System Information
Distribution in Massive
MIMO Systems
Simon Sörman
Master of Science Thesis in Communication Systems
Simon Sörman
LiTH-ISY-EX--16/4952--SE
Communication Systems
Department of Electrical Engineering
Linköping University
SE-581 83 Linköping, Sweden
v
Sammanfattning
Den 5:e generationens mobila nätverk (5G) håller för tillfället på att specifice-
ras och utvecklas med stora förväntningar och krav på datatakt och effektivitet.
4G, och mer specifikt LTE, kan utgöra en grund för designen av nätverket men
det finns flera delar som borde förbättras. En sak att förbättra är den frekven-
ta statiska signalleringen, där systeminformation så som synkroniseringssigna-
ler, nätverksfrekvenser, operatörer, konfigureringar etc. utgör en del. Det har ti-
digare visats att den statiska signalleringen kräver både mycket energi och tid-
frekvensresurser. Eftersom systeminformationen inte är riktad till en specifik an-
vändare så sänds den frekvent till hela cellen, så att alla användare, oavsett antal,
alltid har tillgång till att läsa informationen om det behövs.
5G kommer att använda sig av tekniken massiv MIMO, där basstationen är ut-
rustad med ett stort antal antenner som tillsammans kan användas för att rik-
ta signaler, vilket kallas lobformning. Denna uppsats presenterar en ny metod
för distribueringen av systeminformation som kan utnyttja möjligheten att lob-
forma. En enkel modell tillsammans med kanalstatistik från simuleringar av ur-
bana 4G-scenarion används för att visa att den nya metoden kräver mindre tid-
frekvensresurser än vad som krävs då informationen alltid sänds över hela cellen.
Med höga krav på låg latensen för systeminformationen är vinsten med den nya
metoden särskilt stor.
vii
Acknowledgments
I would like to express my gratitude to Martin Hessler and Erik Eriksson at Er-
icsson Research in Linköping for guiding me and supporting my work with this
thesis. Whenever I was stuck, you would present many ways forward for me to
choose from. I would also like to thank the entire Ericsson LINLAB team for a
great time while writing my thesis. It has been inspiring to work in a such stim-
ulating environment. Another thanks goes to my fellow master thesis writers at
Ericsson who I have enjoyed many moments with during this thesis.
At Linköping university I want to thank Erik G. Larsson and Emil Björnson for
their help and advice. It is a privilege to have the possibility to learn from such
distinguished researchers.
Finally, I would like to thank my fiancée for giving me strength and motivation
whenever I need it. You have had a great part in creating this thesis.
ix
Contents
Notation xiii
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 LTE 5
2.1 Orthogonal Frequency-Division Multiplexing . . . . . . . . . . . . 5
2.2 System Information . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Random Access . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Step 1: Preamble . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Step 2: Random Access Response . . . . . . . . . . . . . . . 10
2.3.3 Step 3: Identification . . . . . . . . . . . . . . . . . . . . . . 11
2.3.4 Step 4: Collision Resolution . . . . . . . . . . . . . . . . . . 11
3 Transmission Methods 13
3.1 Broadcasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.1 Single Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.2 Multiple Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.3 Beam Sweeping . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Dedicated Transmissions . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.1 Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.2 Channel State Information . . . . . . . . . . . . . . . . . . . 21
3.2.3 Reciprocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
xi
xii Contents
5 Simulation 39
5.1 SINR Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.1 Single Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.2 Multiple Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Beamforming Gain Simulations . . . . . . . . . . . . . . . . . . . . 42
6 Results 43
6.1 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.1.1 SINR Simulations . . . . . . . . . . . . . . . . . . . . . . . . 43
6.1.2 Beamforming Gain Simulations . . . . . . . . . . . . . . . . 46
6.2 Method Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2.1 A Minimum on the Bandwidth . . . . . . . . . . . . . . . . 47
6.2.2 Latency Dependencies . . . . . . . . . . . . . . . . . . . . . 50
7 Conclusion 59
7.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.3 The Thesis in a Wider Perspective . . . . . . . . . . . . . . . . . . . 62
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.4.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
List of Figures 65
List of Tables 67
Bibliography 69
xiii
xiv Notation
Notation
Abbreviations
Abbreviation Definition
3gpp 3rd Generation Partnership Project
aoa Angle of Arrival
aod Angle of Departure
bs Base Station
cdf Cumulative Density Function
cp Cyclic Prefix
csi Channel State Information
fdd Frequency Division Duplex
fft Fast Fourier Transform
gob Grid of Beams
ifft Inverse fft
isi Inter-Symbol Interference
itu International Telecommunication Union
los Line of Sight
lte Long-Term Evolution
mib Master-Information Block
mimo Multiple-Input Multiple-Output
mmse Minimum Mean Square Error
mrt Maximum Ratio Transmission
ofdm Orthogonal Frequency-Division Multiplexing
prach Physical Random Access Channel
pss Primary Synchronization Signal
si System Information
sib System-Information Block
sinr Signal-to-Interference-plus-Noise Ratio
sip1 System Information Part 1
sip2 System Information Part 2
snr Signal-to-Noise Ratio
sss Secondary Synchronization Signal
tdd Time Division Duplex
ue User Equipment
zf Zero-Forcing
Notation xv
Notation Definition
C The set of all complex numbers
R+ The set of all non-negative real numbers
U (a, b) Uniform distribution in the interval [a, b[
CN (µ, σ ) Circular symmetric complex normal distribution with
mean µ and variance σ 2
Qp (Y1 ; Y2 ) Random variable that takes the value Y1 with proba-
bility p, otherwise Y2
Defined Parameters
Notation Definition
T1 Interval between broadcasts of sip1
T2 Interval between broadcasts of sip2
T3 Interval between allocated pilot slots
N1 Size of sip1
N2 Size of sip2
N3 Size of pilot (pilot overhead)
pf Full coverage
pc Variable coverage
Ru User arrival rate
A (t) Poisson process of user arrivals
A1 (t) Poisson process of sip2 -covered user arrivals
A2 (t) Poisson process of non-sip2 -covered user arrivals
λ Intensity of the process A2 (t)
L Random variable of latency
LN Random variable of latency, case Normal Detection
LE Random variable of latency, case Early Detection
D Random variable of pilot retries
Si State in a Markov chain where i ues transmit a pilot
ai Probability of adding i users to a Markov chain in one
step
π Stationary probability distribution of a Markov chain
ti Probability of being served a sip2 if i ues transmits a
pilot
Br Bandwidth of reference method
Bn Bandwidth of new proposed method
Bded Bandwidth of dedicated transmissions of sip2
Introduction
1
This chapter is a short introduction to the thesis. It is intended to provide a back-
ground, motivation and purpose of the thesis, as well as presenting the actual
questions that are intended to be answered by it.
1.1 Background
In the last years there has been a large increase in the use of mobile data. As of
today, being connected is a natural thing, and already the number of mobile de-
vices and connections exceeds the total global population. It is predicted that the
growth will continue at a fast pace, with the data traffic increasing with more than
50% every year at least until year 2020 [1]. To be able to support this increase
in traffic, the 3rd Generation Partnership Project (3gpp) is working to create a
specification for the next generation mobile network (5G) that is to be submitted
to the International Telecommunication Union (itu) in 2020 [2].
1
2 1 Introduction
energy at a specific user, see Figure 1.1. By beamforming at several users at once,
the bs can effectively communicate with several users at the exact same time and
exact same frequency. The focusing of energy in space also reduces the energy
that is wasted by transmitting signals in all directions, leading to lower power
consumption [3].
