users, we consider a scenario where the cells reuse pilot sequences, thus resulting
in pilot contamination [9]. For simplicity, we assume that all the cells reuse the same set of
pilot sequences.
Let Hli ∈ CM ×K denote the flat-fading channel matrix between the K UEs in the ith cell and
the M BS antennas in the lth cell. The column hlik in Hli corresponds to the propagation
vector for the k th UE, and can be written as
61
4.3. System Model
p
hlik = βlik qlik , (4.1)
where βlik represents the path loss and large-scale fading coefficient and qlik ∈ CM ×1 , qlik ∼
CN (0, IM ) represents the small-scale fading coefficient vector comprising i.i.d complex Gaus-
sian values. Accordingly, Hli can be written as
where Qli is the small-scale fading coefficient matrix with columns qlik , and D is the path
loss and large-scale fading coefficient matrix respectively. We assume that the βlik values
remain constant over a given coherence interval. We also assume that βlik values are perfectly
known to the lth BS.
The BS employs minimum mean-squared error (MMSE) technique to estimate the channel
matrix Hli from the uplink pilot signals. Let pp be the average transmit power per pilot
symbol. We take the noise variance to be 1, to minimize notation, but without loss of
generality. Therefore, pp can also be interpreted as the normalized pilot transmit SNR and
is therefore dimensionless. With this convention, the MMSE estimate matrix for the channel
between the k th UE in the cell i and the lth BS is given by [43]
L
X 1
ĥlik = ( hlqk + √ wlk )[D0li ]k,k ,
q=1
τ pp
L
X 1 −1 (4.3)
where, D0li = diag{d0lik }, d0lik = βlik ( βlqk + ) , k = 1 . . . K,
q=1
τ pp
wlk ∼ CN (0, IM ),
In (4.3), D0li ∈ CK×K is a normalization matrix and wlk ∈ CM ×1 represents additive white
62
4.3. System Model
Gaussian noise (AWGN). The channel estimate ĥlik follows Gaussian distribution with zero
mean and covariance σĥ2 IM , where σĥ2 is given by Proposition 1. Also, by the property
lik lik
of MMSE technique, the channel estimation errors h̃lik corresponding to the channel hlik is
given by
where h̃lik follows Gaussian distribution with zero mean and covariance σh̃2 IM and σh̃2 is
lik lik
given by Proposition 1.
Proposition 1. If ĥlij ∈ CM ×1 and h̃lij ∈ CM ×1 are respectively the MMSE estimate and
the estimation error for the channel vector between the lth BS and the j th UE in cell i, their
covariance matrices are given by
From (4.3), the following mathematical relationship between ĥlik and ĥllk can be observed
βlik
ĥlik = ĥllk . (4.6)
βllk
Note that the BS l only needs to estimate ĥllk and the role of (4.6) is only to facilitate a math-
ematical simplification in deriving the rate bound in Proposition 2. The above relationship
allows us to decompose the inter-cell interference into correlated terms, i.e., ĥlik , i = 1 . . . L,
and uncorrelated terms, i.e., ĥlij , i = 1 . . . L, j = 1 . . . K, j 6= k, with respect to the desired
user’s channel estimate ĥllk , i = 1 . . . L.
63
4.3. System Model
Once the channel estimates are available, we assume that the BS employs maximum-ratio
combining (MRC) technique to detect the data symbols transmitted by the UEs. Let pu
be the average transmit power per data symbol. We take the noise variance to be 1, to
minimize notation, but without loss of generality. Therefore, pu can also be interpreted as
the normalized data transmit SNR and is therefore dimensionless. With this convention, if
√
xi , such that pu xi ∈ CK×1 and E(xi xH
i ) = IK , is the vector of data symbols transmitted
by the K UEs in the ith cell, the signal yl received at the lth BS is given by
L
√ X
yl = pu Hli xi + nl , (4.7)
i=1
where nl ∼ CN (0, IM ) represents AWGN. When MRC technique is employed to detect the
symbol xlk transmitted by the k th UE, the BS multiplies the received signal yl with the
conjugate transpose of the corresponding channel estimate ĥllk . Post this, we obtain
L L
√ X √ X
ĥH
llk yl = pu ĥH
llk Ĥli xi + pu ĥH
llk H̃li xi + ĥH
llk nl
i=1 i=1
L K L K
√ √ X X √ XX
= pu ĥH
llk ĥllk xlk + pu ĥH
llk ĥlij xij + pu ĥH
llk h̃lij xij + ĥH
llk nl
i=1 j=1,j6=k i=1 j=1
(4.8)
The MMSE channel estimate ĥllk and the channel estimation error h̃llk are uncorrelated and
independent due to joint Gaussianity of both vectors, ∀j = 1 . . . K [43]. To obtain a lower
bound on the achievable uplink rates, we attempt the following procedure based on prior
works [44]-[47].
Let us consider the achievable uplink rate between the k th UE and the BS in cell l.
The BS knows the channel estimate ĥllk and treats it as the true channel. The BS also
knows the variance of estimation error h̃llk (c.f. (4.5)) because we assume that the BS knows
the large-scale fading coefficients, i.e, βlik . The interference-plus-noise term is modelled as
64
4.3. System Model
additive Gaussian noise independent of xlk . The channel estimation error h̃llk is also treated
as part of additive Gaussian noise (See Appendix A.2 for an illustrative example inspired
from [45] on why treating the channel estimation error as part of additive Gaussian noise
yields a lower bound on the achievable rates). Since Gaussian additive noise is the worst for
mutual information [48], we can obtain a lower bound on the achievable rate by treating the
resulting additive noise, i.e., comprising both interference-plus-noise and channel estimation
errors, as a worst case Gaussian noise with the same variance. A similar analogy has been
used in [24] and [49] to derive lower bounds on achievable rates for single-cell scenario and
a symmetric multi-cell scenario respectively.
By following the procedure discussed above, the BS classifies the received signal vector
into desired and interference-plus-noise components as follows
√
ĥH
llk yl = pu ĥH
llk ĥllk xlk +
| {z }
desired
interference + noise
z }| { (4.9)
L K L K
√ XX H √ XX H
pu ĥllk ĥlij xij + pu ĥllk h̃lij xij + ĥH
llk nl
(i,j)6= (l,k) i=1 j=1
Using (4.9), a lower bound on the uplink achievable rate for user k in the same cell is given
by [44]
τ
Rlk = B (1 − ) E{log2 (1 + γlk )},
T
pu ||ĥllk ||4
where, γlk = PL PK PL PK ,
H 2 2 2
pu i=1 j=1,(i,j)6=(l,k) |ĥllk ĥlij | + pu ||ĥllk || i=1 j=1 σh̃ + ||ĥllk ||2
lij
(4.10)
where E(.) is with respect to the channel estimates. Observe from (4.10) that we have (i)
modelled the channel estimation errors with variance σh̃2 = (1 − d0lij )βlij given by (4.5) as
lij
65
4.3. System Model
part of worst-case AWGN, and (ii) utilized the property that MMSE channel estimates and
estimation errors are independent of each other. Since the expression for achievable rate Rlk
0
in (4.10) is mathematically intractable, we obtain a lower bound Rlk on the achievable rate
Rlk as follows
0 τ 1
Rlk ≥ Rlk , B (1 − ) log2 (1 + (E{ })−1 ), (4.11)
T γlk
0
where the bound Rlk is obtained by exploiting the convexity of log(1+1/x) and using Jensen’s
inequality, i.e., E{log(1 + 1/x)} ≥ log(1 + 1/E{x}). Finally, as shown in Proposition 2, we
0
can express Rlk as a function in our design parameters (M, pp , pu ) (c.f. (4.12)). The rate
0
expression Rlk in (4.12) is obtained by simplifying (4.10) using information on the statistics
of the channel estimates and channel estimation errors (c.f. (4.5)) (see Appendix A for a
detailed derivation). The rate bound expression given in Proposition 2 can also be found in
[30], but the derivation is not provided.
