You are on page 1of 82

Ram Sunil Kanumalli

Investigation of Carrier Frequency Oset


Estimation Techniques for Unique Word
(UW) - OFDM Systems
Masterarbeit
zur Erlangung des akademischen Grades
DiplomIngenieur

Studium Information Technology



Alpen-Adria-Universitt Klagenfurt
Fakultt fr Technische Wissenschaften

Begutachter: Univ.Prof. Dr. Mario Huemer


Betreuer 1: Dipl.Ing. (FH) Christian Hofbauer
Betreuer 2: Univ.Ass. Dipl.Ing. Alexander Onic

Institut fr Vernetzte und Eingebettete Systeme


Eingebettete Systeme und Signalverarbeitung
Klagenfurt, im September 2012
Declaration of Honor

I hereby conrm on my honor that I personally prepared the present academic work and
carried out myself the activities directly involved with it. I also conrm that I have used
no resources other than those declared. All formulations and concepts adopted literally
or in their essential content from printed, unprinted or Internet sources have been cited
according to the rules of academic work and identied by means of footnotes or other
precise indications of source.
The support provided during the work, including signicant assistance from my super-
visor has been indicated in full. The academic work has not been submitted to any other
examination authority. The work is submitted in printed and electronic form. I conrm
that the content of the digital version is completely identical to that of the printed version.
I am aware that a false declaration will have legal consequences.

(Unterschrift) (Ort, Datum)

i
Abstract
Unique Word (UW) - Orthogonal Frequency Division Multiplexing (OFDM) is a novel
OFDM signalling technique [1], where the usual cyclic prexes are replaced by determin-
istic sequences called UW. It has proven to be an appealing alternative to conventional
CP-OFDM because of its predominant features. However, it also has to overcome sev-
eral technical challenges. The most well-known issue among those is carrier frequency
oset (CFO) caused due to the mismatch of the oscillator's frequency and the Doppler
shifts between the receiver and the transmitter. A small frequency oset will destroy the
orthogonality among the subcarriers leading to a signicant degradation in the system
performance. Hence, it is necessary at the receiver to estimate and compensate the CFO
accurately and keep its impact at a minimum.
One major aim of this work is the investigation of various CFO estimation techniques
for an UW-OFDM system. A comparison is carried out which evaluates the performance
of time domain against frequency domain based estimators. In the time domain, UWs are
utilized for estimating the CFO, whereas in the frequency domain, the data symbols or
dedicated pilot tones are used. For the latter, a modied version of the UW-OFDM signal
generation is derived in this work to enable the insertion of pilot symbols in the frequency
domain. Eects of various parameters like noise, channel characteristics, UW length, and
the amount of CFO on the performance of the estimators are evaluated. The performance of
the CFO estimators is analyzed by the resulting mean squared error. The signalling concept
of UW-OFDM itself oers a unique opportunity to exploit the inherent redundancy caused
by introducing UWs in the time domain by applying appropriate data estimators like e.g.,
the LMMSE (Linear Minimum Mean Square Error) estimator or the BLUE (Best Linear
Unbiased Estimator). Due to that, the inuence of noise on the frequency domain symbols
can be substantially reduced. Therefore, as a proof, the performance of the frequency
domain estimators utilizing the frequency domain pilot and data symbols after equalization
with dierent possible equalizers is analyzed and compared.
Additionally, the sensitivity of the UW-OFDM system on the CFO is analyzed and
compared to that of the CP-OFDM scheme. Furthermore, the dependency of the perfor-
mance on the specic design of the UW in the presence of CFO is investigated.

ii
Acknowledgments
I take this opportunity to express my deepest gratitude to my professor Dr. Mario Huemer
for allowing me to work on this thesis, providing nancial support, and the wholehearted
condence he has shown in me. I am also very thankful for his scientic advices and enthu-
siastic support. I nd myself very fortunate to have been a part of his group. His impact
on me has reached far beyond this thesis, which has helped me to shape my career in a
very positive way.

I am highly indebted to my supervisor Mr. Christian Hofbauer for his invaluable guid-
ance and for monitoring my progress despite the distance. His eorts and time on revising,
discussing, and rewriting the draft manuscripts have resulted in the successful completion
of this thesis. His advice and support given throughout this thesis is unforgettable. I have
learnt how to look and analyze a research problem, as a result of countless intermediate
discussions with him. I also take immense pleasure in thanking Mr. Alexander Onic who
has been a source of encouragement and advice. He is the person who helped me in get-
ting into the topic and laid foundation of this thesis with his strong support and interest.
Finally, yet important, I would like to express my heartfelt thanks to my family for their
endless love and blessings, and to my friends for their help and support throughout my
Masters.

iii
Contents

1 Introduction 1
2 Review of UW-OFDM 4
2.1 Generation of UW-OFDM Baseband Signal . . . . . . . . . . . . . . . . . . 5
2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Multipath Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Additive White Gaussian Noise . . . . . . . . . . . . . . . . . . . . . 10
2.3 Receiver Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.1 Classical Data Estimators . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Linear Bayesian Data Estimators . . . . . . . . . . . . . . . . . . . . 13

3 Carrier Frequency Oset 15


3.1 CFO Eect in Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 CFO Eect in Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . 20

4 The Pilot Based UW-OFDM Transceiver Model 22


5 Carrier Frequency Oset Estimation 26
5.1 Acquisition Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2 Tracking Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2.1 Time Domain CFO Estimation . . . . . . . . . . . . . . . . . . . . . 30
5.2.2 Frequency Domain CFO Estimation . . . . . . . . . . . . . . . . . . 34

6 Results and Discussion 38


6.1 Simulation Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.2 AWGN Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.3 Time-Dispersive Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.4 CFO Impact on UW-OFDM and CP-OFDM . . . . . . . . . . . . . . . . . . 63

7 Conclusions 71
Bibliography 74

iv
Chapter 1

Introduction
In recent years, Orthogonal Frequency Division Multiplexing (OFDM) has become an im-
portant modulation technique in the eld of telecommunications [2]. The receiver design
architecture of OFDM is simple, providing a very ecient and simplied way of equalizing
the eects caused by frequency-selective multipath channels [3]. OFDM has been imple-
mented in Europe for Digital Audio Broadcasting (DAB) and Digital Video Broadcasting
(DVB) systems [4], in Japan for Multimedia Mobile Access Communications (MMAC),
and in IEEE 802.11 and 802.16 Wireless standards [5]. Currently, a major research is
being done on implementing the OFDM technique in cellular mobile communication sys-
tems [6]. The most conventional approach to implement the OFDM technique is Cyclic
Prex OFDM (CP-OFDM), where the last samples of each OFDM symbol are copied
into the guard interval [7]. Some alternative approaches to the conventional CP-OFDM
have been proposed in [1, 8, 9]. In [8], a similar structure to CP-OFDM, the known symbol
padded (KSP)-OFDM, where the random CP is replaced by known symbol (KS) sequences,
was introduced. In case this known sequence is set to zero, KSP-OFDM coincides with
the zero padded OFDM (ZP-OFDM) [9]. In [1], a new OFDM signalling scheme, where
the usual cyclic prexes are also replaced by deterministic sequences was proposed. This
deterministic sequence is often addressed as Unique Word (UW) and the scheme is referred
to as UW-OFDM. The most important dierence between KSP- and UW-OFDM is the
fact, that the UW is part of the DFT interval, whereas the KS is not. The insertion of
the UW within the DFT interval requires to introduce some correlations in the frequency
domain, which can advantageously be exploited by the receiver to improve the bit error
ratio (BER) performance [10], whereas the KSP-OFDM does not feature these correlations.

UW-OFDM has proven to be an eective signalling technique because of its many ap-
pealing features: 1) The idea of substituting CPs by UWs in OFDM has the advantage
of choosing the UW sequence such that it can eectively be used for synchronization (i.e.,
timing oset estimation and CFO estimation) and channel parameter estimation purposes,
thereby inherently serving as a pilot sequence [11]. 2) The redundancy present in the fre-
quency domain can be exploited to reduce the inuence of noise on the data, thus leading
to a substantial improvement over the CP-OFDM in terms of data estimation [1].

Despite these favorable features, it also has to overcome several technical challenges.
The most well-known issue among those is CFO caused due to the mismatch of the os-

1
CHAPTER 1. INTRODUCTION 2

cillators and the Doppler shifts between the receiver and the transmitter. Our simulation
results show that, similar to CP-OFDM, the UW-OFDM is also highly sensitive to the
CFO. A small frequency oset induces two degrading eects: First of all, the orthogo-
nality among the subcarriers is destroyed which leads to Inter Carrier Interference (ICI),
resulting in a signicant degradation of the overall BER performance. Secondly, the signal
is attenuated and rotated [12]. Therefore, it has become a key challenge at the receiver
to estimate and compensate the CFO accurately and minimize its impact. For the past
fteen years there has been a lot of research carried out based on the subject of CFO esti-
mation in CP-OFDM. A good overview on various CFO estimation algorithms can be found
in [1322]. The CFO estimation for CP-OFDM is well investigated and many sophisticated
algorithms have been proposed. However, to the best of my knowledge, CFO estimation
in UW-OFDM has not been investigated till now. In order to make the UW-OFDM to
be carried on a success line, it is necessary to focus on the CFO issue. This thesis mainly
deals with the investigation of various algorithms for CFO estimation, which can exploit
the information present in the UW. In addition, a comparison is made between the time
domain based CFO estimation utilizing UWs and frequency domain based CFO estimation
utilizing traditional pilot tones. Besides this comparison, the sensitivity of these estimators
on the CFO and the channel characteristics is investigated. Furthermore, a comparison of
UW-OFDM and CP-OFDM regarding their sensitivity to CFO is carried out.

The remaining work is organized as follows: Chapter 2 introduces the UW-OFDM


baseband signal generation and lists the advantages of using this scheme over conventional
CP-OFDM. Furthermore, a system model which includes the transmitter, the channel, and
the receiver signal processing is also described in the same chapter. Chapter 3 presents
the problem of CFO and also discusses the eects of CFO in time and frequency domain.
Chapter 4 describes the modelling of the pilot based transceiver model which is used in
this thesis. Chapter 5 provides an overview of various CFO estimation techniques for UW-
OFDM, which are in fact derived from currently available CFO estimation algorithms for
CP-OFDM. As the pilot based UW-OFDM model will include a pilot sequence in both time
(i.e., as UWs) and frequency (i.e., as pilot tones) domain, the time domain (cycliy prex-
based) methods as well as the frequency domain (pilot tone aided and decision-directed)
methods are discussed. Chapter 6 compares the performance of various CFO estimation
techniques for UW-OFDM that are described in the previous chapter. In addition, the
sensitivity of the UW-OFDM system to the CFO, is analyzed and compared to that of the
CP-OFDM scheme. Finally, Chapter 7 concludes this work.
CHAPTER 1. INTRODUCTION 3

Notation
The following notations are used throughout this work. Bold face letters with lower
case represent vectors (e.g., x, y, ...), and bold face letters with upper case indicate matri-
ces (e.g., X, Y,...). Conjugation, conjugate transposition, and transposition are denoted
H T
by () , () , and () , respectively. Frequency domain variables are represented by plac-
ing a tilde above the variable (e.g., x ), which helps for dierentiating time domain
, X
and frequency domain variables. The expectation operator is denoted by E {}. Xi,j de-
notes the element in the i
th row and j th column of the matrix X. The inverse of the

matrix X is denoted by X
1 . I and 0 denote an n n identity and zero matrix, respec-
n n
tively. X = diag {x1 , x2 , , xn } represents a diagonal matrix with the diagonal elements
x1 , x2 , , xn . The set of complex numbers is denoted by C. Finally, arg {} and || denote
argument and magnitude, respectively, while Re {} and Im {} correspond to the real and
imaginary parts of a complex number.
Chapter 2

Review of UW-OFDM
The concept of adopting an UW as an alternative to CP, cf. [23], [11], was rst employed
in single-carrier with frequency domain equalization (SC/FDE) systems. Here, the intro-
duction of UWs is a straightforward procedure, because the data symbols as well as the
UWs are dened in time domain. The idea of substituting CPs by UWs has the advantage
of choosing the UW sequence such that it can eciently be used for synchronization (i.e.,
timing oset estimation and CFO estimation) and channel parameter estimation purposes,
thereby inherently serving as a pilot sequence. Some relevant studies addressing the issues
of CFO tracking and channel estimation using the UW for SC/FDE systems can be found
in [11, 2326].

The UW concept was then successfully adopted to OFDM, as recently proposed in [1].
However, the introduction of UWs in OFDM is not straightforward, as the data symbols
are dened in frequency domain, whereas the UW is dened in the time domain. A good
comparison between UW-OFDM and UW-SC/FDE is given in [27]. One advantage of
the UW-OFDM scheme over conventional CP-OFDM is that the UW itself serves as a
pilot sequence, thus avoiding the need of pilot carriers. Fig. 2.1 reveals the dierences
between the conventional CP-OFDM and the UW-OFDM transmit symbol structures.
From Fig. 2.1, some key dierences between an UW and a CP based OFDM system can

Figure 2.1: Transmit symbol structures for (a) CP-OFDM and (b) UW-OFDM [1].

be pointed out:

The UW lies inside the Discrete Fourier transform (DFT) window, while the CP lies
outside the DFT interval.

4
CHAPTER 2. REVIEW OF UW-OFDM 5

The CP is based on the transmitted data and thus random. Consequently, it varies
from one OFDM symbol to the other. The UW, on the other hand, is deterministic
and therefore the same for all OFDM symbols.

2.1 Generation of UW-OFDM Baseband Signal


In this section, the generation of an UW-OFDM baseband signal is explained. Similar to
the conventional OFDM systems (e.g., CP-OFDM), a vector of complex QAM/PSK data
symbols CNd 1 ,
d which are dened in frequency domain, is considered. In general, zero
subcarriers are inserted at the DC position and at the band edges. Let the number of zero
subcarriers to be inserted be Nz . After inserting these zero subcarriers, the OFDM symbol
in the frequency domain CN 1 can be written as
x

x
= Bd. (2.1)

B CN Nd contains zero row vectors at the positions of zero subcarriers, unit row vectors
at the appropriate positions of the data subcarriers, and N represents the DFT window
length. The time domain OFDM symbol x CN 1 is calculated as

x = F1
N x, (2.2)

j2kl
where FN represents the N-point DFT matrix with [FN ]k,l = e N and k, l = 0, 1, , N
1. The above explained procedure is valid for any OFDM system. However, in UW-OFDM,
a deterministic sequence called UW xu CNu 1 is introduced at the end of each time do-
T
x = xT xT xd C(N Nu )1

main OFDM symbol. This can be formulated as d u . Here,

denotes the vector containing the random time domain samples aected by the data d.
Fig. 2.2 illustrates the generation of an UW-OFDM symbol. In the following, this genera-
tion process is explained in detail.

To generate an UW-OFDM symbol with the desired properties, it is shown in [28] that
a two step approach is benecial, or else the symbol energy will almost explode.
Step 1: Generate an OFDM time domain symbol with a zero UW such that x =
T
xT 0T , where F1

d x can be obtained from the equation x = N x. Note that the vector x

here is dierent from the one dened in (2.1). The construction of x
is explained shortly
below.
Step 2: Now, add the desired UW sequence xu to
T
the vector x to obtain the nal UW-
0  T T
OFDM time domain symbol x =x+ 0 xu .

In the following, the implementation of step 1 is explained in detail: In order to produce


a zero UW in the time domain, along with the data subcarriers and the zero subcarriers a
new set of subcarriers called redundant subcarriers r = CNr 1 is introduced in frequency

domain. Furthermore, the number of redundant subcarriers Nr is set to Nu (i.e., the length
1
of the UW xu ). As a result, to solve the system of equations x = FN x , the number of
data carriers Nd has to be reduced by Nr , such that N = Nd + Nr + Nz . One idea would
be to take the rst Nd subcarriers and load it with data symbols, and then take the next
CHAPTER 2. REVIEW OF UW-OFDM 6

~
d
~
d ~
x x

QAM Mapping

Serial/Parallel

OFDM Symbol, P, B
Binary data

Parallel/Serial
x

Assemble

IFFT
~r 0
Redundant ADD
Subcarriers Unique Word
Calculation, T 0
xu

Figure 2.2: Block diagram of the time-discrete baseband representation of an UW-OFDM


transmitter.

Nr subcarriers and load it with redundant symbols. The frequency domain symbol would
then be constructed as

 
d
x
=B . (2.3)

r
However, this in not recommended for practical systems. It turns out that these redundant
subcarriers should be placed on optimal locations over the available bandwidth (for alloca-
tion, refer to [1]), or else the energy contribution of the redundant subcarriers on the mean
symbol energy is very high compared to the data symbols, which is not desired. In order to
distribute the data and the redundant subcarriers to their respective locations, a permuta-
tion matrix P C(Nd +Nr )(Nd +Nr ) is introduced. The matrix P is selected in such a way
that the mean energy of the redundant subcarriers is minimum. Therefore, P optimally
permutes the data and the redundant subcarrier positions, cf. [1]. An optimal permutation
matrix for a specic setup will be given in chapter 6. After the distribution of the data
and the redundant subcarriers, the zero carriers are inserted with B CN (Nd +Nr ) . This
can be seen in the following equation:


 
d
x
= BP . (2.4)

r

P and B matrices are shown. For that


For a better understanding, examples of generating
an exemplary system setup with a DFT size of N = 16, UW length Nu = 4, Nd = 8 data
subcarriers, Nr = Nu = 4 redundant subcarriers, and Nz = 4 zero subcarriers is chosen.
One zero subcarrier is positioned at DC, and the others at the band edges. The data
symbol vector
d is dened as

T
=

d d0 d1 d2 d3 d4 d5 d6 d7 , (2.5)

and the redundant symbols vector r as


 T
r = r0 r1 r2 r3
. (2.6)

Choosing the indices of the redundant subcarriers to be {1, 5, 10, 12}, and inserting the
CHAPTER 2. REVIEW OF UW-OFDM 7

zero subcarriers at DC (i.e., 0) and at the band edges (i.e., 7, 8, 9), the frequency domain
OFDM symbol x
can be written as

 T
x
= 0 r0 d0 d1 d2 r1 d3 0 0 0 r2 d4 r3 d5 d6 d7 . (2.7)

Consequently, (2.7) can easily be obtained by appropriate matrices P and B, respectively.