A vital part of mobile networks is the distribution of System Information (si),
i.e. information about the system that is not specific to a single user. Examples
of such information can be detection of the network and synchronization to it,
cell identity, timing information, public warning messages, information about
accessing the network etc. Historically, the distribution has been performed by
broadcasting it frequently. For instance, lte defines a number of different signals
that all contain a set of the entire si. All those signals are broadcast with some
periodicities, which is explained further in Section 2.2. Important to note is that
broadcasting is performed without knowledge of whether any users are actually
listening. This is so that users always will be able to read the information when
needed [4].
Clearly, system information will have to be distributed in the next generation
networks as well. It therefore seems motivated to investigate the design of the
distribution given the new technology provided by massive mimo.
1.3 Method
This thesis is written at Ericsson Research in Linköping. The methodology and
content of the thesis are developed under supervision from researchers at the site.
In the beginning, literature was researched to investigate the relevant communi-
cation system theory. The background of the subject of massive mimo is based
on recent academic research and small case studies, as the subject is quite new.
As for the concept of system information distribution, the background is mainly
based on the procedure performed in lte and therefore some information about
lte was also required. Then, distribution methods were defined together with
theoretical models that relied on data extracted from simulations. All results in
this thesis stem from data collected from an internal simulator at Ericsson and
these models. The obtained results were analysed with aim of providing a rigid
foundation of answering the questions posed in the Problem Formulation.
1.4 Limitations
To keep the scope of the thesis to a reasonable size, some limitations have been
put on the thesis. They are as follows:
• Section 3.1 introduces three different methods of broadcasting. However,
only the first two will be considered.
• The thesis assumes that each bs will be equipped with 100 antennas. The
actual number does not show to be crucial, but this is still a limitation.
• Due to restrictions of the simulator, the thesis has only considered one type
of beamforming.
• Simulations of the actual si distribution methods have not been performed,
but the results rely on modelling of a system and theoretical analysis of
these models.
• Since there is a very large amount of parameters that affects the system,
some of them are not varied throughout the thesis. They are however cho-
sen in accordance with expectations of future mobile networks.
• The most complex variant of the models of the proposed method of si dis-
tribution has not been fully analysed due to time issues, it is only used to
motivate some conclusions.
4 1 Introduction
The leading technology for the fourth generation mobile network systems is Long-
Term Evolution. The first specification of lte was released by 3gpp in 2008 and
offers downlink peak rates of 300 Mbps and uplink peak rates of 75 Mbps. How-
ever, it was not fully classified as advanced 4G by itu until a new release in 2011.
lte has enjoyed a great commercial and technological success, therefore it makes
sense to compare the next generations networks to lte, and in some parts base
the design on it [5].
This chapter presents those parts of lte that is of concern to this thesis. It aims
to provide a sufficient background of currently used technologies.
5
6 2 LTE
s0
f0
s1
s0 , s1 , s2 , . . . f1 x(t)
S→P +
···
sN −1
fN −1
the pulse-shape is that it fulfils the Nyquist criterion, which means that it should
have the same zero-crossings as the rectangular pulse. The resulting spectrum
of the subcarriers looks as in Figure 2.2, where each carrier has an individual
line. In the figure it can be seen that at each subcarrier’s peak, the other subcar-
riers have zero signal energy. The spectrum of the total signal is the sum of all
subcarrier spectra.
To see mathematically why the subcarriers are orthogonal, we observe that the
signal on a subcarrier with frequency fn can be expressed in complex baseband
notation by xn (t) = sn ej2πnt/Tu . We then have orthogonality between different
subcarriers (m , n) due to
ZTu ZTu
xm (t)xn∗ (t) dt = sm sn∗ ej2π(m−n)t/Tu dt = 0. (2.1)
0 0
will suffer from Inter-Symbol Interference (isi). This problem is solved by the
insertion of a Cyclic Prefix (cp) in front of every ofdm symbol. The cp is simply
the last part of the ofdm symbol and we denote its length by TCP . Figure 2.3
illustrates that if the demodulator only integrates over the useful time interval,
the orthogonality is preserved even towards signal components that are not time
delayed by more than TCP , since the integration is only performed over one ofdm
symbol together with its cp. As in (2.1), this is also easily shown by an integral,
in which τ ≤ TCP :
ZTu Zτ ZTu
xm (t)xn∗ (t − τ) dt = xm (t)xn∗ (t − τ + Tu ) dt + xm (t)xn∗ (t − τ) dt
0 0 τ
Zτ ZTu
= sm sn∗ ej2π(m−n)t/Tu ej2πnτ/Tu dt + sm sn∗ ej2π(m−n)t/Tu ej2πnτ/Tu dt = 0. (2.2)
0 τ
The simplification of the first term in the second equality is valid due to the fact
that ej2πnTu /Tu = 1, and therefore the shift of Tu disappears. It should be noted
that the use of a Cyclic Prefix entails that only Tu / (Tu + TCP ) of the received signal
energy is used for demodulation [6].
In lte under normal operation we have ∆f = 15 kHz and TCP ≈ 5.1 µs in every
seventh ofdm symbol and TCP ≈ 4.7 µs in all other symbols. The reason for this
is that the symbols should fill the defined time slot of 0.5 s [4].
8 2 LTE
TCP Tu
Direct path
Time-delayed path
Integration interval
Figure 2.3: Time dispersion is handled in ofdm with the insertion of a Cyclic
Prefix.
Since ofdm has been very successful in lte, it is highly likely that ofdm will
also be the foundation of the modulation used in 5G. There are currently some
waveform suggestions for 5G, all of which are multi-carrier modulations that can
be seen as ofdm with some modifications. Thus the time-frequency grid visual-
ization is valid also for those schemes [7] [8].
fN −1
) .
ie
r .. .
r ..
ar f4
u bc
(s f3
y
nc f2
ue f1
eq
Fr f0
... ...
m−3 m−2 m−1 m m+1 m+2 m+3
Before the ue can access the network there is some more si that is needed. The
most important information is located in a so-called Master-Information Block
(mib), in which the content is very small (only 24 bits) and is broadcast with
an interval of 40 ms. However, the mib is extensively channel coded so that even
users with really bad channel conditions are able to decode it. The mib is required
to be able to read the rest of the system information, which is located in a range
of different sibs (System-Information Blocks). These contain different sets of the
total system information and are quite varying in size and transmission interval,
and some are entirely optional for the network to use. But in all cases, both sib1
and sib2 are required before the ue can finally access the network to be able to
operate in it [4].
Apart from the explicit information described above, lte also broadcasts a set
of various reference signals intended to be used for channel estimation, demodu-
lation reference, positioning etc. Unfortunately all these static signals has to be
transmitted at all times, even when there are no users present in the cell. This
leads to some seemingly unwanted results, as shown in [9]. In that study, an
lte network was simulated for 4 scenarios during 24 hours, resulting in the his-
togram in Figure 2.5. The three scenarios with traffic were defined as “The most
relevant European scenario for 2015”, “An upper bound on the anticipated traf-
fic for 2015” and “An extremity for very high data usage in future networks”
respectively. The figure shows that the amount of empty sub-frames in the net-
work is large even under heavy load. The report does not elaborate on why this
is the case, but the result implies that the static signalling takes up very much re-
sources, since adding a large amount of traffic does not decrease the percentage
of empty sub-frames very much. This is compared to the case of no traffic, where
the only used sub-frames are static signalling. Note that the static signals are not
considered to occupy any sub-frames in these percentages. This motivates the
investigation of how to distribute system information more effectively in future
networks.