Proposition 2. A lower bound on the achievable rate for uplink transmissions between the
lth BS and the k th UE in the same cell, when the BS employs MMSE channel estimation and
MRC multiuser detection, is given by
0 τ c1k (M − 1)pp pu
Rlk (M, pp , pu ) = B (1 − ) log2 (1 + ),
T c2k pu + c3k pp + c4k (M − 1)pp pu + c5k pp pu + 1
X L XL L
X
2
where, c1k , τ βllk , c2k , βlij , c3k , τ βlik ,
i=1 j=1 i=1
L
X L
X
2 2
c4k , τ βlik , c5k , c2k c3k − τ βlik
i=1,i6=l i=1
(4.12)
Observe from (4.12) that our design parameters (M, pp , pu ) are strongly coupled with each
0
other in the expression for Rlk .
66
4.4. Realistic Model for Power Consumption
The power expenditure Ptot for uplink transmissions in a massive MIMO cell can be at-
tributed to three major operations, namely, (i) power amplifier (PA) operations at the UEs,
(ii) circuit operations at the BS and the UEs, and (iii) site-specific operations, such as, BS
cooling, control, and backhaul. We derive an expression for Ptot by modeling the power
expenditure for these major operations.
If η is the power efficiency of the PAs embedded in the UEs, the average power spent by the
K UEs at the PAs for pilot and data transmissions can be given as
K
PP A = (τ pp + (T − τ )pu ). (4.13)
ηT
To model the power expenditure on circuit operations, we classify the circuit power PC into
three major components, namely, signal processing at the radio frequency (RF) transceiver
chains, channel estimation and multiuser detection. As practiced in current LTE and LTE-
Advanced networks, we assume that each antenna at the BS and the UE has a dedicated
RF chain. Also, let Ptb represent the power per RF chain at the BS, Ptu the power per RF
chain at the UE, and F the computational efficiency at the BS (in flops/Watt). Then, based
on prior works [33]-[35], the abovementioned components in PC can be modelled as shown
in Table 4.1. Thereby, PC can be accumulated from Table 4.1 as
2M KB
PC = M Ptb + KPtu + (4.14)
F
Observe that, unlike most studies which assume PC to be a constant term independent of
(M, K), we propose a realistic model where PC is an increasing function in (M, K). Such a
67
4.4. Realistic Model for Power Consumption
Table 4.1: Power Expenditure on Different Circuit Operations for Uplink Transmissions in
a Massive MIMO System
realistic model is particularly important for our study because M is an optimization variable
in our system design.
Power spent on different site-specific factors, such as, power supply, active/passive cooling,
AC/DC and DC/DC converters, switches, and backhaul operations, is accumulated into
Psite . For simplicity, we model Psite as a fixed component in Ptot .
Finally, using (4.13) and (4.14), we obtain Ptot as
K 2M KB
Ptot (M, pp , pu ) = (τ pp + (T − τ )pu ) + M Ptb + KPtu + + Psite . (4.15)
ηT F
Observe from (4.15) that Ptot is an increasing function in our design parameters (M, pp , pu ).
68
4.5. Maximizing Energy Efficiency of Uplink Data Transmissions
In the next section, we utilize the expressions derived in Section 4.3 - 4.4 to formulate
a resource allocation problem which maximizes the EE of uplink data transmissions in a
massive MIMO cell.
Transmissions
Using the expressions for achievable rates and power consumption in (4.12) and (4.15), a
lower bound on the bit-per-joule energy efficiency (EE) for uplink data transmissions in the
cell l is obtained as
PK 0
k=1Rlk (M, pp , pu )
EE(M, pp , pu ) = , (4.16)
Ptot (M, pp , pu )
We now attempt to maximize the EE, given in (4.16), by optimizing (M, pp , pu ) subject to
budget constraints on M and the transmission power per coherence interval. Mathematically,
this EE maximization problem can be expressed as P1
PK 0 (M,p ,p )
Rlk p u
maximize EE = k=1
Ptot (M,pp ,pu )
M,pp ,pu
subject to :
P1 = C1 : αK ≤ M ≤ Mmax , M ∈ R+ (4.17)
C2 : τ pp + (T − τ )pu ≤ T Pmax ,
C3 : pu ≥ 0, pp ≥ 0,
69
4.5. Maximizing Energy Efficiency of Uplink Data Transmissions
integer variable, the solution approach incurs additional complexity because the feasibility
set takes both discrete and continuous forms. We avoid this complexity by assuming M to
be a continuous variable and later, rounding off the optimal M to the nearest integer value.
The following observations can be made on P1 .
(i) The objective function has a fractional form where both the numerator and denomi-
nator are non-linear in (M, pp , pu ).
(ii) P1 is non-convex because M, pp , and pu are strongly coupled with each other in both
the numerator and the denominator of the objective function.
As a result of these features, the fractional programming problem P1 is very difficult to
solve under polynomial time complexity. Therefore, we use Jagannathan’s theorem[36] to
derive an equivalent parametric form P2 (θ), given by
maximize K 0
P
k=1 Rlk − θPtot
M,pp ,pu
P2 (θ) = (4.18)
subject to : C1 − C3 ,
PK 0
∗ ∗ k=1Rlk (M ∗ , p∗p , p∗u )
θ = EE(M , p∗p , p∗u ) = (4.19)
Ptot (M ∗ , p∗p , p∗u )
K
X
0
max Rlk (M, pp , pu ) − θ∗ Ptot (M, pp , pu )
(M,pp ,pu )∈C
k=1
K
(4.20)
X
0
= Rlk (M ∗ , p∗p , p∗u ) − θ∗ Ptot (M ∗ , p∗p , p∗u ) = 0.
k=1
70
4.6. Solution Methodology
Theorem 3 shows that we can solve P2 (θ) to obtain an optimal solution for P1 if we can
develop an intelligent strategy to find θ = θ∗ which satisfies (4.19) - (4.20). In the next
section, we use an alternating optimization technique [37] to solve P2 (θ) and an iterative
update strategy based on Dinkelbach algorithm [38] to obtain θ∗ .
71
4.6. Solution Methodology
0
where, Rlk (M ; pp0 , pu0 ) can be obtained by substituting (pp0 , pu0 ) into (4.12) as follows,
0 τ A1k M − A1k
Rlk (M ; pp0 , pu0 ) = B(1 − )log2 (1 + ),
T A2k M + A3k
where A1k , c1k pp0 pu0 ,
(4.22)
A2k , c4k pp0 pu0 ,
and Ptot (M ; pp0 , pu0 ) can be obtained by substituting (pp0 , pu0 ) into (4.15) as follows
Using the expressions (4.22) and (4.23) to simplify P3 (θ), we can make the following propo-
sition.
72
4.6. Solution Methodology
Proposition 4. The subproblem P3 (θ), given in (4.21), can be expressed as the following
difference of convex (DC) programming problem
PK
k=1 flk (M ) − glk (M, θ)
min
M ∈R+
P3 (θ) = (4.24)
subject to : C1 ,
where flk (M ) and glk (M, θ) are convex functions in M and are given by
τ
flk (M ) = −B(1 − ) log2 ((A1k + A2k )M + (A3k − A1k )),
T (4.25)
τ
glk (M, θ) = −B(1 − ) log2 (A2k M + A3k ) + θ(D1 M + D2 )
T
Proof. See Appendix A.4
K
X
min flk (M ) − ĝlk (M, θ)
M ∈R+
k=1 (4.26)
subject to: C1
73
4.6. Solution Methodology
Observe that the optimization problem in (4.26) is convex in M , because the objective and
constraint functions are convex in M . Therefore (4.26) can be solved efficiently using stan-
dard convex optimization techniques [41] and we use the convex optimization software cvx
by Grant and Boyd [1]. See [40] for a detailed proof on the convergence of CCP techniques.