As a rst step, P C1212 shall permute the positions of the data and the redundant
subcarriers such that


 
d
P
 T
= r0 d0 d1 d2 r1 d3 r2 d4 r3 d5 d6 d7 . (2.8)

r

P can easily be determined and (2.8) can be written as


r0 0 0 0 0 0 0 0 0 1 0 0 0 d0

d0
1 0 0 0 0 0 0 0 0 0 0 0
d1


d1
0 1 0 0 0 0 0 0 0 0 0 0
d2


d2
0 0 1 0 0 0 0 0 0 0 0 0
d3


r1
0 0 0 0 0 0 0 0 0 1 0 0
d4

d3 0 0 0 1 0 0 0 0 0 0 0 0 d5

= . (2.9)

r2 0 0 0 0 0 0 0 0 0 0 1 0 d6


d4 0 0 0 0 1 0 0 0 0 0 0 0 d7



r3

0 0 0 0 0 0 0 0 0 0 0 1

r0



d5 0 0 0 0 0 1 0 0 0 0 0 0 r1

d6 0 0 0 0 0 0 1 0 0 0 0 0 r2
d7 0 0 0 0 0 0 0 1 0 0 0 0 r3
P1212
| {z }

The next step is to insert the zero subcarriers by applying a matrix B C1612 such that

 T
x
=B r0 d0 d1 d2 r1 d3 r2 d4 r3 d5 d6 d7 . (2.10)
CHAPTER 2. REVIEW OF UW-OFDM 8

With an appropriate matrix B, (2.10) immediately follows to


0 0 0 0 0 0 0 0 0 0 0 0 0

r0
1 0 0 0 0 0 0 0 0 0 0 0


d0
0 1 0 0 0 0 0 0 0 0 0 0
r0

d1
0 0 1 0 0 0 0 0 0 0 0 0
d0


d2
0 0 0 1 0 0 0 0 0 0 0 0
d1


r1
0 0 0 0 1 0 0 0 0 0 0 0
d2


d3
0 0 0 0 0 1 0 0 0 0 0 0
r1

0 0 0 0 0 0 0 0 0 0 0 0 0 d3

= . (2.11)

0 0 0 0 0 0 0 0 0 0 0 0 0 r2


0 0 0 0 0 0 0 0 0 0 0 0 0 d4



r2

0 0 0 0 0 0 1 0 0 0 0 0

r3



d4 0 0 0 0 0 0 0 1 0 0 0 0 d5


r3
0 0 0 0 0 0 0 0 1 0 0 0
d6

d5
0 0 0 0 0 0 0 0 0 1 0 0
d7
d6 0 0 0 0 0 0 0 0 0 0 1 0
d7 0 0 0 0 0 0 0 0 0 0 0 1
xe161
| {z } | {z }
B1612

Up till now it has only been established that redundant subcarriers are introduced in
order to fulll

   
d xd
F1
N BP = . (2.12)

r 0

Equation (2.12) represents the frequency and the time domain relation of the UW-OFDM
symbol with zero UW. The redundant subcarrier symbols have to be determined now such
that the zero UW is produced in the time domain. For that let M CN (Nd +Nr ) be
dened as
M11 | M12
M = = F1
N BP, (2.13)
M21 | M22

whereas M11 C(N Nr )Nd , M12 C(N Nr )Nr , M21 CNr Nd , and M22 CNr Nr
(note that Nr = Nu ). Considering (2.12) and (2.13), this leads to




 M11 | M12    
d d x
M r = r = 0 . d
(2.14)
M21 | M22

From (2.14) it immediately follows that

+ M22
M21 d r=0 (2.15)

and consequently
r = M1

22 M21 d. (2.16)
CHAPTER 2. REVIEW OF UW-OFDM 9

Dening T = M1
22 M21 (T C
Nr Nd ), (2.16) can then be written as


r = Td. (2.17)

From (2.17) it becomes obvious that the vector


r is just a linear function of the data .
d
Therefore, (2.12) can be expressed as


     
d I x
d
x= F1
N BP = F1
N B
=
d . (2.18)

r T 0

It is worth to mention that the data subcarriers in combination with the redundant
subcarriers form a systematic complex valued codeword c C(Nd +Nr )1
given by


   
d I = Gd,


c=P =P d (2.19)

r T

whereas G C(Nd +Nr )Nd can be interpreted as the code generator matrix. G introduces
correlations among the frequency domain vector x
, which can advantageously be used to
improve the bit error rate (BER) performance, cf. [10].

In order to achieve the nal OFDM transmit signal, the time domain UW xu is added
to the zero-UW OFDM symbol x obtained in (2.18). Hence, the transmit UW-OFDM
0
symbol x follows to
0 T
x = x + 0T xT

u . (2.20)

0
The time domain vector x can also be written as

0
x = F1
N (
x+x
u ) , (2.21)

where u CN 1
x represents the frequency domain version of the time domain UW xu
dened as
T
u = FN 0 T x T

x u . (2.22)

From this it can be noticed that, although the UW in time domain is appended at the
end of the data, in frequency domain it spreads across the entire spectrum, inuencing the
data
d as well.
0
By using (2.18), (2.19), and (2.21), the time domain UW-OFDM baseband signal x can
be formulated as  
0
x = F1
N
+x
BGd u . (2.23)

In the next sections, the system model including the channel and the receiver side is
described in detail.
CHAPTER 2. REVIEW OF UW-OFDM 10

2.2 Channel Model


In a time dispersive environment, the receiving antenna receives multiple copies of the
transmitted signal from dierent transmission paths, where each path shows dierent prop-
agation delays and attenuations. The two fundamental parameters of a multipath channel
are the coherence time and the coherence bandwidth, which are inversely proportional
to the Doppler spread and the delay spread, respectively [29]. Depending on the symbol
period and the bandwidth of a signal, the fading is categorized into four types. When
the signal bandwidth is considerably larger than the coherence bandwidth of the channel,
the channel is considered as frequency-selective, otherwise, as frequency-at. If the burst
period is larger than the coherence time of the channel, then the channel is time-varying,
otherwise time-invariant. In this thesis, the channel is always assumed to be frequency-
selective and time-invariant.

2.2.1 Multipath Propagation

The multipath channel is modelled as a Finite Impulse Response (FIR) lter with the
Channel Impulse Response (CIR)

h = [h0 h1 h2 ... hNk 1 ]T , (2.24)

where h CNk 1 . Each delay tap hk is modeled as a complex Gaussian random variable
having zero mean. All the taps are considered to be statistically independent and having
an exponential decaying power prole [30]. In order to completely eliminate Inter Symbol
Interference (ISI), the length of the channel Nk and the UW length Nu must satisfy the
condition Nk Nu . The CIR is assumed to be constant for one OFDM burst (typically
a burst consists of a preamble and several OFDM symbols). This assumption seems rea-
sonable, because in many applications that employ OFDM, the coherence time is large
compared to the chosen OFDM burst duration.

2.2.2 Additive White Gaussian Noise

In addition to the multipath, the transmitted signal is also aected by Additive White
Gaussian Noise (AWGN). Each element in the noise vector n CN 1 is assumed to be
a zero-mean complex Gaussian random variable, and all the samples are considered to be
uncorrelated resulting in a covariance matrix n2 I.

On considering the multipath and AWGN eects, the received time domain OFDM
signal yr CN 1 at the receiver can be modelled as

0
yr = Hx + n, (2.25)

where H CN N represents the channel matrix. Due to the UW, a cyclic structure is
introduced, which transforms the linear convolution of the channel with the transmitted
signal to a circular convolution. Therefore, the channel matrix H turns out to be a cyclic
CHAPTER 2. REVIEW OF UW-OFDM 11

convolution matrix and is given by

h0 0 0 hNk 1 h1

.. .. .
. . .
h1 h0 0

.
. .. . ..

. . . .
. h1 . hNk 1


. ..

H = hN 1 . .
. (2.26)
. h0 0 0

k
.. .
. .

0 hNk 1 h1 h0 .


. .. .. .. ..

. . . . .
. ... 0
0 0 hNk 1 h1 h0

Note that the matrix H is a circulant matrix which is dened by its rst column

 T
h0 h1 hNk 1 0 0 .

Equation (2.25) holds only if there is no frequency or timing oset, thus representing an
idealized system. However, in practical cases these eects will appear and will cause severe
degradation of the OFDM system performance. These impairments will be investigated in
detail in the next chapter.

2.3 Receiver Model


On considering (2.21) and (2.25), the received OFDM symbol yr follows to

yr = HF1
N (
x+x
u ) + n. (2.27)

A well known property of the circulant matrix is that any circulant matrix can be diagonal-
ized by the pre and post multiplication with the DFT and the IDFT matrices, respectively.
Let us dene a matrix CN N
H such that

H = F1
N HFN , (2.28)

where
H is a diagonal matrix whose diagonal elements represent the sampled frequency
n o T
= diag FN hT 0T

response of the channel (i.e., H ). By substituting H in (2.27),

yr can be rewritten as

yr = F1 1
N HFN FN (x+x u ) + n
1
= F H (
Nx+x
u ) + n. (2.29)

Applying the DFT on yr , the received frequency domain symbol r CN 1


y follows to

r = FN F1
y x+x
N H ( u ) + FN n (2.30)

=H (
x+x u ) + FN n. (2.31)

By using (2.23), (2.31) can be written as


 
y
r = H +x
BGd u + FN n. (2.32)
CHAPTER 2. REVIEW OF UW-OFDM 12

It can be seen that the vector y


r contains zero subcarriers, which can be excluded from
further processing. So, the vector r can be downsized from N 1
y to (Nd + Nr ) 1 by
multiplying y
r with BT , leading to d C(Nd +Nr )1 dened as
y
d = BT y
y r . (2.33)

Thus, (2.32) can be written as

y
d = H +H
d Gd d BT x
u + v
, (2.34)

where Hd dened as Hd =
C(Nd +Nr )(Nd +Nr ) is a downsized version of
H . The
BT HB

matrix Hd contains the channel frequency response coecients only at the data and at the
C(Nd +Nr )1 is a noise vector dened as v
redundant subcarrier positions. v = BT FN n
2
with zero mean and a covariance matrix N n I.

From the above equation it can be observed that the received symbol y
d still contains
d BT x
the inuence of H u . Since the UW is known to the receiver, its inuence can easily be
removed by just subtracting d BT x
H u from y
d , assuming that the channel or an estimate
of the channel is known. The resulting vector y
turns out to be a linear model, and is
given by
y
=H +v
d Gd . (2.35)

Based on this linear model, several linear estimators are introduced in the next sections.
Whether the presented equalizers have much practical aspect or not, they all give an
illustrative view on how they exploit the correlations in the frequency domain introduced
by the matrix G. More detailed information can be found in [10]. Fig. 2.3 depicts the
basic block diagram of the baseband receiver model for UW-OFDM.

yr ~
yr ~
yd ~
y ~
d
QAM Demapping
subcarriers, BT
Serial/Parallel

Parallel/Serial
Estimation, E
Exclude zero

Subtract UW

Linear Data

yr

Influence

Binary data
FFT

Figure 2.3: Baseband Receiver Model for UW-OFDM.

2.3.1 Classical Data Estimators

Let us consider a linear unbiased estimator of the form


d
b
= E
y, (2.36)

where E CNd (Nd +Nr ) represents the equalizer. As the estimator is assumed to be
unbiased, it has to satisfy the relation
h i

E d
b
= E [E
y ] = EH = d.
d Gd (2.37)

This leads to the following Zero Forcing (ZF) criterion for linear equalizers:

d G = I.
EH (2.38)
CHAPTER 2. REVIEW OF UW-OFDM 13

Channel Inversion Equalizer


A straightforward linear zero forcing equalizer that can be applied to the UW-OFDM
system is the Channel Inversion (CI) equalizer given by

1 .
ECI = I 0 P1 H
 
d (2.39)

This diagonal matrix equalizer just inverts the channel frequency response, and does not
exploit the correlations among the subcarriers. Note that the CI equalizer is the opti-
mum data estimator in CP-OFDM, as the frequency domain data in CP-OFDM does not
have any correlations among them. In contrast, the UW-OFDM has correlations in fre-
quency domain, but the CI equalizer does not exploit these, resulting in a sub-optimum
ZF solution.

Best Linear Unbiased Estimator


Based on the linear model given in (2.35), an optimum ZF equalizer, the Best Linear
Unbiased Estimator (BLUE) [31], follows to
 1
H
EBLUE = GH Hd

H d G H
GH Hd. (2.40)

Since the noise vector v


is assumed to be Gaussian with a covariance matrix
 H
Cvv = E v
v = N n2 I, (2.41)

the BLUE is also the Minimum Variance Unbiased (MVU) estimator whose error covariance
matrix is given by
 1
H
Cee = N n2 GH Hd

H d G . (2.42)

Note that, for UW-OFDM EBLUE is a full matrix, whereas for CP-OFDM it is a diago-
nal matrix, which is equivalent to the CI equalizer. In contrast to the CI equalizer, its
implementation leads to a substantially higher computational complexity, as it requires a
Nd Nd matrix inversion. In [10], a complexity reduced version of this estimator is derived.

2.3.2 Linear Bayesian Data Estimators

All before mentioned estimators are based on classical data estimation where the elements
of a data vector are assumed to be deterministic but unknown constants. The Linear
Minimum Mean Square Error (LMMSE) estimator is based on the Bayesian approach,
where the data vector is assumed to be a realization of a random vector [31]. The theory
of Bayesian estimators yields another equalizer for UW-OFDM. Considering the model
in (2.35) as a Bayesian linear model, where
d is the realization of a random vector with
covariance matrix Cdd = d2 I and
v
is a noise vector with zero mean and a covariance
matrix Cvv 2
= N n I, the LMMSE equalizer is given by

ELMMSE = WH1
d . (2.43)

Here, W is a Wiener smoothing matrix given by


1
N n2  H 1

H H
W=G GG + 2 Hd Hd . (2.44)
d
CHAPTER 2. REVIEW OF UW-OFDM 14

ELMMSE can be interpreted as applying the CI equalization 1


H at rst, followed by a
d
Wiener smoothing operation W. The Wiener smoothing operation exploits the correlations
among the subcarriers introduced at the transmitter side and minimizes the inuence of
the noise on the received data. The estimator in (2.43) can also be written as

1
N n2

ELMMSE = H
GH H
d Hd G + 2 I H
GH Hd, (2.45)
d

with the error covariance matrix expressed as

1
N n2

HH
Cee = N n2 G Hd Hd G + 2 I . (2.46)
d

The estimator dened in the latter equation requires less computational complexity com-
pared to the former one. It is obvious that at higher SNRs (i.e., n2 0), ELMMSE
is equivalent to EBLUE . Comparing the three equalizers described in (2.39), (2.40) and
(2.45) in terms of computational complexity, the following order holds:

ELMMSE > EBLUE > ECI . (2.47)

Similarly, when compared in terms of performance, they have again the same order:

ELMMSE > EBLUE > ECI . (2.48)


Chapter 3

Carrier Frequency Oset


A well-known and critical issue of all OFDM systems is the presence of carrier frequency
oset (CFO). In this chapter the causes and eects of CFO are analyzed in detail. Besides
the CFO, there is a need of timing synchronization at the receiver, as the arrival time of
an OFDM symbol is unknown. The main goal in the timing synchronization is to detect
the instant at which the rst sample of an OFDM symbol is received so as to place the
DFT window accordingly. In this thesis, only the frequency oset is considered, therefore,
an ideal timing synchronization is assumed.