100
90
80
Empty sub−frame ratio [%]
70
60
50
40
30
20
10
0
No traffic Low traffic Medium traffic High traffic
Figure 2.5: The ratio of empty sub-frames at an lte bs for four different
traffic scenarios.
ways.
The random access procedure in lte consist of four steps described below.
dedicated uplink and downlink resource blocks. ues that used different pream-
bles in Step 1 will get individual resource blocks, but those that used the same
preamble are still unaware of the collision. The timing correction included in the
random access response is used to compensate for the propagation delay between
the ue and the bs [4].
3.1 Broadcasting
A crucial part of current radio networks is broadcasting, which is a process where
the bs transmits the signal to the entire cell area, so that the transmission reaches
all users that might be listening. This is today vital to supply ues with system
information so that they can operate correctly in the network.
Important to note is that broadcasting can be done with no information at the
bs about where users are positioned, or about if there even are users listening.
Broadcasting is intended to make it possible for all users to read the transmission
at any time, without prior signalling to the bs.
Also, one should know that while broadcasting is commonly made with just a
single antenna, which is also the case assumed throughout this thesis, there is
some research that shows the benefits of using massive mimo for broadcasting
as well. In [10] and [11] it is shown that some special types of precoding (ex-
plained in Section 3.2.1) results in broadcasts that are more resilient to different
bad channel conditions by introducing extra diversity. Broadcasting can also be
13
14 3 Transmission Methods
Figure 3.1: Illustration of Single Cell broadcasting, each cell broadcasts in-
dependently.
16 3 Transmission Methods
Each bs will thus send the same signal s(t), which due to multipath propagation
will result in a total received signal at the ue of
XX
r(t) = αi,j s(t − τi,j ) + n(t), (3.2)
i j
where the outer sum is over all base stations, and the notation is the same as in
(3.1). As can be seen, this appears to the ue as nothing else than multipath prop-
agation of a single transmission of s(t), although the number of paths is larger
than in a single cell case. Furthermore, since the distances to different bss might
differ, the spread of the delays τi,j will be larger than the spread of a single base
station’s signal. The effect of this is that some signal parts from cells that would
be interference in a Single Cell setting are now turned into useful signal energy
instead. But some signal parts are delayed by too much and will still constitute
interference.
This method of broadcasting is also used in lte, although not for transmission of
system information but was intended for multimedia broadcasting. [4]
Important to note here is that the TCP in lte is 4.7 to 5.1 µs, which implies that the
significant delay spread of the channel is not expected to be larger than this value.
A distance of just 1 km corresponds to a signal time delay of 3.3 µs, which then
should imply that the distance between bs i and the ue constitutes the largest
part of most of the τi,j . Because of this, it is motivated to make an approximation.
Let c be the speed of light, B = {bi } be the set of all bss with corresponding
distances di to the ue, such that d0 < d1 < . . . < dM , and lastly dm is the smallest
3.1 Broadcasting 17
In other words, the approximation is that all signals transmitted from bss that
are close enough to the ue, as compared to the closest bs, will contribute fully to
the received signal energy. bss that are further away will only contribute to the
interference. Another property that makes this a good approximation is that αi,j
is almost certainly smaller for larger τi,j due to both longer propagation distance
and that the signal might have gone through more reflections. Thus, the signal
parts that have larger power also have a delay closer to the distance between the
bs and the ue divided by c.
Figure 3.3 illustrates the above approximation in a scenario with 3 base stations.
Each rectangle represents a version of the transmitted signal, with the color inten-
sity indicating energy, and the horizontal position representing the arrival time
at the ue, i.e. in total αi,j s(t − τi,j ). In reality, only those copies that are received
within TCP constitute useful signal energy, and in the figure those rectangles are
plain. The approximation, however, considers all signal parts from close enough
18 3 Transmission Methods
TCP
bs1
bs2
bs3
base stations as useful, thus the vertically striped rectangles will also be included
as signal energy. The rectangles with a diagonal line are interference in both
reality and in the approximation.
The third and last broadcasting method is set in a quite different scenario as
compared to Section 3.1.1 and 3.1.2. To use this method, the bs has to have a
Grid of Beams (gob), which is defined as a fixed set of beams (precoders, see
Section 3.2.1) for possible use [13].
The effect of beam sweeping is that most ues will be able to receive a broadcast
with a larger signal energy than otherwise, since each one should be positioned
inside at least one beam. However, if this type of system is to have the same
latency as a system where the broadcast covers the entire cell, the beam sweep
has to be made quite fast. In fact, a gob with N beams will have to transmit the
broadcast signal a factor N times more frequently than a broadcast that covers
the entire cell.
3.2 Dedicated Transmissions 19
Sweep direction
3.2.1 Precoding
Consider a system where the bs has M antennas, and there are K ues with one
antenna each. If we limit the system to operate within the coherence bandwidth
and the coherence time, the channel between each bs antenna and each ue is
just a multiplication with a complex constant in complex baseband notation. A
20 3 Transmission Methods
In this thesis, we will make use of the mrt precoding and the property of it that
the Signal-to-Interference-plus-Noise Ratio (sinr) scales with M/K for large M
and K [15].
3.2.3 Reciprocity
Reciprocity is as stated a property of the wireless communication channel that
means that the channel is the same in uplink and downlink. There are, however,
different interpretations that can be made of this definition.
The simplest type of reciprocity in a multipath environment is directional reci-
procity, which means that the Angle of Arrival (aoa) of the channel is the same as
the Angle of Departure (aod). This originates in the fact that the most significant
scatterers for a user are the same in both up- and downlink [17] [18]. Scatterers
are the points in space that reflect the signals to produce a multipath channel.
Thus by estimating the few strongest aoas of the pilot signal sent by the ue, the
bs also knows which aods to use to aim the signal energy in direction of the ue.
To exploit this reciprocity, one can use a pilot of small bandwidth but the down-
link transmission can have a much larger bandwidth since the aoas and aods are
quite independent of the frequency of the signal. Obviously, the precoders in Sec-
tion 3.2.1 cannot be used with directional reciprocity, but the precoding matrix
has to be based on the estimated aoas instead. This type of reciprocity should
be possible to exploit in both a Time Division Duplex (tdd) and a Frequency Di-
vision Duplex (fdd) system, as the received aoa includes information for other
frequencies than the received signal.
At the other end of the reciprocity spectra, there is what is here denoted as full
reciprocity. This is the conventional form of reciprocity in massive mimo litera-
ture, where the channel coefficients are the same in up- and downlink [3]. When
22 3 Transmission Methods
the bs receives the orthogonal pilot signals from the ues, it can estimate the chan-
nel coefficient to each ue with some conventional estimation method. It may how-
ever be more expensive to try to utilize full reciprocity rather than directional.
Since the channel coefficient is only constant within the coherence bandwidth,
the downlink transmission can only use as much bandwidth as the pilot signal
used. For instance, if the system should use four times the coherence bandwidth
for the downlink transmission, the ue has to repeat its pilot signal (which is only
as wide as the coherence bandwidth) on four different carrier frequencies. This
type of reciprocity is only applicable in a tdd system, as the reciprocity-based
transmission must be sent on the same frequency as the pilot was received on.
Observe that reciprocity is a property of the physical channel, while the hardware
is not in general reciprocal. This can be countered by calibration of the hardware
chains to achieve full reciprocity [3].