4: Update iteration
n ← n + 1.
5: until Stopping criterion is satisfied
(flk (Mn ) − glk (Mn )) - (flk (Mn+1 ) − glk (Mn+1 )) <
6: return Mn
On similar lines to Algorithm 2, CCP techniques can be used to solve the DC program-
ming problems in steps 3-4 of Algorithm 1. This completes our discussion on solving P2 (θ).
74
4.6. Solution Methodology
To obtain optimal θ, i.e., θ∗ , we utilize the optimality condition in Theorem 3 and develop
an iterative update strategy based on Dinkelbach algorithm [38]. The proposed strategy is
described in Algorithm 3. During each iteration, we invoke Algorithm 1 with the (M, pp , pu )
values from the previous iteration, so as to solve P2 (θ) for the next solution set. The
(M, pp , pu ) returned by Algorithm 1 is substituted into (4.19), so as to generate a θ update
for the next iteration. This is continued until convergence is observed in the θ values. Proof
of convergence for this algorithm follows from the proof of Dinkelbach algorithm given in [38].
Observe from Theorem 3 that the θ value returned by Algorithm 3 is the energy efficiency
value.
1500
Base Station
User Equipment
1000
(with BS in the center cell as reference)
Y coordinate in meters
500
−500
−1000
Figure 4.1: Simulation setup under study: assuming that the same frequency band and the
same set of pilot sequences are reused in all the cells, we investigate the benefits of RA in
the center cell.
75
4.7. Simulation Results
Parameters Values
PA efficiency (η) 0.3
Circuit power per RF chain at the BS (Ptb ) 1W
Circuit power per RF chain at the UE (Ptu ) 0.2W
Computational efficiency at the BS F 5 ∗ 109 flops/W
Site-specific power consumption (Psite ) 5W
In this section, we present simulation results to verify if the proposed resource allocation (RA)
method can yield better EE values than conventional RA schemes. For all the simulation
results, we consider two baseline scenarios for comparison (i) “no RA”, where no resource
allocation is attempted and full power budget is always utilized, and (ii) “RA with pp = pu ,
where RA is performed but with equal pilot and data powers.
We study EE-maximization in a hexagonal cell that is uniformly surrounded by six other
hexagonal cells (c.f. Fig. 4.1). Monte Carlo simulations with random user locations and
shadowing were used to optimize the system EE with respect to our design variables. For
the simulations, we assume operation at the 2 GHz band with bandwidth B = 10 MHz. All
the cells reuse the same frequency band and each cell has a coverage radius of 500m. Unless
stated explicitly, all the cells reuse the same set of pilot sequences in order, i.e., the k th UE
in each cell reuses the same pilot sequence. The channel coherence interval is assumed to be
T = 180 symbols, corresponding to a coherence bandwidth of 180 kHz and a coherence time
of 1 ms.
Each BS is assumed to serve 5 UEs, i.e., K = 5, which are located uniformly within its
cell area such that all the UEs are at least d0 = 100m away from the serving BS. We also
assume a pilot sequence length of τ = K. The large-scale fading coefficients are obtained
as βlik = zlik /||dlik /d0 ||ν , where zlik is a log-normal random variable with standard deviation
of 8 dB, dlik is the distance between the lth BS and the k th UE in cell i, and ν = 3.8
is the path loss exponent. For the constraint C1 , we assume α = 10 and Mmax = 300.
76
4.7. Simulation Results
The convergence thresholds in Algorithms 2 and 3 are chosen as = 0.01 and δ = 0.1
respectively. Also, inspired by prior works [33] - [35], we choose simulation parameters for
the power consumption model as listed in Table 4.2. We refer to Pmax as the average transmit
SNR budget because Pmax is the average power budget available per transmission symbol
when we assume noise variance to be 1.
We begin with the proof of convergence of the proposed solution methodology by plotting
the θ values returned by the outer algorithm, i.e., Algorithm 3, in Figure 4.2. The plot
corresponds to a particular snapshot of βlij values when K = 5 and we attempt to optimize
(M, pp , pu ) assuming that the same set of pilot sequences is reused in all the cells. In the
figure, we choose Pmax ranging from −15 dB to 5 dB in equal intervals of 5 dB and plot
the corresponding θ values as the proposed solution methodology makes progress towards
convergence. Plots in the Figure 4.2 show that the outer algorithm may take up to 6
iterations to achieve convergence. By running the proposed solution methodology multiple
times for a given snapshot of βlij values, we observe that the average number of iterations
required for convergence is about 4.5. Convergence at least to a local optimum is always
guaranteed because (i) the EE metric is bounded because the achievable rates and the power
expenditure are bounded in our study, and (ii) the proposed solution methodology always
ensures that a given EE level is either maintained or improved .
Observations when (pp , pu ) are optimized for EE, given M = 50, 100
In Figure 4.3, we plot the EE values yielded by the proposed RA algorithm as a function
of the transmission power budget available, i.e., Pmax , when the number of BS antennas
are fixed at M = 50 and M = 100 respectively. The corresponding pilot signal SNRs and
data signal SNRs are plotted in Figure 4.4. We compare the performance of the proposed
algorithm against two baseline scenarios (i) no RA is attempted and full power budget is
always utilized (indicated as “no RA” in Figures 4.3 and 4.4) and (ii) RA is performed but
with equal pilot and data signal powers (indicated as RA with pp = pu in the figure). When no
77
4.7. Simulation Results
5
x 10
14
Pmax = −15 dB
Pmax = −10 dB
Pmax = −5 dB
12
Pmax = 0 dB
Pmax = 5 dB
θ value returned by outer algorithm 10
RA is performed, the EE curve attains a peak value and decreases gradually with increasing
SNR budget. This is expected because the available power budget is always fully utilized
(c.f. Figure 4.4). Consequently, power expenditure always increases with the increasing
SNR budget. However, since the achievable rates saturate in the high SNR region (because
MRC detection does not mitigate the increasing inter-user interference in the system), we
observe that the EE values attain peak values, followed by a gradual decrease when no RA
is attempted.
When RA is attempted with equal pilot and data powers, i.e., pp = pu , the available
power budget is not always fully utilized (c.f. Figure 4.4). This is because full SNR budget
utilization may decrease the EE values when the available SNR budget is larger than a
certain value. As a result, we observe that the EE curve for RA with pp = pu attains a peak
value and this value is maintained thereafter.
Observe from Fig. 4.3 that the proposed RA algorithm achieves significantly higher EE
78
4.7. Simulation Results
5
x 10
14
M = 50, Proposed RA
M = 50, RA with pp = pu
M = 50, No RA
12
M = 100, Proposed RA
M = 100, RA with pp = pu
M = 100, No RA
10
Energy Efficiency (bits/Joule)
W1
8
W2
6
4
TH2
TH1
0
−30 −20 −10 0 10 20 30
Transmit SNR Budget Pmax (dB)
Figure 4.3: EE vs SNR budget Pmax for different RA schemes when M = 50, 100.
25
Pilot SNR,
M = 50, Proposed RA
20 Data SNR, W1
M = 50, Proposed RA
15 Pilot SNR,
M = 100, Proposed RA
Data SNR,
Pilot and Data SNR Allocated (dB)
10 M = 100, Proposed RA
Pilot and Data SNR,
M = 50, RA with pp = pu
5
Pilot and Data SNR,
M = 100, RA with pp = pu
0
M = 50, No RA
W2
−5
−10
TH1 TH2
−15
−20
−25
−30 −20 −10 0 10 20 30
Transmit SNR Budget P (dB)
max
Figure 4.4: Pilot and data SNR vs SNR budget Pmax for different RA schemes when M =
50, 100.