In an OFDM system, the data is transmitted on orthogonal subcarriers in parallel


through the channel. However, all the subcarriers are orthogonal if and only if the trans-
mitter and the receiver are operating at the same frequency. In this case the subchannel
spectral component is zero at the frequency positions of all other subchannels. Fig. 3.1
exemplarily shows the individual spectrum of 4 orthogonal subcarriers, whereas each sub-
carrier is loaded with a data symbol. Any mismatch of the oscillators between the trans-

1.5

1
2
|s(f)|

0.5

0
0 50 100 150 200 250 300 350 400 450
Frequency

Figure 3.1: Frequency response of 4 orthogonal subcarriers.

mitter and the receiver, or a Doppler eect caused due to the movement of receiver or
transmitter, will lead to CFO. The causes of a frequency drifting in the oscillators can
be due to temperature, humidity, electromagnetic interference, aging, and pressure. Usu-
ally, the oscillator stability is measured in parts-per-million (ppm) (1ppm = 106 ). It is

15
CHAPTER 3. CARRIER FREQUENCY OFFSET 16

common that the stability of a typical mobile phone crystal varies between 2 ppm and 12
ppm [32], which is equivalent to a frequency shift of 4 to 24 kHz for a carrier frequency
of 2 GHz. The frequency oset due to the Doppler eect mainly depends on the direction
of the movement of the receiver/transmitter with respect to the direction of the arrival of
the incoming signal and the velocity of the transmitter/receiver. The frequency shift fd
introduced by the receiver movement is given by

v
fd = fc cos(), (3.1)
c
where v is the velocity of the receiver, c is the speed of light, fc is the carrier frequency,
and represents the angle of arrival. For example, a mobile receiver moving towards the
transmitter with a speed of 100 km/h will experience a frequency shift of 185 Hz at a
carrier frequency of 2 GHz. The frequency shift in a typical mobile environment varies
between 1 Hz to a few hundreds of Hz. Note that the amount of CFO due to the Doppler
spread is very low compared to that of a mismatch between the oscillators.
Even a small frequency oset f induces two degrading eects on the data:

The orthogonality among the subcarriers is destroyed which leads to Inter Carrier
Interference (ICI), resulting in a signicant degradation of the overall BER perfor-
mance.

The signal is attenuated and rotated.

In the presence of a frequency oset, the receiver cannot sample exactly at the center
frequency of the subcarriers. Fig. 3.2 clearly visualizes the two main eects of CFO.
Therefore, it has become a key challenge at the receiver to estimate the CFO accurately
and keep its impact at a minimum.

1.5

Frequency offset (f )

1
2
|s(f)|

0.5

0
0 50 100 150 200 250 300 350 400 450
Frequency

Figure 3.2: Inter Carrier Interference caused due to the CFO.

The CFO could be many times larger than the intercarrier spacing. It is thus divided
into an integer part and a fractional part. The integer part only leads to a shift of the
subcarriers by this integer number of positions. Hence, there will be no ICI, and the
orthogonality among the subcarriers is preserved. The fractional part, however, causes
ICI which destroys the orthogonality among the subcarriers. It was shown in [19] that
CHAPTER 3. CARRIER FREQUENCY OFFSET 17

the OFDM system performance degrades signicantly when the oset exceeds 4 - 5 % of
the intercarrier spacing, and it was also shown that the SNR degradation is quadraticly
related to the relative oset. The following sections describe the eect of CFO in time and
frequency domain, respectively.

3.1 CFO Eect in Time Domain


Assuming that a frequency oset f is present in the system (for the moment multipath
and AWGN is neglected), due to this oset, the subcarrier frequencies are shifted by f .
This oset leads to a phase rotation of 2f t in time domain. This can be written as

0 0
y (t) = x (t) ej2f t (3.2)

0 0
where x (t) and y (t) are the transmitted and the received time continuous signal, respec-
tively. Thus, the received baseband spectrum will lie around f instead of lying around
f =0 (i.e., the DC position).

Let  be a relative frequency oset which is dened as the actual frequency oset f as
a fraction of the intercarrier spacing fic given by

f
= . (3.3)
fic
The intercarrier spacing fic can be expressed as

fs 1 1
fic = = = , (3.4)
N N Ts TDF T
where fs and Ts denote the sampling frequency and the sampling time, respectively. By
using the relation in (3.4),  can be rewritten as

 = f N Ts . (3.5)

In discrete time domain, the phase rotation factor e2f t can be expressed as e2f nTs with
n = 0, 1, , N 1, when considering only one OFDM symbol. With (3.5), the discrete
representation of the phase rotation factor is given by

j2n
ej2f nTs = e N n = 0, 1, , N 1. (3.6)

Thus, the phase rotation for one OFDM symbol can conveniently be written in matrix
form as h
0 j2(N 1)
iT 
j2
= diag 1 e N e N



1 0 0
j2 .. .

0 e N . .
.
= .

.. ..
. (3.7)
.. . . 0


j2(N 1)
0 0 e N
CHAPTER 3. CARRIER FREQUENCY OFFSET 18

0
CN N represents the phase rotation matrix that models the frequency oset eect on
one OFDM symbol. For simplicity, the phase accumulated by the previous OFDM symbols
0
is ignored (i.e., the oset matrix includes only the phase accumulated by the current
OFDM symbol).

In order to derive a general oset matrix for an UW-OFDM symbol, the phase accu-
mulated by the previous OFDM symbols and the extra appended UW (usually a UW is
appended in front of every UW-OFDM burst to maintain the cyclic structure, cf. Fig. 2.1)
should also be taken into account, leading to



1 0 0
j2 .. .

0 e N . .
j2[N (l1)+Nu ] .
l = e N .
. .. ..
, (3.8)
. . 0

.
j2(N 1)
0 0 e N

where l indicates the index of the lth transmitted OFDM symbol. Each diagonal element
of l CN N represents the corresponding phase rotation factor for each sample of the
transmitted OFDM symbol.

Fig. 3.3 exemplarily illustrates the phase rotation of the UW xu for dierent values of
CFO. Here, all the samples of the UW are set to ones (i.e., xu = [1, 1, , 1]T ) and the
UW shown in the gure belongs to the rst OFDM symbol. It can be observed that the
UW is rotating along the unit circle in the presence of CFO, and when the CFO is absent,
the samples of the UW lie at the point (1,0). It can also be noticed from the same gure
that the rst sample of the UW lies at various points (i.e., at dierent phases) for dierent
CFOs. The reason can be explained as follows. The phase (or frequency oset) of the nth
sample can be calculated as the sum of the phase accumulated by the previous (n 1)
samples and the current phase of the nth sample. This can be expressed as

2(n1)
n o 2
arg nth sample =e N +e N . (3.9)

As the UW is at the end of an OFDM symbol, at the time of the rst sample of the UW
there is already a phase accumulated by the preceding data samples and this accumulated
phase increases with the increase in . This makes the rst sample of the UW start at
dierent points for dierent  values.

Fig. 3.4 reveals that there is a linear phase dierence between the transmitted time
domain signal and the signal inuenced by CFO. Furthermore, it can also be noticed that
the slopes are increasing with increasing CFO.

By taking into account all the above described CFO eects, the lth received time domain
OFDM signal zl with zero UW can be modelled as

zl = l F1
N xl (3.10)
CHAPTER 3. CARRIER FREQUENCY OFFSET 19

1
data1
data2
0.8 data3
=0
0.6 = 0.1
= 0.2
0.4 = 0.4
= 0.6

Imaginary Part
0.2

0.2

0.4

0.6

0.8

1 0.5 0 0.5 1
Real Part

Figure 3.3: Phase rotation of the UW for dierent CFOs ( = 0.1, 0.2, 0.4, and 0.6).

1.5
= 0.01
= 0.02
= 0.05

1
Phase [rad]

0.5

0
0 50 100 150 200 250 300
Time [samples]

Figure 3.4: Eect of CFO on the phase of the time domain signal.
CHAPTER 3. CARRIER FREQUENCY OFFSET 20

3.2 CFO Eect in Frequency Domain


After performing the DFT operation on (3.10), the lth received frequency domain OFDM
symbol
zl can be written as

N l FN x
zl = F1 1
l . (3.11)

By considering the properties of the circulant matrix discussed in the previous chap-
ter, (3.11) follows to

zl = F1
N l FN x
l (3.12)

=e lx
l , (3.13)

where
el represents the frequency domain version of the time domain oset matrix l .
Note that in contrast to l ,
el is no longer a diagonal matrix, consequently, interferences
are introduced among the subcarriers by the non-diagonal elements along with a phase
shift caused by the diagonal elements.

The following derivation elaborates the eects of CFO in frequency domain. Con-
sidering that the lth received time domain UW-OFDM symbol zl contains the samples
{zl [0] , zl [1] , , zl [N 1]}, applying the DFT on zl yields the frequency domain OFDM
symbol zl with the samples { zl [0] , zl [1] , , zl [N 1]}. Mathematically, this can be
written as

N 1
X j2kn
zl [k] = DFT {zl [n]} = zl [n] e N

n=0
N 1 N 1
X 1 X j2(m+)n j2[N (l1)+Nu ] j2kn
= x
l [m] e N e N e N , (3.14)
N
n=0 m=0

where zl is the IDFT of the vector x


l and l is the index of the received OFDM symbol.
On reordering (3.14), it follows to

N 1 N 1
1 j2[N (l1)+Nu ] X X j2(m+k)n
zl [k] = e N x
l [m] e N
N
m=0 n=0
N 1 N 1
1 j2[N (l1)+Nu ] X j2n X j2(mk)n
= e N e N x
l [m] e N
N
n=0 m=0
N 1
1 j2[N (l1)+Nu ] X j2n
= e N x
l [k] e N +
N
n=0
N 1 N 1
1 j2[N (l1)+Nu ] X X j2(mk+)n
e N x
l [m] e N . (3.15)
N
m=0,m6=k n=0

It can be shown that [33]

sin() j(N 1+2[N (l1)+Nu ])


zl [k] = e N l [k] + Il [k] .
x (3.16)
N sin( 
N )
CHAPTER 3. CARRIER FREQUENCY OFFSET 21

I [k] denotes the ICI on the k th subcarrier caused by the remaining subcarriers and is given
by
N 1 N 1
1 j2[N (l1)+Nu ] X X j2(mk+)n
Il [k] = e N x
l [m] e N . (3.17)
N
m=0,m6=k n=0

This shows that the orthogonality of the subcarriers is lost in the presence of CFO, leading
to the ICI. Fig. 3.5 illustrates this eect for  = 0.5. The rst term of (3.16) represents the
amplitude and the phase distortion of the k th subcarrier caused due to the CFO. It can be
observed that this distortions become severe with increasing CFO.

1.5

= 0.5

1
2
|s(f)|

0.5

0
0 50 100 150 200 250 300 350 400 450
Frequency

Figure 3.5: Inter Carrier Interference caused due to the CFO.


Chapter 4

The Pilot Based UW-OFDM


Transceiver Model
The traditional UW-OFDM signalling structure does not incorporate pilot subcarriers in
the frequency domain, as it is assumed that the UW will take over their estimation tasks.
However, as one aim of this work is the comparison of time domain versus frequency do-
main based estimators, without loss of generality, the traditional scheme has been slightly
adapted. Thus, the altered UW-OFDM structure will include a pilot sequence in both
time (i.e., as UWs) and frequency (i.e., as pilot tones) domain. In practice, of course only
the UW or the frequency domain pilots will exist at one time.

When briey recalling the traditional approach, an UW-OFDM symbol can be formu-
0 T
x = xT xT

lated as d u , where xd denotes a vector containing random values aected by
the data, and xu represents the UW vector which is added at the end of each OFDM sym-
0  T T
bol. x is constructed in two steps: 1) An OFDM symbol with a zero UW x = xd 0T
is generated. 2) The UW sequence xu is added to the vector x and the nal UW-OFDM
0  T T
symbol follows to x = x + 0 xT
u .

The point where the modied scheme diers from the conventional UW-OFDM is the
generation of the vector x
. The pilot subcarriers CNp 1
p are accommodated in the fre-
quency domain by introducing them along with the data subcarriers ,
d and furthermore,
to produce a zero UW in the time domain, redundant subcarriers
r are introduced in the
frequency domain. The number of redundant subcarriers Nr is set to Nu . As a result,
to solve the system of equations x = F1
N x, the current number of data subcarriers Nd
should be reduced by Np + Nr , so that both the pilots and the redundant subcarriers can
be accommodated. In this work the pilot subcarriers are distributed according to the IEEE
802.11a WLAN standard [34]. Moreover, the redundant subcarriers should be placed on
optimal locations over the available bandwidth (for details refer to Section 2.1). Note that,
the optimal redundant subcarrier positions here in this modied scheme are dierent to
the ones in the traditional scheme, because the pilot carriers have occupied some of the
traditional redundant subcarrier positions.

To place the data, pilots, and redundant subcarriers on their respective locations, a per-

22
CHAPTER 4. THE PILOT BASED UW-OFDM TRANSCEIVER MODEL 23

mutation matrix P C(Nd +Nr +Np )(Nd +Nr +Np ) is introduced. The matrix P is optimized
in such a way that the energy contribution of the redundant subcarriers is minimum.
After the distribution of ,
d r and p
, the usual zero subcarriers are inserted with B
CN (Nd +Nr +Np ) . The number of zero subcarriers Nz is calculated as N (Nd + Nr + Np ).
The time domain and the frequency domain relationship of the modied UW-OFDM sym-
bol with zero UW can be formulated as



d  
1 x
r = d .
FN BP (4.1)
0
p

The redundant subcarriers


r have now to be determined from (4.1) such that a zero UW
is produced at the IDFT output. Dening the matrix M CN (Nd +Nr +Np ) as

 
M11 M12 M13
M= F1
N BP = , (4.2)
M21 M22 M23

whereas M11 C(N Nr )Nd , M12 C(N Nr )Nr , M13 C(N Nr )Np , M21 CNr Nd ,
M22 CNr Nr , and M23 CNr Np are appropriate sized sub-matrices. Using (4.1) and
(4.2), the following relation is obtained:

r = M1
1
22 M21 d M22 M23 p
. (4.3)

Note that, the redundant subcarriers in this approach depend on both the data and the
pilot subcarriers, whereas in the traditional UW-OFDM system, the redundant subcarriers
only depend on the data. Dening the matrices T1 = M1
22 M21 and T2 = M1
22 M23
(T1 CNr Nd and T2 CNr Np ), the above equation can conveniently be written as

+ T2 p
r = T1 d . (4.4)

It can be observed from (4.4) that


r results from a linear combination of
d and p
. Using
(4.1) and (4.4), this leads to


I 0  
1 x
= d .
FN BP T1 d + T2 p (4.5)
0
0 I

The above equation can be rewritten as


I 0  
1 x
= d ,
   
FN B P1 P2 P3 T1 d + P1 P2 P3 T2 p
(4.6)
0
0 I

where P1 C(Nd +Nr +Np )Nd , P2 C(Nd +Nr +Np )Nr , and P3 C(Nd +Nr +Np )Np are the
sub-matrices constructed from the permutation matrix P. From (4.1) and (4.6) it follows
that


d
+ P1 0 + P2 T2 + P3 p
   
r = P1 + P2 T 1 + P3 0 d
P . (4.7)
p

CHAPTER 4. THE PILOT BASED UW-OFDM TRANSCEIVER MODEL 24

By introducing the matrices


  G C(N

d +Nr +Np )Nd and G C(Nd +Nr +Np )Np dened
 p as
G = P1 + P2 T1 + P3 0 and Gp = P1 0 + P2 T2 + P3 , (4.7) can be formulated as



d

c=P
+ Gp p
r = Gd , (4.8)
p

where c C(Nd +Nr +Np )1


is considered as a complex valued code word. G and Gp can
now be interpreted as code generator matrices for the code word
c.
0
Therefore, the nal transmit symbol x (assuming without loss of generality that the UW
has been set to zero) can be written as

0
 
x = F1
N B Gd + Gp p
. (4.9)

Similar to (2.27), after experiencing multipath and AWGN, the received symbol r CN 1
y
can be modelled as  
yr = HF1
N B Gd + Gp p
+ n, (4.10)

where H CN N and n CN 1 represent the circulant channel matrix and the noise
vector, respectively. On performing the DFT operation on yr , the frequency domain
symbol y
r can be written as
 
y
r = HB + Gp p
Gd + FN n. (4.11)

As y
r contains zero subcarriers, they can be excluded from further processing, leading to
a downsized vector d = BT y
y r with d C(Nd +Nr +Np )1 :
y

y
d = H +H
d Gd d Gp p
+v
. (4.12)

Here, d C(Nd +Nr +Np )1 is a


H downsized version of
H dened as d = BT HB
H . The
d contains the channel
matrix H frequency response coecients only at the data, pilots
and at the redundant subcarrier positions. C(Nd +Nr +Np )1 is a noise vector dened
v
as v= BT F 2
N n with zero mean and a covariance matrix N n I. From the above equation
it can be noticed that the received symbol y
d contains the known portion H d Gp p
(if the

channel Hd is assumed to be known), and this portion can be subtracted to obtain the
linear model

y
=y d Gp p
d H (4.13)

=H +v
d Gd (4.14)

Based on this linear model, several linear estimators were introduced in the previous chap-
ter (see Section 2.3), and the same procedures can be applied here for the data estimation.