System Information Distribution
4
Methods
23
24 4 System Information Distribution Methods
but that the probability of a user experiencing worse channel conditions than
what the system is capable of handling is larger than zero. It therefore makes
sense to define full coverage as a probability of not being able to serve a user, or
equivalently, defining a percentage of coverage. In this thesis we will define full
coverage as pf = 0.9999, i.e. that with full coverage we expect that 99.99% of all
users can be served.
Since this distribution method is what historically have been used as standard, it
will in this thesis be seen as a reference to compare with.
sip1 sip2
Figure 4.1: The set of system information is split into two parts.
This proposed method does not specify anything about what type of information
that should go into which part, other than that sip1 has to contain information on
how to request sip2 if that procedure is not entirely fixed in the protocol. In any
case it might be useful to, in an objective way, think of sip1 as the information
4.2 Proposed Approach 25
that is most critical and is vital to all users, and sip2 as the information that is
good to know but not equally important.
The requesting procedure can be designed in different ways. In this thesis three
different procedures have been identified and are presented in Sections 4.2.1 to 4.2.3.
The first request procedure alternative is also the simplest one. The bs pre-
allocates some time-frequency resources (slots) that can be used by ues to trans-
mit pilot signals in. A pilot signal is a specific sequence that is known a priori
by both the ue and the bs, which makes it possible for the bs to both detect the
request and to estimate the channel to the ue. There might be more than one
possible pilot signal, and in that case the ue chooses one from the pre-defined
set. All pilot signals in this set should be pairwise orthogonal, so that it is possi-
ble to separately estimate channels to users that happens to choose different pilot
signals in the same pilot slot.
In a system with broadcast signals that cover the entire cell, as in a Single/Multiple
Cell setting, the detection and reception of a pilot signal is sufficient to get some
estimate of the csi. But the type of csi might depend on the bandwidth of the
pilot signal. To be able to exploit the full reciprocity, the pilot signal might be re-
quired to have a relatively large bandwidth, and the allocation of wideband slots
could be expensive when only a fraction of the slots are used due to low load on
the network. With a lower bandwidth of the pilots, it is possible that the bs only
can make an estimate of in which direction the transmitting ue is located, and
can thus only utilize the directional reciprocity of the channel.
In a beam sweeping system, the pilot slot does not necessarily have to have a
large bandwidth. Consider a case where pilot slots are allocated directly after
broadcasts of sip2 in each beam. Then the reception of a pilot indicates that a
ue could not decode the sip2 that was broadcast in the beam that the pilot slot
corresponds to. Figure 4.2 illustrates that if a pilot signal is detected in the striped
pilot slot, the bs knows that there exists a user in beam 2 that could not decode
sip2 , and thus that it should make the dedicated sip2 -transmission in that beam.
Beam N sip2
..
.
Pilot slots
Beam 2 sip2
Beam 1 sip2
Estimate
channel
Dedicated sip2
T1 T2 T3
Broadcast sip1
Broadcast sip2
Pilot slots
sip2 on demand
that for some reason needs si again, perhaps due to outdated information being
stored.
One important thing to note is that the value of pf is actually only used in expres-
sions concerning sinr. In expressions such as arrival rates and latency probabil-
ities, we will only consider those users that are actually covered by the network.
It is not interesting to look at the users that will not be able to take part in any
communication with the system at all.
In Sections 4.3.1, 4.3.2 and 4.3.3 we present different aspects of the model that is
used for the new approach. In Section 4.3.4 we present the model for the classical
approach.
Table 4.1: The sizes of different signals are constant in this thesis.
tion based on the estimated channel to the user. This will be further described in
Section 4.3.3.
Note that this model also is valid if sip1 and/or sip2 is split into different seg-
ments that are sent with different intervals as with the system information in lte
(see Section 2.2). The only thing that is important is the total rate of the signal
that is required.
A user that arrives to the network is required to first search for a transmission of
sip1 . After this there is two different cases that are separated:
• Normal Detection - In this case the ue has to try to decode the next sip2
transmission before it is allowed to send a pilot signal.
A ue can as stated decode sip2 with probability pc . ues that need to send a pilot
signal do so in the next available pilot slot, and all subsequent slots until they
receive a dedicated sip2 .
Observe that the splitting property of the Poisson process entails that we can
write A (t) = A1 (t) + A2 (t). Where A1 (t) and A2 (t) are two independent Poisson
processes with intensities Ru pc and Ru (1 − pc ) respectively. Thus A1 (t) is the
arrivals of users which will be able to decode sip2 , and A2 (t) is the arrivals of
users that are not able to do so.
To model the latency for mathematical purposes we first need to define the ‘choice’
random variable Qp (Y1 ; Y2 ) to be a random variable that takes on the value Y1
with probability p and Y2 with probability 1 − p. We can then model the latency
for a user in the Normal Detection scenario as
LN = U1 + U2 + Qpc (0; U3 + T3 D) (4.1)
30 4 System Information Distribution Methods
Also note that the model disregards that the signals are not instantaneous, and
that it assumes that a dedicated sip2 can be received the moment the pilot is sent.
This approximation is motivated by that the total latency will most often be in
the order of seconds, while the transmission and processing times should be in
the order of milliseconds.
that are needed to calculate (4.3) through (4.6). It is easy to realize that
∞
X πk
E {D} = E {D|X = Sk } , (4.10)
1 − π0
k=1
i.e. it is the sum over all states where at least one user is transmitting a pilot, of
the probability of finding the Markov chain in that state, times the conditioned
expected value. The sum starts on 1 user as there will be at least 1 user when a
ue itself wants a dedicated sip2 . Furthermore, for k > 0 it holds that
∞
X
E {D|X = Sk } = tk · 0 + (1 − tk ) 1 +
pik E {D|X = Si } . (4.11)
i=1
This formula captures that if the ue is served, then it has to do zero retries, and
else it has to do one retry followed by a transition in the Markov chain’s state
space leading to the addition of the expected number of retries from that point.
In total, (4.11) gives an infinite linear system of equations for the conditioned
expected values, which can then be summed according to (4.10).
The values of Pr {D = d} can be retrieved by defining the following vectors and
matrix
t = (t1 , t2 , t3 , t4 , . . .)T , (4.12)
(1 − t1 )
(1 − t2 )
(1 − t3 )
T = (4.13)
(1 − t4 )
. .
.
that this is the user’s second pilot transmission. Thus, by recursively applying the
above argumentation we can obtain
Pr {D = d} = cd · t (4.18)
To be able to compute values of (4.10) and (4.18) with finite number of operations
it is necessary to limit the state space of the Markov chain to a maximum state
SN . This is a valid approximation if it can be argued that πi ≈ 0 for i > N , since
the Markov chain then almost never will be in those states that are removed. It
is also intuitive that it should be negative for the system to be in a state with a
too large number of users transmitting pilots. In this thesis, we chose to limit the
state space to the states S0 , S1 , . . . , S99 . This was deemed sufficiently large as sets
of parameters that entailed that πi 0 0 for some i > 25 were not performing very
well as compared to those that were more limited in their state space.
The only thing that needs to be specified in order to get values of P and ti is how
users are served dedicated sip2 . For this we consider a set of cases defined below.
Case 1: Always Serve All Users
The most simple case to consider is a case when all users that transmit a pilot
signal are served a dedicated sip2 . Some different possible realisations have been
identified that motivates this assumption:
(a) A system which does not do any beamforming, but broadcasts the dedicated
sip2 after a pilot slot where at least one pilot signal was received.