79
4.7. Simulation Results
values when compared to the other two baseline scenarios, particularly when the available
SNR budget is larger than 0 dB. This can be attributed to two important factors which are
evident from Figure 4.4. Firstly, the proposed RA algorithm does not always assign pp = pu .
Particularly when Pmax is larger than a certain threshold (labeled as T H1 and T H2 for
M = 50 and 100 in Figures 4.3 and 4.4), pp is increased while adjusting pu such that the
constraint C1 is still maintained. This results in increased EE levels because the accuracy
of M M SE estimation is increased (c.f. (4.3), (4.5)), thus enforcing a relative increase in
P 0
the achievable rates k Rlk w.r.t the power expenditure Ptot . The proposed RA algorithm
continues to increase pp over a budget window (labeled as W 1 and W 2 for M = 50 and
100 in Figures 4.3 and 4.4), until an improvement in the MMSE estimation accuracy can
no longer improve the EE values. Secondly, beyond the highlighted budget windows, i.e.,
when Pmax > (T H1 + W 1) for M = 50 and Pmax > (T H2 + W 2) for M = 100, the pp and
pu values attain saturation and any additional transmit SNR budget is not utilized because
doing so would reduce the system EE.
In Figures 4.5-4.7, we plot the performance of the proposed RA algorithm against two
baseline scenarios when the available SNR budget Pmax is increased from −25 dB to 25
dB. Similar to our previous discussions, the baseline “no RA” refers to the scenario where no
RA is performed and the available power budget is always fully utilized. The baseline “RA
with pp = pu ” refers to the scenario where RA is performed but with equal pilot and data
transmit SNRs. However, unlike in the previous discussions, we assume that the baseline
scenario “no RA” operates with the same number of BS antennas as used by the proposed
RA algorithm. That is, for each SNR budget value, the proposed RA algorithm yields an
optimal M value and the “no RA” scheme operates with the same M value. This is done
so as to allow for a fair comparison against the proposed RA algorithm. When this is not
done, i.e., when the “no RA” scenario is not configured to use the same M as used by
80
4.7. Simulation Results
5
x 10
14
Proposed RA
RA with pp = pu
12 No RA
0
−30 −20 −10 0 10 20 30
Transmit SNR Budget (dB)
Figure 4.5: EE vs SNR budget Pmax for different RA schemes when (M, pp , pu ) are optimized.
the proposed RA, the EE levels achieved by “no RA” will be drastically lower than the
EE levels illustrated in Figure 4.5. This is because the “no RA” scenario always operates
with the maximum number of antennas, i.e., Mmax , thus incurring significantly higher power
expenditure, particularly when optimal M is considerably lower than Mmax .
Figure 4.5 provides a comparison of the EE levels achieved by the proposed RA algorithm
against the two baseline scenarios discussed above. The corresponding pilot and data signal
SNRs are plotted in Figure 4.6. In Figure 4.7, we plot the M values at which the proposed
RA algorithm and the two baseline scenarios operate.
Observe from Figure 4.5 that the proposed RA algorithm always performs at least as
well as the two baseline scenarios. When in the low SNR region, the proposed RA algorithm
performs equally well as the scenario with pp = pu . As we can observe from Figure 4.6, this
is because the proposed RA algorithm finds it optimal to operate with equal pilot and data
SNRs when the total available SNR budget is low. However, when the available SNR budget
is sufficiently large, say beyond 0 dB, we observe that the proposed RA algorithm achieves
81
4.7. Simulation Results
30
Pilot SNR, Proposed RA
Data SNR, Proposed RA
SNR budget utilized, Proposed RA
Pilot SNR, Data SNR, and
20 SNR budget utilized, RA with pp = pu
Power Allocation and Budget Utilized (dB) Pilot SNR, Data SNR, and
SNR budget utilized, No RA
10
−10
−20
−30
−25 −20 −15 −10 −5 0 5 10 15 20 25
Transmit SNR Budget Pmax (dB)
Figure 4.6: Pilot SNR, data SNR, and budget utilization vs Pmax for different RA schemes
when (M, pp , pu ) are optimized.
significantly higher EE levels than the other two baseline scenarios. Also, for any given SNR
budget, note that the EE levels in Figure 4.5 are always higher than or equal to the EE
levels observed in Figure 4.3 because unlike in the case of Figure 4.3, where M is assigned
predetermined values of 50 and 100, Figure 4.5 plots the scenario where M is also optimized
for EE.
As we can observe from Figure 4.6, the EE improvements achieved by the proposed RA
algorithm may be attributed to two major factors, namely, (i) the proposed RA algorithm
optimizes pp and pu separately and achieves high EE values by using higher pp than pu so as
to improve the MMSE channel estimation accuracy, and (ii) the proposed RA algorithm does
not fully utilize the available power budget after it attains a peak EE value because it realizes
that utilizing any additional SNR budget would result in reduced EE levels. Also observe
from Fig. 4.6 that the proposed RA algorithm does not exceed the total available power
82
4.7. Simulation Results
300
Proposed RA
RA with equal pilot and data powers
250
Number of BS Antennas M
200
150
100
50
0
−30 −20 −10 0 10 20 30
Transmit SNR Budget Pmax (dB)
Figure 4.7: Optimal number of BS antennas M vs SNR budget Pmax when (M, pp , pu ) are
optimized.
budget even if it employs a significantly higher pp than pu when in the high SNR regime.
This is because the UEs only use τ symbols for pilot transmissions and the remaining (T −τ )
symbols are used for data transmissions. Since τ = K = 5 and T = 180, we observe that the
total utilized SNR budget is much lower than the total available budget even if the proposed
RA algorithm increases pp beyond pu by about 8 dB so as to achieve high EE levels.
Observe from Figure 4.7 that the proposed RA algorithm and the baseline scenario “
RA with pp = pu ” yield the same optimal M values. Also, the proposed RA algorithm
does not always fully utilize all the available BS antennas, i.e., M is not necessarily equal to
Mmax . Observe from (4.15) that this is mainly because the power expenditure in the system
increases with M . As the power budget is increased, the proposed RA algorithm reduces
M so as to ensure that there is no relative decrease in the achievable rates with respect to
power expenditure in the system. This figure underlines the importance of modelling PC as
a function in M . If PC was modelled as a constant quantity which is independent of M , we
would have always operated with M = Mmax antennas and thus at significantly lower EE
83
4.7. Simulation Results
levels.
In Figure 4.8, we plot EE values yielded for the center cell by the proposed RA algorithm
against the available SNR budget, i.e., Pmax , when different pilot reuse patterns are used in
the system. The “no pilot reuse”, “pilot reuse in 4 cells”, and “pilot reuse in 7 cells” scenarios
highlighted in the figure correspond to the scenarios where the same set of pilot sequences
are reused in 1, 4, and 7 cells in the system respectively. The EE values are obtained by
optimizing all the three design variables, i.e., (M, pp , pu ), for EE. Comparisons are drawn
against the same baseline scenarios as discussed in the previous subsection. As pilot reuse is
increased, we observe that the achievable EE levels become lower. This is expected because
an increase in pilot contamination adversely affects the achievable uplink rates, but has no
additional impact on the power expenditure in the system. The degradation in EE levels is
not considerably large, possibly because we do not consider a large number of UEs.