By observing the result in (4.13) it can be concluded that this model is not practical for
pilot based estimation, as it completely excludes them. However, for the CFO estimation
along with the data, the pilot symbols are also needed at the receiver side, and therefore,
CHAPTER 4. THE PILOT BASED UW-OFDM TRANSCEIVER MODEL 25

a modied model is introduced where the data as well the pilots are taken into account.
For this purpose it is better to treat pilots like random data (although they are in fact
deterministic) and estimate the pilots along with the data. Then these estimated pilots
are used for the CFO estimation. For that reason, it is more convenient to introduce a
generator matrix G0 constructed from G and Gp such that the BLUE and the LMMSE
equalizers discussed in Section 2.3 can be applied in a straightforward manner. Let the
0
matrix G be given by
0  
G = G Gp . (4.15)

Therefore, the received symbol p C(Nd +Nr +Np )1


y containing both the pilots and the
data symbols as well as the redundant symbols is given in the form of the linear model


 
d G0 d
y
p = H +v
. (4.16)
p

Considering a linear unbiased estimator of the form

" #

d
b 0
=Ey
p , (4.17)
p

b

0
where E C(Nd +Np )(Nd +Nr +Np ) represents the equalizer, the BLUE equalizer follows to

 H 1  H
0 0 0 0
EBLUE = G H
H
d Hd G G H
Hd, (4.18)

with the error covariance matrix



0
H  0
1
Cee = N n2 G H
H
d Hd G . (4.19)

Similar to (2.45), the LMMSE equalizer is given by

1  
N n2
 H
0 0 0 0 H
ELMMSE = G H
H
d Hd G + 2 I G H
Hd, (4.20)
d

with the covariance matrix expressed as

1
N n2
 H
0 0
Cee = N n2 G H
Hd Hd G + 2 I . (4.21)
d

The CI equalizer described in (2.39) can be applied to this model as well. After the
bT bT
h i
equalization process the CFO can be estimated out of the obtained vector
d p

which will be discussed in the next chapter.
Chapter 5

Carrier Frequency Oset Estimation


It is a well known fact that OFDM systems are highly sensitive to CFO. A demodulation of
an OFDM signal in the presence of CFO leads to the loss of orthogonality between the sub-
carriers causing ICI, thus leading to a signicant degradation in the BER performance [12].
Therefore, an accurate frequency oset estimation and compensation is necessary to over-
come this degradation. Numerous papers have been published so far on the subject of CFO
estimation in CP-OFDM systems. However, to the best of my knowledge, CFO estimation
in UW-OFDM has not been investigated till now. This chapter gives an overview of various
CFO estimation techniques for UW-OFDM, which are in fact derived from current CFO
estimation techniques for CP-OFDM. In a real system, CFO estimation is usually carried
out in two phases [35]. First, there is an acquisition phase, where a rough CFO estimate is
computed to align the local oscillator frequency to the received carrier frequency. This op-
eration is normally carried out for each received OFDM burst. Then, in the second step, a
ne CFO estimate is computed, which is often referred to as tracking phase. The splitting
of the estimation in two phases allows the design of the whole synchronization task with
a high degree of freedom, thereby, at each phase a separate optimized algorithm can be
used. This means that, during acquisition phase an optimized algorithm can be utilized
to cover a large acquisition range while neglecting the task in tracking phase, whereas in
the tracking phase the algorithm can be optimized to exhibit high tracking performance,
as the large acquisition range is no longer required in this stage. As one aim of this thesis
is to utilize the UWs for the CFO estimation, from here on, the focus is restricted to the
ne CFO estimation techniques. However, a brief introduction to the acquisition phase is
covered in the following section.

5.1 Acquisition Phase


In the acquisition phase a coarse CFO estimation is carried out. Furthermore, the coarse
CFO, in general, is estimated in two steps [18]. In the rst step, the fractional part of the
CFO is estimated and corrected, and in the second step, the integer CFO is estimated and
corrected. The entire acquisition process is normally done by utilizing a preamble that
is sent at the beginning of each burst (a burst is composed of a group of several OFDM
symbols), cf. [34]. The preamble consists of training sequences that are repeated in small
parts and will experience the same eects of the channel, but they dier in the phase shift

26
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 27

due to the frequency oset. Therefore, the oset can conveniently be estimated by mea-
suring the phase dierence between two successive parts in frequency domain. This idea
was rst proposed by Moose in [14]. He used two identical OFDM training symbols and
estimated the CFO out of them. However, the main drawback of this technique is that the
acquisition range is limited to one half of the subcarrier spacing due to the discontinuity
of the arctangent function used in the algorithm. It means that, if suppose  is very close
to +0.5 or -0.5, then, due to the discontinuity of the arctangent and the noise, the esti-
mated value might jump to -0.5 or +0.5, respectively, thus making the estimator a biased
one, and therefore useless in practice. In the same paper, he also described a strategy to
increase the acquisition range by using shorter training symbols. The range will be then
increased on the one hand, but the estimates will have a larger variance compared to the
ones estimated with longer training symbols on the other hand. The reason for that is the
reduced number of samples the average is built over. Another method that extends the
acquisition range with reduced computational complexity and with improved accuracy was
proposed by Schmidl & Cox in [18]. Relevant literature similar to the above ideas can be
found in [3640].

On considering (3.10), the lth received time domain UW-OFDM symbol zl having an
UW xu can be written as

zl = l F1
N (
x+x
u ) (5.1)

= l x ,
0
(5.2)

where x0 is the transmitted signal and l is the frequency oset matrix as dened in (3.8).
To compensate the eect of CFO, let us dene a de-rotating matrix which is given by



1 0 0
j2 .. .

0 e N . .
j2
[N (l1)+Nu ] .
l = e N .
. .. ..
, (5.3)
. . 0

.
j2
(N 1)
0 0 e N

where  is the estimated oset obtained through the coarse estimation. By de-rotating the
vector zl with l , the coarse CFO can be compensated and the resulting vector rl CN 1
follows to
rl = l zl = l l x .
0
(5.4)

Assuming that the estimated oset is equal to the true oset (i.e.,  = ), then l l will
become an identity matrix. Hence, the eect of the oset will be completely eliminated.
However, in practice, there is always some residual oset left in the signal rl . This is where
the tracking phase comes into play. The following section explains the tracking procedure
followed by some CFO estimation techniques.

5.2 Tracking Phase


After the coarse frequency synchronization, some residual CFO will probably remain in the
signal. This residual CFO will vary in the presence of time-varying Doppler shifts. This
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 28

oset, in general, is in the order of 4 to 5 % of the intercarrier spacing, and this induces
ICI and a phase rotation of data symbols (see Fig. 5.4). If this small oset is not compen-
sated accurately, after a few OFDM symbols, the QAM/QPSK data symbols will rotate
and may fall in a dierent quadrant of the constellation diagram, leading to a degraded
system performance. Therefore, the oset should be estimated (tracked) continously and
the received data corrected accordingly. The tracking methods that are currently avail-
able in the literature for CP-OFDM are broadly classied into three categories: those that
base the estimation on the signal's statistical properties (e.g., cyclic prex-based (CPB)
methods), those that base the estimation on pilot tone information (e.g., pilot tone-aided
(PTA) methods), and decision-directed (DD)/data-carrier aided schemes. In traditional
UW-OFDM, since the UW is dened in the time domain, CPB methods can be applied
to this scheme. However, as the modied model incorporates pilot tones in the frequency
domain, PTA techniques and DD schemes are also discussed in this thesis.
Before going into the details of these techniques, it is very important to notice that the CFO
estimation algorithms for UW-OFDM are derived from currently available CFO estima-
tion techniques for the CP-OFDM. As a consequence, only the CFO estimation techniques
for the UW-OFDM are described in this chapter without having a pre-discussion of the
algorithm for CP-OFDM.

The following explains the basic tracking procedure that is being followed in an OFDM
system [19]. Assume that the received signal rl obtained from (5.4) has some residual
carrier frequency oset  (i.e.,  ). In reality, this error may vary from symbol to
symbol. Therefore, this error has to be tracked and compensated continuously to avoid
ICI. Usually the tracking operation is done on a symbol-by-symbol basis, thus leading to
a closed-loop structure as shown in Fig. 5.1 [19]. Here, rl (k) are the received time-domain

rl (k) vl (k) Further


Processsing

j 2l k

e
l el
Error
NCO Loop Filter
Generator

Figure 5.1: Recursive-loop structure for tracking the CFO [19].

samples of the lth OFDM symbol. The error signal el provides the information on , and
is fed to the loop lter to calculate the frequency error estimate using the iterative function


l =
l1 + el , (5.5)

where
l and
l1 are the estimated frequency errors of the lth and the (l 1)th OFDM
symbol, respectively. is the parameter that controls the behavior of the loop. The pa-
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 29

rameter has to be chosen properly such that a reasonable tradeo exists between the
accuracy and the convergence speed in the steady state.

The estimate
l calculated from the loop lter (i.e.,
l =
l1 + el ) is fed to the
j2
l k
Numerically Controlled Oscillator (NCO) which generates an exponential term e N

for the k th sample (time instant) of the lth OFDM symbol. Therefore, the de-rotating
matrix l for the lth OFDM symbol rl is given by



1 0 0
j2l .. .
0 e N

. .
j2
l [N (l1)+Nu ] .
l = e N .
. .. ..
. (5.6)
. . 0

.
j2
l (N 1)
0 0 e N

Most of the available frequency tracking algorithms in the literature will track the oset
based on the above explained closed loop structure. They dier from each another in the
calculation of the error signal el . If el is calculated from the time domain samples (e.g.
the CP) of the OFDM symbol, then the estimation belongs to the time-domain tracking
scheme, whereas if el is calculated from the samples at the DFT output (i.e., pilot symbols
or data symbols) then it is called a frequency domain tracking algorithm. In case of CPB
methods, estimation is done based on the correlation of the CP and the corresponding last
samples of each OFDM symbol (in UW-OFDM, it is the correlation between two succes-
sive UWs). However, the cyclic prex is usually corrupted by the transmission channel,
therefore, accurate estimates are dicult to obtain in real world scenarios. In PTA algo-
rithms, pilot symbols are placed periodically on particular subcarriers and the estimation
is done by correlating the received pilot symbols with their transmitted values. In most
communication standards, the pilot subcarriers are utilized for channel estimation and for
synchronization purposes as well. Furthermore, the data subcarriers are also potential
candidates to improve the CFO estimation at the cost of some increased complexity. In
these techniques, the estimation is done by comparing the received data symbols with the
sliced values obtained from the de-mapper. Therefore, a larger number of estimates are
obtained to average over.
Note that, in this thesis, it is assumed that the CFO does not vary from symbol to symbol
and the estimation is always carried out from the rst two OFDM symbols, this means,
only one iteration is performed. Therefore, the tracking scheme is not implemented in this
thesis. Fig. 5.2 illustrates the signal processing steps involved for the CFO estimation in an
UW-OFDM baseband receiver. Note that, the gure depicts both the time and frequency
domain estimation procedures simultaneously, but in reality, either the time domain or the
frequency domain estimation procedure is carried out at one time. The following section
addresses the time domain CFO estimation algorithms followed by the frequency domain
estimation techniques.
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 30

Preamble

Channel Coarse CFO


Estimation Estimation

Substract
Post Equalizat CFO
UW DFT S/P
processing ion Correction
influence

IDFT

Fine CFO
Extract Extract Estimation/Tracking
Extract UW
Data Pilots

Decision Directed Pilot Tone Based Time Domain


Methods Methods Methods

Figure 5.2: CFO estimation in an UW-OFDM baseband receiver.

5.2.1 Time Domain CFO Estimation

This subsection describes the time domain estimation algorithm for UW-OFDM, bearing
in mind that this technique is originally derived for CP-OFDM. The algorithm that was
proposed in [13] is found to be the basis for many currently available CPB algorithms.
They proposed a Maximum Likelihood (ML) estimation procedure for estimating the fre-
quency error  and the timing error as well. In this method a log-likelihood function is
considered from which the frequency oset estimate
 is derived. A brief description of
the algorithm will be given in the following.

For a while assume that the received signal s was left with some residual oset .
For simplicity, consider that the channel is an AWGN channel. Therefore, the vector
s C(N +Nu )1 can be modelled as

s = q + n, (5.7)

h iT 0
where q = xT x 0T is the transmitted signal containing the OFDM symbol x preceded
u
by the UW belonging to the previous OFDM symbol. C(N +Nu )(N +Nu ) is the oset
matrix modelling the residual oset  and n represents the noise vector. The length of
the guard interval is set to Nu . Here, every sample in the transmitted signal q is assumed
to be an independent, identically distributed random variable (i.i.d.). According to the
central limit theorem, it can thus be assumed that the transmitted signal q is a complex
Gaussian random process. However, this process is not perfectly white, because due to the
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 31

presence of the UW, there will denitely be a correlation between the two UWs present in
the signal q. Therefore, the received signal s is also not a white process. Additionally, s
contains the information about the frequency oset (for instance, assume perfect timing
synchronization i.e., = ). By exploiting this information, the frequency oset can be
estimated as will be discussed in the following.

It can be noticed that, in the signal s, every sample of the rst UW is correlated with
its corresponding sample in the next UW, whereas the rest of the samples are uncorrelated
with each other. This can be written as [13]

2
s + n2 :m=0

E {s (k)s(k + m)} = s2 ej2 :m=N ,
0 : otherwise

where s2 and n2 represent the signal power and the noise power, respectively, and k refers
to the sample index within s. It can be observed from the above equation that the phase
dierence between any two corresponding samples of the two successive UWs in s is 2,
and the samples in each sample pair are separated by the length N. Let the probability
density function (PDF) of the received signal s for a given frequency oset  be dened
as f (s|). Therefore, the log-likelihood function () can be obtained by taking the
logarithm of the above PDF, leading to

() = log f (s|) . (5.8)

The goal is to nd an estimate


 such that the log-likelihood function is maximized. For
that, the derivative of the likelihood function with respect to  is set to zero, and this
results a maximum point which is
. In [13],
 is shown to be

(N 1 )
u
1 X

= arg s (k)s(k + N ) , (5.9)
2
k=0

where k refers to the index of the samples belong to the rst UW and s(k) represents a
sample of the symbol s. Here, arg {} is with respect to the averaging of the correlations
between the samples which are N samples apart (i.e., between the two UWs). Fig. 5.3 shows
the structure of the frequency oset estimator. Since the argument function only returns
values in between (, ), therefore, it can be observed from (5.9) that the estimation range
is limited to 21 <
 < 21 . As the maximum deviation of the frequency oset in tracking
mode will never exceed half of the subcarrier spacing, this algorithm can asymptotically
be used in tracking mode. The simulation results in [13] have proven this.

The performance of the estimator mainly depends on two factors, one is the length of the
UW Nu and the other is the correlation coecient between the corresponding samples
of two UWs. It can be noticed that as the length of the UW increases (i.e., the number
of samples), the estimation accuracy is improved because averaging is done over more
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 32

s(k) s*(k)
()
Moving
Sum arg{}
1
2
Filter

Z K = 0 to u-1
s(k + )
Figure 5.3: Frequency oset estimation structure as proposed in [13].

samples. The correlation coecient is given by [13]

cov (s(k)s(k + N ))
= p
var (s(k)) var (s(k + N ))




E {s (k)s(k + N )}
= r n

o n

o
E |s(k)|2 E |s(k + N )|2

s2
=
s2 + n2
SNR
= . (5.10)
SNR + 1
From the above equation, it can be recognized that the correlation between the samples
(the samples corresponding to the UW) is corrupted in the presence of noise. Hence, the
estimator works well at high SNR values in the AWGN channel. However, in the presence
of multipath (i.e., a time dispersive channel) the situation becomes worse, as the estimator
is not optimal for these channels. The multipath introduces correlations among all the
samples of s, thus violating the assumption of independence. Therefore, the performance
of the estimator will degrade signicantly in multipath channels, which is not desired.
However, the huge advantage of the UW-OFDM is that the UW can be chosen by the
designer. Therefore, the UW can be chosen such that the correlation properties are only
partly destroyed in the presence of a multipath channel, which can be an advantage re-
garding timing oset estimation. However, in the case of CFO estimation, the estimator
performance is independent of the choice of the UW.
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 33

In matrix/vector notation, (5.9) can be written as


(N 1 )
u
1 X

= arg su1 (k)su2 (k)
2
k=0
1  H
= arg su1 su2 , (5.11)
2
where su1 CNu 1 and su2 CNu 1 are two UWs that belong to the vector s. The
variance of the above estimator under AWGN is given by [23]

2 1 1
 = 2 SN R . (5.12)
Nu (2N )

Another unique feature of UW-OFDM is, as the UW is the same for all the OFDM
symbols, in contrast to CP-OFDM, the autocorrelation can be performed between any
two non-subsequent UWs separated by Nsy OFDM symbols. The maximum likelihood
estimator of  follows to [23]

1  H

= arg su1 su2 , (5.13)
2Nsy
with the error variance given by

2 1 1
 = 2 SN R . (5.14)
Nu (2Nsy N )
This shows that the error variance for the latter estimator is low compared to the one
before. This is because the former one has to estimate a small phase for a given SNR,
while the latter one has to estimate a larger phase (because the distance between two non
subsequent UWs is always larger than for Nsy = 1) having the same SNR, thereby reduc-
ing the error variance of the estimator. This means that the latter estimator performs
well for a slow varying frequency oset, leading to a very accurate estimation. However,
1 1
the estimation range of this new estimator is decreased to
2Nsy <  < 2Nsy . For fast
varying osets, this scheme maintains the possibility to perform exactly the same as the
old one (i.e., Nsy = 1). Note that, a vector of estimates can be obtained by performing
the estimation with dierent Nsy . To get a better accuracy, averaging over the dierent
estimates can also be done if necessary.