(b) If the system has a large amount of possible orthogonal pilot signals to
choose from, then the probability of ues choosing the same pilot is small.
If this probability is small enough, then we can approximate it with 0, thus
we make the assumption that ues never chooses the same pilot signal as
others. This is also the case if we use alternative 3 in Section 4.2.3, which
makes it possible for the bs to assign pilot signals to ues in a coordinated
manner.
In this case, the transition matrix of the Markov chain degenerates to a matrix
in which all the columns are (a0 , a1 , a2 , . . .)T . We can show by Foster’s criterion
(4.9) that this chain always has a stationary distribution for finite λ by letting
v (Sk ) = k and Γ = {S0 , . . . , SM } for M being the smallest integer larger than λ,
thus Γ is finite. Then for s ∈ Γ we have
X∞
E {v (X1 ) |X0 = s} = iai = λ < ∞ (4.19)
i=0
and for s < Γ
∞
X
E {v (X1 ) − v (X0 ) |X0 = sk } = (i − k) ai = λ − k ≤ λ − M < 0. (4.20)
i=0
We can again show by Foster’s criterion (4.9) that this chain has a stationary dis-
tribution under the condition that λ < 1 by letting v be the same function as in
the previous case, and Γ = {S0 }. Then for s ∈ Γ we have
∞
X
E {v (X1 ) |X0 = s} = iai = λ < ∞, (4.22)
i=0
and for s < Γ
∞
X
E {v (X1 ) − v (X0 ) |X0 = s} = (i − 1) ai = λ − 1 < 0. (4.23)
i=0
This case is also ’well-behaved’ and properties of it can be calculated with π and
the fact that ti = 1/i, as all transmitting ues have the same probability of being
served. Observe that we require λ < 1 for this case.
Case 3: K User Groups
In the third and final case, we consider a scenario where the set of pilot signals
are split into K groups, each representing that the received sinr from the bs
broadcast is within some interval. These intervals are chosen such that given that
a user cannot decode the broadcast sip2 , it has equal probability of being required
to send a pilot from each group. The bs is in this case assumed to be able to serve
at most one user per group, and it always serves at least one user for every group
in which there is sent a pilot.
Since users that belong to different groups do not interfere with one another, just
with the users that belong to the same group, this can be modelled as each user
4.4 Performance Metrics 35
group has a Markov chain with the same properties as for Case 2. The only
difference is that the user arrival rate to each Markov chain is scaled by K, i.e.
λ = Ru (1 − pc ) T3 /K. The K Markov chains are all equal and independent of each
other, so from a ue’s perspective, this scenario is the exact same as Case 2 with a
smaller λ. Thus the expressions in (4.10) and (4.18) are still applicable.
The model that is used for the reference method described in Section 4.1 is mainly
the same as described in the sections above. The bs is assumed to have a single
antenna capable of broadcasting and users arrive according to the same Poisson
process A (t). For simplicity of comparison, the si for the reference approach
is also split into sip1 and sip2 with broadcasting intervals T1 = 50 ms and T2
(variable) respectively, and also with the sizes N1 and N2 according to Table 4.1,
but both with full coverage pf .
The latency model is the same as before, but results in the simpler expressions
T +T
E {L} = 1 2 2 ,
L = U1 + U2 =⇒ (4.24)
Pr {L ≤ l} = Pr {U1 + U2 ≤ l} .
4.4.1 Bandwidth
Consider a scenario where the users’ sinr is a random variable distributed with
probability density function fG (g) and respective cdf FG (g). The probability dis-
tribution arises from that different users are located differently, and thus have a
different sinr on the large-scale. Then the reference method will need a band-
36 4 System Information Distribution Methods
width of
N1 /T1 N2 /T2
Br = + . (4.26)
−1 −1
log2 1 + FG 1 − pf log2 1 + FG 1 − pf
The first term is the bandwidth required for sip1 and the second term is the band-
width for sip2 . The terms are both derived from (4.25), with sinr such that pf of
the users have a larger sinr, thus they are able to decode the transmission.
In the first case we had two different realisations, which both gives a different ex-
pression of Bded . For realisation (a), where we broadcast the sip2 after a detected
pilot signal, we get
N2 (1 − π0 ) /T3
Bded = , (4.28)
−1
log2 1 + FG 1 − pf
since we make a broadcast each T3 seconds with probability (1 − π0 ).
For the realisation in (b) where the bs allocates the pilot signal to the ues in an
initial handshake, we instead will use
−1 (1−p )
FG c
∞ Z
X N2 /T3
Bded = πd fmin{H (1) ,...,H (d) } (h) dh, (4.29)
d=1 log2 1 + 30d h
−1
FG (1−pf )
where H is a random variable of user sinr given that the user needs a dedicated
sip2 , and fmin{H (1) ,...,H (d) } (h) is the probability distribution of the minimum of d
samples from H. This expression reflects that for each pilot slot, we transmit a
dedicated sip2 to d ≥ 1 users with probability πd , and that the bs link adapt to
the worst user. We are also assuming a static beamforming gain of 30 ≈ 15 dB,
which is split with the current number of users according to how sinr scales with
mrt described in Section 3.2.1. This gain is a perfect gain for 30 antennas, but
imperfect for the assumed number of 100 antennas which accounts for non-ideal
scenarios.
where fM|G=g (m) is the conditional probability density function of the beamform-
ing sinr gain experienced by a user with broadcasting sinr of g. This expression
is an integral over all users that need a dedicated sip2 and their possible beam-
forming gains, since all users will eventually be served individually. It assumes
that the bs is able to link adapt perfectly to the beamforming channel. This as-
sumption is motivated by that the set-up for a dedicated sip2 involves communi-
cation between the ue and the bs, which then can make estimates of the channel.
Case 3: K User Groups
This case has not been considered in this thesis, and no expression of Bded has
been worked out.
4.4.2 Latency
It is obvious that to minimize the bandwidth needed for si distribution, we would
like to have exceedingly large T2 and T3 . This is from a user perspective not
wanted due to the latency that will be experienced. Therefore we use latency as
the second important metric of a system. Latency can however be measured in
different ways, of which two are used in this thesis.
The first and most intuitive way of measuring latency is to take the mean latency
experienced by a large number of ues, which tends to the expected values given
by (4.3), (4.4) and (4.24) when the number of users grows towards infinity. This
metric considers all users’ experiences.
The second metric is to take a certain percentile’s latency, i.e. if we look at the
95th percentile latency, it is the minimum latency that at least 95 % of the ues lies
within. This metric gives a guarantee of that with some probability, a user will at
most experience a certain latency. Observe that this metric does not include the
latencies of the rest of the users at all. This metric can be calculated with (4.5),
(4.6) and (4.24).
5
Simulation
The results of this thesis are generated by the models of the system described in
Section 4.3 together with simulations. The simulations were performed in a state
of the art Ericsson-internal simulator. The collected data was then post-processed
and analysed in MATLAB.
This chapter presents the set up of the simulator, and what data was collected.
39
40 5 Simulation
and users that don’t, as well as users located both outdoors and indoors. The bss
each have 1 antenna per cell. Other parameters that influence the simulation are
summed up in Table 5.1.
This simulation was repeated 10 times to get a total of 10000 observations. Since
the data that is collected is the downlink sinr in a tdd system, the users do
not get any interference from other users, only from other bss. Thus it does not
matter how many users that are in each simulation.