In Figure 4.9, we consider a scenario where K = 10, with the same constraints on the antenna
budget, i.e., α = 10 and Mmax = 300. When compared to our previous discussions, we have
increased the number of UEs K from 5 to 10. When compared to Figure 4.8, where K = 5,
we observe that the peak EE values in Figure 4.9 are lower even if the minimum number
of antennas M is now 100. This is because the increase in system throughput due to the
increase in M and K is counterweighted by (i) the increase in circuit power expenditure and
(ii) the decrease in number of symbols available for data transmissions. When the number
of UEs is increased, we also observe that the relative improvement in EE due to the use of
different pilot and data powers has diminished. This may be attributed to the increase in
pilot sequence length τ because the proposed RA algorithm has to operate within the total
SNR budget available and therefore, cannot increase pp as much as it could in the case of
84
4.7. Simulation Results
5
x 10
14
Proposed RA, no pilot reuse
Proposed RA, pilot reuse in 4 cells
Proposed RA, pilot reuse in 7 cells
RA with pp = pu, no pilot reuse
12
RA with pp = pu, pilot reuse in 4 cells
RA with pp = pu, pilot reuse in 7 cells
0
−25 −20 −15 −10 −5 0 5 10 15 20 25
Transmit SNR Budget (dB)
Figure 4.8: EE vs SNR budget Pmax under different pilot reuse scenarios for K = 5.
5
x 10
14
Proposed RA,
no pilot reuse
Proposed RA,
12 pilot reuse in 4 cells
Proposed RA,
pilot reuse in 7 cells
RA with pp = pu,
10
Energy Efficiency (bits/Joule)
no pilot reuse
RA with pp = pu,
pilot reuse in 4 cells
8
RA with pp = pu,
pilot reuse in 7 cells
No RA, no pilot reuse
6
No RA,
pilot reuse in 4 cells
No RA,
4 pilot reuse in 7 cells
0
−25 −20 −15 −10 −5 0 5 10 15 20 25
Trasnmit SNR Budget Pmax (dB)
Figure 4.9: EE vs SNR budget Pmax under different pilot reuse scenarios for K = 10.
85
4.8. Summary
4.8 Summary
In this chapter, we studied a challenging resource allocation problem for energy efficiency
(EE) maximization of uplink data transmissions in a pilot-contaminated multicell massive
MIMO system with MRC detectors. To model the bit-per-joule EE, mathematical ex-
pressions were derived for the achievable rates and the power consumption in the system.
Thereby, an EE-maximization problem was formulated, wherein the objective function is
non-convex in our design variables, namely, the number of antennas per Base Station, the
pilot signal power, and the data signal power. Since the optimization problem under inves-
tigation is very difficult to solve in its original form, principles from fractional programming
were used to transform it into a weighted-sum problem and an iterative resource allocation
algorithm was proposed, wherein an alternating optimization technique was used in each
iteration to decompose the problem into a sequence of solvable difference of convex (D.C.)
programming subproblems.
Simulation results reveal that higher EE levels can indeed be achieved by optimizing the
pilot and data powers as separate variables, particularly when operating in the high SNR
regime. Since we used a realistic power consumption model, where circuit power expenditure
is an affine function in the number of BS antennas, we found that the optimal number of BS
antennas decreases with increasing SNR budget. This observation shows that the number
of BS antennas should be dynamically activated or deactivated depending on the available
power budget so as to achieve sustained operation at high EE levels. The effect of pilot con-
86
4.8. Summary
tamination was also investigated and it was observed that pilot reuse had an adverse impact
on the achievable EE levels. Lastly, we also observed that the EE improvements obtained
from treating pilot and data signal powers as separate optimization variables diminished
when the number of UEs in the system were increased.
87
Chapter 5
As the amount of traffic demand in the wireless industry increases over the next two decades,
the joules-per-bit expenditure should be cut down to practically affordable values. Therefore,
the bit-per-joule energy efficiency (EE) metric is emerging as a critical design criterion for
the next generation of wireless networks. One of the key technology enabler in this regard is
the recently proposed massive multiple-input multiple-output (MIMO) technology because
it promises multiple orders of EE gains over current LTE networks. However, techniques
for extracting large EE gains from massive MIMO (MM) networks have not been actively
investigated in the existing literature. Here, we have addressed this limitation by (i) review-
ing MM technology from an EE perspective, (ii) critically analyzing the state-of-the-art and
proposing new research directions for EE-maximization in “hybrid MM” networks, where
MM technology operates alongside other emerging 5G technologies, and (iii) proposing a
novel resource allocation scheme for EE-maximization in MM networks.
In Chapter 2, we reviewed various aspects of MM technology to develop an EE perspec-
tive. Available literature on this topic was summarized and put together in an organized
manner to help the reader in developing a critical perspective. As part of the review, we
proposed a simple but realistic power consumption model for MM networks based on three
major operations, namely, power amplifier, circuit, and site-specific operations. Simplicity
of the proposed model facilitates ease of understanding and allows for reproducibility.
In Chapter 3, we critically analyzed the state-of-the-art to identify several promising
research directions for EE-maximization in hybrid MM networks. Three prominent cate-
gories of hybrid MM networks were analyzed, namely millimeter wave based MM networks,
88
Chapter 5. Conclusions and Future Work
89
Scope for Future Work
Firstly, a resource allocation problem similar to the one discussed in Chapter 4 can be
conducted to maximize energy efficiency in massive MIMO systems with ZF and MMSE
detectors. This is a non-trivial task because the achievable uplink rates for different UEs
under imperfect CSI and pilot contamination do not decouple as observed for MRC detectors
in (4.12). In addition, although ZF and MMSE detection methods offer better throughput
rates than MRC in interference-limited systems, these methods also incur higher compu-
tational complexities. As a result of these issues, we expect that the resource allocation
results for ZF and MMSE detectors will be different from those obtained for MRC detectors.
Investigations in Chapter 4 may also be extended to downlink transmissions to study joint
uplink-downlink resource allocation.
Secondly, observe that the power consumption model in Chapter 4 assumes that the
backhaul operations consume a fixed amount of power per coherence interval. This may
90
Scope for Future Work
not be valid in practice because the power expenditure on backhaul operations depends on
the transmission rates. Future studies may consider this issue while investigating resource
allocation for energy efficiency.
Thirdly, novel precoding schemes can be developed for FDD massive MIMO systems by
combining the hybrid analog-digital precoding and two-stage digital precoding techniques
discussed earlier in Section 3.4. Such methods not only reduce the CSI overhead, but also re-
duce the RF chain requirements in the system. Therefore, designing such precoding schemes
can lead to promising results for FDD massive MIMO systems.
Lastly, few other topics open for future work can be found in Sections 3.4.4, 3.5.4, and
3.6.4.
91
Bibliography
[1] M. Grant and S.Boyd. (2014). CVX: Matlab Software for Disciplined Convex Program-
ming (version 2.1) [Online]. Available: http://cvxr.com/cvx.
[3] Ericsson. (2014). Ericsson energy and carbon report [Online]. Available:
www.ericsson.com/res/docs/2014/ericsson-energy-and-carbon-report.pdf.
[6] S. Vishwanath, N. Jindal, and A. Goldsmith, “Duality, achievable rates, and sum-rate
capacity of Gaussian MIMO broadcast channels,” IEEE Trans. Inf. Theory, vol. 49, no.
10, pp. 2658-2668, Oct. 2003.
[7] J. Chung et al., “A random beamforming technique in MIMO systems exploiting mul-
tiuser diversity,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp. 848-855, Jun. 2003.
[8] T. Yoo and A. Goldsmith, “On the optimality of multiantenna broadcast scheduling
using zero-forcing beamforming,” IEEE J. Sel. Areas Commun., vol. 24, no. 3, pp.
528-541, Mar. 2006.
92
Bibliography
[9] T. L. Marzetta, “Noncooperative cellular wireless with unlimited number of base station
antennas,” IEEE Trans. Wireless Commun., vol. 9, no. 11, pp. 3590-3600, Nov. 2010.