Once the estimate is obtained, the time domain signal rl (from (5.4)) is de-rotated
with this estimate using the matrix l C N N . Therefore, the corrected signal vl can be
obtained as

vl = l rl
0
= l l x . (5.15)

l represents the downsized version of the matrix , excluding the rows corresponding to
previous UW. If the residual frequency oset is perfectly estimated (i.e.,
 = ), then
the matrix l l will become an identity matrix, thus the oset is compensated completely.
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 34

5.2.2 Frequency Domain CFO Estimation

The most conventional approach to track the CFO in frequency domain is with the use
of pilot subcarriers that are usually transmitted along with the data subcarriers in an
OFDM symbol. These pilot subcarriers typically serve also for synchronization and chan-
nel estimation purposes. The principles of CFO estimation using pilot carriers have been
presented in [35], [14], and [41]. Besides these PTA techniques, Decision Directed (DD)
methods are also resulting in a robust and a considerable CFO estimation. These methods
are based on comparing the received frequency domain data symbols with their correspond-
ing sliced version symbols. Some good literature on DD based CFO estimation techniques
can be found in [4245]. The following subsections detail the PTA and DD approaches
individually.

Pilot Tone Aided CFO Estimation


The pilot subcarriers p
are transmitted in every OFDM symbol and are usually distributed
uniformly over the entire bandwidth. This kind of distribution is recommendable especially
in frequency selective fading environments. In an OFDM system, the presence of CFO will
introduce a phase shift on each frequency domain data symbol along with the ICI. In [35], a
Maximum likelihood estimation algorithm was introduced where the CFO is estimated by
measuring the phase shift between two corresponding pilots in subsequent OFDM symbols.
Let, pl,n and pl+1,n be the pilot symbols transmitted in the lth and the (l + 1)th OFDM
symbol on the n
th subcarrier. Then, the CFO  can be estimated as [35]
( )
N 1 X



 = arg zl,n zl+1,n pl,n pl+1,n , (5.16)
N + Nu 2
nI

where zl,n and zl+1,n are the pilot symbols received in the lth and the (l + 1)th OFDM
symbols, respectively, and Nu represents the length of the guard interval. The vector I
contains the indices of the pilot subcarriers p
. The same algorithm can be applied to the
UW-OFDM as well. However, in this thesis, the CFO estimation in frequency domain is
carried using only the rst UW-OFDM symbol. Therefore,  can be estimated using the
phase shift introduced on the pilot symbols p
of the rst UW-OFDM symbol.
The lth received frequency domain symbol zl [k] in the presence of oset, AWGN, and
multipath for an UW-OFDM symbol with zero UW can be written as

sin() j(N 1+2[N (l1)+Nu ])


zl [k] = e N l [k] + Il [k] + vl [k] ,
H [k] x (5.17)
N sin( 
N )

where [k]
H and vl [k] represent the frequency response of the channel and the noise on
the k th subcarrier of the lth OFDM symbol, respectively. Il [k] denotes the ICI on the k th
subcarrier caused by the other subcarriers and is given by

N 1 N 1
1 j2[N (l1)+Nu ] X X j2(mk+)n
Il [k] = e N [m] x
H l [m] e N . (5.18)
N
m=0,m6=k n=0
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 35

When considering only the rst UW-OFDM symbol (i.e., l = 1), (5.17) and (5.18) reduce
to
sin() j(N 1+2Nu )
z [k] = e N [k] + I [k] + v [k]
H [k] x (5.19)
N sin( 
N )
and
N 1 N 1
1 j2Nu X X j2(mk+)n
I [k] = e N [m] x
H [m] e N . (5.20)
N
m=0,m6=k n=0

When neglecting the eects of ICI and noise, the received symbol z [k] can be seen as the
product of the channel frequency response [k]
H and the transmitted symbol x
[k], atten-
uated, and rotated by some phase. Therefore, the frequency oset can be estimated by
simply calculating the phase of the received symbols, thus turning into a phase estimation
problem. However, the presence of a multipath channel will also aect the phase of the
received symbol. Therefore, the data symbols as well as the pilot carriers should be equal-
ized rst before utilizing them for estimation purposes. The received symbol
z containing
both the pilots and the data symbols is given in the form of the linear model


 
H d G0 d

z= +v
, (5.21)
p

where d
H is a downsized version of ,
H excluding the rows corresponding to the zero
subcarriers. The equalization can be performed with a linear estimator of the form

" #

d
b 0
=E
z, (5.22)
p

b

0
where E can be any linear equalizer. Here, in this thesis, three particular equalizers are
considered which are described in (4.18), (4.20), and (2.39).
After the equalization process the pilot symbols are extracted from the estimated vector
iT
bT bT
h

d p
. Assuming that perfect channel knowledge is available at the receiver (i.e.,

d = H
d ),
H the CFO can be estimated out of these estimated pilot symbols p

b by

( )
N X 

= arg p [n] p [n] . (5.23)
(N 1 + 2Nu )
nI

The above equation in matrix/vector notation can be written as

N n o
.

= arg pH p
(5.24)
(N 1 + 2Nu )

As the pilots are known to the receiver, the phase ambiguity can be resolved perfectly. In
reality, the receiver does not have perfect channel knowledge, thus limiting the performance
of the estimator.
The above estimator has given equal weightage to all the pilot subcarriers, and this works
absolutely well in at fading environments (or wideband channels). However, in frequency
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 36

selective fading, weighting the pilots equally is not an optimum combination. The estima-
tion accuracy can be increased by weighting each coecient dierently by following some
criterion. One popular way is to weight the estimates with the inverse of the corresponding
elements in the error covariance matrix of the estimator. The estimator with weighting
factors is given by
( )
N X 

= arg p [n] w
[n] p [n] (5.25)
(N 1 + 2Nu )
nI

For example, if the receiver uses a CI equalizer, then, after equalization there will be a
high inuence of noise on the subchannels experiencing higher channel attenuations. This
means, that pilot subcarriers with higher channel attenuation will be given less weightage
compared to that of a pilot with lower channel attenuation. This procedure is called
selection combining. The pilots are weighted with the squared channel response coecients,
and the CFO can be estimated as
( )
N X 2



= arg p [n] abs H [n] p [n] (5.26)
(N 1 + 2Nu )
nI

Considering a matrix p CNp Np ,


H whose diagonal elements represent the corresponding
channel responses on the pilot subcarriers, then the CFO estimate  in matrix/vector form
follows to
N n
H
o

= arg pH Hp H p p
. (5.27)
(N 1 + 2Nu )
In case of BLUE and MMSE equalization, the weighting coecients for the pilots are given
by the inverse of the corresponding diagonal elements in the error covariance matrices given
in (4.19) and (4.21), respectively.

Decision Directed CFO Estimation


Like as the pilot subcarriers, the data symbols also experience a phase rotation in the pres-
ence of CFO, and this phase shift is the same for all subcarriers (when excluding random
phase shifts due to the ICI). Hence, this phase shift can be utilized to improve the CFO
estimation [42, 43, 45]. The estimation procedure using the data symbols is similar to the
pilot tone based approaches, except the fact that in the DD approach the data symbols are
sliced to the nearest QPSK/QAM symbol in the constellation diagram by tentative/hard
decisions, which might lead to the errors in the estimate when wrong decisions are made.



After performing equalization, the estimated data symbol vector d is extracted out
of the equalized vector. As the receiver does not have the knowledge on the sent data
,


d slicing is performed on d . It means that each symbol in d is mapped to the nearest
symbol in the constellation diagram, assuming that each received symbol is originated from


the corresponding mapped alphabet. Dening the sliced vector as d sl , the CFO can be



estimated by measuring the angle between d sl and d given by

N n
H

o

=
(N 1 + 2Nu )
arg d sl .
d (5.28)
CHAPTER 5. CARRIER FREQUENCY OFFSET ESTIMATION 37

In case of selection combining, the weighting coecients are calculated in a similar fashion
described in the previous section. In reality, there might be a chance of making wrong
decisions, if the received symbols are near to the decision boundaries of the constellation
diagram. However, if the number of data symbols is large enough, then the impact of
wrong decisions can be compensated.

Note that, in the presence of CFO, the phase of the received frequency domain sym-
bols will increase from symbol to symbol. Therefore, it is necessary to compensate this
incremental phase, before the estimation is performed with the current symbol. If not
compensated, then, after a few OFDM symbols, the data symbols will cross the decision
boundaries of the constellation and the majority of data symbols will be sliced to a wrong
QPSK/QAM symbol, leading to a signicant error in the estimation. This eect can be
clearly observed from Fig. 5.4. This incremental phase can be compensated by using the
tracking procedure discussed in the previous section. After estimating the angle from the
current symbol, this angle is added to the angle obtained from the previous symbol to get
a new total rotation angle for the next symbol. In this way, the probability of making
wrong decisions will always be small, providing an accurate estimate.
In the next chapter the performance of various CFO estimation procedures is evaluated
and discussed.

1.5

0.5
Imaginary Part

1st OFDM symbol


0 2nd OFDM symbol
3rd OFDM symbol

0.5

1.5
1.5 1 0.5 0 0.5 1 1.5
Real Part

Figure 5.4: Constellation diagram of the received data symbols of the rst three consecutive
OFDM symbols for  = 0.05 and at n2 = 0.
Chapter 6

Results and Discussion


In this chapter, the performance of dierent CFO estimation algorithms for UW-OFDM
in terms of Eb /N0 and length of training sequence is analyzed. In addition, a comparison
is made between the time domain and the frequency domain CFO estimation techniques.
Besides this comparison, the sensitivity of these estimators on the CFO and the channel
characteristics is investigated. The sensitivity of the UW-OFDM system on the CFO is
analyzed and compared to that of the CP-OFDM scheme.

6.1 Simulation Setup


All the simulations are performed with the pilot based transceiver model, where the UW
as well as the pilots are introduced at the same time. This transceiver model is considered
only for a comparison of time domain versus frequency domain based estimators. In prac-
tice, of course only the UW or the frequency domain pilots will exist at one time. The total
number of subcarriers in each OFDM symbol is set to 64, where 32 are data subcarriers,
16 are redundant subcarriers, 4 are pilot subcarriers, and the remaining 12 are chosen to
be zero subcarriers. All the data subcarriers are modulated with a 4-QAM alphabet. The
four pilots are assigned with the values 1, -1, 1, and 1, respectively. The length of the
UW is set to 16. The positions of the redundant subcarriers are calculated according to
the optimization procedure described in [1]. For the modied scheme (i.e., the pilot based
model) the optimized positions are {3, 7, 11, 14, 18, 21, 25, 27, 39, 41, 45, 48, 52, 56, 60, 63}.
The pilot subcarrier indices in the modied scheme are selected to be {8, 22, 44, 58}. With
these indices it is trivial to construct the permutation matrix P. All the remaining param-
eters are chosen like in the 802.11a standard [34]. This scheme serves as a reference system
when comparing UW-OFDM against conventional CP-OFDM. An OFDM burst having
two UW-OFDM symbols is considered for one iteration run, and in total 5000 iterations
are performed at each Eb /N0 point, completing one simulation run. The performance of
the estimator is evaluated by means of its mean squared error (MSE). In case of multipath,
the Finite Impulse Response (FIR) of the channel is modelled as in (2.24), where each delay
tap is modelled as a complex Gaussian random variable having zero mean. All the taps
are considered to be statistically independent and having an exponential decaying power
prole. The channel delay spread is chosen to be 100ns and the total length of the channel

38
CHAPTER 6. RESULTS AND DISCUSSION 39

is shortened to the UW length Nu . Table 6.1 summarizes the most important parameters
used in this simulation setup.

Table 6.1: Important parameters used in this simulation setup.

Parameter Number
N 64
Nd 32
Nr 16
Np 4
Nz 12
Nu 16
Pilots (p
) 1,1,-1,1 (BPSK alphabet)
)
Data (d 4-QAM alphabet
fs 20 MHz
DFT period 3.2 s
Gaurd duration 800 ns
Total OFDM symbol duration 4 s
Subcarrier spacing 312.5 kHz

In the simulations, 3 CFO estimation techniques utilizing UW, pilots, and data symbols
are investigated and analyzed. However, it seems clearer to discuss the results in AWGN
and multipath channels separately. The following section rst discusses the results in
AWGN environment followed by the multipath channel case.

6.2 AWGN Environment


Before diving into the nal estimation results, let us rst analyze the behavior of two
successive UWs in the presence of CFO. The UW is chosen to be xu = 1. Fig. 6.1
depicts the phase rotation of the UWs belonging to the rst and the second OFDM symbol,
respectively, for  = 0.05. Note that the phase shift between each sample in the rst
UW to its corresponding sample in the subsequent UW is 2, as they are N samples
apart. The CFO can easily be calculated by taking the phase dierence between each
corresponding sample pair of two successive UWs and then averaging over the measured
phase dierences. However, in the presence of noise, the correlation properties between the
two UWs are partly lost (i.e., the phase dierence for all the sample pairs is not same, they
vary around the mean value). By averaging the phase dierence over all the samples, the
inuence of the noise can be reduced. Fig. 6.2 shows two subsequent UWs in the presence
of noise and CFO. It can be observed that the UWs are spread around the unit circle, and
this leads to the eect that the corresponding samples of the two UWs have dierent phase
dierences instead of having an equal phase dierence of 2.

Fig. 6.3 shows the performance of the time domain estimator (i.e., using UWs) with
 = 0.05. It can be clearly observed that the estimator performance increases with
CHAPTER 6. RESULTS AND DISCUSSION 40

0.8

0.6

0.4
Imaginary Part

0.2

0.2

0.4
UW1
0.6
UW2
0.8 data3
data4
1 data5
1 0.5 0 0.5 1
Real Part

Figure 6.1: Phase rotation of two successive UWs when  = 0.05.

0.8

0.6

0.4
Imaginary Part

0.2

0.2

0.4
UW1
0.6
UW2
0.8 data3
data4
1 data5
1 0.5 0 0.5 1
Real Part

Figure 6.2: Phase rotation of two subsequent UWs when  = 0.05 and Eb /N0 = 15 dB.
CHAPTER 6. RESULTS AND DISCUSSION 41

increasing Eb /N0 values. This is because at higher Eb /N0 values, the correlation between
the corresponding sample pairs in the subsequent UWs is higher compared to the case
for lower Eb /N0 values (see (5.10)). Note that the estimation is carried out using two
successive UWs belonging to the rst two OFDM symbols.

3
10
= 0.05

4
10
Mean squared error

5
10

6
10
0 5 10 15 20
Eb/N0 [dB]

Figure 6.3: Performance of the time domain CFO estimator for  = 0.05 in an AWGN
channel.

Fig. 6.4 shows the error between the true and the estimated CFO at Eb /N0 = 15 dB
for  = 0.05. As can be seen from the gure, the error follows a Gaussian distribution
having zero mean and a standard deviation of 0.004.

The MSE curves of the estimator for various Eb /N0 values (Eb /N0 = 10, 15, and 20
dB) over dierent UW lengths are plotted in Fig. 6.5. Note that there is a signicant
improvement (i.e., with a larger slope) in the performance of the estimator until a certain
UW length, and after that length the performance is improved very slowly (i.e., with a
lesser slope). However, the estimator performance will continuously be improved with the
increase of the length of the UW (i.e., averaging reduces the MSE).

Fig. 6.6 proves that the time domain estimator performance is almost independent of
the value of , provided that the oset is in the estimation range (0.5 to 0.5) and
additionally, for values close to the borders of the estimation range, the Eb /N0 value is
high. In other words, the estimator performs exactly the same for all osets within this
range. It can be observed from the gure that the estimator performance is identical for
 = 0.01 and 0.4 for all Eb /N0 values, whereas for  = 0.49, the estimator performance
is worse at low Eb /N0 range while performing the same in the high Eb /N0 region. This
CHAPTER 6. RESULTS AND DISCUSSION 42

400

350

300
No. of occurances

250

200

150

100

50

0
0.015 0.01 0.005 0 0.005 0.01 0.015
Error in the estimated values

Figure 6.4: The error distribution of the estimated CFO for  = 0.05 at Eb /N0 = 15 dB
in an AWGN channel .

4
x 10
3.5
Eb/N0 = 10 dB
E /N = 15 dB
b 0
3
Eb/N0 = 20 dB

2.5
Mean squared error

1.5

0.5

0
2 4 6 8 10 12 14 16
UW length

Figure 6.5: Performance of the time domain estimator for dierent UW lengths at  =
0.05 in an AWGN channel.
CHAPTER 6. RESULTS AND DISCUSSION 43

is because at  = 0.49, due to the noise, the phase between each corresponding samples
of the two UWs might exceeds , leading to a discontinuity in the arctan function, thus
resulting in a heavy error in the estimation. However, for high Eb /N0 values, as the noise
eect is minimum, the estimator works very well because the value 0.49 is still in the
estimation range. It can also be noticed that the estimator performance is completely
worse for  = 0.51 due to the fact that  is out of the estimation range, and the
performance will never improve, even in a noise free region (i.e., n2 = 0).

1
10

0
10

= 0.01
= 0.40
1
10
= 0.49
Mean squared error

2 = 0.51
10

3
10

4
10

5
10

6
10
0 2 4 6 8 10 12 14 16 18 20
Eb/N0 [dB]

Figure 6.6: Performance of the time domain estimator for dierent CFOs in an AWGN
channel.