The downlink sinr was also simulated in a Multiple Cell scenario. The set up
of this simulation was almost the same as in the Single Case scenario, but with a
larger hexagonal grid for the users to be placed in, according to Figure 5.2. The
reason for increasing the size of the grid is that the approximation used for sinr
calculation in this scenario, as described in Section 3.1.2, only considers signals
from far away bss as interference. In a small grid, the ues will almost always be
positioned such that all bss in the grid will count towards useful signal energy.
A larger grid thus makes the situation closer resemble the reality, to get realistic
data. This simulation used the values from Table 5.1 except that the number of
bss and cells were 37 and 111 respectively. And just as before, the simulation was
repeated 10 times with 1000 users each time.
5.1 SINR Simulations 41
Parameter Value
Number of simulations 10
Number of users per simulation 1000
Number of bss 7
Number of cells 21
Inter-Site Distance 500 m
Carrier frequency 2 GHz
Carrier bandwidth 5 MHz
bs maximum transmit power 20 W
Simulation time 5s
Start of data logging 1s
ue antenna Isotropic
Minimum distance from bss 35 m
In this chapter, we present all the results that the thesis has produced.
The corresponding cdf and data for the Multiple Cell scenario can be found
in Figure 6.2 and Table 6.2. Although the sinr is significantly larger for this
scenario, we can see the same characteristics in this cdf, i.e. that it has a long thin
tail which implies that lowering the coverage requirement just by a few percent
gives a large increase in sinr.
43
44 6 Results
0.8
0.6
CDF
0.4
0.2
0
-20 0 20 40 60
SINR [dB]
(a) Full plot of the empirical cdf.
0.3
0.25
0.2
CDF
0.15
0.1
0.05
0
-20 -15 -10 -5 0 5 10
SINR [dB]
(b) Closer look at the empirical cdf.
Figure 6.1: Empirical cdf of downlink sinr in the Single Cell scenario.
6.1 Simulation Results 45
Table 6.1: Values of the Single Cell sinr for some percentiles.
0.8
0.6
CDF
0.4
0.2
0
0 20 40 60
SINR [dB]
Figure 6.2: Empirical cdf of downlink sinr in the Multiple Cell scenario.
Table 6.2: Values of the Multiple Cell sinr for some percentiles.
0.6
CDF
0.4
0.2
0
-20 0 20 40 60
SINR [dB]
Figure 6.3: cdfs of sinr for broadcast and beamforming transmissions in
two different interference environments.
30
25
Beamforming gain [dB]
20
15
10
0
-20 -10 0 10 20 30 40
Broadcast SINR [dB]
Figure 6.4: The beamforming gain is correlated in some manner with the
broadcast sinr.
Parameter Value(s)
Ru 5
pc {0.75, 0.90, 0.99}
T3 1/ ((1 − pc ) Ru )
consequently so does the bandwidth required for sip1 broadcasts. The reason
behind the large difference in sinr between the two broadcasting methods is
that the Multiple Cell broadcasting both results in increased useful signal energy
at the users, as well as decreased interference as compared to the Single Cell
broadcasting.
Apart from the fixed size bandwidth, the classical approach only has one more
signal that requires bandwidth, and that is the broadcasts of sip2 . The proposed
approach has two further allocations; the pilot slots for possible pilot signals,
and the transmissions of sip2 upon request. We can see the impact they have on
the total bandwidth with a very simple plot in Figure 6.5. In this figure we have
calculated the bandwidth for both sub-cases of the most simple pilot model (Case
1), in which all users always get served a sip2 upon request. For this calculation
we have used Single Cell broadcasting, fixed parameters according to Table 6.4
and varied T2 . As can be seen in the table, T3 is chosen depending on pc . If we
had used a fixed T3 , then we know that almost the only difference in the resulting
bandwidth for different values on pc would be the difference in broadcast sip2
bandwidth. To make a fair comparison of different choices of coverage on the
sip2 broadcast, we choose T3 such that there will on average be one user in each
pilot slot. This also makes the assumption in the pilot model of being able to
serve all users, equally probable for all values of pc .
In Figure 6.5a we can see that the new method in this case is more efficient than
the baseline even for quite large T2 , even when using the simplest form of dis-
tributing sip2 on demand, broadcasting. When we are able to use beamforming
and link adaption as we have in Figure 6.5b we can gain even a little bit more,
and most prominently is that lower sip2 coverages perform significantly better.
In this simple case, we vary one parameter (T2 ), which leads to that only the
bandwidth required for sip2 broadcasts is varied in Figure 6.5. The other com-
ponents of the total bandwidth are constant for this case and are presented in
Table 6.5. This table shows that a very important factor to obtain a small band-
6.2 Method Comparison 49
60
Baseline
50
45
40
35
30
0 1 2 3 4 5
SIP2 interval [s]
60
55
75% SIP2 coverage
90% SIP2 coverage
Total Bandwidth [kHz]
45
40
35
30
0 1 2 3 4 5
SIP2 interval [s]
Figure 6.5: Simple bandwidth comparison using the simplest pilot model.
50 6 Results
Table 6.5: Values of constant bandwidth parts of the first simple case.
width is to not have to make transmissions with full coverage, exemplified by the
fact that for pc = 75% the bandwidth for dedicated sip2 decreases with four or-
ders of magnitude when we go from broadcasts to link adapted and beamformed
transmissions.
60 1
0.9
55
0.8
Normal Detection
Total Bandwidth [kHz]
SIP 2 coverage
Lower limit 0.6
45
0.5
40 Normal Detection
0.4
Early Detection
35 0.3
0.2
30
0.1
25 0
0 5 10 15 20 0 5 10 15 20
95th percentile latency [s] 95th percentile latency [s]
(a) Total bandwidth for all scenarios. (b) The optimized value of pc .
20 60
18
50
16
14
40
SIP 2 interval [s]
12
10 30
8
20
6
Normal Detection
4
Early Detection 10 Normal Detection
2 Early Detection
0 0
0 5 10 15 20 0 5 10 15 20
95th percentile latency [s] 95th percentile latency [s]
60 1
0.995
55
0.99
Normal Detection
50
Total Bandwidth [kHz]
SIP 2 coverage
Baseline
0.98
45 Lower limit
0.975
40
0.97
35 0.965
Normal Detection
0.96 Early Detection
30
0.955
25 0.95
0 5 10 15 20 0 5 10 15 20
95th percentile latency [s] 95th percentile latency [s]
(a) Total bandwidth for all scenarios. (b) The optimized value of pc .
25 30
15
15
10
10
5
5
0 0
0 5 10 15 20 0 5 10 15 20
95th percentile latency [s] 95th percentile latency [s]
0.98
0.96
CDF
0.94
0.92
0.9
0 20 40 60 80 100
Latency [s]
As shown, placing a requirement that states that a part of the users should have
limited latency might result in system properties that only considers that part
of the users, ignoring what happens to the others. It can then make sense to
consider a latency requirement that concerns the latency of all users, namely the
average latency. In Figure 6.9 and Figure 6.10 we have the same plots as before,
but with a requirement on the mean latency. The limit is chosen to be lower than
the limit of the percentage, since the mean latency should be lower than the 95th
percentile latency. The two figures uses the same values of Ru as before, one low
and one large.
These cases are very similar to when the percentage requirement was used in that
the new approach requires less total bandwidth than the classical approach, there
are some properties that hold under both the mean and the percentage require-
ment. In almost all cases the Early Detection scenario performs marginally better
than the Normal Detection, but that is barely visible in the graphs. Also common
for all cases is that the optimal coverage of sip2 seems to mostly be 90-99.5%. Fi-
nally, it seems that the new method approaches the lower limit a little bit faster
if Ru is low, implying that it performs a little bit better during lower user loads,
although the difference is small.