[10] H. Q. Ngo and E. G. Larsson, “EVD-based channel estimations for multicell multiuser
MIMO with very large antenna arrays,” in IEEE Proc. Int. Conf. on Acoust., Speech
and Signal Process., Kyoto, Japan, 2012.
[12] A. Ashikhmin and T. L. Marzetta, “Pilot contamination precoding in multi- cell large
scale antenna systems,” in IEEE Proc. Int. Symp. Inf. Theory, Cambridge, MA, Jul.
2012.
[15] M. Sharif and B. Hassibi, “On the capacity of MIMO broadcast channels with partial
side information,” IEEE Trans. Inf. Theory, vol. 51, no. 2, pp. 506 -522, Feb. 2005.
[16] H. Liu et al., “Low-complexity downlink user selection for massive MIMO systems,”
IEEE Syst. J., vol. 85, no. 3, pp. 1,12, May 2015.
[17] J. Shen et al., “Requirements of power amplifier on super high bit rate massive MIMO
OFDM transmission using higher frequency bands,” in IEEE GC Wkshps, Austin. TX,
2014, pp. 433-437.
[18] A. Chockalingam and B. Sundar Rajan, Large MIMO systems. Cambridge University
Press, 2014.
93
Bibliography
[19] Massive MIMO Info Point. (2015). Massive (very large) MIMO systems [Online]. Avail-
able: http://www.massivemimo.eu.
[20] L. Lu, et. al., “An overview of massive MIMO: benefits and challenges,” IEEE J. Sel.
Topics Signal Process., vol. 8, no. 5, pp. 742-758, Oct. 2014.
[21] E. Larsson, et. al., “Massive MIMO for next generation wireless systems,” IEEE Com-
mun. Mag., vol. 52, no. 2, pp. 186-195, Feb. 2014.
[22] F. Rusek, et. al., “Scaling Up MIMO: opportunities and challenges with very large
arrays,” IEEE Signal Proc. Mag., vol. 30, no. 1, pp. 40-60, Jan. 2013.
[23] V. Jungnickel, et. al., “The role of small cells, coordinated multipoint, and massive
MIMO in 5G,” IEEE Commun. Mag., vol. 52, no. 5, pp. 44-51, May 2014.
[24] H. Q. Ngo, E. Larsson, and T. Marzetta, “Energy and spectral efficiency of very large
multiuser mimo systems,” IEEE Trans. Wireless Comm., vol. 61, pp. 1436-1449, Apr.
2013.
[27] L. Zhao et al., “An analysis of the tradeoff between the energy and spectrum efficiencies
in an uplink massive MIMO-OFDM system,” IEEE Trans. Circuits Syst. II: Express
Briefs, vol. 62, no. 3, pp. 291-295, Mar. 2015.
94
Bibliography
[30] H. Q. Ngo, M. Matthaiou, and E. G. Larsson, “Massive MIMO with optimal power and
training duration allocation,” IEEE Wireless Commun. Lett., vol. 3, no. 6, pp. 605-608,
Dec. 2014.
[31] K. Guo, Y. Guo, and G. Ascheid, “Energy-efficient uplink power allocation in multi-cell
MU-massive-MIMO systems,” in Proc. 21st Eur. Wireless Conf., May 2015.
[32] A. M. Tulino and S. Verdu, “Random matrix theory and wireless communications,”
Found. Trends Commun. Inf. Theory, vol. 1, no. 1, pp. 1-182, Jun. 2004.
[33] G. Auer et al., “How much energy is needed to run a wireless network?,” Wireless
Commun., vol. 18, no. 5, pp. 40-49, Oct. 2011.
[34] R. V. R. Kumar, and J. Gurugubelli, “How green the LTE technology can be?,” in Int.
Conf. Wireless VITAE, 2011.
[35] H. Yang and T. L. Marzetta, “Total energy efficiency of cellular large scale antenna
system multiple access mobile networks,” in IEEE Proc. Conf. Green Commun., 2013,
pp. 27-32.
[38] W. Dinkelbach, “On nonlinear fractional programming,” Manage. Sci., vol. 13, no. 7,
pp. 492-498, Mar. 1967.
95
Bibliography
[39] P. D. Tao, “The DC (difference of convex) programming and DCA revisited with DC
models of real world nonconvex optimization problems,” in Ann. Operations Research,
2005.
[41] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge Univ. Press, 2004.
[42] A. Goldsmith et al., “Capacity limits of MIMO channels,” IEEE J. Sel. Areas Commun.,
vol. 21, no. 5, pp. 684-702, Jun. 2003.
[43] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. pp. 165-
171, Englewood Cliffs, Prentice Hall, 1993.
[45] M. Medard, “The effect upon channel capacity in wireless communications of perfect
and imperfect knowledge of the channel,” IEEE Trans. Inf. Theory, vol. 46, no. 3, pp.
933-946, May 2000.
[46] A. Lapidoth and S. Shamai, “Fading channels: how perfect need ”perfect side informa-
tion” be?,” IEEE Trans. Inf. Theory, vol. 48, no. 5, pp. 1118-1134, May 2002.
[47] T. L. Marzetta, “How much training is required for multiuser MIMO?,” in Proc. 40th
Asilomar Conf. Signals, Systems and Computers, 2006, pp. 359-363.
[48] S. Ihara, “On the capacity of channels with additive non-gaussian noise,” Inf. Control,
vol. 37, no. 1, Apr. 1978.
[49] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL of cellular
networks: how many antennas do we need?,” IEEE J. Sel. Areas Commun., vol. 31, no.
2, pp. 160-171, Feb. 2013.
96
Bibliography
[50] A. M. Tulino and S. Verdu, “Random matrix theory and wireless communications,”
Found. Trends Commun. Inf. Theory, vol. 1, no. 1, pp. 1-182, Jun. 2004.
[51] X. Gao, et al., “Massive MIMO performance evaluation based on measured propagation
data,” IEEE Trans. Wireless Commun., vol. 14, no. 7, pp. 3899-3911, Jul. 2015.
[52] A. Adhikary et al., “Joint spatial division and multiplexing for mm-Wave channels,”
IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1239-1255, Jun 2014.
[53] E. Bjornson, L. Sanguinetti, and M. Kountouris. (2015). “Deploying dense networks for
maximal energy efficiency: small cells meet massive MIMO,” [Online]. Available: arXiv
preprint arXiv:1505.01181.
[54] W. Liu, et. al.,“Massive MIMO or small cell network: Who is more energy efficient?,”
in IEEE Proc. Wireless Commun. Networking Conf. (WCNC), 2013, pp. 24-29.
[55] E. Bjornson et al., “Optimal design of energy efficient multi-user MIMO systems: Is
massive MIMO the answer?,” IEEE Trans. Wireless Comm., vol. 14, no. 6, Jun. 2015.
[56] J. Nocedal and S. J. Wright, Numerical Optimization. pp. 140-162, Springer, 1999.
[57] A. Kalis, A. Kanatas, and C. Papadias, “A novel approach to MIMO transmission using
a single RF front end,” IEEE J. Sel. Areas Comm., vol. 26, no. 6, pp. 972-980, Aug.
2008.
[58] M.A. Sedaghat, R.R. Mueller, and G. Fischer, “A novel single-RF transmitter for mas-
sive MIMO,” 18th Int. ITG Workshop Smart Antennas, 2014, pp. 1-8.
[59] T. S. Rappaport et al., “Millimeter wave mobile communications for 5G cellular: it will
work!,” IEEE Access, vol. 1, pp. 335-349, 2013.
97
Bibliography
[61] H. Zhao et al., “28 GHz millimeter wave cellular communication measurements for
reflection and penetration loss in and around buildings in New York City,” in IEEE
Proc. Int. Conf. Commun., 2013, pp. 5163-5167.