In UW-OFDM, as the UW is the same for all the OFDM symbols, the autocorrelation
can be performed between any two non-subsequent UWs which are separated by several
OFDM symbols. Fig. 6.7 shows the performance of the estimator using two non-subsequent
UWs with  = 0.05. Here, for this particular simulation, a burst consisting of 4 OFDM
symbols is considered which contradicts the usual case where the simulation is performed
with a burst having only two OFDM symbols. The estimation is performed with dierent
UW pairs (UW1,UW2), (UW1,UW3), (UW1,UW4), where UW1, UW2, UW3, and UW4
belong to the 4 OFDM symbols, respectively. Fig. 6.7 reveals that there is a signicant
improvement of the performance of the estimator when the distance between the UW
pair increases, thereby improving the estimation accuracy. In CP-OFDM, the maximum
possible distance is one OFDM symbol, which is the same as estimating with the pair
(UW1, UW2) in UW-OFDM. However, if the distance between the UWs increases, the
1 1
estimation range decreases (
2Nsy <  < 2Nsy , where Nsy is the number of OFDM symbols

between the considered UWs). It means that the estimation range for the pair (UW1,UW4)
is limited to0.17 <  < 0.17. Fig. 6.8 shows the performance of the estimator with the
above mentioned UW pairs for  = 0.3. It can be observed that the estimation with the
CHAPTER 6. RESULTS AND DISCUSSION 44

UW pairs (UW1,UW3) and (UW1,UW4) fails because  is out of the estimation range,
whereas with the UW pair (UW1,UW2) the estimation is good as  is below 0.5.

3
10
UW1 and UW2
UW1 and UW3
UW1 and UW4

4
10
Mean squared error

5
10

6
10

7
10
0 5 10 15 20
Eb/N0 [dB]

Figure 6.7: Performance of the time domain estimator using two non-subsequent UWs for
 = 0.05 in an AWGN channel.

In the following the focus is laid on the frequency domain estimation techniques, namely
PTA and DD schemes. The pilot based methods are analyzed rst followed by the DD
approaches. The simulations are performed with the pilot based model having zero UW.
At the receiver, the pilots are estimated like they were data symbols by applying any
linear equalizer at rst (e.g., the equalizers introduced in chapter 4), and then the CFO
is estimated out of these estimated pilots using (5.23). The performance of the pilot
based estimator after applying the LMMSE equalizer (using (4.20)) for  = 0.01 and
 = 0.05 is shown in Fig. 6.9. The weighting average is done by weighting the estimates
with the inverse of the corresponding diagonal elements of the error covariance matrix
given in (4.21). It can be noticed that weighting the estimates has no additional benet
compared to the simple average because the error covariance matrix for an AWGN channel
is close to a scaled identity matrix. It can be seen that the MSE curves for  = 0.05
approaches an error oor at high Eb /N0 values. This is due to the result of ICI on the pilot
subcarriers with the increase in CFO. Note that this error oor raises with the increase in
the oset.

Fig. 6.10 compares the performance of CFO estimators for  = 0.01 after equalization
with CI, BLUE, or LMMSE. It can be noticed that the estimator with the CI equalizer
performs worst compared to the others due to the fact that it completely ignores the
information present on the redundant subcarriers. The LMMSE performance is optimum
CHAPTER 6. RESULTS AND DISCUSSION 45

0
10

1
10

2
10 UW1 and UW2
Mean squared error

UW1 and UW3


UW1 and UW4
3
10

4
10

5
10

6
10
0 5 10 15 20
Eb/N0 [dB]

Figure 6.8: Performance of the time domain estimator using two non-subsequent UWs for
 = 0.3 in an AWGN channel.

2
10
Weighted average
Simple average

3
= 0,05
10
Mean squared error

4
= 0.01
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.9: Performance of the PTA estimator after equalization with the LMMSE equal-
izer for  = 0.01 and  = 0.05 in an AWGN channel.
CHAPTER 6. RESULTS AND DISCUSSION 46

among all the three estimators as it employs a Wiener smoother, which reduces the noise
signicantly on the subcarriers. Therefore, the noise inuence on the pilots gets reduced,
thus leading to a better accuracy in the estimate. This can be observed in the gure
particularly in the low Eb /N0 region. However, at high Eb /N0 values the LMMSE performs
N n2
identical to the BLUE, as the term in (4.20) approaches to zero.
d2

2
10
CI equalizer
LMMSE equalizer
BLUE equalizer

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.10: Performance of the PTA estimator after equalization with CI, BLUE, and
LMMSE equalizers for  = 0.01 in an AWGN channel.

Fig. 6.11 plots the MSE curves of the CFO estimators for  = 0.05 after equalization
with CI, BLUE, or LMMSE. Clearly, all the estimators reach an error oor after a certain
Eb /N0 point because of the eect of ICI induced due to the CFO on the pilot subcarriers.
Here, the LMMSE outperforms the CI and BLUE particularly in low Eb /N0 regions.

In the DD approach, the CFO is estimated by measuring the phase dierence between
the received data symbols and its corresponding sliced symbols. The mean power of the
data symbols is normalized to unity and a total number of 32 data symbols are present
in each OFDM symbol. Fig. 6.12 illustrates the performance of the estimators with CI,
LMMSE, and BLUE equalizers for  = 0.01. In addition, it also compares the perfor-
mance between the estimators utilizing PTA and DD schemes. Clearly, the estimators
with the DD scheme follow the same trend like as the ones with the PTA approach. The
LMMSE outperforms the CI and BLUE equalizers. However, at high Eb /N0 values, the
BLUE approaches the LMMSE performance. It can be observed from the gure that the
MSE curves of the data aided estimators always lie below the MSE curves of the PTA
scheme. This is because of the fact that an averaging is done over 32 data symbols in
CHAPTER 6. RESULTS AND DISCUSSION 47

2
10
CI equalizer
LMMSE equalizer
BLUE equalizer

Mean squared error

3
10

4
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.11: Performance of the PTA estimator after equalization with CI, BLUE, and
LMMSE equalizers for  = 0.05 in an AWGN channel.

the DD scheme, whereas in the PTA approach only 4 symbols (in this case pilots) are
included in the estimation. Although the decisions in the DD scheme are hard decisions,
the estimator works very well because for  = 0.01, the phase rotation experienced by the
data symbols is very small, therefore the probability of each data symbol to fall in another
decision region is very low. Consequently, most of the data symbols are demapped to their
corresponding right symbols in the constellation diagram. However, when  = 0.05 as
shown in Fig. 6.13, the MSE curves of all the DD estimators, regardless of their underlying
equalizer, reached an early error oor compared to that of when  = 0.01. The reason is
simply because for  = 0.05 the phase rotation as well as the ICI on the data symbols
is large and the probability that the data symbols fall in another decision region is high,
thus making most of the data symbols to be sliced to a wrong plane, leading to errors in
the estimates. Therefore, in DD schemes the error oor is due to the phase ambiguity of
the data symbols as well as the ICI. In contrast, for the PTA approach, as the transmitted
symbols are known to the receiver, the phase ambiguity is resolved, and the error oor is
obtained only due to the ICI.

The estimates obtained from both the PTA and DD schemes can be combined to get
an estimate that outperforms both individual schemes. Assuming that p
and d
are
the estimates from the PTA and DD schemes, respectively, then, the combined estimate
c
can be obtained as
c =
p + (1 )
d , (6.1)

where is an appropriate weighting factor and is chosen such that p } +


2 MSE {
CHAPTER 6. RESULTS AND DISCUSSION 48

2
10
Pilots with CI equalizer
Pilots with LMMSE equalizer
Pilots with BLUE equalizer
Data with CI equalizer
3
10 Data with LMMSE equalizer
Data with BLUE equalizer
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.12: Performance of the frequency domain estimators after equalization with CI,
BLUE, and LMMSE equalizers for  = 0.01 in an AWGN channel.

2
10
Pilots with CI equalizer
Pilots with LMMSE equalizer
Pilots with BLUE equalizer
Data with CI equalizer
Data with LMMSE equalizer
3 Data with BLUE equalizer
10
Mean squared error

4
10

5
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.13: Performance of the frequency domain estimators after equalization with CI,
BLUE, or LMMSE equalizers for  = 0.05 in an AWGN channel.
CHAPTER 6. RESULTS AND DISCUSSION 49

(1 )2 MSE {
d } is minimized. The value is calculated as

MSE { d }
= . (6.2)
p } + MSE {
MSE { d }

Fig. 6.14 and Fig. 6.15 depict the MSE curves of the individual as well as the combinational
schemes for  = 0.01 and  = 0.05, respectively. It can be observed that the combina-
tional estimator clearly outperforms both individual schemes. Note that the performance
gain of this combined estimator is lower for  = 0.01 compared to  = 0.05.

2
10
Pilots LMMSE
Data LMMSE
Combined

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.14: Performance of the combined estimator (PTA and DD) with LMMSE equalizer
when  = 0.01 in an AWGN channel.

In the following, a comparison between time domain and frequency domain estimators is
illustrated and analyzed. To fairly compare the time domain estimation with the frequency
domain estimation, the energy of the UW is scaled such that it is equal to the energy of
the pilots by using Parseveals theorem, leading to the relation

H p
p
xH
u xu = . (6.3)
N
Fig. 6.16 compares the time domain estimation (i.e., UWs) and the frequency domain es-
timation (i.e., using pilots and data) for  = 0.01. It can be noticed that the frequency
domain estimators (pilot based as well as data aided) outperform the time domain esti-
mators until a certain Eb /N0 range. This is because frequency domain estimators reach
an error oor due to the ICI caused by , whereas the time domain estimator improves
continuously with the increase in Eb /N0 . One can observe that after a certain Eb /N0
CHAPTER 6. RESULTS AND DISCUSSION 50

2
10
Pilots LMMSE
Data LMMSE
Combined

3
10
Mean squared error

4
10

5
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.15: Performance of the combined estimator (PTA and DD) with LMMSE equalizer
when  = 0.05 in an AWGN channel.

value, the time domain estimator outperforms the frequency domain estimators because
of its continuous improvement. It can be veried from Fig. 6.17 and Fig. 6.16 that the
time domain estimator is independent of the specic value  while the performance of
the frequency domain estimators heavily depend on the values of , as the estimators are
highly inuenced by the ICI.
CHAPTER 6. RESULTS AND DISCUSSION 51

1
10
PTA approach LMMSE
DD approach LMMSE
2 Time domain approach
10

3
10
Mean squared error

4
10

5
10

6
10

7
10
0 5 10 15 20 25 30 35 40 45
Eb/N0 [dB]

Figure 6.16: Performance comparison between time domain and frequency domain CFO
estimators for  = 0.01 in an AWGN channel.

1
10
PTA approach LMMSE
DD approach LMMSE
Time domain approach
2
10
Mean squared error

3
10

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.17: Performance comparison between time domain and frequency domain CFO
estimators for  = 0.05 in an AWGN channel.
CHAPTER 6. RESULTS AND DISCUSSION 52

6.3 Time-Dispersive Environment


This section rst analyzes the behavior of UWs in a multipath environment followed by a
discussion on CFO estimation results in the time and frequency domain. Fig. 6.18 presents
two subsequent UWs after propagating through a multipath channel in the presence of
CFO. The CIR is modeled as in (2.24), consisting of 16 taps, and it is assumed that
the CIR is static for the entire OFDM burst. The UWs are chosen to be xu = 1. It
can be observed from the gure that, although the UWs are completely corrupted by the
multipath channel, the phase dierence between the samples of the UWs caused due to the
CFO is clearly visible. However, the UWs are now more complexly correlated with each
other than a simple pairwise correlation seen in the AWGN channel case, thus making
the situation more dicult. Another interesting fact is that the UW is inuenced by the
CIR, and therefore the UW contains the entire channel information. This information can
advantageously be used for channel estimation purposes [46].

UW1
UW2
1 data3
data4
data5

0.5
Imaginary Part

0.5

1.5 1 0.5 0 0.5 1 1.5


Real Part

Figure 6.18: Two subsequent UWs after the multipath channel and when  = 0.05 and
n2 = 0.

Since the UW is part of the DFT window, it can be equalized at the receiver similar
to the data symbols. Therefore, by exploiting this benet, the UWs are equalized rst by
applying the CI equalizer, and then the equalized UWs can be used for CFO estimation.
However, in practice, the equalization is not perfect in the presence of CFO, which will be
analyzed in the following part. Fig. 6.19 shows two subsequent UWs before and after the
equalization for a small and a large frequency oset. It can be easily noticed that for a
small oset the UWs are almost perfectly equalized (i.e., the UWs lie on the unit circle),
and thus the UWs are only left with the phase rotation caused by CFO. However, in case
CHAPTER 6. RESULTS AND DISCUSSION 53

of a large oset, after the equalization the samples of the UWs are still spread around the
unit circle, although there is no noise introduced in the system. Hence, this is just because
of the imperfect equalization.
The CFO estimation in the time domain is carried out by comparing the phase dierence

Figure 6.19: Two subsequent UWs before and after the equalization when (a)  = 0.01
(b)  = 0.05 and n2 = 0.

between the corresponding sample pairs in two successive UW-OFDM symbols. Fig. 6.20
and Fig. 6.21 depict the MSE curves of the time domain estimator in a multipath channel
environment for  = 0.01 and  = 0.05, respectively. For comparison purposes, the
AWGN MSE curve is also plotted as reference. To show the potential of the estimator, the
averaging is done over 5000 dierent multipath channels. The estimator is derived under
the assumption that the received time domain samples, which are lying outside the UW,
are uncorrelated with each other. However, the presence of multipath introduces corre-
lations among them, thus violating this assumption. Therefore, the performance of the
estimator degrades signicantly in a multipath channel compared to that of the AWGN
channel. One can observe that the MSE curve of the estimator using unequalized UWs
reaches an error oor in multipath, because after a certain Eb /N0 value, the inuence of
the noise on the estimator is negligible, and the MSE will not decrease further beyond
this Eb /N0 point. Since the UW is part of the DFT window, the UW can be equalized
by applying the CI equalizer on the received OFDM symbol. However, Fig. 6.20 shows
that the performance of the estimator using the equalized UWs is worse compared to that
of the unequalized UWs, except for low Eb /N0 region. This is probably due to the rea-
son that the CI equalizer is no longer considered to be a valid equalizer in the presence of
CFO. Therefore, it can be concluded that, on equalizing the multipath channel with the CI
equalizer, the performance of the estimator will degrade signicantly at low Eb /N0 values.
An interesting observation that can also be made is that the estimator performance with
the unequalized UWs reaches an error oor after a certain Eb /N0 , whereas the estimator
performance with the equalized UWs goes on improving with the increase in the Eb /N0 . It
is due to the fact that as Eb /N0 increases, the inuence of the noise on the data decreases,
and the CI equalizer will work reasonable in this region.
CHAPTER 6. RESULTS AND DISCUSSION 54

Fig. 6.22 shows that the performance of the estimator is more dependent on the con-
crete amount of the CFO compared to the AWGN case. It can also be noticed that when
 exceeds 0.25, the estimator performance starts degrading (especially at low Eb /N0 val-
ues) and at  = 0.51, the estimator performs completely worse as  is out of range.

1
10
AWGN
Multipath (unequalized UWs)
2
10 Multipath (equalized UWs CI equalizer)

3
10
Mean squared error

4
10

5
10

6
10

7
10

8
10
0 5 10 15 20 25 30 35 40
Eb/N0 [dB]

Figure 6.20: Performance comparison of the time domain estimator between an AWGN
and a multipath environment for  = 0.01.

Fig. 6.23 shows the performance of the estimator using two non-subsequent UWs when
 = 0.01. This reveals that there is a signicant improvement in the performance, if the
distance between the correlating UWs increases on the one hand, but the estimation range
decreases on the other hand.

In frequency domain, the CFO can be estimated using PTA or DD schemes. As the
CFO estimation procedure using these techniques was already discussed in the previous
section (i.e., for the AWGN channel case), the focus is devoted to analyze the results. The
simulations are performed with the pilot based model having zero UW. Fig. 6.24 shows the
performance of the estimator after applying the LMMSE equalizer when  = 0.01. The
weighting of the estimates is done with the inverse of the corresponding diagonal elements
in the error covariance matrix given in (4.21). It is clearly visible from the gure that
weighting the estimates has a huge performance gain over the simple average with regard
to the MSE. This is because as the channel is frequency selective, the pilots that fall into
deep spectral notches are heavily inuenced and attenuated by the noise and channel,
respectively, therefore, the estimates from those pilots should be weighted less. Fig. 6.25
plots the MSE curve of the PTA estimator with the LMMSE equalizer when  = 0.05.
CHAPTER 6. RESULTS AND DISCUSSION 55

1
10
AWGN
Multipath (unequalized UWs)
2
10 Multipath (equalized UWs CI equalizer)

3
10
Mean squared error

4
10

5
10

6
10

7
10

8
10
0 5 10 15 20 25 30 35 40
Eb/N0 [dB]

Figure 6.21: Performance comparison of the time domain estimator between an AWGN
and a multipath environment for  = 0.05.

1
10

0
10

1 = 0.01
10
Mean squared error

= 0.25
= 0.30
2 = 0.40
10
= 0.51

3
10

4
10

5
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.22: Performance of the time domain estimator for dierent CFOs in a multipath
environment.
CHAPTER 6. RESULTS AND DISCUSSION 56

2
10
UW1 and UW2
UW1 and UW3
UW1 and UW4

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.23: Performance of the time domain estimator using two non-subsequent UWs for
 = 0.01 in a multipath environment.

It can be observed that the curves have an early error oor compared to the curves when
 = 0.01. This is due to the fact that the estimator performance is limited by the ICI
caused due to the CFO. The larger the CFO, the more intense the ICI, and the lower is
the performance of the estimator. It can also be noticed from Fig. 6.25 that after a certain
Eb /N0 value, the weighted average scheme performs worse than the simple average. This
is because from that Eb /N0 value the inuence of noise will become negligible compared to
the inuence of ICI. However, the weighting coecients are still extracted from the error
covariance matrix (see (4.21)) which is no longer a useful means as the elements in the
matrix will not consider the ICI.