The reason to have a requirement on the mean latency was to avoid the prop-
erty that a part of the users had an extremely high latency. This is somewhat
improved by the mean requirement, but the cdf (not shown) again has a “knee”
as in Figure 6.8. The knee is however not nearly as sharp as in this figure, and it
is placed at a higher position, meaning that less users are affected and with much
less grave consequences.
Another solution to many users having too large latency could have been to in-
crease the percentage that should have a low latency. While not being presented
here, it is worth mentioning that we get similar result if putting the requirement
Pr {L ≤ l} ≥ 0.999.
56 6 Results
60 1
55
Normal Detection 0.8
Early Detection
Total Bandwidth [kHz]
50
Baseline
SIP 2 coverage
Lower limit 0.6
45
40 Normal Detection
0.4
Early Detection
35
0.2
30
25 0
0 2 4 6 8 10 0 2 4 6 8 10
Mean latency [s] Mean latency [s]
(a) Total bandwidth for all scenarios. (b) The optimized value of pc .
18 45
16 40
Normal Detection Normal Detection
14 Early Detection 35 Early Detection
12 30
SIP 2 interval [s]
10 25
8 20
6 15
4 10
2 5
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Mean latency [s] Mean latency [s]
60 1
55
0.99
Normal Detection
Early Detection
Total Bandwidth [kHz]
50
Baseline 0.98
SIP 2 coverage
Lower limit
45
0.97
40
25 0.94
0 2 4 6 8 10 0 2 4 6 8 10
Mean latency [s] Mean latency [s]
(a) Total bandwidth for all scenarios. (b) The optimized value of pc .
20 25
Normal Detection
Early Detection 20
Normal Detection
15 Early Detection
SIP 2 interval [s]
15
10
10
5
5
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Mean latency [s] Mean latency [s]
As all results in this section so far have been generated with the data from the
Single Cell broadcasting, it is also interesting to look at similar graphs using the
Multiple Cell sinr statistics. Unfortunately there has been no collection of data
of the beamforming gain that can be achieved compared to the broadcast sinr in
Multiple Cell, due to simulation limitations. We therefore assume together with
pilot Case 2 that all users experience exactly zero gain from beamforming. This
should therefore be seen as an underestimate, since beamforming increases the
sinr in general. This can equivalently be seen as that the dedicated sip2 are also
sent via Multiple Cell broadcasting but link adapted to the user being served,
and only decoded by that user. This assumption should lead to an overestimate
of the bandwidth needed for the new distribution method, but it still produces
the graph in Figure 6.11 for the mean requirement. It seems excessive to present
the resulting parameters and more cases, since the result is very similar to the
Singe Cell results, but with a lower total bandwidth. This figure is intended to
show that the results above hold also for the Multiple Cell case.
Baseline
6.5 Lower limit
5.5
4.5
3.5
0 2 4 6 8 10
Mean latency [s]
Figure 6.11: Minimized bandwidth for Multiple Cell when Ru = 10, with
the requirement that E {L} ≤ l.
Conclusion
7
This is the final chapter of the thesis, which discusses the thesis and its results,
both in detail and with a wider perspective. In the last section we present our
answers to the questions stated in Section 1.2.
7.1 Results
The results presented in Chapter 6 shows that the new approach of si distribution
provides a reduction of time-frequency resources needed as compared to only
broadcasting, with the system model that is used. The reason as to why it is
possible to distribute the same information with less resources originates in the
fact that broadcasting with full coverage is very expensive. As can be seen by the
simple comparison in Figure 6.5 there is a very large difference between having
to broadcast the requested sip2 and being able to adapt the transmission of each
sip2 to the users that should receive it.
The new approach exploits this by only broadcasting part of the si with full cov-
erage and the rest of the information is broadcast only to users with favourable
channels, and on demand to the rest. The gain of the new approach is that the
lower coverage on the sip2 broadcasts makes it possible to adapt the transmission
to a better channel. It shows that this gain is larger than the cost of adding allo-
cation of pilot slots for sip2 requests and dedicated sip2 transmissions. That the
gain is larger than the cost is, of course, dependent on the sizes N2 and N3 , but
they are chosen to reflect the expectation that can be put on a real system.
We analysed two different variations of the new approach, regarding how ues
determine whether they need to send a pilot signal to request a dedicated sip2 .
In almost none of the cases presented in Chapter 6 there is any visible difference
59
60 7 Conclusion
between the two different detections. It is only stated that the Early Detection
performs a little bit better in a majority of the cases. One would expect that
Early Detection should always perform better than Normal Detection, since it
only gives an improvement in latency without any additional cost. Those cases
in which Normal Detection performed better thus have to be the result of the
optimizer failing to find the optimal point. In total the detection scheme does
however not constitute a major difference in time-frequency resources.
When analysing the new method in more detail, it seems that there is two differ-
ent strategies of choosing parameters that fit different cases. The most common
way in the presented results is to cover most of the ues with the sip2 broadcasts
and only have a low load on the pilot slots. The other strategy is to not broadcast
sip2 at all and let all arriving users request it. The latter strategy has only been
applied in a minority of the cases; when the user arrival rate is low, and the la-
tency requirements are high. It makes sense that it actually is optimal to do so in
this case, as the strict latency requirement results in a low T2 if pc is larger than
zero, and as stated the broadcasts can be expensive. Instead, one can elect to have
T3 small so that each user that arrives to the network can send a request within
a small time interval. The first strategy is applicable for larger user arrival rates
and/or not as high requirements on the total latency. Furthermore, it looks as the
new approach converges to the classical approach when the latency requirement
disappears.
A concern that might be present is that it can seem plausible that the reason
that the new method is able to perform more efficiently is that the sinr that the
broadcasts has to be adapted to in the Single Case scenario (−12.944 dB) lies far
into the power-limited region of transmission. The channel capacity in (4.25)
then grows linearly with the sinr gain that is held from lowering the coverage
on broadcast sip2 . Figure 6.11 shows that this is not the case, but that the new
method is more efficient also for the Multiple Case scenario, when the sinr that
the broadcasts has to be adapted to is almost an order of magnitude larger. This
shows that the gain of the new method is much more dependent on the fact that
the cdfs of sinr have long and thin tails, than on the magnitude of the sinr at
the tail. The simulations performed in this thesis shows that this can be expected
to be the case for the sinr.
However, for best performance with the new si distribution method it is impor-
tant that the bs can make use of the information that the pilot transmissions sup-
plies. This is shown by Figure 6.5 where the most simple approach to dedicated
sip2 transmissions is broadcasting with full coverage. As can be seen, that incurs
a high cost that can be reduced greatly by being able to make some estimate of
the channel that the sip2 should be transmitted on. The estimate of the channel
could in practice consist of a link adaption to choose modulation and coding that
suits the channel, a beamforming precoder to direct the signal energy in space,
or most preferably both.
7.2 Method 61
7.2 Method
This thesis concerns a procedure in the not yet specified fifth generation’s mobile
telecommunication system. The simulator that was used in the thesis simulated
an lte network, and while it does support some simulations of some early realiza-
tion of 5G, it would require a considerable amount of work to be able to simulate
an arbitrary si distribution protocol. We have therefore been forced to use an
abstract model of how si distribution works in a system. The validity of the re-
sults is dependent on the validity of the models being used and the simulated
statistics.