[62] A. Alkhateeb et al., “Channel estimation and hybrid precoding for millimeter wave
cellular systems,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831-846, Oct.
2014.
[63] L. Liu, Y. Li, and J. Zhang, “DoA estimation and achievable rate analysis for 3D
millimeter wave massive MIMO systems,” in IEEE 15th Int. Workshop Signal Process.
Adv. Wireless Commun., 2014, pp. 6-10.
[64] J. Mo and R. W. Heath, “Capacity analysis of one-bit quantized MIMO systems with
transmitter channel state information,” IEEE Trans. Signal Process., vol. 6, no. 81, pp.
1, Jul. 2015.
[65] S. Hur, et. al., “Millimeter wave beamforming for wireless backhaul and access in small
cell networks,” IEEE Trans. Commun., vol. 61, no. 10, pp. 4391-4403, Oct. 2013.
[66] M. R. Akdeniz et al., “Millimeter wave channel modeling and cellular capacity evalua-
tion,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1164-1179, Jun. 2014.
[67] J. Hoydis et al., “Making smart use of excess antennas: Massive MIMO, small cells,
and TDD,” Bell Labs Tech. J., vol. 18, no. 2, pp. 5-21, Sept. 2013.
[68] M. Kountouris and N. Pappas, “HetNets and massive MIMO: Modeling, potential gains,
and performance analysis,” in IEEE Proc. APS Topical Conf. Antennas Propagat. Wire-
less Commun., 2013.
[69] Summary of the description of candidate eICIC solutions, 3GPP Standard R1-104968,
2010.
98
Bibliography
[71] A. Adhikary, H. S. Dhillon, and G. Caire, “Massive MIMO meets HetNet: interference
coordination through spatial blanking,” IEEE J. Sel. Areas Commun., vol. 33, no. 6,
pp. 1171-1186, Jun. 2015.
[72] E. Bjornson, M. Kountouris, and M. Debbah, “Massive MIMO and small cells: improv-
ing energy efficiency by optimal soft-cell coordination,” in 20th Int. Conf. Telecomm.,
2013, pp. 1-5.
[73] D. Bethanabhotla et al., “Optimal user-cell association for massive MIMO wireless
networks,” IEEE Trans. Wireless Commun., vol. 6, no. 81, pp. 1-1, Jul. 2014.
[74] Y. Xu and S. Mao, “User association in massive MIMO HetNets,” IEEE J. Syst., vol.
6, no. 81, pp. 1-1, Sept. 2015.
[75] D. Liu et al., “Distributed energy efficient fair user association in massive MIMO enabled
HetNets,” IEEE Commun. Letters, vol. 6, no. 81, pp. 1-1, July 2015.
[76] Texas Instruments. (2010). Energy harvesting ULP meets energy harvest-
ing: a game-changing combination for design engineers. [Online]. Available:
http://www.ti.com/corp/docs/landing/cc430/graphics/slyy01820081031.pdf.
[77] J. Chiasson and B. Vairamohan, “Estimating the state of charge of a battery,” IEEE
Trans. Control Syst. Technol., vol. 13, no. 3, pp. 465-470, May 2005.
[78] Z. Zhou et al., “Energy-efficient antenna selection and power allocation for large-scale
multiple antenna systems with hybrid energy supply,” in IEEE Proc. Global Commun.
Conf. (GLOBECOM), 2014, pp. 2574-2579.
99
Bibliography
[79] J. Zhu, R. Schober, and V. K. Bhargava, “Linear precoding of data and artificial noise
in secure massive MIMO systems”, IEEE Trans. Wireless Commun., vol. PP, no. 99,
pp. 1, Nov. 2015. DOI: 10.1109/TWC.2015.2500578
[81] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless information
and power transfer,” IEEE Trans. Wireless Commun., vol. 12, no. 5, pp. 1989-2001,
May 2013.
[82] Y. Zeng, and R. Zhang, “Optimized training design for wireless energy transfer,” IEEE
Trans. Commun., vol. 63, no. 2, pp. 536-550, 2015.
[83] X. Chen, X. Wang, and X. Chen, “Energy-efficient optimization for wireless information
and power transfer in large-scale MIMO systems employing energy beamforming,” IEEE
Wireless Commun. Lett., vol. 2, no. 6, pp. 667-670, Dec. 2013.
[85] T. Takamoto et al., “World’s highest efficiency triple-junction solar cells fabricated by
inverted layers transfer process,” in 35th IEEE Photovoltaic Specialists Conf., 2010, pp.
412-417.
[86] H. Wang et al., “Wireless information and energy transfer in interference aware massive
MIMO systems,” IEEE Proc. Global Commun. Conf. (GLOBECOM), Austin, TX, 2014,
pp. 2556-2561.
[87] Energy efficiency analysis of the reference systems, areas of improvements and target
100
Bibliography
[89] G. Wunder et al., “5GNOW: non-orthogonal, asynchronous waveforms for future mobile
applications,” IEEE Commun. Mag., vol. 52, no. 2, pp. 97-105, Feb. 2014.
[91] S. C. Cripps, RF Power Amplifiers for Wireless Communications, pp. 165-171, Artech
Pub. House, 1999.
[93] H. Prabhu et al., “A low-complex peak-to-average power reduction scheme for OFDM
based massive MIMO systems,” 6th Int. Symp. Commun. Control Signal Process., 2014,
pp. 114-117.
[95] H. V. Cheng, D. Persson, and E. G. Larsson, “MIMO capacity under power ampli-
fiers consumed power and per-antenna radiated power constraints,” 15th Int. Workshop
Signal Process. Adv. Wireless Commun., 2014, pp. 179-183.
101
[96] B. Hu et al., “Energy efficiency of massive MIMO wireless communication systems with
antenna selection,” J. China Univ. Posts Telecommun., vol. 21, pg. 1-8, December 2014.
[97] B. M. Lee et al., “An energy efficient antenna selection for large scale green MIMO
systems,” in IEEE Int. Symp. Circuits Syst., 2013, pp. 950-953.
[99] T. Schenk, RF Imperfections in High-Rate Wireless Systems: Impact and Digital Com-
pensation, pp. 181-185, Springer, 2008.
[100] B. D. O. Anderson, and J. B. Moore. Optimal Filtering. pp. 93-95, Courier Corp., 2012.
102
Appendix A
L
X 1
ĥlij = hlqj [D̂li ]j,j + √ wlj [D̂li ]j,j
i=q
τ pp
= E{(ĥlij )(ĥlij )T }
L
p
X ( βlqj )2 βlij
2 2 2
τ pp 1 2 2 2 2
βlij τ pp
=( PL + (√ ) L
)IM
τ pp (1 + q=1 τ pp βlqj )2
2
P
q=1 (1 + q=1 τ pp βlqj )
( Lq=1 βlqj τ pp + 1)
P (A.1)
2
= (βlij τ pp )IM
(1 + Lq=1 τ pp βlqj )2
P
2
τ pp βlij
=( )IM
(1 + Lq=1 τ pp βlqj )
P
103
A.2. Effect of Channel Estimation Error of Known Variance on Achievable Rates
104
A.2. Effect of Channel Estimation Error of Known Variance on Achievable Rates
Consider a scenario where a single user transmits a symbol X to the BS, where X is
2
Gaussian distributed with zero mean and variance σX . Assuming an AWGN channel with
2
noise variance σN , the received symbol Y at the BS is given by,
Y = HX + N,
where H is the channel strength and N is the noise. We assume that X, H, and N are
statistically independent. Given a channel estimate Ĥ, we can break H into two components
Ĥ and H̃, i.e., H = Ĥ + H̃, where H̃ is the zero-mean channel estimation error at the BS,
with known variance σ 2 . Now, the mutual information I(X; Y ) between Y and X is given
by
where h denotes differential entropy. Since we have chosen X to be Gaussian with variance
2
σX , we know the value of h(X). We now find an upper bound on h(X|Y ) so as to obtain a
lower bound on I(Y ; X). By definition of differential entropy,
Z
h(X|Y ) = h(X|Y = y)pY (y)dy.