Fig. 6.26 and 6.27 illustrate the performance of the PTA based CFO estimator with
CI, LMMSE, and BLUE equalizers when  = 0.01 and  = 0.05, respectively. Both
results show that the estimator with the CI equalizer has the worst performance, because
it completely ignores the correlations among the subcarriers, whereas the LMMSE and
BLUE equalizers perform identical at all shown Eb /N0 values. The MSE curves of all the
estimators at  = 0.05 exhibit an early error oor compared to that of the MSE curves
when  = 0.01 because of the intense ICI induced due to the large CFO.

Fig. 6.28 and 6.29 compare the performance of the PTA estimators with LMMSE equal-
ization in both AWGN and multipath channels for  = 0.01 and  = 0.05, respectively.
It can be noted that the equalizer performs very well in mitigating the multipath eects in
the presence of CFO, and thus the MSE curve for the multipath case almost approaches
CHAPTER 6. RESULTS AND DISCUSSION 57

2
10
Weighted average
Simple average

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.24: Performance of the PTA estimator with the LMMSE equalizer for  = 0.01
in a multipath environment.

2
10
Weighted average
Simple average

3
10
Mean squared error

4
10

5
10
0 5 10 15 20 25 30 35 40 45
Eb/N0 [dB]

Figure 6.25: Performance of the PTA estimator with the LMMSE equalizer for  = 0.05
in a multipath environment.
CHAPTER 6. RESULTS AND DISCUSSION 58

2
10
CI equalizer
LMMSE equalizer
BLUE equalizer

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.26: Performance of the PTA estimator after equalization with CI, BLUE, and
LMMSE equalizers for  = 0.01 in a multipath environment.

2
10
CI equalizer
LMMSE equalizer
BLUE equalizer
Mean squared error

3
10

4
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.27: Performance of the PTA estimator after equalization with CI, BLUE, and
LMMSE equalizers for  = 0.05 in a multipath environment.
CHAPTER 6. RESULTS AND DISCUSSION 59

the AWGN curve. Similar tendencies can be observed with the BLUE as well as with the
CI equalizer (not shown in this work).

2
10
AWGN
Multipath

3
10
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.28: Performance comparison of the PTA estimator with LMMSE equalizer be-
tween an AWGN and a multipath environment for  = 0.01.

The performance of DD schemes in multipath channels are illustrated in Fig. 6.30


and 6.31 for  = 0.01 and  = 0.05, respectively. In the same gures the MSE curves of
a PTA estimator is also plotted for comparison. It can be observed that the MSE curves
of the data aided estimators always lie below the MSE curves of the PTA scheme because
an averaging is done over more symbols. As the equalizers are applied before the CFO
estimation, the multipath eect on the data symbols is almost eliminated, resulting in MSE
curves very similar to the AWGN case (see Fig. 6.12 and Fig. 6.13). For  = 0.05, the
MSE curves of the DD scheme reach an early error oor compared to that of  = 0.01,
because of more wrong decisions made due to the rotation of the received symbols due to
CFO and also with the increased ICI.

The following illustrates the time domain versus frequency domain estimation methods
in multipath channels for dierent CFOs. Fig. 6.32 and 6.33 plot the MSE curves of the
time domain and frequency domain estimators for  = 0.01 and  = 0.05, respectively.
Note that, for comparison purposes, the UW energy is scaled such that it has the same
energy as the pilots. As a result, the energy of the UW used in these simulations is very less
compared to that of the UW used in the previous simulations (see Fig. 6.20 and Fig. 6.21).
It can be noticed from Fig. 6.32 and 6.33 that, unlike in AWGN, the time domain method is
worse compared to the frequency domain estimation methods. Although the performance
CHAPTER 6. RESULTS AND DISCUSSION 60

2
10
AWGN
Multipath

Mean squared error

3
10

4
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.29: Performance comparison of the PTA estimator with LMMSE equalizer be-
tween an AWGN and a multipath environment for  = 0.05.

2
10
Pilots with CI equalizer
Pilots with LMMSE equalizer
Pilots with BLUE equalizer
Data with CI equalizer
3
10 Data with LMMSE equalizer
Data with BLUE equalizer
Mean squared error

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.30: Performance of the frequency domain estimators after equalization with CI,
BLUE, and LMMSE equalizers for  = 0.01 in a multipath environment.
CHAPTER 6. RESULTS AND DISCUSSION 61

2
10
Pilots with CI equalizer
Pilots with LMMSE equalizer
Pilots with BLUE equalizer
Data with CI equalizer
Data with LMMSE equalizer
3 Data with BLUE equalizer
10
Mean squared error

4
10

5
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.31: Performance of the frequency domain estimators after equalization with CI,
BLUE, and LMMSE equalizers for  = 0.05 in a multipath environment.

of the time domain estimation method is independent of the amount of CFO, whereas the
performance of frequency domain estimation methods heavily depend on the CFO, the time
domain estimator has an early error oor in multipath channels compared to frequency
domain estimators. The reason is obvious. In frequency domain, there is a possibility
to remove the inuence of the multipath channel on the data, from which  has to be
estimated, but in time domain the estimation is still carried out based on the multipath
corrupted UW because the equalization of UWs with the CI equalizer is not performing
well.
CHAPTER 6. RESULTS AND DISCUSSION 62

1
10
PTA approach LMMSE
DD approach LMMSE
Time domain approach
2
Mean squared error 10

3
10

4
10

5
10

6
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.32: Performance comparison between time domain and frequency domain CFO
estimators for  = 0.01 in a multipath environment.

1
10
PTA approach LMMSE
DD approach LMMSE
Time domain approach

2
10
Mean squared error

3
10

4
10

5
10
0 5 10 15 20 25 30
Eb/N0 [dB]

Figure 6.33: Performance comparison between time domain and frequency domain CFO
estimators for  = 0.05 in a multipath environment.
CHAPTER 6. RESULTS AND DISCUSSION 63

6.4 CFO Impact on UW-OFDM and CP-OFDM


This section analyzes the sensitivity of the UW-OFDM system on the CFO, and also
compares it with that of the CP-OFDM system. An OFDM burst having 20 OFDM sym-
bols is considered for one iteration run, and in total 2000 iterations (i.e., 2000 bursts)
are performed for each  value, completing one simulation run. For all the simulations, a
traditional UW-OFDM system is considered, and the CP-OFDM system is built like as
in the 802.11a standard [34]. It is assumed that  is constant for the entire burst. The
phase rotation factor due to the CFO is assumed to start from zero at the beginning of
each OFDM burst, and each sample in the burst is multiplied with its corresponding phase
j2n
rotation factor e N , where n represents the index of the sample. The tracking mechanism
discussed in the Section 5.2 is implemented so that the phase accumulation due to the CFO
from one OFDM symbol to the other OFDM symbol is compensated. Briey summarized,
the tracking mechanism is implemented in the following way. The phase rotation of the
received pilot symbols caused due to the CFO is estimated from the current OFDM sym-
bol. Then, all the data symbols in the subsequent OFDM symbols including the current
symbol is de-rotated by the estimated phase angle in the frequency domain. Thereafter,
again the phase of the received pilots from the next OFDM symbol is estimated and the
subsequent symbols are corrected. This process is repeated until the estimation reaches
the last OFDM symbol in the burst. After the correction process, the data symbols are
then estimated out of the corrected OFDM symbols based on the linear model described
in (4.16). The performance of the systems is analyzed by the resulting mean squared error.

Consider that a frequency oset  is introduced in both systems. For the moment
assume that the channel is a distortion and a noise free channel. An UW-OFDM system
having zero UW is considered for the following simulation
 results.
 Fig. 6.34 shows the
2
MSE curves of the frequency domain symbols (i.e., E x

x ) of UW-OFDM and CP-

OFDM systems plotted over an  range of 0 to 1. It can be observed that the MSE of
the frequency domain symbols in both systems increases with increasing  because of the
increasing ICI inuence. It is important to notice that the UW-OFDM system exhibits
less sensitivity to the CFO compared to the CP-OFDM scheme, and this turns out to
be a considerable advantage of the UW-OFDM scheme. Note that in UW-OFDM, the
frequency domain symbol x
is ), redundant subcarriers
composed of data subcarriers (d
(
r), and pilot subcarriers (p
), while for CP-OFDM the frequency domain symbol x
is
composed of only the data and pilot subcarriers. The MSE curve of the data symbols after
performing the CI equalizer coincides exactly with that of the LMMSE equalizer. Note
that, the CI equalizer for a distortion free channel is an Identity matrix. This shows that
the equalization with the LMMSE equalizer has no additional advantage due to the fact
that this equalizer will only account for the multipath propagation and AWGN, but will
not combat over the CFO.

It is also interesting to see the sensitivity of both systems within a short range of CFO,
which is more realistic in case of a ne CFO estimation. Fig. 6.35 shows the MSE curves
of the frequency domain symbols (i.e., x
) of UW-OFDM and CP-OFDM system for an 
range of 0 to 0.1. It can be observed that even in this range, the UW-OFDM is less sensitive
CHAPTER 6. RESULTS AND DISCUSSION 64

2.5
CPOFDM CI equalizer
CPOFDM data CI equalizer
UWOFDM CI equalizer
2 UWOFDM data CI equalizer
UWOFDM redundant CI equalizer
UWOFDM data LMMSE equalizer
Mean squared error

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.34: MSE curves of the frequency domain symbols of UW-OFDM with zero UW
and CP-OFDM for an  range of 0 to 1 in a distortion and noise free channel.

compared to the CP-OFDM system. It can also be noticed that the error contribution of
the redundant subcarriers (
r) on the MSE of x
is low compared to the error contribution
of the data subcarriers ).
(d

The following illustrates the behavior of the UW-OFDM system in the presence of
CFO, when having dierent UW sequences (other than zeros). Here, two particular UW
sequences xu1 = 1 and xu2 [k] {CN (0, 1)} are considered for the simulations. The energy
of the xu2 is scaled such that it has the same energy as the xu1 . Fig. 6.36 and Fig. 6.37
depict the MSE curves of the frequency domain symbols of UW-OFDM and CP-OFDM
systems in the presence of CFO, when utilizing xu1 and xu2 , respectively. It can be observed
that in contrast to the zero UW, in case of a CI equalizer, the frequency domain symbols of
UW-OFDM system with the above UW sequences exhibit very high MSE compared to the
CP-OFDM scheme because of the subtraction of a "wrong" UW. In the presence of CFO,
the UW inuence can be removed completely only if the CFO inuenced UW (i.e., xu1

or xu2 )
is subtracted instead of subtracting the UW itself. However, the subtraction of
the CFO inuenced UW is impossible as  is unknown to the receiver. An interesting fact
) are equalized with the LMMSE equalizer, the MSE of the equalized
is that, if the data (d
data exactly coincides with that of the UW-OFDM system having zero UW, thus making
the MSE of the equalized data to be independent of the chosen UW sequence. This can
be observed in Fig. 6.38.

The following illustrates the sensitivity of UW-OFDM and CP-OFDM systems to  in


CHAPTER 6. RESULTS AND DISCUSSION 65

0.035
CPOFDM CI equalizer
CPOFDM data CI equalizer
0.03 UWOFDM CI equalizer
UWOFDM data CI equalizer
UWOFDM redundant CI equalizer
0.025
UWOFDM data LMMSE equalizer
Mean squared error

0.02

0.015

0.01

0.005

0
0 0.02 0.04 0.06 0.08 0.1
CFO ()

Figure 6.35: MSE curves of the frequency domain symbols of UW-OFDM with zero UW
and CP-OFDM for an  range of 0 to 0.1 in a distortion and noise free channel.

35

30

25
Mean squared error

CPOFDM CI equalizer
20 CPOFDM data CI equalizer
UWOFDM CI equalizer
UWOFDM data CI equalizer
15
UWOFDM redundant CI equalizer
UWOFDM data LMMSE equalizer
10

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.36: MSE curves of the frequency domain symbols of UW-OFDM with xu1 and
CP-OFDM for an  range of 0 to 1 in a distortion and noise free channel.
CHAPTER 6. RESULTS AND DISCUSSION 66

45

40

35

30
Mean squared error

CPOFDM CI equalizer
25 CPOFDM data CI equalizer
UWOFDM CI equalizer
UWOFDM data CI equalizer
20
UWOFDM redundant CI equalizer
UWOFDM data LMMSE equalizer
15

10

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.37: MSE curves of the frequency domain symbols of UW-OFDM with xu2 and
CP-OFDM for an  range of 0 to 1 in a distortion and noise free channel.

2.5
CPOFDM data CI equalizer
UWOFDM data zero UW LMMSE equalizer
UWOFDM data x LMMSE equalizer
u1
2 UWOFDM data x LMMSE equalizer
u2
Mean squared error

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.38: MSE curves of the estimated data symbols of UW-OFDM and CP-OFDM for
an  range of 0 to 1 in a distortion and noise free channel.
CHAPTER 6. RESULTS AND DISCUSSION 67

a multipath environment. The UW is set to zero. Fig. 6.39 presents the MSE curves of the
frequency domain symbols of UW-OFDM and CP-OFDM in a multipath channel. It can
be observed that the MSE of the frequency domain symbols of an UW-OFDM system is
less compared to that of the CP-OFDM scheme, thus making the UW-OFDM system less
susceptible to the CFO in a multipath environment. Note that the data symbols that are
equalized with an LMMSE equalizer result in a lower MSE compared to the case of a CI
equalizer. Therefore, in the presence of CFO and a multipath channel, the LMMSE equal-
izer performs optimum out of all the equalizers (i.e., CI, BLUE, and LMMSE) discussed
in this work.

3
CPOFDM CI equalizer
CPOFDM data CI equalizer
UWOFDM CI equalizer
2.5
UWOFDM data CI equalizer
UWOFDM redundant CI equalizer
UWOFDM data LMMSE equalizer
2
Mean squared error

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.39: MSE curves of the frequency domain symbols of UW-OFDM with zero UW
and CP-OFDM in a multipath environment (without noise).

Fig. 6.40 and Fig. 6.41 compare the MSE curves of the frequency domain symbols of
UW-OFDM and CP-OFDM in a multipath and a no multipath case, for an  range of 0
to 1 and 0 to 0.1, respectively. It can be observed that, the LMMSE equalizer removes
the inuence of the multipath channel very well in the presence of CFO, and as a result
the MSE curve of the data symbols in a multipath environment approaches almost the
no multipath case. However, in CP-OFDM the CI equalizer cannot perform well in the
presence of CFO, leading to a huge dierence between the MSE curves of the data symbols
for a multipath and a no multipath case.

Fig. 6.42 shows the MSE curves of the frequency domain symbols of UW-OFDM uti-
lizing xu1 as UW and CP-OFDM in a multipath environment. Similar to the no multipath
(i.e., distortion and noise free) case, the MSE of the frequency domain symbols of UW-
CHAPTER 6. RESULTS AND DISCUSSION 68

2.5
CPOFDM data CI equalizer multipath
CPOFDM data CI equalizer no multipath
UWOFDM data LMMSE equalizer multipath
Mean squared error 2 UWOFDM data LMMSE equalizer no multipath

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.40: MSE curves of the frequency domain symbols of UW-OFDM with zero UW
and CP-OFDM in a multipath (without noise) and a distortion and noise free case.

0.1
CPOFDM data CI equalizer multipath
0.09 CPOFDM data CI equalizer no multipath
UWOFDM data LMMSE equalizer multipath
0.08 UWOFDM data LMMSE equalizer no multipath

0.07
Mean squared error

0.06

0.05

0.04

0.03

0.02

0.01

0
0 0.02 0.04 0.06 0.08 0.1
CFO ()

Figure 6.41: MSE curves of the frequency domain symbols of UW-OFDM with zero UW
and CP-OFDM in a multipath (without noise) and a distortion and noise free case.
CHAPTER 6. RESULTS AND DISCUSSION 69

OFDM is very high compared to that of the CP-OFDM scheme. It is important to note
that, even the MSE of the data symbols equalized with LMMSE equalizer is high compared
to the CP-OFDM scheme. This means that, in the presence of CFO and multipath propa-
gation, the UW-OFDM system with an UW sequence (other than zeros) is more sensitive
to CFO compared to the CP-OFDM scheme. Fig. 6.43 and Fig. 6.44 depict the MSE of
the equalized data symbols of UW-OFDM with dierent UW sequences, for an  range of
0 to 1 and 0 to 0.1, respectively. It reveals that the performance of the LMMSE equalizer
is no longer independent of the UW sequence. It entirely depends on the UW sequence.

90
CPOFDM data CI equalizer
UWOFDM CI equalizer
80 UWOFDM data CI equalizer
UWOFDM redundant CI equalizer
UWOFDM data LMMSE equalizer
70

60
Mean squared error

50

40

30

20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
CFO ()

Figure 6.42: MSE curves of the frequency domain symbols of UW-OFDM with xu1 and
CP-OFDM in a multipath environment (without noise).
CHAPTER 6. RESULTS AND DISCUSSION 70

20

18

16

14
Mean squared error

12
CPOFDM data CI equalizer
10 UWOFDM data zero UW LMMSE equalizer
UWOFDM data xu1 LMMSE equalizer
8 UWOFDM data x LMMSE equalizer
u2

0
0 0.2 0.4 0.6 0.8 1
CFO ()

Figure 6.43: MSE curves of the equalized data symbols of UW-OFDM with dierent UW
sequences and CP-OFDM in a multipath environment (without noise).