For all bandwidth calculations we have assumed that the bs is able to perform
perfect link adaption, i.e. use the full channel capacity. This is of course not the
case in reality, the link adaption will not achieve full capacity. We have then al-
ways calculated an underestimate of the actual bandwidth required, but as this is
done for both the classical and the new approach the errors should approximately
cancel each other out, and the comparison is still quite accurate.
The time-frequency allocation model that is used for the classical approach should
be very accurate, as it only consists of broadcasting with full coverage. The model
does make use of the two different system information parts which is unnecessar-
ily complicated, but it is equally valid as explained in Section 4.3.1. The corre-
sponding model for the new approach is more difficult to asses the accuracy of.
The parts of the si that are broadcast are as valid as for the classical method.
The pilot slot allocation is a representation of the overhead that is induced by
the transmission of pilot signals, the magnitude of it should be large enough to
provide useful channel estimates. The most difficult part in the model of the
new approach is the bandwidth required for the dedicated transmissions of sip2 ,
where we have three different cases:
- In Case 1(a) the bs broadcast the sip2 on demand which is perfectly func-
tional, but not efficient as shown in the results.
- Case 1(b) is assumed to be able to serve any amount of users, with link
adaption for the worst user, and beamforming that splits an imperfect gain
for 100 antennas equally on the users. This relies on the assumption that
no users will transmit the same pilot signal, which is not possible for large
amount of users, since the number of orthogonal pilot signals that can be
created is limited. This model is furthermore very simplified but since it
is only used with an aim of having only one ue per pilot slot, the model is
quite reasonable and should suffice for our purposes.
- Case 2 limits the amount of users that can be served at once to 1, selecting a
random pilot transmitting user and adapting the sip2 beamforming to that
user. That all users are chosen with the same probability might be valid
if all users choose different pilot signals, and all of those can be detected
at the bs. Due to the limitation to 1 served user per pilot slot, there will
not be much more than one user transmitting a pilot, and the probability of
62 7 Conclusion
- Case 3 was only presented from a functional point-of-view; that the avail-
able pilot signals were split into K groups, which signals the downlink re-
ception quality of the broadcasts. It was argued that this is from a latency
point-of-view equal to dividing λ, the arrival rate of users needing dedi-
cated sip2 , with K. Thus we can support total user arrival rates of a factor
K larger than in Case 2. An expression for the bandwidth required for this
was not calculated, but one might suspect that it should be similar to both
Case 1(b) and 2.
Overall, these cases contain some approximations that we consider quite well mo-
tivated, but one could always improve the model. Regarding approximations of
the Markov chains, they have actually been verified experimentally by simulating
the actual chains and comparing probability distributions with the approximated
ones. For model Case 1(b) and 2 the total bandwidth that is allocated for dedi-
cated sip2 transmissions is quite small, and thus approximations should not have
a major impact on the final results.
The latency model used calculates latency as the time from arrival to the time
of the start of successful sip2 reception. For the users in the new approach that
are not covered by the sip2 broadcasts, the final time is taken as the start of pilot
transmission. This is not completely realistic as there should be some delay be-
tween the pilot transmission and the reception of sip2 . The results presented will
be a bit altered if this delay is large, but otherwise the impact should be minor.
In total, the models that are being used are not excessively complicated for the
scope of the thesis, but they still should be accurate such as to be able to conclude
that there is a gain to be had by using the new distribution method. Furthermore
there are two different strategies for choosing the coverage of sip2 , each one opti-
mal for different settings and systems.
As this thesis has shown that there is a more efficient method of distributing
system information in a system, it should be considered in the 5G development
to include a similar procedure. The system would benefit from it in that it both
opens up more resources to be used for user data, and should induce a lower
energy consumption for the si distribution part. In total this should contribute
to a network that is cheaper to run, and that could have a larger total throughput
than compared to a network that only broadcasts the si. From an environmental
point-of-view, it is of also course beneficial to minimize the energy consumption.
Since the thesis is quite theoretical, there is no ethical implications of the thesis
that has to be addressed.
7.4 Summary 63
7.4 Summary
The aim of the thesis was to propose at least one brand new method of si distribu-
tion and evaluate it. We presented two different methods, one reference method
that has been used classically, and one new. With the analysis that has been done
of these two methods, we now answer the questions posed in the problem formu-
lation in Section 1.2.
1. How can massive mimo be utilized for system information distribution?
A system that employs massive mimo can use the large antenna arrays for
system information distribution in more than one way. As stated in Sec-
tion 3.1, using a certain precoding has been shown to improve the broad-
casts. This is something that definitely should be considered for a system,
but that was not included in this thesis. We have instead shown how mas-
sive mimo can aid the distribution of si by beamforming. For the bs to
be able to make user-specific beamforming, some csi is required at the bs,
which in turn requires some initial communication between the ues and
the bs. The new approach lets some users request the sip2 in some way, and
this request can be exploited to estimate some form of csi. Without any csi
the bs would not be able to beamform other than in random directions. It
was shown in this thesis that while beamforming can be used for transmis-
sions of dedicated sip2 , which improves the efficiency, it is not vital to the
new distribution method to do so. Even when no beamforming gain was as-
sumed, the new approach could perform better than the classical approach.
However, if one has the possibility to beamform, then there is no reason
from a time-frequency resource perspective to not beamform.
2. How efficient are the proposed solutions with respect to time-frequency
resources?
The new proposed method is in this regard more efficient than the classical
method in a large range of cases. However, the total bandwidth required is
very dependent on the bandwidth required for broadcasts of sip1 , which is
the same for the two methods. As shown, there might be a very large gain in
distributing sip1 with Multiple Cell broadcasts, as it increases the downlink
sinr for all users. Another thing that could decrease this bandwidth would
be to minimize the information that is included in sip1 to make it as small
as possible. For the information that is not included in sip1 , we have shown
that it is possible to reduce the resources needed with a new protocol.
3. What drawbacks or limitations exists in the proposed solutions (e.g. latency
issues)?
The explicit case of latency was dealt with in the thesis, and while the new
method induces a larger overall latency for a fixed set of parameters, the
parameters can be chosen to perform better than the classical method, with
the same latency properties. A drawback of the new method is however
that it is more complex in that it needs to be able to handle users that are
64 7 Conclusion
not covered by sip2 broadcasts. In this thesis we have always assumed the
probability of two ues transmitting the same pilot being zero. This is not
the case when there is no handshake before pilot, and even when there
is, there is a probability of a collision in the handshake instead. This is
something that a realisation has to take into account, and be able to resolve.
The random access procedure of lte shows one way of dealing with such
collisions, and something like that is thus required in the new method.
65
66 LIST OF FIGURES
67
Bibliography
[1] Cisco, “Cisco visual networking index: Global mobile data traffic
forecast update, 2015-2020,” 2016. [Online, 2016-02-18]. Available:
http://www.cisco.com/c/en/us/solutions/collateral/
service-provider/visual-networking-index-vni/
mobile-white-paper-c11-520862.html. Cited on page 1.
[2] 3GPP, “Tentative 3GPP timeline for 5G,” 2015. [Online, 2016-02-
23]. Available: http://www.3gpp.org/news-events/3gpp-news/
1674-timeline_5g. Cited on page 1.
[5] J.-G. Remy and C. Letamendia, LTE Standards. ISTE Limited, John Wiley &
Sons, 2014. Cited on page 5.
[6] M. Engels, Wireless OFDM Systems: How to Make Them Work? Kluwer
Academic Publishers, 2002. Cited on page 7.
69
70 Bibliography