105
A.2. Effect of Channel Estimation Error of Known Variance on Achievable Rates
Since the entropy of a random variable with a given variance is upper-bounded by the
entropy of a Gaussian random variable with the same variance, we can re-write (A.5) as
h(X|Y ) ≤ h(X − αY )
(A.7)
1
≤ loge (2πe var (X − αY )),
2
where var(.) represents the variance of random variable under context. (A.7) is also valid
when we minimize the RHS over α. Let us pick α such that αY is the MMSE estimate of
X in terms of Y . Then [100],
E[XY ]
α=
E[Y 2 ]
2
ĤσX
= (A.8)
2 2 2 2
ĤσX + σH̃ σX + σN
(obtained upon simplification using the fact that X and Y are zero mean)
4 2 2 2
σX σH̃ + σN σX
var (X − αY ) = 2 2 2 2
. (A.9)
Ĥ 2 σX + σH̃ σX + σN
4 2 2 2
1 σX σH̃ + σN σX
h(X|Y ) ≤ loge (2πe ) (A.10)
2 2
Ĥ 2 σX + σ 2 σX
2 2
+ σN
H̃
Using this result in (A.3), we obtain a lower bound on the mutual information between X
and Y as follows
106
A.2. Effect of Channel Estimation Error of Known Variance on Achievable Rates
4 2 2 2
1 2 1 σX σH̃ + σN σX
I(X; Y ) ≥ loge (2πeσX ) − loge (2πe )
2 2 2 2 2
Ĥ 2 σX + σ σX + σN 2
H̃
(A.11)
1 Ĥ 2 σX
2
= loge (1 + 2 2 2
)
2 σH̃ σX + σN
From (A.11), we can interpret that the worst effect the channel estimation error can have is
to behave as AWGN. The bound in (A.11) is equal to the capacity of the wireless channel
when we send Gaussian signal with variance Ĥ 2 σX
2
in an AWGN channel with effective noise
2 2 2
whose variance is σH̃ σX + σN .
The result in (A.11) is based on an assumption that the variance of channel estimation
error is known at the BS. Since we assume that the BS knows βlik values, the BS also
knows the variance of channel estimation errors (c.f. (4.5)). Therefore, we use the above
interpretation and model channel estimation error as part of additive Gaussian noise in order
to obtain a lower bound on the rate expression, as given in (4.10).
107
A.3 Proof of Proposition 2
PL PK PL PK
0 τ pu i=1 j=1,(i,j)6=(l,k) |ĥH 2
llk ĥlij | + pu ||ĥllk ||
2
i=1 j=1 (1 − d0lij )βlij + ||ĥllk ||2
Rlk = B (1 − ) log2 (1 + (E{ })−1 )
T pu ||ĥllk ||4
PL PK PL PK
τ pu i=1
H
j=1 |ĥllk ĥlij |
2
− pu ||ĥllk ||4 + pu ||ĥllk ||2 i=1 j=1 (1 − d0lij )βlij + ||ĥllk ||2
= B (1 − ) log2 (1 + (E{ })−1 )
T pu ||ĥllk ||4
PL PK PL PK
2
− d0lij )βlij
PL
j=1,j6=k |ĝlij |
pu H 2 pu
0 τ i=1 i=1 pu |ĥllk ĥlik | i=1 j=1 (1 1
Rlk = B (1 − ) log2 (1 + (E{ + −1+ + })−1 )
T pu ||ĥllk ||2 pu ||ĥllk ||4 pu ||ĥllk ||2 pu ||ĥllk ||2
L K L K L 2
τ X X XX 1 X βlij
= B (1 − ) log2 (1 + ((pu E{|ĝlij |2 } + pu (1 − d0lij )βlij + 1)E{ }+ − 1)−1 ) (using 4.6)
T i=1 j=1,j6=k i=1 j=1 pu ||ĥllk ||2 β2
i=1 llk
τ
= B (1 − )
T
L K 2 L X K 2 L 2
X X τ pp βlij X τ pp βlij 1 X βlik
log2 (1 + ((pu + pu (βlij − ) + 1)E{ }+ ( 2 ) − 1)−1 )
m
E{trace(A−1 )} = , (A.14)
n−m
where A is an m × m central complex Wishart matrix with n (n > m) degrees of freedom. Using this result, we obtain
1 1
E{ }= , for M ≥ 2
pu ||ĥllk ||2 pu (M − 1)d0llk βllk
(A.15)
(1 + Li=1 τ pp βlik )
P
= 2
pu (M − 1)τ pp βllk
L X
X L L
X L
X L
X
2 2 2
c1k , τ βllk , c2k , βlij , c3k , τ βlik c4k , τ βlik c5k , c2k c3k − τ βlik (A.17)
i=1 j=1 i=1 i=1,i6=l i=1
0 τ c1k (M − 1)pp pu
Rlk (M, pp , pu ) = B (1 − ) log2 (1 + ),
T c2k pu + c3k pp + c4k (M − 1)pp pu + c5k pp pu + 1
0 τ
Rlk (M ; pp0 , pu0 ) = B (1 − ) log2 ((A1k + A2k )M + (A3k − A1k ))
T (A.18)
τ
−B (1 − ) log2 (A2k M + A3k ),
T
0
Substituting the expression for Rlk (M ; pp0 , pu0 ) from (A.18) and the expression for Ptot (M ; pp0 , pu0 )
from (4.23) into the expression for P3 (θ) in (4.21), we obtain
K
X τ
B (1 − ) log2 ((A1k + A2k )M + (A3k − A1k ))
T
max k=1
M ∈R
P3 (θ) = τ
−B ((1 − ) log2 (A2k M + A3k ) − (θD1 M + D2 ))
T
subject to : C1 ,
K (A.19)
X τ
−B (1 − ) log2 ((A1k + A2k )M + ((A3k − A1k ))
T
min k=1
M ∈R
= τ
+B (1 − ) log2 (A2k M + A3k ) − (θD1 M + D2 ))
T
subject to : C1 ,
where we have transformed the original maximization problem into an equivalent minimiza-
tion problem. Let us now define
τ
flk (M ) = −B (1 − ) log2 ((A1k + A2k )M + (A3k − A1k )),
T (A.20)
τ
glk (M, θ) = −B (1 − ) log2 (A2k M + A3k ) + θ(D1 M + D2 )
T
Substituting (A.20) into (A.19), we obtain
PK
k=1 flk (M ) − glk (M, θ)
min
M ∈R+
P3 (θ) = (A.21)
subject to : C1 ,
111
A.4. Proof of Proposition 4
To prove that flk (M ) and glk (M, θ) are convex, it is sufficient to prove that their second-order
derivates with respect to M are always positive.
(A1k + A2k )2 B (1 − Tτ )
∇2M flk (M ) =
(A1k + A2k )M + (A3k − A1k )2
> 0 (since both numerator and denominator are always positive)
B (1 − Tτ )A22k
∇2M glk (M, θ) =
(A2k M + A3k )2
> 0 ( since both numerator and denominator are always positive)
Since flk (M ) and glk (M, θ) are convex ∀(l, k), the objective function in (A.21) is the sum-
mation of a set of functions, each of which is in turn a difference of convex (DC) function.
Since the summation of DC functions is a DC function and the constraint C1 in (A.21) is
convex , the optimization problem in (A.21) is a DC programming problem. This completes
the proof.
112