25

20
Mean squared error

15

10
CPOFDM data CI equalizer
UWOFDM data zero UW LMMSE equalizer
UWOFDM data xu1 LMMSE equalizer
5 UWOFDM data xu2 LMMSE equalizer

0
0 0.02 0.04 0.06 0.08 0.1
CFO ()

Figure 6.44: MSE curves of the equalized data symbols of UW-OFDM and CP-OFDM in
a multipath environment (without noise).
Chapter 7

Conclusions
The main aim of this work is to investigate the Carrier Frequency Oset (CFO) issue in
an UW-OFDM system. For that, eects of CFO in time and frequency domain have been
discussed. An introduction of a small CFO leads to a phase rotation of samples in time
domain, whereas in frequency domain, CFO leads to Inter Carrier Interference (ICI) along
with an amplitude distortion and phase rotation of the data symbols. One major goal of
this work is to evaluate the performance of time domain against frequency domain based
CFO estimators. In case of frequency domain estimators, pilot tones are used for the es-
timation. Therefore, for enabling the insertion of pilot tones in an UW-OFDM symbol, a
modied transmit signal structure has been introduced.

CFO estimation is normally carried out in two phases, namely an acquisition phase
(coarse CFO estimation) and a tracking phase (ne CFO estimation). However, this thesis
only focuses on the investigation of algorithms for the ne CFO estimation. In the time
domain, a Maximum Likelihood (ML) CFO estimation algorithm utilizing the correlation
between two successive UWs has been presented. This algorithm was derived under the
assumption that the signal is only eected by the AWGN. The performance of the esti-
mator mainly depends on two factors, one is the length of the UW, and the other is the
correlation coecient between the corresponding samples of two UWs. Note that, the per-
formance of the estimator is independent of the shape of the UW and also, the performance
improves signicantly with the increase of the UW length, as an averaging is done over
more samples. Simulation results proved that the time domain estimator performance is
independent of the value of the CFO in both AWGN and multipath environments, provided
that the CFO is in the estimation range (-0.5 and +0.5). As the estimator was not derived
under the multipath assumption, the performance of the estimator degrades signicantly
in a multipath channel compared to that of the AWGN channel. Since the UW is part
of the DFT window, the UW can be equalized by applying the Channel Inversion (CI)
equalizer on the received OFDM symbol. The simulation results proved that the estimator
using the equalized UWs is worse compared to that of the unequalized UWs. It seems that
a CI equalizer is no longer an appropriate equalizer in the presence of CFO in a multipath
environment.

In contrast to CP-OFDM, UW-OFDM has the unique feature that the guard inter-
val and thus the UW is the same for all OFDM symbols. Hence, the estimation can be

71
CHAPTER 7. CONCLUSIONS 72

performed utilizing two non-subsequent UWs separated by several OFDM symbols. As


a result, the estimator error variance decreases signicantly. This is because the former
one (using two successive UWs) has to estimate a small phase for a given Eb /N0 value,
while the latter one has to estimate a larger phase (because the UWs are now non subse-
quent) having the same Eb /N0 value, thereby reducing the error variance of the estimator.
However, the estimation range of this new estimator is decreased with an increase in the
distance between the two UWs.

In frequency domain, a Pilot Tone Aided (PTA) technique, where the CFO is estimated
using the phase shift introduced on the received pilot symbols, has been presented. At the
receiver, the pilots are estimated like they were data symbols by applying a linear equalizer
at rst, and then the CFO is estimated out of these estimated pilots. In this work, the
LMMSE, the BLUE, and the CI equalizers are considered. Besides this PTA technique, a
Decision Directed (DD) method, in which the CFO is estimated by comparing the received
frequency domain data symbols with their corresponding sliced version symbols, has also
been investigated. The data symbols are also equalized rst before the estimation is carried
out. Since the slicing of the data symbols is done by tentative/hard decisions, this might
lead to errors in the estimate when wrong decisions are made. The estimator accuracy can
be increased by weighting each estimate dierently with the inverse of the corresponding
elements in the error covariance matrix of the estimator.

It turns out that the LMMSE and the BLUE equalizer perform almost identically as
they exploit redundancy introduced among the frequency domain data in a similar way,
whereas the CI equalizer has performed worse as it does not exploit the redundancy. The
equalizers perform very well in mitigating the multipath eects in the presence of CFO,
hence the MSE curves for the multipath case almost approach the AWGN curves. Besides
that, simulations show that the MSE curves of the DD estimators always lie below the MSE
curves of the PTA scheme. In general, the performance of all frequency domain estimators
degrades signicantly and reaches an early error oor with an increasing CFO. This is the
result of an increasing ICI on the data symbols with an increasing CFO.

Comparing time and frequency domain estimation methods with each other, the fol-
lowing conclusions can be drawn: The performance of the time domain estimation method
is independent of the amount of CFO, while the performance of the frequency domain
estimation methods heavily depend on the CFO. In case of a multipath environment, the
time domain estimator is worse compared to the frequency domain estimation methods, as
in time domain the estimation has been carried out based on the multipath corrupted UW
(as the equalizer even worsens the results), whereas in frequency domain the equalizers
mitigate the eect of the multipath channel very well.

Furthermore, the sensitivity of an UW-OFDM system on the CFO has been analyzed.
It is shown that, for an UW-OFDM with zero UW, the LMMSE and CI equalizers perform
almost the same in both distortion free and multipath (no noise) environments. However,
when a UW other than a zero word is considered, the CI equalizer performs completely
worse, while the LMMSE still performs well although the performance depends on the UW
sequence as well as the channel characteristics. Additionally, simulation results proved
CHAPTER 7. CONCLUSIONS 73

that, in a distortion free environment, the sensitivity of the UW-OFDM scheme is inde-
pendent of the UW sequence, and also less sensitive compared to the CP-OFDM scheme,
whereas in a multipath environment, the sensitivity of the UW-OFDM scheme entirely
depends on the UW sequence.
Bibliography
[1] M. Huemer, C. Hofbauer, and J. Huber, The Potential of Unique Words in OFDM,
in Proceedings of the 15th International OFDM Workshop (InOWo' 10), Hamburg,
Germany, 2010, pp. 140144.

[2] J. Bingham, Multicarrier Modulation for Data Transmission: An Idea Whose Time
has Come, IEEE Communications Magazine, vol. 28, no. 5, pp. 514, 1990.
[3] A. Peled and A. Ruiz, Frequency Domain Data Transmission Using Reduced Compu-
tational Complexity Algorithms, inProceedings of the IEEE International Conference
on Acoustics, Speech, and Signal Processing (ICASSP'80), vol. 5, Denver, Colorado,
1980, pp. 964967.

[4] M. Alard and R. Lassalle, Principles of Modulation and Channel Coding for Digital
Broadcasting for Mobile Receivers, EBU review, vol. 224, pp. 4769, 1987.
[5] R. Van Nee, G. Awater, M. Morikura, H. Takanashi, M. Webster, and K. Halford,
New High-rate Wireless LAN Standards, IEEE Communications Magazine, vol. 37,
no. 12, pp. 8288, 1999.

[6] S. Hara and R. Prasad, Multicarrier Techniques for 4G Mobile Communications.


Artech House Publishers, 2003.

[7] R. Prasad and R. Van Nee, OFDM for Wireless Multimedia Communications. Artech
House Publishers, Boston, 2000.

[8] S. Tang, F. Yang, K. Peng, C. Pan, K. Gong, and Z. Yang, Iterative Channel Esti-
mation for Block Transmission with Known Symbol Padding - A New Look at TDS-
OFDM, inProceedings of the IEEE Global Telecommunications Conference (GLOBE-
COM'07), Washington, USA, 2007, pp. 42694273.
[9] B. Muquet, Z. Wang, G. Giannakis, M. De Courville, and P. Duhamel, Cyclic Pre-
xing or Zero Padding for Wireless Multicarrier Transmissions ? IEEE Transactions
on Communications, vol. 50, no. 12, pp. 21362148, 2002.
[10] M. Huemer, A. Onic, and C. Hofbauer, Classical and Bayesian Linear Data Esti-
mators for Unique Word OFDM, IEEE Transactions on Signal Processing, vol. 59,
no. 12, pp. 60736085, Dec. 2011.

[11] H. Witschnig, T. Mayer, A. Springer, A. Koppler, L. Maurer, M. Huemer, and


R. Weigel, A Dierent Look on Cyclic Prex for SC/FDE, in Proceedings of the

74
BIBLIOGRAPHY 75

13th IEEE International Symposium on Personal, Indoor and Mobile Radio Commu-
nications, vol. 2, Lisbon, Portugal, 2002, pp. 824828.
[12] T. Pollet and M. Moeneclaey, The Eect of Carrier Frequency Oset on the Perfor-
mance of Band Limited Single Carrier and OFDM Signals, in Proceedings of the IEEE
Global Telecommunications Conference (GLOBECOM'96), vol. 1, London, England,
1996, pp. 719723.

[13] J. Van de Beek, M. Sandell, and P. Borjesson,  ML Estimation of Time and Frequency
Oset in OFDM Systems, IEEE Transactions on Signal Processing, vol. 45, no. 7,
pp. 18001805, 1997.

[14] P. Moose, A Technique for Orthogonal Frequency Division Multiplexing Frequency


Oset Correction, IEEE Transactions on Communications, vol. 42, no. 10, pp. 2908
2914, 1994.

[15] H. Chen and G. Pottie, A Comparison of Frequency Oset Tracking Algorithms for
OFDM, in Proceedings of the IEEE Global Telecommunications Conference (GLOBE-
COM'03), vol. 2, San Francisco, USA, 2003, pp. 10691073.
[16] N. Lashkarian and S. Kiaei, Class of Cyclic-Based Estimators for Frequency-Oset Es-
timation of OFDM Systems, IEEE Transactions on Communications, vol. 48, no. 12,
pp. 21392149, 2000.

[17] H. Nezamfar and M. Kahaei, A Comparison of Maximum Likelihood Frequency Oset


Estimation Methods for OFDM Systems, in Proceedings of the International Sympo-
sium on Telecommunications (IST 2008), Tehran, Iran, 2008.
[18] T. Schmidl and D. Cox, Robust Frequency and Timing Synchronization for OFDM,
IEEE Transactions on Communications, vol. 45, no. 12, pp. 16131621, 1997.
[19] M. Morelli, C. Kuo, and M. Pun, Synchronization Techniques for Orthogonal Fre-
quency Division Multiple Access (OFDMA): A Tutorial Review, Proceedings of the
IEEE, vol. 95, no. 7, pp. 13941427, 2007.
[20] W. Chin,  ML Estimation of Timing and Frequency Osets Using Distinctive Cor-
relation Characteristics of OFDM Signals Over Dispersive Fading Channels, IEEE
Transactions on Vehicular Technology, vol. 60, no. 2, pp. 444456, 2011.
[21] T. Fusco and M. Tanda, Blind Synchronization for OFDM Systems in Multipath
Channels, IEEE Transactions on Wireless Communications, vol. 8, no. 3, pp. 1340
1348, 2009.

[22] M. Ghogho and A. Swami, Blind Frequency-Oset Estimator for OFDM Systems
Transmitting Constant-Modulus Symbols, IEEE Communications Letters, vol. 6,
no. 8, pp. 343345, 2002.

[23] L. Deneire, B. Gyselinckx, and M. Engels, Training Sequence Versus Cyclic Prex-A
New Look on Single Carrier Communication, IEEE Communications Letters, vol. 5,
no. 7, pp. 292294, 2001.
BIBLIOGRAPHY 76

[24] M. Huemer, H. Witschnig, and J. Hausner, Unique word Based Phase Tracking Algo-
Proceedings of the IEEE Global Telecommunications
rithms for SC/FDE-Systems, in
Conference (GLOBECOM'03), vol. 1, San Francisco, USA, 2003, pp. 7074.
[25] K. Kim and H. Park, Enhanced Phase Tracking for Unique Word Based SC-FDE on
Frequency Selective Channels, inProceedings of the IEEE International Microwave
Workshop Series on RF Front-ends for Software Dened and Cognitive Radio Solutions
(IMWS), Aveiro, Portugal, 2010, pp. 14.
[26] J. Coon, M. Sandell, M. Beach, and J. McGeehan, Channel and Noise Variance
Estimation and Tracking Algorithms for Unique-Word Based Single-Carrier Systems,
IEEE Transactions on Wireless Communications, vol. 5, no. 6, pp. 14881496, 2006.
[27] M. Huemer, C. Hofbauer, and J. Huber, Unique Word Prex in SC/FDE and OFDM:
A Comparison, in Proceedings of the IEEE GLOBECOM Workshops (GC Wkshps),
Miami, USA, Dec. 2010, pp. 12961301.

[28] A. Onic and M. Huemer, Direct versus Two-Step Approach for Unique Word Gen-
eration in UW-OFDM, in Proceedings of the 15th International OFDM-Workshop,
Hamburg, Germany, Sept. 2010, pp. 145149.

[29] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge


University Press, 2005.

[30] J. Fakatselis,  Criteria for 2.4 GHz PHY Comparison of Modulation Methods, Doc-
ument IEEE 1997.
[31] S. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. Prentice
Hall, 1993.

[32] J. H. C.F. Li, Y.S. Chu and W. Sheen, Cell Search in WCDMA Under Large-
Frequency and Clock Errors: Algorithms to Hardware Implementation, IEEE Trans-
actions on Circuits and Systems I: Regular Papers, vol. 55, no. 2, pp. 659671, Mar.
2008.

[33] Y. Cho, J. Kim, W. Yang, and C. Kang, MIMO-OFDM Wireless Communications


with MATLAB. Wiley, 2010.

[34] IEEE Std 802.11a-1999, Part 11: Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specications: High-Speed Physical Layer in the 5 GHz Band,
IEEE Std., 1999.

[35] F. Classen and H. Meyr,  Frequency Synchronization Algorithms for OFDM Systems
Suitable for Communication Over Frequency Selective Fading Channels, in Proceed-
ings of the IEEE 44th Vehicular Technology Conference, vol. 3, Jun. 1994, pp. 1655
1659.

[36] F. Tufvesson, O. Edfors, and M. Faulkner, Time and Frequency Synchronization for
OFDM using PN-sequence preambles, in Proceedings of the IEEE VTS 50th Vehicular
Technology Conference, VTC 1999 - Fall., vol. 4, 1999, pp. 2203 2207.
BIBLIOGRAPHY 77

[37] M. Morelli and U. Mengali, An Improved Frequency Oset Estimator for OFDM
Applications, IEEE Communications Letters, vol. 3, no. 3, pp. 7577, Mar. 1999.
[38] H. Minn, V. Bhargava, and K. Letaief, A Robust Timing and Frequency Synchroniza-
tion for OFDM Systems, IEEE Transactions on Wireless Communications, vol. 2,
no. 4, pp. 822839, Jul. 2003.

[39] K. Shi and E. Serpedin, Coarse Frame and Carrier Synchronization of OFDM Sys-
tems: A New Metric and Comparison, IEEE Transactions on Wireless Communica-
tions, vol. 3, no. 4, pp. 12711284, Jul. 2004.
[40] J. Gonzalez-Bayon, C. Carreras, and A. Fernandez-Herrero,  A Comparison of Fre-
quency Oset Synchronization Algorithms for WiMAX OFDM Systems, in Proceed-
ings of the EUROCON 2007 The International Conference on Computer as a Tool,
Sept. 2007, pp. 9971004.

[41] S. Kapoor, D. Marchok, and Y. Huang, Pilot Assisted Synchronization for Wireless
OFDM Systems Over Fast Time Varying Fading Channels, in Proceedings of the
IEEE 48th Vehicular Technology Conference (VTC), vol. 3, Ottawa, Canada, May
1998, pp. 20772080.

[42] K. Shi, E. Serpedin, and P. Ciblat, Decision-Directed Fine Synchronization in OFDM


Systems, IEEE Transactions on Communications, vol. 53, no. 3, pp. 408412, Mar.
2005.

[43] Y. Tsai, X. Li, and C. Wei, Data-Carrier Aided Carrier Frequency Oset Estimation
for OFDM Systems, in Proceedings of the IEEE Vehicular Technology Conference
(VTC), Marina Bay, Singapore, 2008, pp. 898902.
[44] S. Moridi and H. Sari, Analysis of Four Decision-Feedback Carrier Recovery Loops
in the Presence of Intersymbol Interference, IEEE Transactions on Communications,
vol. 33, no. 6, Jun. 1985.

[45] K. Shi, E. Serpedin, and P. Ciblat,  Decision-Directed Fine Synchronization for Coded
OFDM Systems, in Proceedings of the IEEE International Conference on Acoustics,
Speech, and Signal Processing,(ICASSP'04), vol. 4, Montreal, Canada, May 2004, pp.
365368.

[46] A. Czylwik, Low Overhead Pilot-Aided Synchronization for Single Carrier Mod-
ulation with Frequency Domain Equalization, in Proceedings of the IEEE Global
Telecommunications Conference (GLOBECOM 98), vol. 4, Sydney, Australia, 1998,
pp. 20682073.

You might also like